Enhancing Industrial Predictive Maintenance through Anomaly Detection in Multivariate Sensor Data using Machine Learning Techniques

Abstract

This research looks at the approach of using machine learning algorithms in identifying anomalies in sensor data to aid in the development of predictive maintenance strategies in industrial systems. To diagnose abnormalities in the multivariate time series data derived from the industrial sensor, different techniques of the machine learning models comprising – supervised learning models, unsupervised learning models, and the hybrid models were designed and compared by the authors. The evaluation of the performances of the developed models shows that the predictive model based on the combination of LSTM networks and Isolation Forest has got the highest accuracy of 95. 7% and F1-score of 0. 93 in the anomalies’ identification. The proposed model can decrease the frequency of unplanned downtime to the company’s operations as well as the cost of maintenance. The findings show that using machine learning techniques for anomaly detection can improve the effectiveness and availability of industrial systems due to the implementation of predictive maintenance.