Remaining useful life prediction for rolling bearings based on adaptive aggregation of dynamic feature correlations

It is significant to predict the remaining useful life (RUL) of the bearing to ensure its safe and stable operation. At present, the data-driven method has been successfully applied in the field of bearing RUL prediction. However, the feature correlations between data at different moments may be different, few methods can dynamically identify the change of the feature correlations between input data at different moments, which can impact the performance of the prediction. This article proposes an innovative RUL prediction method based on the adaptive feature correlations aggregation module (AFCA) and gated recurrent unit (GRU) to address this issue. First, statistical features are extracted from the vibration signal, and the fully connected graph is constructed to map the vibration signal data into the graph structure. Subsequently, the AFCA module is designed and constructed, and the AFCA-GRU model is built by combining GRU. A series of constructed fully connected graphs are fed into the model, and the hidden degradation information in graph structure data is mined to realize the prediction of bearing RUL. Among them, AFCA is used to adaptively explore the spatial correlations between graph node features at different moments, and GRU is used to explore the temporal correlations between graph structures. The PHM2012 Challenge dataset is utilized to validate the effectiveness of the proposed method. The comparative experimental results demonstrate that the performance of the method proposed herein surpasses that of other data-driven methodologies, with the capability to accurately predict the RUL of bearings.