Learning Informative Health Indicators Through Unsupervised Contrastive Learning

TL;DR

The paper tackles robust health indicator construction for complex assets by introducing an unsupervised contrastive learning framework that uses operation time as a degradation proxy. It learns invariant yet degradation-sensitive features via a triplet loss and derives a health indicator from the distance to an OC-SVM boundary, followed by smoothing. Evaluations on milling wear and railway wheel fault detection demonstrate improved health-tracking alignment with wear and competitive fault-detection performance, including early detection in partial-observability settings. The approach shows versatility across modalities and tasks, with potential for monotonicity enforcement and semi-supervised extensions to enhance applicability. Overall, the method provides a robust, unsupervised pathway to informative health indicators for diverse CM scenarios.

Abstract

Monitoring the health of complex industrial assets is crucial for safe and efficient operations. Health indicators that provide quantitative real-time insights into the health status of industrial assets over time serve as valuable tools for e.g. fault detection or prognostics. This study proposes a novel, versatile and unsupervised approach to learn health indicators using contrastive learning, where the operational time serves as a proxy for degradation. To highlight its versatility, the approach is evaluated on two tasks and case studies with different characteristics: wear assessment of milling machines and fault detection of railway wheels. Our results show that the proposed methodology effectively learns a health indicator that follows the wear of milling machines (0.97 correlation on average) and is suitable for fault detection in railway wheels (88.7% balanced accuracy). The conducted experiments demonstrate the versatility of the approach for various systems and health conditions.

Figures (8)

• Figure 1: Data acquisition for the railway wheel dataset: First in Fig. \ref{['fig:process']}, all data sources are linked to individual wheels ('Source'), resulting in a first data split into train and test ('Initial Split'). For the defective wheels, the time of defect initiation is provided by domain experts, as shown in Fig. \ref{['fig:annotation']}. The preliminary healthy label of the wheels in the test dataset is challenged by fault detection models and evaluated by domain expert feedback.
• Figure 2: Examples of railway wheel defects
• Figure 3: Illustration of the partial observation of railway wheels provided by a strain gauge sensor in a WSM measurement site.
• Figure 4: Wheel circumference regions monitored by the eight strain gauge sensors (blue regions) in dependency of different diameters compared to the entire wheel circumference.
• Figure 5: Illustration of the proposed framework: In a first step, an encoder model $f$ is trained on with the triplet loss function $L(x_a, x_p, x_n)$, whereby the selection of the data triplets $x_a, x_p$ and $x_n$ is based on the operation time. In a second step, the trained feature space is exploited to construct the health indicator for CM: wear assessment (upper right) and fault detection (lower right)