Study of deep multimodal information fusion–based digital twin method for gearbox fault diagnosis

Abstract

Abstract Faults in gearbox systems occur frequently due to their complexity, while existing simulation models often struggle to provide an accurate description of gearbox fault behaviors. Thereby, accurately and instantaneously determining the operation status of the gearbox has become a significant challenge. To solve this problem, a digital twin (DT) method based on deep multimodal information fusion (DMIF) is proposed by this study. Firstly, two Boltzmann machines are constructed to extract feature-related data separately from sensor measurements and simulations of the rigid-flexible coupled dynamics model. The information from these two modes is then mapped into a high-dimensional space to form a joint representation. Then, the two information source features are combined by a multilayer feed-forward neural network to create a multimodal information fusion (MIF) used for real-time fault detection. Finally, in order to reduce the information gap between the virtual and physical spaces in the practical deployment of DT-driven fault diagnosis (FD) techniques, an adaptive correction model is proposed based on a multi-objective locust optimization algorithm (LOA), which significantly improves the fidelity and accuracy of the virtual space. The experimental results indicate that the proposed DT method can effectively reduce information errors between the physical and virtual spaces, thereby improving the accuracy of FD.