The Open Set Weighted Domain Adaptation Method for Fault Diagnosis of Rolling Bearings

The transfer fault diagnosis method can identify the fault of rolling bearing dataset collected under different working conditions or different equipment, which realized the cross-dataset application of fault knowledge. However, the existing transfer fault diagnosis methods all assume that the labeled training dataset and unlabeled testing dataset have the same fault label set. The above assumption is not realistic. Because the fault label in the unlabeled testing set is unknown, which makes it difficult to determine whether this fault label also belong to the label set of the training dataset. This paper considers an open set fault diagnosis problem, where the label set of the training dataset is a subset of the label set of testing dataset. In order to solve this problem, a novel open set weighted domain adaptation method is proposed for fault diagnosis of rolling bearings. The weighted domain discriminator is mainly designed in propose method by combining domain similarity and prediction uncertainty, which solves the problem of traditional transfer fault diagnosis cannot complete the domain alignment of open set data. Finally, the open set testing module is designed to distinguish between unknown label data and known label data by setting a threshold. The proposed method is validated by the rolling bearing datasets collected from failure test bench of the gearbox bearing. The results show that the proposed method can realize fault diagnosis for open set of rolling bearing dataset.

I. Introduction

Transfer learning aims to store a solution model of an existing problem and apply it to other related but different problems [1] –[3]. According to the advantages of transfer learning, many researchers have proposed the transfer fault diagnosis methods for rolling bearings. With this method, the diagnosis model learned from one kind of labeled fault dataset of rolling bearings (source domain dataset) can be applied to another related but different unlabeled fault dataset of rolling bearings (target domain dataset). And this method breaks the barrier that different domain datasets have different probability distribution by the way of domain alignment [4–5]. Many researchers have proposed different transfer diagnosis model for rolling bearings based on above ideas. For example, Li et al. [6] proposed an adversarial transfer fault diagnosis method based on fault knowledge mapping for rolling bearing to achieve across domain fault diagnosis. Wu et al. [7] proposed an adaptive depth transfer method to achieve domain alignment by introducing long-short term memory recurrent neural networks and joint distribution adaptation methods. Lei et al. [8] proposed a transfer fault diagnosis method by introducing the polynomial kernel maximum mean difference method to improve the performance of domain adaptation.