COC-DAT: a contrastive learning-based dilated attention temporal network for hypersonic flight vehicle fault diagnosis

Abstract Attitude stability during the reentry phase of hypersonic flight vehicle (HFV) relies heavily on aerodynamic control surfaces. Any failure in these components can severely compromise mission safety. However, the acquisition of fault data is prohibitively expensive, and the types of faults are diverse. Existing methods show suboptimal diagnostic performance under these conditions. This article proposes a contrastive learning-based approach for diagnosing faults in aerodynamic control surfaces. The diagnosis task is defined as a multi-class classification problem using multivariate state time series. To this end, a dilated attention temporal network (DAT-Net) is developed within the contrastive orthogonal-center framework (COC-Framework), forming the COC-DAT-Net. The proposed network combines dilated convolutions with attention mechanisms for feature extraction. It employs an orthogonal-center contrastive learning strategy to optimize the structure of the feature space. In addition, a dynamic sampling mechanism is introduced to improve the quality of sample pair construction. Simulation experiments demonstrate that the proposed method performs well in fault diagnosis of aerodynamic control surfaces. It achieves an average F1-Score of 94.04% and a silhouette score of 81.43%. These results confirm the effectiveness of COC-DAT-Net in multi-class and small-sample scenarios. The method provides methodological support for enhancing the safety of HFV reentry.