A Novel Acoustic-Based Framework for Compound Fault Diagnosis in Rotating Machinery With Limited Samples

Acoustic-based intelligent fault diagnosis (AIFD) for rotating machinery has gained significant attention in recent years due to its ability to overcome the limitations of vibration-based fault diagnosis, particularly in contact measurement. However, research on the small sample problem arising from the scarcity of compound faults in AIFD remains limited. To explore this problem, this article proposes a novel acoustic-based framework for compound fault diagnosis in rotating machinery with limited samples. The framework integrates Mel spectrogram mapping to enhance fault feature representation, a latent space-controlled generative adversarial network with a local perception mechanism (LP-LSGAN) to generate high-quality compound fault samples for effective data augmentation, and a multilabel decoupling classifier, capsule network (CapNet), to accurately identify multiple fault labels. Comprehensive experiments involving two distinct cases on a gearbox test platform demonstrate the ability of the framework to handle diverse fault scenarios. The results show that the proposed framework can effectively generate data that is highly similar to real samples, and as the imbalanced dataset is gradually expanded to a balanced dataset, the fault diagnosis accuracy is improved by about 30%, which is better than the other comparison frameworks under limited samples.