RUL Prediction for Bearings Using MSCNN and LSTM Networks

Abstract

To address the challenges of effectively capturing spatiotemporal features from data using a single prediction model and the limitations of artificially constructed degradation indices in accurately representing the bearing's degradation state at specific time points, this study introduces a novel bearing Remaining Useful Life (RUL) prediction model. The model integrates a Multi-Scale Convolutional Neural Network (MSCNN) and Long Short-Term Memory (LSTM) networks, grounded in a monotonic degradation index. Initially, a monotonic optimality criterion is employed to identify an appropriate health index for bearing degradation. Subsequently, a comprehensive spatiotemporal feature set is developed by merging the multi-scale spatial features extracted by MSCNN with the temporal features derived from LSTM. The proposed LSTM-MSCNN model's effectiveness is validated using the XJTU-SY bearing dataset from Xi'an Jiaotong University as a case study.