AT501578B1 - behavior analysis method for analysing the behaviour of complex systems especially combustion engines, by modelling using interpolation of principle influential parameters for all feasible constellations of input variables - Google Patents

Abstract

The method involves selecting different measuring points which correspond to different constellations of the measuring parameters and carrying out measurements for determining measurement parameters for a real system. A model is then formed which represents the dependence of the measurement parameters on input variables and the model is then calibrated with the measurement values obtained for the real system. The model is divided into at least two sub-models. A first sub-model is formed which represents a first subset of the measurement parameters. A first main influential parameter is identified for the first partial model. An optimal value for the main influential parameter is determined for each measurement point. These parameters for all feasible constellations of input variables are interpolated for calibration of the first partial model. A further partial model is generated to show a further subset of measurement parameters in dependence on input variables and the previously derived first subset of measurement parameters. A further main influential parameter is identified for the further model and the optimal value for this parameter is determined for each measurement point. The further influential parameters for all feasible constellations of input variables are interpolated for calibration of the further partial model.

Description

2 AT 501 578 B1

The invention relates to a method for investigating the behavior of complex technical systems, in particular of internal combustion engines, and for obtaining design parameters by forming a model representing different measured variables as a function of input variables, with the following basic steps: selecting different measuring points, the different ones Matching constellations of measures, and performing measurements to determine measures on a real system; - Creating a model that depicts the dependence of the measured variables on the input variables io and calibrating the model based on the measured values of the real system obtained at the measuring points.

In many fields of technology, it is necessary to model complex systems through models in order to gain information about this system and to carry out development work.

A well-known problem is the so-called frontloading, which involves integrating simulation and analysis already in the early concept or design phase of a new product, so that as many important development decisions as possible are made by simulation. virtual attempts to be secured.

This is particularly important in view of the fact that the implementation of measurements on real systems is complex and the measured values are often not obtainable in real time. A standard method for carrying out simulations is that the system to be examined is mapped by a simulation model that reflects the basic behavior of the system. This simulation model is parameterized and calibrated by a number of measured values, which are obtained on a real system, so that a sufficiently exact correspondence between simulation model and real system is achieved. After the simulation model has been available, further measured values can be calculated in large numbers and 30 with little effort.

Although a method of this kind is universally applicable in principle, it is not practical in all cases. For example, in the area of engine development, it is necessary to develop engine control units at a time when at least initially no real engine data is available at all. Only at a later date is it possible to obtain real data on test benches, but the number of these data is usually much smaller than the number of data that can be calculated by a simulation model.

In order to map complex systems to a minimum of accuracy, it is generally necessary to have enough data to calibrate such a model. In the following, measurable variables are referred to as measured variables which can be measured on the one hand on a real system and on the other hand are to be imaged by the simulation model in order to be able to study the behavior of the system without carrying out further measurements. Typical measuring factors in the development of internal combustion engines are: inflowing air mass, indicated mean effective pressure, maximum cylinder pressure, and so on, exhaust gas temperature pre-catalyst or turbine.

Possible input variables are: - Speed 55 - Injection quantity 3 AT 501 578 B1 - Charging temperature - Boost pressure - Exhaust back pressure - Air-fuel ratio in the inlet 5 - Wall temperatures in the combustion chamber - Start of injection - Ignition delay - Burning time - Vibe form factor 10

The first five sizes can be taken from the test bench measurements; the air-fuel ratio is set to infinity, since there is no exhaust gas recirculation and thus no fuel in the inlet; the start of injection is read from the data of the control unit (ECU maps) during test bench measurement. 15

The ignition delay to ultimately determine the actual start of combustion, the burn time, and the shape factor are set to estimated values depending on speed and load to describe the Vibe burn function to simulate combustion. The detailed determination takes place in the course of the optimization. 20

In practice, often only a small number of real measured values are available, so that it is not meaningfully possible with conventional methods to represent the complex and multivariable relationships between the multiplicity of input variables and the multiple measured values in a valid manner. 25

Object of the present invention is to provide a method which avoids these disadvantages and which makes it possible to create meaningful models even with a small number of real data obtained, which have high prognosis quality. In the context of the method according to the invention, the following steps are carried out in detail: subdivision of the model into at least two submodels; - Creating at least a first sub-model, which images a first subset of the measured quantities 35; - identifying at least one first main influence parameter for the first partial model; Determining an optimum value of the first main influence parameter in each measurement point; Interpolating the first main influence parameter for all meaningful constellations of 40 input variables for the calibration of the first partial model; - Creating a further submodel for mapping a further subset of the measured variables as a function of the input variables and the previously determined first subset of the measured variables; Identifying at least one further main influencing parameter for the further submodel; Determining an optimum value of the further main influencing parameter in each measuring point; - Interpolation of the other main influencing parameter for all meaningful constellations of input variables for the calibration of the further submodel. 50

The essential feature of the present invention is that knowledge of the basic behavior of the underlying physical system is generally used in the creation of the simulation model in order to improve the quality of the model. An essential concept here is the division of the overall model into sub-models, which are causally dependent on each other. 4 AT 501 578 B1

In particular, the causality relationship can be optimally taken into account by calibrating several partial models one after the other, the measured variables determined in each partial model representing input variables for subsequent partial models.

As such, it is possible to simplify the calculation by determining a smaller number of measurands in each submodel than are to be calculated in total. However, a particularly simple and efficient calculation is made possible in a preferred embodiment variant in that exactly one measured variable is determined in each case in a submodel.

A further preferred embodiment variant of the method according to the invention provides that a single main influencing parameter is assigned in a measured variable and that the calibration of the associated submodel is effected by full-factorial variation.

Particularly preferred is a variant of the method in which the optimum value of the main influencing parameters in the measuring points is determined by calculating for different discrete values of the main influencing parameter, i. is calculated at predetermined interpolation points, a desired value for the respective measured variable, from the values thus calculated, an interpolation model is created which indicates desired values for the respective measured variable for any values of the main influencing parameter, then a difference between these desired values and the measured actual value of the measured variable is minimized and from this the optimal value of the main influencing parameter is deduced. As an interpolation model, a polynomial approximation can be performed, but neural networks can also be used for this purpose. The basic idea is that the search for the optimum values can be interpreted as minimizing the amount of the difference between the setpoint and the actual value, and that prefabricated tools are often available for such minimization. The minimization of an interpolated magnitude function of the difference, however, leads to relatively large errors, since even in the optimum continuous differentiability is not given. The preferred solution leads to much better results.

This procedure is compiled in a modeling algorithm, which is also called "ModelFormula".

As a result, the invention will be explained in more detail with reference to embodiments illustrated in the figures. 1 shows a diagram in which measuring points are entered in a characteristic map, which is formed from the rotational speed and the injection quantity; FIG. 2 is a block diagram describing the essential components of a motor under test and the measuring system; FIG. FIG. 3 is a block diagram describing a simplified measuring system; FIG. FIG. 4 are diagrams explaining the VIBE factor; FIG. FIG. 5 shows a diagram corresponding to that of FIG. 1, in which the measurement points actually used are entered in the map; FIG. 6 is a diagram explaining the method according to the invention; FIG. FIG. 7 shows diagrams showing the quality of the method according to the invention; FIG. FIGS. 8 to 14 are further diagrams showing the quality of the method according to the invention; Fig. 15 is a bar graph showing an evaluation of various methods; FIG. 16 shows a further diagram corresponding to that of FIG. 1 and FIG. 5, in which the measuring points used for validation are entered in the characteristic field; FIG. 17 shows further diagrams corresponding to those of FIG. 15; FIG. FIG. 18 is a diagram describing the measurement points used; FIG. FIG. 19 is a bar graph plotting the percentage error of air mass plotted versus training points by speed and load; FIG. FIGS. 20, 21 and 22 are graphs describing variations of the flow coefficients of the scaling factor 5 AT 501 578 B1 and the boost pressure offset, respectively; FIG. 23 is a table illustrating optimization results; FIG. FIG. 24 is a bar graph showing the percentage error of air mass at training points before and after optimization; FIG. Fig. 25 is a diagram explaining the charge pressure offset model; FIG. 26 is a further bar graph showing the percentage deviation of the air mass in the validation points; FIG. FIG. 27 is a map of the boost pressure offset; FIG. FIG. 28 is a bar graph showing the percentage deviation of air mass in the validation points after transmission by interpolation; FIG. FIG. 29 is a bar graph showing the percentage errors of the combustion quantities IMEP and PFP; FIG. - Fig. 30 Characteristics for the burning time, the start of burning and the form factor determined by interpolation, plotted by means of speed and load; Fig. 31 is a diagram showing the variation of the burning time at one operating point; FIG. 32 is a diagram showing the variation of the Heat Transfer Calibration Factor at an operating point; FIG. FIG. 33 is a table showing the optimization results for transmitting the burning time; FIG. FIG. 34 is a bar graph showing the percentage error of the maximum cylinder pressure in the training points before and after the optimization; FIG. FIG. 35 is a table showing the optimization results for determining the calibration factor; FIG. FIG. 36 is a bar graph showing the percentage error of the indicated mean pressure in the training points before and after the optimization; FIG. FIG. 37 is a diagram illustrating the combustion duration model (FFN); FIG. FIG. 38 is a diagram describing the Calibration Factor Model (FFN); FIG. FIG. 39 is a bar graph showing the percentage deviation of the maximum cylinder pressure in the validation points; FIG. FIG. 40 is a bar graph showing the percent deviation of the indicated mean pressure in the validation points; FIG. FIG. 41 is a diagram showing a map of the burning period; FIG. FIG. 42 is a diagram showing a map of the calibration factor; FIG. FIG. 43 is a bar graph showing the percentage deviation of the maximum cylinder pressure in the validation points; FIG. FIG. 44 is a bar graph showing the percentage deviation of the indicated mean pressure in the validation points; FIG. FIG. 45 is a bar graph showing the percentage error of the exhaust gas temperature in the base; FIG. FIG. 46 is a diagram showing the variation of the heat transfer factor; FIG. FIG. 47 is a table showing the optimization results for determining the heat transfer factor; FIG. FIG. 48 is a bar graph showing the percent exhaust gas temperature error at the training points before and after optimization; FIG. FIG. 49 is a diagram explaining the heat transfer factor model (FFN); FIG. FIG. 50 is a bar graph showing the percentage deviation of the exhaust gas temperature in the validation points; FIG. FIG. 51 is a diagram illustrating a map of the heat transfer factor; FIG. FIG. 52 is a bar graph showing the percentage deviation of the exhaust gas temperature in the validation points; FIG. FIG. 53 is a diagram showing the measurement points for direct calibration of the simulation models; FIG. FIG. 54 is a block diagram showing the percentage deviation of the outputs in the validation points; and FIG. 55 is a block diagram showing the work flow for the calibration calibration of the calibration data

Represents simulation models.

The experimental vehicle used was a modern diesel commercial vehicle engine which has the following features: - 4 cylinders; - 4 valves; - Approximately 3 liters of displacement; - Wastegate turbocharger with intercooler; io - common rail injection system; - cooled EGR route; - Meets Euro 4 - emission standard.

In the course of the test bench measurement carried out, various operating points 15 were recorded, which are distributed over the entire speed-load characteristic map, as shown in FIG.

The total of 41 measuring points were measured both with and without exhaust gas recirculation (EGR). The boundary conditions, such. B. boost pressure and temperature, Abgasge-20 gendruck and start of injection, the operating points were varied accordingly. This results in a good distribution of the measuring points over the entire map.

Measurements relevant to this work include 25 - the air mass measured in the intake system, - the indicated mean pressure, - the maximum cylinder pressure, - the air-fuel ratio, - the exhaust gas temperature upstream of the turbine. 30

As part of the engine development, a basic model was created, which includes the following components, among other things, the base model has been created with a charge cycle simulation program called " AVL-BOOST " is available. As a result, this base model is therefore also called the BOOST model: 35 - The characteristic map of the friction mean pressure was calculated from measured cylinder pressures. - A full model of the turbocharger was used, which is based on the compressor and turbine maps of the manufacturer. The manufacturer's maps were extrapolated using stationary measurements. - An external model was created to simulate the specific behavior of a wastegate valve on a real engine (increasing boost at speed). 45 - Combustion simulation was based on a model (MMC, ie Mixture Controlled Combustion), with which the firing curve can be calculated from injection rate and internal cylindrical states. This model has been extended by a HSDI Part (High Speed Direct Injection), which simulates premixed combustion and pre-injection. 50 - In order to model the closed-loop EGR control loop, a MATLAB / SIMULINK model was created which incorporates both the intake air excess air mechanism and the EGR shutdown at high loads. This model was coupled to the 55 BOOST model by incorporating a Matlab DLL interface. 7 AT 501 578 B1

Fig. 2 shows the BOOST full model for the engine described above.

The ambient air is drawn in via the system boundary SB1 and directed via the air filter CL1 to the compressor TC1. Thereafter, the compressed air is fed to the intercooler C01. 5 Another pipe directs the control pressure to the actuator of the externally modeled wastegate valve. The tubes 4, 5 and 20 serve as a connection between intercooler and intake manifold PL1; the tubes 6, 7, 8 and 9 represent the intake ports between the intake manifold and the four cylinders C1, C2, C3 and C4. io The combustion products are brought together after the cylinders and directed to the input of the turbocharger TC1. About a throttle point R3, the exhaust back pressure can be adjusted.

The engine is equipped with an exhaust gas recirculation system. The exhaust gas for the EGR is withdrawn at 15 node J2 and finally returned to the intake flow through an EGR cooler C02 and the throttle R1. This throttle point represents the EGR

Valve.

The recalculations of the measured operating points showed very good simulation results. Particularly the sizes effective medium pressure as well as effective consumption could be mapped down to the measuring accuracy (<1%).

As part of the investigation, this BOOST full model was reduced to a core model in which only the four cylinders with the respective connecting pipes are mapped, as shown in FIG.

There are various reasons for this simplification: - The parameterization effort of such a slimmed-down engine model is much lower than with a complete model including the intake and exhaust system. - The calculation times of the simulation are shortened by a multiple. - The use of these core models offers the possibility of creating standard models for the various engine variants.

However, the simulation with a core model also has several disadvantages: The dynamic processes within a work cycle can no longer be mapped since the entire pipe system is not displayed. - Due to the lack of EGR route an air-fuel ratio in the inlet must be specified to simulate exhaust gas recirculation. 45 - There will be constant values at the system boundaries, so that the gas dynamics, which is taken into account in an overall model, can not be mapped.

These deficiencies shown here lead to a worsening of the simulation results of the core model compared to the full model. 50

Since the exhaust gas recirculation is simulated via an air-fuel ratio in the inlet, a relationship between the exhaust gas mass flow and the A / F ratio must be established. For the derivation of the equation, the state of the gas immediately before the intake valve 55 is considered. The entire EGR mass is already supplied at this point in time and has the same composition as the mass in the cylinder at the high-pressure end, the δ ΑΤ 501 578 Β1. The EGR mass is calculated from the formula for the EGR rate:

X

AGR m

AGR

mL + mAGR _ mAGR m, "(1) with: io Xagr Exhaust gas recirculation mAGR AGR Mass flow mL ... Fresh air mass flow min .. total mass flow in the intake duct 15 Further applies to the percentage of fuel in the exhaust gas and thus also in the EGR mass flow: mB _ mAGR, B mB _ 1 _ 1 20 m0ut "mAGR" mL + mB &quot; n \ + 1-A / F + 1 '

mb with: 25 mB ... fuel mass flow; m0ut total exhaust gas mass flow; A / F. Ratio of fresh air mass to injected fuel mass.

Since both excess air and burnt fuel are present in the EGR mass flow 30, the A / F ratio in the inlet results from the ratio between the fresh air mass present and the fuel mass burned: Λ / I- _ min, L _ mL + mAGR mAGR, B / "v ™ Πη" m ~ Z · \ ö)

35 1 ln, B mAGR, B

Assembling equations 1, 2 and 3 gives a relationship between the EGR rate and the air-fuel ratio in the inlet and the ratio of fresh air mass to injected fuel: 40 AGR = 1 + A / F 1+ A / Fin

with A / F = - mB (4) 45 With this Equation 4, the exhaust gas recirculation rate for the given A / F ratio in the inlet, the given injection quantity and the simulated fresh air mass can be determined with hindsight and thus the EGR rate as the input value for the Model can be used.

To use the EGR rate directly as an input, the actual air-to-fuel ratio (A / F or λ) would have to be introduced as another variable of variation. With equation 4, however, the exhaust gas recirculation rate for the predetermined A / F ratio in the inlet, the predetermined injection quantity and the simulated fresh air mass can be determined in retrospect, and thus the EGR rate can be used as an input variable. In the course of the experiments, among other things, the Vibe function was used as a combustion model in BOOST in order to simulate the time course of the fuel conversion.

By observing the combustion process of homogeneous fuel-air mixtures, Vibe has found that an exponential curve is a good approximation to real combustion processes. From these findings, Vibe developed the burn-through function (also called conversion rate), which is defined as: 10 c x = 1-e m + 1 t 'to (5) with: x. Conversion rate (ratio of burned to total fuel mass); C .... conversion parameters; 15 t0 ... burning time; m .. form factor.

Under the arbitrary stipulation that the total fuel should be converted to 0.1% at time to, the numerical value C = -6.9 is obtained for the constant C.

By deriving the burn-through function, the combustion process is obtained, which describes the instantaneous combustion or conversion rate at any time during combustion. 25 X = dx d (t / t0) = 6.9- (m + 1) -6.9 (6)

The Vibe form factor describes the shape of the firing process. The influence is illustrated in FIG. 4.

It can be seen that the larger the m-factor, the later the energy conversion takes place (diagram a). Also the center of gravity, ie the point of the combustion process at which 50% of the fuel was converted into heat, shifts "late" with increasing m-factor direction. A value of the shape parameter m of 2.3 corresponds approximately to a symmetrical combustion curve (diagram b).

In a diesel internal combustion engine, the firing curve is most likely to be described by a form factor of between 0.1 and 1, since the high compression and the high pressures result in rapid conversion of the fuel.

An essential component of the method according to the invention is the use of polynomial regression models. In the case of polynomial regression models, it is attempted to model the output variables of the engine (eg torque, cylinder pressure, etc.) with the input variables (eg speed, injection quantity, start of injection, etc.) using polynomial functions.

At the beginning there is an empirical modeling. With a mathematical model, the input variables of the system are linked to its output variables. For each output a separate model is set up. For the polynomial functions, an order (up to 10) can be specified. In addition to the main effects, the interactions between the input quantities (interaction terms) are also taken into account (equation 7). 55 u = a0 + a ^ + a2x2 + ... + ajXj + ... + b ^ X ;, + ... + ^ x * + ... + Cj xp + ... (7) 1 0 AT 501 578 B1

The accuracy of the model estimation grows with the number of measurements, but it increases the effort.

The actual model estimation is done with the help of the least squares method. 5 The model coefficients are estimated so that the squares of the squares are minimized. The error is equal to the deviation of the model calculation from the support values.

The quality of the resulting model is checked with various statistical test functions. The individual polynomial terms are checked for their significance and eliminated if necessary. This principle of automatic reduction of the model order results in a simplified model that is easy to interpret while reducing the number of degrees of freedom. The non-significant model terms are removed from the model, thus improving the predictive quality. Also, over-determined models can be avoided, which often have very poor behavior between the support points.

In addition, the algorithm can transform outputs to give the best output distribution. This creates a model that is exactly adapted to the data. 20

The basis was a total of 2880 simulation data sets calculated by a full-factorial variation of the parameters shown in Table 1 using BOOST: speed, injection quantity (mB), air-fuel ratio (A / Fint) to exhaust gas recirculation simulate intake manifold pressure (p2) and temperature (T2) and exhaust back pressure (p3). 25 The start of injection (SOI) was taken from a map as a function of speed and load and corresponds to the measured data; the coolant temperature (Tcoolant) was not varied.

Table 1: Variation parameters for BOOST calculation 30

Variation yes yes yes yes yes yes no no Parameter speed mB A / Fint 72 p2 P3 SOI Tcoolant [1 / min] [%] H [K] [Pa] [Pa] [° KW v. OT.] [K] 500 9 150 293 90000 100000 according to characteristic diagram 353 1000 36 250 305 130000 200000 2000 73 350 320 190000 350000 3000 100 100000 250000 3400 Number of Variations of Variations Parameter 5 4 4 3 4 3 1 1 2880

Table 1: Variation parameters for BOOST calculation 45

Furthermore, it must be ensured that the limits of variation are outside the available measurement data: For example, the speed on the test bench was varied between 1000 and 3200 rpm, whereas in the simulation it varied from 500 to 3400 rpm. The reason for this is that the modeling thus requires a support point network distributed as evenly as possible over the test space so that the model quality reaches an acceptable value.

The output quantities that are relevant for the engine component "Cylinder" are determined from the results of the BOOST calculations and are as follows: 55 1 1 AT 501 578 B1 - Air mass on / off (M_Lx) - Enthalpy flow on / off (M_Enthx) - Indexed Medium Pressure (IMEP) - Wall Heat Flow (WHF) 5 - Maximum Cylinder Pressure (PFP) - Air to Fuel Ratio (A / F)

These values are modeled with a simple regression model depending on the eight input variables and thus provided for the real-time simulation. 10 5). The 9 remaining records served map spread and thus make one available in the simulation:

From the 41 measuring points available for the test carrier, 32 operating points were selected as the training data set (see later for the validation of the results.) 15 The selected operating points are, on the whole, good basis for a map-wide improvement.

The input variables are all 8 variables such as 20 - speed (speed), - injection quantity (mB), - EGR rate (EGR), - charging temperature (T2), - boost pressure (p2), 25 - exhaust backpressure (p3), - start of injection (SOI), - Coolant temperature (Tcoolant).

As outputs, only the sizes 30 - incoming air mass, - indicated mean pressure, - maximum cylinder pressure, - air-fuel ratio 35 can be provided. The enthalpy flows and the wall heat flow are not measured and could, if necessary, be calculated by a heat balance.

To summarize the simulation and measurement data for modeling, a number of different possibilities were investigated: 1.) offset to the output, determined by the mean arithmetic error; 2.) offset to the output, determined by the mean squared error; 3.) factor multiplication of the output quantity; 45 4.) Equivalent addition of the measurement data to the simulation data; 5.) Equivalent and repeated addition of the measurement data to the simulation data; 6.) Introduction of a Bockfaktor as another variation parameter (solution according to the invention).

These six options are described in more detail below: 50

The first possibility 1) corresponds to the original assumption: It is assumed that the BOOST simulations can represent a good correlation of the individual engine sizes, but have a map-wide level shift. This level shift should be compensated for by simply raising or lowering the model. 55 12 AT 501 578 B1

In this case, the delta by which the model is raised or lowered is determined by the mean arithmetic error: 1 1 ^ ^ arith - T ^ (L * Meas '* - * Bo0st)' Π n with: io UMess Measurements: Ußoost values of the simulation model with measured input quantities; n ..... number of measuring points.

This delta is added to the respective output after the regression, ie after the model has been formed:

Uneu = U + Aarith = a0 + Aarith + + ... + ajXj + ... + btXtXi + ... + Ct X? + ... + C | X ?. (9)

The variant 2.) corresponds to almost the first possibility 1.). Here, however, the mean 20 square error is used as a basis to calculate the offset:

(10) 25

The mean squared error is multiplied by the normalized arithmetic error of option 1 to preserve the sign, i. whether the model is raised or lowered. The quadratic error causes larger deviations to be weighted higher than smaller ones, so that the model is more closely aligned with the noticeable differences. Possibility 3.): The multiplication by a factor calculated from the relationship between the measured values and the simulation values (equation 11) is based on the idea that the actual air expenditure at high loads is often too low and at low loads correct is shown. 40 1 1 α = -Σ n n r - \ ^ Measuring ^ Boost ^ (11)

If you multiply by the factor calculated in this way (equation 12), it depends on the absolute value of the output quantity by how much the model rises or falls: z. B. larger air masses are corrected more than smaller air masses. 45

Uneu = Ο · U (12) Possibility 4.) is the simplest of the six possibilities. The measured data are inserted equally into the list of simulation data. Subsequently, a normal regression is performed so that a mixed second order model of simulation and measurement data arises.

Even with option 5.), the measurement data are added unchanged to the simulation data prior to modeling. However, the measured data sets are multiplied (10 times in this case) in order to place a greater weighting on the few 32 measured values, in contrast to the 13 AT 501 578 B1, of almost 3000 simulation values.

The sixth and last possibility according to the invention also incorporates the measurement data into the modeling, but a further variation parameter is introduced. This newly added block 5 factor is set equal to zero for the simulation data and equal to one for the measurement data and causes the model to skew towards the measurement points.

In Fig. 6, the operation is shown vividly. In the front the values of a logarithmic function (C = log (A) + 1) are plotted over their input quantity A. These function values are represented by a 2nd order polynomial (broken line) and have the block factor B = 0. For this data, an additional 5 points with B = 1 are added whose values are slightly above or slightly below the respective function value. The model now gets its shape through the many points on 15 of the "B = 0" side. However, on the "B = 1" side, it tilts easily to the five nodes that represent the measurement data.

Due to the fact that the model is described by a 2nd order polynomial, the Bock factor has only a linear influence as well as interactions with the other input quantities. Thus, one can rule out that the model runs on the "B = 1" side in the extrapolated area, ie outside the existing support points to infinity.

With this method, it is possible to weight the measured data particularly, but to specify the general form of the model by means of the simulation data. Also, a slight modification of the model is allowed without adding a constant map-wide offset. For the actual output after modeling, the block factor is set equal to one (equation 13), because only then is the overall model used. 30

Uneu = a0 + d0B + a ^ + ... + a, Xi + ... + b ^ x, + ... + d ^ B + ... + djXjB + + ... + Cjxf = (a0 + d0) + (a! + d ^ X! + ... + (aj + d,) Xi + ... + b ^ Xj + ... + 0 ^ + ... + c, xf (13)

The following results of the training data are shown by a representative rotation number. The other bases show very similar results and support the results explained here.

The following is a summary review of the results based on the 9 validation data. 40

The graphs show the respective normalized output, plotted against the load, at 2480 rpm. For normalization, the highest measured value was divided. 45 are each shown 4 load points, both without EGR (exhaust gas recirculation) and with EGR. For clarity, only 3 options per diagram are shown. The names in the legend correspond to the six possibilities: 1.) SimModell + arith.Offset - &gt; Offset to the output, determined by the average arithmetic error; 2.) SimModel + quadr.Offset - &gt; Offset to the output, determined by the mean quadratic error; 3.) SimModel * Factor - &gt; Factor multiplication of the output variable; 4.) SimMessModell 1fach - &gt; Equivalent addition of the measurement data to the simulation data; 55 14 AT 501 578 B1 5.) SimMess model 10-fold - &gt; Equivalent and multiple addition of the measurement data to the simulation data; 6.) SimMessModel B = 0/1 - &gt; Introduction of a Bock factor as another variation parameter. 5 The curve with the large square corresponds to the measured values, those with the rhombus the simulated without improvement (-> SimModel, basis).

The 4 outgoing air masses, indexed mean pressure (IMEP), maximum cylinder pressure (PFP) and air-fuel ratio (A / F ratio) are compared with the respective io measurement data separately from each other.

The simulated air mass shows very good results even in the starting point in the points without EGR (FIG. 7). Also in the points with EGR the qualitative accuracy is sufficiently good: the course of the curves coincides (Figure 8). 15

It turns out, however, that the simulation of the exhaust gas recirculation with the core model is very problematic: one obtains a large level difference between the measured and the simulated air mass. Due to this level difference, the mean error is quite large and this error determines in the first 3 ways the adjustment. Consequently, one obtains a deterioration in the points where the pure simulation model has already shown very good results (Fig. 7) and hardly any improvement in the large difference points (Fig. 8).

The simple addition of the measurement data, both simple and ten-fold, has positive effects, but here one recognizes the tendency that the model is based neither on the measured data nor on the simulation data.

The idea according to the invention of balancing through the introduction of a block factor provides the best results both in the operating points without EGR (FIG. 7) and in the points with AGR 30 (FIG. 8).

With the indicated mean pressure, it becomes clear that the curve of the unimproved simulation model is not always qualitatively correct. Here, the simulation gives values other than the measurement. Also, one can see that the full load values are well reproduced, while the part load is mapped with large errors. This shows that BOOST is very well suited for the full-load design, but that there are still shortcomings in the forecasting of the partial load.

Due to the fact that the curve of simulation and measurement is not identical, 40 the first three possibilities are not able to compensate for the error, since in all three cases a constant factor is calculated, which is added or multiplied (FIG 9, Fig. 10). A big disadvantage z. For example, when multiplying by one factor, it is precisely the high IMEP values of the full load that are corrected more strongly than the completely incorrect partial load variables. Thus, the advantages of the simulation program BOOST are not used in Volllastberech- 45 tion.

Figures 9 and 10 show that the introduction of the block factor causes the model to be matched to the measurement points, even in the points with large differences. This possibility also provides the best results here. 50

But also by the simple addition of the measured data good results are achieved, whereby the repeated insertion is better.

When adjusting the maximum cylinder pressure, further disadvantages of the first 3 possibilities are evident: 1 5 AT 501 578 B1

Since the measured value is once below and once above the simulated value, the deviations cancel each other out, so that the mean error and thus the offset or the factor is almost zero. For this reason, there is no improvement of the model, with the arithmetically determined offset even a slight deterioration in the 5 high load points (Fig. 11, Fig. 12).

On the other hand, the other three ideas have an improvement, whereby again the error between measurement and simulation can be reduced most effectively with the aid of the block factor (FIGS. 11, 12). The 10-times insertion of the measured data gives better results than the simple addition. This can be explained by the higher weighting of the measuring points.

The output "air-fuel ratio" has a very large error in the simulation at low load. At the high load points the model agrees with the measurements 15.

This 63% variance in both the low load points without EGR (Figure 13) and those with EGR (Figure 14) results in a relatively large mean error being calculated that significantly degrades the model at full load points (Figure 13) , Fig. 14). In particular, the possibility of offset addition, determined by the mean arithmetic error, causes the air-fuel ratio at full load to drop below the stoichiometric air requirement, that is, λ &lt; 1. This leads to high consumption and high emissions in a diesel engine. For this reason, the model matching with an offset addition is not suitable in this case. Modeling by combining the simulation and measurement data before the regression has a very good potential for improvement despite the large differences in the partial load; the already well imaged full load is hardly affected, whereas the partial load is considerably corrected (Fig. 13, Fig. 14). However, it also shows again that the best results can be achieved by introducing the block factor. At the operating points 30 with EGR, the error is even reduced to zero (Figure 14).

After this detailed description of the individual possibilities, their mode of action and their effects on the respective models of the different output variables on the basis of a representative speed (2480 rpm), the following block diagrams show the through-35 average errors over the entire training data for each output variable (Fig. 15). The value of this error is calculated as the sum of the amounts of deviations divided by the number of training points (Equation 14): 1 n 1 n i i 40 Au - ΣΔιΐ | - Z uMess -uneu. (14) Γ1 1 ΓΙ 1

The different bars each represent a possible solution (see legend); the black line indicates the "worst case", d. H. the maximum deviation of the respective output variable from the nominal value (measured value).

The diagrams confirm the previous results. As can already be seen from the detailed analysis, the introduction of the block factor leads to the lowest average deviations: the air mass can be corrected from 11% to an average of 2%, while the indexed mean pressure even has an improvement potential of 45% in the training data. Even with the maximum cylinder pressure and the air-fuel ratio, the average deviation is reduced by 12% or 26%. The possibility of simply combining the measurement and simulation data also leads to good results, whereby the repeated addition of the measured data provides better results. The 3 ideas to carry out the comparison via an offset or factor multiplication show on average 16 AT 501 578 B1 hardly any improvement, in some cases even a deterioration.

If one considers the maximum deviations (black line), the scatter after the introduction of the block factor is lowest here as well. Even with respect to the other two possibilities of a combined simulation measurement model, the idea according to the invention can in most cases stand out positively.

It becomes clear that the comparison between simulation data and measurement data by means of a block factor can reduce the air mass error in the points with which the comparison is made by -Io from an average of 11% to a maximum of 7%, as well as for the indicated mean pressure of 51% to a maximum of 25%, for the maximum cylinder pressure of 16% to a maximum of 11% and for the air-fuel ratio from 31% to a maximum of 19%.

In order to check these results, as mentioned at the beginning, 9 of the 15 41 operating points were not used for modeling. These points are for validation and have been arbitrarily selected from the map (see FIG. 16).

The input variables of these operating points, i. the respective speed, injection quantity, boost pressure, etc., were used in the corresponding models for air mass, indicated mean pressure, maximum cylinder pressure and air-fuel ratio, which result after carrying out the individual solutions.

Again, the average error between the actual measurements and the values obtained for each option was calculated (equation 14). Fig. 17 illustrates the associated block diagrams in the same manner as previously in Fig. 15.

The results of this validation data lead to no new findings, although the image is not as clear as in the training data. But here, too, the smallest absolute deviations result after the introduction of the block factor, considering all four output variables. One can see the error of the indicated mean pressure in these 9 operating points z. From an average of 51% to a maximum of 18%. Contrary to expectations, the errors in the air mass after the offset addition or after multiplication by the factor are relatively low. However, this could be due to the selection of validation data, so it has to be assumed that it was there that the model was well corrected right there. For the other output variables maximum cylinder pressure, indexed mean pressure and air-fuel ratio can be determined namely no clear improvement. 40

On the basis of the results shown it can be said in summary that the measurement data comparison with the help of an offset addition or a factor multiplication is not recommendable. Because you often get a deterioration in the areas where the original simulation models are already very good, and hardly an improvement where the models are very bad. This occurs relatively often in BOOST simulations, since BOOST has its strengths in full-load design and has not yet been designed for mapping the partial load. Occasionally, the offset or factor addition may even lead to the calculation of unrealistic output quantities. B. simulates an air-fuel ratio at full load below the stoichiometric air demand. 50

Furthermore, it was shown that the curves are not always qualitatively correct, so that with a constant value it is therefore impossible to achieve a map-wide improvement. The equivalent addition of the measured data to the simulation data, both simple and even multiple, shows relatively good results. However, this method is questionable, since simulation data with similar input quantities are already available. This creates a ambiguity in some model areas that are difficult to handle in modeling. This is also the reason why the improved model is based neither on the original 5 simulation values nor on the measurement points.

Therefore, the most sensible is the introduction of a block factor as a further variation parameter, which is set to zero for the measurement data and zero for the simulation data. In this case, the model adapts to the measured values in the areas where measuring points are present, even if there are great differences from the original simulation model. Where there are no measurements, z. B. in quasi-transient operating areas, only the simulation data are used and thus not the whole map distorted.

Note that the measurements are in the range of variation of the BOOST-15 calculations, as the models are otherwise distorted.

Furthermore, it is advisable to distribute the measuring points well in the test room, in order to achieve as far as possible a map-wide improvement. 20 The following describes the procedure for modeling using submodels:

Initially, sensitivity studies are performed to determine which parameter has the greatest impact on the particular output. Correction factors may be introduced or scaling factors provided by BOOST may be changed.

The output variables used were those measured variables that are mainly interesting for the application. These include 30 - inflowing air mass (M_L1), - indicated medium pressure (IMEP), - maximum cylinder pressure (PFP), - exhaust gas temperature before catalyst or before turbine (TABG), 35 the latter, especially because the temperature is a major priority in the Parameterization of various maps has (inlet temperature turbine or catalyst).

In these sensitivity studies, it has been found that the areas of charge cycle, combustion (IMEP & PFP) and exhaust gas can be considered separately, thus keeping the optimization problem manageable.

From the above it can be seen that the simulation with the help of the core model has very great difficulties with the mapping of the exhaust gas recirculation. Therefore, only the measuring points without EGR were used in this part, so that of the 41 available 45 operating points only 27 data sets can be used. These were also split into 20 training records and 7 validation records (Figure 18). The validation data are arbitrary in the map and serve as in the first section for later review of the process (generalization). so The core model described above is used to simulate the measurement points in BOOST. For this purpose, the various input variables are set operating point-dependent: - speed - injection quantity 55 - charging temperature 18 AT 501 578 B1 - boost pressure - exhaust back pressure - air-fuel ratio in the inlet - wall temperatures in the combustion chamber 5 - start of injection - ignition delay - burning time - Vibe form factor io The first five sizes can be taken over from the test bench measurements as already described; the intake-side air-fuel ratio is set to infinity, since there is no exhaust gas recirculation and thus no fuel in the inlet; the start of injection is determined from the data of the control unit (ECU maps) during the test bench measurement. 15

The general optimization procedure is identical for all four output variables:

After determining the main influence parameter for the respective size based on the sensitivity tests, a full-factorial variation of this size is made for every 20 points of training using BOOST. This provides several solutions for each operating point depending on the selected parameter. It is important to ensure that the desired setpoint can be achieved by setting the factor. If this is not the case, either the range of variation must be increased or another influencing variable must be found which leads to the required setpoint values. 25

In a next step, the BOOST results are imported into the model to determine the exact parameter setting so that the setpoint is reached. The objective function to be defined in the model results from the requirement that the delta between actual and setpoint should be minimized as follows: 30 |

TargetFun = actual value setpoint | = | uBo0st - uMeSsl = min. (15) with: 35 UBoost ...... simulated output variable, varies with parameter UMess ...... measured value (setpoint for the corresponding operating point)

This objective function is modeled with the modeling algorithm "ModelFormula" described above and mapped exactly. The prerequisite for this, however, is that the corresponding 40 output variable is also mapped with a high model quality. For this one uses either a polynomial regression mode or a neural network, depending on which provides the better model quality.

The actual parameter determination is then carried out by local optimization of all 45 operating points (training data). In this case, the minimum of the magnitude function from equation 15 is determined with the aid of an optimization algorithm selected in advance. Because where the magnitude function is equal to zero, the actual and the setpoint of the respective output match. Thus, the parameter setting is determined for each operating point of the training data, so that the exact same output variable can be simulated as exactly as possible.

In order to check whether this methodology also applies to other areas of the performance map and thus also the transferability for an application is given, as already mentioned, seven 55 validation data sets were not included in the optimization process. On the basis of these operating points, the general validity of the optimization results is tested.

For this purpose, in a first step, a transfer of the found parameters from the few nodes of the training data to the entire map as a function of speed 5 and injection quantity made possible. Two possibilities were examined:

Modeling: With the help of simulation programs a model can be laid through the optimal setting parameters. This model is based on a mathematical formula that links the respective factor with the input variables speed and injection quantity. 10

Linear interpolation: An interpolation is carried out between the individual interpolation points so that the intermediate values are calculated via a linear relationship between the respective neighboring points. 15 Both methods serve to enable the transmission of the optimized parameters and thus a map-wide improvement of the simulation model.

The results of the entire optimization process are documented below in three sections: charge change, combustion and exhaust gas. The order of the descriptions 20 corresponds to the procedure for matching the models with this method, since the different variation parameters partly have an influence on the other output variables. For this reason, the procedure may not be changed arbitrarily.

Charge change 25

In Fig. 19, the delta between measurement and simulation, plotted against the arranged according to speed and load training points represented. A tendency can be seen: At medium speeds too little inflowing air is simulated. 30 This actually higher air expenditure is explained by the dynamic pressure pulsations in the intake tract, which can not be depicted in the core model during a work cycle. These are now deliberately used in the development of modern internal combustion engines to increase the degree of filling. 35 To compensate for this disadvantage of the core model, the pressure gradient across the engine must be adjusted so that more or less air gets into the engine depending on the operating point.

The first idea was to influence the air mass by throttling the intake and exhaust tracts. For this purpose, the so-called flow coefficients in the restrictions of the BOOST model were changed. Flow Coefficients scale the flow area and can be adjusted at any transition point in the piping system, in this case for inlet side throttling at the outlet of the plenum (PL1, Fig. 5) and for outlet throttling at junction J1. If these two parameters are varied, a variation of the air mass is obtained. In Fig. 20, the ratio of simulated air mass to the measured air mass is plotted over 45 of the flow coefficients for one operating point (setpoint = 1).

It can be seen that although one has an influence on the air mass, but can not increase the incoming air. For this reason, throttling is not suitable for this type of application. 50

In the second option investigated, a variation of the so-called scaling factor of the intake valve in the cylinder element was performed. Scaling Factor scales the effective flow area of the valve, changing the inflowing air. However, it will be apparent that even with this possibility, the predetermined target value can not be reached (FIG. 21). In addition, the influence on the air mass is far from being so great; 20 AT 501 578 B1 she can z. B. be changed in the operating point shown here only by a maximum of 2%.

The third and ultimately most effective option is based on the idea of directly changing the pressure gradient across the engine by adding an offset to the boost pressure. Because by raising the boost pressure with the exhaust back pressure remaining the same, an increased degree of filling is achieved. By contrast, lowering the boost pressure causes less air to enter the engine. The diagram in Fig. 22 shows a variation of the offset. It can be seen that by this method, the air mass can be adjusted to its desired value. So, in order to optimize the charge exchange, a full-factorial variation of the charge pressure offset was carried out between -100 mbar and +250 mbar and then the pressure equalization required to achieve the desired air mass was determined by the method described above. 15 The results that result after the optimization are shown in FIG. For each operating point (left columns), an optimal boost pressure offset (middle column) is determined so that the delta between the actual and setpoint value is zero. In this case, the optimization can be performed ideally (right column: error = 0.000). 20 If you verify this boost pressure offset by a new simulation in BOOST, you get the desired air mass in almost all 20 training points. In Fig. 24, for comparison, the percentage error before the addition of the offset is shown hatched once again in light gray; the dark gray bars show the slight deviations from the setpoint after offset addition. 25

However, a small error still exists while the optimizer claims the error is zero (Figure 23). This is due to the fact that models are used during the optimization, which always represent an approximation. Often, however, it is not possible to improve the model quality so that the reality is 100% correct. 30 For validation, i. the transfer to the remaining seven other operating points, two methods can be used: modeling and interpolation.

The model, which is placed on the boost pressure offset as a function of speed and injection quantity 35 (see Fig. 25), reflects very well the disadvantage of the core model, namely that in this test carrier, the boost pressure must be raised at medium speeds, thus the actual air mass can be displayed.

In order to obtain the most accurate information possible about what lies between the supporting points, 40 the model must be modeled with the highest possible quality. In this case, the method of neural networks is used and a very good coefficient of determination of R2 = 0.915 is obtained.

If one determines the corresponding offset for the 7 validation points from the model and 45 calculates the air mass with the help of BOOST, the results show that a transfer to these points is possible. Fig. 26 presents the percentage deviation before (in light gray hatched) and (in dark gray) after the optimization: the air mass error can be reduced from 6% to just over 1%. Thus, the second possibility of the map-wide transmission by interpolation has the advantage that the map determined from it maps exactly the interpolation points and is not falsified by a modeling. In Fig. 27 this map is shown for the boost pressure offset. It also shows the increased boost pressure in the range of average speeds and high loads. 55 The simulation with the resulting pressure additions show even better results: 21 AT 501 578 B1 the error is even increased from 6% to max. 0.8% reduced (Figure 28).

Basically, one can state that a very good map-wide improvement of the simulated air mass can be achieved with this method. 5

combustion

The optimization of the combustion means the adjustment of the output variables IMEP (indexed mean pressure) and PPF (maximum cylinder pressure) to the respective measured values. 10

If one considers the difference between the measured and the simulated quantity in the characteristic diagram (FIG. 29), no clear tendency is discernible. For this reason, the choice of the parameter for the calibration of the simulation model is not unique. 15 Since there are few details about the internal engine relationships in the early development phase, it makes more sense to use a simple combustion model. Here comes z. B. the vibe burning function in question. For the parameterization, the burning time, the start of burning and the Vibe form factor must be specified. 20 First of all, a full-factorial variation of all three parameters was performed, so that the indicated mean pressure and the maximum cylinder pressure could be adjusted at the same time using this objective function: 25 1000 IMEPBoos1 IMEPMess IMEP, + 100 PFPo "nct - PFP,

measuring

Boost ~ PFR

measuring

| TABGB00St -TABG

measuring

measuring

TABG - = min (16)

measuring

Here, the choice of factors, a special emphasis on IMEP, as this size should be reproduced as accurately as possible. The exhaust gas temperature is integrated into this target function so that, with two optima, that optimum is selected whose exhaust gas temperature is closer to the measured value.

However, it has been found that the optimization problem has too many degrees of freedom and therefore can not be solved unambiguously (no reproducible and robust optimum). Furthermore, it has not been possible to lay a suitable characteristic field over the parameters determined, since the values fluctuated too much and no clear tendency was discernible. Fig. 30 shows the maps of the burning time, the start of burning and the shape factor, which resulted after the optimization. 40 The scaling clearly shows that the individual parameters fluctuate greatly depending on the speed and load, resulting in very large gradients between the individual interpolation points. Thus, with this method, no map-wide improvement could be made possible. Therefore, it is necessary to split the optimization process into two parts by first adjusting the maximum cylinder pressure and then reducing the delta between the measured and simulated IMEP. If possible, only one variation parameter should be available for each output variable. For this reason, a part of the parameters must be set constant. 50

Since the start of injection from the ECU map is known, it makes sense to fix the start of combustion. The ignition delay, which describes the time between the start of injection and the actual start of combustion, is estimated to increase with increasing load and speed: 55 22 AT 501 578 B1 SOC = SOI + ZV = SOI + an + b mB + c · n · mB + d. (17)

In the same way one can model the Vibe form factor, as the combustion curve of a diesel becomes "softer" the higher the speed is (combustion center shifts to 5 "late"): mVibe = e-n + f. (18)

The factors a to d were approximated in this case on the basis of the measured data; e and f were determined so that the form factor at low speed gives the value 0.2 and at high speed the value 0.7.

By approximation, the three-dimensional problem could be reduced to only one free parameter: the burning time. About the burning time, there is no information in a 15 simple test bench measurement. Therefore it makes sense to vary the burning time fully factorially. As shown in FIG. 31, this parameter has a great influence on the maximum cylinder pressure, so that the target pressure can be adjusted in this way.

After determining the burning time and consequently adjusting the maximum cylinder pressure, the indicated mean pressure must be adapted to the measured value. This can be done by changing the heat transfer. Because the heat loss during combustion has an almost linear influence on the output IMEP (FIG. 32). The so-called Calibration Factor in BOOST scales the heat loss and thus adapts IMEP to the measured value. 25

The adjustment of the maximum cylinder pressure over the burning time gives very good results in the individual operating points. Fig. 33 shows the associated screenshot of the optimization results. There remains in some points a small difference left, which is due to the choice of the termination criterion of the optimization algorithm. However, this 30 minimum error is not relevant for the simulation in BOOST.

This is because the associated BOOST simulations show a reduction in the percentage error between measured and calculated PFP, which after adjustment is a maximum of 1.5% (see FIG. 34). 35

The results, which result from the adaptation of the indicated mean pressure, show similar results (Fig. 35, Fig. 36). The deviation of the simulated IMEP from the measured value can even be reduced below 0.2% in these training points. This represents a significant improvement in the mapping of this output. 40

All in all, it can be stated that the two parameters combustion duration and calibration factor make it possible to compare the maximum cylinder pressure and the indicated mean pressure combustion quantities with the respective measurement data, namely at the operating points with which the optimization is carried out. 45

For the transmission of the optimized parameters across the map, two new options are shown below.

On the one hand, a model is set by the parameters with the input variables speed and one-50 injection quantity. FIG. 37 shows the model of the burning time, generated by means of neural networks, in order to obtain the best possible model quality. Fig. 38 shows the model for the Calibration Factor, also calculated by the FNN method. In both cases, the model quality shows defects (R2 = 0.825 or R2 = 0.641). Especially with the Calibration Factor model there is a big difference between the model surface and the actual parameter values. Thus, there is no clear tendency to optimize the indicated mean pressure. For the burning time, however, results in a better model, but here, the support points can not be accurately represented. Nevertheless, one can determine the burning time for the 7 validation points with this model. The associated BOOST calculations show that an improvement in the maximum cylinder pressure is possible, even if only by a few percent (Figure 39).

With regard to the second combustion variable, IMEP, however, these results can not be transferred so well: If one simulates the indicated mean pressure in the validation points with the Calibration Factors determined from the model, only a few points show an improvement, but usually one Increase the error between measured and simulated IMEP (Fig. 40). If the second variant of the characteristic-field-wide transmission is carried out, namely the linear interpolation between the individual interpolation points, the following characteristic diagrams result for the burning duration (FIG. 41) and for the calibration factor (FIG. 42). Again, you can see very clearly the strong gradients between the individual nodes. This is especially evident with the Calibration Factor.

It can also be stated that the behavior of the burning time over speed and load 20 in this characteristic map looks different from the previously described model (FIG. 37): While the FNN model describes two peaks at constant load, the waves in the interpolation map are more constant Rotation speed. This suggests that the model quality of the FNN is not sufficiently good, as expected, and the &quot; reality &quot; can not picture. 25 In FIGS. 43 and 44, the percentage deviations for the maximum cylinder pressure and the indicated mean pressure in the validation points are shown, after the respective burn duration and the respective calibration factor have been determined. Contrary to expectations, the interpolation method, despite its exact mapping, leads to the parameters worse results. While there is an improvement in the maximum cylinder pressure in all 7 validation points for the first transmission variant, the output in this method degrades at some operating points (Figure 43). In absolute terms, however, the error range of 30% to max. Reduce 12%.

With respect to the indicated mean pressure, the interpolation also leads to degradations 35 at some operating points, as does the modeling (Figure 44).

These results indicate that it is possible to optimize individual operating points directly with respect to the combustion quantities. 40 The transfer to the entire map, however, is not possible because it can only result in an accidental improvement in the output. For although the modeling does not represent the correct relationships of the parameters, the transmission results in some cases in better results. This can only be a coincidence and therefore depends on the choice of validation points. 45

The fourth output variable was the exhaust gas temperature. Looking at the difference between measured and simulated temperature in Fig. 45, it is found that the exhaust gas is simulated too hot at all operating points. In the case of this test carrier, this can essentially be explained by the fact that in the simulation no calculation of the heat loss in the pipes of the exhaust gas tract was carried out. Another explanation would be e.g. a different position of the measuring point.

To compensate for this deviation, the wall heat flow in the exhaust system must be increased. This can be done via the heat transfer factor in the tubes of the BOOST model. It is important that the wall temperature is lower than the measured exhaust gas temperature, ie below the setpoint. Only under this condition, a cooling of the exhaust gas can be realized. In this case, the wall temperature was set to a constant 300K, so that the temperature difference between the exhaust gas and the wall remains large and thus the heat loss can be influenced as well as possible. 5

In Fig. 46, the ratio of the simulated to the measured exhaust gas temperature is plotted after a continuous increase in the heat transfer factor. Mann realizes that in this way it is possible to match the initial value to its measured value. The full-factorial variation of the heat transfer factor in BOOST and the subsequent minimization of the magnitude function (equation 15) lead to the results shown in FIG. Again, the target function is successfully optimized to zero at all operating points and a corresponding heat transfer factor is determined in each case. 15 The verification of these results by a simulation in BOOST provides the percentage deviations presented in Fig. 48, ie, an optimal improvement of the exhaust gas temperature in all 20 training points.

The validation, ie the map-wide transmission shows 20 satisfactory results with respect to the exhaust gas temperature.

The FNN model of the heat transfer factor (FIG. 49) has a sufficiently good quality (R2 = 0.840). 25 The corresponding BOOST simulations also show a correction of the error in the validation points of a maximum of 13% to almost 6% (Fig. 50). Furthermore, it can be stated that an improvement in the exhaust gas temperature occurs at all points.

An interpolation between the support points results in a similar relationship 30 between the optimal heat transfer factors, the speed and the injection quantity. The map thus calculated, shown in Fig. 51, shows an area with relatively smooth transitions between the individual parameters.

For this reason, as well as through the modeling transfer, one obtains an improvement in the output TABG at all 7 validation points (Figure 52). The error can even be reduced to just 3%. This indicates that the model quality is insufficient to properly transfer the optimized heat transfer factors.

As shown in this chapter, an improvement in the simulation of the exhaust gas temperature could be achieved via heat loss. However, this is only possible if the simulated temperature before adjustment in all points is too high. In other cases, e.g. If it is displayed too low over the entire map, there is the possibility to increase the wall temperature and thus to heat the exhaust gas. In the event that the starting point does not have a uniform image, one could individually set the wall temperature for each operating point: either above or below the setpoint value, depending on whether it should be heated or cooled.

Validation with one data set 50 In order to generalize the methodology presented in the last chapters, the 7 validation points were exchanged in a further step. In Fig. 53, the known measuring points are again plotted on the speed and injection quantity. As can be seen here, the new validation data sets (square) lie partly in the marginal areas of the operating map, so that a different starting point for the optimization results. 55

Claims (10)

25 AT 501 578 B1 If one now determines the parameters supercharging offset, burning time, calibration factor and heat transfer factor via the new training points, the results are similar to those described above: 5 Fig. 54 presents the percentage deviations of each output in the new validation points Here, the error between simulated and measured air mass and between io simulated and measured exhaust gas temperature is successfully minimized: The error can be reduced to a maximum of 2% or 3%. The graphs for the output variables PFP and IMEP show unexpectedly good results in this case. However, due to the difficulties encountered, it has to be assumed that this improvement happened by chance. The adjustment thus depends heavily on the selected points within the speed-load map 15. FIG. 55 shows the entire workflow for the calibration by calibration of the simulation models. 1. A method for investigating the behavior of complex technical systems, in particular of internal combustion engines, and for obtaining design parameters 25 by forming a model representing different measured variables as a function of input variables, with the following fundamental steps: selecting different measuring points which correspond to different constellations of measured quantities, and performing measurements to determine measured variables on a real system; 30 - creating a model that maps the dependence of the measures on the input variables and calibrating the model based on the measurements of the real system obtained at the points of measurement; characterized in that in detail the following steps are performed: - dividing the model into at least two submodels; 35 - creating at least a first submodel that maps a first subset of the metrics; - identifying at least one first main influence parameter for the first partial model; Determining an optimum value of the first main influence parameter in each 40 measurement point; Interpolating the first main influence parameter for all meaningful constellations of input variables for the calibration of the first partial model; - creating a further submodel for mapping a further subset of the measured variables as a function of the input variables and the previously determined first subset of the measured variables; Identifying at least one further main influencing parameter for the further submodel; Determining an optimum value of the further main influencing parameter in each measuring point; 50 - Interpolation of the other main influencing parameter for all meaningful constellations of input variables for the calibration of the further submodel. 2. The method according to claim 1, characterized in that a plurality of submodels are calibrated successively, wherein the measured variables determined in each submodel represent input variables for subsequent submodels. 3. Method according to claim 1 or 2, characterized in that exactly one measured variable is determined in a submodel. 4. The method according to any one of claims 1 to 3, characterized in that in a 5 measurement a single main influence parameter is assigned and that the Kalibrie tion of the associated submodel is done by full factorial variation. 5. The method according to any one of claims 1 to 4, characterized in that the interpolation of the main influence parameter is carried out as a linear interpolation. Method according to one of claims 1 to 5, characterized in that an internal combustion engine is examined and that the following quantities are used as measured variables: - incoming air mass, - indicated mean pressure, - maximum cylinder pressure and - Abgastemperaturvorkatalysator or turbine. A method according to any one of claims 1 to 6, characterized in that the optimum value of the main influencing parameters in the measuring points is determined by calculating for different discrete values of the main influence parameter, i. is calculated at predetermined interpolation points, a setpoint for the respective measured variable, from the values thus calculated, an interpolation model is created which indicates setpoint values for the respective measured variable for any values of the main influencing parameter, then minimizes a difference between these setpoint values and the measured actual value of the measured variable and from this the optimal value of the main influencing parameter is deduced. 8. The method according to any one of claims 1 to 7, characterized in that at least one further measured variable, which depends on a plurality of input variables, and of which a mathematical basic model is present, which maps the fundamental dependence of the further measured variable of the input variables, thereby obtained in that the following steps are carried out: selection of a multiplicity of first vectors, each of which represents a specific constellation of the input variables and which covers the useful working range of the system; - obtaining computational values of the measurand by using the base model to compute simulation values of the measurand associated with the first vectors; Selection of a multiplicity of second vectors, each representing a further constellation of the input variables; - performing measurements to obtain experimental values of the measure associated with the second vectors; Expanding each vector by one dimension by introducing a block variable which is set to a first value for the first vectors and to a second value for the second vectors; - creating a multivariate regression model representing the measurand as a polynomial function of the extended vectors of the input variables, based on the previously determined computational values of the measurand and the experimental values of the measurand; thus - determining at least one third vector representing a constellation of the input variables at which the system is to be examined; Expanding the third vector by one block variable set to the second value; - Calculate the measured variable with the regression model with the extended third vector as input variable. 27 AT 501 578 B1 9. The method according to claim 8, characterized in that the regression model is non-linear. 10. The method according to claim 9, characterized in that the regression model is quadratic. 11. The method according to claim 8 to 10, characterized in that the number of first vectors is greater than the number of second vectors. 12. The method according to any one of claims 8 to 11, characterized in that the system is an internal combustion engine and that are used as an input variable speed, load, and optionally other sizes. 15 plus 35 sheets drawings 20 25 30 35 40 45 50 55