DE19741973C1 - Method of determining the soot conc. of self-igniting internal combustion engines - Google Patents

Abstract

the method involves deriving the soot conc. using a neural network (202) which is trained using input data (201) characteristic of the soot conc. (203) with associated signals representing the soot conc. Finally, the soot conc. is derived from acquired input data using the trained neural network

Description

The present invention relates to a method for determining the soot concentration of self-igniting internal combustion engines.

Various procedures are known, the soot concentration arithmetically by a To determine the modeling of the combustion process and the flow processes. Here the combustion process is analyzed thermodynamically. Under consideration of The soot concentration is calculated from this model.

Furthermore, various methods are known in which, for example, a measurement of the Soot concentration occurs during a combustion process. Such a measurement is in particular for determining the soot concentration of an internal combustion engine comparatively complex due to the fact that in a comparatively short time sequence repetitive ignition processes.

DE 38 39 348 A1 discloses a method in which the soot concentration in the exhaust gas of internal combustion engines is determined based on the absorption of an exhaust gas stream transmitted light beam.

From EP 0 724 073 A2 a method is known in which the air-fuel Mixture ratio value of an internal combustion engine is to be derived. This is done at Use of a neural network, in which first a training of the neural Network using characterizing input data. Subsequently the air / fuel mixture ratio value is determined from the input data determined Use of the trained neural network derived.

From US-PS 5,093,792 it is known to have a tendency to knock or misfire Detect internal combustion engine using a neural network. The neural network is trained using input data representing the cylinder pressure and represent the angle of rotation of the crankshaft. Then knock tendency or Misfires can be detected by measuring the trained neural network Input data are supplied.

In contrast, it is the object of the invention to provide a method for a simple to propose a realizing recording of the soot concentration. This procedure is also intended to meet different operating conditions as flexibly as possible.

This object is achieved with a method according to claim 1, according to which the soot concentration of the internal combustion engine is derived using a neural network, the neural network being trained using the soot concentration characterizing input data with associated signals that the Represent soot concentration, with the soot concentration subsequently determined from Input data is derived using the trained neural network.

To generate the training data, the soot concentration can be determined, for example, after Bosch process can be measured. With this measuring method an averaged value is over the measuring time is generated. So it is possible to get an averaged value per work cycle Determine soot concentration. This advantageously limits the amount of data, which the Limits processing effort when training the neural network.  

In the Bosch method, the soot concentration is measured with, for example an AVL spot smokemeter. The assignment of a paper filter is detected by a optical evaluation.

The associated signals, which represent the soot concentration, can for example be measured or on the basis of other measured quantities - for example by a Model calculation or again using other neural networks - determined have been.

The use of the neural network has the advantage that it is not necessary comparatively complicated models with connections between the Combustion process and flow processes to be detailed in order for each to which the soot concentration is to be determined, a calculation based on these models to be able to perform. Rather, it is sufficient to use suitable quantities as input data for the to find neural network, taking between these sizes and the soot concentration injective connection must exist. The neural network is then able to predetermined input data and associated measured output data generate that the associated output data based on given input data can be determined by the neural network. Another significant advantage of a neural network is that it is stable against noisy input signals Shows behavior. As a result, there are only comparatively low requirements for the Transducer, so that here costs and effort, especially in the case of Difficult measurement conditions on the engine due to environmental influences are limited can be. When using the neural network, it does not matter whether the Relationship between the input data and the output variable is linear or non-linear.

In the method according to claim 2, the soot concentration is one of them Input data characterizing variables characterizing the cylinder pressure.

These variables can, for example, from the maximum value of the cylinder pressure, the Integral of the cylinder pressure over a certain crank angle, the mean indicated Combustion chamber pressure, the maximum pressure increase in relation to the crank angle or similar exist. Choosing one of these or a similar size has the advantage that Only one value is to be processed per work cycle, which means the processing effort both in the Learning phase of the neural network as well as in the subsequent use of the neural Network for determining the soot concentration considerably simplified.  

However, according to the embodiment of the method according to claim 3, it is also possible that for the input data characterizing the soot concentration, the time course of the Heard cylinder pressure.

When using this input data, the processing effort increases, it does but there are advantages in terms of accuracy, because with the processing of the temporal Course of the cylinder pressure also influences the pre-injection on the Soot concentration can be detected. In addition to these influences of the pre-injection other influences can also be taken into account.

In the embodiment of the method according to claim 4, the signal of the cylinder pressure filtered.

Neural networks show stability against noisy signals, but it has with regard to the cylinder pressure shown that the noise of the signal is not white noise is about influences of disturbing factors with a characteristic Frequency spectrum. Accordingly, low-pass filtering has proven to be useful a frequency of 5 kHz, for example, can be used as the basic frequency.

The cylinder pressure can be measured, for example, or also by means of a neural network can be determined from other sizes, as is the case, for example, in the the applicant's previously published application has been described as national Patent application at the German Patent Office under file number P 197 41 884.8. If the cylinder pressure is not required directly as an intermediate variable, it would be, for example conceivable to "summarize" the neural networks. Instead of the "detour", initially by means of a neural network of structure-borne noise signals from the internal combustion engine Determine cylinder pressure and in turn further using a neural network A neural network could therefore determine the soot concentration from the cylinder pressure are used in which the input data characterizing the soot concentration would be the structure-borne noise of the internal combustion engine.

In the design of the method according to claim 5, the soot concentration is one of them characterizing input data firing curve, temperature curve and / or Total burning process.

It has been shown that these input data also characteristically make a statement about the Allow soot concentration.  

When designing the method according to claim 6, the firing process, Temperature curve and / or total combustion curve through a combustion curve calculation determined.

This firing history calculation is a simple method to add the relevant sizes determine. In the firing course calculation, the time course of the implementation of chemically bound fuel energy in thermal energy from the known Combustion chamber pressure curve calculated in an internal combustion engine. The calculation is done at Application of the first law of thermodynamics and thermal Equations of state.

In the design of the method according to claim 7, the soot concentration is one of them characterizing input data at least one of the following variables: Geometry of the piston, geometry of the combustion chamber, injection pressure, Injector geometry, position of the piston during injection, Average mass temperature, speed, load, air-fuel ratio (λ), residual gas, start of combustion.

These sizes are particularly suitable in combination with one of the aforementioned sizes Input data because their additional use improves accuracy a uniqueness between the input data and the soot concentration as Can improve output size.

It has been shown that meaningful results can already be determined if only a part of the quantities mentioned is used as input data. The results but will be improved if the network has more information.

In the method according to claim 8 is based on the trained network from the listed Sizes made a selection based on which a suitable assignment of the Soot concentration to the input data results.

As a result, it can be used in a specific individual case for a specific type of internal combustion engine the sensible input data are selected by the neural network itself. First the neural network is trained with a set of input data in which some Information with a view to determining the output quantity is superfluous. After the training of the network is complete, the neural network can do this again Recognize input data that are superfluous in this sense. This means that for the future use of the neural network the number of input data can be reduced. It can also reduce the input data through the neural network again knowledge about the formation of soot can be gained.  

In the method according to claim 9, as input data of the neural network or several main components of the corresponding input data are used.

This advantageously reduces the amount of data to be processed.

An embodiment of the invention is shown in more detail in the drawing. It shows in detail:

Fig. 1 is a representation of an approach for processing of learning data by the neural network,

FIG. 2 is an overview of the interdependencies in learning, test and recall phase,

Fig. 3: a complementary overview of the effect relationships in the training and testing phase,

FIG. 4 shows a further overview of the effect relationships in the learning, a test and recall phase reference to a preferred embodiment and

FIG. 5 shows an outline diagram for compressing the input data by the principal component analysis.

Fig. 1 shows the representation of a block diagram in which the neural network is first trained in a step 101 with training data.

In step 102 , a check is carried out with data in order to check the error limit of the neural network. This testing of the network is preferably carried out with unknown data. This can advantageously be used to determine whether the error limit has been reached. The error can be defined, for example, in such a way that the measured output variable is compared with that based on the neural network by forming the difference between these two values. For example, the error can be assessed on the basis of the sum of the error squares.

The relationships are shown in more detail by FIG. 3. In block 301 , learning takes place on the basis of the learning data which are supplied in accordance with block 302 . In accordance with the relationships represented by arrows 303 and 304 , testing is carried out in accordance with block 305 . For this purpose, test data are supplied in accordance with block 306 . Based on the test results, it may be possible to learn again. However, it is also conceivable to change the network topology in accordance with the illustration by arrow 307 in block 308 in order to improve the results. Accordingly, the learning in block 301 according to arrow 309 must then be started again. If it has been determined during testing in block 305 that the error is sufficiently small, a transition to block 311 takes place in accordance with arrow 310 , in which the learning and test phase is terminated.

In order to keep the effort for measuring the training and test data within limits, it is possible to compile several sets of data. You can use these records statistical methods are divided into training data and test data. By doing this Splitting is performed several times, so different combinations of Training data and test data used. Overall, the error can be minimized in this way. For example, with approximately 100 patterns, 10 of these patterns can be used statistically as test patterns to be chosen. If this process is repeated about five times, it will be seen that at averaging the results of the errors obtained is kept within reasonable limits can be.

For the acquisition of the training data, it has proven to be useful during the Work cycle under different load conditions and different ignition times, d. H. Injection times to determine the relevant output variables.

In step 103 , the neural network is used to derive the output variables from currently measured input data.

FIG. 2 shows a block diagram in which the interrelationships between the input data and the output variables and the neural network are illustrated again.

Block 201 characterizes the input data that are fed to a neural network 202 . Output variables 203 are generated from this neural network. In a learning phase of the neural network, the neural network 202 is trained, as represented by block 204 . After this network has been trained, the input data can be selected in accordance with the representation by block 205 . For example, variables that correlate with other variables with regard to the output variable can be eliminated.

The procedure for determining the soot concentration by means of neural networks is to be explained schematically with reference to FIG. 2. Initially, no physical model is available. However, there are data sets in which important information about the relationships between the input data 201 and measured values of the output variable 203 (soot concentration) is implied. These data records come, for example, from measurements on the internal combustion engine or from combustion history calculations. The input data 201 consist of different temporal profiles such as the combustion chamber pressure, the mean mass temperature, the combustion profile and individual information such as the speed and the load or the fuel-air mixture λ. The output variable 203 is the soot concentration. The first step is to generate a neural network and randomly initialize the network parameters. This is represented by blocks 204 and 202 . The initial configuration (number of neurons, layers) must be specified in a suitable manner. Then the network training takes place. The input data 201 should already be suitably preprocessed for this (scaling, favorable range of values, ie between 0 and 1 with a corresponding distance from the edges). The learned learning algorithm then controls the transition from the random, "ignorant" initial state to the trained, correctly mapping final configuration. The knowledge implied in the training data is transferred to the network during learning and stored in the weights.

Several different input data or courses of input data are advantageously used for the training. This is necessary because the suitable set of input data cannot yet be exactly defined due to a lack of knowledge of the exact physical relationships. Due to correlations between different input data, an unnecessarily large amount of information is initially created, which leads to a large input layer. After the network training, a neural model is now available for mapping the selected input information to the soot concentration. The implied knowledge contained therein about the relationships between the input data and the network output, ie the output variable, can then be extracted using suitable methods. This newly acquired knowledge can then be used to identify the relevant characteristics of the input data and to estimate their significance. This is shown in block 205 . A neural network is then trained with the suitable input data record found. After that, the application (recall phase) of the neural model obtained and the very rapid calculation of the soot concentration can take place.

For the training of neural networks, the smallest possible dimension of the input data is advantageous. This creates networks with few input neurons and degrees of freedom (Number of weights to be set). If the input data dimension is related to the available training data is too high, there is a risk that the data room no longer can be covered sufficiently and the trained network between the support points predominantly has to be extrapolated (instead of interpolation if appropriate  Data distribution). The number of training data required increases exponentially with the Dimension of the input data. The advantages of small networks with few Input neurons are the better generalization performance for unskilled test data less degrees of freedom. Smaller networks may also allow the extraction of simple rules by analyzing the weights. The network can be trained faster. Of the Mains output can be calculated more quickly in the recall phase, which is particularly important with regard to the Real-time capability is beneficial. In addition, there is a simpler and less expensive Hardware implementation. In the optimal case, however, the network should not be used for Features approximation relevant features are withheld. As part of the Character extraction now provides the relevant information at a lower cost Form for the subsequent network. This concerns the compressed representation as well as suitable ones Value ranges. The creation of the complete timelines, for example the Combustion chamber pressure of all cylinders to a network in a four-cylinder Internal combustion engine initially 804 input neurons at 201 combustion chamber pressure values per Cylinder. The dimension of the input data is clear for the 120 available support points too large (804 dimensions, inputs). Therefore a dimension reduction proves to be advantageous. Optimal dimension reduction involves maximum compression of data with minimal loss of information. For example, averaging of the Gradients take place over the cylinders. This seems sensible in that it is not excessive Differences prevail between the different cylinders and the measured ones Soot concentration results from the sum of the contributions of all cylinders. One disadvantage is that possible loss of combustion information due to the removal of certain cylinder-specific effects through averaging.

In order to further reduce the 201-dimensional data space obtained, one was found Correlation analysis (autocorrelation) takes place within the courses. The 120 pressure values correlated with each other at two different crank angles (e.g. the pressure values at 5 ° crank angle with pressure values at 20 ° crank angle). The calculated mean correlation coefficient (over the crank angle range) is 0.96. This high correlation signals great redundancy of the information and the possibility of further dimension reduction.

The result is a curve shape that is similar except for the area where the combustion begins, with high correlation coefficients (0.96) at combustion chamber pressure and lower correlation coefficients for all the curves calculated thereon. With these correlations, the principal component analysis for dimension reduction (Principal Component Analysis, PCA) lends itself. The PCA tries to find the directions in which the input data vary the most and which are orthogonal to each other. Geometrically, as exemplified in FIG. 5 for a simple two-dimensional case, this means the rotation of the original coordinate system (x1, x2) in the direction of the main axes (a1, a2). The axis a1 now points in the direction of the first main component with the greatest variance of the data and contains the greatest information content. Dimensional reduction can now take place in high-dimensional data rooms by using only the first main component with the largest proportion of the overall variance. The remaining main components are interpreted as noise and ignored. In mathematical terms, the aim of the PCA is to find the most important eigenvectors of the covariance matrix of all input vectors in which the redundancy is expressed. The N main components H (N * P matrix) then result from the transformation of the data (M * P matrix) into the new coordinate system spanned by the eigenvectors E (N * M matrix):

H = E * D,

where P is the number of data, M their dimensionality and N the dimensionality in the new Data room mean. Dimension reduction takes place at N <M using the N Main components with the greatest variance. The PCA results for strongly correlated input data a big dimension reduction. The extracted main components are among each other not correlated and independent. For each main component, the proportion of Overall variance can be specified, which is explained by the main component.

The application of PCA to combustion chamber pressure shows that the first 10 main components of the pressure curve cover over 99% of the total variance. With these values it turns out the loss of information when describing the courses due to only 10 main components low.

It has proven advantageous to use a feedforward type neural network. With this type of network, the effort involved in processing the data is kept within limits. a good accuracy of the initial results can nevertheless be achieved.

For the training of the neural network, a monitored learning process has proven to be Appropriately proven, because through this supervised learning the error with reasonable Processing effort can be minimized. It is suitable for evaluating the errors for example a gradient method with a momentum term or the Levenberg Marquardt optimization.

It has proven to be expedient to use a feedforward network. In the learning phase was both with Levenberg-Marquardt optimization and with learning Momentum and adaptive learning rule worked. There was more to learning with momentum Learning cycles necessary to fall below a certain error limit. The found one The minimum was often not as good as that with the Levenberg-Marquardt optimization  was determined. Part of the time saved by Levenberg-Marquardt optimization goes lost again, however, because more computing time is required per learning cycle. Of the The main disadvantage of Levenberg-Marquardt optimization is that it is considerable Memory requirement due to the increasing number of matrix elements to be stored, depending on the number of input neurons, the number of hidden neurons, the Output neurons and the training pattern.

It has proven to be useful to provide a network with 12 main components and about 25 hidden neurons.

It has proven to be useful to determine the soot concentration, more than just that 12 main components to be provided. For example, good results have been achieved with the other variables speed, load, λ, fuel mass and residual gas as further input variables. An input vector is therefore made up of these five pieces of information and the 12th Main components together.

Fig. 4 shows the relationships of the preferred embodiment in a symbolic representation. The combustion chamber pressure curve is first measured in block 401 . It is also possible to determine the course of the combustion chamber pressure in some other way or to calculate the course of the combustion or the temperature from the course of the combustion chamber pressure in accordance with the representation in blocks 402 or 403 . In accordance with block 404 , this data is preprocessed and fed to the neural network 405 .

Furthermore, for example, from the quantities that were determined in blocks 401 , 402 or 403 , further quantities, which can be, for example, the fuel / air mixture λ, the load, the residual gas mass and the ignition timing are determined in accordance with the representation in block 406 . These variables are also fed to the neural network 405 .

Furthermore, as shown in block 407, it is also shown that the neural network 405 is supplied with the speed as an input signal.

The soot concentration is then determined from the neural network as shown in block 408 .

The neural network 405 thus goes through a learning and test phase. Then, in the recall phase, the soot concentration can be determined from the corresponding input data using the neural network.

The described determination of the soot concentration is suitable, for example, in a Real-time calculation to determine the soot concentration as part of an on-board diagnosis and / or as a controlled variable. For example, it is possible to use a cylinder-selective one Detect a sooty combustion.

It is also possible to map the map in non-stationary operation, including the Soot concentration take place.

The soot concentration during engine development can also be used with this method be determined. This shows a better efficiency in the development because of the Possibility of better design of the engines.

Claims (9)

1. A method for determining the soot concentration of self-igniting internal combustion engines, characterized in that the soot concentration of the internal combustion engine is derived using a neural network ( 202 ), the neural network ( 202 ) being trained using input data characterizing the soot concentration ( 203 ) ( 201 ) with associated signals which represent the soot concentration ( 101 ), the soot concentration ( 203 ) then being derived ( 103 ) from determined input data ( 201 ) using the trained neural network ( 202 ). 2. The method according to claim 1, characterized in that the soot concentration ( 203 ) characterizing input data ( 201 ) include the cylinder pressure characterizing quantities. 3. The method according to claim 2, characterized in that the time characteristic of the cylinder pressure belongs to the input data ( 201 ) which characterizes the soot concentration ( 203 ). 4. The method according to claim 3, characterized in that the signal of the cylinder pressure is filtered. 5. The method according to any one of claims 1 to 4, characterized in that the input data ( 201 ) characterizing the soot concentration ( 203 ) include the combustion profile, temperature profile and / or total combustion profile. 6. The method according to claim 5, characterized in that the course of burning, temperature and / or Total combustion history can be determined by a combustion history calculation. 7. The method according to any one of claims 1 to 6, characterized in that the input data ( 201 ) characterizing the soot concentration ( 203 ) includes at least one of the following variables:

• 1. Geometry of the piston
• 2. Geometry of the combustion chamber
• 3. Injection pressure
• 4. Injector geometry
• 5. position of the piston during injection,
• 6. mean mass temperature,
• 7.speed,
• 8. Last,
• 9. fuel-air ratio (λ),
• 10. residual gas,
• 11. Start of burning.

8. The method according to claim 7, characterized in that on the basis of the trained network ( 202 , 204 ) a selection is made from the listed variables ( 201 ) ( 205 ), on the basis of which there is a suitable assignment of the soot concentration to the input data. 9. The method according to any one of the preceding claims, characterized in that one or more main components as input data of the corresponding input data can be used.