A Multi-Source Fusion Spatiotemporal Neural Network Improved by Koopman Operators for Predicting Remaining Useful Life

Abstract

: The operation of complex equipment is typically monitored by multiple sensors, and the vast amount of status data generated from this monitoring provides strong support for predicting the remaining useful life (RUL). Due to the influence of unstable operational conditions, the degradation trajectory of the equipment often exhibits a high degree of nonlinearity. Conventional approaches for processing univariate time series data often struggle to effectively identify inherent degradation trends and unstable fluctuations, while exhibiting limited capability in comprehensive modeling of multi-source time series data. This paper proposes a novel spatiotemporal neural network for RUL prediction. Firstly, a temporal decomposition block (TDB) is utilized to decompose the multi-source time series into trend and unstable components. Subsequently, temporal dependency features are extracted using gated recurrent units (GRU), and the Koopman operator is employed to linearly model these features in a high-dimensional space. A channel interaction learning block (CILB) is applied to capture dependencies between sensors and enhance feature representation capabilities. Finally, the prediction module utilizes the linear layer of residual structure to generate the final RUL prediction result, and a method combining ensemble learning and kernel density estimation (KDE) is used to obtain the probability density function of the RUL. The experimental results based on the C-MAPSS dataset show that the prediction accuracy of this method is superior to other existing methods, especially exhibiting better performance under complex