End-of-line fault detection for combustion engines using one-class classification

Abstract

With increasing complexity of today's automotive combustion engines, end-of-line (EOL) testing has become an important method to test assembled engines for production faults. Several hundered measurement signals are evaluated for every EOL test, up to 100% of production volume. The difficulty of finding and maintaining accurate test limits makes EOL testing of complex products interesting as a machine learning problem. In this work, we define the EOL application as a one-class classification problem which can be solved by the support vector data description. The focus is on the comparison of three hyperparameter tuning methods that use little or no measurements from rarely occuring faulty engines. Experimental results with real production data show that fault detection rates of 95% are achievable in the industrial environment despite the high dimensionality of the EOL test data.