Remaining Useful Life Prediction Method Based on the Spatiotemporal Graph and GCN Nested Parallel Route Model

In the past few years, deep learning techniques for predicting remaining useful life (RUL) have shown remarkable advancements, but model prediction accuracy and generalization to different data still need to be improved. Moreover, the complex interactions among high-dimensional variables within multidimensional time series data (MTSD) can have an impact on predictive outcomes. To address these challenges, a RUL prediction approach based on the spatiotemporal graph and the graph convolutional network nested parallel route (GCN-PR) model is proposed for high-end equipment components. The multidimensional feature correlation spatiotemporal (MFCST) graph is constructed to implement feature extraction for data in different formats. In the model, one pathway utilizes graph convolutional network embedded long short-term memory (GELSTM) networks to acquire the feature structure spatial pattern of high-dimensional variables. In the other parallel route, stacked long short-term memory (ST-LSTM) network is employed to comprehensively explore local and global time patterns of MTSD. The spatial and temporal patterns are then strategically weighted to enhance the generalization ability of the model and the spatial perception of the feature structure of high-dimensional variables. The effectiveness of the proposed method was tested using datasets from engines, bearings and pantographs. When benchmarked against other mainstream techniques, the proposed method achieved a notable improvement in root mean square error (RMSE) reduction, reaching up to 25.94%, 34.50%, and 56.04% respectively on these datasets. These results demonstrate that the proposed method has the potential for enhancing prediction accuracy in diverse practical settings.