In recent years, the application scenarios of Unmanned Aerial Vehicles (UAVs) have become increasingly widespread. As core components of UAVs, rotor blades’ health status is directly related to flight safety. Aiming at issues such as insufficient feature extraction, weak noise resistance, and low diagnostic accuracy in the crack fault diagnosis of UAV rotor blades, this study proposes a one-dimensional deep network integrating multi-scale convolution, a multi-path residual module, BiLSTM, and a self-attention mechanism, referred to as 1dMC-MPR-SABinet. Taking the triaxial (X, Y, Z) vibration signals of rotor blades as input, the method integrates a multi-scale convolution module and a multi-path residual module, models the bidirectional temporal dependencies of signals through Bi-LSTM, and is combined with a self-attention mechanism to enhance the capture of subtle fault features. Meanwhile, it adopts the Northern Goshawk Optimization algorithm to optimize hyperparameters, thereby improving stability in noisy environments. Experiments are validated based on a self-collected fault vibration dataset, with precision, recall, and F1-score as evaluation metrics. The results show that the proposed model achieves a diagnostic accuracy of 99.37% under noise-free conditions without NGO-based hyperparameter optimization, representing a maximum improvement of 6.50% over the comparative models. Under a strong noisy condition with SNR = 1, the base model achieves 91.95% accuracy, while after NGO-based hyperparameter optimization, the model performance is further improved, with the precision, recall, and F1-score reaching 97.64%, 97.78%, and 97.01%, respectively. Ablation experiments and generalization experiments further verify the rationality and effectiveness of the proposed architecture.
Zhang et al. (Fri,) studied this question.