HMDM: A hybrid GM model-based and MHSA-GRU data-driven method for remaining useful life prediction of rolling bearings

Abstract

Rolling bearings have been extensively used to reduce friction and restrain relative movement between moving parts. However, failures such as wear or fatigue can be caused under continuous operating conditions, which results in extreme economic losses and severe injuries. The failure rate of rolling bearings can be reduced by the remaining useful life prediction, which can perform effective predictive maintenance to reduce the risk of bearing failures and improve reliability. Machine learning (ML) has inaugurated new approaches to implement RUL prediction of rolling bearings, which is essential to the machinery prognostic and health management (PHM) system. Unfortunately, many existing ML, particularly supervised learning methods, have the following drawbacks. On the one hand, the independent identically distributed law between the training data and test data is always difficult to satisfy, which will induce poor generalization and unsatisfactory prediction results. The amount of degradation data and the corresponding labels are increased with the degenerative process, which costs a lot of time and effort for label setting. On the other hand, most of them are incompetent for the task of RUL estimation with insufficient degradation information and the prediction accuracy is often affected by noise. To deal with the above issues, a hybrid model-based and data-driven model (HMDM) architecture that combines the grey model (GM) and gated recurrent unit (GRU) with multi-head self-attention (MHSA) is proposed. The raw failure signal of the rolling bearing is preprocessed by the singular value decomposition (SVD) method for noise reduction. A novel model learning method is applied to train models, which compensates for deficiencies of prediction errors caused by distribution differences between training data and testing data, as well as avoiding the problem of sample labeling to ease the economic burden and labor waste. The characteristics of the GM and MHSA-GRU are inherited by HMDM, which has excellent prognostic capacity at a variety of time sequence lengths and degradation information with different frequency changes. The prediction of the proposed model is verified by the experimental data from the PRONOSTIA platform and XJTU bearing dataset, and a comparison with other methods is conducted to reveal the merit of HMDM on the performance of RUL prediction. The results show that high prediction accuracy of time series with different lengths and frequency changes can be achieved by the proposed HMDM model, which outperforms the existing ML methods and provides a new solution for RUL prognostics of rolling bearings.