METHOD FOR MODEL-BASED CONTROL AND REGULATION OF AN ENGINE

Description

[0001] The invention relates to a method for model-based control of an internal combustion engine, in which, depending on a torque, injection system setpoints for controlling the injection system actuators are calculated via a combustion model, and gas path setpoints for controlling the gas path actuators are calculated via a gas path model. Furthermore, in this method, an optimizer calculates a performance measure as a function of the injection system setpoints and the gas path setpoints. The optimizer minimizes the performance measure by changing the injection system setpoints and gas path setpoints within a prediction horizon. Based on the minimized performance measure, the optimizer sets the injection system setpoints and gas path setpoints as decisive for setting the operating point of the internal combustion engine.

[0002] The behavior of an internal combustion engine is largely determined by the engine control unit (ECU) in relation to a desired power output. For this purpose, corresponding characteristic curves and maps are applied to the ECU software. These are used to calculate the engine's control variables from the desired power output, for example, a target torque. These parameters include, for example, the injection timing and the required rail pressure. The engine manufacturer populates these characteristic curves / maps with data during a test bench run. However, the large number of these characteristic curves / maps and their interactions with each other result in a high degree of tuning effort.

[0003] WO 2017 / 005337 A1 relates to such a method, known per se, for operating an internal combustion engine with timescale separation, wherein at least one first-order control module calculates setpoint values ​​for at least one second-order control module, wherein the setpoint values ​​are used by the at least one second-order control module for controlling and / or regulating the internal combustion engine, wherein a separation of calculation steps is carried out such that the at least one first-order control module performs a static calculation of at least one first process of the internal combustion engine, and the at least one second-order control module performs a dynamic calculation of at least one second process of the internal combustion engine. Target values ​​for the at least one first-order control module can be calculated by a higher-level coordinator module.The higher-level coordinator module performs a model-based calculation, preferably optimization, of the internal combustion engine's operation, taking into account at least one optimization objective – preferably in real time. This makes it possible to define overarching target specifications for the engine's operation and implement them using the coordinator module. The coordinator module receives requirements, in particular specifications from the engine operator. Specifically, these requirements preferably include specifications for the engine's rotational speed and / or torque.

[0004] In practice, attempts are therefore made to reduce the coordination effort by using mathematical models. For example, DE 10 2006 004 516 B3 describes a Bayesian network with probability tables for determining an injection quantity, and DE 10 2016 205 241 A1 describes a method for operating an internal combustion engine by specifying an injection profile defined by adapted injection parameters.

[0005] In DE 10 2013 206 308 A1, it is only generally described that functional models are based on model parameters that are usually preset and are generally not changed after a motor vehicle has been put into service. However, some of the model parameters can be changed using online adaptation performed in the control unit (i.e., adaptation during operation) in order to account for changes in the behavior of the motor vehicle, the engine system, or other components over the vehicle's lifetime. Furthermore, online adaptation should also be used in cases where slight deviations in system behavior result from manufacturing tolerances of individual components or different combinations of components, which must be taken into account in the model parameters.To perform online adaptation, the control units typically incorporate appropriate logic that can adjust the relevant model parameters accordingly. The idea behind the described method for adapting model parameters of a functional model that simulates routes or system functions is to perform a relearning or adaptation / update of model parameters after the vehicle has been commissioned, such that these model parameters are provided to the vehicle externally.

[0006] German patent DE 10 2015 221 819 A1 only describes in general terms the adaptation of a basic model via a correction model for motor vehicles. This initially reveals that no satisfactory methods are currently known for adapting, i.e., modifying at runtime, data-based functional models implemented in a model calculation unit. The possibility of adapting a data-based basic model using an additive error model is known in principle. By continuously executing the adaptation procedure described therein during operation, a convergence of the correction model is to be achieved in order to create an adapted basic functional model that corresponds as closely as possible to the system behavior.

[0007] With regard to the aforementioned procedure, US 2011 / 0172897 A1 describes a method for adapting the injection timing and injection quantity via combustion models using neural networks. Since only trained data is used, this data must first be learned during a test bench run.

[0008] A general principle of a non-linear model-based predictive control of a diesel engine of the type mentioned above is described in essence in Karsten Harder et al.: "Nonlinear MPC with Emission Control for a Real-World Off-Highway Diesel Engine", 2017 IEEE International Conference on Advanced Intelligent Mechatronics (AIM), Munich, Germany, July 3-7, 2017 (2017-07-03), pages 1768-1773, XP033144574, DOI:10.1109 / AIM, 2017, 8014274, ISBN: 978-1-5090-5998-0 [accessed August 21, 2017]. There, in section III, a real-time MPC algorithm is described with respect to a [missing information - likely a specific feature or concept]. Fig. 1 The control loop is specified in principle.

[0009] In WO 2017 / 094349 A1, an internal combustion engine control unit is described which makes it possible to accurately calculate in real time the engine parameters required for controlling an internal combustion engine, in particular combustion parameters, based on a measurement result from a cylinder pressure sensor, and which thus makes it possible to improve the controllability of the internal combustion engine.A control device for an internal combustion engine comprises: a cylinder pressure sensor that detects the cylinder pressure in a cylinder; a plant model provided in an electronic control unit; a combustion model for calculating a heat release rate using the detected cylinder pressure and engine parameters including heat release rate, intake manifold pressure, EGR temperature and EGR pressure, which represent the state of the internal combustion engine, including the heat release rate; and an engine control unit provided in the electronic control unit that controls the internal combustion engine based on the engine parameters calculated by the plant model.

[0010] From the unpublished German patent application DE 10 2017 005 783.4, a model-based control method for an internal combustion engine is known, in which injection system setpoints for controlling the injection system actuators are calculated using a combustion model, and gas path setpoints for controlling the gas path actuators are calculated using a gas path model. An optimizer then modifies these setpoints with the aim of minimizing a performance metric within a prediction horizon. The minimized performance metric then defines the optimal operating point of the internal combustion engine. While the described method has proven effective with regard to significantly reduced tuning effort, it still offers potential for optimization.

[0011] From DE 10 2015 225 279 A1, a model-based control method is also known, in which a time-scale separation of the control modules and a model adaptation for adjusting to deviations of individual engines or engine aging are provided. For example, a fast process time is assigned for injection management (injection start, injection end), a medium process time for exhaust gas recirculation, and a longer process time for exhaust aftertreatment. Furthermore, instead of torques in a torque controller, a speed deviation in a speed controller can be minimized. For calculation purposes, the characteristic parameters of a global model of the internal combustion engine can be calculated in real time on an engine control unit (ECU) for the entire internal combustion engine within the framework of a nonlinear, model-based predictive control.

[0012] The invention is therefore based on the objective of further developing the previously described method with regard to improved quality.

[0013] This problem is solved by the features of claim 1. Embodiments of the method according to the invention are described in the dependent claims.

[0014] In the method according to the invention, injection system setpoints for controlling the injection system actuators are calculated using a combustion model based on a target torque, and gas path setpoints for controlling the gas path actuators are calculated using a gas path model, wherein the combustion model is adapted during the operation of the internal combustion engine. Furthermore, in this method, an optimizer calculates a quality measure as a function of the injection system and gas path setpoints and minimizes the quality measure by changing the injection system and gas path setpoints within a prediction horizon. Once a minimized quality measure has been determined, the optimizer ultimately uses the injection system and gas path setpoints as the decisive factors for setting the operating point of the internal combustion engine.

[0015] According to the invention, the combustion model is adapted via a first Gaussian process model to represent a basic grid and via a second Gaussian process model to represent adaptation data points.

[0016] In particular, the data for the first Gaussian process model are acquired on a single-cylinder test bench. Extreme operating conditions can be simulated on this test bench, such as a very cold environment or a very high altitude. Subsequent physical modeling is used to vary all input variables in order to cover the entire operating range of the internal combustion engine. Due to system limitations, the data values ​​in the first Gaussian process model only roughly represent the engine system in its normal operating range. However, the advantage of this model is that it provides a basic grid with few data points but physically meaningful extrapolation behavior. The data for the second Gaussian process model are generated from a Design of Experiments (DoE) test bench run of the internal combustion engine with a stationary driving range.Due to system limitations, the data values ​​in the second Gaussian process model are therefore only valid for this steady-state range, albeit with high accuracy. The combination of the first and second Gaussian process models thus includes operating ranges with both real, measured, and extrapolated data values.

[0017] According to the invention, the quality of the second Gaussian process model is assessed using a confidence interval.

[0018] A narrow confidence interval indicates high accuracy, while a wide confidence interval indicates lower accuracy. During operation, the position of a current adaptation data point is assessed in relation to the valid confidence interval. Preferably, the confidence interval corresponds to two standard deviations, i.e., a 95% confidence interval.

[0019] If the current adaptation data point lies within the confidence range, it is added to the second Gaussian process model according to the invention. If the current adaptation data point lies outside the valid confidence range, the second Gaussian process model is modified according to a further development by removing adaptation data points from the second Gaussian process model until the current adaptation data point lies within the new confidence range.

[0020] To reduce storage requirements and computation time, the total number of adaptation data points is compared to a threshold value. If the threshold is exceeded, adaptation data points are removed so that the new total number is less than the threshold. Those adaptation data points that have little or no impact on the accuracy of the second Gaussian process model are removed.

[0021] To further reduce computation time, the first Gaussian process model is back-fitted to a second Gaussian process model to represent a basic grid. This back-fit is based on the requirement that the second Gaussian process is equal to zero at the support points of the first Gaussian process. During the back-fit, each data point of the first Gaussian process model is assigned a timestamp. A time rank can be determined based on the temporal change of the timestamp. From this time rank, a period for the continued operation of the internal combustion engine can be estimated. In other words, a defective NOx sensor, for example, causes a temporal drift of the mean value in the first Gaussian process model. The corresponding time rank then defines the remaining period for the model-based continued operation of the internal combustion engine.The time sequence can of course also be used to detect any unauthorized manipulation of the internal combustion engine.

[0022] The invention offers the well-known advantages of adaptation, namely the standardization of internal combustion engines of the same series. In other words, adaptation automatically reduces manufacturing variations. By recalibrating the first Gaussian process model with the second Gaussian process model, a self-learning system with fault detection is implemented. Since the models can be individually tuned and together represent the internal combustion engine, the tuning effort can be further reduced. The previously required characteristic curves and maps are no longer necessary. The models' extrapolation capability allows for the calculation of reliable engine control variables in dynamic, transient operation as well as in rarely used operating ranges. Furthermore, the deviations between the control system's target values ​​and the legal emission limits can be reduced.

[0023] The figures in the drawing illustrate a preferred embodiment. They show: Fig. 1 a system diagram, Fig. 2 a model-based system diagram, Fig. 3 a block diagram, Fig. 4 a program flowchart, Fig. 5 a subroutine, Fig. 6 a first adaptation example, Fig. 7 a second adaptation example and Fig. 8 a third adaptation example.

[0024] The Fig. 1 Figure 1 shows a system diagram of an electronically controlled internal combustion engine 1 with a common rail system. The common rail system comprises the following mechanical components: a low-pressure pump 3 for supplying fuel from a fuel tank 2, a variable intake throttle 4 for influencing the fuel flow rate, a high-pressure pump 5 for supplying the fuel under increased pressure, a rail 6 for storing the fuel, and injectors 7 for injecting the fuel into the combustion chambers of the internal combustion engine 1. Optionally, the common rail system can also be designed with individual accumulators, in which case, for example, an individual accumulator 8 is integrated into the injector 7 as an additional buffer volume. The further functionality of the common rail system is assumed to be known.

[0025] The depicted gas path comprises both the air supply and the exhaust gas discharge. The air supply includes the compressor of an exhaust gas turbocharger 11, an intercooler 12, a throttle valve 13, an inlet 14 for combining the charge air with the recirculated exhaust gas, and the inlet valve 15. The exhaust discharge includes an outlet valve 16, the turbine of the exhaust gas turbocharger 11, and a turbine bypass valve 19. An exhaust gas recirculation path branches off from the exhaust discharge, containing an EGR actuator 17 for adjusting the EGR rate and the EGR cooler 18.

[0026] The operating mode of the internal combustion engine 1 is determined by an electronic control unit 10 (ECU). The electronic control unit 10 contains the usual components of a microcomputer system, such as a microprocessor, I / O modules, buffers, and memory modules (EEPROM, RAM). The operating data relevant for the operation of the internal combustion engine 1 are applied as models in the memory modules. The electronic control unit 10 uses these models to calculate the output variables from the input variables. The key input variable is a target torque M(SOLL), which is specified by an operator as a desired power output. The input variables of the control unit 10 related to the common rail system are the rail pressure pCR, which is measured by a rail pressure sensor 9, and optionally the individual accumulator pressure pES.The input variables of the electronic control unit 10 related to the air path are the opening angle W1 of the throttle valve 13, the engine speed nIST, the charge air pressure pLL, the charge air temperature TLL, and the charge air humidity phi. The input variables of the electronic control unit 10 related to the exhaust path are the opening angle W2 of the EGR actuator 17, the exhaust gas temperature TAbgas, the air-fuel ratio Lambda, and the NOx actual value downstream of the turbine of the exhaust gas turbocharger 11. The other input variables of the electronic control unit 10 not shown are summarized with reference symbol EIN, for example, the coolant temperatures.

[0027] In Fig. 1 The following are shown as output variables of the electronic control unit 10: a PWM signal for controlling the intake throttle 4, a ve signal for controlling the injector 7 (injection start / end), a DK control signal for controlling the throttle valve 13, an EGR control signal for controlling the EGR actuator 17, a TBP control signal for controlling the turbine bypass valve 19, and an OFF output variable. The OFF output variable represents the other control signals for controlling and regulating the internal combustion engine 1, for example, a control signal for activating a second exhaust gas turbocharger in the case of sequential turbocharging or a variable valve train. Fig. 2 This shows a model-based system diagram. In this representation, the input variables of the electronic control unit 10 are a first library Biblio1, a second library Biblio2, measured variables MESS, and the collective reference symbol EIN, which represents the variables in the Fig. 1 The input variables shown are as follows. The first library, Biblio 1, specifies the operation of the internal combustion engine according to the IMO's MARPOL (Marine Pollution) emission class or the EU IV / Tier 4 final emission class. The second library, Biblio 2, specifies the internal combustion engine type and a maximum mechanical component load, for example, the peak combustion pressure or the maximum speed of the exhaust gas turbocharger. The input variable MESS specifies both directly measured physical quantities and auxiliary quantities calculated from them. The output variables of the electronic control unit are the setpoints for the subordinate control loops, the injection start (SB), and the injection end (SE). Within the electronic control unit are a combustion model (20), an adaptation (21), a gas path model (22), and an optimizer (23).

[0028] Both the combustion model 20 and the gas path model 22 represent the system behavior of the internal combustion engine as mathematical equations. The combustion model 20 statically models the combustion processes. In contrast, the gas path model 22 models the dynamic behavior of the air intake and exhaust gas flow. The combustion model 20 includes individual models, for example, for NOx and soot formation, exhaust gas temperature, exhaust gas mass flow, and peak pressure. These individual models, in turn, depend on the boundary conditions in the cylinder and the injection parameters. The combustion model 20 is determined for a reference internal combustion engine in a test bench run, the so-called DoE test bench run (DoE: Design of Experiments).During the DoE test bench run, operating parameters and control variables are systematically varied with the aim of modeling the overall behavior of the internal combustion engine as a function of engine parameters and environmental conditions. The combustion model 20 is supplemented by adaptation 21. The aim of adaptation is to reduce the production variation of an internal combustion engine.

[0029] After activation of the internal combustion engine 1, the optimizer 23 first reads the emission class from the first library, Biblio 1, and the maximum mechanical component loads from the second library, Biblio 2. The optimizer 23 then evaluates the combustion model 20 with regard to the target torque M(SOLL), the emission limits, the environmental boundary conditions (e.g., the humidity phi of the charge air), the operating situation of the internal combustion engine, and the adaptation data points. The operating situation is defined in particular by the engine speed nIST, the charge air temperature TLL, and the charge air pressure pLL. The function of the optimizer 23 is then to evaluate the injection system target values ​​for controlling the injection system actuators and the gas path target values ​​for controlling the gas path actuators. In doing so, the optimizer 23 selects the solution that minimizes a specific efficiency measure.The quality measure is calculated as the integral of the squared target-actual deviations within the prediction horizon. For example, in the form: . J = ∫ w 1 NOx SOLL − NOx IST 2 + w 2 M SOLL − M IST 2 + w 3 … . + …

[0030] Here, w1, w2, and w3 represent corresponding weighting factors. As is known, nitrogen oxide emissions result from the charge air humidity phi, the charge air temperature, the injection timing SB, and the rail pressure pCR. Adaptation 21 intervenes in the actual values, for example, the NOx value or the exhaust gas temperature.

[0031] The efficiency measure is minimized by the optimizer 23 calculating an initial efficiency measure at a first time point, varying the injection system target values ​​and the gas path target values, and using these variations to predict a second efficiency measure within the prediction horizon. Based on the deviation between the two efficiency measures, the optimizer 23 then defines a minimum efficiency measure and sets this as the decisive factor for the internal combustion engine.

[0032] For the example shown in the drawing, these are the target rail pressure pCR(SL), the injection start SB, and the injection end SE for the injection system. The target rail pressure pCR(SL) is the reference variable for the subordinate rail pressure control loop 24. The manipulated variable of the rail pressure control loop 24 corresponds to the PWM signal for actuating the intake throttle. The injection start SB and injection end SE are used to control the injector ( Fig. 1 : 7 ) directly charged.

[0033] For the gas path, the optimizer 23 indirectly determines the gas path setpoints. In the example shown, these are a lambda setpoint LAM(SL) and an EGR setpoint AGR(SL) for the subordinate lambda control loop 25 and the subordinate EGR control loop 26. The manipulated variables of the two control loops 25 and 26 correspond to the signal TBP for controlling the turbine bypass, the signal AGR for controlling the EGR actuator, and the signal DK for controlling the throttle valve. The feedback measured variables MESS are read by the electronic control unit 10. The measured variables MESS include both directly measured physical quantities and auxiliary quantities calculated from them. In the example shown, the lambda actual value LAM(IST) and the EGR actual value AGR(IST) are read.

[0034] The Fig. 3 Figure 1 shows the interaction of the two Gaussian process models for adapting the combustion model in a block diagram. Gaussian process models are known to those skilled in the art, for example from DE 10 2014 225 039 A1 or DE 10 2013 220 432 A1. In general, a Gaussian process is defined by a mean value function and a covariance function. The mean value function is often assumed to be zero or a linear / polynomial curve is introduced. The covariance function describes the relationship between arbitrary points. A first function block 27 contains the DoE data (Design of Experiments) of the complete engine. This data is determined for a reference internal combustion engine during a test bench run by recording all variations of the input variables across their entire operating range in the steady-state range of the engine. This data characterizes the behavior of the internal combustion engine in the steady-state range with high accuracy.A second functional block, 28, contains data acquired on a single-cylinder test bench. This single-cylinder test bench allows for the setting of operating ranges, such as high geodetic altitude or extreme temperatures, that cannot be tested during a DoE (Design of Experiments) test bench run. These limited measurement data serve as the basis for parameterizing a physical model that roughly accurately represents the global combustion behavior. The physical model roughly depicts the behavior of the internal combustion engine under extreme boundary conditions. Extrapolation is used to complete the physical model, so that a normal operating range is roughly accurately described. In the... Fig. 3 The extrapolation-capable model is marked with the reference symbol 29. From this, the first Gaussian process model 30 (GP1) is generated to represent a basic grid.

[0035] The merging of the two sets of data points forms the second Gaussian process model 31. This means that operating ranges of the internal combustion engine, which are described by the DoE data, are also defined by these values, and operating ranges for which no DoE data is available are represented by data from the physical model. Since the second Gaussian process model is adapted during operation, it serves to represent the adaptation points. In general, the following applies to the model value (reference symbol 32): E x = GP 1 + GP 2

[0036] Here, GP1 corresponds to the first Gaussian process model for representing the basic grid, GP2 to the second Gaussian process model for representing the adaptation data points, and the model value E[x] is the input variable for the optimizer, for example, an actual NOx value or an actual exhaust gas temperature value. The double arrow in the Fig. 3 Two information pathways are shown. The first information pathway represents the provision of the basic grid data from the first Gaussian process model 30 to the model value 32. The second information pathway represents the re-fitting of the first Gaussian process model 30 via the second Gaussian process model 31.

[0037] In the Fig. 4 The program flowchart depicts a main program through which the optimizer optimizes the performance measure J within a prediction horizon. This main program includes a subprogram, UP Adaptation, which provides adapted values ​​to the optimizer. The UP Adaptation subprogram has a longer execution time than the main program; that is, new adapted values ​​are not provided with every iteration of the main program. After initialization at S1, S2 checks whether the start process has finished. If it is still running (query result S2: no), the program branches back to point A. If the start process is finished, S3 reads the operator-specified target torque M(SOLL) and the NOx target value NOx(SOLL). Subsequently, S4 records the operating status of the internal combustion engine.The operating situation is defined by the measured variables, in particular the engine speed nIST, the charge air temperature TLL, the charge air pressure pLL, and the charge air humidity phi. The operating situation is further processed in two ways: firstly, in the Optimizer subprogram, step S5, and secondly, in the Adaptation subprogram UP. This subprogram is used in conjunction with the... Fig. 5 explained.

[0038] After calling the UP Optimizer subroutine, the initial values, for example, the injection start time SB, are generated at S6. A first quality measure J1 is calculated at S7 using equation (1), and at S8, a run variable i is set to zero. Then, at S9, the initial values ​​are changed and calculated as new setpoints for the manipulated variables. At S10, the run variable i is incremented by one. Based on the new setpoints, a second quality measure J2 is then predicted at S11 for the prediction horizon, for example, for the next 8 seconds. At S12, the second quality measure J2 is subtracted from the first quality measure J1 and compared with a limit value GW. The further progress of the quality measure is checked by calculating the difference between the two quality measures. Alternatively, the number of times an optimization has already been performed is checked by comparing the run variable i with a limit value iGW. These two limit value checks thus serve as a termination criterion for further optimization.If further optimization is possible (query result S12: no), the program branches back to point C. Otherwise, at S13, the optimizer sets the second quality measure J2 as the minimum quality measure J(min). The injection system setpoints and the gas path setpoints for the corresponding actuators are then derived from the minimum quality measure J(min). Subsequently, at S14, it is checked whether an engine stop has been initiated. If this is not the case (query result S14: no), the program branches back to point B. Otherwise, the program flow is terminated. A detailed description of the optimizer's operating principle, including prediction, is known from the unpublished patent application with the official file number DE 10 2017 005 783.4, to which reference is hereby made.

[0039] In the Fig. 5 The subroutine UP Adaptation is shown. At S1, it is checked whether the current data point lies within the valid confidence interval KB. If it lies outside the valid confidence interval KB (query result S1: no), the program branches to S2 and removes any previously saved adaptation data point. Then, it branches back to point A and checks again at S1 whether the current adaptation data point now lies within the new confidence interval. This case is in the Fig. 6 presented and is used in conjunction with the Fig. 6 As explained, in loops S1 and S2, adaptation data points are removed from the second Gaussian process model until the current adaptation data point lies within the new confidence interval. If S1 determines that the current data point lies within the confidence interval KB (query result S1: yes), then S3 adds the current adaptation data point to the second Gaussian process model. Subsequently, S4 checks whether the total number n of adaptation data points exceeds a threshold value GW. If not (query result S4: no), the program continues at S6. Otherwise, S5 removes the adaptation data point that least influences the mean. The program then returns to point B and checks the total number n again at S4.Loop S4 / S5 therefore removes as many adaptation data points from the second Gaussian process model as necessary until the total number n is below the threshold GW. This results in reduced memory usage and faster processing time.

[0040] At S6, it is checked whether the first Gaussian process model needs to be adjusted to represent the basic grid. If this is not necessary (query result S6: no), the program execution continues at point C. If an adjustment is necessary (query result S6: yes), the first Gaussian process model is adjusted by adjusting the expected value of the first Gaussian process model back to the second Gaussian process model. The program execution then continues at point C. At S8, a time rank (ZR) is checked for exceeding a limit value. Each data point in the first Gaussian process model is assigned a timestamp. A change in the data point, i.e., a time drift, changes the time rank. If S8 determines that the time rank (ZR) is greater than the limit value (GW) (query result S8: yes), a warning message and the remaining useful life are displayed at S9, and the program execution continues at S10.If, however, S8 detects that the time rank ZR is less than the limit value GW (query result S8: no), the program flow continues at points D and S10. A sensor failure, for example of the NOx sensor, can be detected by querying the time rank. It can also detect unauthorized manipulation of the internal combustion engine. Based on the time rank, an estimate is made of how long model-based continued operation of the internal combustion engine is still possible despite the sensor defect. At S10, it is checked whether the adapted values ​​should be used in the main program. If the check is successful (query result S10: yes), the program returns to the main program. Fig. 4 with the results being passed to the main program. If the check fails (query result S10: no), the program returns to the main program. Fig. 4 without the results being passed on to the main program.

[0041] In the Fig. 6 This illustrates the case where the current adaptation data point is not within the valid confidence interval. The valid confidence interval is defined by the mean (expected value My) and the covariance (sigma 2< ). Fig. 6 includes Fig. 6A bis Fig. 6D For clarity, these are represented two-dimensionally. On the abscissa is a quantity X, representing the input variables of the model, such as the injection timing SB, the rail pressure pCR, the charge air pressure pLL, or the charge air humidity phi. On the ordinate is a quantity Y, representing adaptable model values, such as NOx or the exhaust gas temperature. Non-adaptable quantities include, for example, soot, torque, or fuel consumption, which are also represented by equation (2). In practice, the quantities X and Y are therefore multidimensional. Fig. 6A Shown are a first adaptation data point A (2 / 1), a second adaptation data point B (3 / 1), and a current adaptation data point C (2.5 / 0). The current adaptation data point C is not within the valid confidence interval KB, which is defined in the Fig. 6 is shown in hatched areas. Then, the valid confidence interval KB ( is checked. Fig. 6A ) by removing the first adaptation data point A (2 / 1). From the Fig. 6B It becomes apparent that despite removing the first adaptation data point A, the current adaptation data point C would still lie outside the new confidence interval KB 1. Therefore, the first adaptation data point A is not removed, but rather the second adaptation data point B (3 / 1). As in the Fig. 6C As shown, the current adaptation data value C now lies within the new confidence interval KB2. Therefore, the second Gaussian process model is adjusted such that the current adaptation data point C (2.5 / 0) is adopted, while the previously stored first adaptation data point A (2 / 1) remains. Due to the recalculation, a new confidence interval KB results, as shown in the Fig. 6D depicted.

[0042] In the Fig. 7 The case is shown where the current adaptation data point lies within the current confidence interval KB. Fig. 7 includes Fig. 7A und Fig. 7B The quantity X plotted on the abscissa and the quantity Y plotted on the ordinate correspond to those in the Fig. 6 described sizes. In the Fig. 7A The diagram shows a first adaptation data point A (2 / 1), a second adaptation data point B (3 / 1), and a current adaptation data point C (4 / 1). Since the current adaptation data point C lies within the valid confidence interval KB, it is incorporated into the second Gaussian process model, and the new confidence interval KB is then calculated. Due to this recalculation, the confidence interval between adaptation data points A to C is significantly narrower. See [reference to relevant section]. Fig. 7B A narrower confidence range indicates improved quality.

[0043] In the Fig. 8 The case of a refit of the first Gaussian process model via the second Gaussian process model is illustrated. Fig. 8 includes Fig. 8A und Fig. 8B . The Fig. 8A Figure D shows in detail that the first Gaussian process model (solid line) differs from the mean of the second Gaussian process model (dashed line). The re-adjustment is achieved by adjusting the expected value of the first Gaussian process model so that it corresponds to the adaptation data points of the second Gaussian process model, see [reference]. Fig. 8B . REFERENCE MARK LIST

[0044] 1 Internal combustion engine 2 Fuel tank 3 Low-pressure pump 4 Intake throttle 5 High-pressure pump 6 Rail 7 Injector 8 Individual accumulator 9 Rail pressure sensor 10 Electronic control unit 11 Exhaust gas turbocharger 12 Intercooler 13 Throttle valve 14 Inlet point 15 Intake valve 16 Exhaust valve 17 EGR actuator (EGR: Exhaust gas recirculation) 18 EGR cooler 19 Turbine bypass valve 20 Combustion model 21 Adaptation 22 Gas path model 23 Optimizer 24 Rail pressure control loop 25 Lambda control loop 26 EGR control loop 27 First functional block (DoE data) 28 Second functional block (Single cylinder data) 29 Model 30 First Gaussian process model (GP1) 31 Second Gaussian process model (GP2) 32 Model value

Claims

1. Method for model-based control and regulation of a combustion engine (1), in which, depending on a target torque (M(TARGET)), injection system target values for activating the injection system actuators are calculated via a combustion model (20) and gas path target values for activating the gas path actuators are calculated via a gas path model (22), in which the combustion model (20) is adapted during the running operation of the combustion engine (1), in which a quality measure (J) is calculated by an optimizer (23) depending on the injection system target values and the gas path target values, the quality measure (J) is minimised by the optimizer (23) by changing the injection system target values and gas path target values within a prediction horizon, and in which the injection system target values and gas path target values are set by the optimizer (23) on the basis of the minimised quality measure (J(min)) as being decisive for setting the operating point of the combustion engine (1), wherein the combustion model (20) is adapted via a first Gaussian process model (30) to represent a base grid and via a second Gaussian process model (31) to represent adaptation data points, and a quality of the second Gaussian process model (31) is assessed on the basis of a confidence range (KB), wherein a current adaptation data point, which lies within the current confidence range (KB), is adopted into the second Gaussian process model (31).

2. Method according to claim 1, characterised in that, when a current adaptation data point from the second Gaussian process model (31) lies outside the confidence range (KB), previously stored adaptation data points are iteratively removed until the current adaptation data point lies within the new confidence range.

3. Method according to claim 2, characterised in that a total number (n) of the adaptation data points is compared with a limit value (GW), and if the limit value is exceeded (n > GW), a number of adaptation data points are removed such that the new total number is less than the limit value (GW).

4. Method according to claim 3, characterised in that the first Gaussian process model (30) is readjusted via the second Gaussian process model (31) to represent a base grid.

5. Method according to claim 4, characterised in that a time stamp is embossed in each data point of the first Gaussian process model (30), a time rank (ZR) is determined on the basis of the change in the time stamp and further operation of the combustion engine is estimated depending on the time rank (ZR).