Development and validation of a deep neural network algorithm for predicting critical environmental parameters in three-phase induction motors: an experimental study under controlled environments

Abstract

Predictive monitoring of three-phase induction motors is a key challenge in industrial environments due to the complex interactions between thermal, electrical, and mechanical variables. This study proposes and validates a multivariate Long Short-Term Memory (LSTM) neural network model for the simultaneous prediction of stator temperature, bearing temperature, and RMS vibration using experimental data collected under controlled conditions. The methodology includes preprocessing, leakage-free normalization, temporal sequence construction using sliding windows, and model training with temporal validation. Performance was evaluated using MAE and RMSE metrics, complemented by residual analysis, bootstrap confidence intervals, and time-series cross-validation. Results show that the LSTM model achieves consistent performance, with errors of approximately 6.8–7.0 °C for thermal variables and 0.83 mm/s for vibration. Comparison with traditional models indicates that Random Forest outperforms LSTM in thermal prediction, while LSTM provides advantages in temporal modeling and vibration prediction. The model exhibits a smoothing effect, limiting its ability to capture high-frequency fluctuations. Hypothesis testing revealed no statistically significant improvement over a baseline model. Overall, the findings suggest that model performance depends on dataset characteristics, highlighting the potential of hybrid approaches for industrial applications.