Journal Article

An enhanced domain-adversarial neural networks for intelligent cross-domain fault diagnosis of rotating machinery

Lv, Zhe, Zhang, Zhongwei, Zhou, Jilei, Shao, Mingyu, Ma, Chicheng

2022

Overview

In recent years, numerous studies have explored fault diagnosis methods for rotating machinery based on deep learning. For deep learning-based algorithms, the performance of fault diagnosis will be significantly worse while the labeled training data is insufficient or in case existing an obvious difference between the training and the test data. To overcome this thorny research problem, a novel domain adaptation approach is outlined, capable of realizing fault diagnosis of rotating machinery under different loads. For the proposed approach, a deep sparse filtering model is established as an extractor of the fault features to learn the representative and discriminative features of the source domain data. A domain classifier is applied for making the shift across domains to be indiscriminate. Moreover, Z -score standardization and CORAL are employed as the preprocessing tools to help reduce the influence of features with large variance and reduce the offset between the two domains, respectively. The effectiveness of the outlined method is verified via the experimental vibration data from a bearing and a gear dataset.

• Publisher: Springer Netherlands,Springer Nature B.V
• Subject: Algorithms / Automotive Engineering / Classical Mechanics / Control / Deep learning / Domains / Dynamical Systems / Engineering / Fault diagnosis / Feature extraction / Machine learning / Mechanical Engineering / Neural networks / Original Paper / Rotating machinery / Training / Vibration