Real-Time Fault Detection in Induction Motors Using TinyML: An Evaluation of the Edge Impulse Platform

This article evaluates the Edge Impulse platform for developing TinyML-based fault detection models in induction motors, focusing on identifying broken rotor bars using vibration data collected from the motor casing. The study aims to assess the platform’s capabilities in creating non-intrusive, real-time detection systems implemented on low-power microcontrollers, such as the Arduino Nano 33 BLE, for resource-constrained industrial environments. Using the platform’s tools, an artificial neural network model was trained with real vibration data and optimized for embedded implementation. The model achieved 98.6% accuracy for detecting four broken bars and 95.8% accuracy for the more complex case of detecting a single broken bar, operating within the memory limits of the device and maintaining low detection latencies. Quantization, applied through Edge Impulse’s workflow, successfully reduced latency to 1 ms for the four-bar fault without significantly compromising accuracy, but it caused a substantial accuracy drop for the single-bar fault, highlighting limitations in handling more complex scenarios. This evaluation demonstrates the practicality and constraints of using the Edge Impulse platform for industrial IoT applications, particularly in designing TinyML solutions for fault detection.