Small Sample Fault Diagnosis for UAV Based on Siamese Network with Multiple Similarity Loss

The number of UAV fault samples is limited. Training sample pairs pose a redundancy challenge when constructing sample pairs to alleviate the problem of scarce training information. To this end, a Siamese network based on a generalized sample weighting framework with multiple similarity loss is proposed to extract useful information. It compares the similarity of sample pairs with a preset similarity threshold to determine the sample pairs for subsequent loss calculations. Moreover, the weights are adjusted based on the similarity of sample pairs, assigning larger weights to sample pairs with higher similarity to handle redundancy flexibly. Experimental studies have shown that the proposed method effectively alleviates redundancy and performs well in small sample intelligent fault diagnosis for UAVs.