DE102005001428A1 - Control system for internal combustion engine using common rail fuel injection system allows first and second injections of fuel into each cylinder with second injection controlling mixture strength - Google Patents

Abstract

The internal combustion engine control system uses first and second injections of fuel into each cylinder. The first injection (Q1) may be a pilot injection of a constant amount of fuel. The second injection (Q2) may be variable to control mixture strength and power developed by the engine. The engine has common rail fuel injection system and the pressure (P rail) in the common rail may be measured. The time difference (t diff) between the two injections may be measured. The measured input quantities (110) may be fed into a neuronal computer network (120), containing first (121) and second (122) neurons. An output value (dQ2) is produced.

Description

The The invention relates to a method for controlling an internal combustion engine according to the generic term of the independent claims 1 and 2.

One Such method is preferably used in common rail systems. In such common rail systems, the fuel injection a cycle in at least a first partial injection and in at least split a second partial injection. Through these injections arise line pressure oscillations, the so-called pressure waves, systematically the amount of subsequent injections influence.

This oscillation phenomenon decreases with increasing distance between the injections. Of the Influence on the amount of subsequent injections thus decreases with increasing Distance also decreases and approaches for enough long distances the undisturbed Amount that would be obtained with an isolated injection.

These Effects are systematic in nature and essentially depend on that time interval of the injections involved, the injected quantities the injections involved, the pressure of the injected medium, the so-called rail pressure, and the temperature of the injected Medium between the rail and the nozzle from.

When Print size is preferably used in a common rail system, the rail pressure. As fuel quantity can a fuel amount, a drive time for a corresponding actuator or another characterizing the amount of fuel to be injected Size used become.

A method and an apparatus for controlling an internal combustion engine, in which the fuel injection of a cycle is divided into a first and second partial injection and in which to avoid occurring in common rail systems pressure fluctuations, the fuel quantity to be injected is reduced from the DE 101 23 035 A1 known. Here, in the second partial injection, a fuel quantity which characterizes the fuel quantity injected during the second partial injection is corrected depending on at least one pressure variable characterizing the fuel pressure, the fuel quantity and at least one further variable, for example the time interval of the two injections and / or the temperature of the injected fuel quantity, and thereby significantly reduces the influence of pressure fluctuations on the subsequent partial injection.

at This known pressure wave correction is the at a reference system measured quantity influence of a causing injection on the subsequent injection into Maps filed. about a correction amount becomes the injection duration of the subsequent injections or a variable that characterizes it.

The problem is that the functional relationships between a variable characterizing the time interval between the adjacent injections, the quantities characterizing the quantities of the injections involved, a variable characterizing the pressure of the injected medium, and the temperature of the injected fluid between the rail and the nozzle Output quantity, that is, the correction amount, despite a system are highly complex. This is because there are strong interactions between the input variables. For this reason, for example, from the DE 101 23 035 A1 Significant simplifications are made known methods known, for example, a neglect possible non-significant dependencies between the input variables to map the pressure wave phenomenon using less maps. Due to these simplifications, the accuracy of correction is already considerably reduced due to the structure-inherent simplifications.

Task of invention

Of the The invention is therefore based on the object, a method of control an internal combustion engine to convey, in which the influence of the input variables the correction amount with as possible all interactions so that the to be determined correction amount can be determined more accurately.

Advantages of invention

This object is achieved in a method for controlling an internal combustion engine of the type described above, in which the fuel metering is divisible into at least a first and a second partial injection, characterized in that in the second partial injection, a fuel quantity, which injects the injected during the second partial injection correction Fuel quantity characterized, depending on a time characterizing the time interval of the two injections size, each of the injected quantities characterizing quantities of the successive injections, of a fuel characteris size and of a variable characterizing the temperature of the injected medium by means of a neural network.

By this neural network, which is the system of pressure waves / mass waves learned from the measurements for the pressure wave correction (so-called Learning patterns), it is possible the System of pressure wave processes in the common rail system and the dependencies from the input variables without external intervention in the network structure. This also allows complex and hidden relationships recognized from the given learning patterns and in the network structure be filed. In this way is also an associative storage the information possible. There the information in neural networks is content-related, that is associative, stored neural networks have the ability to analyze input patterns that differ from the learning patterns and the associated Output with high Predict accuracy. For this reason, only a part of the later operating range used for learning the network, resulting in a reduction of the administration effort.

The The object is further by a method for controlling a Internal combustion engine solved by that at the second part injection, a fuel quantity, which the amount of correction fuel injected at the second split injection characterized on the basis of a damped vibration with at least a fundamental and at least the first harmonic, whose amplitudes, losses, the frequencies and the phases characterizing parameters in dependence of each of the injected quantities characterizing quantities of successive injections, one of which is the fuel pressure characterizing size and of one characterizing the temperature of the injected medium Size by of a neural network.

at This method is in contrast to the former method no longer the time course of the quantity waves through the neural network in a sense "learned", but only the relationship between the parameters describing the vibrations, So the fundamental and the first harmonic, and the the injected quantities characterizing quantities as well as the fuel pressure characterizing size. It will Accordingly, it has already been assumed that the mass waves oscillate with a fundamental and a first harmonic, so that only nor the variables characterizing the vibration in connection with a reduced number of input variables the neural network must be brought.

There the characterizing the time interval of the two injections Size in this process not considered will reduce the network complexity and thus the required Computing power for the evaluation of the network considerably.

bestimmt, wobei a₁ die Amplitude, d₁ die Dämpfung, f₁ die Frequenz und p₁ die Phase der Grundschwingung und a₂ die Amplitude der Oberschwingung, d₂ die Dämpfung der Oberschwingung, f₂ die Frequenz der Oberschwingung und p₂ die Phase der Oberschwingung sind.Preferably, the amount of fuel quantity that characterizes the correction amount of fuel injected at the second split injection is based on the following equation

determines, where a ₁ the amplitude, d ₁ the damping, f ₁ the frequency and p ₁ the phase of the basic vibration and a ₂ the amplitude of the harmonic, d ₂ the damping of the harmonic, f ₂ the frequency of the harmonic and p ₂ the phase the harmonic are.

One particular advantage of using neural networks lies in their Robustness against noisy data. Neural networks are less sensitive on noisy data as a conventional method.

One greater Advantage is the ability to learn neural networks. These must are not programmed, but learn their optimal topology with the help of a given, as possible huge Number of patterns of the input variables and associated output quantities themselves. The neural network thus forms the system of pressure wave processes in the common rail system and the dependence from the input variables without external intervention in the network structure.

advantageous Embodiments of the method are the subject of the dependent claims.

So For example, in one embodiment, the neural network has at least one neuron on. As the number of neurons increases, the accuracy of the neuron increases Correction of the pressure wave influence.

The Neural networks are preferably stochastic based specific learning patterns. The learning algorithm is based on this conducted Measurements.

It has proved to be advantageous to carry out the learning and optimization of the neural network outside a control unit by means of a computer, preferably by means of a PC. It is very advantageous that the neural networks can be implemented as software within known, existing control devices, so that additional circuitry is unnecessary.

drawing

Further Advantages and features of the invention are the subject of the following Description and the drawing of an embodiment the invention.

In show the drawing:

1 schematically a first, a neural network having circuit for carrying out a variant of the method according to the invention;

2 a representation in which measured, conventionally calculated as well as according to the inventive method quantity waves over the time interval of two injections are plotted and

3 schematically a second, a neural network having circuit for carrying out a further variant of the method according to the invention.

description an embodiment

A method for controlling an internal combustion engine is described below with reference to a in 1 schematically illustrated having a neural network circuit explained in more detail.

The neural network includes input variables 110 , two a neural network 120 forming neurons 121 . 122 and an output size 130 ,

The input variables are, for example, the rail pressure p _Rail , the injected quantities of the successive injection quantities Q1 and Q2 and the time difference of the successive injections t _diff . Furthermore, the temperature of the injected medium T _DWK can also be provided as a further input variable.

each is the input variables connected to each of the two neurons via "synapses".

The output of the first neuron 121 is further connected to an input of the second neuron 122 connected via another synapse.

As output of the second neuron 122 is output a correction amount dQ2, which is characteristic of the change in the subsequent injections, so for example for the energization of a known per se, not shown injector for injecting a fuel quantity.

• a1) Lernvorgang/Generierung des neuronalen Netzes Das Lernen und Optimieren des neuronalen Netzes erfolgt außerhalb eines Steuergeräts des Fahrzeugs an einem PC. Es werden zunächst Lernmuster erstellt und in ein passendes Format gebracht. Die Lernmuster werden stochastisch gemischt, um eine für den Lernalgorithmus störende Systematik, die auf der Reihenfolge durchgeführter Messungen basiert, zu eliminieren. Auf der Basis der zur Verfügung gestellten Lernmuster wird das Netzwerk aufgebaut und eingelernt. Mit steigender Anzahl an Neuronen steigt einerseits die Genauigkeit, mit der die Druckwellenkorrekturmenge dQ2 auf der Basis der Eingangsgrößen vorhergesagt werden kann. Andererseits wird das Netzwerk hierdurch zunehmend komplexer und die Implementierung in einem Steuergerät erfordert damit mehr Rechenzeit und Speicherplatz. Ziel ist es daher, den Algorithmus zur Generierung des Netzes auf eine möglichst geringe Anzahl von Neuronen zu optimieren und das Netz anschließend in einem Fahrzeug auf seine Güte zu überprüfen (sogenanntes "Pruning"). Hierbei werden nicht benötigte Synapsen entfernt, um das Netz zu vereinfachen.
• b1) Überprüfung der Güte des neuronalen Netzwerks Zur Überprüfung der Güte des eingelernten Netzes wird dem Netzwerk analog zum Lernvorgang eine Reihe von Mustern übergeben. Das Netzwerk berechnet nun ohne Kenntnis der Ausgabegröße zu jedem Muster an vorgegebenen Eingangsgrößen die Ausgabegrößen und speichert diese in einer Ergebnisdatei ab. In 2 ist für einen vorgegebenen Betriebspunkt (p_Rail = 1600 bar, Q₁ = 2 mm³, Q₂ = 60 mm³) ein Vergleich zwischen Mengenwellen, die gemessen wurden (mit I bezeichnet), auf der Basis eines aus der DE 101 23 035 A1 bekannten Verfahrens (mit II bezeichnet) sowie mit Hilfe eines neuronalen Netzwerkes (mit III bezeichnet) berechnet wurden, dargestellt. Unter Mengenwellen wird hier die Veränderung der Menge einer durch Druckwellen beeinflußten Einspritzung durch Variation des Abstands der (davor liegenden) Druckwellen verusachenden Einspritzung verstanden. Das hier verwendete neuronale Netzwerk umfaßte 100 Neuronen. Die mit I bezeichneten gemessenen Mengenwellen wurden sowohl zur konventionellen, aus der DE 101 23 035 A1 bekannten Bedatung als auch zum Einlernen des neuronalen Netzes verwendet. Wie der Figur zu entnehmen ist, gibt das neuronale Netz die tatsächliche Mengenwellenform detailgetreuer wieder als die aus der DE 101 23 035 A1 bekannte konventionelle Druckwellenkorrektur. Insbesondere werden im Bereich kurzer zeitlicher Abstände zwischen den Einspritzungen die Mengenwellen durch das neuronale Netz besser abgebildet.
• c1) Implementierung des neuronalen Netzes in einer Steuereinrichtung Das neuronale Netz wird in einem Steuergerät implementiert. Hierzu wird die Funktionalität des eingelernten Netzes in eine Routine der Maschinensprache C umgewandelt und im Steuergerät des Fahrzeugs implementiert.

In order to generate the neural network and finally to be able to use it in the form of, for example, a software function in the control unit, the following steps are necessary.

• a1) Learning process / generation of the neural network The learning and optimization of the neural network takes place outside a control device of the vehicle on a PC. First, learning patterns are created and put into a suitable format. The learning patterns are mixed stochastically in order to eliminate a systematics that interferes with the learning algorithm, which is based on the sequence of measurements carried out. Based on the learning patterns provided, the network is set up and taught in. On the one hand, as the number of neurons increases, the accuracy with which the pressure wave correction quantity dQ2 can be predicted on the basis of the input quantities increases. On the other hand, the network becomes increasingly complex and the implementation in a control unit thus requires more computing time and storage space. The aim is therefore to optimize the algorithm for generating the network to the smallest possible number of neurons and then to check the network in a vehicle on its quality (so-called "pruning"). This will remove unneeded synapses to simplify the network.
• b1) Checking the quality of the neural network In order to check the quality of the learned network, a series of patterns is passed to the network analogously to the learning process. The network now calculates the output variables for each pattern at given input variables without knowing the output size and stores these in a result file. In 2 For a given operating point (p _rail = 1600 bar, Q ₁ = 2 mm ³ , Q ₂ = 60 mm ³ ) is a comparison between mass waves that were measured (denoted by I), based on one of the DE 101 23 035 A1 known method (denoted by II) as well as with the aid of a neural network (denoted by III), shown. Quantity waves are here understood to mean the change in the quantity of an injection influenced by pressure waves by variation of the spacing of the (preceding) pressure waves causing injection. The neural network used here comprised 100 neurons. The measured quantity waves designated I were used both for conventional, from DE 101 23 035 A1 known as well as for teaching the neural network used. As can be seen from the figure, the neural network reproduces the actual quantity waveform in more detail than that from the DE 101 23 035 A1 known conventional pressure wave correction. In particular, in the region of short time intervals between the injections, the quantity waves are better imaged by the neural network.
• c1) Implementation of the neural network in a control device The neural network is implemented in a control device. For this purpose, the functionality of the learned network is converted into a routine of the machine language C and implemented in the control unit of the vehicle.

wobei a₁ die Amplitude der Grundschwingung, d₁ die Dämpfung der Grundschwingung, f₁ die Frequenz der Grundschwingung und p₁ die Phase der Grundschwingung und a₂ die Amplitude der Oberschwingung, d₂ die Dämpfung der Oberschwingung, f₂ die Frequenz der Oberschwingung und p₂ die Phase der Oberschwingung sind.The variant of the method according to the invention described below is based on the empirically obtained finding that the pressure wave triggered by an injection and the resulting quantity wave during the subsequent injection can be represented to a good approximation by a fundamental vibration and a superimposed harmonic according to the following equation:

where a _{1 is} the amplitude of the fundamental, d _{1 is} the fundamental frequency, f _{1 is} the frequency of the fundamental and p _{1 is} the phase of the fundamental and a _{2 is} the amplitude of the harmonic, d _{2 is} the harmonic, f _{2 is} the frequency of the harmonic and p _{2 are} the phase of the harmonic.

• a2) "Fitten" der Meßdaten Das sogenannte Fitten, d.h. Anpassen der Meßdaten, erfolgt wie beim vorbeschriebenen Verfahren außerhalb des Steuergeräts beispielsweise an einem leistungsfähigen PC. Die vorhandenen Lernmuster werden in ein passendes Format gebracht, beispielsweise die Ausgangsgröße aufgetragen über den Eingangsgrößen. Ein Fitalgorithmus wählt dann zuerst die vier Parameter der Grundschwingungen so aus, daß die gewichtete Abweichung, z.B. die quadratische Abweichung, zwischen der Grundschwingung und der anzupassenden Schwingung minimiert wird. Die Grundschwingung wird anschließend von der anzupassenden Schwingung subtrahiert und die erste Oberschwingung in analoger Weise an die Abweichung angepaßt.
• b2) Lernvorgang/Generierung des neuronalen Netzes Der Lernvorgang und die Generierung des neuronalen Netzes erfolgt wie vorstehend unter a1) beschrieben.
• c2) Überprüfung der Güte des neuronalen Netzes Zur Prüfung der Güte des neuronalen Netzes wird dem Netzwerk 320 analog zum Lernvorgang eine Reihe von Mustern übergeben. Das Netzwerk 320 berechnet nun ohne Kenntnis der Ausgabegröße zu jedem durch die Eingangsgrößen 310 gebildeten Muster die Ausgabegrößen 330 und speichert diese in einer Ergebnisdatei ab. Ausgabegrößen sind die Parameter der Schwingungen, die in einer Schaltungseinheit 340 in die gewünschte Größe der Korrekturmenge dQ2 gemäß der folgenden Gleichung
  
  zurücktransformiert werden. Am Ausgang der Schaltung liegt die Größe 350 an, welche die Korrekturmenge dQ2 repräsentiert.

These total of eight parameters are determined from pressure wave correction-Bedatungsmessungen with a so-called known Fitalgorithmus. The complex and analytically difficult to represent relationship between these parameters and the input parameters rail pressure p _rail and the injected quantities of successive injections Q1 and Q2 is in an in 3 schematically illustrated neural network 320 stored here, in contrast to the method described above, not the time course of the mass waves is learned, but only the relationship between the aforementioned eight parameters to describe the two oscillations and the input variables Q1, Q2 and p _Rail . The neural network 320 can be formed by a so-called known Cascade Correl or a Perzeptron network. In particular, the input variable t _diff , which defines the time difference of the successive injections, is dispensed with, as a result of which the network complexity and thus the required computing power for evaluating the network are considerably reduced. To generate the neural network 320 Finally, which is implemented in the form of a software function in the control unit, the following steps are necessary:

• a2) "fitting" of the measured data The so-called fitting, ie adaptation of the measured data, takes place, as in the case of the method described above, outside the control unit, for example on a powerful PC. The existing learning patterns are brought into a suitable format, for example, the output quantity plotted against the input variables. A fitting algorithm then first selects the four parameters of the fundamental so that the weighted deviation, eg, the quadratic deviation, between the fundamental and the vibration to be adjusted is minimized. The fundamental oscillation is then subtracted from the oscillation to be adjusted and the first harmonic is adapted in an analogous manner to the deviation.
• b2) Learning process / generation of the neural network The learning process and the generation of the neural network is carried out as described above under a1).
• c2) Checking the quality of the neural network To test the quality of the neural network is the network 320 pass a series of patterns analogous to the learning process. The network 320 now calculates without any knowledge of the output size to each by the input variables 310 formed patterns the output sizes 330 and saves them in a result file. Output quantities are the parameters of the oscillations that occur in a circuit unit 340 to the desired amount of the correction amount dQ2 according to the following equation
  
  be transformed back. At the output of the circuit is the size 350 which represents the correction amount dQ2.

The basic idea of the method described above is not to "learn" the quantity waves, as it did in the first variant of the method described above, but to start from an oscillation equation of the quantity waves and here the relationship between the eight parameters which describe the fundamental and the harmonic. and determine the inputs Q1, Q2 and p _Rail through the neural network. This reduces the computational effort considerably.

Claims (8)

Method for controlling an internal combustion engine, wherein the fuel metering in at least a first partial injection and a second partial injection can be divided, characterized in that in the second partial injection, a fuel quantity that characterizes the injected during the second partial injection correction fuel quantity, depending on a the time interval of the two injections characterizing size of each of the injected Quantities characterizing quantities of the successive injections, of a size characterizing the fuel pressure and of a size characterizing the temperature of the injected medium by means of a neural network ( 120 ; 320 ) is determined. A method of controlling an internal combustion engine, wherein the fuel metering is divisible into at least a first partial injection and a second partial injection, characterized in that in the second partial injection, a fuel quantity that characterizes the amount of correction fuel injected in the second partial injection based on a damped vibration with at least one fundamental and at least the first, whose amplitude, attenuation, frequency and phase characterizing parameters depend on quantities of successive injections that characterize the injected quantities, on a variable characterizing the fuel pressure and on the temperature of the injected one Medium characterizing size by means of a neural network ( 320 ). A method according to claim 2, characterized in that the fuel quantity quantity characterizing the correction fuel quantity injected at the second split injection is expressed by the following equation

where a _{1 is} the amplitude, d ₁ the attenuation, f ₁ the frequency and p ₁ the phase of the fundamental and a ₂ the amplitude of the first harmonic, d ₂ the attenuation of the first harmonic, f ₂ the frequency of the first harmonic and p _{2 are} the phase of the first harmonic. Method according to one of Claims 1 to 3, characterized in that the neural network ( 120 ; 320 ) at least one neuron ( 121 . 122 ). Method according to Claim 1, characterized in that the neural network ( 120 ; 320 ) is built on the basis of stochastically determined learning patterns. Method according to one of the preceding claims, characterized characterized in that Learning algorithm performed on Measurements based. Method according to one of the preceding claims, characterized in that the learning and optimization of the neural network ( 120 ; 320 ) takes place outside of a control unit by means of a computing unit, preferably by means of a PC. Method according to one of the preceding claims, characterized in that the neural network ( 120 ; 320 ) is software implemented within a controller software.