Anomaly Detection in Asset Degradation Process Using Variational Autoencoder and Explanations

Abstract

Development of predictive maintenance (PdM) solutions is one of the key aspects of Industry 4.0. In recent years, more attention has been paid to data-driven techniques, which use machine learning to monitor the health of an industrial asset. The major issue in the implementation of PdM models is a lack of good quality labelled data. In the paper we present how unsupervised learning using a variational autoencoder may be used to monitor the wear of rolls in a hot strip mill, a part of a steel-making site. As an additional benchmark we use a simulated turbofan engine data set provided by NASA. We also use explainability methods in order to understand the model’s predictions. The results show that the variational autoencoder slightly outperforms the base autoencoder architecture in anomaly detection tasks. However, its performance on the real use-case does not make it a production-ready solution for industry and should be a matter of further research. Furthermore, the information obtained from the explainability model can increase the reliability of the proposed artificial intelligence-based solution.

Keywords

• machine learning
• deep learning
• anomaly detection
• hot rolling
• explainable artificial intelligence