Real-Time Fault Diagnosis of Induction Motors Based on Stator Current Analysis and Embedded Machine Learning on a Cost-Effective Testing Platform
Induction motors are central to industrial processes, yet their unexpected failure incurs significant production losses and maintenance costs. Motor Current Signature Analysis (MCSA) is a well-established non-intrusive technique for identifying electrical and electromechanical faults via frequency-domain analysis of the stator current. However, manual spectral interpretation remains challenging under low signal-to-noise conditions and variable operating regimes. This paper presents a complete, reproducible framework integrating physics-based simulation, interpretable feature engineering, and lightweight machine learning for real-time supply-fault diagnosis. A three-phase squirrel-cage induction motor is modelled in MATLAB/Simulink to validate MCSA sideband signatures under healthy, phase-loss, and voltage-imbalance conditions. A cost-effective test bench equipped with an ACS712 Hall-effect sensor and an Arduino UNO microcontroller acquires stator current data from 15 independent acquisition runs, yielding a balanced dataset of 1440 fixed-length windows (480 per class). All experiments are conducted under no-load conditions (slip $\approx \mathbf{0. 0 2})$, which represents a conservative lower bound on in-service performance since sideband energy grows with load-induced slip. A 1024-sample Hann-windowed FFT extracts an eight-dimensional feature vector combining frequency-domain indicators and time-domain statistics. Random Forest and XGBoost classifiers are evaluated under random-window and leakage-controlled group-aware protocols. Under the group-aware split, XGBoost achieves 91.3% (±1.4%) accuracy and a macro-F1 of $\mathbf{9 0. 7 \%}(\mathbf{\pm 1. 6 \%})$; Random Forest achieves $\mathbf{8 9. 6 \%}(\mathbf{\pm 1. 8 \%})$ accuracy and a macro-F1 of 88.4% (±2.1%). A McNemar test confirms the performance gap between classifiers is statistically significant $(\mathbf{p}<\mathbf{0. 0 5})$. An explicit offline-real-time benchmark reports end-to-end latency (280-295 ms), fault-to-alarm detection delay (~1.8 s), and false-alarm behaviour under a persistence-rule controller, bridging the gap between dataset-level accuracy and embedded deployment constraints.