Constraint Guided Autoencoders for Joint Optimization of Condition Indicator Estimation and Anomaly Detection in Machine Condition Monitoring

Abstract

The main goal of machine condition monitoring is, as the name implies, to monitor the condition of industrial applications. Effective asset monitoring relies on an accurate assessment of its condition, often represented by a Condition Indicator (CI). The CI should exhibit specific behaviors to ensure its reliability: (i) it should maintain a consistent range, enabling clear differentiation between normal and anomalous data, and (ii) it should demonstrate monotonic behavior over time, reflecting the expected gradual degradation of the asset’s condition. This work proposes an extension to Constraint Guided AutoEncoders (CGAE), which is a robust AD method, that enables building a single model that can estimate a CI that shows the aforementioned behaviors. To improve the monotonic behavior of the CI, the proposed extension incorporates a constraint that enforces this behavior during the training of the model. Experimental results, on two datasets containing run-to-failure data from bearings, indicate that the proposed extension retains the performance of CGAE with regards to CGAE, while also improving the monotonic behavior of the CI. Beyond the improved CI, an additional advantage of the proposed extension is its ability to leverage unlabeled data without requiring assumptions about the data. Furthermore, an ablation study revealed that reconstructing unlabeled data also contributed to enhancing the CI.