最新刊期
卷 54 ， 期 6 ， 2026
-
- 摘要：As the complexity of integrated circuit design grows exponentially, traditional electronic design automation (EDA) tools face unprecedented challenges. In recent years, breakthrough advances in large language models (LLMs) in the field of natural language processing have brought new opportunities to the EDA domain. This paper systematically reviews the application of LLMs in EDA, categorizing them into five main directions: register transfer level (RTL) code generation and optimization, high-level synthesis and logic optimization, physical implementation, verification and security, analog circuit design automation, and EDA tool flow automation and assistance. This paper also analyzes the main challenges faced by LLMs in the EDA domain, including technical issues such as scarcity of high-quality domain data and insufficient multimodal processing capabilities, heterogeneous tool chain integration barriers, and verification of generated content correctness and security, as well as application constraints including intellectual property risks, data privacy protection, and ethical compliance. Future research directions mainly focus on domain-specific corpus construction, cross-modal fusion model architecture, standardized tool interface protocols, and industry standards and ethical frameworks.关键词：large language models;electronic design automation;Hardware Description Language;code generation;Retrieval-Augmented Generation317|193|0更新时间：2026-10-16
- 摘要：The radar cross section (RCS) of an array antenna equals the product of the antenna scattering element factor and the array factor. Typically, sparse array arrangement optimization is used to suppress the peak value of the scattering array factor. In recent years, deep learning-based methods for sparse array RCS reduction have effectively improved the real-time stealth performance. However, sparse optimization cannot mechanically remove the elements from the array in real time. The load matching and active cancellation methods are used as the equivalent method to delete the passive elements, but these methods cannot reduce the RCS of passive elements to zero, they can only reduce it to a small value. Existing deep learning-based sparse array RCS reduction methods are designed for scenarios where the RCS of passive elements is zero. So, these existing methods are unsuitable for practical applications. To address this issue, this paper proposes a sparse array dual-subarray RCS reduction method based on a light kernel network with multimodal and augmentation attention mechanism (MAAM-LKN). Specifically, considering that the RCS of passive elements is usually non-zero, a passive RCS factor of element is introduced. A problem model and objective function for sparse array dual-subarray optimization are established, and the MAAM-LKN model simultaneously optimizes the array arrangements of both the active and passive subarrays. This reduces the RCS of both subarrays and ultimately minimizes the overall RCS of the array. Moreover, since existing neural network models for sparse array RCS reduction lack focus on important information, an augmented attention mechanism is designed in the MAAM-LKN model to weight features according to their importance. This addresses the issue of insufficient emphasis on key features during feature propagation. Additionally, to enhance the model’s comprehension capability, the multimodal is introduced to achieve more effective extraction and perception of key features. Simulation results show that under different passive element RCS factor values, dual-subarray optimization can effectively improve the RCS reduction performance compared to single-subarray RCS optimization. The maximum RCS reduction value can be improved by more than 2.72 dB, and the average RCS reduction value can be improved by more than 0.19 dB. When the RCS of passive elements is non-zero, the MAAM-LKN model achieves an accuracy of 91.09% and an average RCS reduction of 5.9 dB, both higher than other neural network models, while maintaining comparable model complexity.关键词：RCS reduction;passive RCS factor of element;dual-subarray of sparse array;deep learning;attention mechanism;multimodal108|117|0更新时间：2026-10-16
- 摘要：This work introduces a hardware parallel-gradient boolean satisfiability solver, Fold-SAT, enabled by compute-in-memory technology and multilevel resistive memory storage. As shown in its name, Fold-SAT folds multiple clauses into one column of the compute-in-memory array. Compute-in-memory can utilize its computing parallelism to accelerate the solving process of satisfiability problems on hardware. A conventional in-memory SAT-solver only stores one clause per column. It shows high sparsity when stored in memory and usually occupies large memory capacity. During computing iterations, the compute-in-memory accelerator conducts multiply-and-accumulate operations on the array to calculate the satisfiability result of each clause. Fold-SAT uses multiple resistance states to represent the structural information of different clauses. It stores the information of multiple clauses in each column and decodes the satisfiability results of the clauses during readout process. Parallel gradient algorithm is an efficient algorithm for satisfiability problems. The clause-folding scheme applies to both the forward-pass array and the backward-path array in parallel gradient algorithm. To explore the compression result of Fold-SAT, this work designs the resistance ratio of multilevel storage and examines its effects on data compression. Results reveal that, when the forward-pass array uses a resistance ratio of 1:4:16, it reduces the array size to 1/3 of conventional methods and reduces the overall storage sparsity from 97.0% to about 92.8%. The storage sparsity refers to the portion of cells that do not store effective information. Moreover, this work explores the data compression in the backward-pass array. It reduces the required array size to half that of conventional methods. This work discusses the impacts of device non-idealities. Simulated results show that the device-programming algorithms should reduce the device errors to less than 1 μS, and the deviation of conductance mean values should be controlled within 10%. There is a series of design spaces regarding the multilevel devices for the backward-pass array. It could reach acceptable computing precision when the resistance ratio is set between 1:8 to 1:16 for the backward-pass array. Simulation results show that this work achieves a median solving time of 19.74 μs, which is close to the leading work in the literature. This illustrates that this work could reduce the memory capacity cost while maintaining comparatively high solving speed.关键词：satisfiability;resistive memory;in-memory computing;data compression;multilevel storage93|106|0更新时间：2026-10-16
- 摘要：Physical reservoir computing (PRC), characterized by its low training cost and high parallel efficiency, has emerged as a promising paradigm for temporal signal processing and complex dynamical system modeling. By leveraging the intrinsic nonlinearity and short-term memory dynamics of physical devices, PRC maps temporal inputs into a high-dimensional state space, generating rich reservoir states that enable accurate solving and prediction of complex state equations. With advantages such as readout-layer-only training, low inference power consumption, and few-shot data driven, PRC offers a novel pathway for real-time solving of chaotic differential and partial differential equations on edge hardware. However, effectively co-designing high-order temporal processing algorithms with the intrinsic modulation mechanisms of physical devices remains a key bottleneck hindering practical deployment. Here, we report a highly efficient PRC system based on indium gallium zinc oxide (IGZO) electrolyte transistors, achieving the solving and prediction of time-dependent chaotic differential equations and partial differential equations. Leveraging the low-power characteristics and continuous nonlinear mapping capability of IGZO devices, we designed a lightweight algorithm co-optimizing reservoir training, inference, and equation solving. The electrical response and ion migration dynamics of the IGZO transistors were accurately modeled and characterized, followed by system-level validation on the field programmable gate array (FPGA). Resource evaluation demonstrates that, compared to a conventional fourth-order Runge-Kutta (RK4) implementation, our approach reduces digital signal processor (DSP) and flip-flop utilization by approximately 64% and 50%, respectively, while maintaining equivalent accuracy. In performance benchmarks using the Mackey-Glass equation, Lorenz chaotic sequence, and one-dimensional heat diffusion equation, our system achieves an inference latency of 11 μs and reduces power consumption by 15%, with a normalized root mean square error (NRMSE≈0.04) comparable to the Runge-Kutta method. For the partial differential equation (PDE) task, computation time is reduced to 1/400 of that required by the second-order Runge-Kutta (RK2) method, while maintaining a comparable NRMSE≈0.008. This work not only validates the efficiency and accuracy of IGZO-based PRC in solving general differential equations, but also provides a scalable architectural foundation and practical engineering reference for next-generation lightweight, low-power edge scientific computing and neuromorphic hardware platforms.关键词：indium gallium zinc oxide (IGZO) electrolyte transistor;artificial synaptic device;reservoir computing;chaotic differential equation solving;partial differential equation solving72|44|0更新时间：2026-10-16
- 摘要：With the rapid development of deep learning technology, convolutional neural networks (CNNs) have demonstrated exceptional performance in image recognition and processing tasks. However, as the network depth increases, the massive transmission of intermediate data imposes tremendous pressure on the on-chip memory and memory access bandwidth of hardware accelerators. The increasingly prominent “memory wall” problem has severely constrained the overall throughput and energy efficiency of the system. To address this issue, existing inter-layer feature data compression methods are mainly divided into two categories. The first category focuses on lightweight hardware implementation: despite low area overhead, their compression ratio is limited by algorithm complexity, making it difficult to effectively alleviate the off-chip bandwidth pressure in high-throughput scenarios. The second category pursues superior compression performance, but incurs excessive hardware area overhead, which is hard to deploy on resource-constrained edge devices. Aiming at the above challenges, this paper proposes a statistic-aware hybrid compression method for CNN inter-layer feature maps, with the core design goal of achieving high compression ratio and low hardware overhead to resolve the difficulty in balancing compression performance and resource consumption. By deeply exploiting the sparsity and distribution characteristics of the data, this method realizes hardware-friendly data coding combined with a hardware-software co-design mechanism of “offline analysis-online compression”. In the offline analysis stage, statistical analysis is performed on the CNN inter-layer feature data to generate the required coding tables and baseline values. In the online compression stage, the feature data are classified into zero-value data and non-zero-value data. For zero-value data, an enhanced zero run-length encoding combined with entropy coding is adopted; for non-zero data, dynamic baseline-delta encoding is applied. This differentiated coding mechanism reduces the hardware area overhead by 58.7% to 72.9% while maintaining a high compression ratio, which solves the problem of high hardware complexity in traditional compression algorithms.We conduct a systematic evaluation of the compression performance of the proposed method under different network structures and data formats, based on compression experiments on inter-layer feature maps of four representative CNNs: AlexNet, VGG16, ResNet34, and MobileNetV2. Experimental results show that, compared with similar studies, the proposed data compression method achieves a maximum improvement of 58.5% in compression ratio under the INT8 quantization format, and a maximum improvement of 36.7% under FP32/FP16 formats. When deploying the VGG16 model on the ALINX AXU5EV target platform, the accelerator based on the proposed data compression method reaches an inference throughput of 242.8 GOPS. Compared with the compression-free baseline architecture, the computing performance and energy efficiency are improved by 41.4% and 27.8%, respectively.The experimental results demonstrate that the proposed method balances the compression ratio and hardware overhead for CNN inter-layer feature map compression, and provides a new solution for the design of CNN accelerators in resource-constrained edge scenarios.关键词：field programmable gate array;convolutional neural network accelerator;hardware accelerator;interlayer feature map;data compression62|40|0更新时间：2026-10-16
- 摘要：Spiking neural networks (SNNs), regarded as the third generation of artificial neural networks (ANNs), encode and transmit information through discrete spike events. Owing to their event-driven nature, spatiotemporal information processing capability, and low-power characteristics, SNNs have shown great potential in edge intelligence, brain-inspired computing, and neuromorphic hardware. With the rapid development of intelligent terminals and internet-of-things devices, the demand for environmental adaptation and continual learning is increasing, making online learning an important research direction for edge intelligence systems. Existing online learning methods can generally be categorized into two groups. Backpropagation (BP)-based methods achieve high accuracy by utilizing global supervisory information, but require the storage of a large number of intermediate states and error signals, resulting in considerable computational complexity, memory overhead, and power consumption. In contrast, spike-timing-dependent plasticity (STDP)-based methods adopt local and event-driven synaptic updates, making them hardware-friendly; however, their learning capability and task performance are often limited due to the lack of effective supervisory information. To address these challenges, this paper proposes an online learning algorithm, termed forward-forward (FF)-STDP, which combines the FF learning paradigm with the STDP mechanism. Inspired by the layer-wise local optimization strategy of FF, the proposed algorithm introduces teacher signals into the spike-based learning process and constructs a local goodness metric based on neuronal firing rates, thereby enabling supervised learning without error backpropagation. By preserving the local weight-update property of STDP while incorporating supervisory information provided by FF, the proposed method achieves both task-oriented learning capability and hardware efficiency. Furthermore, to reduce the computational complexity and hardware implementation cost of conventional FF algorithms in SNNs, spike-count-based gradient approximation and local derivative approximation are adopted, such that synaptic updates depend only on the spike statistics of pre-and post-synaptic neurons, significantly reducing computational and memory overhead.Based on the proposed FF-STDP algorithm, an event-driven SNN accelerator architecture for edge intelligence applications is developed. The architecture adopts an address event representation (AER) communication scheme and integrates parallel post-synaptic neuron processing, dual-port memory, and pipelined execution strategies to enable efficient collaborative execution of inference and learning. To further reduce hardware complexity, a lookup-table (LUT)-based weight update unit is designed, achieving low-overhead and high-throughput online learning. To validate the effectiveness of the proposed algorithm and hardware architecture, the system is implemented on a Xilinx ZYNQ-7010 platform and evaluated using the modified national institute of standards and technology database (MNIST) and Fashion-MNIST datasets. Experimental results demonstrate that the proposed FF-STDP algorithm achieves classification accuracies of 90.43% and 80.53% on MNIST and Fashion-MNIST, respectively. Operating at 100 MHz, the accelerator core consumes only 15 mW of dynamic power and achieves an inference throughput of 1 887 frames per second. Compared with existing BP-based and conventional STDP-based online learning SNN hardware, the proposed approach significantly reduces hardware resource consumption and power overhead while maintaining competitive classification performance, demonstrating excellent energy efficiency and potential for edge deployment.关键词：spiking neural network (SNN);forward-forward (FF) learning;online learning;Field Programmable Gate Array (FPGA);hardware accelerator;hardware-software co-design140|149|0更新时间：2026-10-16
- 摘要：Airborne radio frequency (RF) systems are rapidly evolving toward intelligent skins, multifunctional integration, and intelligence, with exponential growth in system data volume, computational demands, and product iteration speed. Conventional FPGA+DSP custom computing architectures based on SRIO interconnection suffer from strong coupling between computing resources and tasks, inefficient aggregation of heterogeneous computing power, and limited system scalability, making them inadequate to support the requirements of next-generation airborne RF processing. To overcome these bottlenecks, this paper draws on the computational architecture concepts of mainstream intelligent computing centers and introduces them into the airborne RF computing domain, proposing a data processing unit (DPU)-based integrated RF processing system. This system employs a high-precision clock synchronization mechanism combining IEEE 1588v2 and synchronous ethernet (SyncE), addressing the high-precision spatiotemporal alignment of massive multi-source RF data in elastic distributed architectures. Based on remote direct memory access (RDMA), domain-specific architecture direct (DSA Direct) data passthrough is implemented, supporting end-to-end zero-copy transmission from data acquisition and processing to storage, significantly reducing system transmission latency. Through a heterogeneous computing power interconnection mechanism integrating PCIe and RDMA, aggregation and flexible configuration of heterogeneous computing power under a unified memory view are achieved. Leveraging nanosecond-level synchronous access, collaborative heterogeneous computing, and distributed storage capabilities, a data-centric integrated RF processing architecture is constructed. The architecture consists of RF access nodes and a heterogeneous computing resource pool interconnected through high-speed Ethernet switches, integrating intelligent, parallel, general-purpose, and reconfigurable heterogeneous computing nodes together with NVMe-oF distributed storage nodes into a unified computing power network, thereby supporting task-oriented elastic resource composition and dynamic deployment. Using a domestic DPU FPGA prototype, an experimental environment for the integrated RF processing system was established, and typical data flows including multi-channel RF data access, cross-node heterogeneous computing collaboration, distributed storage access, as well as RF data transmission, storage, and computation were validated. Test results indicate that compared to conventional architectures, the system reduces single-copy latency in data transmission to approximately 1/360 of the original; the peak sequential read/write bandwidth of NVMe over fabrics (NVMe-oF) storage nodes with three disks exceeds 5 700 MB/s, representing a 6 to 8 fold performance improvement over conventional methods; and the clock synchronization accuracy based on IEEE 1588v2+SyncE is approximately 8.33 ns. Furthermore, by constructing typical radar and communication signal processing chains, the system demonstrated an access latency as low as 24.92 ms for 128 MB radar raw data in the radar scenario, while stably supporting the real-time access of 38 parallel baseband data streams in the communication scenario, thereby validating its real-time service capability under high-throughput raw data injection and multi-level heterogeneous computing. Experiments demonstrate that this architecture provides a feasible technical pathway for next-generation airborne integrated RF processing systems.关键词：DPU;RF integration;computing power network;signal processing;edge computing97|67|0更新时间：2026-10-16
Innovative Architectures and Electronic Systems for Efficient Artificial Intelligence
- 摘要：With the rapid development of millimeter-wave communication systems, radio-frequency front-end circuits are increasingly required to support high-quality and high-capacity data transmission. As a key building block in the receiver front-end, the low-noise amplifier (LNA) must provide a low noise figure (NF) for high sensitivity while maintaining sufficient linearity to alleviate gain compression and signal distortion under strong interferers and complex modulation scenarios. However, achieving low noise, high gain, and high linearity simultaneously over a wide bandwidth remains challenging for millimeter-wave integrated circuit design. To address these challenges, this paper presents a K/Ka-band high-linearity wideband LNA implemented in a 55 nm complementary metal-oxide-semiconductor (CMOS) process. The proposed LNA adopts a two-stage cascode topology, in which source inductive degeneration is introduced in the common-source transistors to extend the input voltage swing and improve large-signal linearity. The bias-dependent third-order transconductance nonlinearity is quantitatively analyzed, and the main amplifying transistors are biased near the zero-crossing point of gm3, thereby suppressing third-order nonlinear distortion and improving the input 1 dB compression point (IP1 dB). In addition, a gate-capacitance peaking strategy is employed to compensate the high-frequency gain response, further improving the gain performance over a broad bandwidth. A three-dimensional interdigitated layout strategy is further developed to optimize the transistor finger structure and metal-layer arrangement. This approach reduces the parasitic gate resistance and source inductance, which are particularly detrimental at millimeter-wave frequencies, thereby mitigating their impact on noise, gain, and stability. Measurement results show that the proposed LNA achieves a peak gain of 16.8 dB, a 3 dB bandwidth of 8.8 GHz over 21.5~30.3 GHz, and a minimum NF of 3.26 dB. The measured IP1 dB reaches -3.6 dBm, while the core area is only 0.11 mm². These results demonstrate that the proposed design effectively improves the noise-linearity-gain trade-off and provides a promising solution for high-dynamic-range millimeter-wave receiver front ends.关键词：low-noise amplifier;millimeter-wave;high linearity;three-dimensional interdigitated layout structure;complementary metal-oxide-semiconductor119|213|0更新时间：2026-10-16
- 摘要：With the evolution of 5G/6G communication technologies, the application of high-order modulation and wideband signals (in the order of hundreds of MHz) forces radio frequency (RF) power amplifiers (PAs) to operate in near-saturation regions, thereby exhibiting significant strong memory effects and nonlinear characteristics. To achieve high linearity specifications while maintaining high energy efficiency, digital predistortion (DPD) has become a key technology at the baseband processing end. However, traditional memory polynomial behavioral models are constrained by their fixed structures. The number of their kernel coefficients grows exponentially with the increase of memory depth and nonlinearity order, leading to a severe “curse of dimensionality” and resulting in high implementation complexity for DPD systems. Furthermore, existing dynamic generation models still rely on traditional polynomial multiplication mechanisms, which are prone to causing rapid expansion of the dictionary space and massive computational overhead in wideband scenarios. To address these issues, this paper proposes an adaptive basis function generation technique for PA behavioral models based on linear B-spline basis. This technique abandons cumbersome high-order polynomial multiplication operations, utilizing linear B-spline functions—which possess strict local support and excellent numerical stability—as fundamental building blocks. It introduces a greedy iterative search strategy based on subspace coefficient re-estimation to adaptively construct amplitude gain functions that match specific memory depths by detecting the residual direction. Additionally, this technique extends the nonlinear amplitude mapping to the complex domain, significantly enhancing the capability to capture the complex amplitude and phase distortions of wideband signals, thereby achieving precise fitting of PA characteristics. At the implementation level, by fully leveraging the local asymptotic approximation advantages of B-spline functions, the model constructed by this generation technique can be simplified into a hardware-friendly structure combining look-up tables (LUTs) with linear interpolation. This structure successfully transforms complex nonlinear algebraic computations into underlying memory location indexing and simple addition-multiplication combinations, significantly reducing computational complexity and multiplier logic resource consumption while maintaining accuracy. To verify the performance of the proposed technique, a test platform is constructed based on a 100 MHz wideband signal and PAs. Experimental results demonstrate that, while maintaining linearization performance comparable to traditional models, the proposed method reduces the number of model coefficients by nearly 90% and decreases implementation complexity by 74%, significantly improving the efficiency of both behavioral modeling and DPD. Finally, extended tests on PAs with varying nonlinear characteristics indicate that the adaptive algorithm does not rely on the physical a priori features of specific PAs. This further confirms that the proposed model possesses good generalization capabilities and is easy to apply in practical engineering.关键词：power amplifier;behavioral model;digital predistortion;B-spline basis function;look-up table;linear interpolation63|48|0更新时间：2026-10-16
- 摘要：In complex electromagnetic environments, target signals and interference signals heavily overlap in both time and frequency domains, posing significant challenges for co-frequency blind signal separation under single-channel reception conditions. Under complex non-stationary interference conditions, existing convolution-based separation methods primarily suffer from insufficient non-local correlation modeling, limited decoupling capability between target and interference signals in the latent space, and weak generalization to unknown interference patterns. To address these issues, this paper proposes a single-channel co-frequency blind signal separation method based on dynamic graph neural networks and contrastive learning (DGCL). First, the complex baseband mixture sequence is mapped into an in-phase/quadrature (I/Q) dual-channel representation, and dynamic graph neural networks are utilized to effectively extract multi-scale graph temporal features. Through the synergistic effects of multi-scale segmentation, dynamic graph construction, and graph message passing, the representation capability for complex non-stationary interference is enhanced. Second, a contrastive learning autoencoder driven by target signal matching is designed. By introducing contrastive constraints on top of structural matching reconstruction, the separability of the target signal and the interference signal in the latent space is improved. Furthermore, a latent space manifold mixup enhancement strategy is designed during training. By expanding the deep feature distribution, the model’s robustness against unknown interference patterns is elevated. Experiments conducted under single-carrier, multi-carrier, and multi-class interference conditions indicate that DGCL demonstrates significant advantages in both bit error rate and mean square error metrics.Compared with the conventional WaveNet baseline, DGCL reduces MSE (Mean Square Error) by 6.7 dB and decreases BER (Bit Error Rate) by 1.72 × 10-2. These results verify the effectiveness of the proposed method.关键词：single-channel blind signal separation;co-frequency interference;graph neural network;contrastive learning;manifold mixup71|58|0更新时间：2026-10-16
- 摘要：With the rapid development of wireless communication technology, the number of wireless communication emitters has grown exponentially, leading to an increasingly complex electromagnetic environment and increasingly scarce spectrum resources. Against this backdrop, achieving precise identity recognition and secure access among massive devices has become a critical issue concerning social stability and economic development. Traditional identity authentication methods rely on digital credentials at the protocol or application layer, which are susceptible to theft, forgery, or attacks. In contrast, the specific emitter identification (SEI) technology based on radio frequency fingerprinting (RFF) serves as a physical-layer security measure with broad application prospects. This technology authenticates identities by extracting the inherent physical imperfections in the emitter’s hardware circuits, offering natural uniqueness and long-term stability. However, existing deep learning-based methods for communication emitter identification generally face challenges, including heavy reliance on labeled data, limited generalization, and low efficiency in processing long-sequence signals. To address these issues, this paper focuses on the efficient identification of communication emitters under few-shot conditions and proposes a low-complexity self-supervised recognition algorithm. First, to overcome the problem that fingerprint features in the original I/Q signals are easily obscured by channel noise and modulation information, this paper introduces a lossless variational mode decomposition (VMD) method. This approach adaptively converts the original I/Q signals into multiple intrinsic mode function components, enhancing the representational capacity of fingerprint features. Second, to address the quadratic computational complexity bottleneck of traditional sequence modeling methods when processing long-sequence signals, this paper incorporates the Mamba architecture, which has linear computational complexity. A self-supervised network model based on the Mamba linear flow encoder is constructed, utilizing a large-scale auxiliary dataset for pre-training. By integrating contrastive and supervised learning, the model is guided to learn discriminative, generalizable feature representations. Finally, to tackle the scarcity of new device samples and the presence of unknown interference in practical applications, this paper employs a transfer learning strategy. Using a small number of target-domain emitter samples, a lightweight linear classifier is fine-tuned to adapt to specific downstream classification tasks quickly. Simulation results on the WiSig dataset show that the proposed method achieves 91.20% accuracy in a 6-way 1-shot scenario. While significantly reducing computational complexity and dependence on sample size, the method maintains high recognition performance, demonstrating promising potential for real-world applications.关键词：specific emitter identification;few-shot learning;self-supervised learning;variational mode decomposition;Mamba;data augmentation85|181|0更新时间：2026-10-16
- 摘要：Ambient internet of things (Ambient IoT) indoor localization technology holds significant application value in scenarios such as smart warehousing and new retail. However, complex indoor multipath environments give rise to two prominent issues: first, severe imbalance between strong and weak tag samples caused by multipath blockage and tag performance variations. In a typical clothing display scenario, the effective sample size for weak tags is only about 30% of that for strong tags. Second, multipath aliasing significantly degrades the distinguishability of location-dependent signal features, limiting improvements in localization accuracy. To address these issues, this paper proposes a physically constrained data augmentation and multipath-aware transformer (PCDA-MPAT) localization method that integrates physically constrained data augmentation with a multipath-aware Transformer. At the data augmentation level, we propose a physically constrained data augmentation (PCDA) method. Using phase-frequency slope consistency as a physical constraint, we construct a conditional Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to generate weak tag samples that not only match the distribution of measured data but also adhere to the physical laws of wireless propagation, effectively alleviating the imbalance between strong and weak tag samples from the data perspective. At the model design level, we construct a multipath-aware transformer (MPAT) localization model. We design a time-domain local window attention mechanism adapted to the slow-fading characteristics of quasi-static multipath channels, combined with time-domain relative position encoding, to capture the temporal evolution patterns of multipath features, suppress long-term irrelevant noise interference, and enhance the model’s ability to extract and learn multipath features. Experimental results in a 10 m×10 m clothing display scenario demonstrate that the proposed method achieves a tag-level localization accuracy of 83.57%, outperforming the best existing baseline method by 5.71 percentage points. Ablation studies further validate the effectiveness of PCDA and MPAT. PCDA alone improves the tag-level localization accuracy by 4.29 percentage points, and MPAT further increases it by 2.14 percentage points. These results confirm the effectiveness of the proposed method in complex indoor multipath environments, meeting the localization requirements of scenarios such as smart warehousing and new retail.关键词：Ambient IoT;indoor localization;multipath-effect;data augmentation;generative adversarial network;Transformer68|80|0更新时间：2026-10-16
- 摘要：Gesture recognition is a fundamental enabling technology for humancomputer interaction and has been widely adopted in applications such as smart healthcare, virtual reality, and smart homes. Compared with vision-based approaches, radio frequency identification (RFID)-based gesture recognition offers several distinctive advantages, including contact-free operation, low deployment cost, and privacy preservation, making it a promising solution for practical deployments. However, the performance of existing RFID-based gesture recognition models heavily depends on the scale and diversity of training data. In practice, the acquisition of RF data is costly and the resulting data distributions are inherently limited, which significantly constrains the generalization capability of models to unseen samples. To address this challenge, data augmentation has been widely regarded as an effective means to enhance recognition performance. Nevertheless, most existing augmentation methods overlook the intrinsic variability of human behavior, leading to synthesized samples with insufficient distributional diversity. In fact, even within the same gesture category, substantial variations commonly exist across different instances, primarily because humans cannot reproduce the same action in a perfectly identical manner. In gesture recognition scenarios, such behavioral variability intuitively arises from changes in the hand’s distance relative to the sensor as well as variations in motion speed. This intuitive hypothesis is systematically validated in this work through the establishment of a signal propagation model combined with preliminary experimental analysis. Motivated by these observations, we propose a behavior diversity-aware RFID data augmentation framework. Specifically, grounded in the physical principles of RF signal propagation, the proposed method explicitly encodes gesture-related distance and speed as conditional inputs. A novel conditional diffusion model is then designed to perform controlled shifts along the distance- and speed-conditioned directions during the generation process, thereby synthesizing high-quality samples that are both physically consistent and distributed over a broader behavioral diversity space. By introducing synthetic data enriched with behavioral diversity into the training set of gesture recognition models, the generalization capability to unseen samples and the recognition accuracy can be significantly improved. We implement an RFID-based gesture recognition prototype system in real-world scenarios and conduct systematic physical experiments. Experimental results show that, in a ten-class gesture recognition task, the proposed method improves recognition accuracy by more than 20% compared with baseline models and significantly outperforms existing data augmentation strategies. In addition, the effectiveness of the proposed data augmentation approach does not degrade noticeably as the scale of the sensing task increases. Experimental results on identity recognition further demonstrate that the proposed method can be effectively applied to other human-computer interaction applications.关键词：gesture recognition;radio frequency identification;data augmentation;diversity;diffusion model;wireless sensing89|115|0更新时间：2026-10-16
RF Computing
-
- 摘要：With the rapid proliferation of the low-altitude economy, typical low-altitude applications are imposing escalating demands on communication and sensing capabilities, thereby driving the evolution of integrated sensing and communication (ISAC) networks. Although ISAC has been recognized as one of the six flagship scenarios for 6G, existing research predominantly focuses on terrestrial users and exhibits significant limitations in achieving seamless low-altitude coverage, three-dimensional mobility management, and high-precision sensing. These shortcomings hinder the support for large-scale drone swarm collaboration and precise control, necessitating the development of systematic ISAC solutions tailored to the unique characteristics of the low-altitude economy. Furthermore, existing surveys often confine themselves to discrete technical aspects—such as air interface design, networking protocols, or security mechanisms—and lack a holistic logical framework bridging application requirements, system architecture, and key technologies. To address this gap, this paper provides a comprehensive review of the state-of-the-art in ISAC system architectures and key technologies for the low-altitude economy. First, we analyze the differentiated performance requirements—specifically regarding communication rate, latency, sensing accuracy, and range—across six canonical scenarios, including urban management, logistics delivery, and emergency response. Building on this analysis, four core research challenges are identified: system architecture design, detection and tracking of small, slow, and low targets, continuous communication coverage under high-dynamic conditions, and application-oriented computing resource allocation coupled with security management. Addressing these challenges, this paper presents a structured review: we construct an overarching framework for low-altitude ISAC based on UAV-assisted and satellite-assisted networks alongside O-RAN architectures, integrating low-altitude channel measurements and modeling to reveal propagation characteristics and provide a foundational model for system design; we systematically examine critical low-altitude sensing technologies encompassing target detection, recognition, localization, trajectory monitoring, and imaging to enable real-time situational awareness of airborne targets and environmental states; we explore sensing-driven mobility management and coverage enhancement mechanisms, addressing computing resource allocation through trajectory planning, beam management, multi-node handover, task scheduling, and joint communication-sensing-computing optimization, thereby empowering communication decision-making and closing the control loop; targeting security threats such as covert access, malicious jamming, and eavesdropping, we investigate sensing-driven security management protocols and leverage novel aerial platforms equipped with reconfigurable intelligent surfaces and movable antenna arrays to enhance system robustness and future scalability. Finally, we outline promising future directions, including the construction of environmental-channel knowledge bases using generative artificial intelligence, the deep closed-loop collaborative mechanisms for large-scale drone swarms, and robustness enhancement against feature extraction challenges and jamming in complex low-altitude electromagnetic environments. This review aims to provide theoretical support and technical insights for the sustainable development of the low-altitude economy.关键词：low-altitude economy;integrated sensing and communications;channel modeling;low-altitude target sensing;three-dimensional communication coverage203|241|0更新时间：2026-10-16
- 摘要：To support the large-scale and regular operation of the low-altitude economy (LAE), and address the critical technical bottlenecks of low communication reliability, limited target sensing accuracy, and mutual resource constraints between communication and sensing in low-altitude wireless networks (LAWNs) under complicated airspace environments, this paper proposes a novel LAWN system framework based on networked integrated sensing and communication (ISAC). The proposed framework employs multiple unmanned aerial vehicles (UAVs) as flexible aerial nodes to cooperatively provide communication services for multiple ground information users and perform sensing tasks for multiple aerial targets, where a dedicated radar receiver is adopted to acquire and process echo signals from sensing targets, constructing an integrated LAWN that deeply merges communication services and environmental sensing. To maximize the sum rate of users, we jointly optimize the transmit beamformers, receive beamformers, and UAV deployment positions under the constraints of radar signal-to-interference-plus-noise ratio (SINR), communication SINR, and transmit power budget of each UAV. To handle the non-convex problem, a two-stage alternating optimization (AO)-based algorithm is proposed, which decouples the original problem into the beamforming design subproblem and the UAV deployment subproblem. For the beamforming design subproblem, semidefinite relaxation is applied to simplify the subproblem, and the optimal closed-form expressions for the receive beamformers are derived via the Cauchy-Schwarz inequality, followed by a successive convex approximation (SCA)-based algorithm to handle the reformulated subproblem. For the UAV deployment subproblem, we introduce auxiliary variables to equivalently transform the non-convex objective function into a convex form, followed by a trust-region-based algorithm to obtain the UAV deployment positions. We theoretically analyze the convergence, optimality, and computational complexity of the proposed SCA-based and trust-region-based algorithms. Simulation results validate the effectiveness and superiority of the proposed two stage AO-based algorithm in both single-UAV and multi-UAV scenarios. In the single-UAV case, the proposed algorithm closely matches the global optimal solution obtained by exhaustive search, demonstrating the optimality of the proposed algorithm. In the multi-UAV case, the proposed algorithm achieves significant sum rate performance gains compared with the UAV central deployment design and minimum mean square error precoding design.关键词：low-altitude wireless network (LAWN);low-altitude economy (LAE);networked integrated sensing and communication (ISAC);unmanned aerial vehicle (UAV);alternating optimization91|74|0更新时间：2026-10-16
- 摘要：In view of the fundamental technical requirements for the integration of communication and sensing in 6G networks under the background of low-altitude economy, the existing fixed multi-base-station integrated sensing and communication (ISAC) architectures suffer from limited flexibility. Most current unmanned aerial vehicle (UAV)-enabled ISAC studies focus on independent operation without cooperation with ground-based stations. In addition, trajectory planning, user scheduling, and power allocation exhibit strong coupling in air-ground coordinated scenarios, resulting in non-convex and intractable optimization models. To address these issues, this paper proposes an air-ground coordinated ISAC transmission architecture consisting of UAVs and ground base stations, along with a joint multi-dimensional resource allocation scheme. Leveraging the flexible mobility of UAVs, the constructed architecture establishes an air-ground coordinated communication and sensing mechanism to simultaneously provide communication services for edge users and accomplish sensing tasks for moving targets. With the objective of minimizing the total system transmit power, a multi-dimensional joint optimization model involving UAV trajectory, user scheduling, and power allocation is formulated under the constraints of user signal-to-interference-plus-noise ratio (SINR) and target localization Cramér-Rao lower bound (CRLB). To handle the deeply coupled discrete scheduling, continuous power and dynamic trajectory variables that are difficult to solve jointly, a deep reinforcement learning-based two-layer optimization framework is proposed. In the outer layer, a deep deterministic policy gradient (DDPG) algorithm is adopted to iteratively optimize UAV flight trajectories in continuous action space, adapting to dynamic topological variations caused by high mobility. In the inner layer, semidefinite relaxation (SDR) transforms the non-convex resource allocation problem into a tractable semidefinite programming form, so as to efficiently obtain near-optimal solutions of joint user scheduling and power allocation under fixed UAV positions. The minimal transmit power obtained from inner-layer optimization is constructed as a reward function and fed back to the outer layer, realizing bidirectional interactive updating between trajectory strategies and resource allocation. This hierarchical coupling design avoids the drawbacks of pure reinforcement learning methods, such as redundant action space, difficult constraint embedding and unstable training. Moreover, it alleviates the sensitivity to initial values and local optimum defects of traditional mathematical optimization methods, maintaining feasible optimization domains in dynamic environments. Simulation results demonstrate that the proposed air-ground coordinated ISAC scheme achieves 30%~36% power consumption reduction compared with non-coordinated methods. Furthermore, the proposed two-layer optimization scheme reduces power consumption by 39% and 17% compared with the pure deep reinforcement learning scheme and fixed-trajectory scheme, respectively. This research provides a theoretical basis and technical reference for trajectory design and resource configuration of air-ground coordinated ISAC systems in low-altitude economy scenarios.关键词：UAV-ground BS coordination;integrated sensing and communication;deep reinforcement learning;multi-dimensional resource joint allocation;trajectory design;low-altitude economy66|44|0更新时间：2026-10-16
- 摘要：With the rapid development of the low-altitude economy, unmanned aerial vehicles (UAVs), as core nodes in low-altitude networks, play a critical role in enabling ubiquitous connectivity and environmental reconstruction through integrated sensing and communication (ISAC) capabilities. However, in complex urban environments characterized by dense building layouts, low-altitude UAVs face severe challenges such as frequent line-of-sight (LoS) link blockages and drastic fluctuations in channel quality. Traditional trajectory planning methods based on ideal free-space models struggle to meet practical application requirements, necessitating the design of intelligent planning approaches capable of proactively avoiding blockages and dynamically adapting to environmental variations. To address this issue, this paper proposes a multi-stage trajectory and resource joint optimization method based on a probabilistic LoS channel model for complex environments. First, an elevation-dependent probabilistic LoS model is constructed to quantify the canyon effect in urban areas, steering trajectory planning toward the optimization of wireless link reliability. Second, a multi-stage planning framework is designed to dynamically adjust flight strategies through stage-wise updates of target state estimation, thereby addressing environmental uncertainty and state estimation errors. On this basis, considering kinematic and energy constraints, a joint optimization problem is formulated to maximize the expected weighted communication rate and the effective Fisher information matrix for sensing, achieving an adaptive trade-off between communication and sensing performance. To tackle the strong non-convexity induced by the probabilistic exponential terms, an iterative algorithm based on block coordinate descent and successive convex approximation is developed. Simulation results demonstrate that the proposed method effectively guides UAVs to actively avoid blind zones in high-density urban scenarios while maintaining stable links and high sensing accuracy. Compared with conventional approaches, the proposed method achieves an approximately 80.8% improvement in total communication capacity and an 81.4% reduction in target localization error, significantly enhancing the robustness and effectiveness of low-altitude ISAC networks.关键词：unmanned aerial vehicle;integrated sensing and communication;trajectory design;resource allocation;Cramer-Rao Bound59|39|0更新时间：2026-10-16
- 摘要：Low-altitude unmanned aerial vehicle (UAV) swarms in urban airspace operate over rapidly time-varying multipath channels, where severe Doppler shifts and doubly selective fading can degrade data links, sensing reliability, and closed-loop cooperative control. To address these issues, this paper proposes an integrated sensing-communication-control framework based on MIMO-OTFS-non-orthogonal multiple access. Specifically, orthogonal time frequency space (OTFS) modulation is utilized to provide a quasi-static representation in the delay-Doppler (DD) domain, enhancing robust transmission and parameter estimation (e.g., range and velocity) under doubly selective channels. Simultaneously, multiple-input multiple-output (MIMO) and power-domain non-orthogonal multiple access (NOMA) are combined to achieve massive access and spectral efficiency. Furthermore, we construct a DD-domain micro-Doppler structured-mapping mechanism, where rotor micro-Doppler signatures serve as physical-layer fingerprints for identity association, providing a reliable index for robust tracking and cooperative control under dense or crossing trajectory conditions. For the uplink backhaul, a superposition transmission architecture is designed with full-frequency multiplexing for the master UAV and frequency-division access for slave UAVs. By decomposing complex cluster interference into parallel two-user NOMA models and applying a joint detection scheme based on the message passing algorithm and successive interference cancellation, the proposed method effectively mitigates inter-user interference, ensuring air-ground link reliability. Finally, a sensing-aided cooperative trajectory control loop is established using the sensed range/velocity and micro-Doppler feature identity association results to realize the integrated sensing-communication-control process. Simulation results demonstrate that the proposed scheme avoids the bit-error-rate floor observed in baselines based on orthogonal frequency division multiplexing and non-orthogonal multiple access, and achieves a signal-to-noise ratio gain of approximately 7 dB. Moreover, the sensing module realizes stable estimation of range and velocity, with accuracy trends consistent with the Cramér-Rao bound analysis. In the closed-loop cooperative validation, the sensing-feedback-based UAV swarm achieves three-dimensional formation trajectory tracking. During continuous maneuvering cruise, the cooperative position root mean square errors of the four slave UAVs are controlled at 0.13, 0.15, 0.19, and 0.23 m, respectively, and the relative errors do not diverge. This preliminarily verifies the closed-loop stability of the proposed scheme in the integrated sensing-communication-control process at the simulation level.关键词：integrated sensing-communication-control system;orthogonal time frequency spatial modulation;Micro-Doppler feature recognition;non-orthogonal multiple access;control compensation mechanism92|53|0更新时间：2026-10-16
- 摘要：The low-altitude economy, as an emerging economic paradigm, holds significant importance for socioeconomic development, urban management, and three-dimensional transportation. Low-altitude wireless networks (LAWNs) serve as a fundamental enabler of this economy by providing essential communication and sensing capabilities for various low-altitude applications. To further enhance the LAWNs, this paper proposes the integrated secure communication and sensing performances optimization method, addressing the wireless signal eavesdropping issue for low-altitude applications. Specifically, a secure communication and sensing network model for LAWNs is constructed first, and the indicators, including secrecy rate, average age of information (AoI) for transmission data, and beamforming error, are defined to characterize communication and sensing performance. Then, a multi-objective weighted optimization problem is formulated to simultaneously enhance the secure communication and sensing performance of unmanned aerial vehicles (UAVs) in low-altitude applications. To solve this optimization problem, an improved multi-operator differential evolution algorithm is proposed, which optimizes the base station’s data generation rate and power allocation to enhance the overall secure communication and sensing performance. Theoretical analyses of the algorithm’s complexity and convergence are also provided. Simulation results demonstrate that the proposed method can effectively improve the secrecy rate and reduce the average AoI while minimizing beamforming errors, outperforming baseline schemes.关键词：low-altitude wireless networks;secure communication;integrated sensing and communication;age of information;performance optimization124|86|0更新时间：2026-10-16
- 摘要：With the rapid development of the low-altitude economy, the application of unmanned aerial vehicles in logistics, security and other fields is becoming increasingly widespread, and the safety supervision of low-altitude airspace has become an urgent need. Integrated sensing and communication (ISAC) technology has become an important technical means for ensuring low-altitude security due to its ability to cooperatively achieve high-reliability communication and high-precision target sensing. However, the extreme openness of low-altitude wireless channels makes ISAC systems highly vulnerable to physical layer eavesdropping threats when performing sensing and communication tasks. Existing beamforming designs based on high-precision sensing often ignore secure transmission requirements, making it difficult to achieve the coordination of sensing and security. To address this issue, this paper investigates a secure beamforming scheme in low-altitude scenarios that considers both sensing accuracy and anti-eavesdropping performance. To improve sensing performance from the theoretical lower bound, this paper first derives the asymptotic Cramér-Rao bound (CRB) regarding the angle estimation of low-altitude sensing targets, and further establishes a non-convex optimization model with the minimization of the CRB as the objective function. In order to simultaneously balance communication quality of service and high stealth requirements in complex low-altitude environments, this paper further imposes multiple strict system constraints on the model. Specifically, the model introduces a communication waveform similarity constraint to flexibly and efficiently suppress signal interference among multiple users. Meanwhile, to ensure the reliable transmission of communication links, the model introduces a user received energy constraint to strictly guarantee the demodulation threshold required for the normal operation of the receiver. Additionally, in response to the core requirements of secure transmission, the model introduces an eavesdropping direction energy leakage constraint, aiming to eliminate potential eavesdropping threats from the physical layer through spatial nulling of beam energy. Finally, the model must strictly satisfy the total system power constraint to comply with the energy consumption limits of actual hardware. To efficiently solve the aforementioned model, this paper utilizes the augmented Lagrangian function and variable splitting technique to propose an iterative solving algorithm based on the alternating direction penalty method (ADPM), which decomposes the original optimization problem with complex constraints into multiple sub-problems that possess closed-form solutions or high-precision numerical solutions. Simulation results show that, compared with the traditional alternating direction method of multipliers (ADMM), the proposed algorithm not only has a faster convergence speed but is also able to stably converge to a lower objective function value, exhibiting strong numerical stability and robustness, and successfully achieving the collaborative optimization of high-precision low-altitude sensing and physical layer secure transmission.关键词：beamforming;Cramér-Rao bound (CRB);alternating direction penalty method (ADPM);integrated sensing and communication (ISAC);physical layer security;low-altitude economy188|82|0更新时间：2026-10-16
Integrated Sensing and Communications Towards Low\-Altitude Economy
- 摘要：Owing to its exceptional properties such as an ultrawide bandgap, ultrahigh thermal conductivity, high breakdown field, as well as excellent carrier transport and radiation resistance potential, diamond holds significant promise for radiation-hardened high-voltage power devices, and is particularly suitable for power regulation units in space power systems, deep space exploration platforms and high-energy particle radiation environments. However, with respect to single-event effects, which represent a critical failure mode governing the on-orbit reliability of devices, experimental reports on diamond field-effect transistors (FETs) remain absent in the international literature, and the relevant failure mechanisms, tolerance thresholds and transient response characteristics still need to be systematically clarified. In this work, p-channel hydrogen-terminated diamond high-voltage metal-oxide-semiconductor FETs (MOSFETs) are fabricated, and a post-metallization annealing process is employed to improve the interface quality and electrical performance. The devices achieve an off-state breakdown voltage of -980 V with a leakage level superior to that of comparable devices reported abroad, indicating that the fabricated devices have the basic conditions for carrying out high-voltage bias irradiation verification. Based on this foundation, heavy-ion single-event irradiation experiments using Ta ions (LET = 77.6 MeV·cm²/mg) and Kr ions (LET = 37.9 MeV·cm²/mg) are carried out, revealing for the first time the single-event burnout and single-event transient characteristics of diamond FETs experimentally. The experimental results demonstrate that under a bias of 200 V, no single-event burnout occurred for either ion species, and the burnout threshold exceeds 77.6 MeV·cm²/mg, preliminarily verifying the single-event burnout resistance potential of diamond FETs in high-LET heavy-ion irradiation environments. Only one transient pulse with an amplitude of approximately 40 V is captured during Ta-ion irradiation, whereas no discernible transient events were observed under Kr-ion irradiation, indicating a strong tolerance to single-event effects. This work provides a key experimental basis for the design of radiation-hardened diamond power devices and the investigation of their space radiation effects, and lays a foundation for subsequent reliability evaluation of devices under higher bias and higher LET conditions.关键词：diamond;field effect transistor;high-voltage power device;radiation resistance characteristics;Single-Event Burnout;Single-Event Transient116|179|0更新时间：2026-10-16
- 摘要：With the widespread deployment of internet of things (IoT) devices in smart homes, smart buildings, and other application scenarios, intelligent systems are increasingly expected to provide reliable automation control in environments where device states keep changing, user requirements are often ambiguous, and external disturbances may occur during execution. Existing trigger-action rule methods mainly rely on predefined static logic. Therefore, they have limited ability to represent the preconditions of device operations, state dependencies among devices and environmental attributes, and runtime environmental changes. Large language model (LLM)-based IoT agents improve natural language understanding and task planning, but they usually lack a unified structured environmental representation. As a result, such agents may miss important environmental dependencies, select inappropriate devices, or generate operation sequences with unreasonable execution orders. These limitations become more serious in scenarios involving multi-device collaboration, continuous monitoring, and abnormal execution feedback, where they may directly affect the stability of task execution. To address these problems, this paper proposes environment-aware action framework (EnvAct), an IoT automation execution framework based on LLMs and structured environmental representation. EnvAct uniformly models spaces, devices, state attributes, operational constraints, and state dependencies, and constructs an environmental representation mechanism composed of a constraint table and a dependency table. The constraint table describes the preconditions that must be satisfied before a device operation is executed, the expected effects of the operation, and the latency for the corresponding state effects to take place. The dependency table establishes an inverse mapping from environmental states to candidate operations, so that the system can automatically supplement necessary operations when some preconditions are not satisfied. To reduce the modeling cost in the cold-start stage, this paper extracts event-condition-action rules from Home Assistant community blueprint scripts. Based on the current devices, observable states, and physical commonsense, LLMs are used to filter personalized or environment-incompatible rules and generate an environmental representation that can be continuously updated. The EnvAct system consists of a planner, a compliance checker, an executor, a reflector, and an environmental representation layer. The planner integrates user requirements, real-time environmental information, device interfaces, and historical preferences to generate an initial operation sequence. The compliance checker performs legality verification, dependency completion, and partial-order graph construction according to the constraint table and the dependency table, thereby transforming the initial sequence into an executable plan that satisfies environmental constraints. The executor schedules independent operations in parallel according to partial-order relations, and realizes closed-loop execution through retries, execution feedback, and state verification. The reflector revises environmental representation and preference information based on execution traces and user feedback, enabling the system to improve its behavior in subsequent tasks. Experimental results show that, across six types of IoT tasks, EnvAct achieves overall success rates of 88.3% and 90.0% when driven by qwen-plus and GPT-4o, respectively, outperforming baseline methods such as Sasha and SAGE. In dynamic disturbance experiments, the average execution reliability of the system reaches 81.25%. In real sandbox experiments, the system maintains a success rate above 90%, with an average operation response time below 30 s. These results demonstrate that EnvAct can improve the reliability, adaptability, and engineering usability of automation task execution in dynamic IoT scenarios.关键词：Internet of Things;large language model;automated execution;environmental representation;closed-loop control;intelligent system57|133|0更新时间：2026-10-16
- 摘要：Integrated localization and communication (ILAC) networks deeply integrate communication and positioning functions, enabling efficient communication transmission services while providing reliable positioning services. Owing to their ability to support both information exchange and positioning perception within a unified wireless network, ILAC networks have emerged as a key direction for the evolution of future wireless systems. However, in emergency rescue scenarios, traditional ILAC networks that rely on terrestrial base stations may become unavailable or severely degraded due to damage to ground infrastructure. Such infrastructure failures can result in communication coverage interruption, link quality deterioration, insufficient localization anchors, and reduced positioning accuracy, making it difficult to continuously support mission-critical services in disaster-affected areas.To address this issue, this paper proposes an unmanned aerial vehicle(UAV)-swarm-enabled ILAC network, in which multiple UAVs are deployed as flexible aerial platforms to provide temporary, distributed, and reconfigurable communication and localization services for ground users in disaster-affected regions. Furthermore, a distributed framework is proposed for the joint optimization method of UAV association, spectrum allocation, and power control to ensure efficient communication and positioning services, thereby supporting the successful execution of critical tasks such as emergency rescue. First, a user positioning model based on double-sided two-way ranging is established to describe the ranging relationship between aerial UAV anchors and ground users, as well as the impact of ranging measurements and anchor selection on positioning accuracy. On this basis, the coupling relationship among communication access, positioning anchor association, spectrum sharing, and transmit power allocation is further characterized. Then, a joint optimization problem is formulated with the objective of maximizing the minimum user throughput, while considering positioning accuracy requirements, UAV association constraints, sub-channel allocation constraints, and power constraints. Since the formulated problem involves both discrete association decisions and continuous resource allocation variables, and the communication and positioning performances are strongly coupled through shared UAV anchor nodes and wireless resources, the problem is highly non-convex and difficult to solve directly by conventional centralized optimization methods. To tackle this complexity, the original problem is decomposed into two subproblems: UAV association, power control, and sub-channel allocation. For the first subproblem, a three-dimensional fast aerial anchor selection method is proposed. By introducing spatial partitioning, the proposed method reduces the search space of candidate aerial anchors and enables rapid association between user communication access points and positioning anchors, thereby lowering the computational complexity of UAV association while preserving positioning performance. For the second subproblem, a multi-agent proximal policy optimization-based joint power control and sub-channel allocation algorithm is proposed. By leveraging distributed multi-agent learning, users can learn coordinated resource allocation policies under local network observations, so as to jointly determine spectrum and power configurations in dynamic multi-user scenarios.Simulation comparisons with multiple baseline algorithms demonstrate that the proposed method can effectively improve the minimum user throughput and enhance communication fairness among users while satisfying positioning accuracy constraints. Moreover, under the same network scenario settings, the proposed method achieves superior positioning performance, indicating its effectiveness in jointly balancing communication efficiency and positioning reliability. These results show that the proposed UAV swarm-enabled ILAC framework provides a feasible and effective solution for supporting efficient communication and reliable positioning services in emergency rescue scenarios.关键词：ILAC network;UAV swarm;UAV association;resource optimization;distributed reinforcement learning;emergency communications63|145|0更新时间：2026-10-16
- 摘要：In response to the new demands of 4G/5G communication for intelligent connected vehicles, this paper proposes a design for a wideband high-efficiency transparent glass antenna with surface wave self-cancelling function. The large size and high dielectric constant of the vehicular glass make it easy for the antenna to excite strong surface waves. A large number of electromagnetic waves cannot be effectively radiated into free space, resulting in a decrease in antenna radiation efficiency and distortion of the radiation pattern. However, few research works have conducted in-depth studies on the surface wave issue. To address this challenge, this paper conducts an analysis of the radiation characteristics of current with an odd symmetric distribution. It is discovered that the surface waves excited by this current will cancel each other out in the opposite direction of the current. Therefore, it can be concluded that the surface wave fields of the current that is continuous and odd-symmetric in the
- 摘要：With the rapid development of the Internet of Things, the power supply problem for vast numbers of low-power micro/nano devices, such as wireless sensor nodes, has become increasingly prominent. Conventional battery-based power supply faces challenges including frequent replacement, high maintenance costs, and serious environmental pollution, making self-powered technologies a current research hotspot. Piezoelectric vibration energy harvesters exhibit great potential in solving the self-powering issue of low-power devices owing to their high energy density, simple structure, and ease of integration. Inspired by the strong stability regulation capability demonstrated by the hummingbird hawk-moth during hovering flight, this paper proposes a bio-inspired passive omnidirectional adaptive piezoelectric vibration energy harvester. The harvester consists of a rotating component formed by a deep groove ball bearing and a piezoelectric cantilever beam installed at the center of the rotating component. The up-and-down bending vibration of the cantilever beam mimics the flapping of the hummingbird hawk-moth’s wings, while the adaptive rotational motion of the cantilever beam mimics the moth’s adaptive hovering rotation around its center of mass, thereby achieving omnidirectional adaptive adjustment and stability regulation and significantly enhancing adaptability to environmental vibrations from different directions. Based on the energy method, a distributed-parameter electromechanically coupled nonlinear dynamic model is established, deeply revealing the coupling behavior between the bending vibration mode and the rotational motion. Through a combination of numerical simulation and experimental testing, the adaptive orientation adjustment characteristics, stability mechanism, and dynamic response of the harvester under different initial rotational angles ranging from 0° to 360° are systematically investigated. The results demonstrate that the proposed bio-inspired omnidirectional adaptive piezoelectric vibration energy harvester can, under external excitation in any direction, passively and adaptively rotate such that the piezoelectric cantilever beam ultimately stabilizes at a position aligned with the excitation direction while maintaining large-amplitude bending vibration output. Compared with traditional fixed or array-type multidirectional piezoelectric energy harvesters, the proposed harvester requires neither multiple cantilever beam combinations nor complex irregular structures; it achieves efficient omnidirectional energy harvesting from 0° to 360° using only a single piezoelectric cantilever beam, featuring a compact structure, high space utilization, and strong reliability. Experimental results show that under different initial included angles, the stabilized output voltage of the harvester reaches approximately 29 V, the output power stabilizes at around 830 µW, and the average power density is as high as 29 340 µW/cm³. When the external excitation direction is perpendicular to the initial direction of the cantilever beam, traditional linear piezoelectric energy harvesters essentially fail, whereas the proposed harvester can still recover to its optimal working state through adaptive adjustment. The proposed bio-inspired omnidirectional adaptive piezoelectric vibration energy harvester provides important theoretical and technical references for improving the conversion efficiency of piezoelectric vibration energy harvesters and solving the self-powering problem of low-power micro/nano devices.关键词：piezoelectric vibration generator;omnidirectional adaptive hovering mechanism;biomimetic design;dynamic characteristics;self-power issue51|93|0更新时间：2026-10-16
- 摘要：As new-generation wireless networks evolve toward complex networks integrating communication, sensing, and computing, the number of communication devices is experiencing explosive growth. The concurrent access of massive devices triggers intense spectrum conflicts, which, compounded by complex and dynamic electromagnetic interference threats in open channels, leads to difficulties in collaborative communication within large-scale networks. Although existing multi-agent-based anti-jamming algorithms can alleviate these conflicts, their optimization parameters often increase linearly or exponentially with the number of agents, and the resulting massive training overhead and memory consumption make large-scale expansion difficult. To address the challenges of collaboration and limited scalability, this paper proposes an embodied intelligent anti-jamming communication strategy generation method co-driven by general knowledge and specific information. This method constructs a unified shared neural network architecture through a parameter-sharing strategy, decoupling the growth of optimization parameters from the number of agents, thereby significantly reducing the scale of model parameters while enhancing system scalability. On this basis, a shared convolutional neural network (CNN) is utilized to extract features from the electromagnetic environment, learning general knowledge representations from raw spectrum observations that are available to all agents. Meanwhile, to distinguish individuals, the method introduces learnable identity embedding vectors as specific information and performs joint modeling with general knowledge, thereby achieving differentiated decision-making for homogeneous agents and generating efficient collaborative anti-jamming communication strategies. Simulation results demonstrate that this method offers significant advantages in fixed jamming, sweep jamming, and intelligent jamming scenarios, effectively alleviating collaborative conflicts in large-scale anti-jamming communication; particularly in the sweep jamming scenario, the self-adaptive communication efficiency (SACE) of this method improves by approximately 22% and 38% compared to the multi-agent proximal policy optimization (MAPPO) method and the independent deep Q-network (IDQN) method, respectively, verifying the efficiency of the collaborative mechanism. Furthermore, simulation experiments on multi-channel multi-user scenarios of varying scales, conducted without altering the network architecture, show that the method achieves stable convergence and excellent performance, demonstrating good scalability.关键词：embodied intelligence;multi-agent reinforcement learning;collaborative anti-jamming;parameter sharing;embedded identification;distributed decision making67|128|0更新时间：2026-10-16
- 摘要：With the rapid development of space-air-ground integrated systems, intelligent on-board satellite computing has become an important technical foundation for supporting remote sensing perception, situation analysis, and emergency decision making. However, it faces several challenges, including limited on-board computing power and storage, complex noise in multimodal heterogeneous data, and dynamic changes in task semantics. Existing multimodal fusion methods mainly emphasize vision-text alignment or large-model-driven reasoning, while knowledge graph methods focus on structured semantic modeling. Most of these methods are designed for ground-side high-performance computing environments and lack a unified deployment mechanism for satellite edge computing scenarios. To address these problems, this paper proposes a knowledge-guided multi-modal modular satellite network (KM³-SatNet) for space-air-ground integrated systems. Different from existing methods that directly introduce a global knowledge graph or adopt a fixed fusion structure, this paper develops an integrated design around “task-related knowledge pruning, knowledge-constrained robust fusion, budget-aware dynamic reasoning, and edge-cloud collaborative adaptive deployment”. First, a satellite task knowledge graph is constructed, and dynamic knowledge subgraph extraction is achieved through semantic matching, relation-aware expansion, and personalized PageRank. Second, a knowledge-constrained gated fusion mechanism is designed to establish interpretable dynamic interaction constraints among vision, text, and knowledge. Furthermore, a modular and prunable inference structure and a computing-power-aware controller are proposed to realize adaptive module selection under on-board budget constraints. Finally, an edge-cloud collaborative optimization strategy considering link fluctuation and interruption fallback is presented. Experimental results on multimodal remote sensing task datasets show that, compared with typical multimodal fusion models, the proposed method improves accuracy by 6.3% on the object recognition task and by 5.8% on the disaster scene recognition task. Meanwhile, compared with the lightweight remote sensing multimodal large language model lite (RS-MLLM-Lite), the number of model parameters is reduced by 50.0%; compared with Swin-L, the number of parameters is reduced by 98.7%, and inference latency is reduced by 67.7%. The results demonstrate that KM³-SatNet can maintain high accuracy while balancing model scale, latency, energy consumption, and robustness, making it suitable for multimodal intelligent computing deployment in satellite edge environments.关键词：space-air-ground integration;satellite computing;multimodal fusion;knowledge enhancement;task reasoning;edge-cloud collaboration54|78|0更新时间：2026-10-16
- 摘要：In recent years, the growing demand for global sustainable energy development has accelerated the deployment of photovoltaic (PV) systems across residential, commercial, and industrial sectors. To ensure the safe operation of power grids and improve energy dispatch efficiency, accurately forecasting user-side power load variations has become increasingly critical. However, existing methods often overlook that the data collected from user-side smart meters typically contain both PV generation (source) and power consumption (load) components, which are not effectively separated, leading to biased prediction results. To address this issue, this paper proposes a multi-agent reinforcement learning-based source-load decoupling method, namely SoLED, that mitigates prediction bias caused by the unseparated mixed data of load and PV generation, as well as insufficient modeling of complex spatiotemporal dependencies among users. Specifically, we first design a feature extraction module to capture both short-term and long-term variations in user load, thereby enhancing the representation of spatiotemporal characteristics within the power grid. Then, based on the physical topology of the grid, we construct a simulation environment that models voltage responses under different load and PV power conditions. Finally, leveraging the feedback generated from this simulation, we train a decoupling model within a multi-agent reinforcement learning framework to achieve accurate source-load separation. Experimental results demonstrate that the proposed method improves prediction accuracy by 5.7%~24.6% and 6.6%~24.2% under different weather conditions on two real-world urban power grid datasets. These results confirm the effectiveness of the proposed approach in accurately decoupling user-side mixed power data and enhancing the modeling capability of complex spatiotemporal dependencies, thereby reducing the prediction bias in user-side load forecasting.关键词：photovoltaic generation system;spatiotemporal feature modeling;prototype learning;reinforcement learning;multi-agent reinforcement learning;source-load decoupling50|28|0更新时间：2026-10-16
- 摘要：Far-field object localization plays an important role in wireless communications, electronic reconnaissance, object surveillance, and intelligent sensing applications. However, as the distance between the object and receiving stations increases, severe path loss, shadow fading, and Non-Line-Of-Sight (NLOS) effects significantly degrade the received signal quality, thereby deteriorating the performance of conventional passive localization systems. To address this issue, transceiver-aided localization techniques employ transceivers to amplify and forward signals, thereby establishing line-of-sight propagation paths and improving localization performance. Nevertheless, existing transceiver-aided localization studies mainly focus on active or cooperative localization scenarios and rely on multiple measurements of the object-transceiver-receiver propagation time to achieve high localization accuracy. In passive localization applications, such as electronic reconnaissance and covert object detection, the observation system cannot interact cooperatively with the object, rendering existing transceiver-aided localization methods inapplicable. To address this challenge, this paper investigates far-field passive localization with unknown transceiver positions and proposes a joint object and transceiver localization algorithm based on spatio-temporal observations. Specifically, two instantaneous time difference of arrival (TDOA) measurement models are constructed by exploiting the signal propagation paths of object-common transceiver-different receivers and object-different transceivers-common receiver, respectively. In addition, an angle of arrival (AOA) measurement model is established using the angle information obtained at a reference receiver. Subsequently, two optimization problems corresponding to transceiver localization and object localization are formulated based on the TDOA measurements. By applying semidefinite relaxation (SdR) technique, the resulting nonconvex problems are transformed into convex optimization problems, yielding coarse estimates of both transceiver and object positions. Since the first two stages do not fully exploit the available measurements, a third-stage refinement procedure is developed based on TDOA-only and joint TDOA-AOA observations to further improve localization accuracy. Furthermore, the Cramér-Rao lower bound (CRLB) corresponding to TDOA-only and joint TDOA-AOA observations is derived, and the minimum node requirements for localization feasibility are established. Theoretical analysis demonstrates that, under moderate measurement noise and error conditions, the proposed joint localization algorithm can achieve the CRLB performance for both object and transceivers position estimation. Numerical simulation results further validate the theoretical analysis and show that incorporating AOA information from a single receiver can significantly enhance the joint localization accuracy of both the object and transceivers.关键词：far-field object;transceiver;passive localization;time difference of arrival;Cramér-Rao lower bound84|92|0更新时间：2026-10-16
- 摘要：Polarization information is a crucial parameter in radar target recognition and feature extraction. Existing researche on polarized signals mainly focuses on fully polarized signals. Nevertheless, when radar irradiates sea surface ships or aerial maneuvering targets, the coupled effects of continuous target attitude variation, multi-scattering center interference and time-varying environmental clutter endow radar echoes with inherent partially polarized characteristics. The polarization state of such echoes is time-varying rather than constant, and contains abundant physical information of targets. At present, the analysis of widely existing partially polarized signals in actual radar echoes is limited to static statistical description, while the modeling and analysis of dynamic evolution mechanism of time-varying polarization state are insufficient. To address this insufficiency in dynamic characterization, this paper aims to establish a unified dynamic polarization state model applicable to maritime and aerial dynamic targets. The model decomposes the received signal into a fully polarized component and a dynamically fluctuating unpolarized component. By introducing a target attitude dynamic model based on stochastic differential equations, a nonlinear coupling model between polarization parameters and attitude angles, and a dual-scale unpolarized power model, a unified dynamic polarization state framework is constructed. This framework simultaneously accommodates the low-frequency periodic motion of maritime targets and the high-frequency maneuvering characteristics of aerial targets. Based on this framework, the statistical properties and spectral structure of the time-varying degree of polarization and Stokes parameters are theoretically derived, revealing that target micromotion and scattering fluctuations are key physical mechanisms driving the evolution of the polarization state. Validation results against measured data demonstrate that the proposed model effectively simulates the dynamic polarization behavior of partially polarized signals from maritime and aerial targets. The fluctuations in the degree of polarization and spectral features of the simulated signals show good consistency with the measured data, verifying the model’s effectiveness in revealing target micromotion and scattering mechanisms. The model established in this work provides a unified modeling framework with clear mechanisms and strong interpretability for feature extraction and target recognition using dynamic partially polarized signals, facilitating the quantitative inversion of target physical states from radar returns.关键词：partially polarized signal;time-varying polarization state;degree of polarization;unpolarized component;stokes parameter;poincaré sphere66|46|0更新时间：2026-10-16
- 摘要：The existing methods often combine the weighted prediction error (WPE) de-reverberation method with independent component analysis (ICA) and independent vector analysis (IVA) techniques to realize convolutional blind source separation in a reverberation environment. However, ICA and IVA do not construct flexible sound source models, so the independent low rank matrix analysis (ILRMA) technique is proposed, which in turn constitutes the joint optimization of WPE and ILRMA, namely WPE-ILRMA. However, the performance of WPE-ILRMA still depends on the number of iterations and the setting of filter order. When the number of iterations is small, the performance of the method is poor, and the fixed filter order setting cannot cope with the complex and changeable reverberation environment. Moreover, the update of the separation filter still needs to invert the matrix, which consumes a lot of hardware resources. The impulse response remodeling (IRR) model using the gradient descent method to optimize the pre-filter is proposed to shorten the impulse response length and reduce the influence of a high reverberation environment. However, the objective function Structured by these methods based on the IRR model needs to set the norm value, which is only set by experience. Improper settings will easily lead to the decrease of IRR effect. However, it is impractical to continuously adjust the norm value of the objective function according to the IRR effect in a blind signal processing scenario. Aiming at the problem that the performance of existing convolutional blind source separation methods depends on the norm value or filter order setting, the concept of compression transform is introduced into IRR technology for the first time. It is proven that the compression transform can achieve reverberation suppression performance better than p-norm, and then the compression transform can be used to replace p-norm in the traditional IRR model. An innovative impulse response compression (IRC) model combining the compression transform and the infinite norm is designed. The proposed IRC model not only achieves superior reverberation suppression performance, but also does not need to adjust reverberation suppression performance by empirically setting a norm value or filter order like traditional de-reverberation methods. In addition, the iterative source steering (ISS) technology and ILRMA are introduced to avoid the matrix inversion operation, which makes the proposed method easier to implement in hardware. The simulation results show that the proposed method achieves superior signal separation performance compared with other methods in a high reverberation environment, whether for the animal sound signal separation problem with small component difference, the music signal separation problem with complex components, or the speech signal separation problem with large component difference, which shows that the proposed method has strong robustness and a wide application range. The existing methods adopt a fixed norm value or filter order setting. Improper parameter setting will lead to the performance degradation of the method, but the performance of the proposed method does not depend on the adjustment of the norm value or filter order. The proposed method breaks through the application limitations of existing methods, and the IRC technology can also be applied to other engineering problems related to de-reverberation.关键词：de-reverberation;convolutional blind source separation;independent low rank matrix analysis;impulse response remodeling;impulse response compression;iterative source steering47|25|0更新时间：2026-10-16
- 摘要：Constrained multi-objective optimization problems (CMOPs) are widely prevalent in practical engineering and scientific fields. Such problems typically involve multiple conflicting optimization objectives and a variety of complex constraints. In constrained multi-objective evolutionary algorithms (CMOEAs), the design and integration of the environmental selection strategy (ESS) are crucial, as it directly determines whether the algorithm can achieve an effective balance among feasibility (degree of constraint satisfaction), convergence (speed of approaching the Pareto front), and diversity (uniformity of solution set distribution). However, existing methods mostly rely on static or predefined strategy scheduling mechanisms, making it difficult to adaptively adjust selection pressure according to the real-time state of the population during evolution. Consequently, when dealing with practical scenarios such as complex constraints, discontinuous feasible regions, and expensive objective evaluations, these algorithms exhibit significant limitations. On one hand, when the population is entirely infeasible in the early stages of evolution, there is insufficient driving force to guide the search quickly into the feasible region. On the other hand, after feasible solutions are found, it becomes challenging to dynamically achieve a reasonable trade-off between convergence acceleration and diversity maintenance, often leading to premature convergence or loss of diversity. In some extreme constraint conditions, the algorithm may even fail to find any feasible solution for a long time. These issues severely constrain the overall performance and robustness of CMOEAs. In view of this, this paper proposes a multi-environmental selection strategy integration method based on reinforcement learning (RL) to enhance the autonomous decision-making and dynamic adaptability of constrained multi-objective evolutionary algorithms. The proposed method first characterizes the evolutionary state from three dimensions (feasibility, convergence, and diversity) based on the distribution characteristics of the population in both decision space and objective space. The ε-feasible solution ratio is used to reflect the overall level of constraint satisfaction of the population; the normalized centroid displacement is used to indicate the convergence trend; and the proportion of the variance of ε-feasible solutions in the total population variance is used to evaluate diversity. Then, inspired by the principle of interaction forces in physics, this paper designs three distinct yet inherently coupled environmental selection strategies. The first strategy focuses on enhancing feasibility preservation, constructing a repulsion force model by ranking constraint violation degrees to prioritize feasible solutions. The second strategy focuses on accelerating convergence, using objective function ranking to guide the population toward the Pareto front. The third strategy is dedicated to diversity maintenance, generating repulsion forces solely based on spatial distances among individuals to avoid excessive clustering. These three strategies are coupled with each other, such that even if a suboptimal strategy is selected at a certain step, it does not cause a sharp deterioration in algorithm performance, thereby providing a stable decision-making foundation for reinforcement learning. Finally, a mapping model between population states and environmental selection strategies is constructed through reinforcement learning. Experimental results on three benchmark test suites (LIR-CMOP, DASCMOP, DOC) and an analog integrated circuit optimization design problem show that the proposed method significantly outperforms six state-of-the-art constrained multi-objective evolutionary algorithms on most test problems. It exhibits stronger robustness and adaptability, particularly in challenging scenarios with narrow feasible regions, complex constraints, and discontinuous Pareto fronts. It obtains higher-quality solution sets with fewer function evaluations, thereby fully validating its effectiveness, scalability, and engineering application value in handling practical complex constrained multi-objective optimization problems.关键词：constrained multi-objective optimization;Multi-environment selection strategies;evolutionary optimization;reinforcement learning;Analog integrated circuits99|48|0更新时间：2026-10-16
- 摘要：Low-frequency electromagnetic waves (below 300 kHz) have natural advantages such as strong penetration and low transmission loss, enabling stable and reliable wireless communication in complex lossy media including seawater, soil, and buildings. However, limited by the electrical size constraint, the physical dimensions of an antenna need to be comparable to the operating wavelength (typically greater than one-tenth of the wavelength) to achieve sufficient radiation efficiency. Consequently, conventional low-frequency antennas often require enormous sizes of hundreds of meters, making it difficult to meet the urgent demands of modern cross-media communication applications for device portability and rapid deployment. In this work, a miniaturized low-frequency antenna design based on the Tesla coil magnetic resonance principle and high-voltage pulse excitation is proposed, achieving long-distance signal transmission at a size of one-thousandth of the operating wavelength. The design significantly amplifies the current in the secondary circuit through the magnetic resonance structure, and employs an impedance matching transformer to efficiently transfer energy from the signal source to the high-impedance magnetic resonance coil, thereby enhancing the magnetic field. Meanwhile, high-voltage pulse excitation is introduced, enabling the antenna to obtain a higher instantaneous peak magnetic field under the same average power consumption while effectively suppressing the heating problem caused by continuous high current, thus balancing radiation capability and system reliability. Through systematic parameter optimization, this paper fabricated an antenna prototype operating at 200 kHz low frequency (LF) with a size of 70 cm. Experimental results show that when the antenna is driven by high-voltage pulse excitation, the received signal peak voltage reaches 2.07 times that of continuous wave excitation. Moreover, compared with a single-coil structure with the same number of turns, the magnetic resonance coupling structure increases the received power by about 13.8 dB, corresponding to a signal strength enhancement of 4.9 times, verifying the significant improvement of the synergistic effect of pulse excitation and magnetic resonance on the radiation field strength. On this basis, an outdoor long-distance communication test was conducted. Driven by a peak power of 800 W, the antenna achieved a communication distance of 400 m in a complex campus electromagnetic environment, with a received magnetic field strength of 0.27 pT. To meet the demands of emergency communication for ultra-portability and strong penetration capability, a portable antenna operating at 30 kHz very low frequency (VLF) with overall dimensions of only 10 cm × 7 cm × 4 cm was further developed. Under an input power of 20 W, the antenna successfully achieved reliable cross-medium transmission through three reinforced concrete floors over a distance of 15 m. The proposed scheme, through the synergistic design of high-voltage pulse excitation and magnetic resonance amplification, enables both long-range communication and high-penetration cross-medium communication at centimeter-to-decimeter scales, providing a miniaturized solution for applications such as emergency communication, underground rescue, and underwater sensor networks.关键词：low-frequency antenna;Tesla coil;magnetic resonance;cross-medium communication;magnetic induction communication;wireless transmission71|66|0更新时间：2026-10-16
- 摘要：Prompt learning utilizing large-scale unlabeled data has recently received widespread attention. Among these, the true-false label learning paradigm randomly samples a category from the candidate set and leverages pre-trained vision-language models to make binary judgments, i.e., True or False, thereby generating high-precision weak supervision signals, substantially improving annotation efficiency, and driving the rapid development of the prompt learning field. However, its random uniform sampling mechanism inevitably leads to missing supervision samples for certain categories, making it difficult for downstream task models to construct an accurate category semantic space, and existing methods sever the cross-modal joint optimization path between textual and visual features, destroying the original cross-modal generalization capability of pre-trained models. To address the aforementioned limitations, this paper proposes a true-false labels learning method based on multi-modal consistency collaborative prompt tuning (MCCP). This method introduces consistency constraints in both textual and visual modalities, and designs a cross-modal collaborative optimization strategy. Specifically, in the textual modality, a large language model-driven knowledge collaboration mechanism is designed to generate fine-grained textual descriptions for each category. By constraining the semantic consistency between original textual prompts and category textual descriptions, it fills the semantic information blind spots of categories with missing supervision samples, and implicitly mitigates the effects of insufficient supervision information. In the visual modality, a consistency prompt learning mechanism based on self-supervised learning is proposed. By constructing multiple views, it constrains the pairwise feature distribution consistency among original images, weakly augmented views, and strongly augmented views, thereby forcing the visual prompt network to extract deep semantic features insensitive to data augmentation operations, significantly enhancing the robustness and generalization of visual-semantic representations. In terms of cross-modal joint optimization, an epoch-based cosine dynamic weighting strategy is introduced to achieve a smooth transition and gradual fusion between pre-trained prior knowledge and domain-adaptive feature representations, while a staged gradient update scheme is adopted to separately optimize the consistency losses of text- and vision-specific modalities and the empirical risk-consistency cross-entropy loss of cross-modal image classification, effectively avoiding the gradient conflict problem in multi-modal learning and promoting the deep collaborative alignment of cross-modal semantic spaces. Experimental results on nine multi-class image classification datasets, including CIFAR-100, Tiny-ImageNet, Caltech-101, Food-101, Stanford Cars, describable textures dataset (DTD), fine-grained visual classification of aircraft (FGVC-Aircraft), ImageNet-A and ImageNet-R, demonstrate that this method achieves state-of-the-art average classification accuracies of 73.58% and 74.20% on the True-False labels datasets annotated by the contrastive language-image pre-training model (CLIP) and human annotators, respectively, and simultaneously exhibits optimal transfer generalization performance on the vast majority of target datasets under the cross-dataset transfer setting, verifying the effectiveness of MCCP. This paper effectively addresses the missing supervision signal challenge of true-false labels learning paradigm, providing methodological support for achieving efficient and robust multi-modal cross-domain knowledge transfer.关键词：true-false labels;prompt learning;vision-language models;weakly supervised learning;multi-modal consistency;image classification78|24|0更新时间：2026-10-16
- 摘要：This study designed a broadband current probe for transient pulse measurement and developed a corresponding methodology for measuring and calibrating its transfer impedance. Initially, theoretical calculations determined the required magnetic core permeability for achieving specific dimensional targets and desired transfer impedance. The sensor was realized using a novel nanocrystalline magnetic core combined with a toroidal disk resistor load exhibiting low parasitic inductance. Furthermore, the existing measurement setup for characterizing current sensor transfer impedance was improved. The transfer impedance characteristics of the designed sensor were then measured and validated using a vector network analyzer (VNA). Test results demonstrated that the transfer impedance maintained excellent flatness across the measurement frequency range of 20 kHz to 1 GHz. For time-domain validation, transient current waveforms generated by direct discharge from an IEC 61000-4-2 standard generator with a 700 ps rise time were employed. The current waveforms measured by the sensor showed close agreement with the outputs from a standard current target across different test voltage levels. Finally, the designed current sensor was successfully applied to measure injected currents in a mobile terminal system subjected to electrostatic discharge (ESD) gun excitation, and to measure port disturbance currents on an electronic capacitor voltage transformer (ECVT) electronic device within a power system substation under electrical fast transient (EFT)/Burst conducted disturbances. It also realized the injected-current monitoring of transmission line pulse (TLP) test. These applications provided valuable data foundations for further system-level analysis.关键词：transient pulse;electro-static discharge;current probe;calibration setup;transfer impedance60|29|0更新时间：2026-10-16
PAPERS
- 摘要：The collaborative deployment of reconfigurable intelligent surfaces (RISs) and unmanned aerial vehicles (UAVs) provides a promising technological approach to improving the coverage, transmission capacity, and service flexibility of future wireless networks. Its central challenge lies in the joint optimization of UAV trajectories, RIS phase-shift matrices, active beamforming, and resource allocation variables. This problem is characterized by inter-slot coupling, high-dimensional action spaces, nonconvex constraints, and dynamically varying environments. In large-scale or rapidly time-varying scenarios, conventional methods based on alternating optimization, successive convex approximation, and semidefinite relaxation generally suffer from high iterative complexity, limited online adaptability, and a strong dependence on accurate channel state information. Deep reinforcement learning (DRL) can formulate the joint optimization problem as a sequential decision-making process, thereby providing a new approach to online policy generation in complex dynamic environments. This paper presents a systematic review of DRL-based joint trajectory and phase-shift optimization in RIS-assisted UAV communications. First, a unified optimization framework incorporating trajectory, phase-shift, and beamforming variables is established, and the key choices in Markov decision process (MDP) modeling are analyzed from the perspectives of action-space design, state-space construction, and reward-function formulation. Existing studies are then classified into three technical paradigms according to the representation and decomposition of joint actions: discrete decision-making, continuous decision-making, and distributed cooperation. Their algorithmic mechanisms, applicable scenarios, sample efficiency, convergence stability, scalability, and implementation costs are comparatively examined. Four common bottlenecks are further identified, namely limited scalability, idealized channel state information (CSI) assumptions, the simulation-to-deployment gap, and the lack of fair benchmarking protocols. The comparative analysis shows that discrete methods are suitable for small- and medium-scale RISs with low-bit phase-shift control, whereas continuous methods provide higher control precision but are constrained by high-dimensional policy learning and limited sample efficiency. Multi-agent methods are well matched to distributed system architectures but introduce environmental nonstationarity and credit-assignment difficulties. In large-scale scenarios, most existing approaches rely on structured dimensionality reduction or hybrid decomposition combining DRL with conventional optimization, while strictly end-to-end joint optimization remains relatively uncommon. Finally, future research directions are discussed, including structured DRL architectures, offline reinforcement learning and digital twins, standardized benchmarking platforms, near-field and terahertz communications, and integrated communication, sensing, and computing. This review provides a reference for subsequent algorithm design, performance evaluation, and system deployment.关键词：reconfigurable intelligent surface;UAV communication;deep reinforcement learning;trajectory optimization;phase-shift optimization103|187|0更新时间：2026-10-16
SURVEY AND REVIEW
- 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