Generalized Zero-Shot Learning Based on Diffusion Model and Multilabel Network for Compound Fault Diagnosis

In compound fault diagnosis, the scarcity of samples leads to a low fault diagnosis rate. Existing zero-shot compound fault diagnosis methods lack the ability to simultaneously recognize both seen and unseen fault classes, especially when aligning attributes of unseen faults, where a dimensionality explosion occurs. To counter the deficiencies of traditional zero-shot compound fault diagnosis methods, this article introduces an innovative generalized zero-shot compound fault diagnosis approach. This approach pioneers the use of an attribute transformation strategy, establishing a knowledge bridge between single and compound faults through a semantic label definition module, thereby creating a fault attribute set. The fault generation module is then utilized to ingeniously convert the attribute set into training samples that are rich in fault characteristics. These samples are employed to train a multilabel classification module, enabling the model to identify compound faults, including those of unseen classes. Experiments conducted on two real-world bearing datasets have validated the effectiveness and strong generalization capabilities of this method, offering a novel solution and technical support for current zero-shot rotating machinery fault diagnosis.