To address the susceptibility of existing MTS(multivariate time series)anomaly detection models to training set contamination and their limited ability to capture complex spatial-temporal correlations in MTS,a novel anomaly detection model based on spatial-temporal normalizing flow was proposed.This model employed the conditional normalizing flow to estimate the density of patterns in MTS,enabling robust anomaly detection even in the presence of contaminated training data.Additionally,a patched long short-term memory module was introduced to effectively learn long-term temporal dependencies within MTS,and a dynamic graph learning module based on attention mechanisms was devised to model the evolving correlations among different dimensions of MTS.Experimental results on three real-world cyber-physical system datasets demonstrate that the proposed model significantly outperforms state-of-the-art baselines in both detection accuracy and robustness.
The electromagnetic environment serves as the fundamental medium for radio wave propagation and profoundly influences the operational efficiency of global information infrastructure systems,including communications,navigation,and security.The research focused on electromagnetic environment sensing constellations developed within the commercial spaceflight sector and summarized the evolutionary development of electromagnetic environment sensing constellations through three major phases:the military electromagnetic reconnaissance-dominated era,the stage of global collaborative competition,and the phase of commercial and national co-leadership with coordinated development.Representative constellations and their functional characteristics were outlined for each phase.Current in-orbit constellations were categorized into four primary types-propagation environment sensing,broadband radio frequency signal sensing,typical operational signal sensing,and electromagnetic environment integrated sensing-and their functional characteristics,core capabilities,key technologies,and application scenarios were analyzed.Ongoing development trends are identified in commercial space-based electromagnetic environment sensing toward space-ground system integration,multi-satellite network collaboration,multimodal fusion,and real-time artificial intelligence integration.
With the continuous expansion of electromagnetic space applications and increasing operational complexity,spectrum resources show growing dynamism,diverse signal structures,and complex environments-posing serious challenges to traditional spectrum sensing and management.This paper reviewed key limitations in current electromagnetic spectrum cognition,such as static deployment,isolated knowledge,and closed-loop cognition,which restrict adaptability,autonomous decision-making,and collaborative perception.To address these challenges,the embodied intelligent cognition for electromagnetic spectrum space was proposed.This framework integrates distributed mobile embodied agents with programmable RF front-ends to enable multimodal perception and structured understanding of complex environments.It incorporated large models and agent-based mechanisms for task-driven reasoning and decision-making,using federated and continual learning to build an evolving knowledge system,granting"electromagnetic muscle memory".Finally,coordination mechanisms and key technologies across three core modules were analyzed,emphasizing embodied intelligence's role in real-time perception,proactive regulation,and continuous optimization in complex spectrum spaces,providing theoretical foundations and engineering insights for intelligent cognition and spectrum management across multi-scenario,multi-task environments.
To address the issue of insufficient recognition accuracy caused by the continuous emergence of novel modulated signals in dynamic scenarios,an incremental recognition method for modulated signals driven by features and semantics was proposed.A multi-dimensional feature representation of modulated signals was constructed.A class-incremental knowledge distillation learning mechanism was introduced into a parallel temporal convolutional network to mitigate feature drift under multi-task iteration in dynamic environments.Meanwhile,a modulated semantic map was built based on multi-dimensional features,and a nearest neighbor strategy was adopted to classify both new and existing modulated signals.Furthermore,a joint loss function was designed by integrating distance loss,Laplacian eigenvalue optimization loss,and knowledge distillation loss,which enhances intra-class compactness and inter-class separability of different modulated signals in the semantic space,thereby improving recognition accuracy.Experimental results demonstrate that the proposed method achieves an average recognition accuracy of 84.46％across multiple incremental tasks,outperforming conventional incremental recognition methods by 10％.It effectively enhances the capability of incremental recognition of signal modulation types in dynamic scenarios.
Graph combinatorial optimization problems are challenging to solve due to their NP-hard(non-deterministic polynomial-time hard)nature,and both traditional exact algorithms and existing machine learning methods have limitations.To address this,a unified solving framework based on a pre-training and fine-tuning paradigm was proposed,aiming to enhance the model's generalization capability and efficiency across various tasks.The framework reduced different problems to a unified form,constructing a consistent representation space,learning common knowledge through cross-problem pre-training,and adapting to different test distributions via multiple fine-tuning strategies.Experiments demonstrate that the proposed method achieves superior generalization performance and stable efficiency on multiple classical tasks,providing a viable path toward a general-purpose solver.
Single-shot THz-TDS(terahertz time-domain spectroscopy)leverages spatial encoding to rapidly measure THz(terahertz)waveforms,yet its application is limited by relatively low system precision.To address the issue that a detector's detection accuracy is limited by its full-well capacity,the mechanism by which tuning the probe-light intensity suppresses the dominant noise source,namely the shot noise,was elucidated.Based on this mechanism,a high-speed,high-precision sensor dedicated to single-shot THz detection was designed.The sensor can detect higher probe-light intensities without its linear operating range,effectively reducing the relative contribution of shot noise and thereby significantly improving overall system precision.At a system repetition rate of 5 kHz,the minimum detectable terahertz field was measured to be 2.55 × 10-4 kV/cm within one second.Compared with single-point detection,a 30-fold increase in speed was achieved while maintaining the same information content.This work paves the way for further development of high-precision single-shot THz systems and rapid,complex,multi-dimensional THz spectroscopic experiments.
A bidirectional gated recurrent temporal network model was proposed,and a CGCE(classification guided cross-entropy)loss function with misclassification penalties was incorporated to reduce the 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.
Existing machine learning methods addressing MILP(mixed-integer linear programming)problems primarily focus on node features while neglecting edge features,limiting their ability to extract complete constraint information.To address this limitation,an edge-enhanced graph neural network solving framework based on Sinkhorn algorithm acceleration was proposed.This framework effectively fused node and edge representations by incorporating edge features into the attention mechanism to 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 was 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 machine learning method.
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 was 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 was obtained based on the mask,and after the inverse time-frequency transformation,the single-component signal with missing parts was 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 10 dB,the correlation coefficient between each separated signal component and the original signal is higher than 0.98.
To address the issues of visual place recognition being susceptible to environmental influences such as illumination and weather,and low recognition efficiency in practical scenarios,a visual place recognition model based on multi-criterion extraction and multi-level enhancement of scene semantic information was proposed.This model considers multiple influencing factors between pixels to extract robust semantic features.As the network propagates,it performs hierarchical aggregation of shallow features and utilizes a rapid spatial scoring mechanism for efficient retrieval of recognition results.Validation on different datasets and comparison with existing methods demonstrate that the proposed model achieves an average increase of 3.2％in recall rate R@1.Additionally,the runtime is approximately 1.3 s,and the memory consumption is about 0.21 MB,demonstrating that the model achieves superior recognition performance with reduced time overhead and memory requirements.
With the rapid expansion of NGSO(non-geostationary satellite orbit)constellation systems,spectrum resources have become increasingly congested.To address the issue of spectrum sharing between NGSO and GSO(geostationary satellite orbit)systems,an interference-avoidance beam planning method for spaceborne phased array antennas based on spatial isolation zones was proposed.By analyzing interference scenarios between NGSO and GSO systems,the Earth's surface was discretized into a grid based on latitude and longitude for uniform spatial enumeration.A time-sliced aggregate interference analysis model was then established to assess the interference from NGSO systems to segments of the GSO arc.Based on the interference-to-noise ratio threshold criterion,an optimization objective for phased array antenna beam planning was formulated to facilitate spectrum sharing between NGSO and GSO systems.Furthermore,the mapping mechanism between spatial isolation zones and beamforming strategies was elucidated,leading to the proposed method for interference-avoiding beam planning.Simulation results validate the effectiveness of the approach,offering a valuable reference for the design and implementation of LEO(low Earth orbit)satellite constellations.
Aiming at the problems of poor adaptability and low efficiency of existing transfer learning methods for cross-time domain data,a generative model transfer learning algorithm based on multi-scale feature fusion was proposed.The multi-scale depth-separable convolutional network was used to extract the layered fingerprint features in the signal frequency domain,and the channel attention mechanism was 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 framework was 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 method was 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.
To improve the efficiency of concept lattice construction,an incremental concept lattice construction method based on granular concept network was proposed.The dynamic updating mechanism of the granular concept network was investigated,and a cross-level concept fusion strategy was proposed to generate new concept nodes,so as to achieve the incremental expansion of the network structure.Based on this,the concept lattice of updated formal context was obtained from the granular concept network.Experimental results show the effectiveness of the proposed method in terms of concept acquisition task.
To address the challenges of multi-dimensional performance optimization and task adaptation for MAC(multiple access control)protocols in dynamic network environments,a software-defined reconfigurable MAC protocol architecture and algorithm was proposed.To overcome the limitation of isolated performance optimization in existing adaptive MAC protocol methods,a closed-loop mapping model that correlates network characteristics,task requirements,and protocol performance was established,thereby achieving balanced multi-dimensional performance optimization.For the problem of inaccurate network load model,a distributed observable load model was designed to provide support for the reconfiguration algorithm.Furthermore,the study implemented multi-level flexible MAC protocol reconfiguration through software-defined approaches,overcoming the limitations of hardware-coupled protocols.Simulation results demonstrate that the proposed reconfiguration mechanism balanced multiple performance dimensions and diverse task requirements,while dynamically selecting the optimal MAC protocol.This study provides a systematic solution for achieving adaptive reconfiguration of MAC protocols in wireless ad hoc networks.
To address the challenge of enhancing both accuracy and fine-grained decision-making in adversarial settings involving AI agents powered by LLMs(large language models),a novel strategy generation approach was introduced,which integrated expert policies with MCoT(multi-chain-of-thought)reasoning.By fusing real-time visual input(images)with structured observational data(text),and explicitly embedding time-segmented expert strategies and MCoT reasoning modules into the prompt design,the method provided LLMs with richer situational awareness and substantially enhanced the agent's control capability and decision-making 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 high-confrontation scenarios.
To address the difficulties of UAV(unmanned aerial vehicle)nodes in networking on the same channel due to differences in available channels,especially under conditions of intense spectrum competition or dynamic electromagnetic environments,a multi-channel UAV ad hoc networks clustering algorithm based on adaptive node degree difference was proposed.Implemented a hierarchical clustering approach optimized for modularity,the algorithm calculated node similarity by including adaptive node degree difference under multi-channel conditions.The network modularity function was maximized to cluster large-scale UAV nodes,and network throughput was analyzed using the Bianchi model.Simulation results demonstrate that the proposed algorithm not only achieves a more balanced topology but efficiently increases network throughput compared to the Fast Unfolding,JS_CNC,and HVC_MCNC algorithms.
To achieve better robustness of the deep clustering algorithm when facing complex data of different types,this paper combined a selective ensemble strategy with the deep clustering algorithm,proposing a deep clustering algorithm with fusion selective clustering ensemble.This approach effectively enhanced the robustness and clustering performance of the deep clustering algorithm.The algorithm utilized an autoencoder-based deep clustering algorithm with different initialization parameters to generate multiple diverse base clustering results.It constructed measures for ensemble similarity and diversity of base clustering.A certain number of base clustering with higher similarity and richer diversity were selected as candidates for clustering ensemble.The clustering ensemble strategy considered the reliability of clusters to construct a weighted graph consensus function.Experimental results demonstrate that the deep clustering algorithm with fusion selective clustering ensemble shows improved robustness and achieves better clustering results on various types of data compared to many existing clustering ensemble algorithms.
This study proposed a spectrum resource optimization algorithm for UAV swarms under dynamically changing communication tasks and interference environments,based on a convolutional temporal fusion network.Specifically,the study leveraged the local feature extraction capability of convolutional neural networks and the temporal modeling ability of long short-term memory networks to enhance the autonomous learning and adaptability of UAV swarms.By integrating double deep Q-network within a multi-agent framework,distributed online training was performed,enabling each UAV in the swarm to respond quickly to dynamic tasks and interference while optimizing spectrum resources based solely on local observations.Simulation results show that,in environments with dynamic communication tasks and interference,the proposed algorithm outperforms conventional methods,not only improving spectrum resource utilization efficiency but also demonstrating excellent stability.
Weapon target allocation method based on kill chain aimed to change the status quo of the traditional method that did not consider enough the process of intelligence reconnaissance,command and control as well as information transfer in the combat process,which integrated the process of intelligence reconnaissance,command and control as well as information transfer in the kill chain into the traditional method of weapon target allocation,and searched and corrected the probability-of-destruction matrix through the kill chain and combined with the Gurobi solver to realize a more accurate weapon target allocation.The experimental part used simulation data to verify the effectiveness of the method,and the results show that the Johnson-KC algorithm significantly outperforms the traditional depth first search,breadth first search algorithm,and A*algorithm in terms of the efficiency of the kill chain search,and compared with the traditional weapon target allocation method,the kill chain-based weapon target allocation method reduces the target threat by nearly 50％,and streamlines the number of the remaining edges of the kill net by nearly 80％.
The digital twin system for solid rocket motor is conducive to breaking through the traditional design and management model of solid rocket motor that based on experience and semi-experience for a long time.Digital twin system can reduce the difficulty of design,development,maintenance,and meet the development needs of rapid iteration,digitization,and intelligence of solid rocket motor.The development process and key technologies of the solid rocket motor digital twin system were systematically explored in this paper.Starting from the concept of digital twins,the development process of digital twins was divided into four stages:physical twin,computer aided design/simulation,virtual reflection of reality,and combination of virtual and reality.The solid rocket motor digital twin system framework was established at each stage,and key technologies of each stage were summarized.It was hoped that the above content can serve as a technical roadmap,driving the development and accumulation of related disciplines,and providing theoretical reference for the research and application of solid rocket motor digital twin systems.
