A rolling bearing fault diagnosis method based on multi-scale depthwise separable convolution (MDSC) and a convolutional neural network–Transformer hybrid model (CNN-Transformer) is proposed to address the non-stationarity of fault signals and the difficulty of jointly capturing local and global features. First, continuous wavelet transform (CWT) converts one-dimensional vibration signals into two-dimensional time-frequency images to enhance fault representation. Then, multi-scale convolution (MSC) and depthwise separable convolution (DSC) are introduced to extract local impulsive features and fault patterns at different scales with fewer parameters. A CNN-Transformer architecture is further developed, where convolutional neural network (CNN) captures local details and Transformer models global dependencies. In addition, pretraining-finetuning, data augmentation, label smoothing, and normal sample optimization are adopted to improve training stability and diagnostic performance. Experimental results show accuracies of 98.80% on the Xi’an Jiaotong University bearing dataset (XJTU-SY) and 100.00% on the Case Western Reserve University bearing dataset (CWRU), demonstrating strong discriminative ability, stability, and robustness.
Shuai Yang, Yanchao Chen, Yang Yu· Engineering Research Express· 0 citations
Wind turbine gearbox fault diagnosis is essential for ensuring the safe and stable operation of wind energy systems. However, in industrial scenarios, limited fault data and highly complex, non-stationary operating conditions lead to significant distribution shifts, which reduce the generalization ability of traditional methods. To address this issue, we propose a task-oriented fault diagnosis framework based on large language models. Time-domain and frequency-domain features are extracted from vibration signals to construct structured representations. Inspired by few-shot learning, a support–query task construction strategy is introduced, reformulating classification as conditional task reasoning. Low-Rank Adaptation is adopted for parameter-efficient fine-tuning of the pre-trained model. In addition, a global coverage and local balance task sampling strategy is designed to enhance task diversity and mitigate sample imbalance. Experiments on a wind turbine gearbox dataset show consistent performance across multiple backbone models, including LLaMA2-7B, LLaMA3-8B, Qwen-7B, and Baichuan-7B. LLaMA3-8B achieves the best performance, with 98.00% accuracy and an F1-score of 0.9798. These results demonstrate strong robustness and cross-model generalization under few-shot and complex operating conditions.