Liu, Zeyu (Cranfield University), Kong, Xiangqi (Cranfield University), Chen, Yang (Cranfield University), Wang, Ziyue (Cranfield University), Jia, Huamin (Cranfield University), Al-Rubaye, Saba (Cranfield University)

Mitigating No Fault Found Phenomena through Ensemble Learning: A Mixture of Experts Approach

Scheduled for presentation during the Invited Session "Trustworthy Diagnosis and Prognosis in Connected, Cooperative and Automated Mobility" (WeBT10), Wednesday, September 25, 2024, 16:10−16:30, Salon 18

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

In the aviation industry, the reliance on precise fault diagnostic decision-making is critical for equipment maintenance. A significant challenge encountered is the erroneous categorization of components under "No Fault Found" (NFF), which subjects these components to unwarranted repairs or further testing. Such misclassifications not only trap on airlines through costly cycles of unnecessary maintenance but also exacerbate degeneration and potential safety hazards. Consequently, there is a heightened demand for the development of effective fault diagnosis models that are adapting to the aircraft complex systems and adeptly addressing issues related to the NFF phenomenon. In this study, we draw inspiration from ensemble learning and propose a multiple Naive Bayes experts (MNBMoEs) approach based on a mixture of experts (MoEs) model. This method leverages the predictive advantages of each sub-model on specific features, allowing the hybrid expert decision to outperform any single expert. It also includes a quantitative analysis method for the NFF issue, derived from the confusion matrix according to the industrial definition of NFF. Experiments evaluated on public datasets results show that the ensemble learning approach, based on Mixture of Multiple Naive-Bayes expert models, can effectively utilize the strengths of different models, improving fault diagnosis accuracy to 96.96%, with a maximum reduction in NFF occurrence rates of up to 94.17% and 84.2% model performance improvement.