Triplet Loss Guided Adversarial Domain Adaptation for Bearing Fault Diagnosis

Abstract

Recently, deep learning methods are becomingincreasingly popular in the field of fault diagnosis and achieve great success. However, since the rotation speeds and load conditions of rotating machines are subject to change during operations, the distribution of labeled training dataset for intelligent fault diagnosis model is different from the distribution of unlabeled testing dataset, where domain shift occurs. The performance of the fault diagnosis may significantly degrade due to this domain shift problem. Unsupervised domain adaptation has been proposed to alleviate this problem by aligning the distribution between labeled source domain and unlabeled target domain. In this paper, we propose triplet loss guided adversarial domain adaptation method (TLADA) for bearing fault diagnosis by jointly aligning the data-level and class-level distribution. Data-level alignment is achieved using Wasserstein distance-based adversarial approach, and the discrepancy of distributions in feature space is further minimized at class level by the triplet loss. Unlike other center loss-based class-level alignment approaches, which hasto compute the class centers for each class and minimize the distance of same class center from different domain, the proposed TLADA method concatenates 2 mini-batches from source and target domain into a single mini-batch and imposes triplet loss to the whole mini-batch ignoring the domains. Therefore, the overhead of updating the class center is eliminated. The effectiveness of the proposed method is validated on CWRU dataset and Paderborn dataset through extensive transfer fault diagnosis experiments.

Keywords: Wasserstein distance; fault diagnosis; triplet loss; unsupervised domain adaptation.

Figure 1

The architecture of the proposed method. Instead of computing and updating class centers for each class and reducing the distance of the same class center from different domain, the proposed TLADA method concatenates 2 mini-batches from source and target domain into a single mini-batch and imposes triplet loss to the whole mini-batch ignoring the domains.

Figure 2

The architecture of the proposed method.

Figure 3

Test rig used in this paper. (a) CWRU testbed; (b) Paderborn testbed.

Figure 4

Collected signals of Ball Fault with fault depth 0.014 inch under different working conditions. The horizontal axis represents the time and the horizontal axis are the acceleration data. Four vibration signals under load conditions of 0, 1, 2, 3 from the drive end are shown in (a–d), respectively. Four vibration signals under load conditions of 0, 1, 2, 3 from the fan end are shown in (e–h), respectively.

Figure 5

t-SNE results of task ‘FE -> DE’ on CWRU dataset through method (a) WDGRL, (b) TLADA. To better inspect the class-level alignment between domains, we draw the features of both the source domain and the target domain into single images. Two shapes represent two domains (square for source domain and triangle for target domain), and four colors with numbers represent four classes.

Figure 6

Confusion matrix results of task ‘FE -> DE’ on CWRU dataset through method (a)WDGRL, (b) TLADA. The accuracy and recall of each class are added to the matrix as well.

Figure 7

Vibration signals of Paderborn dataset under different working conditions. The horizontal axis represents the time and the horizontal axis are the acceleration data. Health signals (K001) under working conditions of 1, 2, 3 are shown in (a–c), respectively. Outer Race signals (KA04) under working conditions of 1, 2, and 3 are shown in (d–f), respectively. Inner Race signals (KI04) under working conditions of 1, 2, and 3 are shown in (g–i), respectively.

Figure 8

t-SNE results of task ‘PA -> PB’ through the method: (a) WDTRL, (b) TLADA. Two shapes represent two domains (square for source domain and triangle for target domain) and three colors represent three classes.

Figure 9

Confusion matrix of task ‘PA -> PB’ through the method: (a) WDGRL, (b) TLADA.

Figure 10

Comparison of accuracy of TLADA variants on four tasks.

Figure 11

Analysis of parameter sensitivity of threshold in TLADA. Dashed lines show baseline results without TLADA.