Multiscale PatchTCN-Mixer: A New Method for Extracting Spatial and Temporal Degradation Information in Remaining Useful Life Prognosis

Deep learning has made significant progress in predicting the remaining useful life (RUL) of rotating equipment, but there is still room for improvement in the accuracy and generalization of models. Existing RUL prognosis models often concentrate only on the degradation information of individual time series, neglecting the utilization of spatial information at the subseries level. This article introduces a novel patch temporal convolutional network (PatchTCN) layer model built upon convolutional neural networks (CNNs), capable of segmenting time series into patch level subseries and fully extracting the potential degraded semantic information both between and within these patches. Moreover, a patch adaptive semi-shrinkage (PAS) block is proposed, which eliminates the spatial and temporal redundant information within patches to enhance the model’s predictive accuracy. Subsequently, the multiscale PatchTCN-Mixer (MSPT-Mixer) model is developed. It initially segments the time series into multiple patch scales to extract spatial and temporal semantic information, followed by single-scale fusion to fully capture both local features and the global correlation of degradation trends. The IEEE PHM2012 Prognostic Challenge dataset and the Xi’an Jiaotong University (XJTU)-SY Bearing dataset were used to validate and analyze the MSPT-Mixer model, comparing it with the optimal values of contemporary state-of-the-art models. The findings revealed that the model improved the root mean square error (RMSE), mean absolute error (MAE), and Score metrics of RUL prediction by 18.7%, 16.9%, and 17.4%, respectively, which demonstrates that the proposed model effectively extracts spatial and temporal semantic information from degraded data, thereby enhancing prediction accuracy.