An Anomly Detection Method for Rotating Machinery Based on Pre-Trained Speech Recognition Large Model

Rotating machinery is a key equipment in industrial production. Long-term operation can easily lead to aging and failure, affecting safety and efficiency. Therefore, anomaly detection is crucial. In this research, we proposed a new method to use a large amount of annotated speech data sets to train a speech recognition large model and freeze its encoder part to extract deep features of industrial machine signals and realize the cross-domain migration of speech processing technology to mechanical fault diagnosis. In addition, this method uses a monitoring model with fewer parameters to adapt to data changes. In the unsupervised case, the threshold of anomaly detection is determined by using the gamma distribution curve and Quartile, and the AUC value of anomaly detection is evaluated under supervised and unsupervised conditions. Experimental results show that this method improves the accuracy of anomaly detection while enhancing the generalization ability of the model.