Final Project Report Title : Development of mean value engine model using ANN Course : ECE 539 Instructor : Prof. Yu Hen Hu Name : Soo-Youl ParkIntroduction In response to increasing demands for lower emissions and higher fuel economy, modern diesel engines require advanced control systems to coordinate complex engine sub-systems such as exhaust gas recirculation (EGR), turbocharger with a variable nozzle turbine (VNT), and common-rail fuel injection system. Development of such sophisticated control systems can be time consuming and labor-intensive, especially when the process relies heavily on testing new control designs and extensive calibration of hardware prototypes. In order to reduce the development time and costs, computer simulations are highly desirable and often necessary in the early design stage. In order to carry out much of the control design and development off-line using computer simulations, engineers need not only a model of the engine controller that simulates the control algorithms, but also an accurate engine plant model that can be integrated with the controller model. The requirement of this type of engine simulation model consists of detailed dynamics to capture transient response as well as the capability to predict engine status such as pressure, temperature, mass flow, and torque. Furthermore, the control design usually involves large number of iterations and evaluations. Simulation over long driving cycles and hardware-in-the-loop (HIL) are also expected. As a result, developing a fast running engine model with sufficient accuracy for control applications has become a trend in the automotive industry. However, one of the fast models widely used in the academy and industry requires 20~1000 times greater than real time as a calculation time based on current CPU technology. For example, in order to simulate 30 seconds of engine operation, the required calculation time is 30000s (8.3 hours). To overcome the limitation of calculation time and apply the engine model for various research field including control aspect and HIL, it is proposed that use of the artificial neural network(ANN) can decrease the model complexity without sacrificing model accuracy. We call this engine model supported by ANN asmean value engine model (As a contrast of mean value engine model, we call full physics based engine model as detailed engine model.). Many research shows successful application of mean value engine model. The purpose of this study is to review the previous research about ANN application of engine model and to reproduce the results by myself. Methodology for building mean value engine model 1. Building detailed engine model Detailed engine model was developed to give a comprehensive representation of all engine components, such as the induction system, the inter-cooler, the intake manifold, the cylinders, the exhaust manifold, the EGR loop, the turbocharger, and the exhaust system. The engine model has been validated against experimental measurements on engine dynamometer at a wide range of engine operating conditions. Therefore the results from this model can be a reference for testing the mean value engine model. 2. Generating training and test data set for developing ANN Before generating training and test data, it is important to decide which part of engine model can be replaced with ANN model. Generally, the replaced part should be time consuming part during calculation and if it would be replaced by the ANN model, it would not affect the results much. The representative one is cylinder model. Detailed cylinder model contains many kinds of physics and chemistry model. Therefore, it is a representative time consuming part. Moreover, non-existence of cylinder model doesn’t affect the results much because its effect in overall calculation is not coupled to other kinds of model. The other part which can be replaced with ANN model is engine intake manifold. This part is also time consuming part during calculation but this part could affect the results if it is replaced by ANN with out any compensation. Therefore proper compensation is necessary for building the mean value model and details for compensation can be found in reference [1].The fact that some parts of engine is replaced by ANN model means that some table information which composed of input and output is used instead of directly simulating this part. The variables which comprise this table are summarized in table 1 and these variables comprise the input and output vectors for neural network. Table 1 Variables for input and output vectors of neural network 6 inputs 3 outputs Engine speed Volumetric efficiency Intake manifold pressure indicated efficiency Intake manifold temperature Exhaust energy percent Exhaust manifold pressure Injected fuel mass Injection timing For generating the test and training data, DoE(Design of Experiment) method was used. As can be seen in Table 1, the number of input variable is six and variation of this variable is so wide under engine operating condition. Hence, it is effective way to decide the simulation point using DoE to fully cover the entire engine operating point. For data sampling, Latin Hypercube was used to design the simulations in order to cover the defined input ranges with randomness and even dispersion in space. Total 1000 times of calculation were performed and corresponding time is almost 10 hours. Figure 1 shows sampled value for input variable 2(intake manifold pressure) and Figure 2 shows corresponding value of output variable 1(volumetric efficiency). It can be seen that input variable is evenly selected over entire range, but output value of volumetric efficiency is concentrated on certain range.Figure 1 Sampled input value(Intake manifold pressure) for DoE Figure 2 Corresponding output value(volumetric eff.) from DoE This DoE results contain the information necessary to map the six input factors to the three mean value engine quantities and they will be used to training the neural network. 3. Training and testing ANN 3.1 Learning method In this project, linear self organizing map was used to train the neurons. It uses unsupervised learning to produce a dimensional, discretized representation ofthe input space of the training samples, called a map. The map seeks to preserve the topological properties of the input space. In this learning, the output vector can be expressed by following equation. where n is the
View Full Document
Unlocking...