Machine Learning Applied to Anomaly Detection in Electric Motors Aimed at Industry 4.0

In this work, we propose an anomaly detector for electric motors based on an Autoencoder implemented on a lowcost, computationally efficient platform. The model was trained on a dataset of 1,225 normal vibration signals, augmented with Gaussian noise. Through topological optimization using genetic algorithms, a compact Autoencoder with 10,336 parameters was obtained, achieving average sensitivity and specificity above 80% for anomaly detection. The model can operate in realtime (reaching 1,990 μs per inference), classifying normal and anomalous vibration signals captured by an accelerometer based on microelectromechanical systems technology. The results demonstrate the feasibility of employing machine learning for cost-effective, in-situ predictive maintenance in electric motors. As future work, the implementation and validation of a complete online system are proposed, along with the development of adaptive methods to establish dynamic thresholds between normal and anomalous states.