Meta-Learning Based Early Fault Detection for Rolling Bearings via Few-Shot Anomaly Detection

Di WuarXiv 2022 · DOI · arXiv

TL;DR

This paper introduces a novel meta-learning approach for early fault detection in rolling bearings, specifically addressing the challenge of limited data for new bearing types. By leveraging a Relation Network and a Health State Embedding Strategy, the method effectively detects incipient faults with fewer false alarms, even when dealing with unit-to-unit variability.

What Problem Does This Paper Address?

Rolling bearings are critical components in mechanical systems, and their early failure detection (EFD) is crucial for preventing breakdowns and economic losses. Traditional EFD methods often require a substantial amount of historical data from the specific bearing being monitored. However, in real-world scenarios, obtaining such extensive data for new or rarely encountered bearing types is often infeasible. This "few-shot" learning problem is further complicated by "unit-to-unit variability" (UtUV), where even bearings of the same type can exhibit different degradation patterns due to subtle variations in manufacturing or operating conditions. Existing transfer learning methods primarily address distribution shifts across different working conditions but overlook this UtUV, leading to potential inaccuracies and false alarms.

Methods

The proposed method tackles the few-shot anomaly detection (FSAD) problem for EFD by combining meta-learning with a novel Health State Embedding Strategy (HSES).

  • Meta-Learning with Relation Network (RN): The core of the approach is a Relation Network, a meta-learning technique that learns a transferable metric to measure the similarity between normal and potentially faulty bearing data. This RN is trained on historical bearing data from various sources, allowing it to generalize to new bearing types with limited available normal data.
  • Health State Assessment (HSA) with USSCNN: Before meta-learning, an Unsupervised Sequence Segmentation Convolutional Neural Network (USSCNN) is used to identify the health states (normal vs. faulty) of historical bearing data. This is crucial because historical data often lacks explicit labels. USSCNN helps determine the "early fault start time" for each historical dataset.
  • Health State Embedding Strategy (HSES): To mitigate false alarms, especially in the initial stages of monitoring new bearings, the HSES dynamically expands the set of "normal" reference samples. New arrival data identified as normal are added to the support sets, making the detection more robust.
  • Robust Early Fault Alarm Criterion: A voting strategy and a threshold-based criterion are employed to decide when to issue an early fault alarm, further reducing the likelihood of false positives caused by noise or minor anomalies.

Datasets

The proposed method was evaluated on two widely used run-to-failure bearing datasets:

  • IEEE PHM Challenge 2012 dataset: This dataset contains accelerated bearing degradation experiments under various working conditions. The paper specifically uses bearings PHM1_1 and PHM1_3 for testing.
  • XJTU-SY dataset: This dataset comprises experiments on 15 rolling bearings under three different working conditions. Bearing XJD1_5 was used for testing.

The remaining data from these datasets were used for training the meta-learning model.

Key Insights and Ideas

  • Meta-learning for UtUV: The paper demonstrates that meta-learning, specifically with Relation Networks, is a powerful paradigm for handling both limited target data and unit-to-unit variability in EFD. The learned metric can generalize across different bearing units.
  • HSES for Robustness: The Health State Embedding Strategy is a key innovation for reducing false alarms. By continuously updating the set of normal reference samples, the system becomes less sensitive to transient noise and minor deviations, leading to more reliable early fault detection.
  • Bridging the Gap in Few-Shot EFD: The work effectively formulates EFD with limited target data and UtUV as a Few-Shot Anomaly Detection problem, a significant step towards practical EFD solutions for diverse industrial applications.
  • Importance of Historical Data Labeling: The use of USSCNN to automatically label historical data is a practical necessity, as obtaining manually labeled historical datasets for all possible scenarios is often impractical.

Key Results

  • The proposed method achieved earlier detection of incipient faults compared to several baseline methods on the tested datasets.
  • Crucially, the method generated significantly fewer false alarms than most baselines, particularly when using the HSES. For instance, on PHM1_1 and PHM1_3, the proposed method achieved early fault detection with zero false alarms.
  • The fitness scores, an evaluation metric, showed that the proposed method (with and without HSES) achieved the lowest average scores among the tested baselines, indicating superior performance.
  • The RN's feature extraction was shown to create more discriminative feature spaces for distinguishing between normal and abnormal data.

Limitations and Future Work

The authors acknowledge two main limitations:

  • Manual Threshold Setting: The current method requires manually setting a threshold to determine the health states of samples.
  • Platform Transferability: The knowledge learned is specific to the tested platforms and cannot be directly transferred to different industrial platforms without retraining.

Future work will focus on:

  • Developing methods to learn the threshold automatically.
  • Creating an adaptive EFD method that can adjust to different platforms for real-world industrial applications.

Summary updated

BibTeX
@misc{wu2022meta,
  author = {Di Wu},
  title = {{Meta-Learning Based Early Fault Detection for Rolling Bearings via Few-Shot Anomaly Detection}},
  year = {2022},
  eprint = {2204.12637},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2204.12637},
}
Figure