Anomaly detection method based on the deep knowledge behind behavior patterns in industrial components. Application to a hydropower plant

This paper describes a new methodology that aims to cover a gap detected in the area of detection of anomalies and diagnosis of industrial component behaviors: there is a need of robust procedures compatible with dynamic behaviors and degradations that evolve over time. The method proposed is based on the creation of behavior patterns of industrial components using well-known unsupervised machine learning algorithms such as K-means and Self-Organizing maps (SOMs) as a starting point. An algorithm based on local Probability Density Distributions (PDD) of the clusters obtained is used to enhance the characterization of patterns. The joint use of these algorithms facilitates a new way to detect anomalies and the surveillance of their progress. The paper includes an example of an application of the method proposed for monitoring the bearing temperature of a turbine in a hydropower plant showing how this method can be applied in behavior and maintenance assessment applications. The results obtained prove the advantages and possibilities that the proposed methodology has on real world applications. (c) 2020 Elsevier B.V. All rights reserved.

Automated summary

This study introduces a novel approach to anomaly detection and diagnosis of industrial component behaviors by creating behavior patterns using unsupervised machine learning algorithms, and enhancing pattern characterization through local Probability Density Distribution algorithm. This method allows for the surveillance of dynamic behaviors and degradation processes in real world applications.