Multipath Siamese Neural Network for Intelligent Fault Diagnosis Under Small Samples

The scarcity of fault samples severely limits the performance of traditional fault diagnosis methods under small samples. To address this issue, this article proposes a multipath Siamese neural network (MSN)-based algorithm. MSN ensures the superior fault identification despite limited training samples by integrating information from multiple data types. Specifically, the MSN incorporates a multiscale convolutional module (MCM) and wide convolutional kernels to extract richer feature representations while minimizing the noise interference. In addition, a classification network is integrated into the Siamese network (SN) framework, enabling direct category predictions during testing instead of pairwise similarity computations. A metric neural network (MNN) is also introduced to automatically learn feature similarities during training, thereby improving the model’s generalization capability and accuracy. Experimental evaluations on two distinct datasets demonstrate that the MSN achieves accurate fault detection under small-sample conditions with the robust generalization performance.