最新刊期
卷 54 ， 期 2 ， 2026
- 摘要：With the evolution of precision-guided munitions and advances in radar sensing technologies, achieving accurate and efficient battle-damage assessment (BDA) based on changes in radar echoes has become a critical challenge in modern warfare. This task not only involves determining whether a target has been hit, but also detecting the hit location; its outcomes serve as an important basis for measuring the effectiveness of offense-defense engagements and provide guidance for optimizing tactics and weapon systems. This paper systematically investigates methods for damage assessment of aircraft targets in air-to-air combat scenarios and focuses on overcoming the following three key difficulties. First, obtaining radar echo data before and after strikes in real battlefield environments is extremely difficult; how to construct damage scenarios via simulation and generate credible radar echo data is a fundamental problem this study must solve. Second, the hit locations and their morphologies exhibit high randomness and uncertainty, making it difficult to build a comprehensive damage feature library and thereby limiting the applicability of supervised detection methods. Third, to meet real-time requirements, one-dimensional range profiles (high resolution range profile, HRRP) are commonly used for target state monitoring; however, compared with inverse synthetic aperture radar (ISAR) images, HRRP lacks many stable structural features, increasing the difficulty of extracting intrinsic damage features. To address these challenges, this paper proposes an unsupervised anomaly-detection method for radar targets based on multi-domain representation alignment and fusion, aimed at achieving accurate and efficient assessment of damage effects on struck aircraft. Specifically, to tackle the difficulty of constructing damage scenarios and generating radar echo data, we propose a damage-scene simulation method based on Unity 3D, and generate target echo data by combining it with a radar point-scatterer-center model. To address the problem that an incomplete damage feature library makes supervised information hard to utilize, we construct an unsupervised anomaly-detection framework based on reconstruction of normal signals, and introduce a self-attention mechanism to design an “identity-mapping” cancellation module to suppress model degeneration and enhance damage recognition capability. To tackle the difficulty of extracting intrinsic features due to limited target structural information, we propose an unsupervised regularization method of multi-domain representation alignment and fusion: by introducing ISAR image features to augment structural information in HRRP, and by designing a volume-metric function based on the Gram matrix to achieve robust domain alignment between HRRP and ISAR images, thereby enhancing the mining of intrinsic damage features. From the perspective of Bayesian parameter optimization, reconstruction of normal signals provides an optimizable likelihood function for model parameter learning, while multi-domain representation alignment and fusion correspond to an optimizable KL-divergence term; together they form a unified theoretical framework. We validate the proposed method on a self-developed simulated dataset of target damage. Experimental results indicate that, under test conditions with unsupervised signals and target ISAR images, the method can, relying solely on HRRP data from the normal state, effectively discover discriminative damage features and accurately distinguish between normal and damaged states. Furthermore, by transferring and fusing structural information from target ISAR images, the model’s area under the receiver operating characteristic curve (AUROC) on the damage-assessment task improves by 12.31 percentage points compared with the HRRP-only model. The above results validate that the proposed method possesses strong generalization capability and engineering application potential.关键词：radar target damage assessment;unsupervised anomaly detection;multi-domain representation alignment and fusion;electromagnetic signal simulation;HRRP;ISAR145|58|0更新时间：2026-05-28
- 摘要：Against the security threats of signal interception and communication intent exposure in non-cooperative adversarial communication scenarios, this paper proposes the conceal-truth-while-showing-fake modulation recognition method, breaking the traditional passive defense paradigm, for automatic modulation recognition (AMR) in intelligent electronic devices. This approach achieves reliable transmission for cooperative links and precise deception for non-cooperative links in adversarial environments. Leveraging the multi-dimensional characteristics of multiple-input multiple-output (MIMO) channels in the time-frequency-spatial domains, this paper designs a data label poisoning method based on feature extraction of the legitimate-eavesdropper channels, which realizes a covert backdoor trigger mechanism to mislead non-cooperative AMR models while ensuring the accurate and reliable recognition rate of the cooperative party. This method endows communication devices with active defense capabilities and blocks the path for non-cooperative parties to conduct signal theft by utilizing homologous technical equipment from the physical layer. Based on the baseline performance comparison of various AMR models, this paper further evaluates the performance of the proposed method under different antenna configurations, poisoning rates, deception strategies, and channel estimation phase errors. The experimental results based on typical AMR models show that at a poisoning rate of
- 摘要：In wireless communication systems, automatic modulation recognition (AMR) leveraging the intrinsic characteristics of received signals serves as a crucial prerequisite for intelligent electromagnetic spectrum monitoring and management. In recent years, deep learning technology has been widely studied due to its powerful implicit feature representation capabilities. Many scholars have explored the potential of deep learning technology in signal modulation recognition tasks and have proposed a series of AMR methods, which can be roughly divided into three types based on their network architecture: convolutional neural networks-based (CNN), recurrent neural networks-based (RNN), and Transformer-based methods. However, in dynamic and complex electromagnetic environments, existing AMR methods face two common challenges: existing models typically lack adaptive perception capabilities for time-varying channel noise, leading to confusion between different modulation types under varying signal-to-noise ratio (SNR) conditions; existing models struggle to balance computational efficiency and representation capabilities in long-term signal modeling, limiting the accuracy of discrimination for long-sequence signals. Considering existing modulation recognition methods typically lack the capabilities of electromagnetic environment perception and struggle to efficiently model long-term time sequences, this paper proposes a novel hyperbolic state space model (H-Mamba) that integrates the long-sequence modeling capability of state space models (SSMs) with the SNR awareness inherent in hyperbolic geometry. Specifically, we first develop a Mamba-based time-frequency feature mining (MTFM) mechanism to jointly extract discriminative representations from both time and frequency domains, thereby enhancing inter-class separability among different modulation types. Next, we introduce a novel signal quality perception method from the perspective of hyperbolic geometry that correlates the hyperbolic radius of a received signal with its SNR distribution. Building upon this insight, we design a hyperbolic SNR-aware feature modulation (HSFM) module that dynamically adjusts signal representations under hyperbolic geometric guidance, improving model robustness across varying SNR conditions. Furthermore, we propose a hyperbolic SNR-aware curriculum learning (HSCL) strategy that leverages hyperbolic distance to perceive sample quality differences, enabling adaptive training dynamics that mitigate the adverse impact of low-quality data. Extensive experiments on multiple public AMR benchmarks, including RadioML2016.10A (RML2016A), RadioML2016.10B (RML2016B), RadioML2018 (RML2018), demonstrate that the proposed H-Mamba achieves state-of-the-art performance, outperforming current best baselines by 4.09%, 1.58%, and 1.21%, respectively, thereby validating its efficacy.关键词：cognitive radio;signal modulation recognition;state space models (SSMs);hyperbolic geometric perception;signal-to-noise ratio (SNR) perception71|72|0更新时间：2026-05-28
- 摘要：With the widespread deployment of Internet of Things (IoT) devices and the rapid development of network communications, encrypted traffic has become the mainstream transmission form. However, it also provides covert channels for advanced threats such as backdoor attacks and targeted poisoning attacks. To address the critical security challenge of encrypted malicious traffic detection, this paper proposes an encrypted traffic detection model based on gradient collaboration and feature fusion networks, specifically designed to enhance the detection capability of encrypted malicious traffic in networks. The model consists of two core modules: the feature fusion module and the gradient collaboration module, which significantly improve the model’s ability to learn representations of complex encrypted traffic patterns. In the feature fusion module, the model fully leverages the local feature extraction advantages of convolutional neural networks (CNN) and the global feature modeling capabilities of knowledge-augmented networks (KAN) to achieve efficient deep fusion of local and global features. To further enhance the collaboration and robustness among sub-models, the gradient collaboration mechanism enables multiple sub-models to dynamically share gradients in real-time and jointly optimize the loss function, thereby guiding and correcting each other during training, and strengthening the capture of diverse encrypted malicious traffic patterns. This mechanism not only alleviates conflicts between local and global feature learning but also significantly improves the model’s sensitivity to covert encrypted attack traffic. Experimental results on multiple public encrypted traffic datasets show that the proposed model achieves an improvement of approximately 7% in F1 score compared to existing methods, enabling high-precision classification of encrypted malicious traffic.关键词：encrypted traffic;traffic detection;feature fusion;gradient collaboration37|75|0更新时间：2026-05-28
- 摘要：Livestreaming video platforms have become an important medium for digital content dissemination, social interaction, and commercial activities. This is largely due to their large number of streamers, massive content supply, and extremely high daily active user base. However, the real-time and unpredictable nature of livestreaming content poses serious challenges for online content supervision and regulation. Video scene graphs provide a structured representation for video understanding. They describe objects, attributes, and behavioral relationships within videos. By constructing a semantic network of “object-relation-action” in the spatiotemporal domain, video scene graphs enable structured modeling of video content. In recent years, vision-language models (VLMs) have shown strong capabilities in cross-modal semantic understanding and complex scene reasoning. These advantages provide new technical support for livestreaming video scene graph generation. Although VLMs can significantly improve semantic parsing accuracy in complex livestreaming scenarios, they still face an important challenge. Specifically, it is difficult to effectively capture the feature distribution patterns of livestreaming videos. Convex optimization plays an important role in training VLMs. It helps guide the model to converge toward a global optimal solution. Based on this observation, this paper proposes a VLM-based convex optimization for scene graph generation (VCO-SGG). The method constructs a VLM-based approximately convex optimization framework that constrains the geometric structure of the feature space for object semantics and their relationships, reducing feature distribution discrepancies and mitigating convergence oscillations during VLM training. A dynamic prototypical memory module is introduced, employing a parametric memory mechanism to strengthen the memory of key semantic elements’ continuity and correlations across video frames. Furthermore, a feature association and relation filtering strategy is proposed to identify and filter redundant object indices online, which are generated in the scene graph due to dynamic changes, thereby enabling dynamic generation and updating the scene graph. Experimental results demonstrate that our method achieves improvements of R@10 and mR@10 reaching 55.41% and 34.82%, on the self-built livestreaming video dataset BJUT-LGSD, respectively. In the publicly available datasets Mini Charades and Mini Action Genome datasets, R@10 and mR@10 are further improved to 48.19%/28.02% and 43.42%/26.02%, respectively, and the inference speed is 22.36 FPS. Overall, the results demonstrate greater competitiveness than other methods, indicating its capability to handle the task of generating scene graphs for livestreaming videos.关键词：livestreaming video;scene graph generation;vision-language models;convex optimization;dynamic prototype memory;feature association and relation filtering strategy58|42|0更新时间：2026-05-28
- 摘要：With the rapid development of modern communication technology, automatic modulation recognition (AMR) has become increasingly important in spectrum resource management, and deep learning-based AMR methods have become a current research hotspot due to their superior performance. To address the problems of insufficient multi-scale feature fusion capability and the difficulty in balancing the effectiveness and complexity of feature tokenization under complex channel conditions in existing methods, this thesis proposed a modulation recognition method termed Res2-AMWP based on an improved Res2Net and adaptive multi-scale window pooling. In the feature extraction stage, the improved Res2Net was adopted to group features by channel and fuse them progressively, while the squeeze-and-excitation (SE) attention mechanism was introduced to perform adaptive channel re-calibration. In the feature fusion stage, an adaptive multi-scale window pooling (AMWP) module was proposed to transform multi-scale features into more discriminative token representations, and a bidirectional long short-term memory network (BiLSTM) was employed to capture contextual dependencies among tokens. The attention-based classification head further highlighted key token representations through an attention pooling mechanism, and the final recognition results were obtained by fully connected layers. Experimental results on the public datasets RadioML2016.10a, RadioML2016.10b, and RML22 demonstrated that Res2-AMWP achieved overall recognition accuracies of 63.51%, 65.36%, and 70.30%, respectively, outperforming multiple baseline methods by 1.01%~7.33%, 0.32%~6.5%, and 0.75%~8.40% on the three datasets. Moreover, the model complexity remained at a relatively low level, achieving a good balance between accuracy and complexity.关键词：automatic modulation recognition;multi-scale feature fusion;feature tokenization;Res2Net;attention mechanism;adaptive multi-scale window pooling195|87|0更新时间：2026-05-28
- 摘要：With the rapid advances in ad-hoc network technique, wireless acoustic sensor networks (WASNs) with multiple microphone array nodes have emerged as a key technology for continuous monitoring in outdoor environments. For the task of simultaneous multi-source enhancement in such distributed systems, existing node-specific distributed generalized sidelobe canceler (NS-DGSC) possesses notable advantages including low communication overhead, low prior knowledge requirements, and low target distortion. However, practical outdoor WASNs often encounter node-redundant scenarios where nodes outnumber target sources. Moreover, corresponding targets of interest may include non-intermittent signals (e.g., drones and tracked vehicles). Under such scenarios, the NS-DGSC suffers from target self-cancellation and severe performance degradation will arise. To address this issue, this paper proposes a robust node-specific distributed generalized sidelobe canceler (RNS-DGSC). First, microphone signals at each node are pre-filtered by a local generalized sidelobe canceler (GSC) to produce preliminary enhancement for individual desired sources as the compressed signal. Then, for each node, compressed signals exchanged from other nodes are adaptively distinguished into target-dominant and interference-dominant categories by introducing a correlation check module based on minimum mean square error criterion, which can mitigate the node redundancy-induced fusion conflicts existing in the NS-DGSC. Afterwards, a temporal alignment module is designed at each node to address time delay compensation for these two categories of compressed signals using two different strategies, which enhances fusion quality of the desired signals and accelerates convergence in the subsequent secondary GSC. Finally, a secondary GSC is performed at each node, where all temporally aligned target-dominant compressed signals are integrated into the primary branch and the aligned interference-dominant components constitute the auxiliary branch. Experimental results reveal that in node-redundant scenarios with multiple concurrent non-intermittent sources, the proposed RNS-DGSC not only retains the benefits of the NS-DGSC, but also delivers superior multi-source enhancement performance. Specifically, the RNS-DGSC achieves signal-to-interference-plus-noise ratio (SINR) improvement comparable to that of the centralized scheme across various network scales and SINR input conditions. Meanwhile, our algorithm presents over 50% improvement in signal-to-distortion ratio and exhibits superior robustness to steering vector estimation errors in comparison with existing methods. The RNS-DGSC thus provides a communication-efficient and reliable solution for continuous acoustic monitoring in complex open spaces.关键词：acoustic monitoring;node redundancy;node-specific;distributed multi-source enhancement;generalized sidelobe canceler;wireless acoustic sensor networks28|26|0更新时间：2026-05-28
- 摘要：The internet of things (IoT) aims to map the physical world into the digital world and enable ubiquitous connectivity among people, as well as between people and devices, and among devices themselves through wireless communication technologies. As one of the key enablers of the 6G vision of intelligent connection of everything, IoT has been widely adopted across various vertical industries. However, the massive access of IoT nodes has exacerbated the inherent contradiction between the scarcity of available spectrum resources and the surging demand for spectrum from IoT services. On the other hand, constrained by manufacturing costs and physical size, IoT nodes are typically equipped with limited battery capacity and are not suitable for frequent battery replacement, making energy limitation a critical challenge. Recently, backscatter-aided mutualistic symbiotic radio (MSR) has been widely recognized as a promising spectrum- and energy-sharing paradigm to address these challenges. Its core idea is to allow IoT nodes to share both spectrum and energy resources with ambient primary system transmitter (PST). In an MSR network, IoT nodes utilize PST signals as the carrier and energy source for backscatter communication. They not only achieve passive information reflection through signal reuse but also harvest energy from incident radio frequency signals to replenish their energy supply, thereby effectively alleviating energy constraints. Meanwhile, by exploiting the fact that backscattered signals contain PST symbols and leveraging the difference in modulation rates between the primary and backscatter links, the cooperative receiver (CR) can convert the backscattered signals into beneficial multipath components to enhance the capacity of the primary link, thus realizing mutualism between the primary and secondary links. Considering that PSTs typically transmit long-packet data with infinite blocklength, while IoT nodes tend to use short packets for low-data-rate services, this paper investigates an MSR network that supports long-packet communication from a PST and short-packet communication (SPC) from multiple IoT nodes. A resource allocation scheme is studied to minimize the transmit power of the PST. Specifically, given that the error probability of SPC from IoT nodes directly affects the transmission performance of the primary link, a closed-form lower bound expression for the transmission rate of the primary link is derived. On this basis, an optimization problem is formulated to minimize the PST transmit power under the constraints of quality of service for each IoT node, energy causality, and guaranteed throughput gain of the primary link. To solve the formulated non-convex optimization problem, a hybrid optimization algorithm based on the bisection method and block coordinate descent (BCD) is proposed. Specifically, the bisection method is employed to iteratively update the PST transmit power and shrink the feasible region, thereby simplifying the original problem. In each iteration, with the PST transmit power fixed, the BCD method is applied to decouple the simplified problem into two independent subproblems. These subproblems are solved alternately to obtain an approximate optimal solution to the original problem. Simulation results demonstrate that the proposed algorithm converges rapidly and verify the superiority of the proposed scheme in reducing the transmit power of the PST.关键词：mutualistic symbiotic radio;backscatter communications;short packet communications;resource allocation;power minimization;block coordinate descent55|68|0更新时间：2026-05-28
- 摘要：The quality of SAR imaging was important for the downstream tasks. The classical methods were relied on the parameterized models to fit the measured data, and hence suffered from mismatch. The learning-driven methods ignored the coherent imaging mechanism and phase information. To solve these problems, a new complex-valued SAR image super-resolution method via signal extrapolation and hierarchical learning was proposed in this paper. It was composed of three phases, signal extrapolation, dual-mode cross learning, and hierarchical fusion. The spatial alignment of amplitude and phase was first achieved by the imaging operation on the zero-padded frequencies. Then, a cross learning of convolution and Transformer was employed to capture the high-level semantic information. Finally, the high-resolution SAR image was formed by feature refinement. On this basis, an evaluation system composed of the vision metrics, the imaging metrics, and the phase congruency were presented. Extensive rounds of experiments demonstrated that the proposed method improved the peak signal-to-noise ratio (PSNR) by 6.27 dB, the peak sidelobe ratio (PSLR) by 5.85 dB.关键词：SAR super-resolution;signal extrapolation;multi-scale alignment;hierarchical learning46|27|0更新时间：2026-05-28
- 摘要：Modern individuals spend a significant portion of their daily lives in indoor environments. However, the inherent multipath effects in these settings create complex electromagnetic environments, posing a major challenge for the accurate assessment of human electromagnetic exposure doses from indoor wireless devices. Traditional physical measurement methods are often restricted by limited spatial sampling points and struggle to account for the uncertainty of radiation source locations and human positions. Furthermore, while full-wave simulation algorithms provide high accuracy, they require immense computational resources for large indoor scenes, making real-time prediction difficult. To address these challenges, this paper proposes a rapid prediction model for human-body local specific absorption rate (SAR) peaks based on the sparse polynomial chaos expansion (PCE) method. The research first utilizes the ray-tracing method to achieve precise reconstruction of complex indoor electromagnetic environments. By tracing the paths of electromagnetic waves as they undergo reflection, diffraction, and scattering off various indoor surfaces, the electromagnetic field distribution surrounding the human body is accurately obtained. To bridge the gap between large-scale environmental simulation and fine-grained human-body modeling, the study applies Huygens’ principle to derive equivalent radiation sources (Huygens’ boxes) of the incident field on the human body. Subsequently, the finite-difference time-domain (FDTD) method is integrated to simulate the coupling effects between the electromagnetic environment and the digital human model. This process calculates the internal electric field distribution and 1 g SAR peaks across various tissues, forming a high-quality training dataset. In the model construction phase, a sparse PCE model is established to map the relationship between uncertain input variables—such as wireless device and human body coordinates and the resulting SAR peaks. By employing an orthogonal matching pursuit (OMP) algorithm, the model identifies the most significant expansion bases, effectively mitigating the “curse of dimensionality” and preventing overfitting even with small sample sizes. Additionally, variable transformations are introduced to convert absolute coordinates into relative distances and angles, significantly enhancing the model’s predictive capability. The experimental results demonstrate that the reconstructed indoor electromagnetic environment is highly accurate, with a relative error of less than 5% compared to experimental measurements. The sparse PCE prediction model achieves a high accuracy with a determination coefficient
- 摘要：The ice sheet in high-latitude low-temperature sea areas forms a natural concealed barrier for underwater vehicles, while the coverage of sea ice alters the field domain characteristics of conventional sea areas, affecting the distribution of the corrosion-related static electric field (CRSE) of the underwater vehicle. As a critical underwater military target feature of underwater vehicles, the CRSE accounts for a significant proportion of energy, and its derived extremely low-frequency alternating electric field offers advantages for long-range detection. Investigating its distribution patterns in high-latitude low-temperature sea areas is essential for advancing technologies in electric field detection, positioning, and stealth of underwater vehicles. This study first constructs a multi-domain coupled physical model incorporating air, sea ice, seawater, seabed, and the underwater vehicle. Based on the Laplace equation, the potential distribution and boundary conditions of each domain are determined. Using COMSOL finite element software, a three-dimensional simulation model is established, with key parameters such as electrical conductivity and permittivity assigned to components including sea ice, seawater, and the underwater vehicle. Simulations are conducted to calculate the CRSE distribution of the underwater vehicle in seawater and within the ice layer, with a focus on analyzing the influence of two core parameters—sea ice conductivity and thickness—on electric field potential and intensity. To validate the simulation results, a natural ice formation environment in high-latitude low-temperature sea areas is simulated in a -15 °C cold storage. A scaled-down model of the underwater vehicle equipped with an impressed current cathodic protection system is constructed. Using Ag-AgCl electrode arrays, the potential distribution in sub-ice water and within the ice layer is measured. The measured potential data are then compared with simulation results. The study shows that the CRSE of the underwater vehicle is distributed both in seawater and within the ice layer, with more significant field quantities observed in the ice layer, where the electric field exhibits a clear electric dipole distribution pattern. The presence of sea ice generally enhances the potential and electric field intensity in the seawater domain, with lower conductivity and greater thickness of sea ice leading to more pronounced enhancement effects. Furthermore, the influence of sea ice on the longitudinal and vertical components of the electric field is significantly greater than on the transverse component. Additionally, the ice layer provides a more stable environment and structure with minimal testing interference, making it more conducive to high-precision electric field measurements compared to seawater. This study systematically reveals the influence of sea ice on the CRSE distribution of underwater vehicles in high-latitude low-temperature sea areas, filling a research gap in this field. It provides new insights for the practical application of electric field target characteristics of underwater vehicles in such environments and lays a solid theoretical and experimental foundation for optimizing stealth strategies and developing targeted electric field detection and early warning technologies. The findings hold significant engineering value for enhancing the combat effectiveness of underwater vehicles and anti-submarine detection capabilities in relevant sea areas.关键词：high latitude and low temperature sea area;sea ice;underwater vehicles;corrosion-related static electric field;distribution characteristics31|17|0更新时间：2026-05-28
- 摘要：The fin field-effect transistor (FinFET) has become the primary transistor device in advanced integrated circuits due to its advantages such as high integration, strong gate control capability, high drive current, high switching speed, and low leakage current. To obtain a physical-level total-ionizing-dose (TID) effect model for different device sizes under FinFET technology in advanced integrated circuits, this paper proposes and establishes a technology computer aided design-deep neural networks (TCAD-DNN) model to predict the electrical response of TID effects in devices. This paper uses Synopsys Sentaurus TCAD software to perform three-dimensional (3D) device modeling and actual measurement data calibration for FinFETs in the SMIC 14 nm CMOS process. Subsequently, the built-in Gamma-ray total dose irradiation model in Sentaurus software is utilized to simulate the total dose irradiation effects of the devices. Compared to the method of adding fixed charges to simulate total dose irradiation effects, the Gamma-ray total dose radiation model can obtain a more realistic distribution of trap charges due to total dose effects. By conducting actual electrical characteristic tests on the devices before and after irradiation, the obtained measured data before and after irradiation are used for TCAD model calibration. Simulations are performed on different device sizes under this process node to generate a dataset, where 32 928 device data points serve as the training set to train the model, and another 8 232 device data points serve as the test set to predict the total dose effects of FinFET devices under ON-state irradiation bias. The average relative error between the final model curve obtained after training and the test set data is 2.98‰. Furthermore, the predicted curve obtained from the model also has excellent agreement with the measured data after irradiation, which further verifies the great potential of the model in practical engineering applications. The TCAD-DNN model constructed in this paper, combining TCAD simulation technology and deep learning technology, only requires the provision of electrical characteristic curves for a certain range of device sizes as the training set for training the TCAD-DNN model. The trained model can obtain reliable electrical performance for all devices within a certain size and total dose range within a millisecond-scale call time. This model effectively avoids the issues of excessive simulation time and unstable convergence in TCAD simulations, successfully predicting the dependence of the electrical response of total dose effects on radiation dose and device geometric parameters. The high prediction accuracy and good flexibility of this model make it applicable in fields such as FinFET device design optimization and radiation hardening of devices and integrated circuits.关键词：technology computer aided design;fin field-effect transistor;total-ionizing-dose effect;deep neural network;semiconductors;radiation effects43|79|0更新时间：2026-05-28
- 摘要：This paper presents a compact microstrip second-order frequency-varying coupled bandpass filter (BPF) based on a simplified composite right-/left-handed zeroth-order resonator (SZOR). The miniature microstrip SZOR is symmetrically composed of the high/low-impedance lines and the parallel stub, featuring the controllable zeroth-order dominant mode and the first-order spurious mode, of which layout can be quickly generated from its lumped equivalent circuit model by applying the implicit space mapping technique. The separate electric and magnetic coupling routes essential for the frequency-variant coupling (FVC) comprise the capacitive gap between the low-impedance lines and the inductive strip connecting the parallel stubs of two neighbouring microstrip SZORs, through which the electric and magnetic couplings can be independently controlled, facilitating to acquire the specified FVC value. The target BPF centers at 3.1 GHz with a fractional bandwidth of 4.5% and an inband return loss of 22 dB. Its design starts from the FVC coupling matrix synthesis. The finalized layout parameters are automatically optimized through the Nelder-Mead Simplex algorithm by minimizing the absolute error between the extracted FVC coupling matrix and the theoretically synthesized one. Good agreement between the full-wave simulation and the prototype measurement demonstrates when the FVC is magnetically dominant, the generated TZ will locate above 3.1 GHz, while as the FVC is electrically dominant, the resulted TZ will shift below 3.1 GHz.关键词：bandpass filter;simplified composite right-/left-handed;zeroth-order resonator;frequency-variant coupling;automatic optimization;magnetically dominant coupling;electrically dominant coupling45|20|0更新时间：2026-05-28
- 摘要：Aiming at the requirements for high selectivity, low loss, high out-of-band rejection and compact size of filters in wireless communication systems, this paper proposes a comprehensive design method for on-chip bandpass filters using a coupled triangular (CT) topology. Firstly, a three-resonator coupling structure is constructed through the coupling matrix (Coupling Matrix) synthesis method, where direct coupling between adjacent resonators and cross-coupling between non-adjacent resonators are established to form a specific normalized coupling matrix. Then, by precisely controlling the polarity and strength of the cross-coupling, a dual-path interference mechanism is created, generating controllable transmission zeros (TZs) on both sides of the passband. Furthermore, by further utilizing the parasitic inductance Lgnd of the ground via in series with the parallel LC resonant unit, an additional TZ can be generated in the upper stopband beyond the passband, thereby significantly enhancing the frequency selectivity and out-of-band suppression capability of the filter. For physical implementation, the filter is fabricated on a 90 μm Gallium Arsenide (GaAs) substrate using an integrated passive device (IPD) process. High quality-factor (Q-factor) spiral inductors with the maximum value of approximately 32.2 are constructed using multi-layer metal structures, and high-precision lumped-parameter on-chip bandpass filter electromagnetic model is achieved in conjunction with metal-insulator-metal (MIM) capacitors. Measured results of the on-chip filter through the probe station and vector network analyzer show that the bandpass filter achieves a center frequency of 4 GHz, a fractional bandwidth of 20%, and an insertion loss of 2.35 dB. The stopband rejections are better than 44 dB from lower stopband DC to 2 GHz and 31 dB from upper stopband 6~11 GHz, showing excellent stopband rejection performance. The overall size of the on-chip bandpass filter is only
- 摘要：Reflectarray antennas have been widely used in satellite communication, microwave remote sensing, and radar systems due to their advantages of high gain, low cost, and flexible beam control. However, with the ever-increasing demands for anti-interference capability, signal quality, system reliability, and operational safety in modern wireless communication and sensing systems, achieving low sidelobe level (SLL) with high aperture efficiency has become a critical research focus in reflectarray antenna design. While conventional phase-only synthesis methods are limited in their ability to suppress sidelobes, existing amplitude-phase joint control techniques often suffer from significant efficiency losses due to the poor consistency of element reflection amplitude. To address the above problems, this paper proposes a metasurface-based reflectarray antenna design with low SLL and high aperture efficiency. A novel non-uniform line-width I-shaped metasurface reflection element has been developed. By non-uniformly designing the line widths of the symmetric split ring and cut wire to control the surface current distribution under different resonant modes, the fluctuation of reflection amplitude during the phase tuning process is effectively suppressed. Full-wave simulation results demonstrate that the proposed element not only achieves continuous phase tuning of 0°~360° and amplitude control of 0~1, but also restricts the reflection amplitude fluctuation to less than 0.03 during phase tuning. This superior amplitude consistency is beneficial for achieving high aperture efficiency in reflectarray antennas. Based on the proposed element and Taylor weighting technique, a linearly polarized reflectarray antenna with circular aperture of 193.3 mm in diameter (
- 摘要：Phased array systems with wideband and wide-beam scanning capabilities have become critical equipment in the fields of electronic warfare and broadband communication. This study designs and implements a wideband and wide-beam phased array system, aiming to meet the design requirements for wideband beam broadening in complex environments while achieving system miniaturization and lightweight construction. The system adopts a 6 × 14 array configuration, utilizing miniaturized log-periodic dipole antennas as the basic radiating elements. By bending the dipole arms to extend the current path and combining this with dielectric substrate perforation techniques, the physical dimensions of the antenna elements and the overall system weight are effectively reduced. Additionally, the use of a carbon fiber ground plane further decreases the system weight. The core innovation of this paper lies in constructing a phased array system capable of simultaneously achieving wideband and wide-beam performance. By integrating attenuators, phase shifters, and true time delay lines within the array, combined with beamforming algorithms, stable wide beams and precise wide-angle scanning are achieved under instantaneous wide bandwidth conditions. This design simultaneously overcomes the impact of excessive mutual coupling between elements on active standing wave ratio, which is a challenge in tightly coupled array solutions, thereby ensuring radiation stability across the wide frequency band. The multi-channel T/R modules adopt a “tile-type” structure to enhance integration density. The true time delay lines compensate for phase differences introduced by phase shifters at different frequencies, working together with the phase shifters to form a wideband beam steering network. This effectively avoids the problems of beam squint and signal dispersion that occur in traditional phased arrays during wideband and wide-angle scanning due to the frequency sensitivity of phase shifters. System-level test results demonstrate that the phased array system operates stably across the full 0.8~2 GHz band. Within every 500 MHz instantaneous bandwidth, it achieves beam scanning capabilities of ±20° in the elevation plane and ±40° in the azimuth plane. Concurrently, the azimuth beamwidth is no less than 10°, and the elevation beamwidth is no less than 20°, exhibiting stable wide-beam performance across the broad frequency range. Regarding radiation performance, with an element input power of less than 10 W, the system’s effective isotropic radiated power exceeds 41 dBW, and the overall system weight is significantly reduced compared to traditional designs. The consistency between simulated and measured results validates the design’s effectiveness. The measured stable beam pointing and controllable sidelobe levels confirm the system’s engineering practicality in complex electromagnetic environments. Furthermore, the system maintains good radiation efficiency and stable gain characteristics across the wide frequency band, further verifying its reliability and adaptability in practical applications. The wideband and wide-beam phased array system developed in this study demonstrates significant advantages in beam coverage range, scanning stability within instantaneous bandwidth, and system lightweighting. It provides an effective solution for the engineering application of wideband arrays domestically, and this design approach holds promise for application in communication and detection systems operating at higher frequency bands.关键词：phased array antenna system;active;broadband antenna;array antenna;beamforming;wide beamwidth45|88|0更新时间：2026-05-28
- 摘要：Arrhythmia, as a key trigger for sudden cardiac death in cardiovascular diseases, holds significant clinical importance for improving patient outcomes through its early, precise identification and dynamic classification. However, due to factors such as individual variability, differences in recording devices, and variations in recording environments, significant domain shift issues are prevalent across ECG signals from different patients and databases. Additionally, clinical ECG data commonly exhibit highly imbalanced class distributions, severely limiting the generalization performance of existing models across patients and databases. To address these challenges, this paper proposes a lightweight, fast inter-patient domain adaptation framework for arrhythmia recognition tailored to class-imbalanced ECG signals. First, at the feature level, we introduce BLS-FDDA (broad learning system-feature distribution domain adaptation), a domain adaptation method based on the broad learning system (BLS). By leveraging covariance normalization and distribution reconstruction techniques, it thoroughly analyzes the offset between source and target domain features, successfully aligning their feature spaces. By aligning the distributions of the BLS feature expansion matrix, this method avoids reliance on complex models inherent in traditional deep learning frameworks while ensuring effective feature information transfer. Second, at the data level, this paper further proposes a reversible data domain adaptation method, BLS-DRDA (BLS-data reversible domain adaptation). Integrating error perturbation theory, it derives the data transformation relationship between source and target domains. Based on this theoretical derivation, the method achieves rapid adaptation to new data domains without retraining the BLS main model, significantly reducing transfer costs. Moreover, BLS-DRDA preserves the discriminative capability of the original signal during data transformation while effectively preventing information distortion. At the decision layer, addressing the severe class imbalance in arrhythmia data, a cost-sensitive decision algorithm is designed. By introducing concepts of class center distance and sample distribution weights, this algorithm establishes a weighted decision mechanism that effectively mitigates misclassification issues of minority samples during cross-domain transfer. Finally, multi-patient, multi-database, and continuous domain transfer experiments conducted on the MIT-BIH and INCART public datasets demonstrate that the proposed methods achieve recognition performance approaching 100% in metrics such as accuracy, F1_score, and G_mean, significantly outperforming the original width learning model and various comparative methods. Theoretical analysis and experimental results demonstrate that the proposed BLS-FDDA and BLS-DRDA methods exhibit superior performance across multi-source data, cross-device, and continuous domain adaptation scenarios. This validates the framework’s effectiveness and practicality in complex clinical ECG applications, particularly in the task of identifying arrhythmias across multiple patient categories. The proposed methods substantially enhance recognition capabilities for minority classes and demonstrate exceptional robustness under complex domain shifts and class imbalance challenges.关键词：arrhythmia detection;domain adaptation;class imbalance;broad learning system65|83|0更新时间：2026-05-28
- 摘要：This paper investigates the adaptive output consensus problem of heterogeneous open multi-agent systems under actuator attacks. In complex application scenarios such as the industrial internet and modern cluster systems, multi-agent systems often exhibit open characteristics due to task reconfiguration, fluctuations in communication links, node faults, and dynamic joining/leaving. Specifically, the node set and the overall system dimension may evolve over time, and the communication topology may exhibit time-varying behavior, intermittent connectivity, and reconstruction-induced switching. These characteristics make both the collective dynamics and the available information structure strongly time-varying and uncertain. Meanwhile, distributed control signals typically need to be transmitted through networked channels and received and executed at the actuator. Once an attacker injects malicious data into the actuators, the control inputs can be superimposed with unknown disturbances or even tampered into misleading commands, causing the agents’ actual behaviors to deviate from their intended actions and thereby destroying cooperative consensus. These issues significantly increase the difficulty of distributed controller design and consensus analysis. To address these challenges, an attack-free reduced-order observer is designed based solely on output information. Without explicitly estimating actuator attacks, the proposed observer provides the local state information required for control design, thereby reducing the estimation cost. Building upon this, a fully distributed adaptive resilient control protocol is developed. The protocol relies solely on local observations and interaction information exchanged among neighboring agents to adjust the coupling gains and attack-compensation parameters online, achieving both suppression of attack effects and adaptive handling of topology uncertainties. Notably, the proposed method requires neither any global topology information nor prior knowledge of attack bounds. Thus, it is well suited to open network environments with dynamically changing numbers of nodes and topology switching, offering excellent scalability and practical implementability. To characterize the structural properties and error propagation mechanism of heterogeneous open multi-agent systems, a unified analytical framework is established to describe the evolution of the tracking error. Based on this framework, an average dwell-time condition is derived to guarantee output consensus of the system under actuator attacks, thereby revealing the relationship between the topology switching rate and the system’s convergence performance. A numerical simulation is provided to demonstrate the effectiveness of the proposed approach and to compare it with existing algorithms. The results show that, under the same attack conditions and network settings, the proposed method effectively mitigates the impact of attacks on collective behaviors and improves the convergence speed, thereby demonstrating the effectiveness and advantages of the proposed reduced-order observer and fully distributed adaptive resilient control strategy.关键词：open multi-agent system;resilient consensus;distributed tracking;adaptive control;actuator attack42|24|0更新时间：2026-05-28
- 摘要：The SM9 cryptographic algorithm, independently developed by China, serves as a critical commercial cryptography standard and national standard, playing a key role in the localization and substitution of cryptographic algorithms. However, the original SM9 digital signature, as an identity-based signature mechanism, suffers from the limitation of not supporting hierarchical features. In large-scale network environments, this can easily lead to excessive pressure on the key generation center and network congestion due to the surge in the number of users. To address this challenge, this paper proposes the first hierarchical identity-based signature scheme based on the national cryptographic SM9 algorithm. The scheme innovatively introduces hierarchical signature technology, distributing the key generation task across multiple levels of nodes, effectively alleviating the private key generation and distribution pressure on the key generation center. It is ideally suited for large-scale, multi-layered network scenarios such as the Internet of Vehicles and blockchain. In terms of technical implementation, to adapt to the SM9 algorithm, this scheme is based on prime-order groups and employs efficient hierarchical technology to allocate signature private keys from the upper level to the next level, thereby forming a hierarchical private key update mechanism. Subsequently, users generate digital signatures based on their signature private keys. The resulting signature value consists of only three group elements, adding just one more group element compared to the original SM9 algorithm, and achieves a constant signature length independent of the hierarchical depth. Under the random oracle model, a rigorous security proof of the scheme is provided, demonstrating that the scheme satisfies existential unforgeability under chosen message and identity attacks, and its security can be reduced to the (q,n)-SDH hardness problem. Theoretical analysis and experimental results show that the proposed scheme has significant advantages in signature generation and verification efficiency. As the number of system layers k increases, the signature generation and verification time of this scheme tends to remain constant, significantly outperforming existing hierarchical identity-based signature schemes based on bilinear pairings. Specifically, when the number of system layers k ranges from 2 to 10, the signature generation and verification times are approximately 2.24 ms and 36.08 ms, respectively, improving efficiency by 0.03 to 2.79 times and 0.87 to 1.27 times compared to the current optimal hierarchical signature schemes. Moreover, as k increases, the efficiency improvement becomes more pronounced: when k is 100, the signature generation and verification efficiency are enhanced by approximately 34 times and 4.5 times, respectively, compared to the existing optimal hierarchical signature schemes. Finally, the proposed scheme is applied to the identity authentication scenario of the Internet of Vehicles, successfully resolving network congestion caused by the surge in users in the IoV environment. It realizes a lightweight and decentralized identity authentication mechanism, providing crucial technical support for building efficient and secure large-scale network environments with localized cryptographic solutions.关键词：hierarchical identity-based signature;SM9;identity-based signature;internet of vehicles identity authentication;fixed length signature27|29|0更新时间：2026-05-28
- 摘要：The rapid advancement of large language models (LLMs) has created new opportunities for smart contract vulnerability detection. To exploit LLMs’ semantic understanding while overcoming single-modal limitations, a multimodal smart contract vulnerability detection method (MVul-L) is proposed. The proposed method integrates textual, structural, and visual modalities for comprehensive modeling of smart contracts. First, a task prompting template for LLM-based vulnerability detection is designed. It contains seven key fields that clearly define analytical objectives, prompting strategies, and input-output specifications, reducing model understanding bias. Second, a semantic-enhanced textual feature extraction method based on LLM reasoning and CodeBERT encoding is developed. The LLM interprets contract logic through the task template and generates semantically annotated textual outputs. Both source code and annotations are fed into CodeBERT to obtain enriched textual representations. Finally, a graph attention network (GAT) and a convolutional neural network (CNN) are employed to model structural and visual features, respectively. A Transformer-based multimodal fusion mechanism is further adopted to achieve deep cross-modal integration. Experimental results on public datasets demonstrate that the overall performance of MVul-L surpasses existing methods. In reentrancy, timestamp dependency, and integer overflow vulnerability detection tasks, the F1 score is improved by 3.51%~9.40%, confirming the effectiveness of the proposed method.关键词：large language model;smart contract;vulnerability detection;multimodal fusion;Semantic Enhancement59|37|0更新时间：2026-05-28
-
Identification and Migration of Silent Security Patches in Blockchain Systems via Information Fusion
摘要：Blockchain is a novel distributed system integrating cryptography and smart contracts. It has been widely applied in fields such as financial transactions and copyright protection. Currently, thousands of different blockchains are running worldwide. For compatibility and convenience, many developers fork or reuse the open-source code of mainstream blockchains for further development. However, such a practice also leads to the rapid propagation of security vulnerabilities. Meanwhile, silent security patches refer to security fixes in open-source projects that are not publicly disclosed in vulnerability databases. At present, the transparency of security patches in blockchain projects is insufficient, and there are a large number of silent security patches, further exacerbating the remediation delays in downstream software systems and reducing the reliability of the entire blockchain ecosystem. Therefore, it is urgent to design an automated method for identifying silent security patches targeting the blockchain ecosystem covering multiple programming languages, to promptly detect and fix potential known security issues. To this end, this paper proposes BlockPatch, the first framework for identifying and migrating general silent security patches in the blockchain ecosystem. By fusing multimodal change information through the large language model (LLM), it achieves accurate identification of security patches in multilingual blockchain systems. Specifically, taking code commits as input, BlockPatch extracts commit messages, modified code blocks, and abstract syntax tree (AST) edit actions to obtain multimodal change representations, so as to capture fine-grained change contents and processes. Subsequently, it utilizes the advanced representation capabilities of LLMs to perform semantic embedding on the three types of information and combines neural networks to achieve feature fusion and learning, thereby enhancing the identification capability of security patches. To verify the effectiveness of the method, this paper constructs a patch dataset containing mainstream public and consortium blockchain projects. The experiments show that BlockPatch can achieve a precision of 94.02%, a recall of 94.58%, and an F1-score of 94.29%, outperforming existing state-of-the-art methods by 5.03 percentage points on the F1-score, and achieves good results in identifying different types of security patches. The ablation study further demonstrates the effectiveness of multimodal information fusion. Finally, BlockPatch migrates the identified security patches to downstream blockchain systems for security checks. Based on recent code commits from Bitcoin and Ethereum repositories, BlockPatch identified 16 silent security patches and discovered 28 unpatched security vulnerabilities in downstream projects, highlighting the importance of identifying and applying silent security patches.关键词：blockchain system;security patch identification;large language model;patch information fusion;multimodal learning;patch migration35|26|0更新时间：2026-05-28 - 摘要：Retrieval-augmented generation (RAG), which integrates large language model (LLM) with retrieval systems, has become a mainstream solution for LLM deployment due to its advantages in traceability, interpretability, and low cost for knowledge updates. However, before deployment, RAG systems require rigorous evaluation: developers must construct large-scale knowledge bases and conduct comprehensive tests with thousands of queries, involving intensive retrieval computations and repeated LLM calls. This results in extremely time-consuming evaluation processes, severely hindering the development and iteration of enterprise-level AI systems. To address this bottleneck, we propose RGE-Pipeline (Retriever-Generator-Evaluator Pipeline), a high-throughput, scalable, and fast evaluation framework for LLM-based RAG algorithms. We first conduct preliminary experiments to quantitatively analyze the performance bottlenecks in typical RAG systems, identifying the retriever as the initial bottleneck. By replacing the traditional BM25 algorithm with BM25S, we reduce the retrieval time proportion to below 4%, shifting the bottleneck to the generation and evaluation stages. Building on this, RGE-Pipeline performs system-level optimization from three aspects: (1) decoupling the evaluation workflow into three modules—retriever, generator, and evaluator—and introducing a pipeline parallel architecture to eliminate serial waiting and repeated model loading overhead; (2) designing a fine-grained hardware resource management scheme based on the vLLM inference framework to support concurrent deployment of multiple LLM instances on the same set of GPUs; (3) constructing a mathematical model that reveals the quantitative relationship between the memory allocation ratio of the generator and evaluator and the overall system throughput, proposing three GPU resource allocation strategies—shared VRAM, full-card allocation, and hybrid partitioning—to maximize throughput by balancing the computational load between the two stages. Experiments conducted on the CRUD-RAG dataset, which covers tasks such as text continuation, summarization, multi-document question answering, and hallucination modification with a total of 6 400 queries, demonstrate the significant acceleration achieved by RGE-Pipeline. Under the fixed configuration using BM25S and Qwen2.5-7B for both the generator and evaluator, the hybrid partitioning scheme reduces the total evaluation time from approximately 95 hours (original serial workflow) to 1.3 hours, achieving a speedup of 71.7×, and from approximately 10.8 hours (model-preloading workflow) to 1.3 hours, achieving a speedup of 8.2×. Furthermore, extensibility experiments confirm that RGE-Pipeline maintains strong adaptability across different knowledge base sizes (ranging from 18 KB to 60 MB) and on small-scale query sets. In summary, RGE-Pipeline not only significantly reduces the validation cost of RAG algorithms but also provides a reference design for system optimization in multi-LLM parallel inference scenarios.关键词：Large Language Model (LLM);retrieval-augmented generation (RAG);Best Matching 25 (BM25);knowledge-based QA;vLLM;CRUD-RAG35|30|0更新时间：2026-05-28
- 摘要：This paper addresses the issue of low computational efficiency in solving electromagnetic responses for multiple right-hand sides in metal-dielectric composite structures by proposing an efficient and stable direct solver algorithm. First, based on the M-HODLR framework, strategies for grouping basis functions and constructing the M-HODLR algorithm suitable for metal-dielectric composite structures are investigated. Subsequently, high-order electromagnetic modeling is integrated into the proposed M-HODLR algorithm for such structures, significantly reducing the number of unknowns while ensuring high solution accuracy, thereby effectively improving computational efficiency. Finally, to address the basis function grouping issue in high-order electromagnetic modeling, a grouping strategy based on clusters of high-order basis functions is developed. By reordering the basis functions across different layers and incorporating aggregation operations, upward aggregation for metal-dielectric composite problems within the M-HODLR framework is achieved. The proposed M-HODLR algorithm with high-order electromagnetic modeling significantly accelerates the solution of electromagnetic responses for metal-dielectric composite structures. Theoretical analysis and numerical examples validate the effectiveness of the proposed method.关键词：metal-dielectric composite structures;high-order basis functions;M-HODLR;multiple right-hand sides31|61|0更新时间：2026-05-28
- 摘要：Pedestrian trajectory prediction is crucial for improving the decision-making capabilities of autonomous vehicles and service robots, as well as mitigating potential future collision risks. However, due to the differences and complexity of social interactions within and outside pedestrian groups, existing studies often fail to explicitly distinguish and independently model interactions within and outside groups. This leads to confusion of different types of interaction features during model learning, making it difficult to accurately depict the real motion patterns of pedestrians in complex scenarios, thus restricting further improvement in model prediction performance. Therefore, this paper proposes a pedestrian trajectory prediction model based on an adaptive group masked graph convolution network (AGMGCN). By independently modeling in-group and out-group interactions, the accuracy of model trajectory prediction is improved. The model first constructs a social graph and processes it using a self-attention mechanism to obtain an attention matrix that initially represents the interactions between pedestrians. Subsequently, a time-frequency domain convolution module is designed to further process the attention matrix in both the time and frequency domains, generating a time-frequency interaction matrix to characterize the spatial-temporal interactions of pedestrians, thus achieving a more accurate portrayal of complex dynamic interactions. To effectively distinguish and independently model in-group and out-group interactions, an adaptive group masking module is designed. This module adaptively determines a threshold based on the feature similarity between pedestrians and generates in-group and out-group masking matrices through threshold processing, providing support for the subsequent independent modeling of in-group and out-group interactions. Building upon this foundation, the time-frequency interaction matrix is combined with out-group and in-group masking matrices, and graph convolution is applied to respectively capture in-group and out-group interaction features, thereby achieving independent modeling of pedestrian in-group and out-group interaction relationships. Finally, a feature fusion module is designed to dynamically weight and fuse in-group and out-group interaction features and then temporal convolution networks are used to predict the future trajectory of pedestrians. Experimental results on the ETH, UCY and SDD datasets demonstrate that, with only 23.9 K model parameters, the proposed method reduces the average displacement error by 12% and the final displacement error by 20% compared to DSTIGCN, validating the advantages of the proposed method in terms of prediction accuracy and computational cost.关键词：pedestrian trajectory prediction;adaptive group masking;graph convolution network;time-frequency domain convolution module;temporal convolution networks38|23|0更新时间：2026-05-28
- 摘要：Wind turbine blades are the core aerodynamic components responsible for wind energy capture and energy conversion in the wind turbine systems. Surface defects such as cracks, erosion, and delamination can deteriorate aerodynamic performance and consequently reduce power generation efficiency. In real-world wind farm inspection scenarios, blade surface detection is typically characterized with small scales and low contrast, and are often accompanied by complex backgrounds, illumination variation, and imaging noise. These challenges significantly limit the performance of existing end-to-end detection methods exhibit in small-scale defect detection tasks for wind turbine blades. Although the benchmark evaluations under laboratory conditions have achieved localization accuracies exceeding 99%, the strong coupling between localization and classification under strong background interference and micro-cracks makes these methods can hardly be further improved. To overcome these challenges, we propose Spatial-FineDef (Spatial-Fine Defect Detection Approach), a hybrid two-stage detection method incorporating multi-scale perception and adaptive enhancement. Through explicitly decoupling defect localization from fine-grained classification, the proposed method optimizes candidate region extraction and defect recognition in a staged manner. In the first stage, Spatial-Net incorporates task-oriented data augmentation strategy and improved localization methods to enhance the accuracy of spatial filtering for potential defects. In the second stage, FineDef-Net utilizes ConvNext backbone with lightweight multi-scale linear attention mechanism which enhances the discrimination capability while maintaining low computational complexity. Compared with end-to-end detection methods, by adopting a two-stage strategy of localization followed by classification, Spatial-FineDef effectively suppresses background interference while enabling stable region selection and defect classification for small-scale defects. On a field-collected wind turbine blade fault dataset, Spatial-FineDef achieves an overall accuracy of 96.71% in the classification of four small-scale defect types including surface pitting, coating shedding, edge cracking, and surface cracking. Experimental results demonstrate outperforming multiple representative baseline models across several evaluation metrics. Ablation studies further validate the effectiveness of the decoupled two-stage strategy and the multi-scale linear attention mechanism in handling small defects under complex backgrounds. The proposed method provides a deployable and reliable solution for on-site blade inspection, facilitating real-time fault diagnosis and enhancing the reliability and intelligent operation and maintenance of wind turbine systems.关键词：wind turbine blade;defect detection;end-to-end;multi-scale perception;adaptive enhancement38|34|0更新时间：2026-05-28
- 摘要：Remote sensing imagery is often severely degraded along the imaging chain by heterogeneous factors such as atmospheric scattering, sensor noise, and extreme illumination conditions. Existing all-in-one restoration approaches predominantly rely on implicit feature learning, lacking explicit modeling of the physical frequency-domain properties of degradation and the ability to capture higher-order semantic interactions. To address these limitations, we present Aether, a unified remote sensing image restoration framework that integrates frequency-domain physical perception with higher-order semantic fusion. Aether introduces a harmonic-adaptive degradation analyzer (HADA), which replaces fixed-basis transforms with a data-driven learnable harmonic filter bank. This enables adaptive parsing and precise extraction of spectral fingerprints associated with diverse degradation types. In addition, we design a higher-order nonlinear interaction fusion module (HONIF) grounded in the Kolmogorov-Arnold representation theorem. HONIF constructs a high-dimensional mapping space via spline-based function networks, overcoming the representational bottleneck of conventional linear attention and facilitating deep semantic alignment between degradation priors and image features. Experiments on three benchmark datasets—MD-RSID, MD-RRSHID, and MDRS-Landsat—demonstrate that Aether achieves state-of-the-art performance across haze removal, denoising, deblurring, and low-light enhancement tasks. Notably, on the MDRS-Landsat dataset, Aether surpasses the second-best method by 3.63 dB in peak signal-to-noise ratio (PSNR) for dehazing and by 1.60 dB for low-light enhancement, while improving learned perceptual image patch similarity (LPIPS) by 75.2%, effectively addressing the long-standing challenge of unified and generalizable restoration in complex remote sensing scenarios.关键词：remote sensing image restoration;all-in-one image restoration;harmonic-adaptive;higher-order nonlinear interaction41|70|0更新时间：2026-05-28
- 摘要：To address the issue of high entangling and insufficient discriminative ability of the features extracted by image classification networks, which limits the expressive power of key features, this paper proposes an image classification network with attenuation disentangling mechanism (ADMNet). Firstly, based on the differential response characteristics of biological neurons to signal intensity, an attenuation disentangling mechanism is proposed: the spatial attenuation disentangling (SAD) module is designed to decompose the feature map into independent subspaces and perform attenuation transformation with different thresholds, effectively disentangling and purifying key features and filtering redundant information; the channel disentangling (CD) module is designed to use multi-scale one-dimensional convolution to model the channel entangling relationships within different ranges, dynamically enhancing feature channels related to classifications and suppressing irrelevant channels; then, by integrating the SAD and CD modules, the feature attenuation disentangling (FAD) module is formed, which achieves effective disentangling of image features through the joint operation of dual-branches, enhancing the discriminability of key features, and thereby improving the nonlinear expression ability of image features. Then, a feature aggregation pooling (FAP) module is constructed. It aggregates multi-scale features extracted by different convolutions, enriches feature representations, improves spatial information utilization, and reduces the size of feature maps. Finally, the FAD and FAP modules are embedded into the main path and residual branches of the residual block, respectively, allowing the network to learn features with lower coupling and stronger discriminability, as well as rich contextual information. This reduces information loss during network transmission and enhances the network’s generalization ability. The method proposed in this paper achieves classification accuracies of 96.6%, 80.6%, 97.5%, 89.6%, and 83.1% on CIFAR-10, CIFAR-100, SVHN, Imagenette, and Imagewoof datasets, respectively. Experimental results show that ADMNet can effectively decouple image features, enhance feature discriminability, reduce the risk of information loss, and improve image classification ability.关键词：image classification;attenuation disentangling;feature disentangling;multi-scale features;residual branch58|83|0更新时间：2026-05-28
- 摘要：The widespread adoption of TLS 1.3 and QUIC renders payload content invisible, shifting traffic analysis toward side-channel features. However, rigorous justification for “why side-channel leakage is inevitable in encrypted communications” has long been lacking. This paper establishes a strict foundation from information theory and system design by constructing a formal model
- 摘要：Joint multimodal aspect-based sentiment analysis (JMASA), a crucial research direction in fine-grained sentiment analysis, aims to jointly identify specific aspect terms and their corresponding sentiment polarities from image-text pairs, and has garnered increasing attention in recent years. Although this task holds significant application value in areas such as social media analysis and product review mining, existing methods primarily face two challenges: First, when leveraging pre-trained language models to fuse multimodal information, models often exhibit excessive trust in certain irrelevant visual or textual tokens, allocating unnaturally high attention scores, which interferes with capturing key emotional cues; Second, existing methods struggle to explicitly model the complex relationships between objects within an image and lack an effective mechanism to mine deep semantic interactions and dependencies at the object level between images and text. To address these issues, this study proposes a novel scene graph-enhanced method for joint multimodal aspect-based sentiment analysis based on attention penalty and adaptive learning, named APALSG, which utilizes scene graph generation (SGG) to enhance the analysis. Specifically, the method primarily employs a specially designed attention penalty strategy to penalize and attenuate high attention scores that exceed a predefined threshold, redistributing the attenuated attention values to neighboring tokens within their contextual window. This strategy dynamically adjusts the model’s attention distribution, effectively mitigating the over-focus on irrelevant information, thereby extracting more precise key object features. Furthermore, a scene graph modeling module is designed, which incorporates graph convolutional networks (GCN) to perform message propagation and aggregation on this scene graph, obtaining visual representations enriched with contextual information about inter-object relationships. Finally, an adaptive learning strategy is also designed, enabling the model to adaptively focus on the potential dependencies between the image-text pair relevant to the current aspect, achieving deep cross-modal alignment and fusion. The fused multimodal features are then fed into a classifier to simultaneously perform joint prediction for aspect term extraction and sentiment classification. To comprehensively validate the effectiveness of APALSG, experimental results on multiple publicly available benchmark datasets demonstrate that APALSG significantly outperforms existing state-of-the-art methods, confirming its efficacy. Compared to existing JMASA models, APALSG shows superior performance on the Twitter-2015, Twitter-2017, and MACSA datasets, improving precision by 1.46%, 2.18%, and 1.19% respectively.关键词：joint multimodal aspect-based sentiment analysis;attention penalty;scene graph generation;graph convolutional network;pre-trained language models43|25|0更新时间：2026-05-28
- 摘要：Current scheduling algorithms for the multipath quick user datagram protocol (UDP) internet connections (MP-QUIC) protocol overlook the priority relationships among streams, making it ineffective in distinguishing critical streams from ordinary ones in heterogeneous networks. This results in the blocking of critical webpage loading streams and significantly impacts user experience. Therefore, this paper proposes a priority-weighted and packet arrival time based scheduling (PW-PATS) algorithm, which enhances the performance of MP-QUIC for critical and overall service transmission in heterogeneous network environments. The PW-PATS algorithm quantifies the weight of quick UDP internet connections (QUIC) streams into a priority factor (PF) and uses it as a weighting factor in the calculation of packet arrival time (PAT), forming the core path selection criterion of weighted PAT (W-PAT). This prioritizes the scheduling of high-priority packets to network paths with higher channel quality. Experimental results based on webpage simulation responses demonstrate that compared to the existing lowest round-trip time first (LowRTT) scheduling algorithm, the proposed algorithm significantly improves the transmission efficiency of critical streams. In the traditional webpage access pattern scenario, it reduces the completion time of high-priority hypertext markup language (HTML) streams by 69%, while improving the overall webpage rendering time by 24%. In webpage parallel loading mode scenarios, it reduces the completion time of high-priority cascading style sheets (CSS) streams by 79.8%, and also achieves an improvement of up to 48.9% in overall webpage rendering time under various network conditions.关键词：MP-QUIC;stream priority;heterogeneous network;scheduling algorithm;path scheduling39|58|0更新时间：2026-05-28
- 摘要：Recently, self-attention-based sequential recommendation models have demonstrated remarkable effectiveness in user behavior modeling. However, these models tend to suffer from an over-smoothing problem during deep encoding. Repeated aggregation across multi layers makes high-order representations become increasingly similar, which gradually weakens personalized signals. Meanwhile, the high correlation among feature dimensions introduces redundancy and noise propagation, which weakens the model’s ability to identify important features and consequently limits its generalization capability. To address these challenges, this work proposes feature masking and contrastive learning integrating multi-dimensional decorrelation in sequential recommendation (MCMD-SR), a feature masking and contrastive learning model integrating multi-dimensional decorrelation in sequential recommendation. Firstly, we design a feature masking mechanism based on self-attention. This mechanism measures the contribution of each feature dimension with attention scores. It then selectively masks features with low-contribution which are prone to inducing representation homogenization. In addition, we also introduce a logarithmic mask-rate decay strategy across layers. This strategy applies stronger perturbations in shallow layers to break high-similarity features locally. In deeper layers, it maintains moderate perturbations to continuously suppress excessive aggregation. Furthermore, a contrastive learning task is constructed between the masked final-layer representations and the original shallow-layer representations. The proposed method pulls together positive pairs from the same sequence and pushes apart negative pairs from different sequences or feature dimensions. The proposed method reinforces discriminative and personalized semantics. Thereby it improves the separability of the embedding space. Secondly, we propose a multi-dimensional adaptive decorrelation module. Based on the attention-masked feature matrix, Pearson correlation coefficients (PCC) are computed from both column-wise and layer-wise perspectives. Penalty weights are adaptively assigned according to the correlation strength. This suppresses redundant dimensions and inter-layer dependencies. Meanwhile, it keeps the overall regularization strength controllable. This dual-view decorrelation strategy reduces feature redundancy from both local (column-wise) and global (layer-wise) perspectives resulting in improving the identification of key features. Finally, the self-attention masking mechanism, the contrastive learning module, and the multi-dimensional adaptive decorrelation module are jointly optimized in a multi-task learning framework. These complementary constraints stabilize training and improve embedding quality as well as model generalization. Extensive experiments are conducted on 4 public datasets, where the proposed method is compared with 11 classical and state-of-the-art sequential recommendation models. Experimental results show that MCMD-SR achieves average improvements of 2.13% and 1.67% over the strongest baseline in terms of hit ratio (HR) and normalized discounted cumulative gain (NDCG), respectively. In addition, ablation studies and parameter sensitivity analysis further verify the necessity of each module and their synergistic effectiveness, thereby further clarifying the strong generalization capability of our model.关键词：sequential recommendation;feature masking;contrastive learning;decorrelation;self-attention mechanism;correlation metrics220|76|0更新时间：2026-05-28
PAPERS
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0