A Cross-Domain Stacked Denoising Autoencoders for Rotating Machinery Fault Diagnosis under Different Working Conditions

Abstract

In fault diagnosis of rotating machinery, the shift in domain distributions caused by working condition fluctuations poses a major obstacle for accurate diagnosis. Due to the lack of domain adaptation ability, the diagnosis performance of existing deep learning-based methods degrades significantly when confronting other unseen working conditions. To address this problem, we develop a cross-domain stacked denoising autoencoders (CD-SDAE) with a new adaptation training strategy. Taking advantages from both domain adaptation and manifold learning, the adaptation training strategy consists of two successive paradigms: 1) unsupervised adaptation pre-training to correct marginal distribution mismatch and 2) semi-supervised manifold regularized fine-tuning to minimize conditional distribution distance between domains. In this way, the marginal distributions between the source and target domains are first matched. Then, on this basis, the conditional distributions can be matched more effectively thus makes the model become more adaptable to the target domain. The CD-SDAE is evaluated on gearbox and engine rolling bearing fault datasets. The experimental results show that CD-SDAE is superior to not only conventional deep learning method but also state-of-the-art deep domain adaptation method in terms of diagnostic accuracy.