Mix-VAEs: A novel multisensor information fusion model for intelligent fault diagnosis

Abstract

Multisensor information are usually required to recognize the health condition of machinery by domain experts, since redundancy and complementarity of multisensor information can enhance robustness of fault diagnosis. However, the mainstream approaches of intelligent fault diagnosis based on deep neural networks only focus on applications of single sensor. In this paper, a novel intelligent fault diagnosis model using mixture of Gaussians (MoGs) and variational auto-encoders (VAEs), named Mix-VAEs, is proposed to leverage redundancy and complementarity of multisensor information. The proposed Mix-VAEs mainly contains two modules. First, feature extracting module is constructed by several parallel and independent VAEs which are used to learn data representation from multisensor data respectively, and their latent variable distributions extracted by VAEs are used for fusion. Second, in multisensor information fusion module, a novel fusion method is proposed where latent variable distributions are combined to form MoGs, and the fused features are obtained in a sampling manner from MoGs. The performance of the proposed model is verified through two datasets consisting of multisensor data. Experimental results demonstrate the robustness of the proposed model in two scenarios, sensor failure and local signal missing, while also showing the better performance than other fusion models.