A novel knowledge transfer network with fluctuating operational condition adaptation for bearing fault pattern recognition

Abstract

Data-driven based intelligent fault pattern recognition methods of rolling element bearings have made fruitful achievements in recent years. However, for real-world diagnostic occasions, a hypothesis of identical distribution between training and test datasets of current deep learning approaches is natural to be violated. Though reported transfer learning models have boosted diagnostic performance through knowledge transfer, yet most of them suffer from the following limitations: (a) Partial vibration datasets in the target domain are needed to assist the training of the models. (b) Operational conditions information is not adequately considered in the current diagnostic models. (c) Knowledge transfer-based methods of bearing fault pattern recognition are always implemented under stationary operational conditions. To remedy these limitations mentioned above, a new knowledge transfer network with a sparse auto-encoder and a deep convolutional neural network (KTN-SAEDCNN) is proposed in this paper. In KTN-SAEDCNN architecture, input instantaneous rotating speed (IRS), as operational condition information, is fed into the sparse auto-encoder (SAE) in the target domain so that the operational information can be included into the model training rather than make use of the partial vibration dataset only. Then deep convolutional neural network (DCNN) is utilized to extract features from raw vibrations. Finally, a knowledge transfer model of KTN-SAEDCNN is established through the combination of SAE and DCNN two sub-models. The proposed model enables the capacity of fault pattern recognition under fluctuating operational conditions for rolling bearing fault detection. Lastly, experiments result in bearing demonstrate the effectiveness of the proposed method.