Physics-Informed Deep Learning and Partial Transfer Learning for Bearing Fault Diagnosis in the Presence of Highly Missing Data

TL;DR

The paper tackles bearing fault diagnosis under partial-set and missing-data conditions by introducing PTPAI, a physics-informed deep learning framework that generates synthetic labeled data in the source domain and leverages unlabeled real data in the target domain. It combines MK-MMSD and CDAN for distribution alignment, an auxiliary discriminator and rebalanced RF-Mixup to handle imbalanced and outlier classes, and a comprehensive weighting scheme to address PSFD. Experimental results on CWRU and JNU demonstrate that PTPAI consistently outperforms state-of-the-art methods across varying imbalance levels and missing-data scenarios, with ablation studies highlighting the importance of RF-Mixup and the weighting blocks. The approach offers a practical, scalable solution for real-world bearing fault diagnosis where labeled faulty data are scarce or unavailable and data are often incomplete, enabling more reliable maintenance decision-making.

Abstract

One of the most significant obstacles in bearing fault diagnosis is a lack of labeled data for various fault types. Also, sensor-acquired data frequently lack labels and have a large amount of missing data. This paper tackles these issues by presenting the PTPAI method, which uses a physics-informed deep learning-based technique to generate synthetic labeled data. Labeled synthetic data makes up the source domain, whereas unlabeled data with missing data is present in the target domain. Consequently, imbalanced class problems and partial-set fault diagnosis hurdles emerge. To address these challenges, the RF-Mixup approach is used to handle imbalanced classes. As domain adaptation strategies, the MK-MMSD and CDAN are employed to mitigate the disparity in distribution between synthetic and actual data. Furthermore, the partial-set challenge is tackled by applying weighting methods at the class and instance levels. Experimental outcomes on the CWRU and JNU datasets indicate that the proposed approach effectively addresses these problems.

Figures (10)

• Figure 1: The effect of class-imbalance and distribution discrepancy in the fault diagnosis task. The feature vectors $V_{normal}$ and $V_{faulty}$ correspond to the learned feature representations of the normal and faulty classes, respectively. $D_{ideal}$ denotes the ideal decision boundary of the classifier, while $D_{learnt}$ represents the actual decision boundary learned from the provided datasets.
• Figure 2: Overall framework of the proposed PTPAI method
• Figure 3: The comparison of the RF-mixup with several DBs. With $m<1$, the RF-mixup allocates the label from the minority class to a greater value. When $m>1$, RF-mixup produces labels that support the majority class. As a result, this paper only considers $m<1$.
• Figure 4: Different techniques' diagnosis b-accuracy for different ranges of balance rate in the CWRU dataset
• Figure 5: Different techniques' diagnosis F1-score for different ranges of balance rate in the CWRU dataset