To address the challenges of sensor installation limitations, severe background noise interference, and low model deployment efficiency in rolling bearing fault diagnosis in industrial environments, this paper proposes a lightweight, progressive fusion and knowledge-distillation diagnostic framework that integrates vibration and sound signals. First, considering the differences in physical characteristics between vibration and sound signals, a feature-extraction network for heterogeneous modality perception is designed: the vibration branch employs a large-kernel one-dimensional convolutional neural network, while the sound branch uses a small-kernel stacked two-dimensional convolutional neural network, with depthwise separable convolutions introduced for lightweight modification. Second, an attention-gated progressive feedback fusion strategy is proposed. Learnable gating units are used to filter the confidence of the fused features, feeding them back to the original input as residuals, effectively suppressing noise accumulation and improving fusion quality. Finally, a cross-architecture knowledge-distillation scheme is constructed, transferring the fault feature-discrimination ability from the deep heterogeneous fusion network (teacher network GAF-Net) to the lightweight LightGBM (student network Distilled-LGB). Combined with a normal sample statistical feature alignment mechanism, the student model can independently complete end-to-end fault diagnosis only with online-extractable handcrafted features, achieving microsecond-level pure model inference speed while ensuring diagnostic accuracy, fully meeting industrial edge deployment requirements. Experiments on a self-built industrial dataset and the public UOEMD-VAFCVS dataset show that GAF-Net achieves 97.89% (A → B) and 96.72% (15 Hz → 30 Hz) accuracy. Distilled-LGB achieves 21 ms inference time and 4.2 MB model size with <1% accuracy loss, demonstrating noise robustness, cross-condition generalization, and edge deployment capability.

Full fine-tuning of large-scale vision-language models for video action recognition incurs prohibitive computational cost and often degrades pre-trained spatial representations. To address this, we propose VETA-CLIP, a Video Efficient Temporal Adaptation framework that enhances temporal modeling while preserving cross-modal alignment. By incorporating lightweight adapters into a frozen backbone, VETA-CLIP introduces only 3.55M trainable parameters (a 98% reduction compared to full fine-tuning). Our approach features two key innovations: (1) an Efficient Spatio-Temporal Attention (ESTA) mechanism with a parameter-free boundary replication temporal shift (BRTS) module, which explicitly decouples spatial and temporal attention heads to capture inter-frame dynamics while minimizing disruption to the pre-trained spatial representations; and (2) a novel Variation Loss that maximizes both local inter-frame differences and global temporal variance, encouraging the model to focus on action-related changes rather than static backgrounds. Extensive experiments on HMDB-51, UCF-101, and Something-Something v2 demonstrate that VETA-CLIP achieves competitive performance across zero-shot, base-to-novel, and few-shot protocols, while and remains competitive on the Kinetics-400 dataset. Notably, our eight-frame variant requires only 4.7 GB of peak GPU memory and 2.47 ms of inference per video, demonstrating exceptional computational efficiency alongside consistent accuracy gains.

With the rapid development of the Industrial Internet of Things, data-driven deep learning has achieved remarkable success in bearing fault diagnosis. However, traditional static models suffer from catastrophic forgetting when facing continuously emerging fault categories and limited edge storage. Existing class-incremental learning frameworks expose critical limitations when applied to 1D vibration signals on micro edge devices, including feature space oscillation, difficulty in anchoring lightweight classifiers, and prototype drift over long incremental cycles. To address these challenges, this paper proposes a novel end-to-end class-incremental fault diagnosis method based on lightweighting and feature calibration tailored for severe memory-constrained conditions. Specifically, a lightweight feature extraction mechanism based on an $L_2$ constraint is introduced to replace computationally expensive similarity distillation, effectively suppressing feature space oscillations and providing stable spatial coordinates for old knowledge. Moreover, a mandatory balanced center–margin hybrid replay (CAHM) strategy is designed to balance class representation while proportionally retaining class center prototypes and marginal hard examples, balancing the anchor accuracy of the Nearest Class Mean (NCM) classifier and the discriminability of the decision boundary. Furthermore, an ultra-low-cost linear prototype calibration module is constructed using a learnable affine transformation to actively redirect shifted old class prototypes with negligible inference latency. Extensive long-tail incremental experiments on the CWRU bearing dataset demonstrate that the proposed method forms a highly synergistic anti-forgetting closed loop. Under an extremely limited memory budget ($K=40$), the proposed framework achieves an outstanding final average accuracy of 98.92% after five incremental stages, significantly outperforming mainstream baselines such as iCaRL, PRIL, and SCKD and exhibiting exceptional robustness for continuous online monitoring on industrial edge devices.

In continuous sign language recognition (CSLR), existing methods predominantly adopt frame-wise feature extraction, neglecting temporal continuity and motion trajectory modeling, thereby struggling to capture complete spatio-temporal dynamics. Meanwhile, global cross-modal attention approaches typically directly model interactions between text and the entire video sequence but lack structural constraints, making them susceptible to interference from redundant frames, which leads to attention distribution dilution and undermines fine-grained motion alignment capability. To address these issues, this paper proposes CMST-Net, a Cross-modal Interaction and Spatio-temporal Feature Enhancement Method for Continuous Sign Language Recognition. CMST-Net comprises two principal modules: the Local–Global Cross-modal Fusion Module (LGCFM) and the Spatio-Temporal Feature Enhancement Module (STFEM). LGCFM introduces a synergistic modeling mechanism that combines local sliding-window attention with global attention, achieving a unified fusion of structured local alignment and global semantic modeling. STFEM incorporates multi-scale spatial dilated convolution and coordinate attention to extract fine-grained spatial features while leveraging channel partitioning and a hierarchical residual structure to enhance long-range temporal modeling capability; the two modules collaboratively yield high-quality spatio-temporal feature representations. Experiments on three public benchmark datasets (PHOENIX2014, PHOENIX2014-T, and CSL-Daily) demonstrate that CMST-Net can effectively improve continuous sign language recognition performance, achieving state-of-the-art performance on the PHOENIX2014 and CSL-Daily datasets and competitive results on the PHOENIX2014-T dataset.

In complex environments, sign language recognition (SLR) is easily affected by background clutter, motion blur, and rapid movement. These factors can obscure subtle gesture patterns and weaken the discriminative spatiotemporal features required for recognition. To address these challenges, we propose SST-Net, a spatiotemporal collaborative attention enhancement framework for SLR. First, CGSA introduces input-dependent gates to dynamically fuse spatial and channel attention, rather than using fixed fusion weights. Second, the TFM and SFM both exploit inter-frame differences but use them differently: the TFM enhances motion-sensitive temporal features, whereas the SFM suppresses motion-dominated components to preserve spatial details; both are further refined through feature reweighting. Third, experiments on CSL-100, INCLUDE, and AUTSL show that SST-Net improves accuracy by 2–4 percentage points over strong baselines while using 35 M parameters and 115 GFLOPs, demonstrating a competitive balance between recognition performance and computational cost.

In few-shot action recognition (FSAR), limited annotated data and large scene variations make it difficult for models to learn stable spatial semantics and reliable temporal dynamics. As a result, spatiotemporal representations tend to be weak, and models often fail to focus on discriminative motion regions or capture frame-to-frame changes accurately. Furthermore, the insufficient fusion of local details and global context renders the learned features more susceptible to background noise and scene bias. These issues become more pronounced when background clutter is severe or when different action classes share locally similar segments, leading to unreliable support–query matching and shifted similarity distributions, which ultimately result in class confusion. To address these challenges, we propose a bidirectional adaptive spatiotemporal modeling method integrated with contrastive learning for FSAR. The method constructs attention-guided bidirectional differencing features to model inter-frame variations with semantic alignment, while suppressing background noise. It introduces a local–global interactive channel attention module to strengthen both local and global dynamic representations, and integrates dynamic distance adjustment with hard negative mining during tuple-level matching. This combination imposes contrastive constraints that enhance intra-class compactness and inter-class separability, thereby mitigating interference from cross-class similar segments. Experiments under the standard 5-way 1-shot/5-shot protocol demonstrate consistent improvements across multiple datasets, and the proposed method achieves the best performance under the 5-shot setting while remaining competitive under the 1-shot setting.