Few-shot fault diagnosis for bearing based on lightweight siamese networks
WANG Qingnan
CHEN Qian
LI Tao
TU Jihui
Abstract：To address the challenge of fault diagnosis under limited labeled data,this paper proposes a lightweight Siamese network tailored for few-shot bearing condition monitoring.Raw one-dimensional vi-bration signals are first converted into two-dimensional time-frequency images via Continuous Wavelet Transform,enabling effective denoising and salient feature representation.A metric learning paradigm constructs positive and negative pairs to enhance data diversity and model generalization.The network le-verages MobileNetV3S as the backbone,augmented with an Efficient Channel Attention mechanism to se-lectively emphasize informative channels while maintaining low complexity.Additionally,skip connections in deep Bneck modules are incorporated to improve feature propagation,and global average pooling re-places fully connected layers to further reduce parameters.Experimental validation shows that the pro-posed approach achieves 91.67%accuracy with only 1.16 M parameters and 0.05 GFLOPs,demonstrating superior performance in accuracy,efficiency,and real-time applicability under extreme sample constraints.
Keywords：bearing-diagnosisfew-shotsiamese-networkslightweight
Publication Date：2026-02-25
Online Publishing Date：2026-08-28(First online date of this platform, not the publication date of the document)
Pages：13( 11-23 )