Intelligent multi-fault diagnosis for a simplified aircraft fuel system

Abstract

Machine learning (ML) techniques are increasingly used to diagnose faults in aerospace applications, but diagnosing multiple faults in aircraft fuel systems (AFSs) remains challenging due to complex component interactions. This paper evaluates the accuracy and introduces an innovative approach to quantify and compare the interpretability of four ML classification methods—artificial neural networks (ANNs), support vector machines (SVMs), decision trees (DTs), and logistic regressions (LRs)—for diagnosing fault combinations present in AFSs. While the ANN achieved the highest diagnostic accuracy at 90%, surpassing other methods, its interpretability was limited. By contrast, the decision tree model showed an 82% consistency between global explanations and engineering insights, highlighting its advantage in interpretability despite the lower accuracy. Interpretability was assessed using two widely accepted tools, LIME and SHAP, alongside engineering understanding. These findings underscore a trade-off between prediction accuracy and interpretability, which is critical for trust in ML applications in aerospace. Although an ANN can deliver high diagnostic accuracy, a decision tree offers more transparent results, facilitating better alignment with engineering expectations even at a slight cost to accuracy.