A knowledge and data collaboration-driven method with modified stacked broad autoencoder for few-shot bearing fault diagnosis

Abstract

Data-driven fault diagnosis methods have made substantial advancements in bearing prediction and health management. However, the scarcity of fault samples hinders their application in engineering practice. To tackle this challenge, we put forward a novel knowledge and data collaboration-driven method with modified stacked broad autoencoder (MSBAE) for few-shot bearing fault diagnosis. First, 30 domain-knowledge-based fault features are derived from the time, frequency, and time-frequency domains. Then, by introducing the maximum mean discrepancy and manifold regularization into the original stacked broad autoencoder, MSBAE is constructed to integrate the prior knowledge into its self-learning process, thereby mining more discriminative and robust features from massive unlabeled data. Finally, a least squares classification layer is employed on top of MSBAE for fault recognition, and the structure parameters of MSBAE are fine-tuned with limited labeled fault samples. Extensive experiments on three bearing fault datasets confirm that our method exceeds other cutting-edge methods, particularly under small-sample conditions.