Abstract:

Electromagnetic spectrum resources are strategically scarce national resources， and the electromagnetic spectrum space has become the sixth-dimensional operational domain. Against the backdrop of the development of the sixth-generation mobile communication （6G） and the acceleration of the low-altitude economy， the scale of frequency-using equipment continues to expand， and the demand for spectrum use is becoming increasingly complex and diverse. The spectrum environment exhibits highly dynamic and time-varying characteristics， leading to the intensification of spectrum resource scarcity， the increasingly prominent spectrum security threat， and more intense spectrum confrontation situations. Traditional spectrum management systems based on passive sensing face significant bottlenecks in terms of sensing accuracy， response timeliness， and collaborative capabilities. Electromagnetic spectrum cognitive intelligent management and control， driven by emerging technologies such as large spectrum models， relies on multimodal spectrum sensing， multi-level spectrum cognition， and multi-factor spectrum decision-making to establish an integrated “Observe-Cognition-Decide-Act” cognitive decision-making loop. It promotes the evolution of spectrum management from static planning and passive response to dynamic allocation and proactive regulation， gradually moving toward higher levels of intelligence and automation， to support spectrum management needs in future complex electromagnetic environments. This paper systematically reviews the technical system architecture， the cognitive decision-making loop mechanism， and key technologies of electromagnetic spectrum cognitive intelligence， and offers an outlook on future trends. It aims to provide theoretical reference and technical guidance for addressing the challenges of efficient spectrum management in complex electromagnetic environments and for constructing the next-generation， fully autonomous cognitive intelligent spectrum management system.Highlights： 1. This paper innovatively proposes an integrated cognitive decision-making closed loop driven by a cognitive intelligence engine， promoting the paradigm shift of spectrum management from static planning and passive response to dynamic allocation and proactive control. 2. This paper systematically constructs the cognitive intelligence technology system for the electromagnetic spectrum and the evolution path of next-generation full-domain autonomous control systems， providing theoretical references and technical guidance for addressing the challenges of efficient spectrum management in complex electromagnetic environments.

Abstract:

Intelligent signal recognition technologies can effectively enhance the performance of individual unmanned aerial vehicle （UAV） identification. However， their practical deployment is significantly constrained by time-varying channel effects and feature distribution drift across domains. With the rapid development of the low-altitude economy， UAV safety supervision urgently demands reliable identity authentication mechanisms. Radio frequency fingerprint （RFF） identification， while inherently difficult to forge owing to hardware uniqueness， suffers from severe feature drift in time-varying channels typical of low-altitude intelligent networks. This drift leads to the catastrophic degradation of generalization in pre-trained models. To address this issue， we propose a broad learning-driven method for cross-time-domain incremental identification of individual UAVs. This method adopts a residual network integrated with multi-scale asymmetric convolutions as the backbone， aiming to extract robust and multi-granularity fingerprint features directly from IQ signals. A broad learning system is subsequently introduced as an incrementally updatable classifier； it rapidly updates model weights for new time-domain data by leveraging the generalized inverse matrix， thereby circumventing catastrophic forgetting. Furthermore， a learnable feature fusion module and an experience replay mechanism are synergistically designed to suppress feature drift across time domains. Extensive experiments are conducted on real-world UAV RF signal datasets collected over multiple time spans， with intervals ranging from days to weeks. The results demonstrate that the proposed method achieves an identification accuracy exceeding 90% on both the source and cross-time domains， outperforming baseline algorithms by over 20%. Meanwhile， it maintains stable recognition performance on data from earlier time periods. The proposed approach effectively mitigates the adverse effects of time-varying domain shift， offering reliable technical support for continuous UAV identity recognition and the detection of unauthorized UAVs in complex environments.

Abstract:

Specific emitter identification （SEI） leverages the inherent hardware imperfections of wireless device RF front-ends to achieve device identification， serving as a crucial technology for ensuring wireless communication security. However， in complex electromagnetic environments， various perturbations such as carrier frequency offset， sampling clock deviation， gain fluctuation， time shift， and chirp modulation collectively cause distribution shifts in signal features， thereby weakening the stability of RF fingerprint representation and degrading identification performance. To address these issues， this paper proposes a perturbation-aware modulated convolutional neural network （PAM-CNN）. This method first utilizes a perturbation-aware branch to jointly estimate the perturbation state and its parameters in the input signal， and then performs sample-adaptive modulation of the convolution kernels based on the estimation results. This enables the network to structurally suppress the impact of perturbations during the feature extraction process. Simultaneously， a multi-task joint training framework is constructed， incorporating device identification， perturbation detection， and parameter regression， to enhance the model’s robust representation capability under complex perturbation conditions. Experimental results on a real over-the-air ADS-B baseband dataset and its offline perturbation-augmented data demonstrate that， under various superimposed perturbation conditions， the proposed method achieves an identification accuracy of 95.39% at a signal-to-noise ratio （SNR） of 15 dB and outperforms comparison methods across the full SNR range. The results indicate that this method can effectively enhance the robustness of SEI in complex electromagnetic environments.Highlights: 1.This paper proposes a perturbation-aware modulated convolutional neural network for robust specific emitter identification, where five typical perturbations, including carrier frequency offset, sampling clock deviation, gain fluctuation, time shift, and chirp modulation, are explicitly modeled and used to guide sample-adaptive convolution kernel modulation.2.This paper builds a multi-task learning framework that jointly optimizes emitter classification, perturbation detection, and parameter regression. Experiments on real over-the-air ADS-B baseband data and perturbation-augmented data demonstrate that the proposed method improves recognition robustness under multiple perturbations and different SNR conditions.

Abstract:

This study aims to address the critical challenges of energy diffusion， resource conflicts， and high-dimensional action spaces inherent in multi-jammer cooperative jamming within complex electromagnetic environments. Conventional omnidirectional jamming suffers from severe energy inefficiency， while independent decision-making among multiple jammers frequently results in interference overlap. Furthermore， the joint optimization of beam direction， jamming channel， and transmit power creates an exponentially growing action space that traditional reinforcement learning methods struggle to handle. To overcome these limitations， we propose a collaborative decision-making framework based on deep reinforcement learning to achieve three-dimensional joint resource optimization with minimal communication overhead. The proposed method constructs a multi-agent architecture featuring “centralized training with decentralized execution”（CTDE）， where each jammer utilizes an independent deep Q-network to approximate action-value functions based on local observations. Centralized training is achieved through a shared global reward signal defined as the total number of successfully jammed users， aligning individual policies with system-wide objectives without high-bandwidth data exchange. To mitigate Q-value overestimation， double target networks with soft parameter updating are integrated. An adaptive Boltzmann exploration strategy with exponentially decaying temperature is employed to dynamically balance the exploration and the exploitation. The action space is formulated as a three-dimensional joint space integrating beam direction， frequency channel， and power level assignment. Comprehensive simulations conducted in a 400 m×400 m scenario with four communication user pairs and two intelligent jammers demonstrate the effectiveness of the proposed approach. Quantitative results indicate that the jamming success rate reaches approximately 90%， representing a 50% improvement over independent deep reinforcement learning and an 80% improvement over independent Q-learning. This approach effectively resolves resource conflicts in multi-jammer systems through global reward sharing while ensuring low communication overhead. The integration of double target networks and adaptive Boltzmann exploration successfully addresses training instability in high-dimensional spaces. By achieving joint optimization of spatial， spectral， and power resources， the method significantly enhances energy utilization efficiency， providing a robust technical foundation for intelligent electronic countermeasures.Highlights:1. A novel “distributed execution with centralized optimization” multi-agent architecture is proposed to achieve collaborative jamming with minimal communication overhead and exposure to risk.2. An improved deep Q-network algorithm integrating double target networks and adaptive Boltzmann exploration is designed to address Q-value overestimation and balance exploration-exploitation trade-offs.3. A three-dimensional joint optimization framework for beam direction， jamming channel， and transmit power is proposed， and simulation results validate that the proposed method achieves approximately 90% jamming success rate， outperforming independent learning.

Abstract:

This paper proposes an adaptive rotation-based enhancement method for direction-of-arrival （DoA） estimation of co-frequency multiple emitters using a uniform linear array （ULA）. Addressing the limited estimation accuracy of traditional fixed arrays in multi-source scenarios， a dynamic rotation mechanism is introduced to establish an analytical relationship between the rotation angle and the Cramer-Rao bound （CRB）， leading to the derivation of a closed-form solution for the optimal rotation angle. The trace of the CRB is employed as the loss function， representing the theoretically achievable lower bound of the mean squared error. By minimizing the CRB， the system observability is enhanced， thereby improving the estimation accuracy of DoA algorithms. Given that the CRB in multi-source scenarios is a correlated matrix， a simplified diagonal form of the CRB is derived under the condition of a sufficiently large number of snapshots. This simplified diagonal CRB is utilized for optimization， significantly reducing computational complexity. To achieve high-precision DoA estimation， prior information about the signal incident angles is first obtained using a fast algorithm. This prior is derived from covariance matrix computations after the array receives and samples signals from multiple radiation sources. Subsequently， the rotation angle that minimizes the loss function is determined and adopted as the optimal array orientation. After adjusting the array to this angle， the DoA estimation algorithm is reapplied at the same position to obtain the final DoA results. The proposed method maintains high estimation accuracy under multi-source conditions and effectively enhances overall direction estimation performance. Simulation results demonstrate that the method significantly reduces the CRB， improves localization accuracy， and enhances the resolution of spectral search-based algorithms.

Abstract:

The electromagnetic spectrum is a core strategic resource of modern information systems. As the physical-layer foundation of spectrum cognition， a high-performance wideband RF receiving platform directly bounds the performance ceiling of subsequent sensing and cognition algorithms through its front-end hardware performance. This paper presents a highly integrated software-defined radio platform for spectrum cognition applications based on radio frequency system-on-chip （RFSoC） technology， integrating Gsps-class high-speed AD/DA converters， FPGA programmable logic， multi-core ARM processors， DDR4 high-capacity memory， and diversified high-speed peripheral interfaces onto a single 150 mm×100 mm board. A systematic RF receiving performance evaluation is conducted via a direct signal-source connection method， covering key dimensions including basic receiving parameters， ADC core dynamic performance， and dynamic range with sensitivity. Experimental results demonstrate that the platform supports a maximum receiving frequency of 4 GHz， maintains an effective number of bits （ENOB） of 7.97—8.32 bits， achieves a noise spectral density better than -151 dBFS/Hz， and delivers a system dynamic range exceeding 71 dB. With its comprehensive advantages in high integration density， wide instantaneous bandwidth， and diversified high-speed interfaces， the platform offers an effective hardware foundation for spectrum cognition intelligence.

Abstract:

Electromagnetic （EM） battlefield has become increasingly complex due to the proliferation of heterogeneous communication systems， diverse radar waveforms， and a wide array of data link protocols. Accurate and real-time spectrum situation awareness critically depends on the effective extraction of discriminative features from multi-source， multi-modal EM signals and their fusion into consistent， high-level representations—enabling robust classification and radiation source identification. To address these challenges， this paper proposes a comprehensive recognition framework integrating spectral feature parameter extraction， modulation recognition， protocol identification， and multi-source heterogeneous data fusion. The framework achieves high-fidelity signal characterization under low signal-to-noise ratio （SNR） conditions. First， a hierarchical modulation recognition method is developed based on envelope characteristics， spectral symmetry， and spectral peak count， enabling reliable discrimination among five representative signal types—SSB， FM， FSK， MSK， and AM—as well as TACAN signals. Second， domain-specific communication system features are extracted to construct a data link recognition model with enhanced interpretability and generalization. Third， to handle multidimensional spectral feature fusion， a signal preprocessing pipeline and a dimensionality-reduction fusion model are designed to preserve salient information while reducing redundancy. Furthermore， transfer learning and few-shot learning strategies are integrated to mitigate performance degradation under limited and imbalanced training samples for novel radiation sources. Extensive simulations demonstrate that the proposed framework maintains high recognition accuracy across diverse SNR levels and exhibits strong robustness and generalization capability， effectively overcoming the challenges of low-data regimes and class imbalance.

Abstract:

The effect of head movement on the performance of active noise control headrest systems in reverberant sound fields is investigated in this work. Based on the rigid-sphere scattering model， the performance of active noise control headrests in reverberant fields is analyzed. A prediction formula for noise reduction degradation due to head movement is proposed through numerical simulations， and the accuracy requirements for ear positioning systems applied to active noise control headrests are discussed. Comparisons are made with the case where the primary noise is a plane wave. Simulation results indicate that higher frequencies， greater initial noise reduction levels at the head position， and larger movement distances lead to more significant degradation in noise reduction. For 125 Hz， 250 Hz， and 500 Hz， with an initial noise reduction of 20 dB， the allowable head movement ranges to ensure noise reduction degradation not exceeding 3 dB are 3.1 cm×4.1 cm×1.4 cm， 2.9 cm×2.5 cm×1.2 cm， and 1.4 cm×1.2 cm×1.0 cm， respectively. These ranges are smaller than those for plane wave incidence， which are 3.4 cm×9.1 cm×1.7 cm， 3.1 cm×4.8 cm×1.5 cm， and 1.9 cm×2.3 cm×1.1 cm， respectively. Finally， experiments conducted in a reverberation chamber to validate the simulation results. The findings in this study provide guidance for the design of active noise control headrests applied in reverberant acoustic environments such as aircraft cabins and train compartments.

Abstract:

Low Light Image Enhancement， which is to restore the image acquired under the condition of insufficient light to the normal exposure image. Most of the existing low-light image enhancement algorithms obtain good enhancement effect by designing complex network structure， and the computational efficiency is low. The enhanced image will still have problems such as increased noise， color distortion and detail loss， which will affect visual perception and subsequent advanced visual tasks. Therefore， a lightweight low-light image enhancement method based on multi-attention feature fusion is proposed in this paper. Simple gate attention module is used to extract the global features of low-light images effectively， and the computational overhead is reduced and image details are preserved by simplifying the channel attention and gating unit. The multi-attention fusion module is used to integrate the information of global features and local features extracted from local receiving fields， and enhance the representation of channel attention and spatial attention to global and local features through pixel attention， so as to better restore image color and suppress noise. In addition， the joint loss function is used to constrain the enhancement task， and extensive experiments on real data sets show that the performance of the proposed method exceeds the current advanced low-light image enhancement methods， and has good computational efficiency and generalization ability.

Abstract:

Adaptive model fusion is particularly important for dynamically responding to the evolutionary characteristics of data and tasks. However， existing model fusion methods still have issues such as static weights being difficult to adapt to data similarity， dynamic fusion being driven by single factors， and being susceptible to data distribution drift. To address these shortcomings， this paper proposes an adaptive model fusion method driven by data similarity and model reliability. The method captures the similarity between samples through feature semantic alignment to obtain a similarity matrix， and further obtains the sample matching degree coefficient. Then， based on the base model selection algorithm of performance-diversity， the generalization ability and local performance of the base models are evaluated through multi-dimensional metrics to obtain the reliability coefficient of the base models. Finally， the fusion weight is calculated based on the data similarity coefficient and the reliability coefficient of the base models to obtain the final fusion model strategy. Experimental results on public datasets demonstrate the effectiveness of the proposed method.Highlights:1.Propose an adaptive model fusion method driven by data similarity and model reliability. By incorporating data distribution characteristics and model reliability into model fusion， the method maximizes the prediction performance of the model and realizes adaptive model fusion.2.Propose a precise mixed data similarity measurement module.It achieves semantic alignment of numerical and categorical heterogeneous features through deep embedding， integrates improved K-Prototypes clustering to output sample-cluster similarity vectors， and underpins sample-level local dynamic adaptation.3.Design a performance-diversity dual-goal optimization-based base model selection mechanism， leveraging multi-dimensional evaluation， diversity quantification， and dynamic decaying weights to automatically prune redundant models， adjust reliability weights， and boost fusion robustness.4.Propose an adaptive model fusion framework that does not rely on scenario-specific prior distribution assumptions. It can flexibly adapt to data drift and heterogeneous model fusion requirements in different fields， providing an adaptive fusion solution for complex tasks.

Abstract:

As an unsupervised learning method， although random forest clustering demonstrates strong robustness in processing high-dimensional and complex data， it still faces challenges such as weak discriminability of original data caused by the introduction of negative samples and the interference of noisy decision trees on clustering performance. To address these issues， this paper proposes a random forest clustering based on valley detection and three-way ensemble selection （VDTES-RFC） method. First， the valley detection technology is utilized to identify potential split points for generating training data， and the Gini index is calculated to determine the optimal split points to complete the training of the classification forest. Second， each decision tree is treated as a base clusterer to extract its similarity matrix， and a three-way ensemble selection strategy is adopted to select high-quality decision trees to construct a new forest. Finally， a consensus function is used to integrate the similarity matrices to obtain the final clustering result. Experimental results demonstrate that this method effectively improves clustering accuracy and robustness， achieving dual optimization of efficiency and performance.Highlights:1. The paper proposes a valley detection and three-way ensemble selection-based random forest clustering （VDTES-RFC） method to overcome original data discriminability loss and noisy decision tree interference.2. The paper develops a dual-stage clustering scheme centered on potential split point optimization and dynamic tree filtering. It aligns Gini index-based data partitioning with similarity matrix extraction to ensure high-quality base clusterers.3. The paper adopts a three-way ensemble selection strategy combined with a consensus function to filter high-quality decision trees. It achieves a dual optimization of clustering efficiency and performance.

Abstract:

In text classification tasks， effectively extracting text features while improving computational efficiency is a critical challenge. However， traditional methods often struggle to balance feature richness and computational efficiency. To address this issue， this paper proposes a novel text classification model， i.e.， the linear attention text classification by combining text features and word frequency implicit factors （LTTW）， which introduces a linear attention mechanism to capture key features in the text. Specifically， the model leverages non-negative matrix factorization （NMF） to extract word frequency implicit factors from the term frequency matrix， capturing latent semantic information. Simultaneously， it utilizes pre-trained models to extract semantic features of the text， which are then fused with the word frequency implicit factors to construct a richer text representation. Based on this representation， the linear attention mechanism is applied to effectively capture global dependencies and enhance the processing efficiency of long text sequences. Experiments conducted on public datasets demonstrate that the proposed model outperforms mainstream methods in terms of both accuracy and computational efficiency， with particularly significant efficiency advantages when handling long sequences. The study highlights that the integration of word frequency implicit factors complements the shortcomings of pre-trained models in semantic feature extraction， while the linear attention mechanism effectively captures key textual features and improves sequence processing efficiency. Together， these contributions significantly enhance the performance and efficiency of text classification.

Abstract:

To address the issues of parameter inflation and soaring computational complexity in mainstream speech enhancement models， a lightweight speech enhancement network based on gated hybrid dilated convolution is proposed in this paper. Firstly， a gated hybrid dilated convolution module is designed， which integrates gated linear units with hybrid dilated convolution to achieve multiscale feature extraction of speech signals and precise suppression of noise-sensitive regions， thereby effectively preserving both long-term and short-term speech characteristics while enhancing model robustness. Secondly， a hierarchical channel attention module is proposed to enhance the capture of speech feature correlations in channel dimensions through hierarchical feature fusion， while maintaining low parameter complexity. Experimental results on the VoiceBank+DEMAND dataset demonstrate that the proposed model， with only 0.41 million parameters， achieves competitive performance on the perceptual evaluation of speech quality （PESQ）， the short-time objective intelligibility （STOI）， cepstral signal-to-noise ratio （CSIG）， cepstral background noise（CBAK） and cepstral overall loudness （COVL）， thus achieving an organic integration of model lightweighting and high-precision performance.Highlights：1. Propose a lightweight speech enhancement network with gated hybrid dilated convolution.2. Integrate multiscale feature extraction， channel attention， and Ghost convolution for efficient feature modeling.3. Achieve a good balance between enhancement performance and model complexity on VoiceBank+DEMAND.

Abstract:

Traditional attribute reduction methods construct a unified attribute reduction set for decision classification， ignoring the differentiated representation of attributes among various decision classes， which often results in the pivotal attributes of specific classes being redundantly covered and leads to suboptimal classification accuracy for those specific classes. To address these issues， this paper proposes a class-specific attribute reduction method driven by conditional entropy based on fuzzy rough sets， considering the advantages of fuzzy rough sets in handling widely existing numerical and fuzzy data. Firstly， by integrating the decision inclusion degree of fuzzy rough sets with information entropy theory， a class-specific conditional entropy is defined to quantify the local discriminative power of conditional attributes with respect to the target class. Secondly， the paper presents a class-specific attribute reduction condition based on conditional entropy and defines both internal and external attribute significance measures based on this class-specific conditional entropy. Furthermore， forward （FA-CE） and backward （BA-CE） attribute reduction algorithms are proposed based on attribute significance. Finally， the class-specific attribute reduction is conducted on seven UCI datasets and two feature selection datasets， and comparative analyses are performed against methods based on neighborhood conditional entropy， mutual information， neighborhood rough sets， and a conventional dependency-based reduction approach. The classification accuracy and F₁-score of the proposed method are evaluated using support vector machine（SVM）， K-nearest neighbor（KNN） and classification and regression tree（CART） classifiers， demonstrating the rationality and effectiveness of the proposed class-specific attribute reduction approach.

Abstract:

In view of the current problems that facial expressions are difficult to recognize under conditions of lighting changes and occlusion， as well as the low recognition rate of pessimistic emotions， this paper proposes a facial expression recognition algorithm based on graph convolutional cascade classification based on improved dense connection network and fusion of facial key point features. Since different deep learning models have their own advantages in facial expression recognition， dense connection network has a high accuracy rate in recognizing optimistic and calm expressions， but has a weak recognition effect on pessimistic expressions. Therefore， this paper first uses wavelet transform， key part mask attention mechanism and binary tree classifier to improve the dense connection network， and constructs the I-Densenet （Improved-DenseNet） module for the rough division of optimistic， calm and pessimistic facial expressions to improve the recognition rate of rough division； Secondly， the graph convolutional neural network based on the fusion of facial key point features is used to fine-grainedly divide the pessimistic expression of the face to improve the recognition rate of pessimistic expression. Finally， this paper constructs the D-GFK network （DenseNet-GCN and face key point network） by cascading the improved dense connection network with the graph convolutional neural network based on key point feature fusion， combining the advantages of different models to comprehensively improve the accuracy of facial expression recognition. Experiments show that the model proposed in this paper has achieved good recognition results in facial expression recognition tasks.

Abstract:

The rapid growth of social media has enabled rumors to spread swiftly through extensive online interactions， thereby significantly undermining public trust and destabilizing social order. However， existing rumor detection methods face notable limitations in modeling the dynamic semantic evolution of text and accurately capturing complex propagation patterns， and they often struggle to distinguish between ambiguous rumor categories. To address these challenges， we propose DySGCL （Dynamic semantic and graph feature fusion for contrastive rumor detection）， a novel contrastive learning framework that fuses dynamic semantic representations with graph-based structural features. Specifically， we employ a hierarchical Transformer to extract global semantic embeddings from users’ past posts， while a temporal convolutional module improves sensitivity to fine-grained semantic shifts. For structural modeling， we first simulate adversarial perturbations via edge removal， then leverage a graph attention network （GAT） to highlight critical interaction pathways in the propagation network. Finally， an integrated contrastive objective combining self-supervised and supervised signals further enhances the model’s discriminative power. Experiments on the Twitter15 and Twitter16 benchmarks show that DySGCL outperforms state-of-the-art baselines by 1.8% and 2.0% in accuracy， respectively， validating its effectiveness in dynamic and complex rumor detection scenarios.Highlights:1. A dynamic semantic and graph-structure fusion framework is proposed for rumor detection.2. A hierarchical Transformer and temporal convolution are integrated to capture semantic evolution in users’ historical posts.3. Self-supervised and supervised contrastive learning mechanisms are jointly used to improve the discrimination of ambiguous rumor categories.

Abstract:

Sepsis refers to a systemic inflammatory response resulting from infections， and it carries a high risk of mortality in intensive care settings. Existing predictive models often rely on extracting single feature subsets from a larger set， failing to fully utilize the complex interactions between feature subsets， known as structural mutual information. This limitation reduces prediction accuracy. Structural mutual information not only captures dependencies between features at the same level of granularity but also reveals complex relationships across different granularities， enabling more precise detection of subtle changes in a patient’s condition. To address this limitation， this study presents a novel sepsis prognosis model that deeply explores the structural mutual information within electronic health records， significantly enhancing the accuracy of mortality risk predictions. Experimental results show that the proposed model achieves notable improvements in predictive accuracy， providing clinicians with more dependable mortality risk assessments and clearer decision-making support.

Abstract:

To address the heavy dependence on labeled data in practical radio environments， which limits reliable recognition of previously unseen （unknown） modulation types， this paper proposes an unsupervised modulation recognition method for communication signals based on bootstrap your own latent （BYOL） self-supervised representation learning and a contrastive clustering mechanism. Conventional modulation recognition algorithms are predominantly supervised and require large amounts of labeled data， which is often costly or even infeasible to obtain in real-world scenarios. Unsupervised and self-supervised approaches can alleviate this issue， but existing methods typically suffer from insufficient representation learning capacity and suboptimal clustering performance， and thus struggle to cope with complex channel conditions. The proposed method does not rely on any manual labels. First， we employ a BYOL framework with a dual-branch architecture to encode different sub-segments of the same signal， thereby learning intrinsic and stable representations in a self-supervised manner. Second， instance-level and cluster-level contrastive learning modules are introduced： The former enhances feature consistency across different augmented views of the same signal， while the latter improves the separability of different modulation types in the feature space， thereby enabling high-accuracy blind clustering of unknown modulation types. Experiments conducted on the public RadioML2018.01A dataset show that the proposed method outperforms existing algorithms by more than 10% in various clustering evaluation metrics. Ablation studies further confirm the critical roles of the BYOL module and the contrastive clustering mechanism in improving overall performance. Confusion-matrix analysis demonstrates that， at 10 dB， the proposed method achieves near-ideal recognition accuracy for typical modulation types such as amplitude modulation double-sideband with carrier（AM-DSB-WC），frequency modulation（FM）， and Gaussian minimum shift keying（GMSK）， and also exhibits strong robustness and anti-confusion capability for other more challenging modulation types. In summary， the proposed unsupervised modulation recognition method effectively alleviates the problem of label scarcity in real wireless communication scenarios and shows strong potential for practical deployment.

Abstract:

To address the limited energy supply and insufficient transmission performance of heterogeneous bit and semantic communication networks （HBSCNs）， this paper investigates a movable antenna enabled wireless powered HBSCN （WP-HBSCN）. In the considered network， a hybrid access point （HAP） equipped with multiple movable antennas （MAs） first broadcasts radio-frequency energy signals to bit users and semantic users. Then， the users transmit bit and semantic information to the HAP by time-division multiple access using the harvested energy. By dynamically adjusting the positions of the MAs， additional spatial degrees of freedom are exploited to construct favorable channels for both downlink wireless energy transfer and uplink information transmission. To enhance the bit transmission performance while guaranteeing the quality-of-service requirements of semantic users， the total number of bit data maximization problem is formulated by jointly optimizing the energy beamforming vector， user transmit power， time-slot allocation， and MA positions. The formulated problem is challenging to solve directly because of the coupled optimization variables， nonlinear energy harvesting model， and non-convex antenna position constraints. To address this difficulty， an alternating optimization algorithm is developed based on the block coordinate descent framework. Specifically， the energy beamforming and power allocation subproblem is handled by successive convex approximation， the time-slot allocation subproblem is solved through convex optimization， and the MA position optimization subproblem is addressed using particle swarm optimization. Simulation results verify the convergence of the proposed algorithm and show that the MA-enabled scheme achieves a higher total number of bit data than the benchmark schemes， including the sparrow search-based scheme， equal time-slot allocation scheme， random beamforming scheme， and fixed-position antenna scheme. These results demonstrate the effectiveness of introducing MAs into the WP-HBSCN.