Using Machine Learning (ML) to Predict Potential Failures from Harmonics Historical Data of The Broken Motor

Abstract - A predictive maintenance strategy is proposed for a 0.37 kW, 380 V AC induction motor affected by internal anomalies such as broken rotor bars. The main objective is to identify potential faults using machine learning (ML) algorithms trained on harmonic features derived from stator current signals. Data acquisition was performed across three motor conditions: Healthy, Broken 1 Bar (BB1), and Broken 2 Bars (BB2), operating under three-phase current. The Fast Fourier Transform (FFT) technique was used to extract the 1st , 3rd, 5th, and 7th order harmonic components from each phase (R, S, T), generating twelve features per data sample. To ensure dataset reliability, a harmonic-based filtering method was applied by enforcing frequency thresholds at 250 Hz and 350 Hz to select valid Healthy samples. The processed data were then used to train three machine learning models: Linear Regression (LR), Random Forest (RF), and Decision Tree (DT). Evaluation results demonstrated excellent classification performance, with both Random Forest and Linear Regression achieving approximately 99% accuracy, while Decision Tree reached a perfect score of 100%. This synergy between spectral analysis and machine learning offers a reliable and non-invasive approach for real-time fault prediction, making it highly applicable for industrial monitoring environments.