Towards real-time and memory efficient predictions of valve states in diesel engines

To reduce production costs of current engines, car manufacturers strive to replace built-in sensors by software solutions. However, the limitations of current micro controllers require time and memory efficient algorithms. In this paper, we propose a real-time framework for the detection of engine valve states based on wavelet analysis of in-cylinder pressure curves. Extracted wavelet features are then filtered out using mutual information such that only the most relevant wavelet coefficients become the input of the chosen support vector regressor. A further speedup is achieved by an approximation of the support vector solution which comprises less support vectors. We show that the combination of relevant feature selection and the regressor model simplification results in a significant decrease of the recall phase complexity while retaining good generalization performance.