Email alert
2025 Vol. 47, No. 3
Display Method:
2025, 47(3): 582-589.
doi: 10.11999/JEIT250203
Abstract:
Objective Wireless positioning technology has gained increasing attention in the Internet of Things (IoT), Intelligent Transportation Systems (ITS), and Location-Based Services (LBS). However, Non-Line-of-Sight (NLOS) errors remain a major obstacle to positioning accuracy. When Line-of-Sight (LOS) propagation between mobile and static sensors is blocked by obstacles, ranging measurement errors increase substantially. Suppressing or mitigating NLOS errors is therefore essential for improving wireless positioning performance. Although existing approaches—such as Kalman filtering, hybrid Time Difference of Arrival (TDOA)/Angle of Arrival (AOA) algorithms, and reinforcement learning—have shown some effectiveness, each faces limitations. Algorithm performance can be affected by network topology, lack adaptability in complex environments, or require high computational costs. Moreover, the statistical behavior of NLOS errors remains poorly characterized, making accurate positioning difficult in large-scale settings. This study proposes an Iterative Weighted Least Squares (IWLS) algorithm based on Time of Arrival (TOA) measurements. By defining position residuals and incorporating a residual-based weighting strategy into the WLS framework, the method suppresses NLOS errors effectively. Compared with traditional approaches, the proposed algorithm achieves higher positioning accuracy and better adaptability in NLOS scenarios, while retaining the ease of implementation offered by TOA-based techniques. Methods This study defines a new position residual based on TOA measurements from two Mobile Sensors (MSs). The residual typically approaches zero under LOS conditions but tends to increase significantly under NLOS conditions. As this residual effectively reflects deviations induced by NLOS errors, it is used to assign weights to individual equations within the linear positioning system. A residual-based weighting strategy is proposed, in which each weight is computed from the corresponding position residual, and the Weighted Least Squares (WLS) method is applied to regulate the influence of each equation. The position is estimated by iteratively updating the residuals, computing the associated weights, and applying WLS, thereby progressively reducing the positioning error and yielding an accurate estimate of the MS location. Results and Discussions The performance of the proposed algorithm is evaluated through computer simulations under varying Signal-to-Diffraction Ratio (SDR) and maximum NLOS error (NLOSmax) conditions. The simulation results indicate the following: (1) When the number of NLOS-affected static nodes is two, the Cumulative Distribution Function (CDF) of positioning error for the proposed IWLS algorithm is below 92%@5m, outperforming other tested algorithms and maintaining a consistent advantage (Fig. 6 ). (2) In the NLOSmax scenario (Fig. 7 ), the IWLS algorithm achieves better positioning accuracy than conventional methods when the number of NLOS-affected nodes is small. As this number increases, the error of the proposed algorithm grows more gradually. (3) In the SDR scenario (Fig. 8 ), although all algorithms show degraded performance as SDR increases, the IWLS algorithm consistently yields the lowest Root Mean Square Error (RMSE) and remains closest to the Cramér-Rao Lower Bound (CRLB). Conclusions This study proposes an IWLS localization algorithm inspired by the relationship between position residuals and the reliability of localization equations. A position residual is defined using range measurements from two static sensors, and a residual-based weighting strategy is developed to suppress the influence of NLOS errors. During each iteration, the weighting vector downregulates the contribution of equations affected by large NLOS errors, thereby improving positioning accuracy. Simulation results show that the IWLS algorithm outperforms conventional localization methods under NLOS conditions and achieves RMSE values close to the CRLB. Notably, when two static sensors are affected by NLOS errors, the localization RMSE can be reduced to approximately 2 m, representing 2% of the coverage radius.
2025, 47(3): 573-581.
doi: 10.11999/JEIT240804
Abstract:
RF-based drone detection is an essential method for managing non-cooperative drones, with Drone Individual Recognition (DIR) via RF signals being a key component in the detection process. Given the current scarcity of DIR datasets, this paper proposes an open-source DroneRFb-DIR dataset for RF-based DIR. The dataset is constructed by capturing RF signals exchanged between drones and their remote controllers using a Software-Defined Radio (SDR). It includes signals from six types of drones, each with three different individuals, as well as background signals from urban environments. The captured signals are stored in raw I/Q format, and each drone type consists of over 40 signal segments, with each segment containing more than 4 million sample points. The RF sampling range spans from 2.4 GHz to 2.48 GHz, covering Flight Control Signals (FCS), Video Transmission Signals (VTS), and interference from surrounding devices. The dataset is annotated with entity identifiers (e.g., drone type and individual) and environmental labels (line-of-sight vs. non-line-of-sight). A DIR method based on fast frequency estimation and time-domain correlation analysis is also proposed and validated using this dataset. Objective: Drones are increasingly used in sectors such as geospatial mapping, aerial photography, traffic monitoring, and disaster relief, playing a significant role in modern industries and daily life. However, the rise in unauthorized drone operations presents serious threats to national security, public safety, and privacy, especially in urban areas. While existing methods emphasize general drone detection and classification, they struggle to distinguish individual drones of the same type, which is crucial for distinguishing friend from foe, analyzing swarm dynamics, and implementing effective countermeasures. This study addresses this gap by introducing the DroneRFb-DIR dataset, a large-scale, open-source RF signal dataset for non-cooperative DIR. Additionally, a novel method based on fast frequency estimation and time-domain correlation analysis is proposed to achieve accurate drone identification in urban environments. Methods: The DroneRFb-DIR dataset is developed using SDR device to capture RF signals in an urban environment with interference from devices like Wi-Fi and Bluetooth. It includes signals from six drone types, each with three individual units, as well as background reference signals. The dataset is collected at an 80 MHz sampling rate in the 2.4～2.48 GHz band and stored in raw I/Q format for detailed analysis. Each signal is annotated with identifiers (e.g., drone type and individual) and scene labels (line-of-sight and non-line-of-sight). For algorithm validation, the dataset is partitioned into training and testing sets. The proposed method consists of three key stages: (1) Signal Detection: A dynamic bandpass or band-stop filter isolates drone control signals from background noise and interference. (2) Frequency Localization: Adaptive filtering and frequency estimation to identify the spectral location of drone signals. (3) Identity Feature Extraction: Correlation analysis extracts identity features from control signal segments to differentiate individual drones, focusing on unique frequency modulation patterns. Results and Discussions: The dataset comprises 4,690 signal segments, each containing with over 4 million sample points. Experiments demonstrated the effectiveness of the proposed method (Table 3), showing high rejection rates of background signals and accurate identification of specific drone types. However, performance varied across drone types due to factors such as signal quality, environmental interference, and control signal characteristics. For instance, drones with low-SNR signals or less distinct frequency modulation patterns posed greater challenges for identification. Despite these difficulties, the method achieved competitive accuracy in identifying individual drones, even in non-line-of-sight conditions. These findings underscore the importance of advanced filtering and feature extraction for robust DIR in complex urban environments. Conclusions: This study addresses the critical need for DIR technologies by introducing the DroneRFb-DIR dataset and a novel identification method. Featuring six drone types, 18 individual drones, and one background signal class, the dataset is the first large-scale open-source resource for non-cooperative DIR in urban scenarios (Table 2). The proposed method effectively separates drone signals from interference and accurately identifies individual drones. Future work will focus on expanding the dataset with more diverse drone types, additional environmental scenarios (e.g., multipath interference and dynamic drone states), and machine learning models for improved recognition. Optimization of non-learning methods will also be explored to enhance feature extraction and identification rates, especially for drones with weaker signal characteristics.
2025, 47(3): 590-599.
doi: 10.11999/JEIT240840
Abstract:
Objective This research develops an efficient Bayesian Expectation Propagation (EP) detection method for Differential Spatial Modulation (DSM) systems using Multi-Phase Shift Keying (MPSK). DSM systems are notable for their advantage of not requiring Channel State Information (CSI), yet signal detection complexity remains a significant challenge. The detection problem is reformulated as a parameter estimation task, where a prior and a posterior distribution parameters are iteratively estimated to improve detection accuracy. By decoupling antenna-domain detection from constellation-domain information, computational complexity is reduced while maintaining high performance. Additionally, the traditional EP method is extended to account for variable noise variance, dynamically adjusting the noise term’s second-order estimate to enhance robustness. This research is essential for improving the practical applicability and performance of DSM systems, enabling efficient, low-complexity signal detection in modern wireless communication networks. Methods This research applies an EP approach to enhance the detection of DSM signals. The detection process is reformulated as a parameter estimation problem, where the a priori and a posteriori distribution parameters of the antenna domain and constellation domain are iteratively optimized. The EP algorithm decouples these domains, allowing independent iterative detection of antenna indices and optimal demodulation of constellation bits. This method effectively reduces computational complexity compared to existing detection schemes. Additionally, the traditional EP algorithm is extended by incorporating a variable noise variance mechanism. The second-order moment estimation of noisy random vectors is refined iteratively, improving detection robustness under varying noise conditions. Simulation experiments are conducted to evaluate the proposed scheme, and the results demonstrate superior detection performance and faster convergence across different system configurations. Results and Discussions Three detection algorithms—Zero-Forcing (ZF) detection, Minimum Mean Square Error (MMSE) detection, and Soft-input Soft-output (SISO) detection—are selected for performance comparison . Bit Error Rate (BER) comparisons for 3×3 (Figure 1 ), 4×4 (Figure 2 ), and 5×5 (Figure 3 ) antenna configurations are presented. Simulation results show that the proposed EP algorithm maintains similar BER performance across different antenna configurations, offering an advantage over existing linear schemes. Using a 4×4 MIMO antenna configuration, the proposed EP method outperforms the MMSE linear detection scheme across various modulation orders, with a significant performance gain observed from QPSK to 16PSK (Figure 4 ). Regardless of the antenna configuration, BER performance remains nearly unchanged after 1～3 iterations, with rapid convergence. Compared to a single iteration, three iterations provide a performance gain of approximately 1.5 dB (Figure 5 ). A comparison of BER performance between the constant noise variance in traditional EP and the non-uniform variance proposed in this study (Figure 6 ) shows that the non-uniform noise correction method outperforms the traditional approach, validating the effectiveness of the noise vector correction. Conclusions A detection algorithm based on Bayesian EP is proposed for use in DSM systems. The antenna domain and signal domain are estimated through iterative updates of the a prior and a posterior distribution parameters. The proposed algorithm outperforms traditional linear detection methods in terms of performance while offering lower complexity compared to conventional high-complexity maximum likelihood detection. Additionally, it can be extended to joint detection and decoding systems for enhanced performance.
2025, 47(3): 600-611.
doi: 10.11999/JEIT240874
Abstract:
Objective The Internet of Vehicles (IoV) is a global innovation focus, enabling ubiquitous interconnection among vehicles, roads, and people, thereby reducing traffic congestion and improving traffic safety. Vehicle-to-Vehicle (V2V) communication represents one of the most prominent application scenarios in IoV. This study addresses the reduced efficiency of V2V communication caused by environmental obstacles such as buildings and trees. It proposes the deployment of Reconfigurable Intelligent Surfaces (RIS) on Unmanned Aerial Vehicles (UAVs), leveraging their high mobility and on-demand deployment capability to enhance V2V communication under 6G networks. The model improves communication link quality and stability by utilizing the reflective properties of aerial RIS to mitigate signal attenuation and interference. This research develops a geometry-based Three-Dimensional (3D) dynamic channel model that incorporates the effects of UAV rotation, trajectory movement, and attitude changes on channel characteristics, enabling adaptation to dynamic and non-stationary communication scenarios. The findings provide a theoretical foundation for designing and optimizing RIS-assisted wireless communication systems through statistical analyses in the temporal, spatial, and frequency domains. Methods RIS can regulate incident electromagnetic waves to optimize communication system performance and are regarded as a crucial innovation in Sixth Generation (6G) wireless communication technology. Deploying RIS on UAVs effectively addresses reduced information transmission efficiency caused by obstacles such as trees and buildings, leveraging UAVs’ flexible trajectories and on-demand deployment capabilities. This study proposes a geometry-based 3D dynamic channel model, considering the UAV’s trajectory, three degrees of rotational freedom (pitch, yaw, and roll angles), and attitude changes. Channel propagation components are divided into aerial RIS array components and Non-Line-of-Sight (NLoS) components. Each RIS unit is modeled as an independent reflector capable of altering the propagation path by adjusting its phase and amplitude. The model incorporates time-varying spatial phases and Doppler frequency shifts, capturing the characteristics of dynamic propagation environments. Mathematical expressions for the Complex Impulse Responses (CIRs) are derived, along with analytical formulas for spatial Cross-Correlation Functions (CCFs), temporal Auto-Correlation Functions (ACFs), Frequency Correlation Functions (FCFs), and channel capacity. Various V2V communication scenarios are simulated by adjusting the velocity, direction, and acceleration of transmitters, receivers, and UAVs. Numerical simulations validate the proposed model’s effectiveness by defining four UAV trajectories and various vehicle motion states. Additionally, the temporal, spatial, and frequency correlation characteristics under different motion states are investigated. Finally, the effects of RIS physical attributes, such as the number and size of units, and UAV altitude on channel capacity are analyzed, along with dynamic variations in the power delay profile. Results and Discussions Simulation results demonstrate that the proposed channel model accurately captures channel characteristics. Specifically, the model presents various UAV flight trajectories (Fig. 5 ) and analyzes the temporal autocorrelation properties under different motion states of the transmitter and receiver (Fig. 6 ). It is observed that the temporal correlation exhibits significant non-stationarity across different motion states. However, the introduction of RIS significantly mitigates the decline in correlation. The model also compares the temporal autocorrelation properties corresponding to different UAV flight attitudes and altitudes (Fig. 7 , Fig. 9 ). It is found that as the UAV’s initial altitude increases, multipath effects decrease, and the rate of decline in temporal autocorrelation function values gradually slows. Subsequently, the spatial cross-correlation of the proposed channel model is investigated for different propagation paths, revealing an increase in correlation with the Rician factor (Fig. 8 ). The frequency correlation function values are also examined under varying distances between the transmitter and receiver (Fig. 10 ), showing that while the correlation declines, it gradually stabilizes as the frequency interval increases. Finally, the impact of the RIS’s physical properties on channel capacity and the power delay profile is studied (Fig. 11 , Fig. 12 ). It is observed that increasing the size and number of RIS array elements enhances channel capacity. Additionally, as delay increases, the power exhibits multiple smaller peaks before gradually decaying. These findings provide a valuable theoretical foundation for the future design and optimization of RIS-assisted wireless communication systems. Conclusions This paper presents a geometry-based 3D non-stationary channel model for V2V communications, innovatively incorporating aerial RIS implemented by UAVs equipped with RIS. The model accounts for the time-varying motion trajectories of ground vehicle terminals and UAVs, as well as the fading effects due to UAV attitude variations. Analytical expressions for spatiotemporal-frequency correlation functions and channel capacity are derived from the proposed model, ensuring the accuracy of channel transmission characteristics. By adjusting the model’s parameter configurations, it can accurately characterize the effects of various motion trajectories, dynamic states, UAV flight altitudes, and rotational angles on channel properties. These findings provide valuable insights for the design and performance analysis of RIS-assisted V2V communication systems.
2025, 47(3): 612-622.
doi: 10.11999/JEIT240596
Abstract:
Objective The resource scheduling in Low Earth Orbit (LEO) satellite communication systems using Beam Hopping (BH) technology is a continuous, long-term allocation process. Unlike geostationary earth orbit (GEO) satellites, LEO satellites exhibit high-speed mobility relative to the ground during communication. The design of BH patterns typically occurs within regular time windows, ranging from tens to hundreds of milliseconds, leading to the switching of satellite-to-cell interaction links during certain BH periods. This switching implies that cells migrate between satellite coverage areas, each with varying capacity and delay requirements, which inevitably affects the performance of the receiving satellite. Additionally, the requirements of migrating cells during the switching time slot are directly related to the resource tilt provided by the source satellite before the switch. Therefore, there is a strong correlation between the BH pattern design strategies for different satellites, requiring multi-satellite joint resource scheduling to maintain service quality of cells in regions affected by migration. Methods In order to characterize the demands of joint scheduling for multiple satellites and maximize the minimum traffic satisfaction rate, an optimization problem is proposed for dynamic scenarios involving satellite-to-cell interaction link switching. This optimization problem simultaneously considers co-channel interference, traffic demands with differentiated temporal and spatial distributions, and traffic delay—all factors that affect the service quality of BH systems. To solve this NP-hard problem, a design algorithm of Multi-Satellite Joint BH Pattern based on Resource Adaptive Tradeoff Allocation (RATMJ-BHP) is proposed. First, an inter-satellite joint scheduling framework is proposed to model the complex impact of cell migration on satellite resource scheduling, transforming the multi-satellite scheduling problem into a single-satellite BH pattern design problem. Then, within this framework, a weight design method for multi-satellite joint scheduling is proposed, which quantifies the intensity of service urgency based on the capacity and delay requirements of cells. Finally, this joint scheduling weight is used to design the BH pattern. Results and Discussions Based on the optimization problem modeled in this paper, the satellite optimization region is divided into two areas: the stable region and the immigration region. A comprehensive evaluation, considering both regions within individual satellites and across adjacent satellites, is essential for analyzing the performance of the proposed algorithm. Thus, this paper examines the simulation results from two perspectives: the minimum traffic satisfaction rate and the variation in the minimum traffic satisfaction rate across different regions. Additionally, convergence speed is a key indicator of the algorithm’s performance; therefore, the number of iterations required to produce results for each time slot is counted. The key contributions of this research are as follows: Firstly, the average and maximum convergence times of the proposed algorithm are significantly lower than those of the enumeration method, demonstrating its efficiency in terms of time complexity (Table 3 ). Specifically, with three satellites, the maximum complexity value of the proposed algorithm is 39.05, compared to that for the enumeration method. Secondly, the proposed algorithm outperforms the comparison algorithms in terms of minimum traffic satisfaction rates under different load rates, with a minimum value above 69.34% across various satellites (Fig. 3a ) (Fig. 4a ) (Fig. 5a ). These results show that the RATMJ-BHP algorithm effectively ensures high traffic satisfaction rates for cells in affected regions, demonstrating robustness across different traffic demand rates. Thirdly, the proposed algorithm exhibits a smaller disparity in minimum traffic satisfaction rates across regions, with values remaining close to zero, unlike other algorithms. This indicates its ability to maintain high traffic satisfaction rates for most cells in service areas (Fig. 3b ) (Fig. 4b ) (Fig. 5b ). Finally, simulation results from both perspectives demonstrate consistent performance across different satellites and varying traffic demand rates, highlighting the general applicability of the proposed algorithm in LEO satellite BH systems. Conclusions This paper addresses the design of BH patterns for dynamic scenarios involving satellite-to-cell interaction link switching. To meet the demands of multi-satellite joint resource scheduling in such scenarios, while considering performance factors such as co-channel interference, traffic demands, and traffic delay, the RATMJ-BHP algorithm is proposed. Simulation results show that the proposed algorithm effectively ensures the service quality of cells in migration-affected areas, and its lightweight design demonstrates broad applicability within LEO constellations. This paper contributes to the design strategy of BH patterns in dynamic scenarios during long-term resource scheduling processes, offering a solution to maintain continuous high-quality service to cells throughout prolonged satellite motion. It provides a reference for the design of long-term beam scheduling strategies in LEO satellite BH systems. However, several challenges remain in resource scheduling strategies for LEO satellite BH systems. For instance, the relationship between resource scheduling across BH periods and its impact on long-term system performance has yet to be fully explored. Additionally, while the proposed algorithm focuses on resource scheduling for the forward link of LEO satellite systems, further research is needed for uplink scenarios.
2025, 47(3): 623-632.
doi: 10.11999/JEIT240732
Abstract:
A robust beamforming method based on imperfect Channel State Information (CSI) is proposed for dense Low-Earth Orbit (LEO) satellite network-assisted terrestrial wireless communication systems to enhance spectral efficiency. Specifically, in scenarios where multiple LEO satellites use full frequency reuse, a multi-LEO satellite downlink sum rate maximization problem is formulated, considering constraints on satellite transmit power, satellite-User Terminal (UT) association, and satellite feeder link capacity. To solve the optimization problem, it is decomposed into two subproblems: satellite-UT association and satellite transmit beamforming. Weighted minimum mean-squared error and successive convex approximation methods are then employed to address the non-convex challenges. Simulation results confirm that the proposed multi-satellite full frequency reuse scheme and robust beamforming design effectively improve system throughput, even under non-ideal channel conditions. Objective As the LEO satellite constellation becomes denser, spectrum resources will become scarcer, and co-channel interference among satellites will intensify. Therefore, transmission methods with higher spectrum efficiency are needed. To mitigate the effects of severe satellite-terrestrial wireless channels and enhance system throughput, multi-beam beamforming and phased-array antennas are employed in LEO satellites to achieve higher antenna gain. However, most existing studies assume perfect knowledge of CSI, which is often impractical. Therefore, considering the complex satellite-terrestrial channel conditions and channel estimation errors, a robust beamforming method is preferable. Under dense satellite constellations, the increasing number of satellites and the presence of inter-satellite co-channel interference make the design of robust transmission methods for multi-LEO satellite networks essential. Thus, this paper aims to investigate the design of an efficient robust beamforming method for dense LEO satellite networks under given channel uncertainty. Methods To enable frequency reuse across multiple LEO satellites, this paper proposes a dense LEO satellite network architecture that incorporates a gateway or Geostationary Earth Orbit (GEO) satellite as the centralized controller. In this system architecture, multiple LEO satellites can reuse spectrum, thereby improving spectral efficiency. Additionally, a system sum-rate maximization problem is formulated, considering imperfect Angular-Of-Arrival (AoA) based CSI. The problem incorporates constraints on satellite-user association, multi-satellite downlink transmit beamforming, and satellite feeder link capacity. Results and Discussions The simulation results show that the system sum-rate increases with the satellite transmit power budget, as higher transmit power improves received signal quality (Fig. 3 ). Additionally, the proposed robust beamforming method significantly enhances the system sum-rate compared to existing methods (Fig. 3 ). Furthermore, the results indicate that the communication rate of the UT is constrained by satellite feeder link capacity, and higher feeder link capacity leads to an increase in the system sum-rate (Fig. 4 ). However, the rate of increase in the system sum-rate slows once the satellite feeder link capacity exceeds a certain threshold. The results also reveal that larger AoA uncertainty reduces the system sum-rate, highlighting the significant impact of AoA uncertainty on system performance (Fig. 5 ). Lastly, increasing the number of antennas effectively improves channel quality and further increases the system sum-rate (Fig. 6 ). Conclusions This paper investigates a robust beamforming method for dense LEO satellite networks and proposes a full frequency reuse scheme to enhance spectral efficiency and increase system throughput. Given the challenges in obtaining accurate CSI, an angular-information-based channel uncertainty model is adopted to reflect non-ideal channel conditions. A system sum-rate maximization problem is then formulated to evaluate system performance, considering satellite-UT association and satellite transmit beamforming. To address the non-convex optimization problem, the WMMSE and SCA methods are employed. Simulation results demonstrate that channel uncertainty significantly impacts system performance. When channel uncertainty is small, the performance gap between the proposed robust beamforming method and the ideal CSI case is minimal. Furthermore, the proposed multi-LEO satellite robust beamforming method outperforms other existing schemes.
2025, 47(3): 633-644.
doi: 10.11999/JEIT240344
Abstract:
Objective The rapid development of wireless communication and the Internet of Things (IoT) has led to significant growth in compute-intensive and delay-sensitive applications, which impose stricter latency requirements. However, local devices often face challenges in meeting these demands due to limitations in storage, computing power, and battery life. Mobile Edge Computing (MEC) has emerged as a key technology to address these issues. Despite its potential, the dynamic and complex nature of edge networks presents significant challenges in task offloading and resource allocation. DIgital Twin Edge Networks (DITEN), which map digital twins to physical devices in real-time, offer a promising solution. By integrating MEC with Digital Twin (DT) technology, this approach not only alleviates resource limitations in devices but also optimizes resource allocation in the digital domain, minimizing physical resource waste. This paper tackles the End-to-End (E2E) optimization problem in the offloading, computation, and result feedback process within edge computing networks. A DT-assisted joint task offloading, device association, and resource allocation scheme is proposed for E2E delay optimization, providing theoretical support for improving resource utilization in edge networks. Methods The optimization problem in this paper involves a non-convex objective function with both binary and continuous constraints, making it a mixed integer non-convex problem. To address this, the original problem is decomposed into four subproblems: computation and communication resource optimization, device association optimization, offloading decision optimization, and transmission bandwidth optimization. Within the Alternating Optimization (AO) framework, the Internal Convex Approximation (ICA) method is applied to convert the non-convex problem into a convex one. Additionally, the many-to-one matching problem is transformed into a one-to-one matching problem, and the Hungarian Algorithm (HA) is employed to solve the device association subproblem. Finally, the ICA-HA-AO is proposed to address the E2E delay optimization problem effectively. Results and Discussions The ICA-HA-AO algorithm approximates non-convex constraints as convex ones through constraint transformation and iteratively solves the original problem, determining optimal strategies for task offloading, device association, and resource allocation. Simulation results show that the ICA-HA-AO algorithm achieves optimal performance across varying task resource requirements, bandwidth, edge processing rates, and task volumes. Compared to the worst-performing benchmark scheme, delays are reduced by approximately 0.8 s, 1.5 s, 0.5 s, and 1.2 s, respectively (Fig. 5 –Fig. 8 ). As the DT deviation increases, the delay also increases more significantly, with a rise of about 0.13 s when the deviation increases from 0.01 to 0.02, emphasizing the importance of setting the DT deviation (Fig. 9 ). When the number of devices remains constant and the number of Access Points (APs) increases, the delay continues to decrease, highlighting the significance of AP deployment in practice. Additionally, when the number of APs remains fixed and the number of devices increases, the delay increases accordingly. However, the ICA-HA-AO algorithm effectively controls the rate of delay increase. For instance, when the number of devices is 10, 15, and 20, the delay increase is reduced from 0.39 s to 0.21 s (Fig. 10 ). These results demonstrate that the ICA-HA-AO algorithm can more efficiently utilize and schedule resources, achieving optimal resource allocation. Conclusions This paper investigates the joint optimization problem of task offloading, device association, and resource allocation in DITEN. Firstly, within the edge computing network, physical and DT models are established for a network comprising sensors, edge servers, and actuators. A comprehensive task model is developed, and the E2E delay for tasks is derived. The optimization problem for minimizing E2E delay is then formulated, subject to constraints such as power and energy consumption. Secondly, to solve the proposed mixed integer non-convex optimization problem, the original problem is decomposed into four subproblems. Based on the ICA and HA methods, an ICA-HA-AO algorithm is proposed to solve the problem iteratively. Finally, simulation results demonstrate that the proposed ICA-HA-AO algorithm significantly reduces E2E delay and outperforms benchmark schemes. Future work may explore integrating this method with techniques to improve spectrum utilization, thereby further enhancing spectrum efficiency and overall performance in DITEN systems.
2025, 47(3): 645-656.
doi: 10.11999/JEIT240741
Abstract:
Objective Federated Learning (FL) represents a distributed learning framework with significant potential, allowing users to collaboratively train a shared model while retaining data on their devices. However, the substantial differences in computing, storage, and communication capacities across FL devices within complex networks result in notable disparities in model training and transmission latency. As communication rounds increase, a growing number of heterogeneous devices become stragglers due to constraints such as limited energy and computing power, changes in user intentions, and dynamic channel fluctuations, adversely affecting system convergence performance. This study addresses these challenges by jointly incorporating assistance mechanisms and reducing device overhead to mitigate the impact of stragglers on model accuracy and training latency. Methods This paper designs an FL architecture integrating joint edge-assisted training and adaptive sparsity and proposes an adaptively sparse FL optimization algorithm based on edge-assisted training. First, an edge server is introduced to provide auxiliary training for devices with limited computing power or energy. This reduces the training delay of the FL system, enables stragglers to continue participating in the training process, and helps maintain model accuracy. Specifically, an optimization model for auxiliary training, communication, and computing resource allocation is constructed. Several deep reinforcement learning methods are then applied to obtain the optimized auxiliary training decision. Second, based on the auxiliary training decision, unstructured pruning is adaptively performed on the global model during each communication round to further reduce device delay and energy consumption. Results and Discussions The proposed framework and algorithm are evaluated through extensive simulations. The results demonstrate the effectiveness and efficiency of the proposed method in terms of model accuracy and training delay.The proposed algorithm achieves an accuracy rate approximately 5% higher than that of the FL algorithm on both the MNIST and CIFAR-10 datasets. This improvement results from low-computing-power and low-energy devices failing to transmit their local models to the central server during multiple communication rounds, reducing the global model’s accuracy (Table 3 ).The proposed algorithm achieves an accuracy rate 18% higher than that of the FL algorithm on the MNIST-10 dataset when the data on each device follow a non-IID distribution. Statistical heterogeneity exacerbates model degradation caused by stragglers, whereas the proposed algorithm significantly improves model accuracy under such conditions (Table 4 ).The reward curves of different algorithms are presented (Fig. 7 ). The reward of FL remains constant, while the reward of EAFL_RANDOM fluctuates randomly. ASEAFL_DDPG shows a more stable reward curve once training episodes exceed 120 due to the strong learning and decision-making capabilities of DDPG and DQN. In contrast, EAFL_DQN converges more slowly and maintains a lower reward than the proposed algorithm, mainly due to more precise decision-making in the continuous action space and an exploration mechanism that expands action selection (Fig. 7 ).When the computing power of the edge server increases, the training delay of the FL algorithm remains constant since it does not involve auxiliary training. The training delay of EAFL_RANDOM fluctuates randomly, while the delays of ASEAFL_DDPG and EAFL_DQN decrease. However, ASEAFL_DDPG consistently achieves a lower system training delay than EAFL_DQN under the same MEC computing power conditions (Fig. 9 ).When the communication bandwidth between the edge server and devices increases, the training delay of the FL algorithm remains unchanged as it does not involve auxiliary training. The training delay of EAFL_RANDOM fluctuates randomly, while the delays of ASEAFL_DDPG and EAFL_DQN decrease. ASEAFL_DDPG consistently achieves lower system training delay than EAFL_DQN under the same bandwidth conditions (Fig. 10 ). Conclusions The proposed sparse-adaptive FL architecture based on an edge-assisted server mitigates the straggler problem caused by system heterogeneity from two perspectives. By reducing the number of stragglers, the proposed algorithm achieves higher model accuracy compared with the traditional FL algorithm, effectively decreases system training delay, and improves model training efficiency. This framework holds practical value, particularly for FL deployments where aggregation devices are selected based on statistical characteristics, such as model contribution rates. Straggler issues are common in such FL scenarios, and the proposed architecture effectively reduces their occurrence. Simultaneously, devices with high model contribution rates can continue participating in multiple rounds of federated training, lowering the central server’s frequent device selection overhead. Additionally, in resource-constrained FL environments, edge servers can perform more diverse and flexible tasks, such as partial auxiliary training and partitioned model training.
2025, 47(3): 657-665.
doi: 10.11999/JEIT240447
Abstract:
Objective With the rapid development of wireless communication technologies, the demand for spectrum resources has significantly increased. Cognitive Radio (CR) has emerged as a promising solution to improve spectrum utilization by enabling Secondary Users (SUs) to access licensed spectrum bands without causing harmful interference to Primary Users (PUs). However, traditional CR networks face challenges in achieving high spectral efficiency due to limited control over the wireless environment. Intelligent Reflecting Surfaces (IRS) have recently been introduced as a revolutionary technology to enhance communication performance by dynamically reconfiguring the propagation environment. This paper aims to maximize the sum rate of SUs in an IRS-assisted CR network by jointly optimizing the active beamforming at the Secondary Base Station (SBS) and the passive beamforming at the IRS, subject to constraints on the maximum transmit power of the SBS, the interference tolerance of PUs, and the unit modulus of the IRS phase shifts. Methods To address the non-convex and highly coupled optimization problem, a Deep Reinforcement Learning (DRL)-based algorithm is proposed. Specifically, the problem is formulated as a Markov Decision Process (MDP), where the state space includes the Channel State Information (CSI) of the entire system, the Signal-to-Interference-plus-Noise Ratio (SINR) in the SU network, and the action space consists of the SBS beamforming vectors and the IRS phase shift matrix. The reward function is designed to maximize the sum rate of SUs while penalizing violations of the constraints. The Deep Deterministic Policy Gradient (DDPG) algorithm is used to solve the MDP, owing to its ability to handle continuous action spaces. The DDPG framework consists of an actor network, which outputs the optimal actions, and a critic network, which evaluates these actions based on the reward function. The training process involves interacting with the environment to learn the optimal policy, and the algorithm is fine-tuned to ensure convergence and robustness under varying system conditions. Results and Discussions Simulation results show that the proposed scheme achieves comparable sum rate performance with lower time complexity after optimization, compared to traditional optimization algorithms. The proposed algorithm significantly outperforms the no-IRS and IRS-random phase shift schemes (Fig. 5 ). The results demonstrate that the proposed algorithm achieves a sum rate close to that of alternating optimization-based approaches (Fig. 5 ), while substantially reducing computational complexity (Fig. 5 , Table 2 ). Additionally, the impact of the number of IRS elements on the sum rate is examined (Fig. 6 ). As expected, the average reward increases with the number of reflecting elements, while the convergence time remains stable, indicating the robustness of the proposed algorithm. The DRL-based algorithm, starting from the identity matrix, can learn and adjust the beamforming vectors and phase shifts to approach the optimal solution through interaction with the environment (Fig. 7 ). It is also observed that the variance of the instantaneous reward increases with the transmit power. This is due to the larger dynamic range of the instantaneous reward at higher power levels, resulting in greater fluctuations and slower convergence. The relationship between average reward and time steps under different transmit power levels is presented, highlighting the sensitivity of the algorithm to high signal-to-noise ratios (Fig. 8 ). Moreover, it can be observed that a learning rate of 0.001 yields the best performance, while excessively high or low learning rates degrade performance (Fig. 9 ). The discount factor has a relatively smaller impact on performance compared to the learning rate (Fig. 10 ). Conclusions This paper proposes a DRL-based algorithm for joint active and passive beamforming optimization in an IRS-assisted CR network. The algorithm utilizes the DDPG framework to maximize the sum rate of SUs while adhering to constraints on transmit power, interference, and IRS phase shifts. Simulation results demonstrate that the proposed algorithm achieves comparable sum rate performance to traditional optimization methods, with significantly lower computational complexity. The findings also highlight the impact of DRL parameter settings on performance. Future work will focus on extending the proposed algorithm to multi-cell scenarios and incorporating imperfect CSI to enhance its robustness in practical environments.
2025, 47(3): 666-677.
doi: 10.11999/JEIT240773
Abstract:
Objective To meet the differentiated service requirements of users in dynamic Edge Computing (EC) network scenarios, network slicing technology has become a crucial enabling approach for EC networks to offer differentiated edge services. It facilitates flexible allocation and customized management of communication and computation resources by dividing network resources into multiple independent sub-slices. However, traditional slicing resource management methods cannot handle the time-varying wireless channel conditions and the randomness of service arrivals in EC networks. Additionally, existing intelligent slicing resource management schemes based on deep reinforcement learning face challenges, including the need for extensive information sharing, privacy leakage, and unstable training convergence. To address these challenges, the integration of Multi-Agent Reinforcement Learning (MARL) and Federated Learning (FL) allows for experience sharing among agents while protecting users’ privacy. Furthermore, Generative Adversarial Network (GAN) is used to generate state-action value distributions, improving the ability of traditional MARL methods to learn state-value information. By modeling the joint bandwidth and computing slicing resource management optimization problem as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP), collaborative decision-making for slicing resource management is achieved by sharing only the generator network parameters of each agent through the combination of FL and GAN. This study provides a federated collaborative decision-making framework for addressing the slicing resource management problem in EC scenarios and offers theoretical support for enhancing the utilization efficiency of edge slicing resources while preserving users’ privacy. Methods The core concept of the proposed federated slicing resource management scheme is to first employ both GAN technology and the D3QN algorithm for local training within a multi-agent framework. The FL architecture is then used to share the generator network parameters of each agent, facilitating collaborative decision-making for joint bandwidth and computing slicing resource management. In this approach, each Access Point (AP) agent collects data on the total number of tasks to be transmitted and the number of Central Processing Unit (CPU) cycles required for computing tasks in each associated slice as local observations during each training time slot. Each agent subsequently selects the optimal local bandwidth and computing resource management action, obtaining the system reward, which consists of the average service waiting delay and service satisfaction rate, as well as the observation for the next time slot to train the local network. During the training process, each AP agent maintains its own main generator network, target generator network, and discriminator network. In each training episode, the D3QN algorithm is applied to decompose the state-action values, and GAN is used to perform multi-modal learning of the state value distribution, thus completing the local training. After each training episode, the AP agents upload their main generator network parameters for federated aggregation and receive the global main generator network parameters for the next training episode. Results and Discussions By employing the D3QN algorithm and integrating the advantages of GAN within the MARL framework, alongside leveraging FL to share learning experiences among agents while protecting users’ privacy, the proposed scheme reduces the long-term service waiting delay and improves the long-term average service satisfaction rate. Simulation results demonstrate that the proposed scheme achieves the highest average cumulative reward after approximately 500 episodes (Fig. 3 ), with a notable improvement of at least 10% in convergence performance compared to the baselines. Furthermore, the scheme strikes a better balance between average service waiting delay and average service satisfaction rate (Fig. 4 ). Additionally, it delivers superior performance in terms of user average service satisfaction rate, with at least an 8% improvement under varying user numbers (Fig. 5 ), highlighting its effectiveness in resource management under different task loads. Moreover, the proposed scheme reduces the average service waiting delay by at least 28% (Fig. 6 ) under varying numbers of agents. Conclusions This paper investigates the joint bandwidth and computing slicing resource management problem in dynamic, unknown EC network scenarios and proposes a federated slicing resource management scheme based on GAN-assisted MARL. The proposed scheme enhances the agents’ ability to learn state-value information and promotes collaborative learning by sharing the training network parameters of agents, which ultimately reduces long-term service waiting delays and improves long-term average service satisfaction rates, while protecting users’ privacy. Simulation results show that: (1) The cumulative reward convergence performance of the proposed scheme improves by at least 10% compared to the baselines; (2) The average service satisfaction rate of the proposed scheme is more than 8% higher than that of the baselines under varying user numbers; (3) The average service waiting delay of the proposed scheme is reduced by at least 28% compared to the baselines under varying agent numbers. However, this study only considers ideal, static user scenarios and interference-free communication conditions. Future work should incorporate more real-world dynamics, such as time-varying user mobility and complex multi-user interference.
Fast Sensing Method Based on Beam Squint and Beam Split of Terahertz Reflective Intelligent Surfaces
2025, 47(3): 678-686.
doi: 10.11999/JEIT240789
Abstract:
Objective Reflecting Intelligent Surface (RIS)-aided Terahertz (THz) communications are considered a key technology for future Sixth-Generation (6G) mobile communication systems addressing issues such as signal attenuation and Line-of-Sight (LoS) link blockage issues, due to their ultra-large bandwidth and low power consumption. However, the frequency independent characteristics of RIS elements can cause beam squint effects, where beams of different carriers are directed at different angles. Although this reduces the beam gain received by users, it can be leveraged to enhance sensing capabilities in sensing applications. Specifically, beam squint allows for simultaneous sensing of a target using multiple carrier beams directed in different directions. Existing studies have explored beam squint for beam training. For example, by studying near-field beam squint and True Time Delay (TTD) to generate beams that focus at multiple positions across different frequencies, enabling rapid beam training with reduced overhead. Additionally, combining TTD with beam squint and beam split for sensing extends the beam coverage area and enables the quick acquisition of user locations through feedback. However, there is no research on jointly utilizing beam squint and beam split for sensing in RIS-assisted THz systems. This paper aims to conduct detailed research on the use of beam squint for sensing in such systems. Methods To address the time-consuming issue of beam scanning in RIS-assisted THz systems, a fast sensing method based on RIS beam squint and split effects is proposed. Each RIS element is equipped with a TTD mechanism to dynamically adjust the degree of beam squint, while the large array RIS units are spaced to induce the beam split effect. By combining beam quint and beam split, the method enables rapid sensing of the target area. Specifically, the sensing area is divided into multiple sub-areas, with the TTD and the phase shift at the RIS elements optimized to cover each sub-area based on beam squint. The beam split effect is then used to seamlessly cover multiple sub-areas, significantly reducing time overhead compared to single beam scanning. To further mitigate echo signal path loss, active sensing elements are configured at the RIS for direct reception and analysis of the echo signals. The estimation of the sensing target’s angle, along with its Root Mean Square Error (RMSE), is derived based on this approach. Results and Discussions Consider the RIS-assisted THz sensing system model (Fig. 1 ). By deriving the channel and beam gain expressions, the beam patterns under the beam squint effect are analyzed (Fig. 2 ). Based on the internal structural diagram of the RIS (Fig. 4 ), the beam split effect is examined by varying the spacings between RIS elements (Fig. 5 ), with corresponding beam patterns (Fig. 3 ) presented for different spacings. Next, the RIS structure utilizing TTD (Fig. 6 ) allows for flexible adjustment of the beam squint and split degrees, significantly expanding the beam coverage area compared to traditional beam squint and split methods (Fig. 7 , Fig. 8 ). Additionally, to fine-tune the gaps between adjacent split beams, the ATDS method is proposed. By combining beam squint and beam split, this method achieves near-seamless coverage of all subareas (Fig. 9 ). Finally, the target direction is estimated by analyzing the echo signals received at the RIS-SE, based on the RSME. The simulation results demonstrate the relationship between sensing accuracy and the number of carriers (Fig.10 , Fig. 11 ), confirming the effectiveness and feasibility of the rapid sensing method combining beam squint and split. Conclusions This paper investigates the issues of beam squint and beam split in RIS-assisted THz systems and proposes a rapid sensing method that combines both effects. Specifically, TTD is used to adjust the direction of subcarrier beams based on beam squint. To expand the sensing area, the combined effects of beam squint and beam split, divide the sensing area into multiple subareas, which are simultaneously covered by multiple carrier beams within a single OFDM block. The target direction is then estimated based on echo signals received at the RIS-SE, with sensing error measured using the RMSE between the true and estimated values. Simulation results demonstrate the feasibility and effectiveness of the proposed rapid sensing method. However, it is found that while the beam squint effect significantly reduces beam gain and communication performance, it expands the beam coverage area and enhances sensing capabilities. Therefore, in an integrated sensing and communication system, the impact of beam squint should be considered at different stages. Future research will focus on improving the performance of such integrated systems.
2025, 47(3): 687-695.
doi: 10.11999/JEIT240631
Abstract:
Objective The rapid growth in the number of wireless communication devices has led to the expansion of frequency bands to higher frequencies, resulting in increased overlap between communication and radar systems. Dual-Functional Radar-Communication (DFRC), which shares spectrum resources on the same hardware platform, is an effective solution to address spectrum congestion. The integration of Multiple-Input Multiple-Output (MIMO) technology, which employs multi-antenna techniques, with DFRC is crucial for achieving both high-precision detection and large-capacity communication. Beamforming technology plays a key role in efficiently allocating resources between these two requirements, further enhancing the collaborative gain of DFRC systems. Beampattern gain, a critical performance metric for target detection, makes it essential to investigate beamforming designs that maximize this gain in MIMO DFRC systems. Methods An MIMO DFRC system is considered, which simultaneously achieves target detection and Multi-User (MU) communication. First, a beamforming problem is formulated to maximize the beampattern gain in the target direction, while satisfying MU Signal-to-Interference-plus-Noise Ratio (SINR) and total power constraints. To address this beamforming design problem, two methods based on Semidefinite Relaxation (SDR) and Majorization Minimization (MM) are proposed to solve for the transmit beamforming vectors. Specifically, the SDR-based method transforms the beamforming problem into a semidefinite programming problem by introducing auxiliary variables and relaxing the rank-one constraint. The MM-based method, on the other hand, uses the first-order Taylor expansion to construct a cost function from the objective function, transforms the SINR constraint into a second-order cone constraint, and iteratively solves the simplified problem. Results and Discussions The convergence curves of the SDR-based and MM-based beamforming design schemes are shown (Figure 2 ). The results indicate that the MM-based method can achieve almost the same beampattern gain as the SDR-based method. Under the same number of transmit antennas, a higher SINR threshold results in a smaller beampattern gain. This phenomenon reflects the performance trade-off between communication and radar in MIMO DFRC systems. Under the same SINR threshold, increasing the number of transmit antennas leads to a greater beampattern gain. This is because an increase in the number of transmit antennas provides additional degrees of freedom for the radar. The comparison of the single CVX running time of the SDR-based and MM-based methods under different numbers of transmit antennas is shown (Figure 3 ). The results demonstrate that the single CVX running time of the MM-based method is shorter than that of the SDR-based method for the same number of transmit antennas, and as the number of transmit antennas increases, the complexity reduction of the MM-based method becomes more significant than that of the SDR-based method. The variation curves of beampattern gain with SINR threshold for different numbers of transmit antennas in the MM-based and SDR-based methods are shown (Figure 4 ). The beampattern gain obtained by the MM-based method is slightly lower than that obtained by the SDR-based method. However, as the number of transmit antennas increases, the difference between the two methods gradually decreases. Moreover, the more transmit antennas there are, the greater the SINR achievable by the communication user. When the number of antennas is fixed, the relationship between beampattern gain and transmit SNR obtained by the radar using the MM-based method is presented (Figure 5 ). When the SINR threshold remains unchanged, the relationship between them is shown (Figure 6 ). The results illustrate that, compared with the radar-only scenario, the beampattern gain performance of MIMO DFRC systems is lower, and a larger SINR threshold results in a smaller beampattern gain. Additionally, within a certain range, when the transmit SNR is constant, beampattern gain is directly proportional to the number of transmit antennas. Conclusions This paper addresses the beamforming design problem for MIMO DFRC systems with the objective of maximizing beampattern gain. By jointly optimizing the communication and radar transmit beamforming vectors, the beampattern gain in the target direction is maximized while satisfying the SINR constraint for communication users and the total transmit power constraint. To solve this problem, the SDR-based and MM-based beamforming design methods are proposed. Simulation results demonstrate that the MM-based method offers lower complexity and achieves nearly the same beampattern gain as the SDR-based method. Moreover, as the number of transmit antennas increases, the complexity reduction of the MM-based method is more significant compared to the SDR-based method.
2025, 47(3): 696-706.
doi: 10.11999/JEIT240818
Abstract:
Objective Small target recognition on the sea surface is a critical and challenging task in maritime radar surveillance. The variety of small targets and the complexity of the sea surface environment make their classification difficult. Due to the small size of these targets, typically occupying only one or a few range cells under high-resolution radar systems, there is insufficient spatial scattering structure information for classification. The primary information for classification comes from the target’s Radar Cross Section (RCS) fluctuation and radial velocity change. This study proposes a classification network model based on multidomain radar echo data fusion, providing a theoretical foundation for small target recognition in complex sea surface environments. Methods A small marine target classification network model is proposed, based on multidomain radar echo data fusion, incorporating both time domain and time-frequency domain. Given that data from different domains hold distinct physical significance, a Time-domain LeNet (T-LeNet) neural network module and a time-frequency feature extraction neural network module are designed to extract features from the amplitude sequence and the Time-Frequency Distribution (TFD), respectively. The amplitude sequence primarily reflects the fluctuation characteristics of the target’s RCS, while the TFD captures both the RCS fluctuations and variations in the target’s radial velocity. By extracting deep information from small sea surface targets, effective differential features are obtained, leading to improved classification results. The advantages of the multidomain data fusion approach are validated through ablation experiments, where the amplitude sequence is fused with the input TFD, or the TFD is fused with the input amplitude sequence. Additionally, the effect of network depth on recognition performance is explored by using ResNet architectures with varying depths for time-frequency feature extraction. Results and Discussions A dataset containing four types of small sea surface targets is constructed using measured data to evaluate the effectiveness of the proposed method. Six evaluation metrics are used to assess the model’s classification ability. The experimental results show that when only the TFD is input, the best recognition performance is achieved by the ResNet18 network. This is due to ResNet18’s ability to prevent gradient vanishing and explosion through residual connections, enabling a deeper network capable of more effectively extracting differential features between targets. When only the amplitude sequence is input, the recognition performance of the T-LeNet network improves significantly compared to the performance with only the TFD input. Fusing the amplitude sequence with the T-LeNet network, based solely on the input of the TFD, leads to a notable increase in recognition performance. Thus, incorporating information from other domains, such as time-domain information (amplitude sequence), and extracting abstract features from one-dimensional data with T-LeNet, while also capturing deeper target features from multidomain and multidimensional aspects, significantly enhances the network’s recognition capability. The best recognition performance occurs when both the amplitude sequence and TFD are input using the ResNet18 network, achieving an accuracy of 97.21%, which represents a 21.1% improvement over the TFD-only input with the Vgg16 network (Table 3 ). The confusion matrix reveals that Class I and Class II targets are more accurately classified when using only the amplitude sequence, with average accuracy improvements of 5.5% and 85.1%, respectively, compared to the TFD-only input. Class IV targets are better classified when using only the TFD, with an average accuracy improvement of 5.5% compared to the amplitude sequence input. There is no significant difference in the accuracy of Class III targets (Fig. 5 ). Comparing the classification results of different ResNet networks shows that increasing the depth of the ResNet network does not significantly enhance recognition performance (Table 4 ). Analyzing the loss and accuracy of the various experiments in both the training and validation sets reveals that combining the T-LeNet network improves performance further. Specifically, the accuracy of AlexNet, Vgg16, and ResNet18 in the validation set improves by approximately 7.7%, 5.3% and 3.6%, respectively, while the loss in both the training and validation sets decreases (Fig. 6 ). Conclusions This paper proposes a small sea surface target classification method based on Convolutional Neural Networks (CNN) and data fusion. The method considers both the time domain and time-frequency domain, leveraging their distinct physical significance. It constructs the T-LeNet network module and the time-frequency feature extraction network module to extract deep information from small sea surface targets across multiple domains and dimensions. The abstract features jointly extracted from the time domain and time-frequency domain are fused for multidomain and multidimensional classification. The experimental results demonstrate that the proposed method exhibits strong recognition capability for small sea surface targets.
2025, 47(3): 707-719.
doi: 10.11999/JEIT240528
Abstract:
Objective Feature detection has become an effective approach for detecting small targets in sea clutter environments, attracting significant attention and research. Previous studies primarily focused on extracting differential features between targets and clutter from the current pulse frame for detection. Recent methods have integrated temporal information from multiple frames with current frame features, demonstrating improved detection performance. However, these methods rely on fixed-order Auto Regressive (AR) models, which do not effectively adapt to the time-varying nature of sea clutter. Moreover, the use of static weighting algorithms for feature fusion fails to account for clutter characteristics in the current scene, leading to suboptimal utilization of temporal information. To address these issues, this study proposes a feature AR modeling and one-step prediction method based on a model-stable modified Burg algorithm, enabling adaptive pole distribution adjustment and enhancing the accuracy of sea clutter feature prediction. Additionally, a dynamic weighting algorithm is developed by solving multivariable extreme value problems to obtain minimum variance fused features, fully leveraging historical frame temporal information and improving radar target detection performance. Methods This study employs a modified Burg method to predict sea clutter, incorporating a stability factor in the derivation of reflection coefficients to constrain the model’s poles within the unit circle. This enhances model stability, improving its adaptability to the time-varying nature of sea clutter and increasing the accuracy of feature prediction. A dynamic weighting algorithm is introduced to adaptively adjust fusion weights based on data volatility around the current frame by solving a multivariable extremum problem, thereby minimizing the local variance of fused features. Temporal fusion is performed using the features Relative Average Amplitude (RAA), Frequency Peak to Average Ratio (FPAR), and Relative Doppler Peak Height (RDPH) to generate a fused feature. The fused clutter features are then used to construct a three-dimensional convex hull decision region, where target presence is determined by assessing whether the detection unit’s feature point lies within this region. Detection results are compared with commonly used feature detection methods. Additionally, the study evaluates the boundary performance of the proposed method and contrasts it with the traditional energy-domain CFAR method, providing a comprehensive analysis of its usability and effectiveness. Results and Discussions The proposed method achieves the following results: (1) For clutter data, the temporal fusion algorithm reduces data variance by an average of 0.024 5 compared to no temporal fusion and by 0.003 5 compared to the original temporal fusion algorithm. For target data, it reduces data variance by an average of 1.126 6 compared to no temporal fusion and by 0.179 compared to the original temporal fusion algorithm. (2) The Bhattacharyya distance of the proposed temporal fusion algorithm improves by an average of 0.237 3 compared to no temporal fusion and by 0.109 3 compared to the original temporal fusion algorithm. Under VV polarization, the Bhattacharyya distance improves by an average of 0.219 9 compared to no temporal fusion and by 0.090 8 compared to the original temporal fusion algorithm. (3) The proposed method outperforms other feature detectors in detection performance by effectively utilizing temporal information from historical frames, thereby enhancing the echo information used. Compared to energy-domain CFAR methods, it maintains a strong competitive advantage. Conclusions This study presents innovative solutions to two key challenges in existing sea clutter feature modeling and fusion methods. First, to address the time-varying nature of sea clutter features, a model-stable modified Burg method is proposed for Autoregressive (AR) feature modeling. This approach enables adaptive adjustment of model pole distribution, improving the accuracy of one-step sea clutter feature predictions and simplifying model order estimation. Second, to enhance the utilization of inter-frame temporal information during feature fusion, a dynamic weighted fusion algorithm is introduced to integrate predicted and observed features. This method reduces the variance of fused features and fully exploits historical temporal information. Validation using the IPIX dataset and the shared dataset from the Naval Aeronautical University demonstrates that the fused features obtained through these methods exhibit improved separability compared to the original features, significantly enhancing detector performance.
2025, 47(3): 720-728.
doi: 10.11999/JEIT240517
Abstract:
Objective The Internet of Vehicles (IoV) plays a pivotal role in the development of modern intelligent transportation systems. It enables seamless communication among vehicles, road infrastructure, and pedestrians, thereby improving traffic management, enhancing driving experiences, and optimizing resource utilization. However, existing IoV systems face a range of complex and urgent challenges. A major issue is the high false positive rate in identifying malicious vehicles. These vehicles, intending to disrupt network operations, may engage in harmful activities such as dropping packets or delaying transmissions. This not only compromises data transmission integrity but also poses a serious threat to the overall security and reliability of the IoV network. Furthermore, inaccurate identification may lead to the wrongful penalization of legitimate vehicles, disrupting their normal operations. Another challenge stems from the diverse and complex service requirements within IoV. These range from entertainment services that enhance user experience, to traffic efficiency services aimed at optimizing traffic flow, and highly sensitive services related to traffic safety and privacy. Unfortunately, existing solutions fail to adequately address these varied needs, leading to suboptimal service delivery and potential security risks. Traditional consensus algorithms also face significant limitations in the dynamic IoV environment. The high resource consumption and low efficiency of these algorithms not only waste valuable computational resources but also hinder timely and accurate information processing, affecting the overall performance of the IoV system. To address these issues, it is critical to develop an innovative solution to enhance the security, reliability, and adaptability of IoV systems. This paper proposes a collaborative trust management scheme based on blockchain technology, which aims to address these challenges and improve the overall performance of IoV. Methods To address the challenges outlined above, a comprehensive set of methods is designed. First, a trust management model based on the Dirichlet distribution is developed. This model classifies vehicle trust and collaborative services into multiple levels, each representing a different degree of trustworthiness and service quality. The trust level thresholds for different service types are finely tuned. For example, traffic safety and privacy-related services, which require high security and reliability, are assigned higher trust level thresholds, ensuring that only vehicles with a sufficient trust level can provide these critical services. Second, a trust level evaluation algorithm integrated with a feedback mechanism is developed. This algorithm considers four key factors: the current state of the collaborating vehicle, neighbor recommendations, historical trust data, and service quality. The evaluation process occurs in two distinct but complementary stages: before and after collaboration.Before collaboration, the vehicle’s current state is thoroughly assessed, including its computing power, which determines its capacity to handle complex tasks; propagation delay, which indicates the timeliness of communication; and familiarity with the requesting vehicle, which can influence collaboration reliability. These factors, along with neighbor recommendations and historical trust data, contribute to an initial trustworthiness assessment. After collaboration, a feedback mechanism based on packet delivery ratio and time delay is applied. The packet delivery ratio measures the proportion of successfully delivered packets, while time delay reflects the responsiveness of the vehicle during communication. These metrics are used to adjust the vehicle’s trust level, providing a more dynamic and accurate evaluation of its trustworthiness. Third, the traditional Proof of Work (PoW) consensus algorithm is enhanced by introducing a task priority index. This dynamic adjustment of block creation difficulty for miner nodes allows blocks containing critical trust information or high-priority service data to be added to the blockchain more quickly. This enhancement improves blockchain efficiency. Results and Discussions The simulation results provide compelling evidence for the effectiveness of the proposed scheme. In terms of malicious vehicle identification, as shown in (Fig. 3 ), although the initial identification rate of malicious vehicles is slightly lower than that of some binary-evaluation-based schemes, the proposed scheme demonstrates a significant reduction in the false positive rate. The comparison of false positive rates, presented in (Fig. 4 ), clearly illustrates that the proposed scheme outperforms existing methods. This improvement is attributed to the carefully designed trust level thresholds, which prevent ordinary vehicles with low-quality services from being misclassified as malicious when performing high-level services. Regarding the collaboration success rate, (Fig. 5 ) indicates that the proposed scheme performs better across various service scenarios and different proportions of malicious vehicles. Even when the proportion of malicious vehicles reaches 50%, the collaboration success rate for the three-level services remains above 80%, emphasizing the robustness and reliability of the proposed scheme. In terms of consensus efficiency, as shown in (Fig. 6 ), the improved algorithm outperforms the traditional PoW consensus algorithm. By dynamically adjusting to the actual conditions, the enhanced algorithm allows the RoaSide Unit (RSU) responsible for the area to generate blocks more quickly when the task priority index is larger. This leads to faster processing of critical information and better alignment with the dynamic needs of the IoV collaborative scenario. Conclusions The collaborative trust management scheme based on blockchain proposed in this paper effectively addresses critical challenges in IoV systems, including malicious vehicle identification, service adaptability, and the applicability of consensus algorithms. By accurately classifying service types and vehicle trust levels, and by employing a comprehensive trust evaluation algorithm along with an enhanced consensus algorithm, this scheme significantly improves the security and trustworthiness of IoV systems. Furthermore, it provides a scalable solution for future IoV deployments, facilitating the broader adoption of IoV technology.
2025, 47(3): 729-738.
doi: 10.11999/JEIT240782
Abstract:
Objective In 2017, the PFP algorithm was introduced as an ultra-lightweight block cipher to address the demand for efficient cryptographic solutions in constrained environments, such as the Internet of Things (IoT). With a hardware footprint of approximately 1355 GE and low power consumption, PFP has attracted attention for its ability to deliver high-speed encryption with minimal resource usage. Its encryption and decryption speeds outperform those of the internationally recognized PRESENT cipher by a factor of 1.5, making it highly suitable for real-time applications in embedded systems. While the original design documentation asserts that PFP resists various traditional cryptographic attacks, including differential, linear, and impossible differential attacks, the possibility of undiscovered vulnerabilities remains unexplored. This study evaluates the algorithm’s resistance to related-key differential attacks, a critical cryptanalysis method for lightweight ciphers, to determine the actual security level of the PFP algorithm using formal cryptanalysis techniques. Methods To evaluate the security of the PFP algorithm, Satisfiability Modulo Theories (SMT) is used to model the cipher’s round function and automate the search for distinguishers indicating potential design weaknesses. SMT, a formal method increasingly applied in cryptanalysis, facilitates automated attack generation and the detection of cryptographic flaws. The methodology involved constructing mathematical models of the cipher’s rounds, which are tested for differential characteristics under various key assumptions. Two distinguisher models are developed: one based on single-key differentials and the other on related-key differentials, the latter being the focus of this analysis. These models automated the search for weak key differentials that could enable efficient key recovery attacks. The analysis leveraged the nonlinear substitution-permutation structure of the PFP round function to systematically identify vulnerabilities. The results are examined to estimate the probability of key recovery under different attack scenarios and assess the effectiveness of related-key differential cryptanalysis against the full-round PFP cipher. Results and Discussions The SMT-based analysis revealed a critical vulnerability in the PFP algorithm. A related-key differential characteristic with a probability of 2–62 is identified, persisting through 32 encryption rounds. This characteristic indicates a predictable pattern in the cipher’s behavior under related-key conditions, which can be exploited to recover the secret key. Such differentials are particularly concerning as they expose a significant weakness in the cipher’s resistance to related-key attacks, a critical threat in IoT applications where keys may be reused or related across multiple devices or sessions.Based on this finding, a key recovery attack is developed, requiring only 263 chosen plaintexts and 248 full-round encryptions to retrieve the 80-bit master key. The efficiency of this attack demonstrates the vulnerability of the PFP cipher to practical cryptanalysis, even with limited computational resources. The attack’s relatively low complexity suggests that PFP may be unsuitable for applications demanding high security, particularly in environments where adversaries can exploit related-key differential characteristics. Moreover, these results indicate that the existing resistance claims for the PFP cipher are insufficient, as they do not account for the effectiveness of related-key differential cryptanalysis. This challenges the assertion that the PFP algorithm is secure against all known cryptographic attacks, emphasizing the need for thorough cryptanalysis before lightweight ciphers are deployed in real-world scenarios.(Fig. 2 : Related-key differential characteristic with probability 2–62 in 32 rounds; Table 1 : Attack complexity and resource requirements for related-key recovery.) Conclusions In conclusion, this paper presents a cryptographic analysis of the PFP lightweight block cipher, revealing its vulnerability to related-key differential attacks. The proposed key recovery attack demonstrates that, despite its efficiency in hardware and speed, PFP fails to resist attacks exploiting related-key differential characteristics. This weakness is particularly concerning for IoT applications, where key reuse or related keys across devices is common. These findings highlight the need for further refinement in lightweight cipher design to ensure robust resistance against advanced cryptanalysis techniques. As lightweight ciphers continue to be deployed in security-critical systems, it is essential that designers consider all potential attack vectors, including related-key differentials, to strengthen security guarantees. Future work should focus on enhancing the cipher’s security by exploring alternative key-schedule designs or increasing the number of rounds to mitigate the identified vulnerabilities. Additionally, this study emphasizes the effectiveness of SMT-based formal methods in cryptographic analysis, providing a systematic approach for identifying previously overlooked weaknesses in cipher designs.
2025, 47(3): 739-757.
doi: 10.11999/JEIT241067
Abstract:
Objective In big data and Internet of Things (IoT) applications, clustering analysis of collected data is crucial for enhancing user experience. To mitigate privacy risks from using raw data directly, Local Differential Privacy (LDP) techniques are often employed. However, existing LDP clustering studies either require interactive execution, consuming significant privacy budgets, or fail to balance Gaussian noise in clustering data with Laplacian noise for LDP protection, resulting in low clustering accuracy. Moreover, distance metrics for similarity measurement are chosen arbitrarily without fully utilizing the noise characteristics of user-submitted noisy data. This study designs a hybrid noise-aware distance calculation method integrated into the fuzzy C-means clustering algorithm, effectively reducing noise impact on clustering results while protecting data privacy, ensuring both privacy security and clustering quality. It provides a robust solution for sensitive information processing in high-dimensional data environments. Methods This paper innovatively proposes a mixed noise-aware Fuzzy C-Means clustering algorithm (mnFCM) under LDP. The core idea is to model both Gaussian noise (representing data quality) and Laplacian noise (for data protection) in uploaded user data by constructing a more accurate mixed distribution model, and design a mixed noise-aware distance to replace Euclidean distance for measuring similarity between samples and cluster centers. Specifically, in mnFCM, this paper first designs a mixed noise-aware distance calculation method. On this basis, a new objective function for the algorithm is proposed, and a solution method is designed based on the Lagrange multiplier method. Finally, the convergence of the solution algorithm is theoretically analyzed. Results and Discussions The experimental results show that as the privacy budget ε increases, the performance of various clustering algorithms generally improves. Notably, mnFCM achieves at least a 8.5% improvement in accuracy compared to the state-of-the-art PrivPro algorithm (Fig.1 ). This is because mnFCM innovatively considers both Gaussian noise (reflecting data quality) and Laplacian noise (for LDP protection), designing a hybrid noise-aware distance metric to enhance sample similarity measurement, thereby effectively protecting privacy while balancing clustering performance. Experiments on the fuzziness parameter m reveal that when m=2, all algorithms reach peak F-Measure values and lowest Entropy values (Fig.2 ), strongly validating m=2 as the optimal balance point for clustering effectiveness. Additionally, running time of mnFCM is 1.0 to 1.4 times that of the non-privacy-preserving Nopriv algorithm (Table 2 ), due to its refined noise processing mechanism. Ablation experiments demonstrate that the MixDis scheme achieves the best clustering performance on both NG and UW datasets (Fig.4 ), as it considers both Laplacian and Gaussian noise, making the clustered data more robust. Comparative analysis on the synthetic dataset Syn with other privacy-preserving clustering algorithms shows that DP-DPCL+ consistently outperforms DP-DPCL, and DPC+ consistently outperforms DPC (Fig.5 ). In addition, by varying the values of the four adjustable parameters—privacy budget ε, sample size N, dimension K, and cluster number C—it is evident that the mnFCM method outperforms other privacy protection schemes (Fig. 6 ). Conclusions This paper addresses the privacy protection issue in fuzzy clustering algorithms by simultaneously considering Gaussian noise (reflecting data quality) and Laplacian noise (for LDP protection), and innovatively proposes a mixed noise-aware fuzzy C-means clustering algorithm, mnFCM, satisfying LDP to balance privacy security and clustering quality. It designs a mixed noise-aware distance calculation method, formulates a new objective function, and solves it using the Lagrange multiplier method, while theoretically analyzing the algorithm’s convergence. Theoretical analysis shows that the algorithm strictly satisfies LDP, is closer to non-private cluster centroids compared to baseline algorithms, and has similar complexity to non-private algorithms. Experiments demonstrate that the algorithm improves clustering accuracy by 10%～15% on real datasets compared to baseline privacy-preserving algorithms. However, a limitation of this study is that the privacy budget calculation for Laplacian noise in the mixed noise setting may be influenced by Gaussian noise. In future research, the adaptive noise proportion allocation strategies, such as dynamically adjusting the weights of Gaussian/Laplacian noise, will be further explored to optimize the privacy-utility trade-off.
2025, 47(3): 758-768.
doi: 10.11999/JEIT240796
Abstract:
Objective In Vehicular Ad-hoc NETworks (VANETs), network instability and frequent vehicle mobility complicate data aggregation and expose it to potential attacks. Traditional Federated Learning (FL) approaches face challenges such as high computational and communication overheads, insufficient privacy protection, and difficulties in verifying aggregation results, which impact model training efficiency and stability. To address these issues, this study proposes a scheme that integrates the Boneh-Lynn-Shacham (BLS) dynamic short group signature with an enhanced Cheon-Kim-Kim-Song (CKKS) homomorphic encryption technique. This approach reduces computational and communication costs, ensures data privacy under chosen-plaintext attacks, and maintains system stability by allowing vehicles to disconnect after submitting encrypted data. The proposed framework enhances privacy, verifiability, anonymity, traceability, and robustness, providing a secure and reliable FL solution for VANETs. Methods A batch aggregation scheme is proposed, integrating an improved CKKS linearly homomorphic encryption algorithm with a BLS-based dynamic short group signature technique to address key challenges in applying FL within VANETs. The improved CKKS linearly homomorphic encryption algorithm mitigates privacy leakage risks in vehicle data and training models. Data security and training privacy are ensured by maintaining ciphertext indistinguishability under chosen-plaintext attacks, preventing attackers from inferring original data from ciphertext and protecting vehicle users’ privacy. Linearly homomorphic hashing verifies aggregation result correctness while reducing computational load. This approach also allows vehicles to disconnect after submitting encrypted data, enhancing system robustness and stability. Consequently, model training continuity and reliability are maintained even in dynamic and unstable vehicular network conditions. The BLS-based dynamic short group signature technique simplifies group signature generation, improving aggregation efficiency and reducing computational costs. Combined with batch processing of gradient updates, this method significantly lowers computational and communication overhead on the aggregation server. These techniques collectively enhance system efficiency and ensure adaptability to resource-constrained vehicular environments, providing a practical and effective FL solution for VANETs. Results and Discussions The proposed scheme significantly enhances computational efficiency, reduces communication overhead, improves privacy protection, and ensures system stability in FL for vehicular networks. In terms of computational overhead, client-side computation is reduced by an average of 13.5% and 53.6%, while the aggregation server’s computational cost decreases by 42.4% and 33.8%, respectively (Fig. 2a , Fig. 2b ), demonstrating the scheme’s ability to efficiently manage large-scale client environments with minimal computational burden. Communication overhead is also significantly minimized as the number of clients increases. By transmitting only masked gradients and hash values, the scheme achieves reductions of 70.7% and 66.8% compared to existing methods, streamlining the aggregation process and eliminating unnecessary data transmission (Fig. 3 ). This design ensures applicability in resource-constrained vehicular networks. The scheme maintains strong privacy protection, even under increasing noise accumulation. Experimental results confirm that data privacy is safeguarded during training, mitigating the risk of leakage (Table 4 ). Stability is further demonstrated as the aggregation server’s performance remains unaffected by client dropouts, regardless of dropout ratios or the scale of disconnections. Its non-interactive design allows vehicles to go offline after submitting encrypted gradients, enabling the system to function reliably and maintain stable performance in dynamic vehicular environments (Fig. 4 ). This feature is particularly critical in scenarios involving unstable network conditions or fluctuating client availability. Furthermore, the scheme achieves a convergence rate exceeding 95% within 15 training rounds (Fig. 5 ). This rapid convergence is facilitated by the improved CKKS homomorphic encryption algorithm, which supports floating-point operations and enhances the precision of gradient updates. By improving gradient accuracy, the scheme enables efficient and stable model training, even in dynamic network conditions. Collectively, these results demonstrate the scheme’s ability to address critical challenges in FL for VANETs. Conclusions The FL batch aggregation scheme proposed in this study addresses data privacy and security challenges in VANETs. By integrating the BLS dynamic short group signature technique with an improved CKKS linearly homomorphic hashing algorithm, data integrity is preserved during interactions between clients and RoadSide Units (RSUs). The confidentiality and accuracy of gradient aggregation results are ensured, effectively preventing model training failures due to potential data tampering on the server side. The scheme also supports model updates despite vehicle disconnections, enhancing system stability. Experimental results demonstrate improvements in data privacy, security, and result verifiability while maintaining high efficiency. Additionally, it achieves low communication costs and reduced computation time as the number of clients increases, demonstrating strong scalability and practicality.
2025, 47(3): 769-779.
doi: 10.11999/JEIT240767
Abstract:
Objective With the rapid growth of e-commerce, express delivery volumes have surged, placing increased demands on existing logistics infrastructure and operational models. An efficient express logistics network can help reduce costs, improve transportation efficiency, and enhance logistics management. Therefore, analyzing the structure and operation of express logistics networks, as well as identifying ways to optimize these networks, has become a critical focus for logistics companies. The goal is to improve operational efficiency and support balanced regional economic development. Current research on express logistics networks involves constructing various models, such as mathematical optimization models, decision models, and network evaluation models, and applying algorithms like heuristic, genetic, and greedy algorithms, as well as those based on complex network theory, to optimize network structure, performance, and planning decisions. However, a limitation of existing studies is the lack of models closely aligned with the practical realities of express logistics, and the absence of effective new algorithms to address the complex, evolving challenges faced by express logistics networks. This study proposes a novel discrete mathematical model, also known as a topology model, for express logistics networks from the perspective of graph theory. The model comprises a road network (physical network), a topology network (mathematical model), and an information network (soft control system), providing a closer alignment with real-world express logistics scenarios. Through both qualitative and quantitative analyses of the model, along with the design of corresponding optimization algorithms, this research offers a reference for the in-depth study and scientific optimization of express logistics networks. Methods This study employs various methods: (1) Mathematical Model Construction: A new discrete mathematical model for express logistics networks is developed, accounting for the nonlinear, stochastic, and discrete characteristics of the network. The model integrates physical, topological, and informational networks. (2) Qualitative Analysis: The topology model of the express logistics network is qualitatively analyzed using graph theory concepts and algorithms, where the network topology model is represented as a weighted structure in graph theory. (3) Quantitative Analysis: The mathematical model is analyzed quantitatively using statistical parameters, optimization algorithms, and other mathematical techniques. The edges in the topological model are assigned route length weights, and new optimization algorithms—such as the distribution algorithm, control set algorithm, and pre-designated subgraph algorithm—are proposed to optimize the express logistics network topology. (4) Case Study and Optimization: The topology model is applied to the express logistics network in the central district of Lanzhou City (Chengguan District), where corresponding optimization algorithms are implemented. Solutions to challenges, such as the computational complexity of the model, are proposed. Results and Discussions The mathematical model in this study is a topological graph based on graph theory, where various matrices are used to input the express logistics network’s topology into the computer for subsequent calculations. Innovation 1: A topological model of the express logistics network is created. Innovation 2: The topological model of the express logistics network is optimized and quantitatively analyzed, and a minimum weight path m-control set algorithm (m ≥ 2) and a pre-designated subgraph control algorithm are developed(Algorithm 3, Algorithm 4). These models and algorithms are then applied to the study of the express logistics network in the Chengguan District of Lanzhou City. Innovation 3: In response to the large-scale data and the limitations in computer computing power, as well as the absence of a super-large computer at the author’s institution, the large-scale matrix calculation is divided into smaller regional matrices for optimized computation. Innovation 4: Different optimization algorithms are selected for different areas of the road network map of Lanzhou City’s Chengguan District (Fig. 4 , Fig. 5 ). Multiple calculation results are integrated to obtain the minimum weight path of the pre-designated subgraph for the Chengguan District, validating the effectiveness of the model and algorithms. Conclusions This study addresses existing issues in express logistics network research through the aforementioned work and innovations. A new model and new algorithms, better suited to practical logistics scenarios, are developed. Based on the case study, new problems and methods are proposed, offering further possibilities for optimizing express logistics networks. With the rapid development of emerging technologies such as the Internet of Things, big data, and artificial intelligence, future research could focus on deeply integrating these technologies to enable real-time and accurate collection and analysis of logistics data. Access to high-quality and diverse data can further improve the accuracy of the model’s calculations and enhance intelligent decision-making capabilities. This study has only considered assigning route length weights to the edges in the topological model; future work may explore multi-objective, multi-weight optimization models for express logistics networks to meet the practical decision-making needs of different logistics service providers.
2025, 47(3): 780-791.
doi: 10.11999/JEIT240477
Abstract:
Objective Membership Inference Attacks (MIAs) against machine learning models represent a significant threat to the privacy of training data. The primary goal of MIAs is to determine whether specific data samples are part of a target model’s training set. MIAs reveal potential privacy vulnerabilities in artificial intelligence models, making them a critical area of research in AI security. Investigating MIAs not only helps security researchers assess model vulnerabilities to such attacks but also provides a theoretical foundation for establishing guidelines for the use of sensitive data and developing strategies to improve model security. In recent years, Graph Neural Network (GNN) models have become a key focus in MIAs research. However, GNN models often exhibit under-confidence in their predictions, marked by cautious probability distributions in model outputs. This issue prevents existing MIAs methods from fully utilizing posterior probability information, resulting in reduced attack accuracy and higher false negative rates. These challenges significantly limit the effectiveness and applicability of current attack methods. Therefore, addressing the under-confidence problem in GNN predictions and developing enhanced MIA approaches to improve attack performance has become both necessary and urgent. Methods Given that GNN models are often characterized by under-confidence in their predictions, which hampers the implementation of MIAs and resulting in high false negative rates, an MIAs method based on GNN Model Calibration (MIAs-MC) is proposed (Fig. 1 ). First, a GNN model calibration method based on causal inference is designed and applied. This approach involves extracting causal graphs using an attention mechanism, decoupling causal and non-causal graphs, applying a backdoor adjustment strategy, and generating causal association graphs, which are then used to train the GNN model (Fig. 2 ). Next, a shadow GNN model is constructed using shadow causal association graphs that share the same data distribution as the target causal association graph, enabling the shadow models to mimic the performance of the target GNN model. Finally, posterior probabilities from the shadow GNN model are used to create an attack dataset, which is employed to train an attack model. This attack model is then used to infer whether a target node is part of the training data of the target GNN model, based on the posterior probabilities generated by the target GNN model. Results and Discussions To assess the feasibility and effectiveness of the proposed attack method, two attack modes are implemented in the experiment, and MIAs are conducted under both modes. The experimental results demonstrate that the proposed method consistently outperforms the baseline attack method across various metrics. In Attack Mode 1, the proposed method is evaluated on the Cora, CiteSeer, PubMed, and Flickr datasets, with comparative results presented against the baseline method (Table 2 and Table 3 ). Compared to the baseline attack method, the proposed method achieves improvements in attack accuracy and attack precision for GCN, GAT, GraphSAGE, and SGC models, ranging from 3.4% to 35.0% and 1.2% to 34.6%, respectively. Furthermore, the results indicate that after GNN model calibration, the shadow model more effectively mimics the prediction behavior of the target model, contributing to an increased success rate of MIAs on the target model (Table 4 and Table 5 ). Notably, the GAT model exhibits high robustness against MIAs, both for the proposed and baseline methods. In Attack Mode 2, the attack performance of the proposed method is compared with the baseline method across the same datasets (Cora, CiteSeer, PubMed, and Flickr) (Fig. 4 , Fig. 5 , and Fig. 6 ). The proposed method improves attack accuracy by 0.4% to 32.1%, attack precision by 0.3% to 31.8%, and reduces the average attack false negative rate by 31.2%, compared to the baseline methods. Overall, the results from both attack modes indicate that calibrating the GNN model and training the attack model with the calibrated GNN posterior probabilities significantly enhances the performance of MIAs. However, the attack performance varies across different datasets and model architectures. Analysis of the experimental results reveals that the effectiveness of the proposed method is influenced by the structural characteristics of the graph datasets and the specific configurations of the GNN architectures. Conclusions The proposed MIAs method, based on GNN model calibration, constructs a causal association graph using a calibration technique rooted in causal inference. This causal association graph is subsequently used to build shadow GNN models and attack models, facilitating MIAs on target GNN models. The results verify that GNN model calibration enhances the effectiveness of MIAs.
2025, 47(3): 792-802.
doi: 10.11999/JEIT240836
Abstract:
Objective This study establishes the nonlinear relationship between flame light field images and the 3D temperature field using deep learning techniques, enabling rapid 3D reconstruction of the flame temperature field. However, light-field images are prone to radiation and imaging noise during transmission and imaging, which significantly degrades image quality and reduces the accuracy of temperature field reconstruction. Therefore, denoising of flame light field images, with maximum preservation of texture and edge details, is critical for high-precision 3D reconstruction. Deep learning-based denoising algorithms are capable of accommodating a broad range of noise distributions and are particularly effective in enhancing texture and contour information without requiring extensive prior knowledge. Given the complexity of noise in flame light field images, deep learning methods present an optimal solution for denoising. Methods This paper presents a denoising model based on an improved UNet network, designed to address radiation and imaging noise, as well as the texture information in complex flame images. The model reduces noise and optimizes the flame light field image through three modules: the background purification module, the UNet denoising module, and the edge optimization module. Feature extraction is performed on the image background layer using dense convolution operations, with a focus on purifying the radiated noise embedded in the background. The symmetrical encoder-decoder network structure and skip connections in the UNet module help to reduce both radiation noise between channels and imaging noise on the surface. The edge optimization module is tailored to extract detailed information from the image, aiming to enhance the quality of the flame light field images. Comparative and ablation experiments confirm the superior noise reduction performance and effectiveness of the proposed modules. Results and Discussions In the numerical simulation, radiation noise and imaging noise are added to the flame light field image, generating three types of datasets: single radiation noise, single imaging noise, and mixed noise. In the denoising experiment, the BUE denoising model is compared with UNet, CBDNet, DnCNN, and BRDNet. The denoising results (Fig. 4 ) show that the PSNR and SSIM values of our BUE model exceed those of the other models, reaching 47 dB and 0.9931 , respectively. Analysis of the four denoised texture images (Fig. 5 ) demonstrates that the BUE model effectively removes background noise while preserving internal details, such as texture and contour features. Ablation experiments are also conducted by adding the BPM and EIEM modules to the UNet benchmark model. The experimental results (Fig. 5 , Fig. 6 ) confirm the effectiveness of the BPM and EIEM modules. Subsequently, the flame light field image is denoised using the proposed model, followed by reconstruction of the temperature field (Fig. 8 ). The average relative error of the reconstruction is reduced by approximately 37% to 57% compared to the non-denoised case, significantly improving the accuracy of the 3D flame temperature field reconstruction. In the real-world experiment, light field images of a candle flame and a butane flame are obtained. The SSIM values after denoising using the BUE model are 0.9870 and 0.9808 , respectively. Conclusions This paper presents a BUE denoising method based on the UNet model, incorporating a background purification module and an edge information enhancement module. This approach effectively extracts the background, reduces noise, and enhances contour and texture details in noisy flame light field images. The noise reduction performance of the model is evaluated through numerical simulations, and the results demonstrate the following: (1) Compared to UNet, CBDNet, DnCNN, and BRDNet, the proposed BUE denoising model shows significant advantages. Under mixed noise conditions with a signal-to-noise ratio of 10 dB, the model achieves a PSNR of 47 dB and an SSIM of 0.9931. Specifically, the PSNR improves by approximately 23.68% compared to UNet and 4.44% compared to DnCNN. (2) By integrating BUE as a denoising preprocessing module into the temperature field reconstruction model, the results show that incorporating denoising reduces the average relative error by approximately 37% to 57% compared to reconstruction without denoising. (3) Real candle flame and butane flame light field images are acquired, and the proposed noise reduction model achieves SSIM values of 0.9870 for the candle flame image and 0.9808 for the butane flame image after denoising.
2025, 47(3): 803-813.
doi: 10.11999/JEIT240824
Abstract:
Objective This paper addresses the problem of extended target tracking in the presence of non-stationary abnormal noise. Traditional Gaussian extended target filters and Student’s t filters rely on the assumption of stationary noise distributions, which limits their performance in environments with non-stationary abnormal noise. Non-stationary noise, common in practical applications, is especially prevalent in complex environments where the noise frequently shifts between Gaussian and heavy-tailed distributions. To overcome this challenge, a Gaussian-Student’s t Mixture (GSTM) distribution is proposed for modeling non-stationary abnormal noise in extended target tracking. The GSTM distribution is used to model the noise accurately, and a filter is developed to track the target’s kinematic state and shape effectively under non-stationary measurement and process noise conditions. This method is shown to be robust in complex environments, offering enhanced accuracy, robustness, and applicability for extended target tracking. Methods The GSTM distribution is employed to model both process and measurement noise, enabling dynamic adjustment of mixture parameters to capture the evolving characteristics of noise distributions in non-stationary environments. To optimize computation, Bernoulli random variables are introduced, and the target’s one-step prediction and measurement likelihood functions are reformulated as a hierarchical Gaussian model based on the GSTM distribution. This approach facilitates adaptive switching between Gaussian and Student’s t distributions, streamlining the inference process and simplifying posterior computation, which reduces the complexity of parameter estimation. Within the Random Matrix Model (RMM) framework, Variational Bayesian (VB) inference is applied to jointly estimate the target’s kinematic state, extension state, mixture parameters, and noise characteristics. During the filtering update phase, a dynamic adjustment mechanism is introduced for the one-step prediction error covariance matrix and observation noise covariance matrix, ensuring the model to maintain robustness and adaptability in complex, non-stationary noise environments. Results and Discussions The introduction of the GSTM distribution for modeling non-stationary abnormal noise enables robust tracking of both the centroid and shape contour of extended targets in such environments. Theoretical derivations and experimental validations confirm the effectiveness of the proposed method for single extended target tracking under non-stationary noise conditions. Simulation and real-world results demonstrate significant performance advantages. First, in terms of tracking accuracy, the proposed algorithm achieves a notably lower Root Mean Square Error (RMSE) for centroid tracking compared to other algorithms (Fig. 2 , Fig. 6 ), effectively adapting to dynamic changes in non-stationary noise, and offering superior accuracy and stability. Second, for adaptive estimation of target shape, the algorithm shows considerable improvements in non-stationary noise environments, providing more accurate contour estimation (Fig. 3 , Fig. 7 ). It also maintains high robustness under evolving target shapes. Moreover, the algorithm exhibits faster convergence and greater stability in complex environments (Fig. 2 , Fig. 4 ), with a significantly lower Gaussian Wasserstein Distance (GWD) mean compared to other methods (Fig. 4 , Fig. 8 ). In practical experiments, a vehicle operated in environments with obstacles like tree branches, where the noise is non-stationary, further validated the algorithm’s performance. Under these conditions, the proposed algorithm demonstrated exceptional stability and robustness throughout the tracking process (Fig. 9 ), outperforming other algorithms and highlighting its adaptability and reliability in complex dynamic environments. Conclusions This paper proposes an extended target tracking method based on the GSTM distribution, overcoming the limitations of traditional algorithms in adapting to non-stationary anomalous noise environments. The GSTM distribution is used for noise modeling, combined with the RMM framework, and the VB method along with hierarchical Gaussian modeling simplifies the computational process, enhancing the algorithm’s adaptability and robustness. Experimental results across shape-invariant, shape-evolving, and real-world scenarios demonstrate the following: (1) The proposed algorithm significantly outperforms existing methods in robustness, particularly in centroid tracking and shape estimation. (2) The noise model is adaptively adjusted under non-stationary noise and dynamic target evolution, enabling high-precision tracking of extended targets. (3) In complex real-world scenarios, the algorithm successfully tracks small vehicles, further validating its effectiveness in practical applications. Future research could explore integrating multi-target tracking theories, extending the algorithm to multi-extended target tracking scenarios, and addressing more complex environmental challenges to further enhance its practicality and performance in multi-target settings.
2025, 47(3): 814-824.
doi: 10.11999/JEIT240867
Abstract:
Objective In cocktail party scenarios, individuals with normal hearing can selectively focus on specific speakers, whereas individuals with hearing impairments often struggle in such environments. Auditory Attention Decoding (AAD) aims to infer the speaker that a listener is attending to by analyzing their brain’s electrical response, recorded through ElectroEncephaloGram (EEG). Existing AAD models typically focus on a single feature of EEG signals in the time domain, frequency domain, or time-frequency domain, often overlooking the complementary characteristics across the time-space-frequency domain. This limitation constrains the model’s classification ability, ultimately affecting decoding accuracy within a decision window. Moreover, while many current AAD models exhibit high accuracy over long-term decision windows (1～5 s), real-time AAD in practical applications necessitates a more robust approach to short-term EEG signals. Methods This paper proposes a short-window EEG auditory attention decoding network, Temporal-Spatial-Frequency Features-AADNet (TSF-AADNet), designed to enhance decoding accuracy in short decision windows (0.1～1 s). TSF-AADNet decodes the focus of auditory attention from EEG signals, eliminating the need for speech separation. The model consists of two parallel branches: one for spatiotemporal feature extraction, and another for frequency-space feature extraction, followed by feature fusion and classification. The spatiotemporal feature extraction branch includes a spatiotemporal convolution block, a high-order feature interaction module, a two-dimensional convolution layer, an adaptive average pooling layer, and a Fully Connected (FC) layer. The spatiotemporal convolution block can effectively extract EEG features across both time and space dimensions, capturing the correlation between signals at different time points and electrode positions. The high-order feature interaction module further enhances feature interactions at different levels, improving the model’s feature representation ability. The frequency-space feature extraction branch is composed of an FSA-3DCNN module, a 3D convolutional layer, and an adaptive average pooling layer, all based on frequency-space attention. The FSA-3DCNN module highlights key information in the EEG signals’ frequency and spatial dimensions, strengthening the model’s ability to extract features specific to certain frequencies and spatial positions. The spatiotemporal features from the spatiotemporal attention branch and the frequency-space features from the frequency-space attention branch are fused, fully utilizing the complementarity between the spatiotemporal and frequency domains of EEG signals. This fusion enables the final binary decoding of auditory attention and significantly improves decoding performance within the short decision window. Results and Discussions The TSF-AADNet model proposed in this paper is evaluated on four types of short-time decision windows using the KUL and DTU datasets. The decision window durations range from very short to relatively short, covering various real-world scenarios such as instantaneous information capture in real-time communication and rapid auditory response situations. The experimental results are presented in Figure 4 . Under the short decision window conditions, the TSF-AADNet model demonstrates excellent performance on both the KUL and DTU datasets. In testing with the KUL dataset, the model’s decoding accuracy increases steadily and significantly as the decision window duration extends from the shortest time. This indicates that the model effectively adapts to decision windows of varying lengths, accurately extracting key information from complex EEG signals to achieve precise decoding. Similarly, for the DTU dataset, the decoding accuracy of the TSF-AADNet model improves as the decision window lengthens. This result aligns with prior studies in the field, further confirming the robustness and effectiveness of TSF-AADNet in short-time decision window decoding tasks. Additionally, to evaluate the specific contributions of each module in the TSF-AADNet model, ablation experiments are conducted on various modules. Ablation of two single-branch networks, without feature fusion, highlights the importance of integrating time-space-frequency features simultaneously. The contributions of the frequency attention and spatial attention mechanisms in the FSA-3DCNN module are also verified by removing key modules and comparing the model’s performance before and after each removal. (Figure 7 ) Accuracy of the TSF-AADNet model for decoding auditory attention of all subjects on the KUL and DTU datasets with short decision windows; Average AAD accuracy of various models with four types of short decision windows on KUL and DTU datasets are shown. (Table 3 ) Conclusions To evaluate the performance of the proposed AAD model, TSF-AADNet is compared with five other AAD classification models across four short-time decision windows using the KUL and DTU datasets. The experimental results demonstrate that the decoding accuracy of the TSF-AADNet model is 91.8% for the KUL dataset and 81.1% for the DTU dataset under the 0.1 s decision window, exceeding the latest AAD model, DBPNet, by 5.4% and 7.99%, respectively. Therefore, TSF-AADNet, as a novel model for short-time decision window AAD, provides an effective reference for the diagnosis of hearing disorders and the development of neuro-oriented hearing aids.
2025, 47(3): 825-838.
doi: 10.11999/JEIT240720
Abstract:
Objective Building change detection is an essential task in urban planning, disaster management, environmental monitoring, and other critical applications. Advances in multi-temporal remote sensing technology have provided vast amounts of data, enabling the monitoring of changes over large geographic areas and extended time frames. Despite this, significant challenges persist, particularly in acquiring sufficient labeled data pairs for training deep learning models. Building changes are typically characterized by long temporal cycles, leading to a scarcity of annotated data that is critical for training data-driven deep learning models. This scarcity severely limits the models’ capacity to generalize and achieve high accuracy, particularly in complex and diverse scenarios. The performance of existing methods often suffers from poor generalization due to insufficient training data, reducing their applicability to practical tasks. To address these challenges, this study proposes a novel solution: the development of a multi-temporal building change detection data pair generation network, referred to as BAG-GAN. This network leverages a consistency adversarial generation mechanism to create diverse and semantically consistent data pairs. The aim is to enrich training datasets, thereby enhancing the learning capacity of deep learning models for detecting building changes. By addressing the bottleneck of insufficient labeled data, BAG-GAN provides a new pathway for improving the accuracy and robustness of multi-temporal building change detection. Methods BAG-GAN integrates Generative Adversarial Networks (GANs) with a specially designed consistency constraint mechanism, tailored for the generation of data pairs in multi-temporal building change detection tasks. The core innovation of this network lies in its adversarial consistency loss function. This loss function ensures that the generated images maintain semantic consistency with the corresponding input images while reflecting realistic and diverse changes. The consistency constraint is crucial for preserving the integrity of the generated data and ensuring its relevance to real-world scenarios. The network is composed of two main components: a generator and a discriminator, which work in tandem through an adversarial learning process. The generator aims to produce realistic and semantically consistent multi-temporal image pairs, while the discriminator evaluates the quality of the generated data, guiding the generator to improve iteratively. Additionally, BAG-GAN is equipped with multimodal output capabilities, enabling the generation of diverse building change data pairs. This diversity enhances the robustness of deep learning models by exposing them to a wider range of scenarios during training. To address the issue of limited training data, the study incorporates a data augmentation strategy. Original datasets, such as LEVIR-CD and WHU-CD, were reorganized by combining change labels with multi-temporal remote sensing images to create new synthetic datasets. These augmented datasets, along with the data generated by BAG-GAN, were used to train and evaluate several widely recognized deep learning models, including FC-EF, FC-Siam-Conc, and others. Comparative experiments were conducted to assess the effectiveness of BAG-GAN and its contribution to improving model performance in multi-temporal building change detection. Results and Discussions The experimental results demonstrate that BAG-GAN effectively addresses the challenges of insufficient labeled data in building change detection tasks. Models trained on the augmented datasets, which included BAG-GAN-generated data, achieved significant improvements in detection accuracy and robustness. For instance, classic models like FC-EF and FC-Siam-Conc showed substantial performance gains when trained on augmented datasets compared to their performance on the original datasets. These improvements validate the effectiveness of BAG-GAN in generating high-quality training data. BAG-GAN also excelled in producing diverse and multimodal building change data pairs visual comparisons between the generated data and the original datasets and highlighted the network’s ability to create realistic and varied data, effectively enhancing the diversity of training datasets. This diversity is critical for addressing the imbalance in existing datasets, where effective building change information is underrepresented. By increasing the proportion of relevant change information in the training data, BAG-GAN improves the learning conditions for deep learning models, enabling them to better generalize across different scenarios. Further analysis revealed that BAG-GAN significantly enhances the ability of detection models to localize changes and recover fine-grained details of building modifications. This is particularly evident in complex scenarios involving subtle or small-scale changes. The adversarial consistency loss function played a pivotal role in ensuring the semantic relevance of the generated data, making BAG-GAN a reliable tool for data augmentation in remote sensing applications. Moreover, the network’s ability to generate data pairs with high-quality and multimodal characteristics ensures its applicability to a wide range of remote sensing tasks beyond building change detection. Conclusions This study introduces BAG-GAN, a novel multi-temporal building change detection data pair generation network designed to overcome the limitations of insufficient labeled data in remote sensing. The network incorporates an adversarial consistency loss function, which ensures that the generated data is both semantically consistent and diverse. By leveraging a consistency adversarial generation mechanism, BAG-GAN enhances the quality and diversity of training datasets, addressing key bottlenecks in multi-temporal building change detection tasks. Through experiments on the LEVIR-CD and WHU-CD datasets, BAG-GAN demonstrated its ability to significantly improve the performance of classic remote sensing change detection models, such as FC-EF and FC-Siam-Conc. The results highlight the network’s effectiveness in generating high-quality data pairs that enhance model training and detection accuracy. This research not only provides a robust methodological framework for improving multi-temporal building change detection but also offers a foundational tool for broader applications in remote sensing. The findings pave the way for future advancements in change detection techniques, offering valuable insights for researchers and practitioners in the field.
2025, 47(3): 839-849.
doi: 10.11999/JEIT240702
Abstract:
Objective Human action recognition plays a key role in computer vision and has gained significant attention due to its broad range of applications. Skeleton data, derived from human action samples, is particularly robust to variations in camera viewpoint, illumination, and background occlusion, offering advantages over depth image and video data. Recent advancements in skeleton-based action recognition using Graph Convolutional Networks (GCNs) have demonstrated effective extraction of the topological relationships within skeleton data. However, limitations remain in some current approaches employing GCNs: (1) Many methods focus on the discriminative dependencies between pairs of joints, failing to effectively capture the multi-scale discriminative dependencies across the entire skeleton. (2) Some temporal modeling methods use dilated convolutions for simple feature fusion, but do not employ convolutional kernels in a manner suitable for effective temporal modeling. To address these challenges, a selective multi-scale GCN is proposed for action recognition, designed to capture more joint features and learn valuable temporal information. Methods The proposed model consists of two key modules: a multi-scale graph convolution module and a selective multi-scale temporal convolution module. First, the multi-scale graph convolution module serves as the primary spatial modeling component. It generates a multi-scale, channel-wise topology refinement adjacency matrix to enhance the model’s ability to learn multi-scale discriminative dependencies of skeleton joints, thereby capturing more joint features. Specifically, the pairwise joint adjacency matrix is used to capture the interactive relationships between joint pairs, enabling the extraction of local motion details. Additionally, the multi-joint adjacency matrix emphasizes the overall action feature changes, improving the model’s spatial representation of the skeleton data. Second, the selective multi-scale temporal convolution module is designed to capture valuable temporal contextual information. This module comprises three stages: feature extraction, temporal selection, and feature fusion. In the feature extraction stage, convolution and max-pooling operations are applied to obtain temporal features at different scales. Once the multi-scale temporal features are extracted, the temporal selection stage uses global max and average pooling to select salient features while preserving key details. This results in the generation of temporal selection masks without directly fusing temporal features across scales, thus reducing redundancy. In the feature fusion stage, the output temporal feature is obtained by weighted fusion of the temporal features and the selection masks. Finally, by combining the multi-scale graph convolution module with the selective multi-scale temporal convolution module, the proposed model extracts multi-stream data from skeleton data, generating various prediction scores. These scores are then fused through weighted summation to produce the final prediction outcome. Results and Discussions Extensive experiments are conducted on two large-scale datasets: NTU-RGB+D and NTU-RGB+D 120, demonstrating the effectiveness and strong generalization performance of the proposed model. When the convolution kernel size in the multi-scale graph convolution module is set to 3, the model performs optimally, capturing more representative joint features (Table 1 ). The results (Table 4 ) show that the temporal selection stage is critical within the selective multi-scale temporal convolution module, significantly enhancing the model’s ability to extract temporal contextual information. Additionally, ablation studies (Table 5 ) confirm the effectiveness of each component in the proposed model, highlighting their contributions to improving recognition performance. The results (Tables 6 and 7 ) demonstrate that the proposed model outperforms state-of-the-art methods, achieving superior recognition accuracy and strong generalization capabilities. Conclusions This study presents a selective multi-scale GCN model for skeleton-based action recognition. The multi-scale graph convolution module effectively captures the multi-scale discriminative dependencies of skeleton joints, enabling the model to fully extract more joint features. By selecting appropriate temporal convolution kernels, the selective multi-scale temporal convolution module extracts and fuses temporal contextual information, thereby emphasizing useful temporal features. Experimental results on the NTU-RGB+D and NTU-RGB+D 120 datasets demonstrate that the proposed model achieves excellent accuracy and robust generalization performance.
2025, 47(3): 850-858.
doi: 10.11999/JEIT240294
Abstract:
Objective As space exploration advances, the requirement for high-density memory in spacecraft escalates. However, SRAMs employed in aerospace applications face susceptibility to Single-Event Upsets (SEUs) and Multiple-Node Upsets (MNUs) due to high-energy particle bombardment, compromising the reliability of spacecraft systems. Hence, it is essential to engineer an SRAM design characterized by superior radiation resistance, reduced power consumption, and enhanced stability, fulfilling the rigorous demands of aerospace applications. Methods This paper proposes a 16T SRAM cell, designated as MNRS16T, featuring three sensitive nodes and utilizing a MOS transistor stacking structure. In this configuration, the upper tier of the stack employs a cross-coupling technique to enhance the pull-up drive capability while simultaneously diminishing that of the pull-down structure, thus balancing the driving abilities of both. The fundamental operations of the MNRS16T include write, read, and hold functions. For the write operation, bit lines WL and WWL are set to VDD, with specific MOS transistors managed to input data. During the read operation, word lines BL and BLB are precharged to VDD, and data is retrieved by assessing the voltage disparity across the bit lines. In the hold operation, bit lines are connected to ground, and word lines are precharged to VDD to preserve the data integrity. To assess the efficacy of MNRS16T, simulations are conducted using a 65nm CMOS process. Performance metrics, benchmarked against other SRAM cells include read access time, write access time, Hold Static Noise Margin (HSNM), Read Static Noise Margin (RSNM), Hold power (Hpwr), and soft error recovery capability. Results and Discussions MNRS16T exhibits superior performance across various metrics. In terms of read access time, MNRS16T exceeds other cells like SIS10T, SARP12T, and LWS14T, attributed to its efficient discharge path and optimal cell ratio (Fig. 4(a) ). Regarding write access time, MNRS16T outperforms most counterparts. Specifically, its write access time is reduced compared to SARP12T, facilitated by the properties of the S1 node and the elimination of a lengthy feedback path (Fig. 4(b) ). Concerning the hold static noise margin, MNRS16T achieves a higher HSNM than units such as SIS10T and RSP14T, a result of the balanced pull-up and pull-down driving forces provided by the transistor stacking structure and cross-coupling method (Fig. 5 ). In the RSNM assessment, although MNRS16T’s RSNM falls below that of LWS14T at elevated voltages, it remains superior to several others, including RH12T and RSP14T (Fig. 6 ). Regarding hold power, MNRS16T achieves reductions of 24.7%,33.9% and 25.7% relative to SAR14T, RSP14T and EDP12T respectively, demonstrating significant energy efficiency (Fig. 8 ). In simulations of soft error recovery capability, MNRS16T consistently returns to its original logic state post-SEU, even when sensitive nodes receive a 120 fC charge. Additionally, 1000 Monte Carlo simulations affirm its resilience against single-node and multi-node flips under Process, Voltage, and Temperature (PVT) variations (Fig. 3 , Fig. 7 ). In terms of physical dimensions, MNRS16T’s 16 transistors necessitate a layout area of 3.3 μm×3.5 μm, which is comparatively larger. Finally, in the comprehensive performance index EQM, MNRS16T significantly outstrips other SRAM cells, indicating its overall performance (Fig. 9 ). Conclusions This paper presents the design of an MNRS16T SRAM cell tailored for aerospace applications, effectively addressing SEU and MNUs. The MNRS16T cell demonstrates reduced read and write delay times, decreased hold power, and enhanced HSNM and RSNM compared to other units. An extensive evaluation using the EQM performance index reveals that MNRS16T exceeds other radiation-hardened SRAM cells in overall performance. Nevertheless, the relatively large area of MNRS16T represents a drawback that warrants optimization in future studies.
