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

Abstract

Method (10) for creating an emission model of an internal combustion engine, comprising the following steps: Providing (S11) a plurality of recorded measurement series on an internal combustion engine; Filtering (S12) the measurement series with different low-pass filters; Optimizing (S13) a plurality of emission models for different combinations of filtered measurement series as inputs for the emission model, such that a deviation of the predicted emissions of the emission model to the corresponding recorded emissions is minimized; Selecting the emission model from the majority of emission models that achieves the smallest deviation between the emission model's predicted emissions and the measured emissions; and Output (S15) of the optimized emission model with the smallest deviation.

Description

[0001] The invention relates to a method for creating an emission model of an internal combustion engine using at least one low-pass filter, the use of the model, as well as a device, a computer program and a machine-readable storage medium. State of the art

[0002] Models are known that are used to predict emissions at the exhaust outlet. These models are typically designed to predict the emissions of an engine, particularly a combustion engine, and then, based on this, to predict emissions at the exhaust outlet using a model of the exhaust system.

[0003] These models for predicting engine emissions are typically generated by taking measurements at different, predefined engine operating points. The emissions prediction model is then calculated based on these measurements. This can be done, for example, with the ETAS ASCMO software.

[0004] Furthermore, it is known that the emissions prediction model can be configured to consider multiple previous measurements, particularly time series, enabling it to predict the dynamic behavior of the engine's emissions more accurately. This is achieved using a feature matrix. This matrix contains information on which measurement points from the data series should be used for the model prediction. The number of columns in this 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 to the model, such as measured engine parameters.In a processing step during model creation, which can be called "automatic feature selection", the relevant features of the feature matrix are extracted by training and evaluating the model for each combination of features using a brute-force approach.

[0005] DE 10 2008 057 494 A1 discloses a method for determining an emission model of an internal combustion engine using principal component analysis, wherein the internal combustion engine is subjected to a test operation, whereby at least one parameter of the internal combustion engine is varied.

[0006] US 2010 / 0251699 A1 discloses a control system for providing an optimal estimate of NOx emissions in the exhaust gas during a selective catalytic reduction process. The control system includes a continuous emission monitoring sensor configured to generate a response signal that represents a first estimate of the NOx emission, wherein the response signal has a first time delay between a time of NOx emission measurement and the time at which the corresponding response signal is provided from the sensor to the continuous emission monitoring sensor, and the continuous emission monitoring sensor has a first time constant. Advantages of the invention

[0007] The feature matrix mentioned above becomes very large if the model requires a high sampling rate and a long history. If the measurements are recorded with high resolution, e.g., with a sampling rate of 100 ms, and the maximum time delay is set to, for example, 3 minutes (=180 seconds), this results in 1800 columns. Consequently, a significant amount of computing time and power is required to test all possible combinations (1800 columns multiplied by the number of different measurements in the feature matrix) during model training and to achieve sufficient model prediction accuracy.

[0008] Furthermore, the current approach considers the exact measurement value at the time in question in the past. This is often unnecessary and increases the noise in the models. In most cases, it is sufficient to consider the order of magnitude of the measurement value in the past.

[0009] These problems are solved by the features of the independent patent claims. Advantageous embodiments of the invention are the subject of the dependent claims.

[0010] According to the independent claim, the use of a filter to aggregate past information within the measurement series is proposed. From the unfiltered and, optionally, the filtered measurement series, the "Automatic Feature Selection" can determine the best combination of features with regard to achieving the best model quality.

[0011] Both model quality and model robustness were significantly improved by this approach. Additionally, the computation time for automatic feature selection is drastically reduced.

[0012] Consequently, a highly accurate engine emissions model is provided, achieving exceptionally high accuracy across the entire engine operating range, even under transient conditions. Furthermore, the filtering process makes the model more robust in its predictions, as the previous emissions model was highly sensitive due to its reliance on historical data. Disclosure of the invention

[0013] In a first aspect, the invention relates to a computer-implemented method for creating an emission model of an internal combustion engine, comprising the following steps: The process begins with providing a plurality of recorded, particularly unfiltered, measurement series from an internal combustion engine. This is followed by filtering the measurement series with different low-pass filters and optimizing a plurality of emission models for different combinations of filtered measurement series as input variables for the emission model, such that the deviation of the emission model's predicted emissions from the corresponding recorded emissions is minimized.

[0014] The next step involves selecting the emission model from the majority of models that achieves the smallest deviation between the model's predicted emissions and the measured emissions. This can then be followed by a step that selects, from the filtered measurement series, those series that, when used as input for the emission model during optimization, result in the smallest deviation between the model's predicted emissions and the measured emissions.

[0015] This is followed by the output of the optimized emission model with the smallest deviation and preferably the filtered measurement series used for this purpose and the respective associated low-pass filters that were used to filter the selected filtered measurement series.

[0016] The inventors discovered that it is not the exact position of a measurement within a series that is crucial, but rather whether a measured quantity has exhibited a high or low value within the last few seconds. A vague recollection of previous measurements is therefore sufficient. Surprisingly, this has also led to the emission model achieving particularly good predictions even in transitional regions.

[0017] A measurement series is a sequence of measurement points, in particular sensor values, which were each recorded at a predetermined time within a sequence of times.

[0018] It is suggested that the different low-pass filters differ in that they have different time constants, e.g., 1s, 3s, 5s, 10s, 30s. The time constants define how far back in the past the signal is considered and, in particular, weighted during the filtering process.

[0019] Furthermore, it is proposed that first-order low-pass filters be used. The advantage of this is that, by aggregating the information from the measurement series and concentrating it in the filtered signals, it is no longer necessary to select an ideal combination at all previous time points; instead, only a suitable combination from the filtered signals needs to be found.

[0020] Furthermore, it is proposed that each recorded measurement series is assigned to a measured quantity, wherein the measured quantity characterizes a quantity of the internal combustion engine, in particular an engine speed, relative cylinder air filling, lambda, ignition angle, cylinder ignition information, wherein the predicted emission is a particulate emission, in particular mass and / or number, or a gaseous emission, in particular NOx, THC, CO, and / or temperature / pressure of the emissions.

[0021] Furthermore, it is proposed that when the low-pass filters are applied to the first acquired measurements and no previous measurements are yet available, e.g., when the engine is started, additional fictitious measurements are provided as input for the low-pass filters. These additional fictitious measurements characterize a constant operating point of the engine at low load. The low load is 5 to 40% relative to full load, preferably 20 to 35%. A speed close to the idle speed is particularly preferred for the constant operating point, e.g., 1000–1900 rpm. Fictitious measurements are therefore artificially generated measurements.

[0022] In further aspects, the invention relates to a device and a computer program, each configured to perform the above methods, and a machine-readable storage medium on which this computer program is stored.

[0023] It shows: Fig. 1 schematically a flowchart of an embodiment of the invention;

[0024] Fig. Figure 1 schematically shows a flowchart of a procedure (10) for creating an emission model of an internal combustion engine.

[0025] The method (10) begins with step S11. In this step, measurement series of various parameters characterizing the internal combustion engine, preferably a state of the internal combustion engine, are recorded. One of the parameters can be, for example, engine speed. Additionally, the emissions of the internal combustion engine corresponding to each of the recorded parameters are also recorded as a measurement series. These recorded data subsequently serve as training data.

[0026] Step S12 follows. This can be described as "filtering." Here, the raw data from the measurement series in step S11 are processed. Among other calculations, they are filtered using a multiple of different low-pass filters. The outputs of the various low-pass filters are then provided to optimization software in addition to the standard inputs. The optimization software is preferably ETAS ASCMO. It is conceivable that the different low-pass filters are added to the feature matrix as additional rows.

[0027] Preferably, the low-pass filters differ in that they filter a different number of measurement points within the measurement series. First-order low-pass filters are particularly preferred.

[0028] Preferably, the low-pass filter is described by the mathematical formula: outt=outt−1+(in−outt−1)*dT / T

[0029] In the subsequent step S13, the optimization software determines an optimal combination of the different low-pass filters and conventional inputs. This combination ensures that the emissions model predicts the corresponding emissions from the combustion engine, measured in step S11, as accurately as possible. For example, the optimization software can combinatorially test which combinations of the different low-pass filters result in the best emissions model. This can be achieved, for instance, by selecting a combination of different low-pass filters and using these as inputs to optimize the emissions model, thereby minimizing the deviation between the emissions predicted by the emissions model and the emissions measured in step S11.

[0030] Preferably, the emission model is a GP-NARX, which is trained, for example, using supervised learning. Other machine learning systems and model types are also conceivable. Neural networks, especially RNNs, have proven to be particularly suitable.

[0031] It should be noted that the selection of different low-pass filters and subsequent optimization of the emission model is performed multiple times for different combinations of these filters. Thus, a multiple emission models with different input variables are optimized.

[0032] In another embodiment, in step S13, the time constants of the low-pass filters can also be optimized using, for example, a genetic algorithm.

[0033] Subsequently, the emission model that achieves the smallest deviation between the predicted emissions of the emission model and the emissions recorded in step S11 is selected from the majority of optimized emission models.

[0034] After step S13, the optional step S14 can be executed. In this step, the selected emissions model can be tested with validation data. The validation data consists of additional measurement series not included in the training data and is used to test whether the emissions model has learned the training data correctly. If step S14 reveals that the selected emissions model has not been trained correctly, step S13 can be repeated. It is conceivable that this testing with validation data in step S13 could be performed after predefined optimization steps.

[0035] The selected emission model with the smallest deviations is then output as the most accurate emission model in step S15. Preferably, step S15 also outputs the associated measured variables whose corresponding measurement series were used as input variables for this emission model. Preferably, step S15 also outputs the associated low-pass filters whose filtered measurement series were used as input variables. Preferably, an associated feature matrix is ​​also included.

[0036] After step S15 has been executed, procedure (10) is complete. The emission model output from step S15 can then be used for various applications. For example, this emission model can be used to predict emissions from the internal combustion engine. It is also conceivable that this emission model could be used to predict emissions at the outlet of an exhaust system connected to the internal combustion engine. In this case, for example, another model could be used to further process the predicted engine emissions by taking into account physical effects of the exhaust system in order to then calculate the emissions at the exhaust system outlet.

[0037] When using the emission model to predict emissions, measurement series can be recorded which are assigned to the corresponding output measurement quantities from step S15 and are filtered with the output low-pass filters from step S15 and then used as input quantities for the emission model.

Claims

[1] Method (10) for creating an emission model of an internal combustion engine, comprising the following steps: Providing (S11) a plurality of recorded measurement series on an internal combustion engine; Filtering (S12) the measurement series with different low-pass filters; Optimizing (S13) a plurality of emission models for different combinations of filtered measurement series as inputs for the emission model, such that a deviation of the predicted emissions of the emission model to the corresponding recorded emissions is minimized; Selecting the emission model from the majority of emission models that achieves the smallest deviation between the emission model's predicted emissions and the measured emissions; and Output (S15) of the optimized emission model with the smallest deviation. [2] Method according to claim 1, wherein the different low-pass filters differ in that they have different time constants, the time constants characterizing which previous measured values ​​the respective low-pass filter takes into account. [3] Method according to any of the preceding claims, wherein the different low-pass filters are first-order low-pass filters. [4] Method according to one of the preceding claims, wherein when optimizing (S13) the emission models, at least the most recent measurement points of a plurality of the recorded measurement series are additionally taken into account as input variables and / or used for the combination of the input variables. [5] Method according to one of the preceding claims, wherein each recorded measurement series is assigned to a measured quantity, wherein the measured quantity characterizes a quantity of the internal combustion engine, in particular an engine speed, relative cylinder air filling, lambda, ignition angle, cylinder ignition information, wherein the predicted emission is a particulate emission, in particular mass and / or number, or a gaseous emission, in particular NOx, THC, CO, and / or temperature / pressure of the emissions. [6] Method according to claim 3, wherein when the low-pass filters are applied to the first measured values, additional fictitious measured values ​​are provided as input for the low-pass filters, wherein the additional fictitious measured values ​​characterize a constant operating point of the motor at low load. [7] Using the emission model created according to one of the preceding claims to predict the emission of the internal combustion engine or at the outlet of an exhaust connected to the internal combustion engine, wherein the measurement series are filtered with the selected low-pass filters and the emission model predicts the emissions depending on the selected input variables. [8] Device configured to carry out the method according to any of the preceding claims. [9] Computer program configured to execute the method according to any one of claims 1 to 7. [10] Machine-readable storage medium on which the computer program according to claim 9 is stored.

Images