To address the challenges of modeling maneuvering targets and achieving stable tracking in multi-target scenarios, this paper proposes the spatial-temporal hybrid attention network (STHA-Net), an end-to-end multi-target tracking model based on the Transformer architecture. STHA-Net incorporates an STHA mechanism to capture the temporal continuity of target motion and characterize inter-target interactions, thereby enhancing motion modeling capability. Further, a dynamic gating fusion mechanism is employed to adaptively balance temporal and spatial attention heads, leveraging their respective strengths. This design enables the model to maintain high tracking accuracy under various dynamic scenarios. Extensive experiments in complex multi-target tracking environments demonstrate that STHANet outperforms baseline methods reducing tracking error by 12.98% and 20.51% compared to Transformer and attention based bidirectional-long short-term memory methods, respectively. Furthermore, compared to traditional filtering algorithms, STHA-Net achieves a 54.43% and 51.43% improvement in tracking accuracy.
With the continuous improvement of 5G network infrastructure, the informatization process in the industrial sector is accelerating at an unprecedented pace. The market demand for gateways integrating 5G networks with multi-access edge computing (MEC) technology is growing increasingly strong. There is also a rising requirement for these gateways to possess capabilities such as high reliability, large bandwidth, fast processing speed, and localized data processing. This paper, in response to the practical management needs of photo voltaic power stations and industrial fields, proposes the design of a 5G artificial intelligence (AI) edge computing intelligent gate way with dual 5G modules. The design is based on the i.MX8M Plus processor and integrates dual 5G module design. It supports multiple communication modes, including 5G, Wi-Fi, Ethernet, and serial ports. The software incorporates edge computing scripts, various communication protocols, and a lightweight AI model, enabling real-time data acquisition and processing at the edge. It supports for several industrial standard protocols ultra low latency acquisition, and intelligent operation features like remote over-the-air upgrade, secure shell secure operation, and configuration hot updates. The system was tested and validated in a typical photovoltaic application scenario. The results demonstrate that, compared with the traditional single-module Internet of Things gateways, the dual 5G AI edge gateway proposed in this paper reduces end-to-end latency by more than 70%, improves data reliability tenfold, enhances bandwidth by 100%–200%, extends coverage capability by 50%, effectively alleviates the bandwidth pressure on the core network, and enhances the overall reliability and security of the system.
In modern radar systems, there is a significant demand for inverse synthetic aperture radar (ISAR) imaging from random missing sampling (RMS). The matrix completion (MC) methods offer a novel approach for ISAR reconstruction from RMS data, one that avoids the discrete grid effect. However, existing MC approaches based on Hankel transformation exhibit limited computational efficiency due to its anti-diagonal structure. To address the need for high-quality and efficient reconstruction in RMS ISAR imaging, this paper proposes a fast reconstruction algorithm utilizing a Hermitian-Toeplitz (HT) structured matrix formulation. Numerical simulations demonstrate that the HT matrix exhibits superior low-rank property and faster convergence than the Hankel matrix, which is more favorable for rapid reconstruction. As for the case of azimuthal RMS, a range-cell sliding window is employed to construct the HT matrix from the target reconstruction matrix, thereby preserving coherence in the range dimension. By leveraging the strong low-rank property and structural feature of the HT matrix, the proposed algorithm employs truncated singular value decomposition for fast iterative updating, significantly improving computational efficiency. The simulation experiments verify that the proposed algorithm achieves faster reconstruction efficiency compared to the Hankel matrix transformation and maintains robust reconstruction capability under low signal-to-noise ratio conditions. Additionally, the effectiveness of the algorithm has been confirmed through measured data.
Compressive sensing (CS) has been successfully combined with deep-learning neural networks to generate models for image-depth CS. Many existing models face significant limitations, such as their inability to handle tasks with multiple CS ratios and suboptimal performance in image reconstruction. To overcome these challenges, an optimized deep network model, named global multi-attention condition module (GMCM) iterative shrinkage-thresholding algorithm (ISTA)-Net++ is proposed. In this model, the CS ratio of an image is incorporated as a network parameter, enabling the network to effectively handle multiple CS ratio tasks. Additionally, a hybrid attention mechanism is employed to preprocess the reconstructed image for feature extraction and to calculate the image complexity score. These elements are then combined with the CS ratio to jointly predict the parameters and iterative reconstruction steps, which are subsequently broadcast to all iterative stages of the model. Extensive experiments on high- and low-resolution image datasets demonstrate that the proposed GMCM ISTA-Net++ model achieves a superior performance compared to traditional models. Given its flexibility, enhanced reconstruction quality, and adaptability, this model holds significant potential for deployment in resource-constrained mobile image-processing terminals.
The monopulse technique can acquire target azimuth information independently of the Doppler bandwidth, making it useful for solving the issue of narrow Doppler bandwidth in for ward-looking radar imaging. However, the presence of angular glint causes interference between the scatters in the azimuth direction, which will result in poor resolution in the monopulse forward-looking imaging procedure. Therefore, the resolution enhancement in the azimuth direction has become an urgent problem to be solved. To address the aforementioned problem, this paper proposes an approach for monopulse forward-looking radar imaging based on difference channel maximum likelihood estimation algorithm. The proposed method simultaneously exploits both amplitude and phase information for angle estimation, providing superior resolution, and offers four configurations that can be flexibly selected to accommodate various imaging scenarios and system requirements. The simulation results confirm the superiority of the proposed method over the traditional method in terms of undistortion, resolution, and robustness under low signal-to-noise ratio.
In aerial target recognition tasks, single-view inverse synthetic aperture radar (ISAR) methods are highly susceptible to target attitude variations and partial occlusions, which often result in incomplete structural representation and consequently limit recognition accuracy. To address this issue, this paper proposes a multi-view attention fusion network (MAFNet) for multiview ISAR image recognition. MAFNet employs a shared EfficientNet backbone to extract deep features from multiple viewing angles, adaptively fuses multi-view features using a multihead attention mechanism, and further exploits view-level max pooling to capture globally discriminative representations, enabling end-to-end multi-view ISAR target recognition. Comparative experiments are conducted on datasets comprising six classes of simulated aircraft targets and seven classes of measured aircraft targets. Experimental results demonstrate that the MAFNet achieves an average recognition accuracy of 98.27% ± 1.72% on the simulated dataset at a signal-to-noise ratio (SNR) of 20 dB, and maintains a high accuracy of 95.07% ± 4.01% even under severe noise conditions with an SNR of −10 dB. On the measured dataset, MAFNet attains an accuracy of 95.62% ± 5.12%, consistently outperforming conventional single-view and existing multi-view recognition methods. Moreover, MAFNet contains only 10.6×106 parameters and requires 1.3 G floating point operations per sample, achieving an effective balance between recognition performance and computational efficiency. These results validate the proposed method's superiority in terms of accuracy, robustness, and efficiency, providing a reliable technical solution for high-precision aerial target recognition in complex operational scenarios.
This paper proposes a two-dimensional (2D) angular super-resolution framework for sparse arrays under the single snapshot condition. The azimuth-elevation 2D angular super resolution model is established, which shows the relationship between the 2D angular super-resolution image and the signal. Using this model, the 2D angular super-resolution problem is transformed into the beyond linear optimization problem. In order to efficiently and accurately address this optimization problem, we propose the sparse multi-layer iterative (SMLI) algorithm based on beyond linear signal processing (BLiSP) theory. During the layered iterative process, the solution region is continuously narrowed. The constructed nonlinear weighting matrix and sparse constraint coefficient enhance the ability to differentiate between the effect of signal and noise in solving the beyond linear optimization problem. Additionally, the nonlinear weighting matrix can be adaptively updated during the solving process, ensuring high-performance angular super-resolution results under different signal-to-noise ratios (SNRs). Experiments confirm the proposed method's effectiveness and robustness.
Aiming at the issues where traditional direction-of-arrival (DOA) estimation algorithms experience substantial performance degradation in low signal-to-noise ratio environments, and deep learning-based DOA estimation methods rely on massive training data with prolonged model training cycles, this paper proposes two efficient and high-precision DOA estimation methods based on ensemble learning. By formulating DOA estimation as a multi-label classification problem and leveraging the classification chain paradigm, data-driven models, classification chain-random forest (CC-RF) and classification chain-eXtreme gradient boosting (CC-XGBoost), are constructed, which are capable of handling multi-label classification tasks. To verify the effectiveness of the proposed methods, a multi-dimensional comparative experiment is designed to benchmark their performance against the traditional multiple signal classification (MUSIC) algorithm and a convolutional neural network (CNN) model. Experimental results indicate that in both single-source and multi-source scenarios, the proposed CC-RF algorithm exhibits excellent performance, achieving DOA estimation accuracy comparable to the MUSIC algorithm; in multi-source estimation scenarios, both proposed models demonstrate strong noise adaptability. Compared with the traditional MUSIC and CNN algorithms, the estimation error of the CC-XGBoost and CC-RF models is reduced by up to nearly 30 times while maintaining low time complexity, with the single estimation time reduced by approximately 90% compared to traditional methods. This study provides a technical pathway for DOA estimation in complex environments and holds significant applicationvalue in fields such as radar detection and wireless communication.
When detecting weak targets, radar systems often encounter the interference from complex clutter in the field scene, leading to a high false alarm rate and a low detection rate when the traditional radar methods are employed. To address this issue, this paper proposes a clutter suppression method based on generative adversarial networks (GANs) to improve the detection performance of weak targets. This method designs a clutter identification network based on the characteristics of clutter in the range-Doppler domain. By adopting a divide-and-conquer strategy, a clutter suppression network is selected based on the clutter identification results. This method transforms the clutter suppression process into a mapping problem from the clutter-affected data domain to a clutter-free data domain. Experiments results show that the proposed method achieves a 10% improvement in detection rate in rain clutter sce narios, and reduce the false alarm rate by 90% in ground clutter scenarios.
Accurately sensing the channel state of heterogeneous networks is key to matching users' diverse service communication demands with the channel state, and is an effective way to improve the utilization efficiency of network resource. However, existing channel state perception methods are not suitable for heterogeneous network, and their perception performance is easily affected by interference uncertainty. In order to achieve channel state perception of heterogeneous networks, this paper adopts a centralized collaborative perception model, where each node obtains local channel state perception results based on statistical pulse parameters at the physical layer. In order to reduce the impact of interference on perception performance, this paper uses the Jousselme distance to quantify the degree of difference among nodes caused by interference. Using the average credibility as a threshold, nodes in the sensing area are classified. On this basis, the local perception results of each node are performed classification-based correction to improve the accuracy and reliability of channel state perception. Simulation results indicate that the proposed method has good adaptability for channel state perception in complex electromagnetic environments. The perception results can accurately reflect the actual channel state, which is conducive to improving the network throughput.
To address the limitations of traditional coherent sources direction-of-arrival (DOA) estimation methods, which rely on rank recovery of covariance matrices and fail to perform effectively in complex environments, and the instability of neural network models caused by insufficient spatiotemporal feature extraction during coherent DOA estimation, this paper proposes a hybrid neural network combining convolutional neural network (CNN) and Transformer to jointly extract spatiotemporal features from array-received signals. The proposed model treats DOA estimation as a regression problem, directly modeling the mapping relationship between received signals and DOA. By using array-received signals as input, the model employs CNN to extract spatial features firstly. The resulting feature maps are restructured and fed into a multi-head attention mechanism to capture spatial relationships between array elements. Finally, the learned features are input into a Transformer model to capture long-range dependencies in received signals across different snapshots from the same array element, thereby suppressing temporal coherence. By combining CNN�s local feature extraction and Transformer's global dependency modeling, this approach overcomes the limitation of single-model architectures in simultaneously capturing local and global features, significantly improving the representation capability for complex coherent signals. Experimental results demonstrate that the proposed method exhibits strong robustness and reliability under challenging conditions such as low signal-to-noise ratio, limited snapshots, and multiple sources.
To address the challenge of recognizing active electronically scanned arrays (AESAs) radar work modes under few labeled data, this paper proposes a semi-supervised algorithm based on multi-task contrastive learning. The method employs a dual-stage collaborative optimization framework combining self-supervised pre-training and supervised fine-tuning to effectively leverage latent feature information from unlabeled data. In the self-supervised pre-training stage, a dual-path data augmentation strategy is designed to generate mode-consistent pulse sequence samples, where an expander and a base encoder work synergistically to capture both temporal dependencies and cross-pulse deep correlation features. Additionally, three self-supervised tasks, contextual contrastive learning, pulse sequence reconstruction, and augmentation feature prediction, are introduced to optimize feature space distribution, disentangle noise interference, and enforce consistency constraints across augmented samples, respectively. During the supervised fine-tuning stage, a classifier integrates self-supervised representations with few labeled data to enhance recognition performance. Comparative experiments demonstrate that the proposed algorithm achieves significantly higher recognition accuracy on the AESA radar work mode dataset compared to other main-stream semi-supervised methods under few labeled conditions. Ablation studies validate the strengthening effect of the multi-task collaborative mechanism on model performance. Additionally, data augmentation comparison experiments identify the optimal augmentation strategy for this algorithm.
Azimuth ambiguity significantly degrades the quality of synthetic aperture radar images. Sub-look spectral analysis (SSA) is a common ambiguity-detection method, but its performance is limited by threshold sensitivity and the high correlation of specific ambiguities across sub-looks. To overcome these specific limitations, this paper proposes an improved detection method. It first increases the number of sub-looks and constructs a high-dimensional multi-look matrix to enrich the coherence differences between targets and ambiguities. Non-negative matrix factorization is then employed to decompose this matrix, effectively separating the coherent target components from the variably coherent ambiguity components without relying on predefined thresholds. Experimental results on real data demonstrate that the proposed improvements achieve superior azimuth-ambiguity-detection performance compared with conventional SSA methods.
Traditional optimization algorithms from operations research are not well-suited for addressing the vehicle routing problem (VRP) with a large number of instances and high real time requirements. Deep reinforcement learning (DRL) introduces a new approach to solving VRP, characterized by its rapid problem-solving capabilities and robust generalization abilities. In the domain of VRP variants, numerous benchmarks are utilized to evaluate the performance of the solving algorithms. However, the limited number of instances in these public benchmarks is insufficient for training DRL algorithms. To address this issue, an effective benchmark augmentation method using distribution derivation is proposed. By modeling the distributions of key VRP variables and estimating their parameters, the method derives distributions closely resembling the original benchmark and subsequently generates a large number of synthetic instances accordingly. The proposed method holds promise for widespread applications across various benchmarks. Experiments on VRP with time windows (VRPTW) and heterogeneous fleet VRP (HFVRP) verify the generalization abilities. The results show that DRL trained using the augmented benchmarks performs much better on public benchmarks than trained using the data customized by the algorithm authors. In addition, the DRL algorithms trained using the augmented benchmarks achieve performance comparable to state-of-the-art algorithms in public benchmarks for various VRP variants while significantly reducing the computation time.
Unmanned aerial vehicle (UAV) swarms operate as collaborative, networked systems with collective intelligence, finding applications in agriculture, transportation, and military surveillance. Traditional human-in-the-loop control of UAV swarms suffers from high operator cognitive load and slow decision cycles. We proposes a digital twin-driven framework that enables intelligent and reliable decision-making for UAV swarms. By establishing a real-time, interactive, and co-evolving virtual environment, the framework transitions swarm supervision from “human-in-the-loop” to “human-on-the-loop” mode. Experimental validation on a multiple-UAV testbed demonstrates that the proposed framework maintains centimeter-level trajectory accuracy, reduces threat response time by 37.3%, and achieves near-zero manual intervention during nominal missions. The framework thus enables faster, more reliable, and scalable autonomous swarm operations by shifting operators from direct control to high-level supervision.
To address the large-scale imaging satellite mission scheduling problem (LISMSP), this paper studies the mission scheduling model and solution algorithm. The analysis of internal relationships among satellite resources and modeling are completed first. Then, this paper proposes a data-driven adaptive multistage dynamic optimization algorithm (DDA-MDO). This algorithm represents a combination of satellite grouping and mission allocation (SGMA), multi-stage dynamic optimization (MDO), and data mining technology. The SGMA aims to reduce the negative impact caused by the massive imaging resources and mission requirements. The MDO dynamically adjusts the observation opportunities of missions to optimize the solution. Historical data are utilized for frequent pattern mining to generate optimum solutions, thereby enhancing the adjusting effect and improving the algorithm's performance further. Experimental results show that the DDA-MDO outperforms meta-heuristics and state-of-the-art methods in large-scale scenarios and obtains the same results as the exact algorithm in general scenarios. With the increase in problem size, it shows better robustness and solution quality.
Aerial gravity measurements in polar regions require higher accuracy and stability from satellite navigation systems. The performance of conventional Global Positioning System (GPS) significantly deteriorates as one moves from lower latitudes to polar areas. First, a regional comparative analysis of BeiDou Navigation Satellite System (BDS) and GPS positioning performance in polar regions is conducted. Results indicate that in polar environments, BDS demonstrates certain advantages regarding satellite availability and geometric distribution. Based on this analysis, a method for extracting airborne gravity anomalies using BeiDou signals in polar regions is proposed and improved. The process begins by optimizing BeiDou observation data using an altitude angle weighting method combined with precise ephemeris data. An error correction model is constructed to enhance system adaptability in the polar environment. Following this, high-precision position and acceleration estimation is achieved using the precise point positioning (PPP) technique, integrated with inertial navigation information, to extract aerial gravity anomalies in polar regions accurately. Results from Antarctic aerial gravity measurement experiments indicate that the BDS-based model improves accuracy by 30% compared to GPS under the same conditions.
Resolving conflict and achieving consensus among social groups with diverse opinions becomes a critical issue in today's extensively connected society. Despite the ubiquitous heterogeneity of connection or contact patterns, the study of how topological characteristics of network structure affect opinion convergence is still insufficient. Based on Deffuant and colleagues' bounded confidence model and the transformable network structure between random network and typical complex network types, including small-world network and scale-free network, we analyze the critical factors affecting continuous opinion convergence. We find that the network density plays a crucial role in the aggregated process of opinions in the social group, followed by the modularized level and the average shortest path length of the social network. However, the structural features have little impact on the consensus phase transition threshold. The further simulation experiments under real networks can be well understood based on the interplay of these three main factors. These findings confirm the paramount importance of creating a high-frequency and widely communicated atmosphere to mitigate conflict and efficiently reach consensus.
The evaluation of air combat decision-making has garnered significant attention due to its potential to effectively mitigate losses resulting from erroneous decisions. However, existing research primarily focuses on static evaluation methods. Therefore, this paper proposes a dynamic multi-round decision evaluation method based on the characteristics of multi-round unmanned aerial vehicle air combat under opponent's optimal strategy. In order to determine objective weights, an improved multi-attribute decision making method is proposed, which incorporates the proximity as a correction coefficient for evaluation indicators, utilizing the cosine similarity instead of Euclidean distance, and incorporating both actual and theoretical objective weights to prevent data mutations. Subsequently, the game theory is employed to reasonably adjust subjective and objective weights to obtain comprehensive weights. To address the issues related to the ambiguity and randomness during the evaluation process, a reverse cloud generator is utilized to determine the center of gravity of the cloud model using comprehensive weights while employing the weighted deviation degree for evaluating air combat decision-making effectiveness. By activating the cloud generator through the cloud model, the optimal strategies for each round of air combat are determined, thereby completing the dynamic evaluations for multi-round sequential decision-making processes. Finally, the feasibility and effectiveness of the proposed method are verified through simulations.
With the improvement of the informatization and intel-ligence level of logistics equipment, the interactive and collaborative relationships between equipment entities become complex, and the uncertainty problems emerge in the equipment system-of-systems. Herein, a heterogeneous network model is built to describe logistics equipment system-of-systems, which considers the heterogeneity and complex connections of different logistics equipment nodes. Next, the topological structure properties of this model are analyzed. On this basis, the experiments on the logistics equipment system-of-systems under attack strategies including degree attacks, betweenness centrality attacks and random attacks are taken to assess the changes of structural invulnerability. Results show that the logistics equipment system-of-systems heterogeneous network has similar topological structure characteristics of typical complex networks, namely small-world effect and scale-free characteristics, indicating that the flow, sharing, and synchronization between logistics equipment entities in the network are relatively easy. Meantime, the key logistics equipment nodes with large values such as degree, closeness centrality, and betweenness centrality should be protected in the logistics equipment system-of-systems heterogeneous network against deliberate attacks. The current work provides a perspective for demonstration and affords the theoretical support for development and decision-making of logistics equipment system-of-systems.
