For the current wearable devices in the application of cross-diversified user groups, it is common to face the technical difficulties of static sign language recognition accuracy attenuation, weak anti-noise ability, and insufficient system robustness due to the differences in the use of users. This paper proposes a novel static sign language recognition method enhanced by a self-attention mechanism. The key features of sign language gesture classification are highlighted by the weight function, and then the self-attention mechanism is combined to pay more attention to the key features, and the convolutional neural network is used to extract the features and classify them, which realizes the accurate recognition of different types of static sign language under standard gestures and non-standard gestures. Experimental results reveal that the proposed method achieves an average accuracy of 99.52% in the standard static sign language recognition task when tested against the standard 36 static gestures selected within the reference American Sign Language dataset. By imposing random angular bias conditions of ±(0°–9°] and ±(9°–18°], the average recognition rates in this range were 98.63% and 86.33%. These findings indicate that, compared to existing methods, the proposed method not only maintains a high recognition rate for standard static gestures but also exhibits superior noise resistance and robustness, rendering it suitable for static sign language recognition among diverse user populations.

Rolling bearings are critical components in industrial rotating machinery, with their fault diagnosis accuracy directly impacting operational safety. Existing intelligent diagnostic methods face challenges such as low efficiency in modeling long time-series signals, insufficient attention to key fault features, and high sensitivity to model hyperparameters. To address these issues, a triple-collaborative diagnostic framework is proposed, integrating a dynamic chaotic sparrow optimization algorithm with attention mechanisms. The framework incorporates the dynamic chaotic sparrow optimization algorithm, which adaptively adjusts the population roles and utilizes chaotic perturbation strategies to fine-tune the model’s parameter space. An informer-SENet hybrid architecture is then employed, leveraging a probabilistic sparse self-attention mechanism for efficient feature extraction from long-sequence vibration signals, while a channel-sensitivity weighting mechanism improves the representation of fault-sensitive frequency bands. This model offers global diagnostic capabilities with parameter-adaptive optimization. Experimental evaluation on the Case Western Reserve University bearing dataset demonstrates that the proposed method achieves 100% accuracy, F1 score, recall, and precision under standard operating conditions. In cross-load testing at a 3HP high-load condition, the method maintains diagnostic accuracy at 98.53 (± 0.16) %. Furthermore, cross-condition validation on the XJTU-SY bearing dataset achieved an average accuracy range of 94.35%–99.20% across different scenarios, demonstrating the strong generalization and robustness of the proposed framework under variable speed, load, and noise environments. This approach provides an efficient and robust solution for real-time, precise fault diagnosis of industrial bearings.

Dynamic gesture recognition technology for wearable devices is a key enabler of efficient human-machine interaction within the embedded devices. However, existing methods still face challenges such as low recognition accuracy in small sample conditions, high computational overhead, limited robustness, and inadequate real-time performance on embedded devices, particularly in the context of wearable devices where real-time processing and low-power consumption are crucial. This paper proposes a lightweight and adaptive dynamic gesture recognition method based on data glove input that fuses time–frequency domain features to overcome these challenges. The proposed approach integrates dynamic-weight convolution with a Gaussian kernel-optimized shared cross-attention mechanism, effectively combining both time-domain and frequency-domain features while capturing global and local information from gesture signals acquired by a custom data glove equipped with bending and attitude sensors. Additionally, parameter sharing and a lightweight parameter design are employed to reduce computational complexity. Experimental results on the ‘National Standard Sign Language Dictionary’ dataset reveal a test-set recognition accuracy of 98.44%, a gesture recognition time of 2.5 ms per gesture, and a model size of 29.75 k parameters, occupying only 0.12 MB of storage. Tests on the Jetson Nano embedded device show that the proposed method maintains a high recognition accuracy of 96.89%, with a recognition time of 5.7 ms per gesture and an average power consumption of 2.58 W in low-power mode, confirming its efficiency, real-time capability, and feasibility for embedded devices, especially in resource-constrained environments.

Electroencephalography (EEG) provides a noninvasive basis for depression detection. However, building robust models remains challenging due to the nonstationary temporal-spectral dynamics, subject-dependent inter-channel relationships, and weakly constrained fusion across heterogeneous EEG representations. This paper proposes a Temporal-Spectral and Graph Cross-Attention Network (TS-GCAN) for EEG-based depression detection within a sensor signal processing framework, and develops a structurally constrained collaborative analysis paradigm. The proposed framework integrates multi-scale temporal-spectral encoding, prior-constrained adaptive graph learning, structure-aware cross-domain attention fusion, and prototype-based aggregation into a unified pipeline, thereby enhancing coordinated modeling across complementary representations. Experimental results on two public resting-state EEG datasets (PRED+CT and HUSM) show that TS-GCAN achieves 94.52% accuracy (Acc), 93.73% specificity (Spe), 94.56% F1-score (F1), and 0.948 AUC on PRED+CT, and 91.89% Acc, 91.54% Spe, 91.62% F1, and 0.970 AUC on HUSM. Paired statistical hypothesis testing across folds further confirms that the proposed method achieves significant performance improvements over competing baselines (p < 0.05). In addition, TS-GCAN maintains a moderate model size and computational cost while achieving millisecond-level inference latency. It should be noted that the results on PRED+CT are obtained from a subject-level balanced subset, with labels defined by BDI questionnaire thresholds. In contrast, HUSM provides complementary evidence under clinically grounded DSM-IV diagnostic labels. Therefore, the reported findings should be interpreted in the context of the current subject-level evaluation protocol and differing label definitions.

In brain tumor magnetic resonance imaging (MRI) segmentation, conventional U-Net structures are limited by insufficient multimodal fusion, weak multi-scale semantic representation, and deep feature overlapping. This results in suboptimal precision when identifying and segmenting complex lesion areas. To address these limitations, this paper proposes a multimodal brain tumor segmentation method based on multi-branch fusion and pyramid attention mechanisms. Specifically, cross-scale attention modules are integrated into all skip connections, leveraging split pyramid convolution and efficient channel attention to achieve dynamic fusion of multi-scale features. Furthermore, a multi-branch modality feature fusion module is introduced in the first three stages of the skip connections to effectively integrate complementary information across different MRI modalities, thereby enhancing feature representation capabilities. A prototype-enhanced Atrous spatial pyramid pooling (ASPP) module is employed at the bottleneck layer, where a prototype-aware block (PAB) is embedded within the conventional ASPP. By extracting class prototypes using masks and calculating their similarity with feature maps, the PAB performs prototype-guided reweighting of multi-scale Atrous features, thereby directing the network’s attention toward highly discriminative regions. Comparative experiments on the public BraTS2020 and BraTS2021 datasets demonstrate that the proposed method achieves favorable segmentation results. On BraTS2020, the proposed method achieved dice scores of 90.45%, 86.65%, and 79.85% for the whole tumor, tumor core, and enhancing tumor (ET) regions, respectively. On BraTS2021, it achieved dice scores of 92.82%, 87.91%, and 83.71%, with corresponding HD95 values of 4.38 mm, 6.81 mm, and 5.07 mm. In addition, boundary-oriented evaluation in the ET region further demonstrated good delineation capability. These results suggest that the proposed method has certain advantages in complex lesion segmentation, boundary delineation, and small-volume lesion identification.