A shorter time to market is a clear business advantage for any company. Car
manufacturers also thrive to bring their new cars to the market at their planned time.
However, emission tests are heavy blockades for any new car. If a new car line cannot
pass an emission test, the manufacturer cannot sell it. Finding the reasons why a car fails
an emission test is hence an important task. Until now, to analyze the problem, the
engineers must look at all signals produced from sensors built into the car. That process is
thus very time intensive. In this thesis, we proposed a technique that ranks these signals
based on their relevance to the emission outcome. The technique uses machine learning
to automate the ranking process, thus helping the engineer to decrease the analysis time.
The thesis is divided in three main parts. In the first part, we conducted a literature
review to assess the existing anomaly detection techniques. After that, three techniques
are chosen to be implemented and benchmarked. We found that the IsolationForest
technique performs best for automobile data. In the second part, we analyzed the
problem proposed above and introduced a technique to solve it, using the anomaly
detection techniques that we assessed in the previous part. In the last part, the proposed
solution is implemented in form of a recommender system and gets evaluated using real
data from emission tests.