DTST: A Dual-aspect Time Series Transformer Model for Fault Diagnosis of Space Power System

Fault diagnosis is one of the key technologies for maintaining the reliability and safety of space power systems. High-precision fault diagnosis is crucial to ensuring the normal operation of the system. In recent years, fault diagnosis methods based on traditional deep learning models have matured, but these models have problems capturing long distance dependencies in sequences and are limited to modeling in the temporal dimension. To address these challenges, this paper proposes a novel fault diagnosis method for space power systems, namely Dual-aspect Time Series Transformer (DTST). DTST first adopts a token sequence generation method to decompose the data into two sets of sequence tokens in the temporal and spatial dimensions. Then, by introducing the Transformer, it obtains class tokens for these two sets of sequence tokens and merges them into a global class token for performing fault diagnosis tasks. To validate the rationality of the DTST structural design, this paper conducts comprehensive experiments on the space power system dataset and real telemetry dataset. The experimental results show that, compared to single-structure models, DTST with a dual-structure design performs superiorly in diagnostic performance. Meanwhile, the fusion of dual-structure design has also been adequately demonstrated. Compared to traditional deep learning models and Transformer variant models, DTST demonstrates superior performance and robustness.