J. Barry Oaks Advancement Award Presentation: Sparsity-assisted machinery fault diagnosis

Abstract: Sparsity-assisted machinery fault diagnosis uses sparsity-assisted signal processing techniques to improve the performance of fault feature extraction and thus to promote the ability of diagnosis, which is realized by exploring sparsity priors in a certain domain of fault-related condition information. In recent years, sparsity-assisted signal processing techniques have been widely studied for machinery fault diagnosis. In this representation, fault-induced sparsity prior is introduced to model the vibration signal to enhance the fault information for fault feature extraction, and moreover the optimization algorithm to solve the sparsity-assisted signal model is unrolled to construct a fault-information-aware interpretable neural network for mechanical fault diagnosis. More specifically, this presentation introduces a new sparsity-assisted model under the assumption that the noise in the signal obeys a mixture of generalized Gaussian (MoGG) distribution. Also, the Lp norm (0＜p≤1) is adopted as the regularization to keep the model sufficiently sparse and adjustable. Thus, this sparsity-assisted model is named a MoGG noise distribution enabled sparse representation (MoGG-SR) model. The mixed distribution characteristic can make this model more adaptive, and the use of the generalized Gaussian function as the basis function makes it more robust to outliers. Furthermore, the presentation introduces an interpretable neural network to provide high-performance and credible mechanical fault diagnosis results. The sparsity-assisted interpretable network is mainly generated by unrolling the nested iterative soft thresholding algorithm (NISTA) for a sparse coding model and it is named NISTA-Net. Therefore, the network architecture of NISTA-Net has a clear theoretical basis, and users know how it is designed. Additionally, a visualization method is introduced for NISTA-Net to examine whether the network has learned meaningful features. This method helps users better understand how NISTA-Net performs classifications. These two aspects of transparency/interpretability allow NISTA-Net to be more credible when applied for mechanical fault diagnosis. We carried out simulation and experiment study to verify the performance of sparsity-assisted machinery fault diagnosis. The results reveal that sparsity-assisted methods can well extract the fault features of the concerned bearings and gears. As a consequence, it achieves the best performance compared with other advanced networks.