A Class Incremental Learning Approach Based on Autoencoder Without Manual Feature Extraction for Rail Vehicle Fault Diagnosis

Traditional intelligent diagnosis methods and current popular deep learning based diagnosis methods basically adopt the approach of batch learning, which may waste time and computing resources since they need to discard the previous learned model and retrain a new model based on the newly added data and prior data. Moreover, manual feature extraction is often a necessary step for intelligent diagnosis, and such a process relies much on prior knowledge. To solve the above mentioned problems, this paper proposes a fault diagnosis method based on class incremental learning without manual feature extraction. Based on denoising autoencoder, the proposed method obtains the autoencoders using the raw data acquired for each health state. In the class incremental learning process, only the autoencoder of new health state need to be trained while the former trained autoencoders are retained. Test data is classified according to the minimal reconstruction error calculated through the autoencoders. At the end of this paper, the proposed method is validated through vibration data of rolling bearings for rail vehicle. The experimental results show that the presented method is effective.