Quick Look

Feasibility Investigation of a Model-Based Virtual Cylinder Pressure Sensor with Individual Variable Oriented Independent Estimators, 03-9440

Principal Investigators
Junmin Wang
Ryan C. Roecker
Charles E. Roberts Jr.

Inclusive Dates:  10/20/03 - 02/20/04

Background - As one of the most important parameters for the combustion process of the internal combustion engine, the cylinder pressure can provide tremendous amounts of useful information to improve the engine performance. Moreover, for some alternate combustion modes, such as homogeneous charge compression ignition (HCCI) and low-temperature combustion (LTC), the cylinder pressure signal is even more important to control the combustion processes. This project aims at developing a virtual cylinder pressure sensor (VCPS) to estimate the cylinder pressure related variables. The VCPS can be used as the fault diagnostic model for the physical cylinder pressure sensor as well.

Notice that for each engine combustion cycle, the cylinder pressure is a function of the current engine operation condition (e.g., intake manifold pressure, temperature) and the engine control commands (e.g., injection timing, injection mass), and so are these cylinder pressure representative variables. However, as a complex and highly nonlinear function, it is difficult to describe the cylinder pressure as a physics model.

Approach - The idea of the virtual cylinder pressure sensor with individual variable-oriented independent estimators is to use the physical cylinder pressure sensor to develop the independent model for every cylinder pressure related representative variable respectively by using the neural network technology. Then the well-developed models can be used on the production engines for the engine control systems and fault diagnostic modules.

Figure 1 shows the structure of the virtual cylinder pressure sensor. The items indicated with dash lines are used for the model development and training purpose only, which are preformed on a development engine with physical cylinder pressure sensor. After all the training processes are finished and all the models are verified, these parts will be removed. The rest will be ready for the production engines to use. Notice that more variable estimation modules could be added at the request of the engine control systems and fault diagnostic modules.

By using the independent models to estimate and predict different cylinder pressure related variables, the research team can tune each model separately and to simplify the modeling task and improve the robustness and accuracy of the VCPS.

Accomplishments - The VCPS was developed at 2,000 rpm from light load to high load on a light-duty diesel engine. Two neural network-based cylinder pressure related variable independent estimators are developed and verified, namely the cylinder pressure maximum rising rate position model and the maximum cylinder pressure value model. The results show that the models can predict the variables correctly compared with the extracted variables from the measured physical cylinder pressure signal. Very good generalization capabilities of the developed models are observed in the sense that the models work well not only for the training data set but also for the new inputs to which they have never been exposed. Figure 2 shows the verification of one VCPS estimator.

The results of this study indicate that the VCPS with individual variable-oriented estimators can greatly reduce the modeling task and improve the robustness of the models. The VCPS can be used for the fault diagnostics of the physical cylinder pressure sensor as well as engine predictive control.

	Figure 1. Structure of the VCPS with Individual Variable-Oriented Independent Estimators. More estimators could be added.
Figure 2. Verification of the Developed Maximum Cylinder Pressure Estimation Model. The estimated and measured values match very well.

2004 Program Home