Enhanced triplet network for fault diagnosis with parallel convolution under sample imbalance

Due to prolonged high-speed operation and challenging operating environments, bearings in equipment are prone to failure. Fault diagnosis plays a crucial role in ensuring the operation of equipment by establishing a mapping between monitoring data and different health states to detect faults promptly. The difficulty of unbalanced training samples is often encountered in the fault diagnosis process. It is difficult to perform accurate fault diagnosis. This paper addresses the issue of bearing fault diagnosis under imbalanced sample conditions and proposes a Parallel Convolution Module with Improved Triplet Network (PCM-ITN) method based on metric learning. The proposed method focuses on extracting information from minority-class fault samples and optimizing both intra-class and inter-class distances. Parallel Convolution Module enhances feature extraction capabilities by extracting signal features at multiple scales. A triple loss function is used to make the classification boundaries clearer.. Additionally, the focal loss (FL) function is incorporated to improve diagnostic efficacy and perform fault classification tasks directly. Comparative experiments conducted on two experimental platforms show that the PCM-ITN method achieves the highest classification accuracy at all imbalanced rates, with average improvements of of 14.15% and 18.14%. Ablation experiments validate the effectiveness of the proposed method.