Two-step residual-error based approach for anomaly detection in engineering systems using variational autoencoders. (July 2022)

• Record Type: Journal Article
• Authors: González-Muñiz, Ana; Díaz, Ignacio; Cuadrado, Abel A.; García-Pérez, Diego; Pérez, Daniel
• Is Part Of: Computers & electrical engineering. Volume 101(2022)
• Publication Date: 2022-07
• Subjects: Anomaly detection -- Novelty detection -- Deep autoencoder -- Variational autoencoder -- Engineering systems; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108065
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: Anomaly detection is a crucial task in the engineering systems field. However, there is usually little or no information about all possible abnormal modes in systems. Hence, a common approach is to build a model of healthy behaviour, based on normal operation data, so that anomaly detection would depend on how well new data fit this model. According to this idea, we propose a residual-error based approach consisting of: a variational autoencoder, used to model the probability density function of the system's healthy behaviour; and a two-step classification algorithm, which classifies the incoming samples based on their residuals, and reports not only their normal/anomalous nature but also that of their components. We have tested this proposal in three different engineering contexts and we have compared its performance with that of state-of-the-art approaches, demonstrating its capability to successfully detect and characterize anomalies. Highlights: Variational autoencoder (VAE) is used to detect anomalies in engineering systems. Anomalies are detected using a two-step classifier based on VAE residuals. The classifier indicates the nature of the samples and also that of their components. This classification provides an explainable diagnostic of the anomaly decision. The proposal has been tested in three different engineering contexts. Graphical abstract: