Noise-Robust and Edge Deployable ML Framework for Predictive Maintenance of Industrial Motors

Industrial three-phase induction motors are critical to manufacturing operations, yet their failure can cause costly downtime and maintenance overhead. This work introduces a machine learning framework tailored for fault diagnosis and severity assessment in such motors, using synthetic data that replicates six realistic electromagnetic and mechanical fault conditions. The system employs synchronized vibration and current signals, and evaluates several classifiers-including LSTM, GRU, TCN, RNN, and Random Forest-under varying industrial noise profiles such as phase jitter, frequency drift, and transient spikes. Experimental results demonstrate that time-series models consistently outperform classical approaches in noisy environments. For deployment on resource-constrained edge devices, structured pruning and quantization are applied to reduce model size and latency. Beyond motor fault classification, the framework also detects and visualizes approximately spherical clusters to comprehend fault severity levels. The proposed system enables scalable fault monitoring and low-latency predictive maintenance optimized for embedded industrial environments.