[Paper Review] Bearing fault diagnosis under varying working condition based on domain adaptation

Bo Zhang, Wei Li+2 more|arXiv (Cornell University)|Jul 31, 2017

Machine Fault Diagnosis Techniques25 references21 citations

TL;DR

This paper proposes an unsupervised domain adaptation method based on subspace alignment (SA) for bearing fault diagnosis under varying working conditions, where training and testing data distributions differ. By aligning the feature subspaces of labeled source data and unlabeled target data via PCA and linear mapping, the method enables a single classifier to generalize across domains, achieving near-100% accuracy in classifying fault types and severities on both benchmark and new experimental datasets.

ABSTRACT

Traditional intelligent fault diagnosis of rolling bearings work well only under a common assumption that the labeled training data (source domain) and unlabeled testing data (target domain) are drawn from the same distribution. When the distribution changes, most fault diagnosis models need to be rebuilt from scratch using newly recollected labeled training data. However, it is expensive or impossible to annotate huge amount of training data to rebuild such new model. Meanwhile, large amounts of labeled training data have not been fully utilized yet, which is apparently a waste of resources. As one of the important research directions of transfer learning, domain adaptation (DA) typically aims at minimizing the differences between distributions of different domains in order to minimize the cross-domain prediction error by taking full advantage of information coming from both source and target domains. In this paper, we present one of the first studies on unsupervised DA in the field of fault diagnosis of rolling bearings under varying working conditions and a novel diagnosis strategy based on unsupervised DA using subspace alignment (SA) is proposed. After processed by unsupervised DA with SA, the distributions of training data and testing data become close and the classifier trained on training data can be used to classify the testing data. Experimental results on the 60 domain adaptation diagnosis problems under varying working condition in Case Western Reserve benchmark data and 12 domain adaptation diagnosis problems under varying working conditions in our new data are given to demonstrate the effectiveness of the proposed method. The proposed methods can effectively distinguish not only bearing faults categories but also fault severities.

Motivation & Objective

• To address the challenge of bearing fault diagnosis when training and testing data are drawn from different distributions due to varying working conditions.
• To leverage existing labeled training data without requiring costly re-annotation of new labeled data for each new condition.
• To improve the domain generalization of fault diagnosis models across different loads and speeds.
• To develop a transfer learning-based solution that minimizes distribution shift between source and target domains using unsupervised adaptation.
• To validate the method on both public benchmark data and newly collected experimental data with diverse operating conditions.

Proposed method

• The method uses Fast Fourier Transform (FFT) to convert vibration signals into spectrum images for feature representation.
• Principal Component Analysis (PCA) is applied to both source and target domains to extract low-dimensional feature subspaces.
• Unsupervised Subspace Alignment (SA) learns a linear transformation to align the source subspace with the target subspace, minimizing domain discrepancy.
• A common linear classifier (SVM or neural network) is trained on the aligned source data and applied directly to the target domain.
• The alignment process is performed without any labeled target data, making it unsupervised.
• The method is evaluated on 60 domain adaptation tasks from Case Western Reserve University data and 12 from newly collected data with varying loads and speeds.

Experimental results

Research questions

• RQ1 Can unsupervised domain adaptation with subspace alignment effectively reduce distribution shift between labeled training data and unlabeled testing data under varying working conditions?
• RQ2 Can the proposed method achieve high fault classification accuracy across different loads and speeds without retraining or re-annotating data?
• RQ3 How does the proposed method compare to traditional fault diagnosis approaches that assume identical source and target data distributions?
• RQ4 To what extent can the method generalize to new fault severities and operating conditions not seen during training?
• RQ5 Does the alignment of feature subspaces improve the discriminability of fault categories and severities in cross-domain settings?

Key findings

• The proposed method achieved 100% classification accuracy on 60 domain adaptation problems from the Case Western Reserve University benchmark dataset under varying loads and speeds.
• On the newly collected experimental dataset with varying speeds, the method achieved near-100% accuracy, demonstrating strong generalization across speed variations.
• After subspace alignment, the distributions of the same fault class across different domains were significantly pushed closer together, as shown by visualization.
• The method successfully distinguished both fault categories and fault severities (0.007in to 0.028in) across multiple working conditions.
• The approach outperformed prior methods that relied solely on invariant feature extraction, as it actively aligns domains using unlabeled target data.
• The results confirm that unsupervised domain adaptation via subspace alignment is effective for real-world bearing fault diagnosis where data distributions shift due to operational changes.