Cylinder Pressure-based Detection of Engine States

Recently, the Beru AG (today: BorgWarner Beru Systems GmbH) introduced the Pressure Sensor Glow Plug [1] [2] which enables the acquisition of in-cylinder pressure curves. The aim of our cooperation project with BorgWarner BERU Systems is to verify whether these pressure curves can be used to predict the states of actuators within a turbocharged Diesel engine. Passive devices such as valves influence the pressure signal by cancelling out certain noise parts. Thus, it is expected that a varying setting of the underlying actuators is reflected in high and low pressure areas of the cylinder pressure curve. Modelling this interrelationship using neural networks or support vector machines we are able to predict actuator settings using the actual in-cylinder pressure curve only ([Komma09],[Shutty09]). We thereby can omit further sensors that fulfil this task. In total, we predict the state of four engine valves: the exhaust throttle EGR valve module, the high pressure EGR valve, the throttle valve, and the variable turbine geometry valve of a VTG system. Further studies focus on the prediction of other engine states such as the intake pressure, intake temperature, and the turbo charger speed (cf. figure below).



Data Preprocessing and Model Generation

The general procedure is split into two phases. Both phases rely on preprocessed in-cylinder pressure curves acquired during a complete working cycle: in the model generation phase, a model learns the correct assignment of known engine states given the pressure curves. In the recall phase, pressure curves are applied to the learnt model in order to predict the current engine states. Below, details of the pressure curve preprocessing stage and model generation phase are provided.


Haar Wavelet Analysis

Internal combustion engine acoustic measurements provide information about the engine's operating parameter and physical characteristics. However, the acquired signals are complex and superimposed by backward noise, demanding accurate processing. In our approach, we employ the Haar wavelet transform for feature extraction. As wavelets are localized in both space (time) and scale (frequency) domains, they can detect local features in a signal. Furthermore, wavelet analysis has a runtime complexity of O(n) making it feasible in our domain.


Feature Selection

Feature selection techniques are necessary for reducing the input dimensionality to avoid unwanted effects like overfitting. One solution for the feature selection process is to assign each feature a statistical relevance measure. If the chosen measure is independent from the employed regression model this technique is referred to as the filter or feature ranking approach. The applied filter is based on mutual information which is a non-parametric measure of shared information or dependency between two (or more) variables. Given a degree of mutual information between the target value and each pressure curve feature we only select those features which provide the maximum amount of shared information.


Model Simplification

The machine learning model is represented by a support vector machine (SVM). To meet real-time and memory constraints, the established SVM has to be simplified by reducing the support vector set (cf. red dots in the figure below). In our framework, we adopt the technique of Downs et al. [3]. There, they remove those support vectors that can be expressed as a linear combination of other ones in feature space. By modifying the weights of the remaining support vectors, the generalization performance is preserved while the complexity of the regressor is decreased. Our experiments showed that 90% of all support vectors could be removed without a significant increase in prediction performance.


Publications

  • [Komma09] Philippe Komma and Andreas Zell. Towards real-time and memory efficient predictions of valve states in diesel engines. In IEEE Workshop on Computational Intelligence in Vehicles and Vehicular Systems (CIVVS 2009), pages 8-15, Nashville, TN, USA, 2009.

  • [Shutty09] John Shutty, Wolfgang Wenzel, and Philippe Komma. Control strategy optimization for hybrid EGR engines. In 9. Internationales Stuttgarter Symposium 2009 - Automobil- und Motorentechnik, Stuttgart, Germany, March 2009.

Contact

Philippe Komma, Tel.: (07071) 29-78970, philippe.komma at uni-tuebingen.de