Fully Simulated-Data-Driven Transfer-Learning Method for Rolling-Bearing-Fault Diagnosis

Abstract

Transfer learning has been applied to deal with the insufficient labeled target dataset problem in data-driven fault diagnosis. However, most existing solutions cannot work well when real data is completely unusable in the training process, which often occurs in engineering practice. Considering this challenge, a fully simulated-data-driven transfer-learning method is proposed for rolling-bearing-fault diagnosis. The proposed method's key feature is using a domain-invariant data-transform method to convert domain-variant datasets to domain-invariant datasets so that common features from simulated and real datasets can be shared. The transform process relies on the physics knowledge of bearing faults and is implemented using the hidden Markov model (HMM). The proposed method is fully driven by simulated data without using real datasets under the fault-diagnosis-model construction process. This approach is a new way to implement transfer learning for rolling-bearing-fault diagnosis. The experimental results demonstrate the proposed method's effectiveness with real data.