Design and analysis of a learning-based testing system for certification of vehicle systems

Sammanfattning: In this work, a learning-based testing system is designed and evaluated in terms of its perfor-mance and feasibility of use in testing of safety-critical vehicle systems; the objective is to reduce testing time and costs. A literature study was conducted on the AMASS project, model-based testing and machine learning; based on which a design of the testing system was developed. The finished testing system uses a genetic algorithm for generating solutions of high fitness, which in this application implies test cases that provoke failures in a target system under test, in order for the developers to detect system defects. The target testing system is a model of Volvo’s Brake-By-Wire ABS module. It was concluded that the testing system is effective in increas-ing fitness of solutions through iteration; the performance of the machine learning algorithm is dependent on parameters such as the mutation rate and the size of the populations into which solutions are clustered.