Contrastive Learning for Fault Detection and Diagnostics in the Context of Changing Operating Conditions and Novel Fault Types

Katharina Rombach, Gabriel Michau, Olga FinkSensors 2021 · DOI

TL;DR

This paper introduces contrastive learning, specifically using triplet loss, as a novel approach for fault detection and diagnostics in industrial machinery. The method aims to create feature representations that are robust to changing operating conditions while remaining sensitive to new, previously unseen fault types, a critical challenge in real-world industrial settings.

What Problem Does This Paper Address?

Industrial assets are monitored to detect faults and ensure efficient operation. However, data-driven fault detection methods face two major hurdles:

  1. Changing Operating Conditions: External factors like temperature or load can alter sensor readings, leading to false alarms even when the machinery is healthy. Training data often doesn't capture all possible operating conditions.
  2. Novel Fault Types: Faults are rare, and training datasets rarely include all possible fault types or their variations. This makes it difficult for models to detect new, unencountered faults.

Existing methods struggle to be both invariant to normal operational variations and sensitive to new fault signatures. This paper aims to bridge this gap by developing a feature learning approach that addresses both challenges simultaneously.

Methods

The core of this work is the application of contrastive learning, a technique that learns to group similar data points together and push dissimilar ones apart in a learned feature space. Specifically, the authors employ the triplet loss function.

  • Triplet Loss: This loss function works with triplets of data points: an "anchor" (a data sample), a "positive" (a sample similar to the anchor, e.g., same fault type), and a "negative" (a sample dissimilar to the anchor, e.g., a different fault type). The goal is to minimize the distance between the anchor and positive samples while maximizing the distance between the anchor and negative samples.
  • Feature Encoder Network: A neural network (encoder) is trained using the triplet loss to map raw sensor data into a low-dimensional feature space (16 dimensions in this study). The idea is that this learned representation will capture the essential semantic meaning of the data, filtering out irrelevant variations.
  • Evaluation: The learned feature representations are then evaluated using two methods:
  • Classification: A Support Vector Machine (SVM) is trained on the features to assess how well known fault types can be distinguished, particularly under changing operating conditions.
  • Clustering: Algorithms like OPTICS and k-means are used to group the features. This is crucial for identifying novel fault types, as they should ideally form distinct clusters or be identifiable as outliers.

Datasets

The study uses the Case Western Reserve University (CWRU) Bearing Dataset. This is a publicly available benchmark dataset commonly used in Prognostics and Health Management (PHM) research. It contains accelerometer measurements from bearings under various loads (representing operating conditions) and health states, including healthy conditions and different types and severities of faults (inner race, outer race, ball faults). The authors carefully select subsets of this dataset to simulate scenarios with novel operating conditions and novel fault types.

Key Insights and Ideas

  • Contrastive Learning for PHM is Promising: The paper demonstrates that contrastive learning, particularly with triplet loss, can learn feature representations that are remarkably robust to changes in operating conditions.
  • Invariance vs. Sensitivity Trade-off Addressed: The key insight is that contrastive learning can achieve both invariance to normal operational variations (like temperature or load shifts) and sensitivity to novel fault signatures, a difficult balance to strike with traditional methods.
  • Low-Dimensional Features are Key: Using a constrained, low-dimensional feature space (16D) forces the encoder to learn more semantically meaningful representations, effectively filtering out noise and irrelevant variations.
  • Clustering for Novelty Detection: While classification is good for known faults, clustering proves essential for identifying novel fault types. The learned features allow for better separation, enabling the detection of previously unseen faults as distinct clusters or outliers.

Key Results

  • Invariance to Operating Conditions: The triplet encoder achieved near-perfect classification accuracy (100%) on test data with novel operating conditions, showing minimal degradation compared to data from training conditions. Other methods, like autoencoders and FFT-based features, showed significant accuracy drops.
  • Sensitivity to Novel Fault Types: In Case Study 2, the triplet encoder's feature space, when clustered using OPTICS, identified a high percentage (94%) of outliers as belonging to novel fault types, with a low false alarm rate. This suggests it can effectively flag unknown issues.
  • Clustering Performance: While both OPTICS and k-means were used, OPTICS showed a particular advantage in identifying novel faults as outliers when applied to the compact feature spaces learned by the triplet encoder.

Limitations and Future Work

  • Dataset Balance: The effectiveness of the outlier detection approach for novel faults relies on a reasonable balance between known and novel data in the training and testing sets. If the novel data heavily outweighs the known data, novel faults might be misclassified as new "normal" conditions rather than outliers.
  • Semi-supervised/Unsupervised Settings: The authors suggest future work should explore how contrastive learning can be applied in more realistic semi-supervised or fully unsupervised settings, which are common in industrial PHM where labeled data is scarce.
  • Hyperparameter Sensitivity: While efforts were made to tune hyperparameters, the performance of clustering algorithms can still be sensitive to these choices, especially for density-based methods like OPTICS.

Summary updated

Abstract

Reliable fault detection and diagnostics are crucial in order to ensure efficient operations in industrial assets. Data-driven solutions have shown great potential in various fields but pose many challenges in Prognostics and Health Management (PHM) applications: Changing external in-service factors and operating conditions cause variations in the condition monitoring (CM) data resulting in false alarms. Furthermore, novel types of faults can also cause variations in CM data. Since faults occur rarely in complex safety critical systems, a training dataset typically does not cover all possible fault types. To enable the detection of novel fault types, the models need to be sensitive to novel variations. Simultaneously, to decrease the false alarm rate, invariance to variations in CM data caused by changing operating conditions is required. We propose contrastive learning for the task of fault detection and diagnostics in the context of changing operating conditions and novel fault types. In particular, we evaluate how a feature representation trained by the triplet loss is suited to fault detection and diagnostics under the aforementioned conditions. We showcase that classification and clustering based on the learned feature representations are (1) invariant to changing operating conditions while also being (2) suited to the detection of novel fault types. Our evaluation is conducted on the bearing benchmark dataset provided by the Case Western Reserve University (CWRU).
BibTeX
@inproceedings{rombach2021contrastive,
  author = {Katharina Rombach and Gabriel Michau and Olga Fink},
  title = {{Contrastive Learning for Fault Detection and Diagnostics in the Context of Changing Operating Conditions and Novel Fault Types}},
  booktitle = {Sensors},
  year = {2021},
  doi = {10.3390/s21103550},
}
Figure