- 學位論文
以故障注入與傷害等級模型評估自動煞車系統之風險
Risk Assessment of Automated Braking Systems Using Fault Injection and Injury Severity Models
摘要
本研究提出一種使用故障注入和傷害等級模型來評估自動煞車系統(AEB)風險的方法。研究的主要目標是建立一套針對先進駕駛輔助系統/自動駕駛系統的量化風險評估方法。研究方法包括構建AEB系統和故障模型，其中包括感測器、控制器和通訊網絡的故障注入模型；採用基於系統理論過程分析(STPA)的策略性故障注入方法，以提高危害事件觸發效率；在歐盟新車安全評鑑協會之測試場景下進行模擬，包括行人縱向行走和後車追尾兩種情況；使用擴展Delta-V估算速度變化量，並結合過往研究和新開發的傷害等級模型，評估危害事件的嚴重性和可控性。 主要研究結果顯示，基於STPA的故障注入方法能觸發更多危害事件，且形成更危險的情況。通訊網絡延遲、控制器運算位元翻轉和運算延遲是導致最多危害事件的三種故障類型。使用機器學習方法(如隨機森林、自適應增強和堆疊模型)開發的新傷害等級模型表現較好。大多數單一故障導致的危害事件嚴重程度為無傷害(AIS 0)。在有人為操作的情況下，大多數危害事件被評估為完全可控(C0)。 本研究為AEB系統的風險評估提供了一種新的量化方法，可幫助提高系統安全性。未來研究可進一步改進傷害等級模型和擴展到更多場景，以更全面地評估自動駕駛系統的安全性。
並列摘要
This study proposes a method for evaluating the risks of Automated Emergency Braking (AEB) systems using fault injection and injury severity models. The main objective is to establish a quantitative risk assessment approach for advanced driver assistance systems and autonomous driving systems. The research methodology includes constructing AEB system and fault models, encompassing fault injection models for sensors, controllers, and communication networks; adopting a strategic fault injection method based on System-Theoretic Process Analysis (STPA) to enhance the efficiency of hazard event triggering; conducting simulations in Euro NCAP test scenarios, including pedestrian walking and rear-end collision situations; and using extended Delta-V to estimate velocity changes, combined with existing research and newly developed injury severity models to assess the severity and controllability of hazard events. The main findings indicate that the STPA-based fault injection method triggers more hazard events and creates more dangerous situations. Communication network delays, controller bit-flip errors, and computational delays are the three fault types causing the most hazard events. New injury severity models developed using machine learning methods (such as Random Forest, AdaBoost, and Stacking Classifier) show better performance. Most single faults lead to hazard events with no injuries (AIS 0). In scenarios with human operation, most hazard events are assessed as controllable in general (C0). This study provides a new quantitative method for risk assessment of AEB systems, which can help improve system safety. Future research could further improve the injury severity models and extend the approach to more scenarios, enabling a more comprehensive safety assessment of autonomous driving systems.