A Novel Method for Fault Diagnosis of Rolling Bearings Based on Domain-Adversarial Partial Transfer Learning

With the successful application of deep learning, fault diagnosis of rotating machinery has a significant accuracy improvement. However, these methods based on deep learning need massive labeled samples for training, while it is quite difficult to obtain labeled fault data in the real scenario of engineering. Lacking labeled fault samples, many intelligent methods can not be applied to fault diagnosis. Fortunately, the classical datasets of rotary machines nearly contain all the common fault types. Moreover, transfer learning provides an alternative approach for fault diagnosis of rotating machinery under the situation of lacking labeled fault data. Therefore, a novel intelligent method based on domain-adversarial partial transfer learning is proposed in this paper for the settlement of the above problem, named DA-PTL. The proposed model consists of three parts: feature extractor, domain discriminator, and sample classifier. By transfer learning, the method can realize the fault diagnosis of the target bearings using an open bearing dataset and the gathered data of the target bearings. The experimental results indicate that the proposed method can achieve a good result for fault diagnosis of rolling bearing that is a lack of massive labeled fault samples.