Intelligent fault diagnosis scheme via multi-module supervised-learning network with essential features capture-regulation strategy

Abstract

The performance of data driven-based intelligent diagnosis method greatly depends on the quantity and quality of data. Nevertheless, due to realistic limitations, failure data is hard to acquire, which makes the training process of numerous intelligent models unsatisfactory and leads to performance degradation Aiming at this problem, considering the local impulse characteristics as minimum diagnosable units, this paper proposes a signal adaptive augmentation network (SAAN) to effectively construct artificial samples for amplifying fault data volume. The SAAN consists of three sub-structures: impulse extractor, regulator, and classifier. The impulse extractor combines inner product matching principle to extract the local impulse features from insufficient samples to construct massive initial artificial samples. The regulator adopts convolution and deconvolution frameworks to regulate and reconstruct the initial artificial samples by specially designed synthetic loss function, which makes artificial samples have same characteristic distribution as real samples. The augmented method is used for validation on three bearing data with some advanced algorithms. Besides, a focal normalized network is designed for classification under small samples. Relevant experiments indicate that the SAAN shows a competitive performance with existing state-of-art diagnostic methods, which can helpfully improve recognition accuracies of various diagnostic models by 5%–35% under small sample prerequisite.