A Novel Compound Fault Diagnosis Method by Combining Feature Conditional Diffusion and Capsule Networks Under Scarce Samples

In machinery fault diagnosis, traditional fault diagnosis methods often face the challenge of scarce compound fault samples. To overcome this challenge, a novel compound fault diagnosis method combining feature conditional diffusion and capsule networks (FCD-CNs) is proposed. First, to tackle the imbalance problem caused by scarce compound fault samples, a conditional diffusion model is constructed to generate high-similarity compound fault pseudo-features. Then, a capsule network is designed to decouple and classify faults from the perspective of their coupling relationship between different faults. Furthermore, a three-phase training strategy is proposed to reduce the distributional differences between generated and real features. Finally, the proposed method is substantiated through comprehensive case studies in the Hangzhou Dianzi University (HDU) dataset. The results demonstrate its effectiveness and merits in the extreme scarcity of compound fault samples.