Abstract
The operational continuity of electric motors is essential for industrial productivity, as unexpected failures result in economic losses and safety risks. This study proposes a predictive diagnostic system based exclusively on Motor Current Signature Analysis (MCSA) with on-device inference using TinyML, targeting resource-constrained environments. The design includes current signal acquisition through a non-invasive transducer, analog conditioning, preprocessing via root mean square calculation in overlapping windows and normalization, and the training of a lightweight one-dimensional convolutional neural network optimized for microcontroller execution. The prototype was evaluated using a class-balanced dataset, applying standard classification metrics and resource usage profiling. The results show perfect discrimination between normal and abnormal conditions associated with power electronics disturbances, with inference times compatible with real-time monitoring and low memory consumption. It is concluded that MCSA, combined with edge inference, is a viable and low-cost alternative for predictive maintenance, particularly in facilities with infrastructure limitations, and that its integration into multivariable systems could expand coverage to mechanical failure modes.

This publication has 29 references indexed in Scilit: