Open Access

Lv Wang, Junyu QiORCiD, Yi QinORCiD

• Compound faults are becoming more frequent in rotating machinery due to the increasing complexity of industrial equipment. However, acquiring fault data, particularly compound fault data, remains challenging. Recent progress in zero-shot learning has provided techniques capable of recognizing unseen compound faults even when no corresponding samples are available during training. Nevertheless, existing approaches are limited due to domain shift and feature confusion problems, which compromise diagnostic accuracy. To address these problems, a triple-prototype fusion (TPF) framework is developed for zero-shot fault diagnosis. First, a prototype fusion framework is constructed to address the domain shift issue. This framework fuses the generated prototype of the simulated faults, the feature prototype derived from the actual faults, and the theoretical prototype constructed by the prior knowledge, thereby alleviating the problem of domain shift caused by insufficient information from the single prototype. Then, a multi-distribution fusion contrastive learning mechanism is developed to solve the feature fusion problem. This mechanism introduces the multivariate prior information of distributions when generating positive samples, making the constructed positive sample more discriminative. Additionally, a model-sample stabilization diffusion model is proposed to enhance the stability of sample generation. The model introduces stability mechanisms from model training and sample denoising perspectives to comprehensively improve the stability of generated samples. The effectiveness of the TPF framework is verified through two experimental cases. The TPF framework enables the diagnosis of unseen compound faults without requiring any compound fault data for training. Benchmarking against eight typical models and advanced zero-shot learning methods, the TPF achieves favorable performance. The code will be available at https://github.com/QinYi-team/TPF.