One-shot transfer learning with limited data sample for bearing component fault diagnosis

Rolling element bearings are crucial components in the machine, so it is important to maintain the bearings' health. The classic deep learning model needs bulky quality data for the model to achieve high performance. However, it is difficult for the industry to obtain bulk-quality data due to machinery systems working in harsh conditions and the current one-shot learning model has limited capabilities in transfer learning. Thus, a one-shot learning with rhombus Siamese neural network (RSNN) is proposed for a small data size fault diagnosis. RSNN in this study focuses on a large number of classes with a small sample data size and transfer learning without pre-training the target data. The results proved that the one-shot RSNN has high prediction accuracy for the limited data fault detection and diagnosis (FDD) by achieving 90.63% performance with just four training data per class for the 64 classes for the CWRU dataset.