Research on bearing fault diagnosis integrating deep learning and knowledge graphs

Abstract

To overcome the problems of scattered knowledge in traditional fault diagnosis, difficulty in identifying non-stationary signals, and poor interpretability of results, an intelligent diagnosis method integrating deep learning and a knowledge graph is proposed. This paper designs a VMD Transformer BiGRU model, combines mechanism knowledge with vibration signals, and improves fault recognition ability under complex working conditions through multi-scale decomposition and bidirectional temporal feature modeling. A multi-level bearing fault knowledge graph is constructed based on classification results and mechanism rules, and the KG-BERT-RGCN dual channel model is used to achieve semantic and structural joint modeling, breaking through traditional completion dependencies. Finally, an intelligent decision support system is built. The experiment shows that this method achieves a diagnostic F1 score of 95.59% and outperforms mainstream methods in link prediction performance.