Implementation of a Predictive Maintenance System using Unsupervised Anomaly Detection

Abstract

This work presents the implementation and architecture of a predictive maintenance system using unsupervised learning, aiming to provide an approach where its benefits are returned to the users faster than traditional system found in the literature. Equipment faults and maintenance affect directly efficiency of industrial plants. Maintenance management can be optimized by modeling and predicting problems. The implementation of a system able to model problems or anomalies and to inform operators and supervisors in case of alarms or notifications is intrinsically connected with the fourth industrial revolution. Although the architecture proposed here can be generalized to many different devices, a MPU6050 for sensing and ESP32 as the middleware were used to implement the proposed concept. Three different methods for anomaly detection were implemented and deployed to Google Cloud, using Compute Engine service. The developed web-server that provides a dashboard to visualize time-series of the sensed physical magnitudes and historical data about all anomalies detected is presented. Therefore, resulting in a platform with intelligence to detect and report problems and abnormalities to employees of industrial plants.