Human Senses Mimicking on Virtual Public Roads

Abstract

Automobile industry is continuously innovating and progressing towards autonomous solutions. An essential requirement for true vehicle autonomy is the capability to autonomously detect mechanical faults, as this function is critical for maintaining safety and ensuring that autonomous vehicles can operate independently, without an human intervention. As a part of a PhD research project aimed to develop a reliable diagnostic method concerning structural behavior of a vehicle and mitigate the malfunctions, this thesis is intended to investigate the influence of confounding variables on a set of pre-defined faults through multi-body simulations and signal processing technique. This project also incorporates the elements of machine learning framework to evaluate the performance metrics such as False Alarm Rate, Fault Detection Rate and Accuracy. A comparative study is done based on Score which is a harmonic mean of these performance metrics. The two mechanical faults related to Anti-Roll Bar and Knuckle were induced. The confounding variables were shortlisted based on their influence and feasibility to replicate them in ADAMS Car. This study highlights the inherent trade-off between false alarm rate and fault detection capability in autonomous vehicle fault models, emphasizing the need to balance sensitivity and specificity for safe deployment. The analysis demonstrates that confounders affect fault detection performance differently depending on the component and fault type. Additionally, certain operating conditions can reduce the distinguishability between fault types, leading to similar performance outcomes. Importantly, incorporating confounders during model training significantly improves robustness by reducing false positives and enhancing sensitivity, enabling more reliable differentiation between true faults and external influences, an essential requirement for safe and effective autonomous vehicle operation.