Abstract:Resting state functional magnetic resonance imaging (rs-fMRI) technology is widely used for predicting and diagnosing brain diseases. Researchers have proposed a large number of diagnostic methods based on rs-fMRI functional connectivity networks, but these methods have problems such as not fully considering the complex dependency relationships of brain network data and high computational costs. To address the aforementioned issues, this paper proposes a brain disease diagnosis method based on the bidirectional selective state space model (SSM) framework, Bi-BrainSSM。This method designs an attentive bi-Mamba (ABM) module that combines multi head attention mechanism with a bidirectional selective state space model to enhance the representational ability of brain functional connectivity networks and improve the computational efficiency of the model. Simultaneously introducing orthogonal clustering readout to solve the problem of information loss in traditional aggregation methods. This method was validated on the publicly available ABIDE autism dataset and REST-MDD depression dataset. The experimental results show that compared to other comparison methods based on rs-fMRI, Bi-BrainSSM achieves more accurate diagnostic performance, while also having higher computational efficiency and lower GPU memory occupancy.

Abstract:Due to factors such as medical ethics, recording conditions, participant cooperation, and annotation costs, high-quality Parkinson's disease (PD) speech annotation data is relatively scarce, which to some extent restricts the application of supervised learning methods in speech-based PD detection. Although the above problems can be overcome through cross-lingual detection and the use of multi-types of corpora, most existing studies have adopted homogeneous corpus alignment strategies, ignoring the complexity and diversity of speech expression in natural contexts, making it difficult to fully extract the deep pathological features of PD. To this end, a continuous self-supervised cross-lingual PD detection method is proposed in this paper, aiming to fully utilize speech data resources for PD detection while avoiding time-consuming and laborious manual annotation. First, based on multi-type speech data in the source language, a sequential fine-tuning strategy is adopted to adaptively train the top-level of the self-supervised feature extraction model. To prevent corpus conflicts and forgetting issues, a continuous learning replay buffer and feature distillation mechanism are introduced to enhance the plasticity of the self-supervised feature extraction model while improving its ability to retain old knowledge. Finally, the model is transferred to the target language and the ability of self-supervised models to model "language independent" features is utilized to achieve effective recognition of PD in cross-lingual scenarios. The experimental results on Italian and Chinese PD datasets show that the proposed method extracts more discriminative features and has better classification accuracy and generalization ability than existing methods in cross-lingual PD detection.

Abstract:To address the challenges of low positioning accuracy and poor robustness caused by the strong nonlinear coupling of piezoelectric hysteresis, mechanical friction, and transmission backlash in the macro-micro composite-driven cell injection mechanism, an adaptive fast non-singular terminal sliding mode (FNTSM) control strategy is proposed. This strategy aims to eliminate dependence on precise dynamic models and achieve cross-scale precise motion control. First, a dynamic model of the piezoelectric-driven cell injection mechanism is established, considering uncertainties such as hysteresis, friction, and other perturbations. Then, to overcome the dependence on the prior knowledge of disturbance bounds and the chattering issue in traditional sliding mode control, a FNTSM controller is designed that integrates time delay estimation (TDE) technology with adaptive gain adjustment. This controller uses TDE to estimate and compensate for lumped disturbances in real-time, ensuring system states converge within a finite time through FNTSM. Additionally, a PID-type sliding mode surface is introduced, and an adaptive law is designed to dynamically adjust the sliding mode surface parameters, enhancing the system’s ability to suppress strong time-varying disturbances such as friction coefficient changes and load mutations, as well as improving response speed. Experimental results show that in tracking a sine wave trajectory, the controller's root mean square error percentage is 1.39%, with a maximum tracking error percentage of 4.79%. Compared to the traditional PIDSM controller, the maximum tracking error is reduced by 50.9%. In comparison with the recently proposed PIDSMC-RBF advanced controller, the maximum tracking error percentage is reduced by 9.28%. The proposed controller effectively improves the trajectory tracking accuracy of the cell injection mechanism under cross-scale motion, providing a reliable technological solution for micro-manipulation operations.

Abstract:To address the challenge of accurately and efficiently detecting cells and injection needle tips with large scale discrepancies in automated microinjection, this paper proposes a high-efficiency instance segmentation model named DELD-YOLO. Built upon the YOLO11 architecture, the proposed model incorporates structurally re-parameterized diverse branch blocks into the backbone to enhance feature representation. An efficient lightweight multi-scale feature pyramid network is introduced to improve cross-scale feature fusion, and a detail-enhanced lightweight segmentation head is designed to preserve fine structural information while reducing computational cost. Experimental results demonstrate that DELD-YOLO outperforms the baseline YOLO11 in both cell and injector segmentation tasks, achieving mAP values of 98.7% for cells and 75.2% for injectors at IoU thresholds from 0.5 to 0.95. With only 2.08M parameters and 10.0 GFLOPs, the proposed model exhibits a favorable balance between accuracy and efficiency, making it well suited for automated cell injection and edge-device deployment.

Abstract:The steady-state visual evoked potential (SSVEP) technology is rapidly developing in the field of brain-computer interfaces. This paper targets a four-class frequency recognition task to address the low recognition rate caused by insufficient feature extraction and the influence of noise and other non-target signals in short time windows. To this end, a convolutional neural network for EEG signal classification integrating a sparse attention mechanism with time–frequency dual-domain features (SATF-CNN) is proposed. Firstly, a dynamic sinusoidal position encoding module is embedded at the input end. Then, a channel extraction path with a Top-k hybrid sparse attention mechanism is constructed, and key features are selected through multi-threshold sparse fraction selection. Finally, a dual-branch feature extraction network is built, with a time-domain convolutional network branch and a fast Fourier transform frequency-domain network branch capturing time-domain and frequency-domain features respectively. The Kolmogorov-Arnold network is used to achieve nonlinear feature fusion. Experimental results show that in a four-class cross-subject experiment with a 1-second time window, the accuracy rate is as high as 93.54%, and the information transmission rate is 93.13 bit/min. The SATF-CNN model demonstrates outstanding classification performance in SSVEP recognition tasks, providing a new solution for clinical neuroengineering and the development of intelligent interaction devices.

Abstract:Remote Sensing Image Change Captioning (RSICC) aims to generate natural language descriptions that reflect land-surface alterations using multi-temporal images, holding substantial value in domains such as land-resource monitoring and disaster assessment. However, in practical deployment, bi-temporal images are frequently vulnerable to complex non-semantic interference, such as seasonal successions and sudden illumination fluctuations, which easily induces "semantic hallucinations" in deep networks. Moreover, the ubiquitous simple differencing mechanisms adopted by existing methods struggle to effectively suppress these confounding factors and lack sufficient sensitivity to highly localized change signals, thereby yielding imprecise descriptive captions with poor contextual alignment. To tackle these multi-faceted challenges, this paper proposes a robust remote sensing image change captioning framework grounded on multi-scale feature compensation and change enhancement. To effectively alleviate the severe interference stemming from pseudo-change factors like lighting variations and environmental dynamics, a novel Feature Compensation Network (FCN) is designed. Leveraging a bidirectional guided compensation mechanism via cross-temporal attention, the FCN adaptively modulates bi-temporal features to enforce strict semantic and stylistic consistency across non-change background regions, thereby achieving the explicit suppression of illumination and shadow pseudo-changes. Concurrently, to overcome the perceptual deficiency regarding subtle change signals, a Discriminative Change-Aware Guided Module (DAGM) and a multi-scale fusion strategy are introduced. By synchronously reinforcing highly discriminative representations across both spatial and channel dimensions, this module substantially elevates the framework's operational sensitivity to the minor and intricate evolution of complex ground objects. Comprehensive quantitative comparison experiments conducted on the challenging LEVIR-CC and Dubai-CC benchmark datasets demonstrate that the proposed method consistently and significantly outperforms mainstream state-of-the-art algorithms across multiple core evaluation metrics, such as BLEU-4 and CIDEr, validating its distinct superiority in generating accurate and grammatically coherent text descriptions. Furthermore, qualitative analysis and comprehensive visualization results provide intuitive, empirical evidence supporting the robust cross-temporal interpretability of the model. Meanwhile, rigorous ablation studies systematically verify the theoretical necessity and practical effectiveness of each key proposed component. Ultimately, both experimental and theoretical investigations indicate that the proposed Feature Compensation Network (FCN) successfully achieves bidirectional interactive alignment and explicit interference suppression within non-change regions. Meanwhile, the seamlessly integrated Change-Aware Guided Module (DAGM) and multi-scale feature fusion strategy can effectively mine, preserve, and reinforce fine-grained details within genuine change regions, offering a highly promising and generalizable solution for high-precision earth observation and semantic change interpretation tasks.

Abstract:Sound Event Detection (SED) plays a critical role in intelligent perception systems, with broad applications in industrial monitoring, ecological surveillance, and public safety. To address the limitations of low-resource SED scenarios, such as insufficient labeled data and limited temporal resolution, this paper proposes a frame-level few-shot sound event detection method based on multi-task learning and dynamic convolution. The proposed method integrates a foreground/background sound classification (FBSC) task alongside the primary SED objective, and combines a NetMamba encoder with a Frequency Dynamic Convolution (FDC) module to enhance long-range temporal modeling and frequency-adaptive feature extraction. Additionally, a linear time-domain perturbation strategy, TimeFilterAug, is designed to simulate noise interference in complex acoustic environments, further improving the model's generalization capability. The proposed method achieved an F1-score of 56.7% in the DCASE 2024 Challenge, ranking second. Experimental results confirm its effectiveness in detecting transient events and maintaining robustness under noisy conditions, demonstrating strong potential for practical deployment.

Abstract:Aiming at the problem of unbalanced quality of experience (QoE) caused by limited storage space of roadside unit (RSU) in edge caching, a vehicle user (VU) activity-aware cloud-edge collaborative data caching method is proposed to improve the fairness of content caching services and ensure QoE. The three-tier cloud-edge collaborative caching system model of the Internet of Vehicles is constructed to maximize the net revenue of the system and ensure the timeliness and effectiveness of content delivery by jointly optimizing service mode and cache replacement. VU is clustered based on different activity to obtain personalized content prediction results. The content request prediction model is designed, and the global model update is realized by using the federated learning framework. The problem of maximizing the long-term profit of the system is transformed into a multi-agent markov decision process (MDP), and the multi-agent deep deterministic policy gradient (MADDPG) algorithm is used to obtain the optimal decision. The simulation results verify the effectiveness of the solution. Compared with other baseline methods, the average system revenue and cache hit rate are improved, and the system fairness is effectively ensured.

Abstract:Entity alignment is a critical task for integrating multi-source knowledge graphs, aiming to identify semantically equivalent entities across heterogeneous graphs. Existing methods typically rely on an entity's neighborhood structure or pretrained semantic information for representation learning, but they perform poorly in complex heterogeneous scenarios and generally fail to effectively utilize multi-hop path information. To address these limitations, we propose the entity alignment with path-integrated structure-aware attention (PISA). This model first extracts semantically representative multi-hop paths between entities to construct a path-triple-based extended graph structure. Then, a structure-aware multi-head attention mechanism is introduced to learn entity representations from the fused graph, enabling the model to more comprehensively capture semantic relations between entities across complex structures and diverse relations. On the DBP-15K FR-EN, JA-EN, and ZH-EN subsets, PISA’s Hits@1 improves over the baseline by 1.90%, 0.76%, and 1.79%, respectively. On the WK31-15K EN-DE (V1), EN-DE (V2), EN-FR (V1), and EN-FR (V2) subsets, PISA’s Hits@1 improves over the baseline by 1.78%, 2.01%, 1.46%, and 1.27%, respectively, which demonstrates its effectiveness and stability in heterogeneous knowledge-graph entity alignment.

Abstract:Aiming at the problem of insufficient prediction accuracy of models caused by complex nonlinear dependencies and temporal shifts between channels in multivariate time series forecasting, this paper proposes a multivariate time series forecasting model based on dilated convolution and channel attention mechanism (Dilated Convolutional Time Series Transformer, DCTST). Firstly, an inverted embedding module independently encodes the entire historical sequence of each variable into high-dimensional tokens, preserving global temporal features. Secondly, a channel attention module is designed, combining a self-attention mechanism with a learnable routing strategy, and a dilated convolution module is introduced to expand the receptive field along the variable dimension to capture multi-scale local dependencies. Finally, cross-layer residual connections integrate the original input with deep-level features, enhancing model stability. Theoretical analysis and experiments show that compared to mainstream channel-independent models (e.g., PatchTST) and channel-dependent models (e.g., iTransformer), DCTST achieves a significant reduction in both Mean Squared Error (MSE) and Mean Absolute Error (MAE) on multiple real-world and synthetic datasets. The experimental results indicate that DCTST can effectively model inter-channel dependencies, improving the prediction accuracy and robustness of multivariate time series, making it suitable for practical engineering scenarios such as smart grids and traffic flow forecasting.