PRIMUS: A convolutional model for predictive maintenance of induction motors

Abstract

In the Fourth Industrial Revolution, there is a growing use of digital technologies for the continuous monitoring of industrial equipment, aiming to anticipate failures, prevent unplanned shutdowns, and increase the efficiency of maintenance execution. However, traditional monitoring methods often prove insufficient to prevent the occurrence of electrical or mechanical failures, resulting in significant financial and strategic losses for large industries. Three-phase induction motors are extremely important in numerous industrial processes, being responsible for converting electrical energy into mechanical motion in conveyor belts, ventilation systems, pumps, and compressors. Their robustness, constructive simplicity, and low maintenance cost have consolidated their massive presence in the global industrial landscape. Nevertheless, continuous operation under variable load and environmental conditions makes them susceptible to progressive failures, such as bearing wear, misalignment, electrical imbalance, and winding short-circuits. The early detection of these anomalies is a critical challenge for ensuring operational reliability. In this context, this work presents PRIMUS (PRedictive maintenance for Induction Motors Using Sound), a model based on a Deep Convolutional Neural Network (CNN) designed for predictive maintenance in industrial induction motors. The system is capable of monitoring the equipment's operation in different starting and operating architectures, specifically in delta and star connections, configured in both series and parallel, using exclusively acoustic signals as a diagnostic source. The audio recordings that compose the database were performed in a three-dimensional circular format, with the microphone positioned at a constant distance of 15 centimeters from the motor shaft, ensuring standardization in the capture of characteristic sound signals for each operational condition. The main contribution of this research is the development of a monitoring method for industrial induction motors that, compared to traditional techniques, does not require intrusive sensors or multiple data sources for its operation. The proposed system uses a single source of information, the motor's acoustic signals, combined with advanced deep machine learning methods. This approach results in a low-cost solution, both for implementation and operational maintenance, offering a practical and scalable alternative for predictive maintenance in Industry 4.0.