Method and apparatus for creating an emission model of an internal combustion engine

Abstract

The invention relates to a method (10) for creating an emission model of an internal combustion engine. The method (10) starts with providing (S11) a plurality of measurement sequences on an internal combustion engine. The measurement sequences are then filtered (S12) with different low-pass filters and from the filtered measurement sequences a measurement sequence is derived which, as an input variable for an emission model, achieves a minimum deviation of the predicted emissions of the emission model from the measured emissions when optimizing the emission model.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for creating an emission model of an internal combustion engine using at least one low-pass filter, and to a device, a computer program and a machine-readable storage medium. BACKGROUND

[0002] Some models are known which are used to predict emissions at the outlet of an exhaust tract. These models are usually designed in such a way that they predict the emissions of an internal combustion engine, in particular, and from this the emissions at the outlet of the exhaust tract are then predicted by means of a model of the exhaust tract.

[0003] These models for predicting the emissions of a motor are usually generated in such a way that measurements are carried out at different predefined operating points of the motor, and from these measurements a model for predicting the emissions is then calculated. This can be carried out, for example, with the software ETAS ASCMO.

[0004] It is furthermore known that a model for predicting the emissions can be set up in such a way that it can take into account a plurality of previous measurements from the past, in particular a measurement sequence in the form of a time series, so that this model can predict the dynamic behaviour of the emission generation of the motor more precisely. For this purpose a so-called feature matrix is used. This feature matrix contains information about which measurement points of the measurement sequence should be used for the model prediction. The number of columns of the feature matrix corresponds to the number of previous measurement points or time steps considered, while the number of rows corresponds to the number of different inputs in the model, for example the measured measurement variables of the motor. In a processing step which can be called "automatic feature selection" when creating the model, the relevant features of the feature matrix are extracted by training and evaluating the model for each combination of features with the Brute-Force method. SUMMARY

[0005] The feature matrix mentioned above becomes very large if a high sampling frequency and a long history are required for the model. If the measurements are recorded with a high resolution, for example with a sampling frequency of 100 ms, and a maximum time delay is adjusted to 3 minutes (= 180 seconds), then a column number of 1800 results therefrom. A great deal of computing time and computing power must be applied in order to test all possible combinations (1800 columns times the number of different measurement variables of the feature matrix) in the model training and in order to achieve a sufficient model prediction quality.

[0006] Furthermore, in the method hitherto, the exact measurement value at the time observed in the past was taken into account. This is often not necessary. And the noise of the model is increased. In the past, it was often sufficient to take into account the order of magnitude of the measurement value.

[0007] According to the present invention, it is proposed to use filters to aggregate past information within a measurement sequence. From unfiltered and optionally filtered measurement sequences, "automatic feature selection" can determine the optimal combination of features in terms of best model quality.

[0008] This method significantly improves both model quality and robustness. The computation time for automatic feature selection is also significantly reduced.

[0009] Therefore, a highly accurate motor emission model is provided, which achieves particularly high accuracy across the entire operating range of the motor, even under unstable conditions. Furthermore, through filtering, the model becomes more robust in its predictions, as the emission model is highly sensitive to past observations.

[0010] In a first aspect, the present invention relates to a computer-implemented method for creating an emission model of an internal combustion engine, comprising the following steps:

[0011] The method begins by providing multiple detected, particularly unfiltered, measurement sequences from an internal combustion engine. The measurement sequences are then filtered using different low-pass filters, and multiple emission models for different combinations of the filtered measurement sequences are optimized to minimize the deviation of the predicted emissions from the emission models relative to the total detected emissions from the detected measurement sequences, where the filtered measurement sequences serve as input parameters for the emission models.

[0012] Then, select the following emission models from a pool of emission models that achieve the minimum deviation of the predicted emissions from the measured emissions. This is followed by selecting those measurement sequences from the filtered measurement sequences if the emission model achieves the minimum deviation of the predicted emissions from the measured emissions when used as input parameters for optimizing the emission model.

[0013] The optimized emission model with minimal deviation is then output, along with a filtered measurement sequence preferably used thereto and a corresponding low-pass filter, which has been used to filter the selected filtered measurement sequence.

[0014] The inventors have discovered that the precise location of a measurement within a measurement sequence is not decisive; rather, it is whether the measured parameter already has a high or low value within the last second. Therefore, a "fuzzy memory" of previous measurements is sufficient. Surprisingly, this has also led to emissions models achieving particularly good predictions even in transition regions.

[0015] A measurement sequence is a series of measurement points, in particular sensor values detected at a predefined time in a series of times.

[0016] It is proposed that the different low-pass filters differ in that they have different time constants, for example 1 s, 3 s, 5 s, 10 s, 30 s. The time constant defines how long in the past observations are observed and taken into account, in particular weighted, when filtering.

[0017] It is further proposed that the filter is a first-order low-pass filter. In this regard it is advantageous that the information from the measurement sequence is now centrally present in the filtered signal by aggregation, and that it is now no longer necessary to find the ideal combination in all previous times, but only a suitable combination of the filtered signal has to be found.

[0018] It is further proposed that each detected measurement sequence is assigned to a measurement variable, wherein the measurement variable characterizes a variable of the internal combustion engine, in particular a motor rotational speed, a relative cylinder charge, lambda, an ignition angle, a cylinder ignition information, wherein the predicted emissions are particulate emissions, in particular mass and / or quantity, or gaseous emissions, in particular NOx, THC, CO and / or a temperature / pressure of the emissions.

[0019] It is further proposed that, when the low-pass filter is applied to the detected first measurement value and no previous measurement value is available, for example at start-up of the motor, an additional virtual measurement value is provided as an input variable for the low-pass filter, wherein the additional virtual measurement value characterizes a constant operating point of the motor at low load. Low load is 5 to 40%, preferably 20 to 35%, relative to full load. It is particularly preferred that, for the constant operating point, a rotational speed in the vicinity of the idling rotational speed is used as a virtual measurement value, for example 1000-1900 revolutions / minute. The virtual measurement value is thus an artificially generated measurement value.

[0020] In other aspects, the invention relates to a device and a computer program, respectively, which are set up to carry out the above-described method, and to a machine-readable storage medium on which a computer program is stored. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flow chart schematically showing an embodiment of the invention;

[0022] Figure 1 A flow chart schematically showing a method (10) for creating an emissions model of an internal combustion engine. DETAILED DESCRIPTION

[0023] The method (10) starts with a step S11. In this step, a measurement sequence of very different measurement variables is detected on the internal combustion engine, which characterize the internal combustion engine, preferably the state of the internal combustion engine. One of the measurement variables can be, for example, the motor speed. The emissions of the internal combustion engine, which are related to the respectively detected measurement variables, are also detected as measurement sequences. These detected data are used as training data below.

[0024] Step S12 is then carried out. This step can be referred to as "filtering". The raw data from the measurement sequences in step S11 are processed here. In addition to other calculations, the raw data are filtered by means of a plurality of different low-pass filters. In addition to the conventional inputs, the outputs of the different low-pass filters are then provided to the optimization software. The optimization software is preferably ETAS ASCMO. It can be considered to add the different low-pass filters to the feature matrix in the form of further columns.

[0025] It is preferred that the low-pass filters differ in that they filter different numbers of measurement points within the measurement sequences. It is particularly preferred that the low-pass filters are first-order low-pass filters.

[0026] It is preferred that the low-pass filters are described by a mathematical formula:

[0027] .

[0028] In a step S13 which follows immediately, the different low-pass filters are then optimized with the optimization software in combination with the conventional inputs, which best combination results in the following: the emissions model predicts the detected emissions from step S11 of the same belonging internal combustion engine as precisely as possible. To this end, the optimization software can, for example, experimentally test which combination of different low-pass filters results in the best emissions model. This can be implemented, for example, by selecting a combination of different low-pass filters and optimizing the emissions model with the combination as input variable for the model, so that the deviation between the predicted emissions of the emissions model and the detected emissions from step S11 is minimized.

[0029] It is preferred that the emissions model is a GP-NARX, which is trained, for example, by means of supervised learning. Other machine learning systems and other model types can also be considered. It has been found that neural networks, in particular RNNs, are also particularly suitable.

[0030] It should be noted that the selection of different low-pass filters and the subsequent optimization of the emissions model are carried out several times for respectively different combinations of low-pass filters. Thus, a plurality of emissions models with respectively different input variables are optimized.

[0031] In another embodiment, the time constant of the low-pass filter can additionally be optimized in step S13 by means of, for example, a genetic algorithm.

[0032] Subsequently, the following emission model is selected from the plurality of optimized emission models, which emission model achieves the smallest deviation between the predicted emissions of the emission model and the detected emissions from step Sll.

[0033] After step S13, an optional step S14 can be performed. Here, the selected emission model can be tested with validation data. The validation data are further measurement sequences which are not included in the training data and thus serve to test the emission model whether it has learned the training data correctly. If step S14 shows that the selected emission model has not been trained correctly, step S13 can be performed again. It can be considered that such a test with validation data is performed in step S13 after a predefined number of optimization steps.

[0034] Subsequently in step S15, this selected emission model with the smallest deviation is output as the most precise emission model. Preferably, in step S15 also the associated measurement variable is output, the measurement sequence of which is considered as input variable for this emission model. Preferably, also the associated low-pass filter is output in step S15, the filtered measurement sequence of which is used as input variable. Preferably also the associated feature matrix is.

[0035] After performing step S15, the method (10) ends. Subsequently, the outputted emission model from step S15 can be used for different applications. For example, this emission model can then be used for predicting the emissions of the internal combustion engine. It can also be considered that this emission model is used for predicting the emissions at the outlet of the exhaust tract connected to the internal combustion engine. Here, for example, another model can be considered with which the predicted emissions of the motor are then further processed by taking into account the physical effects of the exhaust system in such a way that the emissions at the outlet of the exhaust system are then calculated.

[0036] When using the emission model for predicting the emissions, the measurement sequence assigned to the associated outputted measurement variable from step S15 and filtered with the outputted low-pass filter from step S15 can be detected and then used as input variable for the emission model.

Claims

1. A method (10) for creating an emission model of an internal combustion engine, comprising the following steps: Multiple detected measurement sequences are provided on the internal combustion engine (S11); The measurement sequence is filtered using different low-pass filters (S12). The emission models for multiple different combinations of filtered measurement sequences are optimized (S13) to minimize the deviation of the emission predicted by the emission model from the detected emission of the corresponding detected emission from the detected measurement sequence, wherein the filtered measurement sequence is used as the input parameter for the emission model. The following emission models are selected from the plurality of emission models, wherein the emission models obtain the minimum deviation of the predicted emissions of the emission models relative to the measured emissions; and Output (S15) the optimized emission model with the smallest deviation.

2. The method of claim 1, wherein the difference between the different low-pass filters is that they have different time constants, wherein the time constants characterize which prior measurements are taken into account by the respective low-pass filters.

3. The method according to any one of the preceding claims, wherein the different low-pass filters are first-order low-pass filters.

4. The method according to claim 1 or 2, wherein when optimizing (S13) the emission model, at least the nearest measurement points of a plurality of detected measurement sequences are additionally considered as input parameters / combinations of said input parameters.

5. The method according to claim 1 or 2, wherein each detected measurement sequence is associated with a measurement parameter, wherein the measurement parameter characterizes a parameter of the internal combustion engine. The predicted emissions are particulate emissions or gaseous emissions and / or the temperature / pressure of the emissions.

6. The method according to claim 5, wherein, The measured parameters represent the motor speed, relative cylinder charge volume, λ, ignition angle, and cylinder ignition information.

7. The method according to claim 5, wherein, The particulate emissions include their mass and / or quantity.

8. The method according to claim 5, wherein, The gaseous emissions include NOx, THC, and CO.

9. The method according to claim 3, wherein, If the low-pass filter is applied to the first measurement, then an additional virtual measurement is provided as an input parameter for the low-pass filter, wherein the additional virtual measurement represents the constant operating point of the motor at low load.

10. An emission model for predicting emissions from the internal combustion engine or at the outlet of an exhaust manifold connected to the internal combustion engine, the emission model being created using the method according to any one of claims 1 to 9, wherein the measurement sequence is filtered with a selected low-pass filter, and the emission model predicts emissions based on selected input parameters.

11. An apparatus for performing the method according to any one of claims 1 to 9.

12. A computer program product configured to perform the method according to any one of claims 1 to 9.

13. A machine-readable storage medium on which the computer program product according to claim 12 is stored.

Images