Abstract:To enhance the sustained mission capability of large-scale UAV swarms facing both soft and hard-kill threats, this research focuses on swarm topology reconstruction methods. Addressing the issue of node failures, a dynamic control authority migration mechanism under a centralized-distributed hybrid control architecture is proposed to improve reconstruction efficiency; concurrently, an SDN-based elastic networking architecture is designed, integrating intent-driven principles to enable intelligent and dynamic configuration of network resources. Simulation comparisons were conducted between dual-mode reconstruction strategies: neighbor autonomous compensation and resource-pool dynamic scheduling. When facing small-scale node loss, neighbor compensation leverages its advantage in local decision-making, reducing the average reconstruction latency by 38.5% compared to resource-pool scheduling. However, as the number of failed nodes increases, the combined centralized-distributed controller strategy achieves the shortest reconstruction time; notably, under persistent electromagnetic interference environments, this combined strategy demonstrates significant advantages, reducing latency by 21.7% compared to a purely distributed approach. This research provides theoretical support and practical reference for topology reconstruction in large-scale UAV swarms.

Abstract:Aiming at the problems of poor adaptability and low efficiency of existing transfer learning methods to cross-time domain data, this paper proposes a generation model transfer learning algorithm based on multi-scale feature fusion. The multi-scale depth-separable convolutional network is used to extract the layered fingerprint features in the signal frequency domain, and the channel attention mechanism is combined with the adaptive focusing of the key components of the inherent distortion of the hardware to enhance the discrimination of key fingerprint features. At the same time, a Bi-directional Generative Adversarial Networks (Bi-GAN) framework is used to realize the potential spatial alignment of the features of the source domain and the target domain using bi-directional mapping constraints. The Maximum Mean Discrepancy (MMD) method is used to help align the feature distribution of source domain and target domain. Based on the real acquisition radar data set verification, the recognition accuracy can reach about 90%, and the time complexity is low, which can adapt to the requirements of practical application scenarios.

Abstract:Aiming at the problems that current signal separation methods usually require a known number of signal components and have poor separation performance in the case of severe aliasing such as crossover in the time-frequency domain, an intelligent signal separation method based on diffusion generation is proposed. Firstly, perform semantic segmentation on the time-frequency graph of the aliased signal to obtain each signal region corresponding to the non-overlapping parts of time and frequency, and form a signal mask. Furthermore, the time-frequency graph of the single-component signal is obtained based on the mask, and after the inverse time-frequency transformation, the single-component signal with missing parts is obtained. Finally, taking this as a condition, the improved latent diffusion model was concatenated with noise. The improved model achieved the reconstruction of each signal component by removing the training module of latent variables, improving the network parameters, and designing the loss function. The proposed method does not require the known number of signal components. Experimental results show that it can adapt to three FM signal aliasing scenarios. When there is severe overlap in the time-frequency domain and the signal-to-noise ratio is 10dB, the correlation coefficient between each separated signal component and the original signal is higher than 0.98.

Abstract:A bidirectional gated recurrent temporal network model was proposed, and a CGCE (classification guided cross-entropy) loss function was incorporated to mitigate misclassification of high-risk intentions. The bidirectional gated recurrent temporal network model was composed of a BiTCN (bidirectional temporal convolutional network), a BiGRU (bidirectional gated recurrent unit), and an attention module. The BiTCN was used to capture global features, the BiGRU enhanced understanding of complex data through bidirectional learning, and the attention module dynamically assigns feature weights to emphasize key information. The introduction of CGCE enhanced the model's sensitivity to high-risk misjudgments, thereby reducing the misjudgment of high-threat intentions. The proposed model achieves an accuracy of 98.58% and outperforms the comparison methods across accuracy, precision, and F1-score. Furthermore, the incorporation of CGCE further improves the model’s accuracy and significantly reduces misclassifications of high-threat intentions, validating the effectiveness of the proposed model and CGCE in aerial and maritime target intention recognition.

Abstract:To address the challenge of enhancing both accuracy and fine-grained decision-making in adversarial settings involving AI agents powered by large language models (LLMs), this study introduces a novel strategy generation approach that integrates expert policies with multi-chain-of-thought (CoT) reasoning. By fusing real-time visual input (images) with structured observational data (text), the method provides LLMs with richer situational awareness. Time-sensitive expert strategies and parallel reasoning chains are embedded into prompt design, significantly improving the agent’s control and tactical precision. Experimental validation in high-difficulty StarCraft II scenarios demonstrates that the method achieves a 95% win rate without any additional model training. Results indicate that the approach enables interpretable and fine-tuned decision outputs in highly dynamic adversarial environments, offering a compelling pathway for leveraging LLMs in strategic behavior generation under competitive conditions.

Abstract:To address the limitations of LLMs (large language models) in reasoning and decision-making within complex dynamic game scenarios, the DCoRD (dynamic-chain-of-reasoning-and-decision) method was proposed based on the limitations of text generation mechanism and reasoning of large language models. The DCoRD consisted of a reasoning-decision framework and a dynamic decision option library, serving as structured prompt engineering to enhance the reasoning and decision-making abilities of LLMs. By incorporating task objectives to constrain output formats and content scope, the method reduced model hallucinations and improves decision accuracy. Four approaches were compared: free-generation mode, traditional chain-of-thought, chain-of-draft and the proposed DCoRD method, in a StarCraft II environment. Experimental results demonstrate that DCoRD significantly reduces token consumption and response latency while enhancing decision accuracy and task alignment, offering novel theoretical and methodological insights for applying LLMs to game-theoretic decision tasks.

Abstract:To improve the efficiency of constructing concept lattices, we propose an incremental concept lattice construction method based on granular concept network in this paper. The proposed method explores the update mechanism of granular concept network in the formal context where attributes are constantly increasing. Specifically, the updating mechanism of concepts in each layer of the granular concept network was first explored, and then new concept nodes were generated through cross level concept fusion strategy to achieve incremental expansion of the network structure. Furthermore, the concept lattice of updated formal context was generated from the granular concept network. Finally, the effectiveness of the method proposed in this paper was verified through numerical experiments in terms of concept acquisition task.

Abstract:Recent advancements in Artificial Intelligence of Things (AIoT) technologies have brought about an increasing popularity in leveraging deep learning algorithms to detect potential failures in cyber-physical systems (CPS). Typically, an anomaly detection model is deployed to monitor the multivariate time series (MTS) generated by sensors to identify abnormal operation states. However, the contemporary unsupervised deep learning models for MTS anomaly detection are susceptible to contamination in the training dataset and are incapable of capturing the spatial-temporal correlations in MTS, result in suboptimal practical detection performance. In this paper, we propose a novel framework called Spatial-Temporal Normalizing Flow (STNF) to tackle the above problems. Our framework extends the conditional normalizing flow for MTS density estimation, aiming to achieve robust anomaly detection against training dataset pollution. Additionally, we introduce a patched Long Short-Term Memory (LSTM) module to effectively learn robust representations of long-term dependencies within MTS. Moreover, a dynamic graph construction module is devised to model the complex and evolving correlations among different dimensions of MTS. We evaluate our approach on three real-world CPS datasets and report improvements over the state-of-the-art approaches in terms of both performance and robustness.

Abstract:Mixed-Integer Linear Programming (MILP) is a key technology for solving a wide range of real-world combinatorial optimization problems. However, existing machine learning methods addressing these problems primarily focus on node features while neglecting edge features, limiting their ability to extract complete constraint information. To address this limitation, this paper proposes a novel solution framework named SHARP based on an edge-enhanced Graph Neural Network accelerated by the Sinkhorn algorithm. This framework effectively fuses node and edge representations by incorporating edge features into the attention mechanism to better learn the underlying patterns of MILP. Furthermore, to overcome the shortcomings of traditional methods in problem scale generalization and hyperparameter tuning, an adaptive Regret-Greedy algorithm is designed to enhance solution feasibility and quality by dynamically adjusting variable assignment strategies. Experimental results on combinatorial auction and item placement datasets show that the proposed framework achieves performance improvements of 24.88% and 5.86% on the Primal Integral metric compared to the Gurobi and SCIP solvers, respectively, and a 17.19% improvement over the current state-of-the-art ML method.

Abstract:To enhance the survivability of unmanned combat aerial vehicles in multi-stage air combat missions,, a data-knowledge driven intelligent avoidance decision-making method was proposed. The method employed Markov decision process to formally model the avoidance decision-making process and introduced the self-attention mechanism of blocked interaction with self-enemy situation based on reinforcement learning. And a data-knowledge driven policy update approach was adopted to train the agent. The avoidance performance and air combat effectiveness of the proposed model were evaluated based on the simulation platform. It was proved that the proposed model has significant theoretical significance and reference value for improving the survivability of unmanned combat aerial vehicles, by comparing it with the traditional reinforcement learning deep Q-network and rule-based maeuvering model.