Sparse Sample Train Axle Bearing Fault Diagnosis: A Semi-Supervised Model Based on Prior Knowledge Embedding

Data-driven fault diagnosis models often exhibit limited generalization abilities when trained on small sample sizes, as commonly encountered in complex working environments such as train bearing systems. To address this challenge, incorporating domain knowledge into machine learning models has been proposed. Specifically, prior knowledge in the form of wavelet packet decomposition and information entropy features was extracted to characterize fault-related physical phenomena. Convolutional neural networks (CNNs) supplemented this approach by learning general features. An attention mechanism fused the two feature types, emphasizing the prior knowledge. Extreme gradient boosting (XGBoost) classification was then applied for fault discrimination. In addition, a self-training method semi-supervised the model during training. The approach was evaluated on three bearing datasets against deep learning and machine learning baselines. Results showed the method achieved over 70% accuracy with only 30 labeled samples per class, outperforming the semi-supervised baseline. This lightweight, knowledge-driven solution provides a foundation for efficient fault diagnosis from limited data in complex fault scenarios. Overall, the study demonstrated an effective method for addressing small sample size challenges in machine fault diagnosis.