Incipient Fault Diagnosis Based on DNN with Transfer Learning

Abstract

Diagnosis of incipient fault is critical for safe operation of the system because it can prevent disastrous accidents from happening by diagnosing the early fault before deterioration. Deep learning is efficient in feature extraction but it requires a large number of samples to train traditional deep neural network (DNN). It is thus inevitable that the efficiency of DNN will be affected when it is applied to incipient fault diagnosis for there are usually a very limited number of incipient fault samples. Furthermore, a large amount of information involved in significant fault samples was not adequately used for incipient fault diagnosis. To solve this problem, this paper proposes an incipient fault diagnosis model with DNN-based transfer learning. The model can extract fault feature involved in a large number of significant fault samples and apply it to extract insignificant fault feature with a small number of incipient fault samples. In this way, the proposed transfer learning method can efficiently diagnose incipient fault in the case when only a limited number of incipient fault data is available. The efficiency of the proposed model is demonstrated by utilizing the Case Western Reserve University bearing data set.