[Paper Review] Meta-Learning Based Early Fault Detection for Rolling Bearings via Few-Shot Anomaly Detection

Wenbin Song, Di Wu+2 more|arXiv (Cornell University)|Apr 27, 2022

Machine Fault Diagnosis Techniques4 citations

TL;DR

This paper proposes a meta-learning-based few-shot anomaly detection method for early fault detection in rolling bearings, addressing unit-to-unit variability (UtUV) by learning a generic metric via a Relation Network and using health state embeddings to reduce false alarms. It achieves earlier incipient fault detection with lower false alarms than baselines on two real-world datasets.

ABSTRACT

Early fault detection (EFD) of rolling bearings can recognize slight deviation of the health states and contribute to the stability of mechanical systems. In practice, very limited target bearing data are available to conduct EFD, which makes it hard to adapt to the EFD task of new bearings. To address this problem, many transfer learning based EFD methods utilize historical data to learn transferable domain knowledge and conduct early fault detection on new target bearings. However, most existing methods only consider the distribution drift across different working conditions but ignore the difference between bearings under the same working condition, which is called Unit-to-Unit Variability (UtUV). The setting of EFD with limited target data considering UtUV can be formulated as a Few-shot Anomaly Detection task. Therefore, this paper proposes a novel EFD method based on meta-learning considering UtUV. The proposed method can learn a generic metric based on Relation Network (RN) to measure the similarity between normal data and the new arrival target bearing data. Besides, the proposed method utilizes a health state embedding strategy to decrease false alarms. The performance of proposed method is tested on two bearing datasets. The results show that the proposed method can detect incipient faults earlier than the baselines with lower false alarms.

Motivation & Objective

• Address the challenge of early fault detection (EFD) in rolling bearings when only limited target bearing data are available.
• Recognize and model Unit-to-Unit Variability (UtUV), which existing transfer learning methods often overlook despite its impact on fault detection performance.
• Formulate the EFD task under limited data and UtUV as a few-shot anomaly detection problem to improve generalization to new bearings.
• Develop a method that learns a generic similarity metric across normal and target bearing data to detect incipient faults earlier.
• Reduce false alarms through a health state embedding strategy that enhances anomaly discrimination.

Proposed method

• Employ a Relation Network (RN) to learn a generic metric that measures similarity between normal reference data and new target bearing data.
• Use meta-learning to train the RN on support sets of normal data from multiple bearings, enabling fast adaptation to new bearings with few samples.
• Introduce a health state embedding strategy that encodes the current health condition of a bearing to improve anomaly detection sensitivity.
• Train the model using episodic training, where each episode simulates a new bearing with a few support samples and query samples for few-shot adaptation.
• Optimize the RN to minimize the contrastive loss between similar (normal) and dissimilar (anomalous) samples during meta-training.
• Apply the trained model to detect incipient faults by comparing query data from a new bearing to the learned metric and embeddings.

Experimental results

Research questions

• RQ1 Can a meta-learning-based few-shot anomaly detection framework effectively detect incipient faults in rolling bearings with limited target data?
• RQ2 How does modeling Unit-to-Unit Variability (UtUV) improve early fault detection performance compared to methods that ignore bearing-specific variations?
• RQ3 To what extent does the health state embedding strategy reduce false alarms in early fault detection?
• RQ4 How does the proposed method compare to existing transfer learning and anomaly detection baselines in terms of detection timeliness and accuracy?

Key findings

• The proposed method detects incipient faults earlier than baseline methods on two real-world bearing datasets.
• The method achieves lower false alarm rates due to the integration of health state embeddings, which improves anomaly discrimination.
• The use of meta-learning enables fast adaptation to new bearings with only a few target samples, making it suitable for few-shot EFD.
• The Relation Network-based metric learning effectively captures similarity between normal and anomalous bearing states under UtUV.
• The method demonstrates robust performance across different working conditions and bearing units, indicating strong generalization.
• Empirical results confirm that considering UtUV leads to improved detection reliability and earlier fault identification compared to methods ignoring this variability.