Remaining Useful Life Prediction for Bearings in Highway Machinery Equipment Based on ICEEMDAN-ISOMAP-DCLSTM Framework

Accurate remaining useful life (RUL) of rolling bearings constitute a critical technical challenge for operational safety assurance as highway infrastructure transitions to lifecycle maintenance. The accuracy of data-driven models for the RUL prediction of rolling bearings is heavily reliant on the feature extraction from raw vibration signals and the optimal selection of features. To address this challenge, this study introduced a novel hybrid model that integrates the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and the Isometric Feature Mapping (ISOMAP) method with the Divide-and-Conquer Long Short-Term Memory Method (DCLSTM) for RUL prediction under variable working conditions. Initially, the ICEEMDAN algorithm was utilized to mitigate modal aliasing during the signal decomposition process and to perform noise reduction; subsequently, time-domain and frequency-domain features of the denoised vibration signals were extracted. The ISOMAP method was then employed for feature dimensionality reduction and reconstruction, with an emphasis on minimizing reconstruction error and residual variance. Finally, DCLSTM was established and evaluated against other commonly used neural network models using benchmark bearing datasets. Compared to the second-ranked LSTM model, the proposed method achieved reductions of 37.42%, 40.62%, 34.39% in Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (sMAPE), and an enhancement of 1.28% in Nash-Sutcliffe Efficiency coefficient (NSE). The experimental results demonstrated that this approach can accurately and effectively predict the RUL of bearings, providing a new technical avenue for the health management of bearings.