A multi-scale convolutional Siamese network for few-shot fault diagnosis of unmanned aerial vehicle rotor

Unmanned Aerial Vehicles (UAVs) often face infrequent fault occurrences and high manual annotation costs, resulting in a critical shortage of valid fault samples for diagnostic research. Traditional fault diagnosis methods struggle with small sample sizes. This paper proposes a novel deep metric learning method, the Multi-Scale Convolutional Siamese Network (MSCSN), to address the few-shot learning problem in UAV rotor fault diagnosis. First, discrete wavelet transform (DWT) is used to compress and normalize the vibration signals, enhancing the prominence of signal features. Then, based on the multi-scale convolutional neural network (MS-CNN) model, the network automatically extracts multi-level features from rotor fault vibration signals, improving its adaptability to complex data. Finally, the Siamese network structure, with shared parameters and identical architecture, processes sample pairs and incorporates a small number of support samples for few-shot learning. Experimental results show that the proposed model achieves a highest accuracy of 94.42 % in few-shot tasks. In cross-domain transfer learning tests, the model achieves an average accuracy of 90.28 %, demonstrating its superior generalization ability and robustness across different environments. We also validated the model's stability using the publicly available MVS-UAV-BF dataset.