Machine Learning Based Fault Detection Of Automated Driving Functions In Time-series Data

Advanced Driver Assistance Systems (ADAS) mark a major leap forward in the realm of au tomotive innovations, designed to elevate both the safety and the comfort of driving. ADAS provides additional information from the car surrounding environment to support a driver and assist in implementing critical actions. The synchronization of a driver’s actions and the information from the environment is essential for the efficient performance of the various ap plications of ADAS[2]. Thesesystemsworktodiminishthechancesofcollisionsandtoenrich thedrivingexperience. However,thesophisticatednatureofADASintroducesobstacleswhen it comes to their incorporation into vehicles, their validation for use, and the examination of any errors that mayarise. This is further compoundedbytheintricate network of various Elec tronic Control Units (ECU)s involved and the substantial volumes of data they handle. To address these challenges, ADAS must undergo thorough and meticulous testing. This ensures that these systems are both reliable and safe before they are integrated into vehicles for con sumer use. This thesis addresses the challenges associated with the increasing complexity of ADAS by proposing an approach to predict errors by creating a informative, good, and big enough vehicle data, using Supervised Machine Learning Models to recognize the language and the characteristic of the cars. This was done by studying the communication between the ECUs and finding the reasons why something abnormal happens in the first place from the created data, analyzing the specific traces related to special function in the vehicle recorded live. The adapted model aims to read the four month’s-long prepared data, analyze the traces of the vehicles facing errors and the vehicles driven completely fine. then, using this ’data-driven approach’ todoareliableprediction ofwhetheranerrorwillhappeninthevehicleornot. The mainpurposeandfocusofthethesiswasonthecreationofthedata,thatthedatasetshouldbe good and interepretable enough to be used for the Mahcine Learning model to do the predic tion. The results show that with the both Binary-Classification-Models, the accuracy is good enough and the precision and recalls are reliable enough to state that the created dataset is a goodrepresentative enough dataset that one might beabletousetheactual datainstead of just the error codes for the analysis and predictions as they are not completely reliable.