Method for determining a particle load of a particle filter, particularly for use in an exhaust gas aftertreatment system for an internal combustion engine

Description

Technical area

[0001] The invention relates to particulate filters for cleaning a gas mass flow, in particular to soot particle filters for use in exhaust gas aftertreatment systems of motor vehicles. Furthermore, the invention relates to methods for more precisely determining the particle loading of such a particulate filter.

[0002] German Patent Application DE 10 2004 055 605 A1 discloses a method for determining the soot load of a particulate filter of an internal combustion engine. Among other things, the amount of ash accumulated in the particulate filter and a loading value for the soot load of the particulate filter are determined from a pressure difference across the particulate filter. Furthermore, an autoadaptive model for correcting the influence of ash deposits when determining the soot load in the particulate filter is disclosed.

[0003] A method for determining the filtration capacity of a diesel particulate filter is known from published patent application CN 110725738 B. The method includes, among other things, determining a pipeline air leakage correction factor, correcting a value detected by the differential pressure sensor by the pipeline air leakage correction factor, and then obtaining a corrected differential pressure value. Based on the relationship between the corrected differential pressure value and a lower differential pressure limit, it is determined whether the filtering capacity of the diesel particulate filter is normal or not.

[0004] Published patent application WO2006 / 055992 A2 discloses an emissions model that, based on engine operating data such as engine speed, torque, and other variables, provides a current ideal value for the particulate mass emitted by the engine. In addition, an integrating particle sensor determines a measured real integrated particulate mass during a measurement period. The modeled particulate mass and the measured particulate mass are fed into a comparison algorithm. A value for the deviation between these two results is further processed to correct the emissions model. For correcting the emissions model, one suggestion is, for example, to correct the particulate emission value using a factor. Technical background

[0005] Particle filters are typically used to reduce particle density in gas mass flows. For example, particle emissions from internal combustion engines in motor vehicles are reduced using particulate filters, which make a significant contribution to reducing soot particle emissions from diesel and gasoline engines. During operation of the internal combustion engine, the particulate filters become loaded with ash and soot particles, which settle in the walls and filter cake. In a regeneration process, the soot load is burned off at regular intervals with the addition of oxygen and at elevated exhaust temperatures to regenerate the particulate filter and maintain it in optimal operating condition. Disclosure of the invention

[0006] According to the invention, a method for determining a loading of a particle filter, in particular for cleaning combustion exhaust gas of an internal combustion engine, according to claim 1, and a computer program product and a machine-readable storage medium according to the independent claims are provided.

[0007] Further embodiments are specified in the dependent claims.

[0008] According to a first aspect, a method for determining a particle loading of a particle filter as at least one loading variable for an internal combustion engine is provided, comprising the following steps: Detecting or providing a plurality of state variables of the supplied gas mass flow, which comprise at least one particulate mass flow and a pressure difference indication for an actual pressure difference between an inlet side and an outlet side of the particulate filter; determining at least one modeled loading variable using a physical particulate filter model as a function of one or more state variables of a supplied gas mass flow; modeling a correction variable for the at least one modeled loading variable using a data-based error model as a function of the state variables of the supplied gas mass flow and the pressure difference indication, wherein the error model comprises or corresponds to a temporal convolutional network (TCM); determining the at least one loading variable as a function of the at least one modeled loading variable and the modeled correction variable.

[0009] The loading of a particulate filter is a crucial parameter that significantly influences the system behavior in which the particulate filter is operated. Excessive loading results in high gas backpressure through the particulate filter, which can lead to increased system energy consumption or even mechanical damage to the particulate filter. When the particulate filter is used as an exhaust particulate filter for a combustion engine in a motor vehicle, this leads to increased fuel consumption of the combustion engine.

[0010] Furthermore, the filter efficiency for a particle filter that is underloaded is low, so that a minimum loading of particles is necessary for sufficient filter efficiency.

[0011] It is therefore important to keep the loading of the particulate filter at a moderate level, i.e. within a certain range, in order to optimize the filter efficiency of the filter in terms of filtration and the energy consumption of the entire system.

[0012] This requires knowledge of the current particulate filter load in order to be able to influence it appropriately. The degree of particulate filter load, which can be specified with a load variable, can be adjusted or changed by activating or avoiding a regeneration cycle.

[0013] Such a regeneration cycle, for example, provides for reducing the load on a soot particle filter for an internal combustion engine by supplying oxygen and increasing the temperature in the particle filter in order to burn off the soot particles and thereby remove them from the filter.

[0014] However, the loading of a particulate filter cannot be measured directly, so it must be determined using a particulate filter model based on available variables during system operation, such as the engine system with the combustion engine. Conventional particulate filter models are generally physically based models that depict the complex behavior of a particulate filter depending on state variables. However, common particulate filter models are limited in their accuracy, particularly during a regeneration cycle. Although detailed simulation models for particulate filters are also known, they are not suitable for the online calculation of a loading variable.

[0015] Therefore, the above method provides for the use of a physical particle filter model that is corrected using a data-based error model. The data-based error model can provide a correction variable, e.g., in the form of a correction offset or a correction factor, for correcting the at least one modeled load variable.

[0016] The error model can be trained to specify a respective correction value for the respective load value depending on the course of the state variables and the pressure difference value, wherein the respective correction value is designed to apply to the respective load value, in particular to apply it additively or multiplicatively.

[0017] The error model can be trained by offline simulations of the particulate filter depending on the state variables of the system and depending on a difference between a real value of the at least one load variable and a modeled value of the at least one load variable.

[0018] Since the physical particulate filter model determines the loading based on the accumulation of a particulate mass flow, it is necessary to structure the error model in such a way that the correction value is determined taking into account the dynamic behavior of the particulate filter. This allows the error model to also represent accumulated errors of the particulate filter model, whereby the time series to be considered for determining the loading value can be very long.

[0019] Surprisingly, it was found that the use of a temporal convolutional network (TCN) as an error model is particularly suitable for representing the deviations of a modeled loading variable from the actual loading of the particle filter. Especially for long time series, the use of a temporal convolutional network as an error model has proven to be more accurate and less complex than conventional recurrent neural networks, such as LSTM or GRU, which can only represent dynamic behavior with good accuracy for shorter time horizons.

[0020] The state variables can include a gas volume flow, a temperature and / or a pressure.

[0021] The at least one load size and the at least one modeled load size may include: a wall loading quantity that indicates a quantity of particles deposited on surfaces of a filter material of the particulate filter, and a modeled wall loading quantity that indicates a modeled quantity of particles deposited on surfaces of the filter material of the particulate filter; and / or, a cake loading quantity that indicates a quantity of particles deposited in a cake on a surface of a filter element of the particulate filter, and a modeled cake loading quantity that indicates a modeled quantity of particles deposited in a cake on a surface of a filter element of the particulate filter; and / or a filter efficiency that indicates the cleaning efficiency of the particulate filter, and a modeled filter efficiency that indicates the modeled cleaning efficiency of the particulate filter.

[0022] It can be provided that a modeled pressure difference indication between the inlet-side and outlet-side pressure across the particulate filter is determined using the physical particulate filter model, wherein the correction variable for the at least one modeled loading variable is further carried out using the data-based error model as a function of the modeled pressure difference indication, wherein the error model is trained to indicate the respective correction variable for determining the respective loading variable as a function of the courses of the state variables, the pressure difference indication and the modeled pressure difference indication, wherein the respective correction variable is designed to apply to the respective modeled loading variable, in particular to apply to it additively or multiplicatively.

[0023] It can be provided that the method is carried out cyclically in time steps, wherein in each time step the values ​​of the state variables, the pressure difference information and the modeled pressure difference information are temporarily stored, wherein time series of the state variables, the pressure difference information and the modeled pressure difference information for a previous evaluation period are provided to the error model.

[0024] Furthermore, depending on at least one loading variable, the particulate filter can be regenerated or scheduled regeneration of the particulate filter can be prevented.

[0025] According to one embodiment, the particulate filter can be used as an exhaust gas particulate filter in an exhaust gas aftertreatment device of a motor vehicle.

[0026] According to a further embodiment, the temporal convolutional network of the error model can be formed with between 5 and 30 residual layers, which have a first sub-network for normalization and a second sub-network for causal convolution with a dilation rate increasing towards further downstream residual layers. Brief description of the drawings

[0027] Embodiments are explained in more detail below with reference to the attached drawings. They show: Figure 1 shows a schematic representation of an engine system with an internal combustion engine and a particulate filter for cleaning the combustion exhaust gas; Figure 2 shows a schematic representation of the structure of the particulate filter; Figure 3 shows a block diagram illustrating the function for providing a corrected charge quantity; Figure 4 shows a flow chart illustrating a method for determining a charge quantity of a particulate filter and for operating the engine system of the Figure 1 ; and Figure 5 a more detailed representation of the architecture of the error model as a temporal convolutional network. Description of embodiments

[0028] Figure 1shows a schematic representation of an engine system 1 with an internal combustion engine 2, as an example of a system with a particulate filter. The internal combustion engine 2 can be a diesel or gasoline engine operated by a control unit 3. Via an exhaust gas discharge section 4, the internal combustion engine 2 expels combustion exhaust gas, which is fed to an exhaust aftertreatment device 5 with a particulate filter 6.One or more sensors 7 can be arranged in the exhaust gas discharge section 4 in order to detect state variables of the operation of the engine system 1 in a sensory or model-based manner, such as a temperature sensor 9 for detecting an exhaust gas temperature T and an inlet pressure sensor 7 for detecting an exhaust gas back pressure p in , which is provided as the inlet pressure of the particle filter 6, wherein depending on the operating point and the load and speed of the internal combustion engine 2 further state variables, such as an exhaust gas mass flow V, as well as a soot load . soot, in of the combustion exhaust gas, can be modelled using suitable models implemented in the control unit 3.

[0029] A further pressure sensor 8 can be arranged on the outlet side of the particulate filter 6 to measure a downstream exhaust pressure p out . Alternatively, the ambient pressure can be assumed as the downstream exhaust pressure. Alternatively, a differential pressure sensor can also be used to directly detect the differential pressure reading.

[0030] In Figure 21 shows a schematic representation of the structure of the particulate filter 6. The particulate filter has an inlet channel 61 and an outlet channel 62, which are separated by a filter element 63 made of filter material. The filter element 63 is porous and filters the exhaust gas mass flow flowing through the filter material 63. The filter material of the filter element 63 offers a large surface area for the deposition of soot particles. During the separation process of soot particles, soot particles accumulate on surfaces in the filter material. This loading is specified as the wall particle loading by a wall loading variable. Furthermore, during the separation process of soot particles, soot particles accumulate in a filter cake on the surface of the filter element 63. This loading is specified as the cake particle loading by a cake loading variable. Soot particles are also deposited to a lesser extent on the walls of the inlet channel 61.

[0031] In conjunction with the functional block diagram of the Figure 3 and the flow chart of the Figure 4 A method for operating the particulate filter 6 is described in more detail below. The method can be implemented in the control unit 3 as hardware and / or software and serves to determine a wall loading variable, a cake loading variable, and a filter efficiency as loading variables depending on the state variables of the engine system 1.

[0032] The block diagram of the Figure 3 shows a loading model 30 with a physical particle filter model 31, which is based on state variables T, ṁ soot, in , V̇ , p in the operation of the engine system 1 a modeled wall loading size m soot,wall,mod and a cake loading size m soot,layer,mod Furthermore, the particle filter model 31 can provide a modeled pressure difference value Δ p mode, e.g. as differential pressure p in - p out across the particle filter 6 and furthermore a filter efficiency the way model.

[0033] Furthermore, the state variables T, soot, in , V̇ , p in and the modeled pressure difference Δ p mode from the particle filter model 31 and a measured pressure difference Δ p measure (e.g. as the difference between the measured inlet pressure p in and the measured outlet pressure p out of the particle filter 6) is fed to a shift register 37 in order to store time series of the corresponding variables and to provide them to a downstream error model 32. The error model 32 is data-based and trained to generate a correction variable for the wall loading variable k soot,wall , a correction variable for the cake loading variable k soot,layer and a correction variable k η for the filter efficiency depending on the time series of the state variables T, soot, in , V̇, p in , the modeled pressure difference Δ p mode from the particle filter model 31 and the measured pressure difference Δ p measure to provide.

[0034] The correction quantities k soot,wall , k soot,layer and k η can be provided as a correction offset and / or correction factor and are used in corresponding loading elements 33, 34, 35 of the modeled wall loading size m soot,wall,mod, the modeled cake loading size m soot,layer,mod and the modeled filter efficiency the way In this way, the corrected wall loading size is obtained m soot,wall, Cake loading size m soot, layer and filter efficiency η as output variables. Furthermore, the particle filter model 31 can provide a modeled pressure difference value Δ p mode , e.g. model as differential pressure p in - p out across the particle filter 6.

[0035] In the flowchart of the Fig. 4, which is executed cyclically, the state variables T, soot, in , V̇ , p in recorded or modeled or provided.

[0036] The physical particle filter model 31 accumulates in step S2 the current particle mass flow in the combustion exhaust gas depending on the state variables to the modeled wall loading variable m soot,wall,mod, the modeled cake loading size m soot,layer,mod and further determines the modeled filter efficiency the way . Furthermore, the physical particle filter model 31 models the pressure difference Δ p mode provided.

[0037] In step S3, the state variables T, ṁ soot,in , V̇ , p in , which is stored in the shift register 37 and a measured pressure difference value Δ p measure buffered in a suitable shift register of the error model at each time step. In this way, time series of the state variables T, ṁ soot,in , V̇ , p in the measured pressure difference Δ p measure and the modeled pressure difference Δ p mode provided.

[0038] Furthermore, in step S4, the time series of the above variables are fed to the error model 32, which, depending on this, calculates the corresponding correction variables k soot,wall , k soot,layer and k η provides.

[0039] In step S5, the corrected wall loading size m soot,wall, the corrected cake loading size m soot, layer and the corrected filter efficiency η as output variables from the modelled wall loading size m soot,wall,mod , the modeled cake loading size m soot,layer,mod and the modeled filter efficiency the way calculated using the appropriate correction variables, in particular by additive or multiplicative loading.

[0040] In step S6, depending on the load sizes determined in this way, m soot,wall , m soot,layerand the filter efficiency η an intervention in the operating strategy for the regeneration of the particulate filter 6 can be made. For this purpose, it can be checked in step S6 whether the sum of the wall loading size m soot,wall and the cake loading size m soot, layer exceeds a predetermined threshold value. If this is the case (alternative: yes), a regeneration cycle is started in step S7, in which the operation of the internal combustion engine 2 is modified such that an increased oxygen input into the combustion exhaust gas occurs and the combustion exhaust gas temperature is also increased.

[0041] If the threshold value is not exceeded (alternative: No), a check is carried out in step S8 to determine whether the sum of the wall loading size m soot,wall and the cake loading size m soot, layerfalls below a predefined second threshold value. If this is the case (alternative: Yes), a regular regeneration cycle is inhibited in step S9 until the second threshold value is exceeded, and a scheduled regeneration is not performed if necessary. The method then continues with step S1.

[0042] The error model 32 can be implemented as a data-based error model. In particular, it is proposed herein to implement the error model as a temporal convolutional neural network (TCN). The architecture of such a neural network is described in the paper Bai S. et al., "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling," arxiv:1802.01271 v2, April 19, 2018.

[0043] The error model is trained based on simulations of the particle filter loading in conjunction with the provided physical particle filter model. For this purpose, time series of the above-mentioned state variables are measured or modeled during operation of a real combustion engine, and a simulation is used to determine the corresponding particle loading curve of the particle filter. In particular, simulated curves of the wall loading variable are obtained. m soot,wall,mod, the cake loading size m soot,layer,mod and filter efficiency the way .

[0044] Furthermore, based on the state variable curves, the physical particle filter model is evaluated in order to determine the curves of the modeled wall loading variable m soot,wall,mod, the modeled cake loading size m soot,layer,mod and the modeled filter efficiency the way and the modeled pressure difference Δ p modeThe respective correction values, particularly for additive or multiplicative loading, are now determined for the individual support points of these modeled and simulated curves.

[0045] This results in training data sets for training the error model 32, the courses of the state variables T, soot, in , V̇ , p in , the modeled pressure difference Δ p mode and the measured pressure difference Δ p measure and assign them to a correction value for a current time step that corresponds to the last time step of the curves. The error model can be trained offline using these training data sets. In this way, the error model 32 can be trained based on the highly accurate loading simulations of a particulate filter.

[0046] The state variable curves considered in the loading model correspond to time series of the listed state variables for a predetermined evaluation period or a predetermined time series length. For example, the time series can correspond to a range between 10 s and 60 s and a number of data sets between 200 and 2,000.

[0047] The error model 32 can be designed as a neural network with an input layer 41, an output layer 42 and a number of, for example, between 5 and 30 residual layers, as in Figure 5 Each layer comprises a number of neuron elements that are functionally connected to neuron elements of other layers.

[0048] The input layer 41 represents a representation of the time series of the state variable curves, the modeled pressure difference value and the measured pressure difference value for the evaluation period. The output layer 42 represents a linear classifier in the form of a dense layer in order to calculate the relevant correction variables, ie the correction variable for the wall loading variable k soot,wall , the correction variable for the cake loading variable k soot,layer and the correction variable for the filter efficiency k η to provide.

[0049] Each of the residual layers 43 comprises a series of mathematical operations, which are explained below. Each residual layer 43 has several subnetworks. A first subnetwork comprises a weight normalization layer, a second subnetwork comprises a causal convolution layer, and a third subnetwork comprises a nonlinear activation function and a test-time dropout layer.

[0050] A state variable curve with a time series of x t t = 1 d ∈ ℝ d is thus first processed in a causal convolutional layer with normalized weights. The weight kernel of the convolutional layer can be: w = w t t = 1 K ∈ ℝ K , where K is the kernel size. Weight normalization decomposes the channel weights into normalized representations w = g ‖ v ‖ v , where g is the magnitude of the weighting vector w and v v represents its direction.

[0051] The corresponding output signal y = y k k = 1 d ∈ ℝ d of causal convolution is as follows: y t = g ‖ v ‖ ∑ k = 1 K x t − k ⋅ r ⋅ υ k where r corresponds to the dilation rate and increases geometrically from one residual layer 43 to the next residual layer, thus maximizing the effective range of the evaluation period for the rear residual layers 43. Subsequently, the convoluted signal y is subjected to a non-linear activation function (e.g., ReLu).

[0052] Using an optional dropout process, a model parameter can be randomly set to zero with a certain probability in all neuron layers of the neural network except the output layer. This measure avoids overfitting the model during training of the error model. When applying the error model for prediction, repeated calculations with the same input variables and the dropout measure can generate a correspondingly stochastic output signal, which can be evaluated to determine the correction value. See also Yarin Gal et al., "Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning," University of Cambridge, https: / / arxiv.org / pdf / 1506.02142.pdf. z t = max 0 , y t ⋅ c t where ct ~ Bernoulli(1 - p ) with p a given dropout rate.

[0053] The residual layers 43 are traversed one after the other and the final output of the residual layer 43 is added element by element to the input of the respective residual layer 43.

Claims

1. Method for determining a particle loading of a particle filter (6) with at least one loading variable, by performing the following steps: - detecting (S1) or providing a number of state variables (T, ṁsoot,in, V˙, pin) of a supplied mass flow of gas, which comprise at least one mass flow of particles and a pressure difference indication, in particular a measured pressure difference indication (Δpmess), for an actual pressure difference between an input side and an output side of the particle filter (6); - determining (S2) at least one modelled loading variable (msoot,wall,mod, msoot,layer,mod) with the aid of a physical particle filter model (31) depending on the one or more state variables (T, ṁsoot,in, V˙, pin) of the mass flow of gas supplied; - modelling (S4) a correction variable (ksoot,wall, ksoot,layer, kη) for the at least one modelled loading variable(msoot,wall,mod, msoot,layer,mod) with the aid of a data-based error model (32) depending on the state variables (T, ṁsoot,in, V˙, pin) of the mass flow of gas supplied, the pressure difference indication (Δpmess), wherein the error model (32) corresponds to a temporal convolutional network (TCM) or comprises such a network; - determining (S5) the at least one loading variable (msoot,wall, msoot,layer) depending on the at least one modelled loading variable (msoot,wall,mod, msoot,layer,mod) and the modelled correction variable (ksoot,wall, ksoot,layer, kη).

2. Method according to Claim 1, wherein the state variables (T, ṁsoot,in, V, pin) comprise a gas volume flow (ṁsoot,in), a temperature (T) and / or a pressure.

3. Method according to either of Claims 1 and 2, wherein the loading variable (msoot,wall, msoot,layer) and the modelled loading variable (msoot,wall,mod, msoot,layer,mod) comprise: - a wall loading variable (msoot,wall, msoot,wall,mod),which indicates an amount of particles accumulated on surfaces of a filter material (63) of the particle filter (6), and a modelled wall loading variable, which indicates a modelled amount of particles accumulated on surfaces of the filter material (63) of the particle filter (6); and / or - a cake loading variable (msoot,layer, msoot,layer,mod), which indicates an amount of particles deposited in a cake on a surface of a filter element of the particle filter (6), and a modelled cake loading variable, which indicates a modelled amount of particles deposited in a cake on a surface of a filter element (63) of the particle filter (6) - a filter efficiency (η), which indicates the cleaning efficiency of the particle filter (6), and a modelled filter efficiency, which indicates the modelled cleaning efficiency of the particle filter (6).

4. Method according to one of Claims 1-3, wherein the error model (32) is trained to indicate the respective correction variable (ksoot,wall, ksoot,layer, kη) for the respective loading variable depending on progressions of the state variables (T, ṁsoot,in, V, pin), the pressure difference indication (Δpmess) and the modelled pressure difference indication (Δpmod), wherein the respective correction variable (ksoot,wall, ksoot,layer, kη) is designed to apply, in particular additively or multiplicatively apply, the respective modelled loading variable (msoot,wall,mod, msoot,layer,mod).

5. Method according to one of Claims 1-4, wherein a modelled pressure difference indication (Δpmod) between the input-side pressure and the output-side pressure over the particle filter (6) is ascertained with the aid of the physical particle filter model (31), wherein the correction variable (ksoot,wall, ksoot,layer, kη) for the at least one modelled loading variable (msoot,wall,mod, msoot,layer,mod) is also implemented with the aid of the data-based error model (32) depending on the modelled pressure difference indication (Δpmod), wherein the error model (32) is trained to indicate the respective correction variable (ksoot,wall, ksoot,layer, kη) for ascertaining the respective loading variable (msoot,wall, msoot,layer) depending on progressions of the state variables (T, ṁsoot,in, V, pin), the pressure difference indication (Δpmess) and the modelled pressure difference indication (Δpmod), wherein the respective correction variable (ksoot,wall, ksoot,layer, kη) is designed to apply, in particular additively or multiplicatively apply, the respective modelled loading variable (msoot,wall,mod, msoot,layer,mod).

6. Method according to one of Claims 1-5, wherein the method is performed cyclically in time increments, wherein in each time increment the values of the state variables (T, ṁsoot,in, V, Pin), the pressure difference indication (Δpmess) and, if applicable, the modelled pressure difference indication (Δpmod) are buffer-stored, wherein the error model (32) is provided with time series of the state variables (T, ṁsoot,in, V, Pin), the pressure difference indication (Δpmess) and, if applicable, the modelled pressure difference indication (Δpmod) for a previous evaluation time period.

7. Method according to one of Claims 1-6, wherein, depending on the at least one loading variable (msoot,wall, m soot, layer), the particle filter (6) is regenerated or a planned regeneration of the particle filter (6) is prevented.

8. Method according to one of Claims 1-7, wherein the particle filter (6) is used as an exhaust particle filter in an exhaust gas aftertreatment device of a motor vehicle.

9. Method according to one of Claims 1-8, wherein the temporal convolutional network has between 5 and 30 residual layers, comprising a first subnetwork for a normalization and a second subnetwork for a causal convolution with a dilation rate increasing towards further subordinate residual layers.

10. Computer program product, comprising commands which, when the program is executed by a control unit of an internal combustion engine, cause it to perform the steps of the method according to one of Claims 1 to 9.

11. Machine-readable storage medium, comprising commands which, when executed by a control unit of an internal combustion engine, cause it to perform the steps of the method according to one of Claims 1 to 9.