Fault Diagnosis of UAV Based on Adaptive Siamese Network with Limited Data

The multi-channel data of UAV is coupled and limited. Traditional diagnostic methods relying on precise mathematical models and expert knowledge are challenging to complete the fault diagnosis of UAV with limited data. In addition, noise and distortion in multi-channel data can lead to the degradation of data quality and has an influence on the fault diagnosis method based on deep learning. To overcome these deficiencies, an adaptive siamese network is constructed by two paralleled feature extractors with shared weights. Compared with other networks, the proposed network takes advantage of channel attention mechanism to weigh the features of different channels adaptively. Therefore, the proposed network can adaptively enhance feature weights based on channel importance and reduce the coupling effect of multi-channel data. In addition, the utilization rate of sample data can be improved by random and repeated extraction, and the problem of limited data can also be effectively solved. Finally, the proposed siamese network is applied to fault diagnosis of UAV with limited data, and the effects of the channel number, training samples, and noise on diagnosis are discussed. The experiment results show that the proposed siamese network has a high accuracy, low data dependency, and strong anti-noise. The obtained research results provide an effective method for solving the coupled multi-channel data with limited data in the fault diagnosis of UAV.