A Multistage Deep Transfer Learning Method for Machinery Fault Diagnostics Across Diverse Working Conditions and Devices

• Published in: IEEE access Vol. 8; pp. 80879 - 80898
• Main Authors: Zhou, Jian, Zheng, Lian-Yu, Wang, Yiwei, Gogu, Christian
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; convolutional neural network; Devices; Domains; Engineering Sciences; Fault diagnosis; Fault diagnostics; Feature extraction; Machine learning; Mathematics; maximum mean difference; Mechanical engineering; Mechanics; Model accuracy; multistage training; Protocol (computers); Robustness; Statistics; Training; transfer learning; Vibration; Working conditions
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.2990739

Deep learning methods have promoted the vibration-based machinery fault diagnostics from manual feature extraction to an end-to-end solution in the past few years and exhibited great success on various diagnostics tasks. However, this success is based on the assumptions that sufficient labeled data are available, and that the training and testing data are from the same distribution, which is normally difficult to satisfy in practice. To overcome this issue, we propose a multistage deep convolutional transfer learning method (MSDCTL) aimed at transferring vibration-based fault diagnostics capabilities to new working conditions, experimental protocols and instrumented devices while avoiding the requirement for new labeled fault data. MSDCTL is constructed as a one-dimensional convolutional neural network (CNN) with double-input structure that accepts raw data from different domains as input. The features from different domains are automatically learned and a customized layer is designed to compute the distribution discrepancy of the features. This discrepancy is further minimized such that the features learned from different domains are domain-invariant. A multistage training strategy including pre-train and fine-tuning is proposed to transfer the weight of a pre-trained model to new diagnostics tasks, which drastically reduces the requirement on the amount of data in the new task. The proposed model is validated on three bearing fault datasets from three institutes, including one from our own. We designed nine transfer tasks covering fault diagnostics transfer across diverse working conditions and devices to test the effectiveness and robustness of our model. The results show high diagnostics accuracies on all the designed transfer tasks with strong robustness. Especially for transfer to new devices the improvement over state of the art is very significant.