Machinery multi-sensor fault diagnosis based on adaptive multivariate feature mode decomposition and multi-attention fusion residual convolutional neural network

Abstract

Due to the complex and rugged working environment of real machinery equipment, the resulting fault information is easily submerged by severe noise interference. Additionally, some informative features may be omitted if the feature learning concerns only a single sensor of machinery vibration data. Therefore, to mine more comprehensive fault information and achieve more robust fault diagnosis results, this study proposes a machinery multi-sensor fault diagnosis method based on adaptive multivariate feature mode decomposition and multi-attention fusion residual convolutional neural network. As an extension of the feature mode decomposition (FMD), the adaptive multivariate feature mode decomposition (AMFMD) with the improved whale optimization algorithm (IWOA) is firstly presented to automatically decompose the collected multi-sensor vibration data into a group of multichannel mode components, which both inherit the anti-noise robustness of the original FMD and overcome the obstacles of artificial parameter selection of FMD. Subsequently, multichannel mode components containing the most abundant fault information are selected via an impulse sensitive measure hailed as multichannel comprehensive index (MCI), and the frequency slice wavelet transform (FSWT) of the selected multichannel mode components is further calculated and organically fused to generate the colored multichannel time–frequency representation (MTFR) containing multi-sensor important signatures. Finally, by integrating the advantages of feature learning of residual network (ResNet) and convolutional neural network (CNN), a multi-attention fusion residual convolutional neural network (MAFResCNN) with squeeze-excitation module (SEM) and convolutional block attention module (CBAM) is constructed to simultaneously capture global and local feature information from the fused multichannel time–frequency representation and implement automatic discrimination of machinery fault states, which can both enhance machinery fault information and whittle down the useless information, even promote the feature learning performance without significantly increasing the computational burden of the model. The validity of the proposed approach is verified by a diagnosis case of a real wind turbine, demonstrating that the proposed approach has superiority in machinery fault identification compared with some similar techniques.