A Dual Transformer–Based Feature Fusion Network for Gas Turbine Engine Fault Diagnosis

Accurate fault diagnosis plays a crucial role in implementing condition-based maintenance for gas turbine engines, enhancing their operational reliability and mitigating associated costs. Despite notable advancements achieved by deep learning–based diagnostic approaches, challenges remain in effectively managing multisensorial long time-series data when conducting gas-path (GP) fault diagnosis: (1) the conventional learning paradigms are insufficient for processing long time-series data; (2) current diagnostic methods often neglect or inadequately exploit the inherent correlations between multisensorial data channels; and (3) the reliance on intricate preprocessing techniques often compromises the end-to-end fault diagnosis efficiency. To address these issues, this paper proposes a dual transformer–based feature fusion network to extract the spatiotemporal features of GP fault data to obtain a more comprehensive reflection of engine operational conditions. First, a data preparation layer was designed to directly manipulate raw data, where an adaptive convolutional filtering mechanism is employed for better antinoise performance and wider attention to both global and local information. Two enhanced encoders were then set to extract the temporal and spatial correlations from the input GP fault data in parallel. Finally, a feature fusion layer is presented to adaptively combine spatiotemporal features for classification, further improving the feature quality and model efficiency. The effectiveness of the proposed method was verified on GP fault data sets of a two-shaft turbofan engine, achieving an accuracy of 99.47% in offline training and an average latency of 0.0064 s in online testing. The influence of hyperparameters on the diagnostic performance, computational complexity, and number of parameters was analyzed in detail. The results demonstrate the wide applicability and practical value of the proposed model for multisensorial GP fault data analysis.