Lightweight Machine Learning for Edge-Based Machinery Fault Diagnosis
First author · ISRM 2026 · Springer Nature, 2027
Summary
Deploying machinery fault diagnosis on affordable edge devices is constrained by the memory, latency, and computational demands of heavyweight AI. This study presents a compact framework that combines frequency-domain signal features with optimized classical machine-learning models for fast, on-device diagnosis without relying on the cloud.
Key findings
Signal processing. Two-axis vibration signals sampled at 25.6 kHz are divided into non-overlapping 8,192-sample windows. FFT is then used to extract ten compact spectral features.
Lightweight models. An optimized Random Forest and Softmax Regression classify five machine conditions: normal, rotor unbalance, shaft misalignment, inner-race fault, and outer-race fault.
Results. The 10-tree Random Forest achieves 99.52% test accuracy using 194,380 bytes of RAM, with an estimated latency of 33.6 ms per window on an ESP32. Optimization cuts both memory use and inference latency by more than 50%.