Application of a dual-branch recursive-time series joint model based on a multi-scale Transformer and multi-modal fusion to fault diagnosis of fixed-wing UAV actuators

Fixed-wing UAV actuators are prone to compound failures under complex multimodal dynamic environments, posing challenges to fault diagnosis accuracy and robustness. To address this, a dual-branch recursive-time-series joint model based on a multi-scale Transformer and multimodal fusion is proposed. The model employs a gated recurrent unit (GRU) branch to capture temporal dependencies, and a convolutional branch to extract spatial features from recursive graphs derived from time-series signals. A Transformer-based fusion module adaptively integrates six sensor modalities, including rudder angle, thrust, and gyroscope, enabling accurate modelling of nonlinear patterns and dynamic interactions. Data augmentation techniques, such as Gaussian noise, translation, and scaling, are introduced to improve generalisation. Experimental validation on 5,000 samples across one normal and ten fault modes shows that the proposed model achieves 95% F1-score and 98% AUC on a balanced dataset, and 91% F1 and 95% AUC on unbalanced data. Compared with a range of classical and recent deep learning baselines, including RF (81%), SVM (83%), RNN (85%), LSTM/GRU (88%), Transformer (91%), Transformer-XL (92%), and CNN-BiLSTM (89%), the proposed method exhibits superior classification performance, especially for rare and compound faults. Ablation studies confirm the critical role of each module. The results demonstrate the model's strong potential for real-time, reliable fault diagnosis in UAV actuator systems, offering theoretical and engineering value for autonomous flight safety and maintenance planning.