CN114647973A

CN114647973A - Method and control device for estimating smoke quality

Info

Publication number: CN114647973A
Application number: CN202111540180.9A
Authority: CN
Inventors: B·阿尔特
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2020-12-17
Filing date: 2021-12-16
Publication date: 2022-06-21
Also published as: DE102020216145A1

Abstract

A method and control apparatus for estimating soot mass. The invention relates to a method for estimating soot mass during loading of a particle filter in an exhaust gas mass flow of an internal combustion engine and/or soot mass during burn-up by means of a physical model describing a physical interrelation during operation of the internal combustion engine on the basis of measured parameters and/or characteristic data, wherein the physical model has a plurality of submodels, each submodel comprising a physical submodel having as input data measured parameters and/or characteristic data for the respective submodel for determining a respective submodel value, wherein at least one submodel is assigned a trained data-based submodel which is trained to supplement or correct the submodel value of the physical submodel concerned as input data depending on at least a part of the measured parameters and/or characteristic data, wherein at least one physical submodel to which the data-based submodel is applied is corrected by the corresponding data-based submodel The model of (2) is output.

Description

Method and control device for estimating smoke quality

Technical Field

The invention relates to a method for estimating a soot mass during loading of a particle filter in an exhaust gas mass flow of an internal combustion engine and/or a soot mass during burn-up by means of a physical model. The invention also relates to a computer program product and a control device.

Background

All industrialized countries now set standards relating to air pollution control in the form of limit values or target values for the emission of pollutants, in particular in the form of particle limit values for internal combustion engines. These standards also relate to the pollution and consumption of internal combustion engines, particularly motor vehicles and road vehicles, which are equipped with said internal combustion engines.

The measures inside the engine of an internal combustion engine alone no longer enable the strict limits of today's emission legislation. In order to reduce the exhaust gas values to the level required by legislators, particulate filters are used in new vehicles. In this case, the particle filter is usually a wall-flow filter, to which soot and dust are applied during the life cycle of the vehicle, which soot and dust in turn decisively increase the filtration efficiency of the particle filter. Although the dust is continuously retained in the particle filter and thus provides a constant filter fraction, the filter fraction is dependent in particular on the soot particle load due to the soot. Thus, optimal filtration efficiency can be achieved by targeted establishment or elimination of soot particle layers under specific conditions, such as high filter temperatures or excess oxygen.

Empty particulate filters have limited filtration efficiency, and it is therefore important to preload these particulate filters with dust or soot as quickly as possible to increase the filtration efficiency to the desired level. For this purpose, it is expedient to preload the particle filter in a targeted manner before installation or to inject a large amount of soot into the particle filter shortly after the engine has been started. The filtering efficiency increases during driving, since the individual pores in the particle filter are increasingly filled anyway.

In case of sufficiently high temperatures and strong oxygen supply, the soot particle layer can burn off very quickly, which leads to a severe loss of filtration efficiency. It is therefore essential to stop the oxidation of the soot particle layer in time and not to regenerate the particle filter too much and thus to maintain the filtration efficiency at the required level. Over the life of the vehicle, the particle filter becomes increasingly clogged with dust which can no longer be burned off by itself, thus being responsible for the permanently high filtration efficiency. However, the exhaust gas counterpressure increases in the case of an excessively high load, which leads to consumption disadvantages. Therefore, the particle filter should be replaced from a critical dust load.

The pressure difference in the particle filter is an important measurement variable which is calculated from the pressure difference at the inlet and outlet of the particle filter. The change in the flow resistance of the filter can be derived from the change in the pressure difference. However, it is difficult to determine from the corresponding change in the flow resistance what share of the dust or soot is loaded into the particle filter and what the current efficiency of the particle filter is actually.

For this purpose, a physical model for soot formation and dust formation has been developed, which can provide a decision about the dust and soot load in the particle filter on the basis of the measured differential pressure and other measured or influencing variables (for example the mileage of the vehicle, the type of oil used, etc.). In this way, under certain driving conditions, smoke can be more strongly accumulated, smoke can be eliminated, or the exothermic smoke burnup reaction can be stopped.

Various methods for detecting the load in a particle filter in a motor vehicle are known from the prior art, wherein the pressure difference across the particle filter is measured.

EP 2065582 a1 is mentioned as an example, which discloses a method for detecting the load on a particle filter, in particular for filtering the exhaust gas of an internal combustion engine, wherein a pressure drop across the particle filter is determined and a variable characterizing the flow resistance of the particle filter is derived on the basis of the pressure drop and the load on the particle filter is derived from the flow resistance. In this case, the variable characterizing the flow resistance is corrected as a function of a variable characterizing the operating point of the internal combustion engine.

Disclosure of Invention

According to the invention, a method for estimating the soot mass during the loading of a particle filter and/or the soot mass during burn-up in the exhaust gas mass flow of an internal combustion engine by means of a physical model according to claim 1 and a corresponding control device according to the parallel independent claims are provided.

Advantageous embodiments and developments which can be used individually or in combination with one another are the subject matter of the dependent claims.

According to a first aspect, a method is provided for estimating soot mass during loading of a particulate filter in an exhaust gas mass flow of an internal combustion engine and/or soot mass during burn-up by means of a physical model describing a physical interrelation during operation of the internal combustion engine on the basis of measured parameter and/or characteristic data, wherein the physical model comprises various physical sub-models having the parameter and/or characteristic data measured for the respective sub-model as input data for determining the respective sub-model value, wherein at least one sub-model is modeled as a trained data-based sub-model using an artificial intelligence (KI) method, and wherein at least a part of the measured parameter and/or characteristic data is transmitted to the at least one data-based sub-model as input data, and wherein the at least one sub-model to which the data-based is applied is corrected by the corresponding data-based sub-model And model output of the submodels of the model.

The object on which the invention is based is therefore to determine the soot mass in the particle filter and/or the soot mass during burn-up in the exhaust gas mass flow of an internal combustion engine by means of at least one physical model, so that simpler determinations can be made of the dust and soot load and burn-up in the particle filter.

The physical model is also referred to as a soot model, which essentially determines the amount of soot in the particle filter taking into account the pressure drop in the exhaust gas flow using given and known measured parameter/characteristic data and the underlying physicochemical process. Such a soot model preferably comprises various sub-models.

The particle filter calculates unmeasurable variables, load and filtration efficiency, in particular on the basis of measurable or other calculated variables such as exhaust gas mass flow, pressure and temperature upstream of the particle filter, temperature in the particle filter, oxygen concentration in the exhaust gas mass flow and mass flow of the primary emissions. The product of the filter efficiency and the incoming raw exhaust mass flow can then be used to calculate the actual particulate mass flow loading the particulate filter.

The parameter and characteristic data are, for example, measured variables or influencing variables/characteristic variables (for example, vehicle mileage, type of oil used, etc.).

Typically, most parameters are known from the provided data table and can be used directly for parameterization. Certain parameters, such as permeability or activation energy, may be experimentally determined based on selected measurements (e.g., measurements from a single particulate filter).

According to the invention, a method for estimating the soot mass during loading of a particle filter in an exhaust gas mass flow and/or the soot mass during burn-up by means of a trained artificial intelligence is described. The physical model is here divided into various submodels. According to the invention, at least one of these submodels is supplemented by a data-based submodel, and then the corresponding value that the corresponding submodel would return can be modeled as a model value of the data-based submodel.

Thus, according to the invention, the data-based submodel serves as a complement to or can replace the physical submodel so far. In particular, in certain embodiments, a data-based submodel is used as the output error model.

The physical model may be supplemented or replaced by designing sub-models with data-based sub-models.

By means of the data-based submodel, changes in the number of particles during loading and/or during burn-up can be simulated at least in part.

By using the data-based submodel, the model values of the data-based submodel and the obtained data-based submodel may also be continuously improved and more accurate values achieved throughout the life cycle of the vehicle.

The physical basic structure of the physical model may be intentionally retained in whole or in part. However, the physical model or the respective submodel may also be replaced if the data-based submodel has been taught, e.g. by further training, to provide better values.

This results in simpler and, if appropriate, more accurate information about the load on the particle filter of the internal combustion engine (gasoline or diesel), so that measures for improving the filter efficiency, such as adaptations in the fuel system, for example, to make larger/smaller soot particles appear, and/or measures for protecting components, such as overflow shutoff prohibition, i.e., injection to reduce excess oxygen despite the overflow phase, can be initiated in a targeted manner. By the high possible filtration efficiency which can be achieved by means of the invention, it is possible to comply with the limit values required in exhaust gas regulations. Furthermore, fuel consumption and carbon dioxide emissions can be optimized (reduced) by means of the invention.

The invention makes it possible to determine the load on the particle filter and the filter efficiency of an internal combustion engine (gasoline or diesel) more accurately during continuous operation. Measures such as adaptation in the fuel system or the activation of an overflow shutoff prohibition can thereby be carried out early, and in time, the filtration efficiency can be increased or an excessive accumulation of soot particles in the particle filter due to component protection reasons (for example, high temperature loads during combustion) can be counteracted. Whereby increased fuel consumption and thus increased carbon dioxide emissions can be counteracted.

Preferably, a trained artificial neural network is used as the artificial intelligence method or the machine learning method. The trained artificial neural network can be designed as a self-learning network, which is continuously improved. Whereby a continuously improved result is to be expected. Artificial neural networks are particularly suitable for use with a large number of parameters/characteristics.

Furthermore, the artificial neural network may be constructed as a multi-layer neural network. All existing data-based submodels can also be constructed as artificial neural networks. Different artificial intelligence approaches may also be used for the respective data-based sub-models.

Alternatively or additionally, a gaussian process model is used as an artificial intelligence approach. The gaussian process model is particularly suitable for the use presented herein, since it finds application in mathematical modeling of the behavior of non-deterministic systems based on observations.

Preferably, at least one data-based submodel is modeled in addition to the associated physical submodel. The results between the conventional submodel and the data-based submodel can be compared. The results of the conventional physical submodel can be verified or improved. The load on the particle filter and the filter efficiency of the internal combustion engine (gasoline or diesel) can thus be determined more precisely during continuous operation.

Preferably, the at least one data-based submodel is supplemented to at least one physical submodel as an output error model.

Here, the output error model represents a measure of the deviation between the model behavior and the measured behavior of the particle filter. Output errors can be minimized by the data-based submodel, resulting in a good parallel model with the physical submodel and the real particle filter.

Preferably, in case that the KI pressure difference model value corresponding to at least one data-based sub-model deviates from the corresponding measured pressure difference, the corresponding data-based sub-model is adapted such that the deviation is minimized. Thus an artificial neural network or other KI method can continue to learn that the corresponding data-based sub-model provides better and better results. The load on the particle filter and the filter efficiency of the internal combustion engine (gasoline or diesel) can therefore be determined more and more accurately during continuous operation, and measures such as regeneration can be initiated early.

Preferably, the adaptation of the at least one modeled data-based sub-model is performed offline. In particular, the at least one modeled data-based sub-model may be trained and adapted with sufficiently representative measurement data to test the fleet of vehicles during the application phase. Very good modeling results can thus be achieved in advance. Furthermore, external adaptation can be performed on a powerful computer.

In particular, this may be accomplished by additional measurements of the particulate filters of multiple engines within the scope of testing prior to training/adapting the modeled data-based sub-model. This makes it possible to create such a modeled data-based submodel particularly well or simply.

Alternatively or additionally, the at least one modeled data-based submodel is adapted online in a control device of the respective vehicle. This makes it possible to achieve a constant improvement in a stable and adaptable manner to the respective vehicle.

Further, the data-based sub-model may include at least one KI efficiency model, at least one KI burn-up model, and at least one KI differential pressure model, the KI efficiency model models KI efficiency model values by means of at least a portion of the measured parameter and/or feature data and supplements or replaces the physical efficiency model as an output error model, the KI burn-up model models KI burn-up model values by means of at least a portion of the measured parameters and/or characteristic data and supplements or replaces the physical burn-up model as an output error model, wherein the KI pressure differential model simulates a KI pressure differential model value associated with the modeled KI burnup model and/or the modeled KI efficiency model with the aid of a KI burnup model value and/or a KI efficiency model value, and wherein the KI differential pressure model supplements or replaces the physical differential pressure model as an output error model.

The efficiency model describes the change in the number of particles during loading. Instead, the change in mass of the particles during the burn-up process is calculated by means of a burn-up model. The difference consisting of two mass changes is then integrated. From this, the mass of particles (number of particles) remaining in the particle filter is derived. If the mass or quantity of particles in the filter is known, a model value of the pressure difference can finally be determined by means of the pressure difference model.

Furthermore, the exhaust system has a differential pressure measuring device for providing a differential pressure across the particulate filter, in particular a differential pressure at the inlet and the outlet.

This means that the physical model or individual physical submodels can be completely eliminated if necessary. The hybrid model structure can be created by three data-based sub-models. Individual data-based submodels may be used to change physical submodels in a modular fashion. Alternatively, a data-based submodel may be used as an output error model and used to validate or refine model values calculated by conventional submodels.

In other designs, if the measured differential pressure deviates from the KI differential pressure model value with the simulated KI efficiency model value input to the KI differential pressure model, the KI efficiency model is adapted such that the deviation is minimized.

In other designs, if the measured differential pressure deviates from the KI differential pressure model value with the simulated KI burnup model value input to the KI differential pressure model, the KI burnup model is adapted such that the deviation is minimized.

Preferably, the KI pressure differential model is adapted such that the deviation is minimized if the measured pressure differential deviates from the KI pressure differential model value in case the modeled KI burn-up model value is input into the KI pressure differential model, and if the pressure differential deviates the same or similar to the KI pressure differential model value in case the modeled KI efficiency model value is input into the KI pressure differential model.

Thereby, the respective data-based submodels are improved during operation and a better estimation can be achieved over time.

By using the data-based submodels, it is possible to additionally use individual data-based submodels instead of the corresponding physical conventional submodels, if for example in other life cycles another submodel with other input variables should be used. Furthermore, the data-based submodel can be updated repeatedly during ongoing operation, so that the simulated pressure difference corresponds to the measured sensor value, and the load and the filter efficiency of the particle filter of the internal combustion engine (gasoline or diesel) can therefore be determined more accurately during ongoing operation.

In particular, the data-based submodel can be used as an output error model to supplement the static physical submodel in a targeted manner. This means that the KI efficiency model is adapted as soon as there is a deviation between the simulated pressure difference and the measured pressure difference during loading. Conversely, if there is a deviation between the two pressure differences during regeneration/burn-up, it is preferable to adapt the KI burn-up model. If the same deviation exists both during loading and during burn-up, the KI pressure difference model is preferably adapted. It should be noted here that the adaptation of the data-based submodel can be carried out both offline and online, for example in the control device of the respective vehicle.

The object is also achieved by a computer program product having program code means for carrying out the method described above. The computer program product, such as the computer program means, may for example be provided or delivered as a storage medium, such as a memory card, a USB stick, a CD-ROM, a DVD, or may also be provided or delivered in the form of a file downloadable from a server in a network. The advantages already explained in connection with the method result in connection with the computer program product.

The object is also achieved by a control device for controlling an internal combustion engine having a particle filter, wherein the control device is designed to estimate a soot mass for loading the particle filter with soot particles from an exhaust gas mass flow of the internal combustion engine and/or a soot mass for regeneration on the basis of the method described above.

In this case, the control device is provided to carry out the method, that is to say has instructions on the basis of which the method according to one of the preceding embodiments is carried out. The advantages already described in connection with the method are achieved in connection with the control device. In particular, the control device has a processor on which the data-based submodel can be run and adapted. Furthermore, the control unit preferably has a differential pressure measuring device which is designed to measure a differential pressure and is arranged in the exhaust gas line.

Drawings

Preferred embodiments of the present invention are explained in more detail below based on the drawings.

Figure 1 shows an overview of a physical model with respective efficiency model, burn-up model and difference model for a particle filter according to the prior art,

figure 2 schematically shows a data-based submodel according to the invention in overview,

FIG. 3 shows an overview of a model with respective KI efficiency model, KI burnup model and KI difference model, in accordance with the present invention, and

fig. 4 schematically shows a particle filter in operation.

Detailed Description

Fig. 1 shows an overview of a physical model for a particle filter according to the prior art with various conventional physical submodels, which are here designed as a conventional efficiency model 200, a conventional burn-up model 300 and a conventional differential pressure model 400. Such a particle filter has a filter wall and already present soot particles layers which are regularly burnt off. The soot is burnt off both in the soot particle layer and in the filter wall. Here, regeneration is initiated when the soot quality is high. In this case, the non-measurable variables such as load and filtration efficiency are preferably determined in the physical model on the basis of measurable or other calculated variables, such as exhaust gas mass flow, pressure and temperature upstream of the particle filter, temperature in the particle filter, oxygen concentration in the exhaust gas mass flow and initially emitted mass flow.

In addition, the soot mass distribution between the filter wall and the established soot particle layer is also determined in the calculation of the filtration efficiency.

For this purpose, both the soot mass used for loading and the soot mass used for regeneration of the filter wall and soot particle layer, respectively, have to be determined in the physical model.

For this purpose, the current physical model is calculated with the soot particle mass as a basis variable. This calculation can also be converted into the number of particles.

The physical model is now used to calculate the soot mass for loading and the soot mass during regeneration in the filter wall and soot particle layer. The physical model is typically constructed as a non-linear model describing the physical interaction. The physical model includes a plurality of sub-models, referred to herein as an efficiency model 200, a burn-up model 300, and a differential pressure model 400, which will be described below.

The efficiency model 200 uses the filter efficiency and the exhaust gas mass flow

And pressure before the particle filter

And the temperature before the particle filter

Describes the change in mass of a particle filter

。

From this, the adapted efficiency can be calculated

And taking into account the raw exhaust gas mass flow

Determining the adapted efficiency

. Instead of particle mass, particle number may also be used. In this case, the efficiency model 200 describes the change in the number of particles during loading.

Mass change of smoke mass in smoke particle layer after burning off

By means of exhaust gas mass flow through the burn-up model 300

Pressure before the particle filter

Temperature before the particle filter

Oxygen in the exhaust gas mass flow

And the average temperature T in the particle filter_gpfAnd (4) determining. Thus, by means of the burn-up model 300, the change in mass of the particles during the burn-up process is calculated. Instead of particle mass, particle count may also be used. In this case, the burn-up model 300 describes the change in the number of particles in the burn-up process.

The difference between the two mass changes is then integrated and from this the mass of the particles remaining in the particle filter is finally derived.

The actual particle mass (particle number) applied to the filter wall and the soot particle layer in the particle filter can thus be calculated with the aid of the efficiency model 200. By means of the burn-up model 300, the corresponding change in mass of the particles (particle mass or particle number) during the burn-up process is calculated. The difference between the corresponding mass flow changes is then integrated. The mass of particles remaining in the filter wall and in the soot particle layer in the particle filter is ultimately determined from this.

If the particle mass or the particle quantity in the walls of the particle filter and in the soot particle layer is known, the model value of the pressure difference Δ p can finally be calculated by means of the further pressure difference model 400. However, this physical model requires a large amount of computation. Furthermore, not all interactions may be mapped into physical submodels.

Fig. 2 shows data-based submodels 2, 3 and 4, respectively, according to the present invention. These data-based submodels 2, 3 and 4 form, together with the physical submodel, a hybrid structure for the respective submodel. The outputs of the efficiency model 200, the burn-up model 300 and the pressure difference model 400 (physical submodel) are thus applied with the outputs of the correspondingly assigned data-based submodels 2, 3 and 4, in particular added or multiplied with the outputs of the correspondingly assigned data-based submodels 2, 3 and 4.

In addition to the conventional

physical submodels

200, 300, and 400 (FIG. 1), there may be various data-based submodels 2, 3, and 4. These data-based sub-models are here divided into a KI efficiency model 2, a KI burn-up model 3 and a KI differential pressure model 4.

The KI efficiency model 2 is for example based on an artificial neural network or another machine learning method (machine learning model). The artificial neural network may be a multi-layer model. The KI efficiency model 2 may be trained and adapted in a manner known per se based on representative measurement data from the testing phase. For example, the training is performed on the basis of training data that maps at least some of the parameters and the characteristic variables as correction variables by which the model values of the efficiency model are corrected. By this design, the physical associations may be mapped in an improved manner.

The artificial neural network may be updated repeatedly in continuous operation and thus used as an output error model. Instead of the artificial neural network, another method from the field of machine learning may also be used.

As with the traditional efficiency model 200 (fig. 1), the trained artificial neural network or KI efficiency model 2 obtains the following measured parameters: efficiency of filtration

Exhaust gas mass flow

Pressure before the particle filter

And temperature before the particle filter

. Instead of particle mass, the number of particles may also be used, if the neural network is trained to do so. In this case, the KI efficiency model 2 describes the change in the number of particles during loading.

If it is determined that the measured differential pressure deviates from the KI differential pressure model value 4, the KI efficiency model 2 is adapted such that the deviation is minimized. Thereby, the artificial neural network is further trained or tracked. This can be performed offline as well as online in the control devices of the individual vehicles.

The KI burn-up model 3 is likewise based on an artificial neural network or another artificial intelligence method (machine learning model). The artificial neural network may be a multi-layer model. The KI burn-up model 3 may be trained and adapted in a manner known per se on the basis of representative measurement data from the testing phase. For example, the training is performed based on training data that maps at least a part of the parameters and the characteristic parameters as correction parameters by which the model values of the burn-up model 300 are corrected. By this design, physical associations may be mapped in an improved way.

The artificial neural network may be updated repeatedly in continuous operation and thus used as an output error model. Alternative methods from KI may also be used in place of the artificial neural network.

As in the case of the conventional burn-up model 300 (FIG. 1), the KI burn-up model 3 or the trained artificial neural network relies on the exhaust gas mass flow

Pressure before the particle filter

Temperature before the particle filter

Oxygen in the exhaust gas mass flow

And the average temperature T in the particle filter_GpfDetermining the change in mass of soot in a soot layer after combustion

. Thus, burnup can be calculated by means of the KI burnup model 3The mass of the particles changes during the process. Instead of particle mass, particle number may also be used. In this case, KI burn-up model 3 describes the change in the number of particles during the burn-up process.

If it is determined that the measured pressure difference deviates from the KI pressure difference model value 4, the KI burn-up model 3 is adapted such that said deviation is minimized. Thereby, the artificial neural network is further trained or tracked. This can be performed offline as well as online in the control apparatus (control device 12) of the individual vehicle.

The KI difference model 4 is again based on an artificial neural network or another artificial intelligence approach. The artificial neural network may be a multi-layer model. The KI difference model 4 may be trained and adapted in a manner known per se based on representative measurement data from the testing phase. For example, the training is performed based on training data that maps at least a part of the parameters and the characteristic parameters as correction parameters by which the model values of the difference model 400 are corrected. The physical association can be mapped in an improved way by the design.

If the mass or quantity of particles in the walls of the particulate filter and in the soot particle layer is known, a KI differential pressure model value for the differential pressure Δ p can likewise be calculated by the KI differential pressure model 4. If there is a deviation between the differential pressure measured in the case where the modeled KI burn-up model value is input to the KI differential pressure model 4 and the KI differential pressure model value, and if there is a deviation that is the same as or similar to the differential pressure and the KI differential pressure model value in the case where the modeled KI efficiency model value is input to the KI differential pressure model 4, the modeled KI differential pressure model 4 is adapted so that the deviation is minimized. This can be performed offline as well as online in the control devices of the individual vehicles.

Fig. 3 shows a model 1 according to the invention with data-based submodels 2, 3, 4 according to the invention in addition to conventional

physical submodels

200, 300 and 400. It has been recognized that the efficiency model 200, the burn-up model 300, and the pressure differential model 400 have heretofore been based on complex physical associations that heretofore could only be described with simple models. Thus, measures for improving the filtration efficiency and for protecting the components may not be activated in a targeted manner, and, for example, there may be unnecessary flooding shut-off prohibition with increased fuel consumption and increased carbon dioxide emissions, if necessary.

Thus, according to the invention, these complex associations are now represented by data-based submodels 2, 3, 4, wherein the data-based submodels 2, 3, 4 are based on artificial intelligence, such as artificial neural networks, gaussian process models or other machine learning models.

The submodels/data-based submodels can therefore also be exchanged modularly if, for example, further submodels/data-based submodels with further input variables are to be used in further life cycles. Furthermore, the data-based submodels 2, 3, 4 may be updated repeatedly in continuous operation, so that the modeled pressure differences correspond to the measured pressure differences (sensor values). Model values of the data-based submodels may be compared to the measured values. The respective data-based submodels 2, 3, 4 can thus be adapted within the hybrid model structure such that the deviation between the measured values and the KI submodel values is minimized.

The data-based submodels 2, 3, 4 according to the invention can be supplemented in a targeted manner, i.e. the KI efficiency model 2 is adapted as soon as there is a deviation between the simulated KI pressure difference and the measured pressure difference during the loading process. Conversely, if there is a deviation between these two pressure differences during regeneration/burn-up, the KI burn-up model 3 is adapted. The KI pressure difference model 4 is adapted if the same deviation between the two pressure differences exists during loading and during combustion.

It should be noted here that the adaptation of the data-based submodels 2, 3, 4, respectively, can be performed off-line (e.g. using measurement data representative enough to test a fleet of vehicles during an application phase) and on-line in the control devices of the individual vehicles.

Fig. 4 schematically shows a particle filter 10 in operation, in which the method according to the invention can be used together with a control unit 12 according to the invention. The internal combustion engine 15, which is operated with fuel, delivers fresh air via the intake channel 13. The exhaust gases of the internal combustion engine 15 are discharged through an exhaust gas line 14 and are purified, in particular, by means of a catalytic converter 16 and a particle filter 10. When the internal combustion engine 5 is running, the particulate filter 10 filters soot particles from the exhaust gas.

If the particle filter 19 is loaded with particles, the pressure difference of the exhaust gas mass flow through the particle filter 10 increases. A regeneration process is then required to combust the particles. For this reason, the load on the particulate filter 10 should be set as accurately as possible. This regeneration process can be controlled by means of the control unit 12 according to the invention.

Claims

1. Method for estimating soot mass during loading of a particle filter (10) in an exhaust gas mass flow of an internal combustion engine and/or soot mass during burn-up by means of a physical model describing physical interrelationships during operation of the internal combustion engine on the basis of measured parameters and/or characteristic data, wherein the physical model has a plurality of sub-models, wherein each sub-model comprises a physical sub-model (200, 300, 400) having as input data measured parameters and/or characteristic data for the respective sub-model for determining respective sub-model values, wherein at least one sub-model is assigned a trained data-based sub-model (2, 3, 4) trained to supplement or correct the value of the physical sub-model concerned according to at least a part of the measured parameters and/or characteristic data, wherein the model output of at least one physical submodel (200, 300, 400) to which the data-based submodel (2, 3, 4) is applied is corrected by the corresponding data-based submodel (2, 3, 4).

2. The method of claim 1, wherein the at least one data-based sub-model comprises a trained artificial neural network or a gaussian process model.

3. Method according to any of the preceding claims, wherein the output of the at least one physical submodel (2, 3, 4) is applied with the output of the assigned data-based submodel, in particular by adding or multiplying with the output of the assigned data-based submodel.

4. The method according to any of the preceding claims, wherein the at least one data-based submodel (3, 4, 5) is modeled in addition to and/or independently of an associated physical submodel (200, 300, 400).

5. The method according to claim 4, wherein at least one physical submodel (200, 300, 400) is supplemented with the at least one data-based submodel (3, 4, 5) as an output error model.

6. The method according to claim 5, wherein in case of a deviation of the KI pressure difference model value corresponding to the at least one data-based sub-model (2, 3, 4) from the corresponding measured pressure difference, the corresponding data-based sub-model (2, 3, 4) is adapted such that said deviation is minimized.

7. Method according to claim 6, characterized in that the adaptation or training of at least one modeled data-based sub-model (2, 3, 4) is performed offline or online in a control device of the internal combustion engine.

8. The method according to any of the preceding claims, wherein the data-based submodel (2, 3, 4) comprises:

-at least one KI efficiency model (2) modeling KI efficiency model values by means of at least a part of the measured parameter and/or characteristic data, and wherein the KI efficiency model (2) supplements or replaces the physical efficiency model (200) as an output error model, and/or

-at least one KI burn-up model (3) modeling KI burn-up model values by means of at least a part of the measured parameter and/or characteristic data, and wherein the KI burn-up model (3) supplements or replaces the physical burn-up model (300) as an output error model, and/or

-at least one KI pressure difference model (4) modeling KI pressure difference model values by means of KI burnup model values and/or KI efficiency model values, and wherein the KI pressure difference model (4) supplements or replaces the physical pressure difference model (400) as an output error model.

9. The method according to claim 8, wherein the KI efficiency model (2) is adapted such that a modeled KI efficiency model value is minimized if the measured differential pressure deviates from the KI differential pressure model value if the KI differential pressure model value is input into the KI differential pressure model (4).

10. The method according to claim 8 or 9, wherein the KI burn-up model (3) is adapted such that a deviation of a measured differential pressure from a KI differential pressure model value is minimized if the deviation is present if the KI differential pressure model value is input into the KI differential pressure model (4).

11. The method according to claim 10, wherein the KI pressure differential model (4) is adapted such that the deviation is minimized if the measured pressure differential deviates from the KI pressure differential model value if a modeled KI burn-up model value is input into the KI pressure differential model, and if the pressure differential deviates from the KI pressure differential model value the same or similar if a modeled KI efficiency model value is input into the KI pressure differential model (4).

12. Control device (12) for controlling an internal combustion engine (15) having a particle filter (10), wherein the control device (12) is configured to estimate the mass of soot in the particle filter (10) and/or in the burned particle filter (10) in the exhaust gas mass flow of the internal combustion engine based on the method according to any of the preceding claims.

13. Computer program product comprising instructions which, when said program is executed by at least one data processing apparatus, cause said at least one data processing apparatus to carry out the steps of the method according to any one of claims 1 to 11.

14. A machine-readable storage medium comprising instructions which, when executed by at least one data processing apparatus, cause the at least one data processing apparatus to perform the steps of the method according to any one of claims 1 to 11.