Abstract: Deconvolution methods are powerful tools for mechanical equipment fault diagnosis; however, traditional research still relies on shallow feature extraction, making it difficult to handle extremely low signal-to-noise ratios. To address this issue, by introducing the idea of feature learning into the traditional deconvolution theory, a Gini index (GI) based sparse deep deconvolution (GI-SDD) method is proposed for early fault diagnosis of mechanical equipment. First, a band-averaging strategy is adopted to initialize the input layer filter, providing direction for subsequent deconvolution. Next, GI that can represent sparse features of mechanical faults is utilized as the loss function to guide the training of the deep network. Weight optimization is implemented based on the generalized eigenvector algorithm (EVA), thereby learning weak fault features layer by layer. Finally, correlation coefficients and envelope kurtosis (EK) criteria are utilized to evaluate the fault information, reducing dimensionality to output the most significant fault components. Simulation and experiment results demonstrate that the proposed method is robust against strong background noise with fault features being greatly enhanced. Furthermore, the EK of the proposed method improves by 163.43% and 187.11% compared with traditional MED and MGID results respectively.

Key words: Gini index (GI), eigenvector algorithm (EVA), deep deconvolution, feature learning, fault diagnosis