PCASTNet: A Physics-Constrained Adaptive Style Transfer Network for Sample Generation in Cross-Machine Small-Sample Fault Diagnosis
Xiaoxi Hu, Jingh Ao Li, Yuhan Huang, Xinyu Zhang, Hengjun Wang, Huan Wang, Yiming He
Abstract
To compensate for the overlooked data realism and physical consistency in traditional cross-machine sample generation methods, this paper proposes a Physics-Constrained Adaptive Style Transfer Network for Sample Generation in Cross-Machine Small-Sample Fault Diagnosis (PCASTNet) for generating diagnostic samples of the monitored machine under small-sample conditions. PCASTNet employs style transfer to decouple fault content from a reference machine and machine style from the monitored machine. Moreover, it further integrates physical priors by introducing a band energy preservation constraint during sample synthesis. Built upon multi-scale wavelet transform, the framework consists of a wavelet encoder, an Adaptive Style Normalization (AdaSN) module, a wavelet decoder, and a multi-objective loss that jointly constrains content fidelity, style consistency, and energy preservation. Experimental results on two cross-machine scenarios demonstrate that PCASTNet can generate a large number of samples from limited monitored machine data and significantly improves diagnostic accuracy under small-sample conditions.