IBRNet: Interpretable Belief Rule Network modeling method for fault diagnosis of redundant inertial navigation systems

Abstract

With the increasing complexity of inertial navigation systems, especially with the incorporation of redundant inertial measurement units, fault diagnosis is facing new challenges. On one hand, it involves the effective integration of system mechanisms and angular increment data. On the other hand, it focuses on the interpretability of the fault diagnosis models and their corresponding diagnostic results. To address these issues, this article proposes a new belief rule network structure. It takes into account interpretability constraints during the optimization process. The proposed structure, called IBRNet, aims to achieve fault diagnosis of redundant measurement units. Specifically, the belief rule network is used to model semi-quantitative information, and interpretability is achieved through two components: parameter optimization based on interpretability constraints and interpretability analysis based on Kernel SHAP. Additionally, a fault diagnosis model based on IBRNet was established for the inertial measurement unit with 5-table redundancy. Experimental results show that the proposed IBRNet method can provide higher accuracy and interpretability compared to several existing methods.