Multi-Sensor Based Remaining Useful Life Prediction of Bearing Motors: A Comparative Study of LSTM and CNN Models

Abstract

Accurate Remaining Useful Life (RUL) prediction is essential for implementing effective predictive maintenance strategies in industrial rotating machinery. Bearing motors are particularly critical components whose unexpected failure may cause severe production losses and safety risks. This study presents a comparative investigation of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) architectures for RUL prediction using multi-sensor monitoring data. The dataset consists of 1000 days of simulated operational data from three bearing motors under varying degradation conditions. Five sensor parameters are considered: vibration (RMS), acoustic emission, temperature, stator current, and rotational speed (RPM). After preprocessing and sliding-window segmentation, 2910 time-series sequences were generated and divided into training, validation, and test sets. Model performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²). Experimental results show that LSTM significantly outperforms CNN, achieving an R² of 0.9877 on the test dataset, while CNN achieved R² below 0.34. The findings confirm the importance of temporal dependency modeling in long-horizon degradation prediction and provide guidance for selecting deep learning architectures in predictive maintenance applications.