DE102020214953A1 - Method and device for training a classifier for controlling an injector of an internal combustion engine - Google Patents

Abstract

Computer-implemented method for training a classifier (60), the method for training comprising the following steps: • determining a training time series (xi), the training time series (xi) comprising a plurality of input signals, each input signal being assigned a position and the training time series (xi) a desired position is also assigned, which characterizes a closure time within the training time series (xi);• determining a plurality of first values using the classifier (60) and based on the training time series (xi), with a first value each having a Corresponds to the input signal of the training time series (xi) and the first value characterizes a probability with which the position assigned to the input signal is the desired position;• Determining a loss value, the loss value characterizing a weighted sum, the summands each being based on a difference the position of an input signal and the desired position and the weighting is based on the first value corresponding to the input signal;• determining a gradient of the loss value with respect to at least one parameter (Φ) of the classifier (60);• adjusting the at least one parameter (Φ) the classifier (60) based on the gradient.

Description

technical field

The invention relates to a method for training a classifier, a method for controlling an injector of an internal combustion engine, a control system for controlling an injector of an internal combustion engine, a training system for training a classifier, a computer program and a machine-readable storage medium.

State of the art

From the unpublished DE10 2020 211 418.8 a method for determining a closing time of an injector of an internal combustion engine by means of a classifier is known.

Advantages of the Invention

A direct injection of fuels in internal combustion engines, such as diesel engines or Otto engines with direct injection, allows advantageous operating properties of the internal combustion engine. One challenge is to control the combustion process as precisely as possible in order to improve the operating properties of the internal combustion engine, particularly with regard to fuel consumption, efficiency, pollutant emissions and smooth running.

To this end, it is essential to operate the injectors (also injection nozzles or injection valves) of the internal combustion engine in such a way that the fuel quantity to be injected can be metered with a high level of repeat accuracy. The injectors can, for example, have an electromagnetic actuator or piezo actuator which can actuate a valve needle of the injector in order to lift it off a needle seat and open an outlet opening of the injector for discharging the fuel into a combustion chamber.

Due to structural differences, structural tolerances and/or different operating conditions such as temperature, fuel pressure or fuel viscosity, there is some uncertainty in determining the exact time the injector is closed, i. H. the point in time at which the needle of the injector closes and no more fuel can get through the injector into the combustion chamber.

One possibility to determine the closing time very precisely is by means of a classifier, for example as disclosed in DE10 2020 211 418.8 . For example, a time series of input signals from a sensor can be transmitted to the classifier, with the input signals each characterizing measurements of the sensor with regard to a deformation of the injector. The individual input signals are preferably sorted in the time series according to their measurement time. At each measurement time, the classifier can then determine a probability with which the respective measurement time is a closing time of the injector.

A challenge with this method is to be able to determine the closure time exactly if the at least two greatest probabilities determined by the classifier are approximately the same and the measurement times are relatively far apart. It is therefore desirable that the output behavior of the classifier is unimodal, ie the greatest probabilities determined by the classifier all correspond to measurement times that are too close together.

The method with features of independent claim 1 allows a classifier to be trained for determining a closing time of an injector, the classifier being trained in such a way that a distribution of probabilities determined by the classifier is approximately unimodal. This makes it possible for the determination of a closure time to be more precise during operation of the classifier.

Disclosure of Invention

• • Ermitteln zumindest einer Trainingszeitreihe, wobei die Trainingszeitreihe eine Mehrzahl von Eingabesignalen umfasst, wobei jedem Eingabesignal eine Position zugeordnet ist, wobei die Position basierend auf einer Sortierung der Eingabesignale gemäß einer zeitlichen Anordnung der Eingabesignale innerhalb der Zeitreihe ermittelt wird, und der Trainingszeitreihe ferner eine gewünschte Position zugeordnet ist, die innerhalb der Trainingszeitreihe einen Verschlusszeitpunkt charakterisiert;
• • Ermitteln einer Mehrzahl von ersten Werten mittels des Klassifikators und basierend auf der Trainingszeitreihe, wobei jeweils ein erster Wert mit einem Eingabesignal der Trainingszeitreihe korrespondiert und der erste Wert eine Wahrscheinlichkeit charakterisiert, mit der es sich bei der dem Eingabesignal zugeordneten Position um die gewünschte Position handelt;
• • Ermitteln eines Verlustwertes, wobei der Verlustwert eine gewichtete Summe charakterisiert, wobei die Summanden jeweils auf einer Differenz der Position eines Eingabesignals und der gewünschten Position basieren und die Gewichtung auf dem mit dem Eingabesignal korrespondierenden ersten Wert basiert;
• • Ermitteln eines Gradienten des Verlustwertes bezüglich zumindest eines Parameters des Klassifikators;
• • Anpassen des zumindest einen Parameters des Klassifikators basierend auf dem Gradienten.

In a first aspect, the invention relates to a computer-implemented method for training a classifier, the classifier being designed to determine a classification of a closing time of the injector based on a time series of input signals from a sensor of an injector of an internal combustion engine, the method for training having the following steps includes:

• • Determining at least one training time series, the training time series comprising a plurality of input signals, each input signal being assigned a position, the position being determined based on a sorting of the input signals according to a temporal arrangement of the input signals within the time series, and the training time series also being a desired one Position is assigned that characterizes a closure time within the training time series;
• • determining a plurality of first values using the classifier and based on the Training time series, a first value in each case corresponding to an input signal of the training time series and the first value characterizing a probability with which the position associated with the input signal is the desired position;
• • determining a loss value, the loss value characterizing a weighted sum, the addends each being based on a difference between the position of an input signal and the desired position, and the weighting being based on the first value corresponding to the input signal;
• • determining a gradient of the loss value with respect to at least one parameter of the classifier;
• • Adjusting the at least one parameter of the classifier based on the gradient.

The method can be understood in such a way that a closing time of the injector is determined based on a time segment of input signals, ie signals or measurements of a sensor, of the injector. The time segment is the time series. The input signals are therefore captured by the sensor and can then be summarized in the time series. Within the time series, the input signals are preferably sorted based on their respective measurement point, ie a point in time at which they were measured by the sensor. This sorting allows each input signal within the time series to be assigned an index that characterizes the position of the input signal within the time series.

In particular, it is possible for the input signals to be determined using a piezo sensor, with the piezo sensor being set up to determine a deformation of the injector.

The inventors were able to find out that, with regard to possible operating conditions of the injector within the internal combustion engine, a piezo sensor is best suited for determining an exact deformation of the injector. Furthermore, the inventors were able to find out that a time series based on input signals of a deformation is extremely suitable for determining a closing time of the injector.

The time series can be in the form of a matrix, for example. In this matrix, the rows preferably contain the respective input signals of the sensor, while the columns each characterize one dimension of the input signal.

The time series is then fed to a classifier. In particular, for each position within the time series, the classifier can determine a probability with which the position characterizes a closing time of the injector. In other words, the classifier can determine at each measurement time characterized by the time series whether the measurement time is a closure time or not. In the method, the first values each characterize this probability.

The classifier can preferably include a neural network, by means of which the first values can be determined. The inventors were able to find out that the use of a neural network leads to the best accuracy when determining the time of closure.

Due to the described construction of the time series, the method can also be understood in such a way that it replaces the problem of regression of the closure time, which would otherwise occur when determining the closure time, with a classification of the closure time.

The training of the classifier based on the loss value can be understood in such a way that during the training it is penalized if a high probability is determined for a position that is not close to the desired position. This gives the classifier the incentive to learn during training that the greatest probabilities determined for a time series in each case should be close to one another with regard to the positions that correspond to them. This construction of the loss value means that the classifier can advantageously be trained in such a way that, after training, the classifier can determine a plurality of first values for a time series, with a distribution of the first values being approximately unimodal. The distribution of the first values can be understood as an empirical distribution in terms of statistics.

It is preferably possible for the steps of the method to be repeated iteratively. The steps can be repeated until a specified number of iterations has been reached. Alternatively or additionally, it is possible for the method to be ended if the loss value in relation to a time series reaches or falls below a predetermined threshold value or an average of loss values in relation to a plurality of time series reaches or falls below a predetermined threshold value. In this case, the time series can originate in particular from a validation data record.

In the step of determining the loss value, a summand can preferably be a power of the diff be reference. In this way, the punishment can advantageously be adapted to the training. This is advantageous because too high a penalty can lead to the classifier not being able to achieve a sufficiently good classification accuracy, since the learning objective (ie minimizing the loss value) cannot be achieved by too high a penalty. The exponent of the power can be understood in particular as a hyperparameter of the training and can be determined accordingly using known methods of hyperparameter search.

It is also possible that in the step of determining the loss value the weighting is a power of the first value.

ermittelt wird, wobei p_y ein erster Wert ist, der mit dem Eingabesignal an Position n korrespondiert, N die Anzahl an Eingabesignalen innerhalb der Zeitreihe ist, c die gewünschte Position ist, α ein Exponent der Potenz des ersten Wertes ist und β ein Exponent der Potenz der Differenz ist.If a single training time series is used to train the classifier, in particular if only one training time series is used per iteration of the method, the loss value can be calculated according to the function $$l\left(p_1,p_2,...,p_N,c,a,\beta \right)={\displaystyle \sum _{n=1}^N{p}_n^a}\cdot {\left(n-c\right)}^{\beta }$$

is determined, where p _y is a first value corresponding to the input signal at position n, N is the number of input signals within the time series, c is the desired position, α is an exponent of the power of the first value and β is an exponent of the power of the difference.

ermittelt werden, wobei p_m,n ein erster Wert ist, der mit dem Eingabesignal an Position n der m-ten Trainingszeitreihe der Mehrzahl von Trainingszeitreihen korrespondiert, M die Anzahl der Zeitreihen der Mehrzahl von Zeitreihen ist, N die Anzahl an Eingabesignalen innerhalb der jeweiligen Zeitreihen ist, α ein Exponent der Potenz des ersten Wertes ist und β ein Exponent der Potenz der Differenz ist.If a plurality of training time series are used to train the classifier, in particular if a plurality of training time series is used per iteration of the method, the loss value with respect to the plurality of training time series can be determined according to the formula $$l\left(p_{1.1},p_{1.2},...,p_{M,N},c,a,\beta \right)=\frac{1}{M}{\displaystyle \sum _{m=1}^M{\displaystyle \sum _{n=1}^N{p}_{m,n}^a}}\cdot {\left(n-c_i\right)}^{\beta }$$

be determined, where p _m,n is a first value that corresponds to the input signal at position n of the mth training time series of the plurality of training time series, M is the number of time series of the plurality of time series, N is the number of input signals within each is time series, α is an exponent of the power of the first value, and β is an exponent of the power of the difference.

The classifier trained with the method described can be used to control an injector after training.

• • Trainieren eines Klassifikators gemäß dem oben beschriebenen Verfahren zum Trainieren des Klassifikators;
• • Ermitteln einer Zeitreihe von Eingabesignalen eines Sensors des Injektors;
• • Ermitteln einer Mehrzahl von ersten Werte mittels des Klassifikators und basierend auf der Zeitreihe;
• • Ermitteln eines Verschlusszeitpunkts des Injektors basierend auf den ersten Werten;
• • Ansteuern des Injektors entsprechend des ermittelten Verschlusszeitpunkts.

In a further aspect, the invention therefore relates to a computer-implemented method for controlling an injector of an internal combustion engine, comprising the steps:

• • training a classifier according to the method described above for training the classifier;
• • determining a time series of input signals from a sensor of the injector;
• • determining a plurality of first values using the classifier and based on the time series;
• • determining a closing time of the injector based on the first values;
• • Activation of the injector according to the determined closing time.

The sensor used in the control method can preferably be a structurally identical sensor to the sensor whose input signals are included in the training of the classifier in the form of time series.

For the method for controlling the injector, the sensor of the injector can preferably determine an input signal at regular intervals. The input signals determined in this way can then be arranged into time series. The time series can therefore be understood as a plurality of sensor signals relating to a specified period of time, it being possible for the specified period of time to be defined via the size of the time series.

The determined closure time can then be used to determine an injection quantity of fuel that was delivered via the injector. The injector can then continue to be controlled based on the injection quantity. For example, it is conceivable that for at least one subsequent injection process by the injector, a setpoint quantity of fuel that is to be injected by the injector in the injection process is changed. The behavior of the internal combustion engine can therefore advantageously be influenced by determining the closing time, for example in such a way that lower consumption is made possible.

• 1 schematisch den Aufbau eines Trainingssystems;
• 2 schematisch den Aufbau eines Steuerungssystems zur Steuerung eines Injektors.

Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings show:

• 1 schematic of the structure of a training system;
• 2 schematically shows the structure of a control system for controlling an injector.

Description of the exemplary embodiments

1 shows an embodiment of a training system (140) for training a classi fikators (60) by means of a training dataset (T). The training data set (T) comprises a plurality of training time series (x _i ) of input signals that are used to train the classifier (60), the training data set (T) also containing a desired output signal (y _i ) for each training time series (x _i ) which corresponds to the training time series (x _i ) and characterizes a position of a closure instant within the training time series (x _i ).

For training, a training data unit (150) accesses a computer-implemented database (St ₂ ), the database (St ₂ ) making the training data set (T) available. The training data unit (150) determines at least one training time series (x _i ) and the desired output signal (y _i ) corresponding to the training time series (x _i ) from the training data set (T), preferably at random, and transmits the training time series (x _i ) to the classifier (60) . Based on the input signal (x _i ), the classifier (60) determines an output signal (ŷ _i ), which characterizes a probability for each position within the training time series (x _i ) that the respective position is a closure time. In the exemplary embodiment, the classifier (60) comprises a plurality of parameters (Φ) which determine the classification behavior of the classifier (60).

In the exemplary embodiment, the classifier (60) for classifying the training time series (x _i ) includes a neural network, to which the training time series (x _i ) is transmitted. The time series is preferably in the form of a matrix, with the rows containing the respective input signals of the time series, while the columns each characterize a dimension of the input signal. In the exemplary embodiment, the neural network is a neural convolutional network. In further exemplary embodiments, the rows or columns of the matrix can also be concatenated in order to obtain a vector representation of the time series. In these further exemplary embodiments, in particular a multilayer perceptron can be used as the neural network. Regardless of the type of neural network, the parameters (Φ) can include in particular the weights of the neural network.

In still other exemplary embodiments, the classifier (60) for classifying the training time series (x _i ) can also include other machine learning models, for example a support vector machine, a random forest or a Gaussian process processes).

Regardless of the machine learning model used, the training time series (x _i ) can be pre-processed by the training data unit (150) before it is made available to the classifier (60). As a pre-processing, the training time series (x _i ) can be normalized in particular with regard to an average time series of at least part of the time series of the training data set.

The desired output signal (y _i ) and the determined output signal (ŷ _i ) are transmitted to a changing unit (180).

Based on the desired output signal (y _i ) and the ascertained output signal (y _i ), the change unit (180) then determines new parameters (Φ′) for the classifier (60). For this purpose, the modification unit (180) compares the desired output signal (y _i ) and the determined output signal (ŷ _i ) using a loss function. The loss function determines a first loss value that characterizes how far the determined output signal (y _i ) deviates from the desired output signal (y _i ).

ermittelt wird, wobei p_y eine Wahrscheinlichkeit charakterisiert, dass eine Position n der Trainingszeitreihe (x_i) einen Verschlusszeitpunkt charakterisiert, N die Anzahl an Eingabesignalen innerhalb der Trainingszeitreihe (x_i) ist, c eine durch das gewünschte Ausgabesignal (y_i) charakterisierte Position eines Verschlusszeitpunkts ist, α ein Exponent einer Potenz der Wahrscheinlichkeit ist und β ein Exponent einer Potenz einer Differenz zwischen einer Position n und der durch das Ausgabesignal (y_i) charakterisierten Position c ist.In the exemplary embodiment, the loss value is calculated according to a loss function $$l\left(p_1,p_2,...,p_N,c,a,\beta \right)={\displaystyle \sum _{n=1}^N{p}_n^a}\cdot {\left(n-c\right)}^{\beta }$$

is determined, where p _y characterizes a probability that a position n of the training time series (x _i ) characterizes a closure time, N is the number of input signals within the training time series (x _i ), c is a position characterized by the desired output signal (y _i ). of a closure instant, α is an exponent of a power of the probability and β is an exponent of a power of a difference between a position n and the position c characterized by the output signal (y _i ).

ermittelt werden, wobei p_m,n eine Wahrscheinlichkeit charakterisiert, eine Position n der m-ten Trainingszeitreihe (x_i) der Mehrzahl von Trainingszeitreihen einen Verschlusszeitpunkt charakterisiert, M die Anzahl der Trainingszeitreihen des Stapels von Trainingszeitreihen ist, N die Anzahl an Eingabesignalen innerhalb der jeweiligen Trainingszeitreihe (x_i) ist, α ein Exponent der Potenz der Wahrscheinlichkeit ist und β ein Exponent der Potenz der Differenz ist.In alternative exemplary embodiments, it is also possible for the classifier (60) to be trained with a respective batch of training time series (x _i ). In these exemplary embodiments, the loss value can be calculated according to a loss function $$l\left(p_{1.1},p_{1.2},...,p_{M,N},c,a,\beta \right)=\frac{1}{M}{\displaystyle \sum _{m=1}^M{\displaystyle \sum _{n=1}^N{p}_{m,n}^a}}\cdot {\left(n-c_i\right)}^{\beta }$$

be determined, where p _m,n characterizes a probability, a position n of the m-th training time series (x _i ) of the plurality of training time series characterizes a closure time, M is the number of training time series of the stack of training time series, N is the number of input signals within the respective training time series (x _i ), α is an exponent of the probability power and β is an exponent of the difference power.

The changing unit (180) determines the new parameters (Φ') based on the loss value. In the exemplary embodiment, this is done using a gradient descent method, preferably Stochastic Gradient Descent, Adam, or AdamW. In alternative exemplary embodiments, the new parameters (Φ′) can also be determined using an evolutionary algorithm.

The determined new parameters (Φ') are stored in a model parameter memory (St ₁ ). The determined new parameters (Φ′) are preferably made available to the classifier (60) as parameters (Φ).

In further preferred exemplary embodiments, the training described is repeated iteratively for a predefined number of iteration steps or iteratively repeated until the loss value falls below a predefined threshold value. Alternatively or additionally, it is also conceivable that the training is ended when an average loss value with regard to a test or validation data record falls below a predefined threshold value. In at least one of the iterations, the new parameters (Φ') determined in a previous iteration are used as parameters (Φ) of the classifier (60).

Furthermore, the training system (140) can comprise at least one processor (145) and at least one machine-readable storage medium (146) containing instructions which, when executed by the processor (145), cause the training system (140) to implement a training method according to one of the aspects of the invention.

2 shows a control system (40) for controlling an injector (20) of an internal combustion engine by means of the classifier (60). A deformation of the injector (20) is detected by a sensor (30) at preferably regular time intervals. In the exemplary embodiment, the sensor (30) is a piezo sensor. In alternative exemplary embodiments, other sensors (30) are also possible for determining a deformation of the injector, for example sensors (30) based on a strain gauge.

A measurement (S) determined by the sensor (30) is transmitted to the control system (40). The control system (40) thus receives a series of measurements (S). From this, the control system (40) determines control signals (A) which are transmitted to a control unit (10) of the injector (20).

The control system (40) receives the sequence of measurements (S) from the sensor (30) in a receiving unit (50) which converts the sequence of measurements (S) into a time series (x) of input signals. The time series can be determined, for example, by selecting previous measurements and the current measurement (S). Alternatively, it is conceivable that the time series includes a predefined number of measurements from the past and the current measurement (S). In other words, the time series (x) is determined depending on the sensor signal (S). The time series (x) of input signals is fed to the classifier (60).

The classifier (60) is preferably parameterized by the parameters (Φ), which are stored in a parameter memory (P) and are provided by this.

The classifier (60) determines an output signal (y) based on the time series (x). The output signal (y) is fed to an optional conversion unit (80), which uses it to determine control signals (A) which are fed to the control unit (10) of the injector (20) in order to control the injector (20) accordingly.

The control unit (10) receives the control signals (A), is controlled accordingly and carries out a corresponding action. The control unit (10) can include a control logic (not necessarily structurally integrated), which determines a second control signal from the control signal (A), with which the injector (20) is then controlled.

In further embodiments, the control system (40) includes the sensor (30). In yet further embodiments, the control system (40) alternatively or additionally also comprises the control unit (10).

In further preferred embodiments, the control system (40) comprises at least one processor (45) and at least one machine-readable storage medium (46) on which instructions are stored which, when they are executed on the at least one processor (45), the control system ( 40) cause the method according to the invention to be carried out.

In alternative embodiments, as an alternative or in addition to the control unit (10), at least one further device (10a) is controlled by means of the control signal (A). The device (10a) can be, for example, a pump of a common rail system to which the injector (20) belongs. Alternatively or additionally, it is conceivable that the device is a control unit of the internal combustion engine. Alternatively or In addition, it is also conceivable that the device (10a) is a display unit, by means of which the information of the classification determined by the classifier (60) can be displayed to a person (eg a driver or a mechanic).

The term "computer" includes any device for processing predeterminable calculation rules. These calculation rules can be in the form of software, or in the form of hardware, or in a mixed form of software and hardware.

In general, a plurality can be understood as indexed, i.e. each element of the plurality is assigned a unique index, preferably by assigning consecutive integers to the elements contained in the plurality. Preferably, when a plurality comprises N elements, where N is the number of elements in the plurality, integers from 1 to N are assigned to the elements.

QUOTES INCLUDED IN DESCRIPTION

This list of the documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• DE 102020211418 [0002, 0006]

Claims (12)

Computer-implemented method for training a classifier (60), wherein the classifier (60) is designed to determine a classification of a closing time of the injector based on a time series (x) of input signals of a sensor (30) of an injector (20) of an internal combustion engine, wherein the method for training comprises the following steps: • determining at least one training time series (x _i ), the training time series (x _i ) comprising a plurality of input signals, each input signal being assigned a position, the position being based on a sorting of the input signals according to a temporal arrangement of the input signals within the time series is determined, and the training time series (x _i ) is also assigned a desired position which characterizes a closure time within the training time series (x _i ); • Determination of a plurality of first values by means of the classifier (60) and based on the training time series (x _i ), with a first value corresponding to an input signal of the training time series (x _i ) and the first value characterizing a probability with which it the position associated with the input signal is the desired position; • determining a loss value, the loss value characterizing a weighted sum, the addends each being based on a difference between the position of an input signal and the desired position, and the weighting being based on the first value corresponding to the input signal; • determining a gradient of the loss value with respect to at least one parameter (Φ) of the classifier (60); • Adjusting the at least one parameter (Φ) of the classifier (60) based on the gradient. procedure after claim 1 , wherein the input signals are determined by means of a piezo sensor (30), wherein the piezo sensor (30) is set up to determine a deformation of the injector (20). Procedure according to one of claim 1 or 2 , wherein the classifier (60) comprises a neural network (61) by means of which the plurality of first values is determined. Method according to one of the preceding claims, wherein in the step of determining the loss value the summand is a power of the difference. A method according to any one of the preceding claims, wherein in the step of determining the loss value the weighting is a power of the first value. A method according to any one of the preceding claims, wherein the loss value is according to the function $$l\left(p_1,p_2,...,p_N,c,a,\beta \right)={\displaystyle \sum _{n=1}^N{p}_n^a}\cdot {\left(n-c\right)}^{\beta }$$

is determined, where p _n is a first value corresponding to the input signal at position n, N is the number of input signals within the time series (x _i ), c is the desired position, α is an exponent of the power of the first value and β is an exponent of the power of the difference. Procedure according to one of Claims 1 until 5 , where the loss value with respect to a plurality of training time series (x _i ) and according to the formula $$l\left(p_{1.1},p_{1.2},...,p_{M,N},c,a,\beta \right)=\frac{1}{M}{\displaystyle \sum _{m=1}^M{\displaystyle \sum _{n=1}^N{p}_{m,n}^a}}\cdot {\left(n-c_m\right)}^{\beta }$$

is determined, where p _m,n is a first value that corresponds to the input signal at position n of the mth training time series (x _i ) of the plurality of training time series, M is the number of time series (x _i ) of the plurality of time series (x _i ), N is the number of input signals within the respective time series (x _i ), α is an exponent of the power of the first value and β is an exponent of the power of the difference. Computer-implemented method for controlling an injector (20) of an internal combustion engine, comprising the steps of: • Training a classifier (60) according to one of Claims 1 until 7 ; • determining a time series (x) of input signals from a sensor (30) of the injector (20); • determining a plurality of first values by means of the classifier (60) and based on the time series (x); • determining a closing time of the injector (20) based on the first values; • Activation of the injector (20) according to the determined closure time. Training device (140), which is set up, the method according to any one of Claims 1 until 7 to execute. Control system (40), which is set up according to the method claim 8 to execute. Computer program which is set up, the method according to one of Claims 1 until 8th to be executed when executed by a processor (45, 145). Machine-readable storage medium (46, 146) on which the computer program claim 11 is saved.

Images