Few-shot fault diagnosis of rolling bearings based on time-frequency enhanced multi-scale metric learning
Abstract
To address the challenges of low diagnostic accuracy, limited feature extraction capability and poor cross-condition generalisation in rolling bearing fault diagnosis under small-sample scenarios, this study proposes an end-to-end CWT-MCNN-SNN framework that integrates the continuous wavelet transform (CWT), a multi-scale convolutional neural network (MCNN) and a Siamese neural network (SNN). The CWT converts one-dimensional vibration signals into two-dimensional time-frequency representations, enhancing the visibility of non-stationary transient features. The MCNN employs complementary small- and large-scale convolutional kernels to fuse local impulsive details with global structural information, addressing the limitations of single-scale feature extraction. The SNN introduces a metric-learning mechanism that strengthens similarity discrimination and mitigates classification difficulties in few-shot conditions. The experimental results show that the proposed model achieves 68.97 ± 1.57% accuracy with only 20 training samples, improving to 86.73 ± 1.29% and 93.63 ± 1.06% with 40 and 60 samples, respectively, and consistently exceeding 99% when more than 300 samples are available. The framework further demonstrates strong robustness and generalisation across multi-condition, cross-condition and cross-dataset evaluations on the Case Western Reserve University (CWRU) dataset and the Southeast University (SEU) dataset, substantially outperforming comparative algorithms. These findings confirm that the unified design of time-frequency enhancement, multi-scale feature fusion and metric-learning optimisation enables the CWT-MCNN-SNN framework to serve as a highly accurate, stable and transferable solution for intelligent fault diagnosis under limited-sample conditions.