Method for generating an injection quantity correction model for a gasoline engine, use of the injection quantity correction model, control unit and vehicle

Abstract

Method for generating an injection quantity correction model (16) for a gasoline engine, the method comprising the steps: - Providing an injection model (10) which demonstrates the effect of an injection correction quantity (Δ inj ) of fuel into the combustion chamber (12) of a spark-ignition engine maps to a combustion air ratio (λ1) over time, wherein the injection model (10) contains the correlation after what time from the injection time and how a changed injection correction quantity (Δ inj ) in the exhaust gas mixture leaving the combustion chamber (12), - Generating a machine learning algorithm trained on a training dataset, wherein the algorithm constructs an injection quantity correction model (16) using the injection model (10), which corrects the injection quantity (K) inj) of fuel into the combustion chamber (12) of the Otto engine as a function of at least one state variable of rotational speed (n) used as an input variable eng ), intake manifold pressure (p in ), Intake camshaft phase ((φ enw ) and exhaust camshaft phase (φ anw ) estimates, and the training dataset includes a variety of state parameters with respect to at least one input variable and associated combustion air ratios (λ2).

Description

[0001] The invention relates to a method for generating an injection quantity correction model for a gasoline engine using a machine learning algorithm and to the use of the injection quantity correction model. Furthermore, a corresponding control unit and a corresponding vehicle are provided.

[0002] To minimize the emission of pollutants such as nitrogen oxides, hydrocarbons, and soot, modern vehicles powered by gasoline engines are equipped with a lambda sensor (λ-sensor). The lambda sensor is the primary sensor in the lambda control loop for catalytic exhaust gas purification. It compares the residual oxygen content in the exhaust gas with the oxygen content of a reference, usually the ambient air, to determine the air-fuel ratio λ (ratio of combustion air to fuel). Based on this determined air-fuel ratio, the fuel injection quantity is regulated to achieve a balanced, i.e., stoichiometric (λ = 1), air-fuel ratio, as any deviation from the stoichiometric air-fuel ratio generally increases fuel consumption and the amount of harmful emissions.

[0003] The implementation of a lambda controller, which corrects the fuel injection based on the determined air-fuel ratio (measured value by the lambda sensor), is typically a component of a gasoline engine's operating strategy. However, by design, the controllers used (P, PI, or PID) can only react to mixture deviations that have already occurred (air-fuel ratio λ ≠ 1). Consequently, the occurrence of mixture deviations is not primarily prevented in the known operating strategy, but rather reduced to the existing mixture deviation over time. In steady-state operation of the gasoline engine, for example, at constant speed and constant intake pressure, the lambda controller can effectively compensate for any mixture deviation that has occurred.In dynamic operating conditions, however, that is, with changing engine speeds and intake pressures, constantly changing mixture deviations occur which cannot be reduced by the lambda controller of the known type, or not to a sufficient extent.

[0004] The publication DE 195 47 496 A1 discloses a method or device for the exact determination of the mass of air drawn into the cylinders of an internal combustion engine as a basis for the measurement of the fuel mass by means of a learning observer.

[0005] The publication DE 10 2020 116 488 B3 relates to a method for operating an internal combustion engine, in particular a method for regulating an air-fuel ratio, using a neural network and a corresponding control unit.

[0006] Document DE 197 06 750 A1 discloses a method for mixture control in an internal combustion engine and a device for carrying it out.

[0007] Publication DE 10 2009 032 064 B3 concerns an internal combustion engine (Otto engine) with controlled or regulated fuel injection.

[0008] Publication DE 10 2016 205 241 A1 discloses a method and a device for operating an internal combustion engine.

[0009] Document DE 696 35 429 T2 concerns internal combustion engines, in particular internal combustion engines in which fuel can be operated according to a predetermined injection profile with one or more pre-injections and one or more main injections.

[0010] The invention is therefore based on the objective of developing an operating strategy for a gasoline engine which prevents or at least further reduces the occurrence of mixture deviations from the stoichiometric air-fuel ratio in gasoline engines.

[0011] The problem according to the invention is solved by a method for generating an injection quantity correction model for a spark-ignition engine using a machine learning algorithm, a method for using the injection quantity correction model, a control unit, and a vehicle according to the independent claims. Preferred embodiments are the subject of the respective dependent claims.

[0012] A first aspect concerns a method for generating an injection quantity correction model for a gasoline engine, in particular a gasoline engine of a motor vehicle.

[0013] In one process step, an injection model is provided that depicts the effect of an injected quantity, more precisely an injection correction quantity, of fuel into the combustion chamber of a gasoline engine on the air-fuel ratio (determined, for example, by a lambda sensor) over time, i.e., the exhaust gas mixture leaving the combustion chamber. In other words, the injection model represents the correlation between the time after injection and how an increased or decreased injection quantity is reflected in the lambda signal at the lambda sensor. Consequently, using the injection model, it can be determined how and at what time the injection quantity must be changed, or should have been changed, to prevent or at least reduce a (future) mixture deviation from the stoichiometric air-fuel ratio.

[0014] In a further process step, a machine learning algorithm trained on a training dataset is generated. Using the injection model, the algorithm builds an injection quantity correction model that estimates a correction to the amount of fuel injected into the combustion chamber of the gasoline engine as a function of at least one state variable used as an input: engine speed, intake manifold pressure, intake camshaft phase, and exhaust camshaft phase. The correction is preferably specified with respect to an injection quantity calculated from the steady-state operation of the gasoline engine using a filling model. In this case, the estimated correction of the injection quantity correction model represents a correction with respect to the calculated injection quantity in steady-state operation.

[0015] The training dataset comprises a multitude of state parameters relating to at least one input variable and associated air-fuel ratios (over time). In other words, the algorithm learns, based on the prepared training dataset, the relationship between a change in the state variables in the air path of the spark-ignition engine and the resulting deviation from the stoichiometric air-fuel ratio at the lambda sensor. The resulting injection quantity correction model thus uses a learned relationship between the behavior of the state variables in the air path of the spark-ignition engine and a correction of the injection quantity, which ideally leads to a stoichiometric air-fuel ratio (λ ≈ 1). Consequently, the injection quantity can be adjusted to achieve a near-stoichiometric air-fuel ratio using the injection quantity correction.In other words, the deviations that occur are understood as the result of a disturbance influencing the system, which must be corrected, similar to an active noise-canceling algorithm. Consequently, the injection quantity correction model designed according to the invention enables the occurrence of a deviation from the stoichiometric air-fuel ratio to be prevented or at least reduced directly, i.e., proactively—and not reactively or with a delay, as with known lambda controllers—using the injection model. Thus, the injection quantities can be effectively corrected not only in steady-state operating conditions of the gasoline engine, but also in dynamic operating conditions, in order to further reduce harmful emissions and excessive fuel consumption.

[0016] It is understandable that any combination of the four state variables mentioned, for example, a multitude of them or all of them, can be included in the training data as input variables. Accordingly, the relevant input variables are then taken into account by the machine learning algorithm when building the injection quantity correction model. The more state variables considered when building the injection quantity correction model, the better the prediction quality and thus the accuracy of the injection quantity correction model. With increasing accuracy of the injection quantity correction model, the occurrence of deviations from the stoichiometric air-fuel ratio is reduced even further.

[0017] In a preferred embodiment, at least one of the state variables is the intake manifold pressure. Although the state variables engine speed, intake manifold pressure, intake camshaft phase, and exhaust camshaft phase correlate with each other in gasoline engine operation, it has been found that a change in the intake manifold pressure is a strong indicator of a resulting deviation from the stoichiometric air-fuel ratio. Therefore, using the intake manifold pressure as a state variable can lead to improved accuracy of the injection quantity correction model.

[0018] In a further preferred embodiment, the training dataset is extended by at least a plurality of associated state parameters of an additional state parameter of an air path of the spark-ignition engine, which differs from the at least one state parameter, and the injection quantity correction model further estimates the correction of the injection quantity as a function of this additional state parameter. In other words, more than one state parameter of the spark-ignition engine's air path is used to train the algorithm, thus further improving the accuracy of the resulting injection quantity correction model.

[0019] Preferably, a further state variable is defined as engine speed, intake manifold pressure, intake camshaft phase, exhaust camshaft phase, aging state, and temperature. The aging state is preferably represented by the coking state of one or more valves and / or injectors and / or the degree of contamination of one or more air filters. Temperature state variables are preferably cylinder wall temperature, exhaust gas temperature, oil temperature, and / or intake air temperature. The state variables engine speed, intake manifold pressure, intake camshaft phase, exhaust camshaft phase, exhaust gas temperature, oil temperature, and / or intake air temperature are preferably measured directly by a suitable sensor. The state variables aging state and / or cylinder wall temperature are preferably determined using a model-based approach, i.e., indirectly using known models.The use of further or additional state variables, especially state variables of the air path of the Otto engine, improves the accuracy of the injection model even further.

[0020] In a further preferred embodiment, the injection model is provided using an FIR filter and / or another machine learning algorithm. An FIR filter is a finite impulse response (FIR) filter, which is well-suited for stable digital signal processing. With the FIR filter, the injection model can be generated relatively easily using data on (varying) injection quantities and associated air-fuel ratios. The FIR filter is preferably trained for different operating points with respect to the state variables of the air path of the spark-ignition engine in order to better represent the different engine behavior. For example, the data used by the FIR filter preferably also includes associated engine speeds and / or intake manifold pressures. Here, the FIR filter is preferably trained for different operating points with respect to engine speed and intake manifold pressure.It has been found that the behavior of the injection quantity is influenced by the air-fuel ratio under varying engine speeds and / or intake manifold pressures, so that the accuracy of the injection model can be further improved by taking the state variables of engine speed and / or intake manifold pressure into account. Building an injection model with the additional machine learning algorithm is comparatively complex, but an even higher level of accuracy can be achieved.

[0021] Preferably, the further algorithm is trained with another training dataset to build the injection model. The prepared training set includes a multitude of state parameters relating to a corrective intervention in the injection quantity and associated combustion air conditions.

[0022] It is also preferred that the training dataset be extended to include associated state parameters relating to the engine speed and / or intake manifold pressure. By taking the engine speed and / or intake manifold pressure into account, the accuracy of the resulting injection model can be further improved.

[0023] In a further preferred embodiment, the algorithm and / or the further algorithm comprises an artificial neural network. A single- or multi-layered artificial neural network is preferably considered as the machine learning algorithm, although the invention is not limited to this. Rather, other relevant machine learning algorithms can also be used. Examples include, among others, support vector machines and dynamic neural networks (LSTMS).

[0024] In a further preferred embodiment, the training data set and / or the additional training data set are generated by the vehicle's control unit. Furthermore, the algorithm and / or the additional algorithm are preferably trained by the vehicle's control unit. Training the algorithm(s) on the vehicle's control unit allows the algorithms to be trained using updated training data, for example, training data generated by the control unit, in order to optimize the accuracy of the generated injection and / or injection quantity correction models. This allows, in particular, recently occurring disturbances affecting the air-fuel ratio to be taken into account in the injection quantity correction model, in order to further reduce the occurrence of deviations from the stoichiometric air-fuel ratio.

[0025] In a further preferred embodiment, the multitude of state parameters relating to the at least one input variable and the associated combustion air ratios, and / or the multitude of state parameters relating to the intervention in the injection quantity and the associated further combustion air ratios, are transmitted to an external control unit. In other words, the corresponding measurement data acquired by the control unit are transmitted to an external computing unit for processing. By outsourcing the creation of the training data set(s) and the training of the machine learning algorithm(s), the computing power and capacity of the vehicle's control unit can be conserved or kept small, and the greater computing power and capacity of an external computing unit can be utilized.

[0026] Preferably, the injection model and / or the injection quantity correction model are received by the vehicle's external control unit. The required memory of the injection and / or injection quantity correction model is comparatively small, so even vehicles with control units of limited computing power and / or capacity can use the injection and / or injection quantity correction model to correct the injection quantities. Consequently, the advantages described herein can be made available to a large number of vehicles. Of course, it is additionally or alternatively preferred that the training data set and / or the further training data set and / or the trained algorithm and / or the further trained algorithm are received by the vehicle's external control unit. Then, the remaining steps to achieve the advantages presented herein can be performed by the vehicle's control unit.

[0027] Another aspect concerns a further procedure, namely a procedure for using the above injection quantity correction model, more precisely the injection quantity correction model built with the above-mentioned procedure.

[0028] In a further step of the procedure, the at least one state variable used as an input from the engine speed, intake manifold pressure, intake camshaft phase, and exhaust camshaft phase of the injection quantity correction model is determined. The choice of the state variable to be determined depends on the input variable(s) of the injection quantity correction model. Therefore, if more than one input variable is used to construct the injection quantity correction model, the corresponding state variables of the input variables are determined analogously.

[0029] In a further step of the procedure, a correction of the injection quantity of fuel into the combustion chamber of the gasoline engine is determined based on an output of the injection quantity correction model using at least one determined state variable as input.

[0030] Furthermore, the amount of fuel injected into the combustion chamber of the gasoline engine is adjusted according to the determined injection quantity correction. The optional features and their advantages described in the above (first) method can be implemented analogously using the further method and can therefore be combined with each other as desired.

[0031] Another aspect of the invention relates to a control unit for a gasoline engine, which is configured to carry out the aforementioned (first) method and / or the aforementioned further method. The gasoline engine is preferably a gasoline engine for a motor vehicle. The optional features and their advantages described in the methods can be implemented analogously with the control unit and are therefore arbitrarily combinable.

[0032] Another aspect concerns a vehicle with a gasoline engine, which includes the aforementioned control unit and at least one sensor for determining at least one of the state variables used as input to the injection quantity correction model generated according to the aforementioned (first) method: engine speed, intake manifold pressure, intake camshaft phase, and exhaust camshaft phase. Furthermore, it is understood that the vehicle includes a means for adjusting the amount of fuel injected into the combustion chamber of the gasoline engine, as well as a means for detecting the air-fuel ratio. Designs of such means are well known to those skilled in the art.

[0033] In a preferred embodiment, the vehicle further comprises an additional sensor for determining a further state variable of an air path of the gasoline engine, which is used as a further input variable of the injection quantity correction model and differs from the at least one state variable. Preferably, the vehicle comprises any combination or all of the sensors necessary for determining the aforementioned state variables. The specific design possibilities of the sensors required for detecting the state variables are generally known to those skilled in the art.

[0034] The aforementioned control unit is preferably implemented using electrical or electronic components (hardware) or firmware (ASIC). Additionally or alternatively, the functionality of the control unit is realized when a suitable program (software) is executed. Equally preferred is the control unit implemented using a combination of hardware, firmware, and / or software. For example, individual components of the control unit for providing specific functionalities are designed as separate integrated circuits or arranged on a common integrated circuit.

[0035] The individual components of the control unit are preferably configured as one or more processes running on one or more processors in one or more electronic computing devices and generated during the execution of one or more computer programs. The computing devices are preferably configured to cooperate with other components, for example, at least one sensor for determining a state variable, in order to implement the functionalities described herein. The instructions of the computer programs are preferably stored in a memory, such as a RAM element. However, the computer programs can also be stored in a non-volatile storage medium, such as a CD-ROM, flash memory, or the like.

[0036] It is also apparent to a person skilled in the art that the functionalities of several computing units (data processing devices) can be combined or combined in a single device, or that the functionality of a particular data processing device can be distributed across a large number of devices in order to realize the functionality of the control unit.

[0037] Another aspect concerns a computer program comprising commands which, when the program is executed by a computer, such as a control unit for a gasoline engine of a motor vehicle having at least one sensor for determining speed, intake manifold pressure, intake camshaft phase and exhaust camshaft phase, cause it to carry out one or both of the methods according to the invention, in particular a method for generating an injection quantity correction model for a gasoline engine or a method for using the injection quantity correction model.

[0038] Further preferred embodiments of the invention result from the other features mentioned in the dependent claims.

[0039] Unless otherwise stated in individual cases, the various embodiments of the invention mentioned in this application can be advantageously combined with one another.

[0040] The invention is explained below using exemplary embodiments with reference to the accompanying drawings. These show: Fig. 1 a schematic representation of a diagram for providing an injection model; Fig. 2a to 2d schematic representations of input and output signals from Fig. 1 over time; Fig. 3 a schematic representation of a training of a machine learning algorithm to build an injection quantity correction model; Fig. 4a to 4d schematic representations of input signals from Fig. 3 over time; Fig. 5 a schematic representation of a control unit with the injection quantity correction model from Fig. 3 according to one embodiment; and Fig. 6a to 6c schematic representations of a determined correction of the injection quantity and associated combustion air conditions over time.

[0041] The following describes a method for generating an injection quantity correction model 16 for a gasoline engine based on the Fig. 1, Fig. 2, Fig. 3 to Fig. 4 explained in more detail. Fig. 5 and Fig. Section 6 describes a method for using the injection quantity correction model 16 with reference to an embodiment of a motor vehicle with a control unit 18 for a gasoline engine, which has the generated injection quantity correction model 16 stored in a memory. With the method according to the invention, a deviation in the air-fuel ratio λ2 of the gasoline engine is corrected by means of a correction of the injection quantity K. inj reduced.

[0042] Fig. Figure 1 shows a schematic representation of a diagram for providing an injection model 10. The injection model 10 represents the effect of an intervention in the injection quantity Δ inj from fuel into a combustion chamber 12 of a gasoline engine to a combustion air ratio λ1 over time.

[0043] Typically, a predetermined injection quantity Inj of fuel, calculated, for example, from a filling model based on the steady-state operation of the gasoline engine, is injected into the combustion chamber 12 of the gasoline engine. To control the catalytic exhaust aftertreatment, gasoline engines include a lambda sensor (not shown) that determines the current deviation of the gas mixture leaving the combustion chamber 12 from the stoichiometric air-fuel ratio (λ = 1). If the determined air-fuel ratio λ deviates from the stoichiometric air-fuel ratio (i.e., λ ≠ 1), then either too much fuel (rich gas mixture, λ < 1) or too little fuel (lean gas mixture, λ > 1) has been injected into the combustion chamber 12.Since deviations from the stoichiometric air-fuel ratio λ increase fuel consumption and the emission of harmful exhaust gases, a lambda controller 14 regulates the air-fuel ratio λ1 based on the detected deviation by adjusting the injection quantity accordingly. For the purpose of the subsequently described setup of the injection model 10 and an injection quantity correction model 16, the lambda controller 14 is deactivated. In the subsequent use of the constructed injection quantity correction model 16, the lambda controller 14 is activated.

[0044] A fundamental correlation between a change in the injection quantity Δ inj and the combustion air ratio λ1 is determined using the schematic representations of input and output signals in the Fig. 2a to 2d are shown. Fig. Figure 2a shows a schematic progression of an exemplary intervention in the injection quantity Δ injof fuel into a combustion chamber 12 of a gasoline engine. An example test signal of the injection quantity Δ is shown. inj in the form of a pseudorandom binary sequence (PRBS), which can serve to obtain a database for the parameterization of the injection model 10. Fig. Figure 2b shows the corresponding schematic curve of the combustion air ratio λ1. Initially, there is no change to the injection quantity Δ. inj and the air-fuel ratio λ1 is stoichiometric (λ = 1). For example, the injection quantity Inj is determined by adjusting the injection quantity Δ inj The air-fuel ratio (λ1) is reduced (for example, by five percent), and a sluggish response to an increasingly lean gas mixture is evident up to a saturation point. This is then achieved by subsequently increasing the injection quantity (Inj) by adjusting the injection quantity (Δ). injThe value of the air-fuel ratio λ1 then drops again into the rich gas mixture range, reaching a minimum. After a further reduction in the injection quantity, the behavior of the air-fuel ratio λ1 essentially repeats itself.

[0045] Based on the in the Fig. 2c and Fig. 2d schematic representations of the rotational speed n eng and the intake manifold pressure p in It is evident that the in Fig. 2a and Fig. The curves shown in 2b were recorded in a steady-state operating mode of the Otto engine, since the rotational speed n eng and the intake manifold pressure p in are and will remain constant. In summary, the Fig. 2a and Fig. 2b therefore, it is noticeable that even in steady-state operating mode a stoichiometric control of the combustion air ratio λ1 is achieved by means of a corresponding intervention in the injection quantity Δ injis made more difficult by the inertia of the system.

[0046] In the dynamic operating mode of the Otto engine, where the speeds n eng and suction pressures p in Dynamically changing the injection quantity involves determining an exact time to intervene in the injection quantity Δ. inj all the more difficult. Therefore, in the method according to the invention, an injection model 10 is first provided which determines the effect of the injection quantity Δ. inj , more precisely, an intervention in the injection quantity Δ inj , from fuel into combustion chamber 12 of a gasoline engine to a combustion air ratio λ1 over time. For this purpose, a machine learning algorithm is used, namely an artificial neural network which is or was trained with a prepared training set. The prepared training set comprises a multitude of state parameters of state variables, namely the intervention in the injection quantity Δ inj ( Fig. 2a), a corresponding combustion air ratio λ1 ( Fig. 2b) and an associated rotational speed n eng ( Fig. 2c) and an associated intake manifold pressure p in ( Fig. 2d). The training data consists of the aforementioned state parameters in dynamic operating mode, i.e., at different values ​​of the rotational speeds n. eng and the intake manifold pressures p in , selected. As a result of the training, the artificial neural network learns a correlation between the injection timing and the required amount of intervention in the injection quantity Δ. inj and the influence on the air-fuel ratio λ1 and builds a corresponding injection model 10 which reflects this correlation. By including the rotational speeds n eng and the intake manifold pressures p inThe injection model 10 is also suitable for a dynamic operating mode of the gasoline engine. However, the present invention is not limited to the use of an artificial neural network. Rather, other mathematical methods can also be used to generate the injection model 10, for example, an FIR filter.

[0047] After the provision of the injection model 10, the next process step follows, namely the generation of an injection quantity correction model 16, which with regard to the Fig. 3 and 4a to 4d are explained in more detail below.

[0048] Fig. Figure 3 shows a schematic representation of training a machine learning algorithm to build the injection quantity correction model 16. Fig. Figures 4a to 4d show schematic representations of input signals over time for a training dataset prepared for training the algorithm. The prepared training dataset includes a multitude of associated state parameters with respect to the input variables rotational speed n. eng ( Fig. 4a) Intake manifold pressure p in ( Fig. 4b), Intake camshaft phase φ enw ( Fig. 4c), exhaust camshaft phase φ anw ( Fig. 4d) and associated combustion air conditions λ2 ( Fig. 6a). Of course, in the simplest case, it is sufficient to use at least one of the aforementioned input variables and the associated combustion air ratios λ2. For the sake of clarity, the procedure is described with regard to the four mentioned input variables, but is not limited to them.

[0049] Based on the prepared training data set, the algorithm learns the relationship between the changes in the state variables of the input signals, which affect the state variables of the air path of the spark-ignition engine, and a resulting deviation from a stoichiometric air-fuel ratio λ1, and builds a corresponding injection quantity correction model 16, which corrects the injection quantity K inj of fuel into the combustion chamber 12 of the Otto engine as a function of the input variables rotational speed n eng , Intake manifold pressure p in , Intake camshaft phase φ enw and exhaust camshaft phase φ anw The algorithm can estimate the injection quantity K. It uses the provided injection model 10 to determine the correct injection timing for the injection quantity correction. injto be able to specify. In other words, the injection quantity correction model 16 uses a learned or acquired relationship between the behavior of the input variables and a correction of the injection quantity K. inj , which leads as close as possible to a stoichiometric air-fuel ratio (λ ≈ 1), so that, based on the measured (exemplary) four input variables, the injection quantity Inj is adjusted to achieve an almost stoichiometric air-fuel ratio λ1 using the correction of the injection quantity K inj can be adapted.

[0050] The injection quantity correction model 16, constructed according to the invention, thus makes it possible to prevent or at least reduce deviations from the stoichiometric air-fuel ratio λ1 directly, i.e., proactively – and not reactively or with a delay, i.e., only after deviations have already occurred. Since the injection quantity correction model 16 is built using the injection model 10, the injection model 10 is no longer required or used when the injection quantity correction model 16 is employed (for example, in a control unit). Therefore, the injection quantities Inj can be effectively corrected not only in steady-state operating conditions of the gasoline engine, but also in dynamic operating conditions, in order to further reduce harmful emissions and excessive fuel consumption.

[0051] Fig. Figure 5 shows a schematic representation of a control unit 18 with the injection quantity correction model 16. Fig. 3 according to one embodiment, which is stored in a memory (not shown) of the control unit 18. The control unit 18 is a control unit of a motor vehicle with a gasoline engine (not shown). The motor vehicle includes a sensor for determining the state variables rotational speed n. eng , Intake manifold pressure p in , Intake camshaft phase φ enw and exhaust camshaft phase φ anw Using the example of a motor vehicle, a method for using the injection quantity correction model 16 is now described.

[0052] In a first step, state parameters of the four state variables used as input variables of the injection quantity correction model 16 are determined using the four sensors of the motor vehicle.

[0053] In a further process step, a correction of the injection quantity K is carried out. injof fuel into the combustion chamber 12 of the Otto engine on the basis of an output of the injection quantity correction model 16 to an input of the determined state variables into the injection quantity correction model 16.

[0054] Finally, the injection quantity Δ inj of fuel into the combustion chamber 12 of the gasoline engine according to the determined correction of the injection quantity K inj adjusted. The influence of the determined correction of the injection quantity K inj will be in view of the Fig. Sections 6a to 6c are described in more detail.

[0055] The Fig. Figures 6a to 6c show schematic representations of a determined correction of the injection quantity K. inj and associated combustion air conditions λ2, λ sum over time.

[0056] The Fig. Figure 6a shows an exemplary combustion air ratio λ2 without applied correction of the injection quantity K. injIt is evident that the combustion air ratio λ2, controlled by the conventional lambda controller 14, exhibits, in some cases, considerable deviations from the stoichiometric combustion air ratio λ1. The exemplary correction of the injection quantity K, corresponding to the combustion air ratio λ2 and output by the injection quantity correction model 16, is shown. inj is in Fig. Figure 6b illustrates this. By taking the injection model 10 into account during the training of the algorithm or during the construction of the injection quantity correction model 16, a sufficient improvement in the air-fuel ratio or mixture behavior can be achieved. This is shown in Fig. 6b is well illustrated by the exemplary temporally shifted maxima compared to those in Fig. The corresponding maxima shown in 6a are evident. Fig. 6c is the result of applying the correction of the injection quantity K inj out of Fig.6b on the injection quantity Inj in the obtained combustion air ratio λ sum The deviations from the stoichiometric combustion air ratio λ1 are shown. Compared to the course of the combustion air ratio λ2, the deviations are significantly reduced. Reference symbol list 10 Injection Model 12 Combustion chamber 14 Lambda controllers 16 Injection Quantity Correction Model 18 Control unit Inj Fuel injection quantity Δ inj Intervention in the injection quantity n eng speed p in Intake manifold pressure λ1 combustion air ratio K inj Correction of the injection quantity λ2 Combustion air ratio without correction of the injection quantity λ sum Combustion air ratio with correction of the injection quantity φ enw Intake camshaft phase φ anw Exhaust camshaft phase

Claims

Method for generating an injection quantity correction model (16) for a spark-ignition engine, wherein the method comprises the steps of: - providing an injection model (10) that maps the effect of an injection correction quantity (Δinj) of fuel into the combustion chamber (12) of a spark-ignition engine on an air-fuel ratio (λ1) over time, wherein the injection model (10) includes the correlation of how long after the injection time and how a changed injection correction quantity (Δinj) is reflected in the exhaust gas mixture leaving the combustion chamber (12), - generating a machine learning algorithm trained with a training dataset, wherein the algorithm uses the injection model (10) to construct an injection quantity correction model (16) that corrects the injection quantity (Kinj) of fuel into the combustion chamber (12) of the spark-ignition engine as a function of at least one state variable of rotational speed (neng) used as an input.The training dataset estimates intake manifold pressure (pin), intake camshaft phase (φenw) and exhaust camshaft phase (φanw), and comprises a variety of state parameters with respect to at least one input variable and associated combustion air ratios (λ2). Method according to claim 1, wherein the at least one state variable is the intake manifold pressure (pin). Method according to claim 1 or 2, wherein the training data set is extended by at least a plurality of associated state parameters of a further state parameter of an air path of the spark-ignition engine which differs from the at least one state parameter, and the injection quantity correction model (16) further estimates the correction of the injection quantity (Kinj) as a function of the further state parameter. Method according to claim 3, wherein the further state variable is rotational speed (neng), intake manifold pressure (pin), intake camshaft phase ((φenw), exhaust camshaft phase (φanw), aging state and temperature. Method according to one of the preceding claims, wherein the injection model (10) is provided using an FIR filter and / or a further machine learning algorithm. Method according to claim 5, wherein the further algorithm is trained with a further training data set to build the injection model (10), the training set comprising a plurality of state parameters relating to an intervention in the injection correction quantity (Δinj) and associated further combustion air ratios (λ1). Method according to claim 6, wherein the further training data set is extended by associated state parameters relating to the rotational speed (neng) and / or the intake manifold pressure (pin) of the engine. Method according to any of the preceding claims, wherein the algorithm and / or the further algorithm comprises an artificial neural network. Method according to one of the preceding claims, wherein the training data set and / or the further training data set is created by a control unit (18) of the vehicle and / or the algorithm and / or the further algorithm are trained by the control unit (18) of the vehicle. Method according to one of the preceding claims, wherein the plurality of state parameters relating to the at least one input variable and the associated combustion air ratios (λ2) and / or the plurality of state parameters relating to the intervention in the injection correction quantity (Δinj) and the associated further combustion air ratios (λ1) are transmitted to a vehicle-external control unit. Method according to claim 10, wherein the injection model (10) and / or the injection quantity correction model (16) are received by the vehicle-external control unit. Method for using the injection quantity correction model (16) according to one of the preceding claims, wherein the method comprises the steps: - Determining the at least one state variable used as an input variable of rotational speed (neng), intake manifold pressure (pin), intake camshaft phase ((φenw)) and exhaust camshaft phase (φanw) of the injection quantity correction model (16), - Determining a correction of the injection quantity (Kinj) of fuel into the combustion chamber (12) of the spark-ignition engine based on an output of the injection quantity correction model (16) using the at least one determined state variable as an input variable, and - Adjusting the injection correction quantity (Δinj) of fuel into the combustion chamber (12) of the spark-ignition engine according to the determined correction of the injection quantity (Kinj). Control unit (18) for a gasoline engine, which is configured to perform the method according to one of claims 1 to 11 and / or to perform the method according to claim 12, wherein the injection quantity correction model (16) of the method according to one of claims 1 to 11 is stored on the control unit (18). Vehicle with a gasoline engine, comprising: the control unit (18) according to claim 13 and at least one sensor for determining the at least one state variable used as an input variable of the injection quantity correction model (16) generated according to the method according to one of claims 1 to 11, namely speed (neng), intake manifold pressure (pin), intake camshaft phase ((φenw)) and exhaust camshaft phase (φanw). Vehicle according to claim 14 further comprising a further sensor for determining a further state variable of an air path of the gasoline engine which is used as a further input variable of the injection quantity correction model (16) and which differs from the at least one state variable.

Images