Controlling the Rate of Heat Release with Learning Algorithms

Transcription

1 Controlling the Rate of Heat Release with Learning Algorithms Increasing degrees of freedom in engine management systems have a high impact on the workload for the calibration of modern diesel engines. To be able to handle this in an efficient way many OEMs aim to cover the diversity of vehicle model ranges with fewer engine series. IAV describes in this article a new calibration method that aims to make derivative calibration of diesel combustion processes more efficient. IAV 52

2 INTRODUCTION AUTHORS Sven Holzendorf, M. Sc. is Development Engineer for System Development and concepts for Combustion Process at IAV in Berlin (Germany). Dipl.-Ing. Katarina Aleksic-Roessner is Development Engineer for Exhaust and Fluid Systems at IAV in Berlin (Germany). Dr.-Ing. Achim Lechmann is Technical Consultant for combustion Process/Diesel Hybrid at IAV in Berlin (Germany). The increasing degrees of freedom in engine management systems are causing constant growth in the workload for the calibration of modern diesel engines, due to the need for comprehensive optimization of increasingly complex and powerful systems. At the same time, many OEMs aim to cover the diversity of vehicle model ranges with fewer engine series. Generating derivatives from a lead project is therefore becoming increasingly significant. IAV has developed a new kind of calibration method that aims to make derivative calibration of diesel combustion process more efficient. The calibration engineer will thus be able to transfer the combustion procedure design from an already calibrated lead project to one or several derivative projects. This can be useful, for example, if there is a change in system supplier for the injection system, or if a combustion procedure is to be calibrated to another output class. This article also examines whether combustion processes can be transferred successfully by equating the heat release profiles in the high-pressure phase of the engine process with that of a lead project. Here it makes sense to use closed-loop combustion control based on in-cylinder pressure. The injection system parameters result automatically as actuating variables and do not have to be optimized separately for the derivative project. The following sections describe the development of a combustion control system including some example results for the presented calibration method. Potential for further development is shown as well as other possible application cases. CLOSED-LOOP COMBUSTION CONTROL WITH STEADY-STATE PARAMETERS Cost functional The presented method was developed on IAV engine test facilities in two steps. The first trials were carried out on a single-cylinder research engine with common rail direct injection and directly driven piezo injector. The setup was actuated by a research control unit for continuous variation of the injection rate by needle seat throttling. The time curve for the electric charge in the piezo stack of the injector was adjusted exactly to a definable target profile. This acts as actuating signal for the first development stage of the combustion control system. The in-cylinder pressure trace is recorded using a test bench indication system with piezo-electric transducers. For every engine cycle j, n pressure values are recorded and used initially to calculate the rate of heat release y j. This acts as feedback signal for the combustion control. Together with the curve of the electric piezo charge u j, it defines the system boundaries or more precisely the input and output variables for the control plant. The control plant is de - scribed below by means of a simple linear model Eq. 1 y j = Gu j + ɛ j whereby G is the system transfer matrix and ɛ j is the sum of model uncertainty and measurement noise. By stipulating a reference heat release profile y ref, it is possible to compute a control error curve for the current engine cycle which Eq. 2 e j = y ref - y j describes the dynamic tracking behavior within the j -th cycle. The variables mentioned can be used to compute a new actuation curve for the next engine cycle with a reduction in control deviation. When used repetitively, it is thus possible to minimize the control error norm step by step over several cycles. This control principle is referred to as iterative learning control (ILC) and is known for example from [1]. Controller synthesis was based initially on a constrained form of the norm-optimal ILC (quadratically optimal design Q-ILC) according to [2]. The actuating signal thus results as solution of the cost functional (1) in TABLE 1. The constraint u j 0 ensures that there is no negative actuation. The quadratic program is solved by an iterative algorithm using the principle of conjugate gradients. EXTENDED CLOSED-LOOP COMBUSTION CONTROL WITH ONLINE SYSTEM IDENTIFICATION Initial studies showed that the quality of the plant model G has a major impact on the tracking behavior of Constraint (1) u * j = arg min uj [e T j W e e j + u T j W u u j + (u j - u j-1 ) T W u (u j - u j-1 )] u j 0 (2) u * j = arg min uj [e T j W e e j + u T j W u u j ] u j 0 u j 0 n I TABLE 1 Overview of the cost functionals used for controller synthesis; W e, W u and W Δu are positive semi-definite weighting matrices; details are described in [2] ( IAV) MTZ worldwide

3 FIGURE 1 Schematic diagram of adaptive combustion control ( IAV) the ILC. Clear control deviations can persist, especially when long ignition delays result in high, variable response times that are only inadequately reflected in a steady-state model. On the other hand, the systems behavior depends on a large number of process parameters. For example, varying engine speed has a significant influence on the system matrix. Online system identification was therefore implemented for automatic adaptation of the system transfer matrix G j for the current working point, FIGURE 1. The cycle-variant plant model is adapted using a Bayesian minimumvariance estimator. To this end, in every engine cycle the statistical distribution p( ) is estimated of the whole impulse response g j of the plant: Eq. 3 p(g u,y) ~ p(u, y g)p(g) Although the impulse response of the combustion process is infinitely long in theory, the limited cycle time allows the system to be modeled as FIR filter. The model can thus depict any orders and response times. The system matrix G j FIGURE 2 Controlling an isobaric process by continuously varying the injection rate (top) and adapting the plant model (bottom) (qualitative) ( IAV) 54

4 FIGURE 3 Schematic diagram showing combustion procedure transfer to the derivative calibration ( IAV) then results from positioning the mean vector of p(g j ) (maximum a-posteriori method) in a Toeplitz structure. An example sequence for adapting g j to a specific operating point is shown in FIGURE 2 (bottom). It transpires that the response time of the system or the sum of mean injection and ignition delay, is successfully learnt. FIGURE 2 (top) shows an example for controlling an isobaric process. CONTROL CONSTRAINTS OF PRODUCTION INJECTION SYSTEMS cannot be handled due to the principle. Tests have shown that problem-adapted modification of controller synthesis as per (2) in TABLE 1 is expedient for the production injection system. Attention should be drawn to a new constraint which restricts the L 0 pseudo-norm u 0. It is thus possible to restrict the maximum number of individual injection events. To solve the functional, a modified form of the matching pursuit algorithm is used [3] that efficiently finds suboptimal yet adequately exact solutions. RESULTS OF COMBUSTION PROCEDURE TRANSFER The objective of derivative calibration consists in reproducing the system behavior of a lead project with minimum calibration effort. Calibration of the engine parameters is evaluated using emission, comfort and consumption variables. The following section also uses these variables to evaluate the transfer of combustion procedure design, FIGURE 3. In this method, the rate of heat release is transferred from the The second development step consists of tests on a standard passenger car diesel engine, equipped with an engine control unit with an in-circuit emulator (ICE) and injectors with solenoid valve technology. One cylinder in the test engine has a piezo-resistive pressure transducer integrated in the glow plug as a standard feature. The signal is read directly from the engine control unit using the ICE interface. The injection parameters are also varied using this interface. While the injection system of the single-cylinder engine realizes nearly any injection profile, there are strong restrictions in the actuating signal in the production engine. In particular, the number of injections per engine cycle is restricted, and continuous variation of the injection rate is not possible without great effort. These characteristics mean that the actuations resulting as a solution of the simple convex functional (1) in TABLE 1 FIGURE 4 Absolute differences in normalized soot emissions (stated as a % of the measured value range) ( IAV) MTZ worldwide

5 FIGURE 5 Actuation and rate of heat release (top) (qualitative) and target variable achievement (bottom) illustrated at a typical operating point ( IAV) FIGURE 6 Actuation and heat release profile (top) (qualitative) and target variable achievement (bottom) illustrated at a typical operating point with considerable deviations from the lead project ( IAV) 56 lead project to the derivatives as one of the key attributes defining the combustion process. The injection parameters to be calibrated automatically result from the associated actuating variables of the heat release controller. In a broad map range, heat release profiles controlled with different injection configurations were recorded at 23 operating points of a virtual lead project. These operating points were then set again by the heat release controller starting in trailing-throttle mode. Start and quantity for each injection were defined solely by the controller. All relevant target values were measured for both the open-loop and closed-loop controlled measuring points. The measured value range of each variable was normalized to the interval % for uniform evaluation at different operating points. For example, FIGURE 4 shows the deviations in particulate matter emission of an automatically generated derivative calibration versus reference calibration of the lead project. The mean absolute error of the normalized variables is below 5 % at 20 operating points. FIGURE 5 shows the rate of heat release and actuating signal of a representative operating point for open-loop and closedloop control mode. There is a high degree of consistency between the characteristic attributes of heat release, such as start of combustion and heat rate profile as well as the actuation curve. This consistency also applies to the manifestation of the target variables. The deviations are partly within the measuring tolerance. In this case, none of the normalized target variables deviates by more than 2 % from the reference measurement in the lead project. In order to examine the limits of the presented method, operating points were systematically selected whose injection configuration exhibits actuating signals resulting in highly timevariant plant behavior within an engine cycle. For example, FIGURE 6 shows an operating point with high response time variance. It can be expected that the ignition delay of the first pilot injection deviates greatly from the mean response time of the time-invariant plant model g j. Despite the low value for the control error norm, the actuations are fundamentally different. The NO x emissions determined mainly by combustion chamber temperature with adequate

6 excess air reveal comparatively good consistency on account of the similar rate of heat release. However, consumption and the emission values for HC and CO resulting from incomplete combustion deviate considerably from their reference values. OUTLOOK The presented method was shown to make derivative calibration more efficient. Problems arise with combustion procedure transfer when partially homogeneous combustion concepts hamper modeling and closed-loop control of the system. However, such problems can be encountered with further development steps. In addition to derivative calibration, other innovative processes are conceivable in combustion process calibration. It is possible, for example, to integrate model-based concepts as described in [4] or [5] for comprehensive optimization of heat release profiles or other variables derived from in-cylinder pressure traces. These processes could be used to apply both digital and also continuous rate shaping without limiting the flexibility of such injection systems with parametric injection profiles. REFERENCES [1] Hinkelbein, J.: Verbrennungscharakteristikregelung mittels Einspritzverlaufsmodulation bei direkt einspritzenden Dieselmotoren. Aachen (Germany), Rheinisch-Westfälische Technische Hochschule, Dissertation, 2010 [2] Bristow, D. A.; Tharayil, M.; Alleyne, A.G.: A Survey of Iterative Learning Control. In: IEEE Control Systems Magazine, 2006, pp [3] Mallat, S. G.; Zhang, Z.: Matching Pursuits with Time-Frequency Dictionaries. In: IEEE Transactions on Signal Processing, 1993, pp [4] Holzendorf, S.: Heat-release-controlled Combustion Calibration. In: MTZworldwide 77 (2016), No. 5, pp [5] Zaccardi, J.-M.; Nicolas, F.; Rudloff, J.; De Paola, G.: Optimized Diesel Engine Emissions and Efficiency by Means of Numerical DoE. International Calibration Conference Data Analytics, Methods and DoE, Berlin (Germany), 2017 MTZ worldwide

7 Driver Assistance Systems From Assistance to Automated Driving 4th International ATZ Conference on Automated Driving 18 and 19 April 2018 Wiesbaden Germany SENSOR SYSTEMS Environment recognition, interfaces, vehicle dynamics models, vehicle platooning, driverless systems MACHINE LEARNING Deep learning, neural networks, integration of human driving data NEW METHODS AND PROCESSES Intelligent testing, cloud-based validation, connected development /// KEYNOTE LECTURES Dr. Michael E. Hafner, Daimler AG Ralph Lauxmann and Christian Schumacher, Continental Prof. Dr. Gernot Spiegelberg, Siemens AG /// PARTNERS /// SPONSORS Abraham-Lincoln-Straße Wiesbaden Germany 58 Phone Fax PROGRAM AND REGISTRATION