Research On Quantitative Assessment Method Based On The Safety Of The Intended Functionality Analysis Of Feature Scenarios

With the challenges bring by autonomous driving,the concept of Safety of the Intended Functional Safety(SOTIF)derived from it aims to deal with the limitations of its "perception-decision-execution" functions.Meanwhile,advanced autonomous driving functions are influenced by external environment.However,due to data shortage,current safety analysis process of Safety of the Intended Functionality stay only in qualitative level,and hardly to identify triggering events effectively,which inevitably leads to biased analysis resultsBased on the rules of scenario evolution in SOTIF and Hazard Analysis and Risk Assessment(HARA)of "Known Unsafe Scenarios",this thesis focus on the data analysis,and provides quantitative data to support the determination of Automotive Safety Integrity Levels(ASIL)by combining the characteristics of Operational Design Domain(ODD).The details of this thesis are as follows:Firstly,based on the "Known Unsafe Scenarios",and ODD attributes,this thesis study corresponding impact features and design a scenario division method for Safety of the Intended Functionality.Then,hazard causes,such as people,vehicles,road,and environment,are discussed.Combined with corresponding factors,such as traffic flow,population density and crime rate,typical area is selected as the analysis representative.At last,severity label is added to road accident data.Combined with road feature,environment factor,several algorithms,such as Logistic Regression,Random Forest,and XGBoost are used to comparative evaluation.Simulation results show that Random Forest model could evaluate scenario better and effectively identify the trigger events.Meanwhile,in view of the complicated workload of the HARA analysis process,a principal component analysis(PCA)based data Dimension-Reduced are done.Experiment results show that 40 important principal components can effectively characterize about 90% of feature differences of the whole data.Among all features,road conditions had the greatest impact on the result evaluation,while the validity of the method for SOTIF scenario evaluation was verified using Random Forest with an accuracy of about 0.8526.