A Bayesian Optimized Stacked Long Short-Term Memory Framework for Real-Time Predictive Condition Monitoring of Heavy-Duty Industrial Motors

In the era of Industry 4.0, condition monitoring has emerged as an effective solution for process industries to optimize their operational efficiency. Condition monitoring helps minimize unplanned downtime, extending equipment lifespan, reducing maintenance costs, and improving production quality and safety. This research focuses on utilizing Bayesian search-based machine learning and deep learning approaches for the condition monitoring of industrial equipment. The study aims to enhance predictive maintenance for industrial equipment by forecasting vibration values based on domain-specific feature engineering. Early prediction of vibration enables proactive interventions to minimize downtime and extend the lifespan of critical assets. A data set of load information and vibration values from a heavy-duty industrial slip ring induction motor (4600 kW) and gearbox equipped with vibration sensors is used as a case study. The study implements and compares six machine learning models with the proposed Bayesian-optimized stacked Long Short-Term Memory (LSTM) model. The hyperparameters used in the implementation of models are selected based on the Bayesian optimization technique. Comparative analysis reveals that the proposed Bayesian optimized stacked LSTM outperforms other models, showcasing its capability to learn temporal features as well as long-term dependencies in time series information. The implemented machine learning models: Linear Regression (LR), Random Forest (RF), Gradient Boosting Regressor (GBR), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Support Vector Regressor (SVR) displayed a mean squared error of 0.9515, 0.4654, 0.1849, 0.0295, 0.2127 and 0.0273, respectively. The proposed model predicts the future vibration characteristics with a mean squared error of 0.0019 on the dataset containing motor load information and vibration characteristics. The results demonstrate that the proposed model outperforms other models in terms of other evaluation metrics with a mean absolute error of 0.0263 and 0.882 as a coefficient of determination. Current research not only contributes to the comparative performance of machine learning models in condition monitoring but also showcases the practical implications of employing these techniques. By transitioning from reactive to proactive maintenance strategies, industries can minimize downtime, reduce costs, and prolong the lifespan of crucial assets. This study demonstrates the practical advantages of transitioning from reactive to proactive maintenance strategies using ML-based condition monitoring.