Deep convolution domain-adversarial transfer learning for fault diagnosis of rolling bearings

Abstract

It is generally difficult to obtain a large number of labeled samples (i.e., samples with known fault types) of rolling bearings installed on large-scale mechanical equipment under current working conditions, which leads to the low accuracy of fault diagnosis for current testing samples using traditional machine learning algorithms. On account of this, a novel transfer learning method termed as deep convolution domain-adversarial transfer learning (DCDATL) is proposed for rolling bearing fault diagnosis in this paper. In the proposed DCDATL, a new deep convolution residual feature extractor is constructed to extract high-level features, which can avoid gradient problems such as gradient disappearance and gradient divergence during training DCDATL, thus improving the convergence and non-linear approximation ability of DCDATL. At the same time, the joint distribution of labeled samples in auxiliary domain and unlabeled samples in target domain is creatively used for domain-adversarial training, which can enhance the adaptability of samples in auxiliary domain to target domain and improve the transfer performance of DCDATL. Moreover, the strategy based on minimizing the joint distribution domain-adversarial total loss function of DCDATL is innovatively presented to improve the fault classification accuracy after high-level feature transfer. The above advantages of DCDATL make it feasible to perform high-precision fault diagnosis on current testing samples by using the historical labeled samples in auxiliary domain when there are no labeled samples in target domain. The fault diagnosis examples of rolling bearings demonstrate the effectiveness of the proposed method.