Encoding Techniques on Multivariate Time Series Signals for Failure Prevention of Industrial Assets with Unsupervised Deep Anomaly Detection

Abstract

This paper addresses the use of encoding techniques of time series data in a Predictive Maintenance (PdM) system for failure prevention of industrial assets. Baseline is a dataset of an electric motor from an industrial application. Goal of this work is to compare the effectiveness of the different encoding techniques with the raw time series using an unsupervised deep learning approach for anomaly detection in multivariate time series based on a Convolutional Autoencoder (CAE). The encoding techniques investigated here are Gramian Angular Field, Markov Transition Field and Recurrence Plot. For this task, 37 experiments have been realized. It has been demonstrated that anomaly detection with almost all of the encoding techniques performed better than with the use of raw time series. Finally, the proposed approaches have been evaluated and further potential research work in this area has been pointed out.