Optimized Machine Learning for Induction Motor Fault Diagnosis Using Vibration and Frequency-Domain Features

Abstract

This research presents an optimized machine learning framework for predictive maintenance in industrial motor systems, utilizing vibration signal analysis and ensemble-based modelling to enhance fault detection accuracy and reliability. A publicly available dataset from Mendeley Data, consisting of vibration signals recorded from a 1.0 kW induction motor under eight operational conditions, was employed to evaluate model performance. The study integrates both time-domain and frequency-domain feature extraction using Fast Fourier Transform (FFT), enabling the identification of critical vibration patterns associated with motor degradation. Three experimental scenarios were designed to assess the impact of feature selection, data balancing, and hyperparameter tuning. Various machine learning algorithms—including Random Forest, Gradient Boosting, AdaBoost, Light Gradient Boosting Machine (LightGBM), and CatBoost—were trained and optimized using GridSearchCV to determine the most effective model configuration. Model performance was evaluated through multiple metrics such as Accuracy, Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). The results demonstrate that the optimized CatBoost model achieved the highest classification accuracy of 87.28%, outperforming other classifiers while maintaining computational efficiency and compact model size. These findings confirm that the integration of FFT-based features, data balancing, and hyperparameter optimization significantly enhances model robustness, making the proposed framework suitable for real-time industrial IoT applications. The study provides a reproducible and scalable approach for intelligent predictive maintenance, supporting the transition toward data-driven and autonomous maintenance strategies in smart manufacturing environments.

-