Intelligent Predictive Maintenance for induction motor Using Deep Learning in Industrial AC

Abstract

Predictive maintenance of industrial equipment is necessary for reducing unplanned downtime, reducing repair expenditures and maintaining the competitiveness of a firm. Spontaneous break-downs of induction motors can cost millions in lost revenues. A conventional reactive maintenance approach and time based preventive maintenance techniques can be too late to be useful or wasteful of resources, whereas today’s data-driven predictive maintenance PdM can predict faults before they happen. We have created and evaluated three deep-learning architectures for PdM of three phase induction motors a pure temporal long short-term memory (LSTM) network, a fuzzy-logic adaptive network based fuzzy inference system (ANFIS) model, and a new end-to-end hybrid of ANFIS and LSTM. We built six-sensor data streams (three-phase currents, three-phase voltages, vibration, temperature and speed) into historical and current time sliding windows, and marked four fault classes: normal, bearing, mechanical, and winding. After separating train/test partitions in a stratified manner, we trained and evaluated based on accuracy, precision, recall, F1-score, and confusion matrix. The hybrid ANFIS+LSTM model combines interpretive fuzzy inference and rapid temporal feature learning in such a way as to both rapidly converge and robustly generalize, reached near perfect fault detection performance. By combining ANFIS and LSTM in an end-to-end design, we were able to combine fast temporal feature learning with interpretable fuzzy rules. The model achieved 98% convergence in fault detection, with rapid convergence and consistent accuracy across all types