Application of structural reliability methods for the safety assessment of autonomous vehicles

Abstract

Advanced Driver Assistance Systems (ADAS) see a constantly growing attention by researchers and industries as more and more vehicles are equipped with such technology. It is generally expected to see first Automated Driving Systems (ADS) in the market in the next years. One of the most important aspects for reaching this goal and releasing ADS is testing and validation.

Several publications stated out, that the required mileage needed to proof the probability of failure of the system is impossible to reach in field operational tests. Therefore, statistical methods, e.g. Monte-Carlo simulation, combined with Software-in-the-Loop (SiL) simulation may help to overcome this limit. The field of reliability analysis, initially developed for structural mechanics, provides algorithms and approaches which can be applied to assess ADS. Due to the different kind of parameters and criteria, available methodologies need to be analyzed and adapted to ADS specific challenges.

In this paper, the presented process is based on event-based simulations where specific traffic scenarios are parametrized, simulated and analyzed by a set of criteria. By using predefined distribution functions for each input parameter, a safety statement can be given by approximating the probability of failure for each traffic scenario by determining the unsafe region in the parameter space. Therefore, multiple steps of different algorithms are combined to ensure trustworthy results by being very efficient in reducing the number of required simulation runs.

Based on established reliability methods, known from civil engineering and from aircraft design, small event probabilities are estimated by varianced-reduced Monte Carlo procedures as Importance Sampling. Since in each investigated scenario different failure modes and regions may occur, which have to be considered, State-of-The-Art procedures for searching for the most probable failure regions have been substantially extended. In this paper, it is shown, which challenges appear in applying these methods for Software-in-the-Loop simulation models.