An Explainable AI approach for detecting failures in air pressure systems

Abstract

The Air Pressure System (APS) plays a crucial role in heavy-duty vehicles (HDVs), supplying pressurized air to essential subsystems such as braking and suspension. APS failures normally lead to vehicles being stranded on the road with associated safety and financial risks. Although detecting these failures is essential to prevent such events, the detection trustworthiness is equally important given the high sensitivity of this issue. This paper addresses the problem of APS failure detection using Explainable Boosting Machine (EBM), a highly intelligible and interpretable glass-box model. A dataset of operational driving data from 110 healthy vehicles, without any APS failures, and 30 faulty vehicles, with detected APS failures, was collected. First, essential preprocessing steps were developed to deal with the hierarchical big data and to extract indicative features. The main objective of EBM is to distinguish faulty vehicles from healthy ones based on those features while providing explanations for its decisions. The model succeeded in detecting most of the faulty vehicles with a small proportion of false alarms (roughly 5%); the overall accuracy was 91.4% and the F1 score was 0.80. In addition, the provided explanations were thoroughly investigated to evaluate the validity and trustworthiness of the model decisions. At the same time, the explanations themselves were assessed based on domain knowledge to prove their efficacy and relevance. When compared with a human expert analysis, these explanations highly align with the experts’ knowledge of the APS problem. The proposed methodology is easily adaptable for other time-series predictive maintenance applications across different fields.