Weighted domain adaptation networks for machinery fault diagnosis

Intelligent fault diagnosis of machines has received much attention in this big data era. Most reported models are constructed under the assumption that the training and testing data are from the same distribution. However, data distribution will shift due to working condition changes, posing challenges on the performance of intelligent models. This study considers the case that out of many known working conditions with labeled historical data, the model is to be used under another unlabelled target working condition. A multiple source domain adaptation method is proposed to learn fault-discriminative but working condition-invariant features from raw vibration signals. Different known working conditions will be assigned with different weights, on the basis of their distributional similarities to the target working condition. Two case studies are carried out to validate the effectiveness of the proposed method, respectively on rotating speed changes and load level changes. (c) 2021 Elsevier Ltd. All rights reserved.

Automated summary

This study focuses on intelligent fault diagnosis of machines in the context of changing working conditions. A multiple source domain adaptation method is proposed to learn fault-discriminative but working condition-invariant features to address the data distribution shift issue.