.png){kind=icon}
.png){kind=icon}
Return
Multivariate cyclostationary deep deconvolution (MCDD): An intelligent multivariate signal processing algorithm
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.131803.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Blind deconvolution (BD) techniques are widely employed to recover fault source signals by formulating an objective function aligned with fault characteristics. However, traditional BD methods are highly susceptible to strong impulses and noise. To address these limitations, this study introduces a novel multivariate cyclostationary deep deconvolution (MCDD) approach, characterized by a multi-input, multi-node, and deep network architecture. First, the input layer takes multi-channel data as input, enabling the fusion of feature information across different channels. Moreover, a filter bank is initialized using a 1/3 binary tree filter bank structure, which divides the frequency spectrum into sub-bands. Secondly, in the hidden layer, generalized Gaussian stationary/generalized Gaussian cyclostationarity is used to guide the feature learning within a deep network, and the eigenvalue algorithm is employed to update the weights. The hidden layers progressively enhance fault features by extracting the characteristics of sparse signals at each layer. Finally, in the output layer, dimension reduction is performed to extract the most salient fault features. Three case studies demonstrated the superiority of the proposed MCDD algorithm over traditional BD techniques in fault feature extraction. The quantitative analysis verified the proposed MCDD algorithm has enhanced robustness and diagnostic precision, particularly under noisy operating conditions.
Keywords:
Blind deconvolution
Cyclostationary signal processing
Deep learning
Fault diagnosis
Multivariate signal processing
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
View more organizations