Method and device for operating a fuel injection valve by means of a machine learning method

Abstract

The invention relates to a method for operating an injection valve (6) by determining an opening or closing instant of the injection valve (6) on the basis of a sensor signal (S), the method having the following steps: - providing (S1, S2) a sequence of evaluation point times (Z) by sampling a sensor signal of a sensor of the injection valve; - using (S3) a first, nonlinear, data-based submodel (21) in order to obtain a first output vector (A1) on the basis of the sequence of evaluation point times (Z), wherein each element of the first output vector is assigned to a determined instant; - using (S4) a second, linear, data-based submodel (22) in order to obtain a second output vector (A2) on the basis of the sequence of evaluation point times (Z), wherein each element of the second output vector (A2) is assigned to a determined instant; - limiting (S5, S6) the instants determined by the first output vector (A1) in dependence on the second output vector (A2) in order to obtain the opening or closing instant.

Description

Technical Field

[0001] The present invention relates to a method for operating a fuel injection valve of an internal combustion engine using a data-based model, particularly for determining the opening and / or closing times to estimate the amount of fuel injected. Background Technology

[0002] Electromechanical or piezoelectric injection valves are used to meter fuel in combustion engines. These injection valves enable direct and precise measurement of fuel supply to the cylinders of the combustion engine.

[0003] The challenge lies in controlling the combustion process as precisely as possible to improve the operating characteristics of the combustion engine, particularly in terms of fuel consumption, efficiency, harmful emissions, and smooth operation. Importantly, this requires operating the injection valves to measure the amount of fuel to be injected with high repeatability, even under varying operating pressures and with the need for multiple injections per duty cycle.

[0004] The injection valve may have an electromagnetic actuator or a piezoelectric actuator that manipulates a valve needle to remove the needle from the needle valve seat and open the injection valve's outlet to release fuel into the combustion chamber. Due to structural differences and varying operating conditions, such as temperature, fuel pressure, and fuel viscosity, there is uncertainty in determining the exact opening and closing times of the injection valve; the opening time is defined as the moment from which fuel reaches the cylinder's combustion chamber through the injection valve; the closing time is defined as the moment from which fuel reaches the cylinder's combustion chamber through the injection valve until that moment. Summary of the Invention

[0005] According to the present invention, a method, apparatus and injection system for operating a fuel injection valve are provided.

[0006] Other construction methods are described in this invention disclosure.

[0007] According to the first aspect, a method is provided for operating an injection valve by determining the opening or closing time of the injection valve based on sensor signals, the method comprising the following steps:

[0008] - By sampling the sensor signals of the injection valve's sensors, an evaluation point time series is provided;

[0009] - Use a non-linear, data-based first sub-model to obtain a first output vector based on the time series of evaluation points, where each element of the first output vector is assigned to a specific time.

[0010] - Use a linear, data-based second sub-model to obtain a second output vector based on the evaluation point time series, where each element of the second output vector is assigned to a specific time.

[0011] - Based on the second output vector, restrict the time determined by the first output vector in order to obtain the on or off time.

[0012] Although the injection valve is controlled according to a pre-defined change process of the control signal, the resulting changes in the opening and closing motion of the injection valve make it impossible to precisely pre-determine the actual opening and closing times for starting and ending fuel injection. This is due to the complex correlation between valve movement and the current operating point.

[0013] To monitor valve movement, a piezoelectric sensor is installed in the injection valve. This piezoelectric sensor is configured as a pressure sensor to detect pressure changes triggered by the operation of the injection valve and to provide a corresponding sensor signal. The measured sensor signal can then be evaluated to determine the actual opening and closing times of the injection valve, thus adapting the control of the injection valve accordingly.

[0014] However, sensor signals are also noisy, and are particularly dependent on the actual fuel pressure in the fuel supply and the duration of the control being measured.

[0015] Data-driven evaluation models can be used to assess sensor signals to determine the opening and / or closing times of the injection valve. These models can correspond to or include probabilistic regression models, neural networks, or classification models.

[0016] Using data-based evaluation models to determine when to turn on or off is advantageous because proper modeling, especially with the aid of physical excitation models, cannot be mapped due to the diverse range of influences and interactions. However, in safety-critical applications, the use of purely data-based models is questionable because it cannot be guaranteed at every operating point that the evaluation model's output will not lead to undesirable system behavior.

[0017] Therefore, according to the method mentioned above, the evaluation model can be set as a hybrid model that independently performs the determination of the opening and / or closing times based on two data-based sub-models, and combines the results of the sub-models with each other to obtain the opening and closing times. For this purpose, a first data-based sub-model is set up, which maps the evaluation point time series to the model output describing the opening and / or closing times, wherein the evaluation point time series is derived from the sensor signal of the piezoelectric sensor in the injection valve. Thus, each sub-model independently describes its respective opening and / or closing times.

[0018] The model output can be an output vector in Logit form. The output vectors of the two sub-models are respectively defined such that the indices of the elements of each output vector indicate the corresponding on or off time. For example, in the case of n evaluation points, the output vector can correspondingly include n elements. Here, the indices of the output vector elements are assigned to consecutive time points within the observed evaluation time period. In particular, the time points assigned to the elements of the sub-model's output vector can correspond to evaluation time points that are evenly spaced in time.

[0019] Therefore, the index of the element with the largest value in the output vector indicates the on and off moments to be determined during the evaluation period. In this way, a sub-model can be constructed to describe the value of each element in the output vector with the probability that the moment determined by the element's index corresponds to the on and / or off moment to be output. Specifically, the on or off moment can be determined by determining the index of the element in the output vector with the corresponding highest value and assigning that index to the corresponding on or off moment.

[0020] The index values ​​are assigned to the corresponding times in advance by training the sub-model, and this is explained in particular with respect to the crankshaft angle, or by referring to the starting time of the injection valve controlled by the control signal for the actuator unit of the injection valve.

[0021] The results from these two sub-models are combined to obtain the actual on or off moments. This allows for a trade-off between the reliability of the second sub-model and the accuracy of the first sub-model.

[0022] The first sub-model based on the data can be constructed as a nonlinear neural network. Nonlinear neural networks typically have multiple layers of neurons, at least one of which has a nonlinear activation function. Preferably, the output layer is configured without a nonlinear activation function. This allows for the modeling of complex nonlinear relationships as well.

[0023] The data-based second sub-model can be constructed as a linear neural network, preferably with only one layer. The linear second sub-model does not have a non-linear activation function, and therefore differs from the non-linear first sub-model in its lower complexity. The data-based second sub-model can be configured as a linear model that assigns the evaluation point time series to a time step in the same manner as the first sub-model. Preferably, the sub-model is constructed such that the output format for turning the time steps on or off is the same.

[0024] In linear neural networks, the input vector is multiplied by a matrix, and the result is provided as the model output. In particular, no non-linear activation function is set.

[0025] Especially in safety-critical applications, such as determining and monitoring the amount of fuel to be injected by the injection valve, an understandable function used to measure the fuel quantity is required. Since the data-based first sub-model corresponds to a nonlinear neural network and thus a black-box model, the function used to determine the output vector cannot be analyzed. The linear, data-based second sub-model, however, is an understandable mapping.

[0026] Since the first sub-model is constructed as a purely data-based classification model, it can typically be trained using training data to determine more accurate information about the turn-on or turn-off times. To avoid making safety-critical, erroneous determinations of the turn-on or turn-off times, the times determined by the first sub-model are constrained by the output of the second sub-model.

[0027] For example, this can be achieved by limiting the deviation from the time determined by the second sub-model, where the time determined by the first sub-model is limited to the limiting range so that the time determined by the first sub-model is output as the on or off time.

[0028] By using a linear second sub-model, the impact of each entry in the time series at the evaluation point can be explicitly understood. In this case, unexpected model outputs may not occur. This is particularly important for safety-critical applications where so-called black-box models, in which the computational process is not understandable, are not permitted.

[0029] In particular, the moment determined by the first output vector can be limited in such a way that the first index value of the first output vector is limited to an index value range that extends within a range defined by a pre-given tolerance threshold around the second index value of the second output vector.

[0030] Alternatively, the index values ​​determined from the first and second output vectors can be assigned to corresponding times, wherein the times determined from the second sub-model are used to pre-define a range of constraints by means of a pre-given tolerance distance, which limits the times determined from the first sub-model, so as to determine the on or off times.

[0031] According to another aspect, a device is provided for performing one of the methods mentioned above.

[0032] Alternatively, a first sub-model and a second sub-model can be created to directly output the on and / or off times, respectively. In this way, at least the data-based first sub-model can also be set as a regression model to output the actual times, i.e., the on or off times, based on the training of the sub-model.

[0033] It can be configured to operate the injection valve according to the opening and / or closing times, wherein the operation of the injection valve is specifically performed such that the opening duration of the injection valve, determined by the determined opening and / or closing times, is set to a pre-given desired opening duration.

[0034] According to another aspect, a method is provided for training a sub-model for a data-based evaluation model used to determine the opening and / or closing times of an injection valve, the method comprising the following steps:

[0035] - Provide training datasets that illustrate the determined opening and / or closing times for evaluation point time series;

[0036] - Using the training dataset, train the first sub-model, which is constructed as a non-linear, data-based model;

[0037] - Using the training dataset, a second sub-model is trained, which is constructed as a linear, data-based model.

[0038] Correspondingly, the sub-model can be trained from a pre-given time series of evaluation points with corresponding labels. In the case of a regression model, the labels may correspond to corresponding on or off times. In the case of a classification model, the labels may correspond to an output vector with multiple elements, where the index values ​​of the elements are assigned to the corresponding on or off times. For example, the label output vector may assign "0" to all elements, and only assign "1" to the element whose index value is to be assigned to a specific on or off time.

[0039] After training, when using the corresponding sub-model for inference, the values ​​of the relevant elements indicate the following probabilities: the assigned time corresponds to the on or off time with these probabilities.

[0040] The linear function of the second sub-model corresponds to a data-based linear model that evaluates the time series at the evaluation points using matrix multiplication. Unlike the calculations in the first sub-model, the calculations in this linear model are understandable due to the linearity of the matrix multiplication. This is particularly important for security-critical applications where so-called black-box models, in which the computational process is not transparent, are not permitted. Attached Figure Description

[0041] The implementation details are then elaborated upon in more detail, based on the accompanying diagrams.

[0042] Figure 1 A schematic diagram of an injection system for injecting fuel into the cylinders of a combustion engine is shown.

[0043] Figure 2 A flowchart illustrating the following method is shown: the method is used to apply a trained, data-based evaluation model to determine the amount of fuel injected;

[0044] Figure 3 A block diagram is shown to illustrate a method for determining the opening or closing timing of an injection valve using a data-based evaluation model; and

[0045] Figure 4 A schematic diagram illustrating an exemplary time-varying process of the evaluation point time series is shown. Detailed Implementation

[0046] Figure 1 The layout of the injection system 1 for a combustion engine 2 in a motor vehicle is shown, with an exemplary representation of one cylinder 3 (particularly one of a plurality of cylinders) for the combustion engine 2. The combustion engine 2 is preferably constructed as a diesel engine with direct injection, but it can also be configured as a gasoline engine.

[0047] Cylinder 3 has an intake valve 4 and an exhaust valve 5 for supplying fresh air and for discharging combustion exhaust gases.

[0048] Furthermore, in order to operate the combustion engine 2, fuel is injected into the combustion chamber 7 of the cylinder 3 via the injection valve 6. For this purpose, fuel is delivered to the injection valve via a fuel delivery device 8, which supplies fuel at high fuel pressure in a manner known per se (e.g., common rail).

[0049] The injection valve 6 has an electromagnetically or piezoelectrically operable actuator unit 61 coupled to a valve needle 62. In the closed state, the valve needle 62 rests on a needle valve seat 63. By manipulating the actuator unit 61, the valve needle 62 moves longitudinally, releasing the valve opening in the needle valve seat 63. This part is designed to inject fuel under pressure into the combustion chamber 7 of cylinder 3.

[0050] The injection valve 6 also has a piezoelectric sensor 65 arranged within it. The piezoelectric sensor 65 deforms due to pressure changes in the fuel guided by the injection valve 6 and generates a sensor signal via a voltage signal.

[0051] Fuel injection is controlled by a control unit 10, which pre-sets the amount of fuel to be injected by energizing the actuator unit 61. The sensor signal is sampled in time using an A / D converter 11 in the control unit 10, specifically at a sampling rate of 0.5 MHz to 5 MHz.

[0052] During the operation of the combustion engine 2, sensor signals are used to determine the correct opening or closing time of the injection valve 6. For this purpose, the sensor signals are digitized into an evaluation point time series using an A / D converter 11 and evaluated using a suitable evaluation model. Based on fuel pressure and other operating variables, the opening duration of the injection valve 6 can be determined, and the amount of fuel injected can be determined accordingly. To determine the opening duration, it is particularly necessary to specify the opening and closing times so that the opening duration can be determined as the time difference between these variables.

[0053] By observing the changes in sensor signals, the on and / or off times can be determined. In particular, the on and / or off times can be determined using a data-based evaluation model.

[0054] Figure 2 A method for determining the opening and / or closing timing of the injection valve 6 of cylinder 3 in engine system 1 is illustrated based on a flowchart. This method is used to sense the control functions of combustion engine 2, and in particular to determine the amount of fuel injected.

[0055] Combination Figure 3 The block diagram illustrates the method in more detail. The method is implemented in a control unit 10, which can be implemented in software and / or hardware.

[0056] The method is described based on observations of the injection valve 6 of a single cylinder 3 in the combustion engine 2. Of course, this method can be applied in parallel to all injection valves 6 of the combustion engine 3, that is, it can also be applied in parallel to a multi-cylinder combustion engine.

[0057] In step S1, a sensor signal is detected using a piezoelectric sensor 65. This signal is typically a voltage signal generated due to pressure changes in the delivered fuel.

[0058] In step S2, the sensor signal is sampled using the A / D converter 11 to determine the evaluation point time sequence Z within the evaluation period. The evaluation period can be defined by the control time window of the injection valve. The control time window is defined by the start of the control actuator unit 61 and a defined duration, which is the maximum duration for which the control signal for the actuator unit 61 pre-given valve opening. Therefore, the control time window has a defined time base, against which the evaluation point time sequence Z is provided as follows: the evaluation point time sequence Z is the basis for further determining the opening or closing time. In particular, the evaluation point time sequence can be determined by downsampling the previously oversampled sensor signal.

[0059] The evaluation period can be set using a fixed time reference for the working cycle of the combustion engine 2, and in particular, the evaluation period can preferably begin within the compression stroke given a predetermined crankshaft position. The evaluation period can be selected such that the entire opening time window of the injection valve can be mapped to it. Figure 4 The figure shows such an evaluation time period with an exemplary evaluation point time series Z.

[0060] In step S3, the evaluation point time series Z is fed to the data-based first sub-model 21. The data-based first sub-model 21 may be a trained neural network with multiple layers and / or something similar. At least one layer has one or more neurons, which contain activation functions, particularly non-linear activation functions. An output layer is configured without activation functions, which provides the model output.

[0061] The first sub-model 21 provides Logit as the first output vector A1. The first output vector A1 includes a plurality of elements, which are respectively assigned to possible on or off times. Preferably, the possible on or off times are assigned according to the rising index value of the elements.

[0062] In step S4, the evaluation point time series Z is fed to the second sub-model 22. As a data-based linear model, the second sub-model 22 can be trained using training data. The data-based second sub-model 22 can be constructed as a linear neural network, preferably with only one layer. In the linear neural network, the input vector is multiplied by a matrix, and the result is provided as the model output. No activation function, especially a non-linear one, is set. By using the linear sub-model, the influence of each entry in the evaluation point time series Z can be explicitly understood.

[0063] Therefore, the second sub-model 22 can also be applied in security-critical fields because the second sub-model 22 is understandable and can exclude unexpected model outputs.

[0064] The evaluation result of the evaluation point time series Z in the second sub-model 22 corresponds to the same format as in the first sub-model 21, namely the second output vector A2. From this second output vector A2, the second on or off time can be determined in the manner described above.

[0065] The first sub-model 21 and the second sub-model 22 are trained using a training dataset containing evaluation point time series Z and associated output vectors, respectively. The output vector encodes the opening and / or closing moments to be trained, which have been pre-determined by measuring the injection valve on a test bench, such that only elements whose index values ​​correspond to the opening and / or closing moments to be trained are assigned high values, especially "1", while the remaining elements are assigned low values, especially "0".

[0066] In step S5, the results of the sub-models (i.e., the first output vector A1 and the second output vector A2) are combined with each other in combination block 23. For this purpose, output vectors A1 and A2 are analyzed separately, given the on or off time to be determined. This is done for these two sub-models (21, 22) by determining the index of the element with the highest value using the arg-max function, where such determined index is assigned to the on or off time.

[0067] Here, using a pre-given tolerance distance, the index values ​​determined by the second sub-model form a constraint range around the corresponding index values ​​determined by the second sub-model 22, wherein the pre-given tolerance distance can be a pre-given integer tolerance distance. The constraint range restricts the output of the first sub-model 21 in such a way that the index values ​​determined from the first sub-model are restricted to the constraint range.

[0068] In other words, if two index values ​​are equal, that index value is used to determine the time of opening or closing. If the index value determined from the first output vector A1 is within the range of the index value determined by the second output vector A2, then the index value determined by the first output vector A1 is used. If the index value determined from the first output vector A1 is not within the range of the index value determined by the second output vector A2, then the index value determined by the first output vector A1 is restricted to the index value determined by the second output vector A2.

[0069] In step S6, the on or off time to be determined in evaluation block 24 is determined by selecting the corresponding assigned time from the constrained index values. This assignment corresponds to a pre-given time grid for the output vectors of the two sub-models.

[0070] Alternatively, the index values ​​determined from the first and second output vectors A1 and A2 can also be assigned to the corresponding times before constraints are applied. From the times determined from the second sub-model, a predetermined constraint range is given by means of a pre-defined tolerance distance, which limits the times determined from the first sub-model in order to determine the on or off times.

Claims

1. A method for operating the injection valve (6) by determining the opening or closing time based on a sensor signal (S), comprising the following steps: - By sampling the sensor signals of the sensor of the injection valve, a time series (Z) of evaluation points (S1, S2) is provided; - Use (S3) a non-linear, data-based first sub-model (21) to obtain a first output vector (A1) based on the evaluation point time series (Z), wherein each element of the first output vector (A1) is assigned to a specific time. - Use (S4) a linear, data-based second sub-model (22) to obtain a second output vector (A2) based on the evaluation point time series (Z), wherein each element of the second output vector (A2) is assigned to a specific time. - Based on the second output vector (A2), limit (S5, S6) the time determined by the first output vector (A1) to obtain the opening or closing time; The first sub-model (21) and the second sub-model (22) are constructed such that the value of each element of the first output vector and the second output vector is described respectively according to the following probability: the time determined by the index value of the element corresponds to the open or closed time to be output with the probability.

2. The method according to claim 1, wherein, The first output vector (A1) and the second output vector (A2) each include a plurality of elements, wherein the first or second index value of the element indicates a specific moment for the opening or closing moment, wherein consecutive index values ​​correspond to consecutive opening or closing moments having equal or variable distances in time, and the same first and second index values ​​correspond to the same opening or closing moment.

3. The method according to claim 2, wherein, The opening or closing time determined by the first output vector (A1) is limited in the following way: the first index value of the first output vector (A1) is limited to an index value range that extends within a range defined by the second index value through a pre-given tolerance threshold.

4. The method according to claim 1, wherein, The index values ​​determined from the first output vector (A1) and the second output vector (A2) are assigned to the corresponding time, wherein a pre-given tolerance distance is used to pre-given a constraint range that limits the time determined from the first sub-model in order to determine the opening or closing time.

5. The method according to any one of claims 1 to 4, wherein, The first sub-model (21) is constructed as a nonlinear neural network, while the second sub-model (22) is constructed as a linear neural network.

6. The method according to any one of claims 1 to 4, wherein, The operation of the injection valve (6) is performed according to the opening and / or closing times, wherein the operation of the injection valve (6) is performed such that the opening duration of the injection valve (6), determined by the determined opening and / or closing times, is set to a pre-given desired opening duration.

7. An apparatus for performing one of the methods according to any one of the preceding claims.

8. A computer program product comprising instructions that, when implemented by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 6.

9. A machine-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 6.

Images