Method and apparatus for training a data-based time point determination model with machine learning methods to determine an opening or closing time point of an injection valve

Abstract

The invention relates to a computer-implemented method for training a data-based time point determination model based on sensor signals for determining an opening or closing time point of an injection valve, having the steps: providing at least one unlabeled evaluation point time series by sampling a sensor signal of an injection valve; training the time point determination model in order to assign time point information to the evaluation point time series, which represents a determined opening or closing duration, wherein the training is performed by means of a first and a second shift function, wherein for the training a consistency loss function is used, which describes a difference between a first time point information and a second time point information, the first time point information corresponding to a model output of the time point determination model for the evaluation point time series shifted in time according to the first shift function, the second time point information describing a time point information shifted in time according to the second shift function, which corresponds to a model output of the time point determination model for the at least one evaluation point time series.

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 time point determination model, particularly for determining the opening or closing time points to estimate the amount of fuel injected. Background Technology

[0002] Electromechanical or piezoelectric injection valves are used to meter fuel in internal combustion engines. These valves can directly and accurately meter the fuel supply to the cylinders of the internal combustion engine.

[0003] One challenge is to control the combustion process as precisely as possible to improve the operating characteristics of internal combustion engines, particularly regarding fuel consumption, efficiency, emissions, and operational smoothness. To this end, it is important to operate the injection valves so that the amount of fuel to be injected can be metered with high repeatability, at varying operating pressures, and, if necessary, with multiple injections in each power stroke.

[0004] The injection valve may have an electromagnetic actuator or a piezoelectric actuator that manipulates a valve needle to lift the needle from its seat and open the injection valve's outlet orifice to discharge fuel into the combustion chamber. Due to structural differences and variations in operating conditions such as temperature, fuel pressure, and fuel viscosity, there is uncertainty in determining the precise opening time of the injection valve—the start time when fuel enters the combustion chamber of the cylinder through the injection valve—and the precise closing time of the injection valve—the end time when fuel enters the combustion chamber of the cylinder through the injection valve. Summary of the Invention

[0005] According to the present invention, a method for training a data-based time-point determination model for use in an injection system to determine the opening or closing time of an injection valve is provided, as well as an apparatus and injection system according to the parallel claims.

[0006] Other design options are described in the dependent claims.

[0007] According to the first aspect, a computer-implemented method is provided for training a data-based time-point determination model based on sensor signals for use in an injection system to determine the opening or closing time of an injection valve. This method comprises the following steps:

[0008] - By sampling the sensor signals of the injection valve's sensor, at least one unlabeled evaluation point time series is provided;

[0009] - Train a data-driven time-point determination model to assign time-point information to the evaluation time series, which represents the determined on or off duration.

[0010] The training is performed using a first shift function to shift the evaluation point time series over time and a second shift function to shift the time point information over time.

[0011] For this training, a consistency loss function is used, which describes the difference between the time-point information of the evaluation point time series shifted over time according to the first shift function and the time-point information of at least one evaluation point time series shifted over time according to the second shift function.

[0012] In particular, a data-based time point determination model can be trained based on a prediction loss function and using at least one training dataset, which assigns the evaluation point time series to specific time point information as labels.

[0013] Although the injection valve is controlled according to a pre-defined change process of the control signal, the resulting opening and closing movements of the injection valve change, making it impossible to accurately pre-determine the actual opening and closing times for starting and ending fuel injection. This is due to the complex dependence of valve movement on the current operating point.

[0014] To monitor valve movement, a piezoelectric sensor is provided in the injection valve. This piezoelectric sensor is designed as a pressure sensor to detect pressure changes in fuel triggered by actuating the injection valve and provide a corresponding sensor signal. The measured sensor signal can then be evaluated to determine the opening and closing times of the injection valve, allowing for corresponding adjustments to the valve's operation.

[0015] An evaluation point time series is obtained by sampling the sensor signal at a pre-defined sampling rate, wherein the evaluation point time series is determined for a pre-defined evaluation period. The remaining sampled values ​​are not part of the evaluation point time series.

[0016] However, sensor signals are also affected by noise and are particularly dependent on the actual fuel pressure during fuel supply and the duration of the operation being measured.

[0017] By using data-driven time-point determination models, sensor signals can be evaluated to determine the opening or closing time of the injection valve. Data-driven time-point determination models can correspond to neural networks, probabilistic regression models, or other data-based models.

[0018] To train this time-point determination model, evaluation point time series, each equipped with different labels, can be pre-defined to form a training dataset. The labels correspond to information about the assigned opening or closing time points of the injection valve. This training data is obtained through measurements on a test bench, where different opening durations are pre-defined for different actuation time points of the injection valve, the evaluation point time series are recorded, and the corresponding actual opening or closing time points are determined using suitable test bench sensors.

[0019] It can also provide unlabeled evaluation point time series, which depict the actual sensor signal changes determined by the piezoelectric sensor, and for these unlabeled evaluation point time series, no on or off time points have been determined.

[0020] Based on these unlabeled evaluation point time series, a time point determination model can now be trained using a consistency loss function. To this end, two shift functions are provided: a first shift function to shift the evaluation point time series over time; and a second shift function to shift the modeled time point information over time.

[0021] The first shift function shifts the index of the evaluation point time series forward or backward by an integer shift value, and, if necessary, fills the empty evaluation points at the beginning or end with the corresponding sampled values ​​of the sensor signal. The second shift function shifts the time point information accordingly. Thus, the shift value of the time point information can be determined by the second shift function according to the time interval predetermined by the sampling rate of two consecutive evaluation points and the integer shift value.

[0022] The consistency loss function can evaluate the difference between modeled first-time-point information and modeled second-time-point information, for example, in the form of a consistency loss value, which represents the difference, L2 norm, or other suitable loss value that evaluates the distance between the first-time-point information and the second-time-point information.

[0023] The predictive loss function can evaluate the difference between time-point information modeled using the evaluation point time series and time-point information assigned to the evaluation point time series that has been determined by bench measurements as markers, for example, in the form of a predictive loss value, which represents the difference, L2 norm, or other suitable loss value that evaluates the distance between the modeled time-point information and the assigned time-point information.

[0024] Therefore, training a data-driven time-point determination model can be based on a training dataset (labeled evaluation point time series) and unlabeled evaluation point time series.

[0025] Unlabeled evaluation point time series can also be determined outside the testbed. In this way, the data-based time point determination model can be determined using a training dataset from testbed measurements, and subsequently, for example, retrained on an operating internal combustion engine / vehicle based on the evaluation point time series determined there, according to a consistency loss function. This training, based on the unlabeled evaluation point time series, can be performed to fine-tune the time point determination model previously trained with the training dataset. This can be achieved using a pre-given number of tuning steps and a reduced learning rate relative to training using the training dataset, i.e., a faster tuning speed for the model parameters of the data-based time point determination model.

[0026] Furthermore, to train a data-driven time-point determination model, a total loss function can be used, which depends on both the consistency loss function and the prediction loss function, wherein, in particular, the evaluation point time series of at least one training dataset is used or assumed to be at least one unlabeled evaluation point time series. Therefore, the training is based on the same evaluation point time series for both prediction and consistency loss values.

[0027] The training dataset and the evaluation point time series can be used separately to train one of the two independent neural networks of the Siamese network.

[0028] Training of a data-driven time-point determination model can be continued or interrupted based on a quality value. Specifically, a quality value that indicates the quality of the time-point determination model can be determined from multiple consistency loss values ​​(evaluations of the consistency loss function) of the trained time-point determination model, based on the evaluation point time series and multiple first and corresponding second shift functions with different shift values. The quality value can be formed by multiple consistency loss values ​​(based on unlabeled evaluation point time series) with respect to multiple time-shifted evaluation point time series, using different first and associated second shift functions, such as the sum of consistency loss values ​​or the maximum value of the consistency loss values.

[0029] If a quality value is found to exceed a pre-defined threshold, the on or off time point is remeasured once or multiple times to create one or more training datasets.

[0030] According to another aspect, a method is provided for determining the opening or closing time of an injection valve to operate the injection valve by means of a data-based time point determination model trained according to the above method, wherein the operation of the injection valve is performed according to the opening or closing time point, wherein the operation of the injection valve is particularly performed such that the opening duration of the injection valve determined by the determined opening and / or closing time point is set to a predetermined target opening duration.

[0031] According to another aspect, an apparatus for performing the above-described method is provided. Attached Figure Description

[0032] The embodiments are then described in more detail with reference to the accompanying drawings. Wherein:

[0033] Figure 1 A schematic diagram of an injection system for injecting fuel into the cylinders of an internal combustion engine is shown.

[0034] Figure 2 A flowchart is shown illustrating a method for training a data-based time-point determination model to determine the opening or closing time of an injection valve;

[0035] Figure 3 A block diagram illustrating the determination of consistency loss values ​​based on unlabeled evaluation point time series is shown;

[0036] Figure 4 An exemplary illustration of the time series at the evaluation points is shown. Detailed Implementation

[0037] Figure 1 The arrangement of the injection system 1 of an internal combustion engine 2 in a motor vehicle is shown, with cylinders 3 (especially multiple cylinders) exemplarily shown for this internal combustion engine. The internal combustion engine 2 is preferably designed as a direct-injection diesel engine, but it can also be provided as a gasoline engine.

[0038] Cylinder 3 has an intake valve 4 and an exhaust valve 5 for supplying fresh air and for discharging combustion exhaust gases.

[0039] Furthermore, fuel for running the internal combustion engine 2 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 supply line, which is supplied at high fuel pressure in a manner known in itself (e.g., common rail).

[0040] The injection valve 6 has an actuator unit 61 that can be operated electromagnetically or piezoelectrically, and this actuator unit is coupled to a valve needle 62. In the closed state of the injection valve 6, the valve needle 62 is located on a needle seat 63. By operating the actuator unit 61, the valve needle 62 moves longitudinally and releases a portion of the valve orifice in the needle seat 63 to inject pressurized fuel into the combustion chamber 7 of the cylinder 3.

[0041] The injection valve 6 also has a piezoelectric sensor 65 arranged within the injection valve 6. The piezoelectric sensor 65 deforms due to pressure changes in the fuel guided through the injection valve 6 and generates a voltage signal as a sensor signal.

[0042] The injection is performed under the control of a control unit 10, which specifies the amount of fuel to be injected by energizing the actuator unit 61. Sensor signals are sampled in time by means of an A / D converter 11 in the control unit 10, specifically at a sampling rate of 0.5 to 5 MHz.

[0043] When the internal combustion engine 2 is running, sensor signals are used to determine the correct opening and / or closing times of the injection valve 6. For this purpose, the sensor signals are digitized into a sensor signal time series by means of an A / D converter 11 and evaluated using a suitable evaluation model. Based on this model, the opening duration of the injection valve 6 can be determined according to fuel pressure and other operating parameters, and the amount of fuel injected can be determined accordingly. To determine this opening duration, in particular, the opening and closing times are needed so that the opening duration can be determined as the time difference between these parameters.

[0044] The turn-on and / or turn-off times can be determined by considering the changes in the sensor signal. In particular, turn-on or turn-off times can be determined using data-based time-point determination models.

[0045] To train a data-based time-point determination model, in addition to the training dataset obtained from measurements, unlabeled evaluation point time series can also be used.

[0046] Figure 2 The flowchart illustrates the method for training this data-based timing determination model, which, when run in the aforementioned engine system 1, is used to determine the opening or closing timing of the injection valve 6 of cylinder 3. This method can be executed in the controller of a test bench. This test bench is capable of measuring the injection valves in the engine system, allowing the corresponding opening and / or closing timings to be accurately detected by means of additional sensing devices based on the control signals for actuator unit 61. For this purpose, the sensor signal of piezoelectric sensor 65 is sampled, and the corresponding sampled evaluation point time series is detected at a resolution, for example, between 5 and 20 µs.

[0047] 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 supplied fuel.

[0048] In step S2, the sensor signal is sampled using an A / D converter 11 to determine the evaluation point time sequence within the evaluation period. This evaluation period can be defined relative to the control time window of the injection valve. The control time window is defined by the start of control of the actuator unit 61 and the specified duration, which is the maximum duration for which the control signal of the actuator unit 61 specifies the valve opening. Therefore, the control time window has a defined time reference for which an evaluation point time sequence is provided, which forms the basis for further determining the opening or closing time points. In particular, the evaluation point time sequence can be determined by downsampling previously oversampled sensor signals.

[0049] An evaluation period can be provided using a fixed time reference relative to the power stroke of the internal combustion engine 2, which preferably begins at a predetermined crankshaft position within the compression stroke. The evaluation period can be selected such that the entire opening time window of the injection valve 6 can be mapped within it. Figure 4 The figure shows such an evaluation period with an exemplary evaluation point time series.

[0050] In step S3, the actual time point is determined as the on or off time point of the evaluation point time series according to the test bench sensor device. This time point is considered as a marker of the previously determined relevant evaluation point time series, thus forming a training dataset.

[0051] In step S4, the next evaluation point time series can be detected without identifying the corresponding time points as on or off points. Therefore, for this training, a training dataset and unlabeled evaluation point time series are provided.

[0052] In step S5, a data-based time-point determination model is trained based on these training datasets and these unlabeled evaluation point time series. The evaluation point time series on which the training dataset is based can also be provided as unlabeled evaluation point time series.

[0053] This training can be performed in a manner known to itself based on a loss function that maps the input vector to the model quality. A common training method for this is backpropagation.

[0054] Training can be performed alternately on the training dataset and unlabeled evaluation point time series. To this end, the prediction loss can be determined based on the training dataset in the form of prediction quality. This prediction quality can depend on the time-point information of the modeled data and the time-point information of the evaluation point time series assigned to the training dataset. Specifically, the model quality can correspond to the difference between the modeled time points and the assigned time-point information, L2 norm, or other distance-determined results.

[0055] Furthermore, the consistency loss value can be determined based on the evaluation point time series and according to pre-given first and second shift functions. This is then combined with... Figure 3 The process is described using a block diagram.

[0056] To this end, a first shift function Φ(x) is applied to a pre-given unlabeled evaluation point time series X, shifting the evaluation point time series by one or more element positions according to an integer shift value. This can be done, for example, by changing the index values ​​of the elements of the evaluation point time series by integer shift values ​​and then supplementing the missing index values ​​at the beginning or end of the evaluation point time series with the corresponding sampled values ​​of the sensor signals. For example, the evaluation point time series can be shifted n evaluation points / samples forward in time, such that the first n elements of the evaluation point time series must be set to arbitrary new values. Here, the last n evaluation points of the original evaluation point time series X are discarded. The resulting time-shifted evaluation point time series corresponds to X'=Φ(X).

[0057] The time series X' of the evaluation points shifted over time is fed to the time point determination model f in order to obtain the first comparison value f(Φ(X).

[0058] The modeled time-point information Y=f(X) is generated by determining the model f according to the time points from the original evaluation point time series X.

[0059] Define a second shift function Ψ(Y) that shifts the modeled time point information by a corresponding duration, derived from the shift value n of the first shift function Φ(X). Assuming that the evaluation points of the evaluation time series are uniformly spaced in time, for example, by time increments of Δt, the second shift function Ψ(Y) shifts the desired time point by +nxΔt, i.e., it applies an offset of nxΔt to the time point information.

[0060] The second comparison value Ψ(f(X)) is determined from the modeled time point information Y according to the second shift function.

[0061] Now, the consistency loss value L_kon is derived from the deviation between the modeled time point Ψ(f(x)) that was moved and the model value of the time series f(Φ(x)) that was moved for the evaluation point, according to L_kon=L(f(Φ(x)), Ψ(f(x))).

[0062] The consistency loss value L_kon can now be used, for example, to train a data-based time-point determination model using the backpropagation algorithm. During initial training on the testbed to obtain a training dataset with labeled evaluation point time series, the data-based time-point determination model can also be retrained based on unlabeled evaluation point time series, i.e., solely based on the consistency loss function, during the use of injection valves in an engine system. Thus, the training dataset obtained on the testbed can be used to pre-train the data-based time-point determination model, which is then retrained based on unlabeled evaluation point time series.

[0063] In step S6, it can be checked whether the time-point determination model based on the data is accurate enough. This can be found, for example, by determining corresponding consistency loss values ​​based on the evaluation point time series for multiple first and second shift functions that implement different time offsets for unlabeled evaluation point time series.

[0064] Now, based on the determined consistency loss values, a quality value can be created, for example, as the sum of the determined consistency loss values ​​or as the maximum value of the determined consistency loss values, and a threshold comparison can be used to check whether the consistency of the data-based time-point determination model is sufficient to terminate the training. If so (option: yes), the data-based time-point determination model is implemented into the control unit 10 of the engine system in step S7; otherwise, the process jumps back to step S1 and one or more broader training datasets are determined on a test bench.

[0065] The quality value can be determined by determining multiple consistency loss values ​​for the model at the time points after training. These consistency loss values ​​are derived from the consistency loss function based on at least one unlabeled evaluation point time series and multiple first and corresponding second shift functions with different shift values ​​n.

[0066] The following describes another possibility for training the time-point determination model. In this case, the time-point determination model is designed as a classification model to output Logit as time-point information. For this training, a training dataset is artificially generated by assigning random labels—that is, random specific time-point information—to the evaluation point time series as possible opening or closing times of the injection valve. This training dataset is then used to train the time-point determination model with a few steps that do not lead to model convergence. Subsequently, the labels for the training data are redefined using those output by the previously determined function. This process is iterated until the labels generated in this way converge at the time points to be trained. Unlabeled and labeled evaluation point time series are used together in this process. The evaluation point time series of the (labeled) training dataset can then be evaluated to see how well the assigned opening or closing times are predicted. If the trained time-point determination model makes sufficiently good predictions, the learned classes are also assigned opening and closing times from the training dataset. Another portion of the unlabeled evaluation point time series is used to determine the consistency loss value and is used to further train the time-point determination model.

[0067] In another implementation, the data-driven time-point determination model can also be designed as a Siamese network, wherein a first sub-network of the Siamese network is trained using a training dataset based on a prediction loss function, and a second sub-network of the Siamese network is trained using unlabeled evaluation point time series based on a consistency loss function. Once the Siamese network is trained, the parameters of the two sub-networks are used together through weight sharing.

Claims

1. A computer-implemented method for training a data-based time-point determination model based on sensor signals to determine the opening or closing time of an injection valve (6), the method comprising the following steps: - By sampling the sensor signal of the sensor of the injection valve (6), at least one unlabeled evaluation point time series (S1, S2) is provided; - Train the data-based time point determination model (S5) to assign time point information (Y) to the evaluation point time series (X), wherein the time point information represents the determined on or off duration. The training is performed using a first shift function (Φ(X)) to shift the evaluation point time series (X) over time and a second shift function (Ψ(Y)) to shift the time point information (Y) over time. The training process uses a consistency loss function that describes the difference between information at a first time point and information at a second time point. The first time point information corresponds to the model output of the time-determining model for the evaluation point time series (X') shifted over time according to the first shift function (Φ(X)). The second time point information describes the time point information shifted over time according to the second shift function (Ψ(Y)), and the time point information corresponds to the model output of the time-determining model for at least one evaluation point time series (X).

2. The method of claim 1, wherein the data-based time point determination model is trained based on a prediction loss function and using at least one training dataset, wherein the at least one training dataset assigns the evaluation point time series to specific time point information as labels.

3. The method of claim 2, wherein, in order to train the data-based time-point determination model, a total loss function is used, the total loss function depending on the consistency loss function and the prediction loss function.

4. The method according to claim 3, wherein the at least one evaluation point time series (X) corresponds to the evaluation point time series (X) of the training dataset.

5. The method of claim 2, wherein the data-based time point determination model is trained alternately using the unlabeled evaluation point time series (X) according to the consistency loss function and using the at least one training dataset according to the prediction loss function.

6. The method of claim 2, wherein the data-based time point determination model is pre-trained using the at least one training dataset according to the consistency loss function and retrained using the unlabeled evaluation point time series (X) according to the consistency loss function.

7. The method according to any one of claims 2 to 6, wherein the at least one unlabeled evaluation point time series corresponds to at least one evaluation point time series of the training dataset.

8. The method according to any one of claims 2 to 6, wherein training of the data-based time-point determination model continues or is interrupted based on a quality value indicating the quality of the time-point determination model, wherein a plurality of consistency loss values ​​of the trained time-point determination model are determined, the consistency loss values ​​being derived from the consistency function based on at least one unlabeled evaluation point time series and a plurality of first and corresponding second shift functions with different shift values, wherein in particular the quality value corresponds to the sum of the consistency loss values ​​or the maximum value of the consistency loss values.

9. A method for determining the opening or closing time of an injection valve (6) to operate the injection valve by means of a sensor signal (S) and a data-based time point determination model trained according to any one of claims 1 to 8, wherein operation of the injection valve (6) is performed according to the opening or closing time point, wherein 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 time point is set to a predetermined target opening duration.

10. An apparatus for performing one of the methods according to any one of claims 1 to 9.

11. A computer program product 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 9.

12. 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 9.

Images