Anomaly Detection in Smart Industrial Machinery Through Hidden Markov Models and Autoencoders

This study addresses the need to develop a sustainable manufacturing process in industrial factories, as the industry desires to remain competitive while it is challenged to adopt eco-friendly practices. A Machine Learning based software is proposed to deal with the environmental issues, aiming to facilitate the monitoring and analysis of industrial machinery, more exactly of CNC woodworking machines. The focus is on two aspects that determine the environmental impact: energy consumption and toxic emissions, which are used to determine the operating modes of the machines and to detect potential working anomalies. This software consists of a pipeline with two main components: the first one aims to categorize the operating modes of the used machines through time series clustering methods, such as Hidden Markov Models. The second component employs Hidden Markov Models again alongside deep learning based Autoencoders to identify huge deviations within the environmental data. For evaluation, a dataset was collected as a time series from a CNC woodworking machine and then the preprocessed data was further analyzed using the implemented software. The experiments have shown that for anomaly detection in machine operating modes, the Hidden Markov Model outperforms the Autoencoder and state-of-the-art models in terms of efficiency and robustness.