DE19706750A1 - Method for controlling the mixture in an internal combustion engine and device for carrying it out - Google Patents

Abstract

A process for controlling the mixture in an internal combustion engine has the following steps: (a) at least one value linked to the air volume that reaches a combustion chamber of the internal combustion engine is measured (so-called input value); (b) at least one output value which controls the amount of fuel to be supplied is determined depending at least on the input value(s) measured under (a), by means of stored representation information; (c) the amount of fuel is supplied according to the output value under (b); (d) a value which gives information on the thus obtained mixture is measured (so-called real value); (e) a deviation between the real value measured under (d) and a nominal value is determined; (f) the stored representation information is altered depending on the deviation determined under (e) for the state of operation measured under (a), so that when steps (a) to (e) are carried out in the future in the same state of operation, the deviation is reduced. This is implemented by a learning process. Also disclosed is a device for carrying out this process.

Description

The invention relates to a method for controlling the mixture in a Internal combustion engine and a device for carrying it out.

Mixture control not only plays an important role in the operating behavior of a Internal combustion engine, but is crucial for achieving lower emissions of harmful exhaust gases. This is how it is transferred to gasoline engines Three-way catalytic converter only satisfactorily harmful exhaust gases (in particular Hydrocarbons, carbon monoxide and nitrogen oxides) in harmless Reaction products if the so-called lambda value (λ) of the mixture within very narrow limits is λ = 1 (stoichiometric ratio). Even with diesel engines certain variable must be achieved to achieve low emissions and high efficiency Mixing ratios can be achieved very precisely. Given the increasing problems in the area of air pollution control and the increasingly strict Emission control regulations therefore play an important role in mixture control to.

In the conventional systems, the mixture is adjusted i. a. by Feedforward control and a higher-level control. Namely, z. B. via an im Exhaust gas flow arranged lambda probe the remaining residual oxygen content in the exhaust gas measured. The desired stoichiometric value of the air-fuel mixture (i.e. λ = 1) leads to a certain residual oxygen content. When deviating from λ = 1 shows a larger or smaller residual oxygen content. The lambda sensor generates a corresponding output signal. This serves as the actual value for a PI controller, which changes the amount of fuel to be injected so that the stoichiometric or any other desired mixing ratio maintained  becomes. The control difference that is unavoidable in the conventional sense To keep it as small as possible, a static map is used operating point dependent z. B. dependent on the throttle valve position and the speed Pre-control of the injection quantity. The maps are at the engine manufacturer in extensive test on test benches. The mean values in the map filed to the series distribution of the decisive for the combustion Compensate engine parameters. As long as there is no control signal, the Mixture control the value for the injection quantity specified in the map. Due to the dead time amounting to several work cycles (until a certain one occurs Injection quantity in the exhaust gas at the location of the lambda sensor), it is this Regulation not possible, the desired after changes in load or speed To maintain the mixing ratio. In practice, this is why Acceleration and deceleration phases completely deactivate the lambda control; the The injection quantity is then only controlled using the map. With dynamic Overall, there are significant deviations from the driving operation Lambda setpoint, which leads to correspondingly high pollutant emissions. Austerity Emissions regulations cannot be complied with or only with difficulty.

In the publication U. Lenz and D. Schröder: Artificial Intelligence for Combustion Engine Control, SAE-Paper No. 960328, February 1996 a method for Determination of the air mass sucked into a cylinder proposed. The Publication by U. Lenz and D. Schröder: Identification of isolated nonlinearities with Neural Networks, GMA Technical Committee 1.4 "Theoretical Methods of Control engineering ", workshop in Interlaken, 1996, relates generally to a so-called Observer approach based on a neural network. In these A method of artificial intelligence, can be used for publications Regulation of internal combustion engines, which describes an "indirect" Represents a regulatory approach by gaining knowledge. (First of all, it should be noted that the In contrast, the present invention is a novel method of the artificial Intelligence teaches that this is a "direct" regulatory approach by the process of Tax law "learns").  

The invention has for its object to provide a method with the above Disadvantages avoided or at least reduced, so that even strict Emission standards can be met. This also includes the provision of a corresponding device.

This object is achieved by the mixture control method according to claim 1, which comprises the following steps:

• a) Measure at least one size with which the in a combustion chamber of the Air mass coming from the internal combustion engine is related (so-called Input variable);
• b) determining at least one controlling the quantity of fuel to be supplied Output variable as a function of at least those measured in a) Input variable / s, with the aid of stored image information;
• c) supplying the amount of fuel according to the initial variable from b);
• d) measuring a quantity which carries information about the mixture thus formed (so-called Actual size);
• e) determining a deviation of the actual variable measured in d) from a target value for this size;
• f) changing the stored image information as a function of that in e) determined deviation for the operating state measured in a), so that at a future execution of steps a) to e) in the same operating state the deviation becomes smaller;

and that realizes a learning process. This includes both the stationary and the dynamic operating states.

Here, the term "measuring" is understood in a broad sense, which the actual physical measurement and possibly deriving a quantity from it. The measured variable can therefore be a directly measured variable or one derived from it Be great. For example, the sensor for the actual size can be a Be a lambda probe, which corresponds to the residual oxygen content of the exhaust gas Emits signal. In this example, "measuring" can also be used to obtain this Signal also the determination of the residual oxygen content and possibly the determination of the  Include lambda values. The i.a. mixture size of interest is the mixture ratio (i.e. the lambda value). It is possible to measure a mixture size directly or one Measurement of an exhaust gas size to make a conclusion about the interested Allow mixture size (e.g. the lambda value).

In short, the invention teaches a learning mixture control which compares the mixture ratio actually contained with a target value and at a Deviation from this, the stored control information is adapted so that at going through the same or similar operating point in the future Deviation is reached. The learning process according to the invention generates one exact representation of the real conditions under all relevant operating conditions, from which the control information is then derived to avoid any control difference becomes; this applies in particular to the dynamic operation of the internal combustion engine. So it's about z. B. using artificial intelligence methods to learn from mistakes, to avoid this in the future.

The mixture formation is explained in more detail below:
In the case of piston engines, performance, consumption and emissions are decisively influenced by the mixture ratio (also air ratio) λ. This mixture ratio λ is defined as the ratio of the air mass (m _acyl and fuel mass (m _fcyl ) in the engine cylinder (combustion chamber),

where K _{λ is} the factor for stoichiometric combustion (K _λ = 14.7 for gasoline). The task of the engine control is therefore to optimally adjust the mixture ratio depending on the operating point according to the above criteria. For example, for an emission-optimized operation of a gasoline engine with a three-way catalytic converter, an air ratio of λ = 1.00 can be achieved, which corresponds to an exactly stoichiometric mixture ratio. In contrast, diesel engines are operated with a variable, lean mixture, ie with a λ greater than one; at full load, the mixture is enriched down to the so-called smoke limit at around λ = 1.5.

The following illustration of the mixture formation refers to an example Internal combustion engines with direct injection. As will be explained below, this is Invention can also be used advantageously in other engines.

Direct injection in diesel engines has been known for some time. Which includes also diesel engines with an ante or vortex chamber, if these are effectively part of the Combustion chamber. Gasoline direct injection for gasoline engines is more than 50 years ago Once in series production for years, Daimler Benz military aircraft engines were included equipped with a mechanical direct injection. After the Second World War one mechanical direct fuel injection in the Mercedes 300 SL and sports car Lloyd small car used, but ultimately the mechanical effort unprofitable. Given the increasing environmental protection requirements, such are Engines that are expected to have low fuel consumption are up to date again.

The advantage of direct injection is compared to a conventional one Intake tube engine in the exact metering of the fuel without the dynamic Storage effect through the so-called wall film formation there.

It is expected that the efficiency of gasoline engines in the partial load range in addition to Direct petrol injection by replacing the throttle valve by means of independent controllable inlet valves (hydraulic, electromagnetic) can be increased.

In the following, the mixture formation is exemplified for various direct-injection piston internal combustion engines described mathematically.

The mixture ratio λ in the piston internal combustion engine is determined by the Air and fuel mass in the cylinder. It is therefore favorable to consider the Mixture formation as the intersection of two paths, the air and the fuel path.  

First, the intake behavior, the so-called air path, is to be described: the engine sucks fresh air out of the intake manifold when the intake valves are open and the piston is running down. As a result of the so-called choked-flow effect, the inflow speed of the air through the inlet valves into the cylinders can reach a maximum speed of sound. As a result, there can be no complete pressure equalization between the intake manifold and the cylinder during the intake cycles, which become shorter with increasing speed; the cylinder therefore contains less air mass than would be possible in the thermodynamic intake manifold state according to its volume. Against this background, the suction behavior of the engine can be described mathematically as a pump with non-linear efficiency, the air mass (m _acyl ) that enters the cylinder per intake stroke results in

With:
T _m = temperature of the air mass
p _m = pressure of the air mass
V _D = stroke volume of a cylinder
The dimensionless efficiency "volumetric efficiency" η _vol is nonlinearly dependent on the intake manifold pressure and the speed of the engine (which determines the duration of the intake stroke) in engines with a throttle valve (gasoline engine, conventional, vacuum cleaner or turbo)

η _vol = η _vol (p _m , n) (3)

or

η _vol = η _vol (p _m , n, α _cs ) (4)

or

η _vol = η _vol (p _m , n, α _sr ) (5)

if the engine is equipped with a camshaft that can be adjusted by α _cs or with a switching intake manifold (α _sr ) (ie an intake manifold that is adjustable in effective length).

If the intake valves are driven independently of a camshaft, that is, the load control occurs, for. B. on the duty cycle a _{a of} the inlet valve opening, the volumetric efficiency is determined by this duty cycle and the speed, ie

η _vol = η _vol (a _a , n) (6)

According to equation 2, the air mass in the cylinder after the intake stroke is also determined by the pressure and the temperature in the intake manifold. If the engine is not equipped with a throttle valve, these thermodynamic conditions correspond to those of the environment, it can be assumed that these fluctuate only slowly, with an adaptation the measurement of both variables can be omitted. If, however, the internal combustion engine is operated in a throttled manner, both variables can quickly change. B. also pressure and possibly temperature in the intake manifold to determine the air mass in the cylinder. The non-linear dependence therefore results for the air mass sucked into the combustion chamber per work cycle

with volumetric efficiency depending on intake manifold pressure and speed or Duty cycle and speed. For switching suction pipes or more variable Camshaft controls also depend on the dependence of volumetric efficiency from these control variables.

The fuel injection, the so-called fuel path, is described below:

Analogous to the intake behavior, the fuel injection can be described by means of an efficiency. If ideal, inertia-free opening and closing of the injection valves and constant gasoline supply pressure were assumed, the fuel mass injected into the cylinder per cycle would be calculated at the duty cycle a _f

m _fcyl = Ka _f (8)

with a constant K, determined by valve opening cross-section and fuel supply pressure. Due to the dynamics of the valve needles and a low supply pressure loss upstream of the valves, an efficiency η _{f will also be} introduced here for the description. This efficiency depends non-linearly on the opening duration, i.e. on the duty cycle of the injection valves a _f and the control frequency 1 / n. This results in the non-linear dependency

NL _f = m _fcyl = η _f (a _f , n) Ka _f (9)

for the fuel mass injected into the cylinder per work cycle.

According to equation 1 and the above derivations for the air path and the fuel path, the mixture ratio (the air ratio λ) is thus determined

In Fig. 1, this relationship is exemplified for a piston internal combustion engine with conventional intake valve control. Namely, Fig. 1 shows a "signal flow" for formation of the mixture ratio in direct-injection internal combustion piston engines with conventional inlet valve control. The additional (thinly depicted) dependency η _vol = f (.,., Α _cs ) applies to engines with variable camshafts (petrol and diesel engines). For a free-sucking diesel engine, for example, p _m ≈ p ₀ (p ₀ = ambient pressure) or for a throttled gasoline engine p _m ≦ p ₀ . In turbo engines, the intake manifold pressure can also exceed the ambient air pressure. If the piston internal combustion engine under consideration is equipped with independently driven intake valves instead of a more conventional intake valve control, the volumetric efficiency depends on e.g. B. from the duty cycle a _a and the speed n. The load can then be controlled by varying the duty cycle, the motor draws in unthrottled. The mixture ratio is then z. B. formed according to the illustration of FIG. 2, which shows a "signal flow diagram" for the formation of the mixture ratio in direct-injection piston internal combustion engines with camshaft-free intake valve control.

When the internal combustion engine is operating, the general task now is to set a desired mixture ratio by means of a control system. This mixture ratio is determined from the quotient of the air and fuel mass in the cylinder. Both the air and fuel mass in the cylinder cannot be measured and are also not easily adjustable (e.g. by calibration, time variance) by varying the control variables p _m or a _a and a _f . However, the intake air mass or the injected fuel mass - in the case of the direct injector explained in more detail here by way of example - is always the same for the same control variables at the same engine operating points if the time variance as a result of aging or e.g. B. slow clogging of a valve is neglected. Both NL _a and NL _f are static (possibly time-variant) non-linearities and can only be determined by a complex calibration. The mixture ratio can be measured: new methods are based on the interpretation of the ionization current in the electric spark ignition (in the gasoline engine), but the measurement using a faster switching lambda probe is conventional. The measurement by means of a lambda probe (e.g. broadband lambda probe or probe with an abrupt course of the probe voltage at λ = 1) in the exhaust manifold (so-called downpipe) is preceded by a speed-dependent dead time caused by the internal combustion engine working method.

From the above explanations it is clear that the mediated for the inventive method Mapping of the air mass size (input size) to the fuel quantity size  (Output quantity) comprises two nonlinear functions that cannot be predicted exactly, namely that of the air and fuel path. To simplify matters, it was used as an example referred to direct injection engines. As explained in more detail below, this is The method according to the invention can also be applied to non-direct-injection engines which have a so-called fuel storage effect.

According to the invention, the mapping which is not exactly known a priori is not - as in the prior art Technology - mediated by a regulation, but by a learning control. The Learning can be done on the fly, in which the different, in practice occurring operating points (such can be a tuple of a speed value and of an intake manifold pressure value). The inventive method is here in every relevant, d. H. achievable operating point. After all Operating points have been run through several times - which in normal operation i. a. quite fast is achieved - the overall mapping for all possible values of the input variable is learned.

After teaching, the control system delivers instantan - d. H. without any Control delay - the correct value of the output variable with high accuracy even after a change in the operating point. The learning process is advantageous continuously continued to enable a continuous adaptation to disturbance variables which are not considered Input variables are recorded. This can be, for example, wear-related Changes, changes in intake air temperature, cooling water temperature, external Act air pressure, the oxygen content of the air, etc.

The method according to the invention is able, after learning the image Approaching new operating points without ensuring any λ deviation. At In contrast, conventional control approaches always have a control difference which then becomes an integral part of the controller, for example Control variable is integrated until the difference has become zero.

The invention thus has the following advantages:

• - due to the self-adaptation of the mapping information, the Requirements for the accuracy with which the image information for the  Mixture control map must be known. This reduces the effort for the development of an engine control and the series development considerable;
• - The control method according to the invention is robust against series spreads and time-varying disturbance variables;
• - The desired lambda value is not only in stationary operation, but also after a change in the operating state (speed and / or load changes of the Motors) adhered to with great accuracy without any time delay;
• - Overall, this results in lower exhaust emissions and a lower one Fuel consumption.

The invention thus makes a contribution to environmental protection and careful handling with the limited resources.

The invention can also be used in the course of a fault diagnosis Use company. If, in fact, the degree of the required adaptation normal series distribution and disturbance factors can be deduced from this Infer error states, such as excessive wear or a defect. By appropriate evaluation of the degree of adaptation, for example by a vehicle diagnostic system, engine damage or partial failures can be detected early during operation.

In principle, the method can be carried out so that one or more of the Method steps a) -f) averaged over several work cycles or cycles of a cylinder be performed. In a multi-cylinder engine with z. B. a common intake manifold and / or a common lambda probe then contribute to those in steps a) and / or d) measured quantities based on the averaging of several cylinders and possibly several Working games at. However, the method is preferred in time with the work cycles of the performed single cylinder (claim 2), d. that is, the sequence of process steps a) -f) is done once in a single work cycle of a single cylinder carried out. Accordingly, the measurement in step a) takes place during the Intake stroke of a cylinder. In step b) the amount of fuel to be supplied is on the Basis of this measurement (and possibly previous measurements: more on this below) determined. The supply of fuel in step c) then takes place, for. B. in the  immediately following compression stroke of the same cylinder, d. H. in the same working cycle as Step a) related to the same cylinder. The following steps d) -f) are generally due to carried out delayed by dead time effects; however, they become the working set of steps a) -c) and the associated cylinder. Basically, it is possible to follow the procedure not in every work cycle, but e.g. B. only in every second, third, etc. game perform. However, an embodiment in which the method is also particularly advantageous steps a) -f) is processed once for each working cycle of each cylinder.

The most common gasoline engine is a throttled engine, which is common is controlled by adjusting a throttle valve arranged in front of an intake manifold. At In such a motor, the input variable or - if there are several input variables - one of the Input variables advantageous the pressure in the intake manifold (claim 3). This pressure determines namely essentially the cylinder filling. It can be an engine with or without Act charging. With a charged engine (e.g. a turbo or Compressor motor), the pressure in the intake manifold can be temporarily or permanently above the Atmospheric pressure.

Some engine designs make the dynamics of the intake air for you Take advantage of charging. In some of these designs, the intake system is changeable trained to adapt the dynamic charging to different operating conditions can (see. Kraftfahrtechnisches Taschenbuch, Bosch, 1991, p. 373, "Schalt-Ansaugsyste me "). For example, the effective intake manifold length can be adjusted to acoustic Use phenomena to increase filling (key word: resonance charging). With these Systems is advantageously a characterizing the instantaneous position of the intake system Size (e.g. the effective intake manifold length), the input variable or one of the input variables (Claim 4).

A further development of the type of throttled engines has an intake valve control adjustable valve timing. For example, by adjusting the camshaft a shift in the opening and closing times and / or a change in Opening time can be achieved (see automotive paperback, see above p. 370). In In this case, one or more valve timing parameters (e.g.  Camshaft rotation and / or axial displacement) the input variable or one of the Input variables (claim 5).

Another type of motor that could become very important in the future has freely operable ones (e.g. electromagnetically actuated) inlet valves. Such an engine doesn't need one Throttling have more, it can before the inlet valves with atmospheric pressure be acted on, because the valves due to the free selectability of the opening and Closing times can completely take over the power control of the motor. Advantageous the duty cycle and / or the closing and / or opening time of the intake valve is one Input variable or one of the input variables (claim 6). The duty cycle is that on the Duration of the intake cycle or of the entire working cycle related opening period.

It was already explained above that the air mass entering the cylinder depends on the speed of the internal combustion engine. The speed is therefore advantageously an input variable or one of the input variables (claim 7).

As already explained above, gasoline engines with 3-way catalytic converters usually operated with a constant mixture ratio of λ = 1, diesel engines on the other hand with a variable target mixture ratio. The method according to the invention is also suitable for the latter mode of operation with a simple addition, namely by that the role of an input variable is assigned to the target mixture ratio (claim 8th). While with an Otto engine the input size range z. B. is two-dimensional (he is spanned by intake manifold pressure and speed, for example) for a diesel engine as e.g. B. add the target mixture ratio to the third input variable, so that a three-dimensional input size space is spanned here (e.g. by Intake manifold pressure, speed and target mixture ratio).

In the internal combustion engine, fuel is preferably supplied by injection. The output quantity controlling the quantity of fuel to be supplied is then advantageously a or several of the following variables: injection duration, duty cycle of the Injector opening, injection pressure, degree of opening of the injection valve (claim 9). With The duty cycle is the opening time based on the duration of a work cycle  or a work cycle. The opening degree of the injection valve can, for. B. about the Stroke of the valve needle can be controlled.

With the most common Otto engines today, fuel is not injected directly into the cylinders, but into the intake pipe upstream. The fuel wets here first the suction pipe wall (so-called wall wetting effect) and then needed for the transition some time in the gaseous phase. As a result, only a portion of the reached the amount of fuel injected into a work cycle in this work cycle Cylinder. The remainder becomes gaseous later and therefore becomes later Work cycles and - with common injection for several cylinders - partly not associated cylinders burned. The method according to the invention is particularly advantageous can be used with an engine that does not have such a fuel storage effect here no such delay needs to be taken into account (claim 10). Especially this is a direct injection engine. Even engines where the injection into one Antechamber or secondary chamber does not have a fuel storage effect, if after the Combustion cycle practically no fuel remains in the antechamber or secondary chamber.

The lack of a fuel storage effect allows a particularly simple design of the Procedure in which the determination of the output variable (s) in step b) is based only on variables, that belong to the work cycle currently being carried out - in particular to the result of the In the immediately preceding step a), the input variable measurement (claim 11).

However, the method according to the invention is also advantageous in the case of internal combustion engines Fuel storage effect can be used. The one at a certain work cycle in a cylinder fuel quantity depends here i. a. also from that in previous bars injected fuel quantity. It is therefore preferred in such an engine quantities from one or more previous work cycles in the determination of the Output variable / n in step b) included (claim 12). This can be, for. B. the in fuel quantities injected in these previous strokes are those in the current one Clock still impact. Their consideration allows a very precise control of the im current stroke fuel quantity to be injected.  

The stored image information advantageously has the form of a map, which contains the output variable (s) directly or indirectly (claim 13). You get a map For example, by discretizing the (usually multidimensional) input size space and each input quantity cell formed by the discretization has an output quantity or - in the case of a multidimensional mapping - assigns several output variables. The The discretion of the entrance area does not have to be done regularly, just as the Size of input size cells may not be constant. The learning takes place so that in Step f) the stored values of one or more adjacent input quantity cells be adjusted in accordance with a detected deviation so that in a future Operation in the same input size cell a smaller deviation occurs.

The input variable (s) are particularly preferably mapped to the output variable (s) and the adaptation of the stored image information by a neural network Algorithm (claim 14). As a result of the parallelization of data processing in so-called neurons forming the network, the use of a neural network Algorithm the calculation time of the output variable based on the stored Mapping information compared to other interpolation methods (e.g. splines, linear or polynomial interpolation) considerably shortened when using adapted hardware will. While using known interpolation methods already by the Selection of the interpolation (possibly nonexistent) prior knowledge of the approximating connection, a neural network can be completely without this previous knowledge get along. In particular, neural networks are characterized by their evaluation and Adaptation algorithm for the present learnable method can be used very advantageously.

According to this preferred embodiment, a neural network NN is set up in such a way that the control variable of the air path (for example the intake manifold pressure) is mapped onto the control variable of the fuel path in such a way as to achieve a desired mixture ratio. From a historical point of view, the control variable of the air path is also regarded as the control variable with which the driver influences the power output of the internal combustion engine. The adaptive neural network NN is supposed to be the control variable, depending on the driver-influenced control variable for the air path and possibly on the engine operating state (for example speed). B. output a _f for the fuel path. The output of the neural network corresponds to z. B. the duty cycle of the injection valves a _f , the input space of the neural network NN in this example consists of intake manifold pressure p _m and engine speed n, so we write

a _f = NN (p _m , n) (11)

The aim of the control is to have a desired air-fuel ratio for each operating point

to reach. In the following we neglect the constant factor K _λ . With the neural network NN as control according to a _f = a _f (p _m , n) = NN (p _m , n), the above equation can therefore be used as

to be written. If a certain mixture ratio _is λ is to be achieved, so the neural network NN an image must be learned so that m = _acyl λ _to m _Fcyl applies, so

NL _a (p _m, n) = λ _should NL _f [NN (p _m, n), n] (14)

In this case, the output variable is advantageously formed by linking a base value vector representing the stored imaging information and an activation vector dependent on the input variable (s) (claim 15). This link is preferably linear, and in particular has the form of a dot product or - in the case of a multidimensional output variable - a vector matrix product (claim 16). According to the example above, the control variable a _f is then obtained

a _f (p _m , n) = Θ ^T A (p _m , n) (15)

where Θ is the base vector and A (p _m , n) is the activation vector.

It is advantageous that the activation vector is standardized and only depends on the distance of the input variable / n to the support points on which the vector representation is based (claim 17). The standardization condition is z. B.

A (x) ^T 1 = 1 (16)

d. H. the sum of all components of the activation vector is always one.

The dependence of the activation vector A only on the distance of the input variable / n to the support points can be expressed in that its components A _i only depend on a variable that corresponds to the distance to the component belonging to the support point under consideration:

A _i = A _i (d _i ) (17)

with the distances

the support points, d. H. are the locations of the neurons in the entrance room.

An embodiment in which the mapping of the input variable (s) is particularly advantageous the output variable / n is essentially local (claim 18). This means that with one certain value of the input variable (s) essentially only that support value (s)  contribute to the mapping to the initial size / s in the immediate vicinity the input variable (s) lies. This is advantageously achieved in that only the component (s) of the activation vector have appreciably large values maintains / maintains a short distance from the input variable (s) while Components at a greater distance are negligibly small or disappear (An Proverb 19).

The components of the activation vector depend particularly advantageously on the distance of the input variable / n from the associated support point according to a center function (claim 20). Examples of advantageous center functions are

where σ is a (possibly variable) width parameter.

Basically, the "range" of learning does not need that of the figure to match. So it is z. B. possible to design the learning process locally (i.e. in a Only adapt the base value components in the vicinity of the deviation point), on the other hand, to perform the mapping non-locally (i.e. also in the mapping operation Include support components that are further away from the one to be mapped Input value). However, both ranges are advantageously essentially the same chosen, d. H. the mapping information is essentially adapted in the same range of distance from a point of deviation, which is also in the illustration of a input variable at this point is included in the output variable (claim 21).

The mapping information is advantageously adapted essentially locally to the point of deviation (claim 22). This can be done e.g. B. by

Δ Θ | _{pm, n} ≈ Δ Θ (19)

express, where Δ Θ is a base value correction vector. Both the adaptation and the mapping are particularly advantageous locally and with the same range.

More precisely, the mapping information is preferably adapted such that a base value correction vector is added to the base value vector, which is proportional to a link between the deviation value and the activation vector (claim 23). The link is in particular the product of the deviation value with the activation vector. The adaptation takes place e.g. B. according

Δ Θ = ηe (p _m , n) A (p _m , n) (20)

The factor η in this equation represents the learning step size. The quantity e is the measured mixture error at the operating point, the

e (p _m , n) = λ (p _m , n) -λ _soll (21)

is defined as a deviation from the target value assumed to be constant here. When using a broadband lambda probe, the size of the deviation can be inferred, whereas this information is only available in the form of a statement “too rich” or “too lean” in the case of a lambda probe with a step characteristic. Metaphorically speaking, the activation of A (p _m , n) selects the individual base values, which according to the relationship

be adapted.

If a process capable of learning, such as the control of internal combustion engines, is used, the learning process is preferably carried out solely for safety and acceptance reasons, preferably with an adaptation law with demonstrable stability and parameter convergence (claim 24). "Convergence" here means the convergence against a global minimum of the distance between λ and λ _{should be} understood in all operating points approached for learning (not just convergence against a local minimum). Convergence means that the learning process can only be completed when the base vector converges against the only possible but unknown base vector. Reference is made to the following description of proof of stability and convergence of a preferred adaptation law.

In terms of the device, the above-mentioned object is achieved by a device for mixture control in an internal combustion engine according to claim 24, which comprises:

• a) at least one device for measuring at least one size with which the air mass entering a combustion chamber of the internal combustion engine There is a connection (so-called input variable);
• b) at least one control device for supplying fuel;
• c) at least one device for measuring a quantity, the information over the resulting mixture or its combustion (so-called Actual size);
• d) at least one memory for holding the variable Mapping information;
• e) and one programmed to carry out the method according to claim 1 and / or hard-wired computers.

With regard to advantageous configurations of the device, reference is made to the above statements referred to the procedure.

The invention will now be described on the basis of exemplary embodiments and the attached drawing illustrated. The drawing shows:

Fig. 1 shows a signal flow diagram for the formation of the mixture ratio in a direct-injection piston internal combustion engine with a conventional lass control;

Fig. 2 is a signal flow diagram corresponding to Figure 1, but for an unthrottled internal combustion engine with freely operable intake valves.

Figure 3 is a schematic representation of a direct injection gasoline engine with conventional intake valve control.

FIG. 4 shows a signal flow diagram corresponding to FIG. 1, but with a neural network algorithm;

. Fig. 5 is a so-called parallel representation of a neural network algorithm;

Fig. 6 is a Signalflußplandarstellung a neural network algorithm with an input dimension;

Fig. 7 diagrams the soft power output as a function of a one dimensional input value representing, tadaption to illustrate the location of the Stützwer;

Fig. 8 is an illustration of the inter- and Extrapolationsverhalten a Neuronales- network algorithm;

FIG. 9 is an easy learning structure of a neural network algorithm:

FIG. 10 is an illustration of a non-linearity by way of example, adopted for the air path;

FIG. 11 shows a representation corresponding to FIG. 10 for the fuel path;

12 is a diagram of the temporal course of the output variable at the beginning of learning.

Fig. 13 is a diagram according to FIG 12, but in the course of learning.:

FIG. 14 shows a diagram according to FIG. 12, but after the learning process has almost been completed;

FIG. 15 is a diagram of the mixture error during the learning process; and

Fig. 16 is a comparison of the to be learned and the learned relationship.

What is important when using approaches with a learning process is the security concept, So the question of whether learning always leads to the right result without making mistakes; a mathematical proof of stability and convergence should therefore be feasible. In the The following exemplary procedure to be described becomes a special neural Network depending on certain measurable state variables of the internal combustion engine and / or the driver's request, the duty cycle of the injection Output direct-injection (solenoid) valves as a control variable. It makes sense the neural network is used as a corrective element in addition to the previously determined stationary knowledge, which is adopted online so that the desired mixture ratio is always adapted becomes.

With regard to the formation of mixture ratios, reference is made to FIGS. 1 and 2 and the associated previous statements.

The following presentation of exemplary embodiments is initially intended to take place on a direct-injection gasoline engine with conventional intake valve control with fixed control times according to FIG. 3.

In this type of combustion engine, the load is controlled via the position angle of the throttle valve: the driver uses it to control the air mass flow _at into the intake manifold. When the inlet valves are open, the cylinder sucks in the air mass flow _av during the downward movement of the piston until the fresh air mass m _acyl in the cylinder is available for combustion after closing. The fresh air mass in the cylinder depends crucially on the thermodynamic intake manifold pressure p _m according to equation 2. Ultimately, the driver uses the throttle valve position to determine the air pressure in the intake manifold and thus indirectly the load in the cylinder.

From the above, it can be seen that the air mass m _acyl in the combustion chamber after the intake valves have closed is non-linearly dependent on the one hand on the engine speed and on the other hand the intake manifold pressure.

The engine speed determines that available for intake Time period, so that as a rule of thumb applies that at lower speed more fresh air same thermodynamic state as in the intake manifold.

Due to the constructive design of the intake manifold and the intake manifold Effect achieved by acoustic phenomena (interference) in certain Speed ranges a higher filling is achieved. This will be at Internal combustion engines with variable intake manifold geometry to increase torque exploited.

When the piston rapidly descends with the inlet valves open, there is inside the cylinder opposite the intake manifold a large vacuum, so that the cylinder is out  the suction pipe. According to the thermodynamic flow equation for compressible media, however, the medium flowing through cannot be higher than that Reach the speed of sound so that the mass flow through the inlet valve for different operating ranges of the engine with increasing pressure difference no longer can increase further (Laval ratio, choked-flow effect).

Overall, the intake air mass _acyl in this example depends non-linearly on the intake manifold pressure and the speed, in the case of variable intake geometry it also depends on its control angle or that of the adjustable intake camshaft.

So we write mathematically

m _acyl = NL _a (p _m , n) (23)

Analogous considerations can be made for direct fuel injection: As a result The inertia of the valve needles also results in a non-linear one Relationship between injected fuel mass and duty cycle with the pulse frequency (reciprocal of half the speed for four-stroke engines). Another Effect that affects the injected fuel mass nonlinearly is the drop in Fuel supply pressure when opening the valve: this also creates in the Fuel supply system pressure fluctuations, which the injection behavior influence depending on the speed. Overall, the injected fuel mass depends these considerations are statically nonlinear from the valve opening time and the Injection frequency (2 / n).

So we write mathematically

m _fcyl = NL _f (a _f , n) (24)

The idea for mixture control is now to create a neural network NN like this assume that the control variable of the air path (here the intake manifold pressure) on the Control quantity of the fuel path is mapped in such a way so as to achieve a desired one To achieve mixture ratio.  

The adaptive neural network NN should output the control variable (here a _f ) for the fuel path depending on the engine operating state (here the speed) and the driver-influenced control variable for the air path (here the intake manifold pressure).

The output of the neural network corresponds here to the duty cycle of the injection valves a _f , the input space of the neural network NN consists of intake manifold pressure p _m and engine speed n, so we are writing

a _f = NN (p _m , n) (25)

The aim of the control is to have a desired air-fuel ratio for each operating point

to reach. In the following we neglect the constant factor K _λ . With the NN as control according to a _f = a _f (p _m , n) = NN (p _m , n), the above equation can therefore be used as

to be written. If a certain mixture ratio _is λ be reached, in NN an image must be learned so that m = _acyl λ _to m _Fcyl applies, so

We use examples of a variety of adaptive neural network architectures here the so-called DANN, which makes the chosen approach particularly plausible; furthermore, this algorithm allows simple evidence for stability and Parameter convergence as well as meaningful interpretability of what has been learned Knowledge. The DANN is a particularly optimal design of a computing time so-called RBF network with local support effect. A detailed description the THEN follows below.

So we start

a _f (p _m , n) = Θ ^T A (p _m , n) (29)

the base values Θ of the neural network must be adapted in order to compensate for both non-linearities in equation 28 at the operating point (p _m , n).

This is supposed to be with the adaptation approach

Δ Θ = ηe (p _m , n) A (p _m , n) (30)

done with guaranteed stability. The factor η in this equation represents the learning step size. Metaphorically speaking, the activation of A (p _m , n) selects the individual base values, which according to the relationship

be adapted. Now the basic values at DANN mainly work locally, we can without major errors

(the size to the right of the corresponding sign is the component of the Θ vector that is closest to the position (P _m , n)).

We are replacing because of the discrete-cycle working method of the piston internal combustion engine in this derivation the continuous derivation (according to time) through the notation with Delta sizes, which is a derivation after the work cycles of the internal combustion engine should correspond.

In order to provide the stability proof mathematically, we examine the effect of this local adaptation according to equation 32 on the mixture error e at the operating point that

e (p _m , n) = λ (p _m , n) -λ _soll (33)

is defined as a deviation from the target value assumed to be constant here. So we decide

from which with equation 27

follows. First, let's consider the first quotient on the right side from the above equation:

is less than or equal to zero! This can be explained quite simply: with constant air mass in the cylinder, the air ratio λ decreases with increasing injected fuel _mass m _fcyl , the mixture becomes richer. And with increasing duty cycle a _f , that is to say a longer valve opening duration at a constant speed, more fuel is of course also injected, but at least as much; less would be absurd. So it applies according to equation 27

because yes

is of course greater (or possibly equal) to zero.

Now we consider the second quotient on the right side of Equation 35: From the properties of the neural network THEN follows

which means that when a base value is adjusted upwards at a certain point in the entrance area of the DANN / GRNN, the output of the network naturally also increases at this point. So that applies

These derivations are possible because in the case of base value adaptation because of the locality of the base value effect of these networks

applies.

We can now summarize:

and follows with the adaptation law

So that's the relationship

Δe (p _m , n) = -kηe (p _m , n) k ≧ 0 (43)

derived with which the stability of the adaptation approach 18 is proven. To we interpret equation 43 verbally:

The sign of the error change is always inverse to the sign of the error; With in other words, if the error is less than zero (e <0), its value increases (Δe <0) The amount becomes smaller. If the error is greater than zero, its value falls, and here too thus the amount is smaller. The error can therefore only converge towards zero.

In addition to the convergence of the mixture error at the operating point to zero, the properties of the DANN also result in the convergence of the base values in the learned operating points against the only possible base values with which λ = λ _{should be} achieved. This derivation applies locally, for the working point (p _m , n) of the internal combustion engine. It is logical, again because of the locality of the base value effect, that the mixture errors are only compensated with high accuracy at these operating points, in which the control system can also learn.

There is also a learning correlation to take into account, namely dead time due to the four-stroke process. In the derivations above, we assumed that when adapting the mixture, the resulting mixture error can be used to adapt the injection control. As a result of the four-stroke working method of the internal combustion engine, however, when using an exhaust gas lambda probe, a measured value for the air ratio is only available after three clocks of dead time (compression, combustion, pushing out). This is shown in Fig. 4. This figure shows a signal flow diagram for the formation of the mixture ratio in throttled direct-injection piston internal combustion engines. The additional (thinly depicted) dependency η _vol = f (.,., Α _cs.sr ) applies to engines with variable camshafts or intake manifolds. There is a dead time of three cycles between mixture formation and lambda measurement due to the four-cycle working process of the internal combustion engine.

For learning correlation, we must therefore take this dead time into account when adapting. This is expediently done by simply delaying the input values (p _m and n) during the adaptation, so that the error signal and the base value to be adjusted correlate with one another. In contrast, the undelayed input values must be used to form the injection signal (a _f ).

The example described above relates to the synthesis of a control law for injection in direct-injection internal combustion engines. The description was based on a throttled direct petrol injector with a constant target mixture ratio. If the target mixture ratio can assume a variable value, for example in the case of diesel engines, this target value can be provided as an additional input dimension of the DANN. Likewise, in the case of intake valves driven without camshafts, their control signal (duty cycle a _a ) can be used as a network input instead of the intake manifold pressure. In a real implementation, it makes sense to use the neural network as a correction element in addition to e.g. B. off-line certain prior knowledge in the form of a linear mapping p _m → a _f . This can significantly reduce the learning time.

The following table gives an overview of the nomenclature used here:

In the following, an example of a suitable learnable neural network Algorithm explains the so-called DANN in more detail. It combines the advantages standardized RBF networks with the advantage of computing time and storage space optimization.

When using the so-called Distance Activation Neural Network DANN according to FIG. 5, which shows a parallel representation of a DANN with two input dimensions, parameter convergence is guaranteed by the local support value effect during the learning process and the problem of so-called persistent excitation is avoided. Parameter convergence means that the neural network can only approximate a certain static non-linearity with one parameter set. Accordingly, there is an unambiguous assignment of the learned context and the basic values of the NN, which makes the knowledge learned interpretable. This decisive advantage in comparison to other algorithms capable of learning is based on the shape of the RBF network, especially the DANN: the parameters to be adjusted during learning mainly act locally and the network thus provides a continuous output function over the input space with defined interpolation behavior. The extrapolation behavior of the DANN at points where knowledge is available in the form of learned parameters is just as defined as it is sensible: the output of the network corresponds to a (weighted) mean value of the learned knowledge in the learned environment; the DANN differs considerably from the original RBF network. The network consists of locally activated neurons, ie mainly the neurons in the immediate vicinity of the network input χ are activated. The structure of the DANN can be divided into activation and weighting. This illustrates the signal flow diagram representation of a DANN shown in FIG. 6 with an input dimension (ie scalar input x,) A x) and ϑ are vectors; (x) = ϑ ^T A (x). This makes it accessible for the so-called delayed activation process during learning.

The algorithm of the DANN will now be briefly described: A scalar estimate at a point x ∈ ⁿ with a given set of q data points (base values ϑ _i at base points ζ _i ) (ϑ _i ζ _i ), i∈) [1, q] and At THEN, beim ∈ ⁿ is defined by the equation

with the distances

Equation 44 guarantees the limitation of activation A (x) by which normalization becomes

A (x) ^T 1 = 1 (46)

achieved, that is, the sum of the activations of all neurons is always equal to one; Similarly, the activation of a single neuron is always between zero and one. From equation 44 it is clear that with a very small smoothing factor a almost only one neuron is activated; almost only one base value thus contributes to the result of the evaluation. If a base value is adjusted, the network output is only changed in its surroundings. We call this state of affairs "locality" of the base value effect. To illustrate, FIG. 7 shows the location of the supporting values action, the network output changes only in the vicinity of the reclined support value. This finding is important for the proof of stability of learning in the method according to the invention.

Equation 44 defines a continuous, arbitrarily non-linear output function = f ( x ) (denotes estimated or adjustable variables) with a defined inter- and extrapolation behavior, as shown in FIG. 8. This figure shows the inter- and extrapolation behavior of the DANN, the crosses indicate the existing data points. In the learned area, the curve approximated by the base values (crosses) matches the sine function to be learned. Beyond the area learned, the estimated value tends to the average of all available knowledge (base values) with increasing distance from the next base value. The estimated value at an evaluation point at a great distance from the stored knowledge (data points) will result from the mean value of the existing knowledge; in the immediate vicinity of a data point, this mainly determines the network output. If the weighting values ϑ _{i are} defined as adjustable, the simplest online structure for learning, shown in FIG. 9, can be derived. This representation applies to continuous-time systems. The THEN shown can learn all static (without internal states such as memory) up to discontinuities (grouped activation method) apart from a small approximation error due to the finite number of base values. In the case of images with memory (e.g. because of the memory effect mentioned above), information about previous events is included in the learning and mapping behavior. The learning structure shown in FIG. 9 is based on a known mathematical error model, for which stability by the direct method according to Ljapunov has been proven.

A one-dimensional non-linearity to be learned as in FIG. 9 is to be represented in the DANN form as the dot product y (t) = ϑ ^T A (x (t)) + d, where x (t), y (t)) and A , ϑ ∈; the constant vector ϑ is the unknown parameter vector of dimension n to be learned, whereby n base values distributed over the input space can represent the real non-linearity to be learned except for the error d → 0 with n → ∞. 9, the observer error is calculated as e (t) = (t) -y (t). If a parameter error vector is defined as Φ (t) = (t) - ϑ , the error equation results

According to the fault model 1

be chosen as a guaranteed stable adaptation law, because of the properties of the THEN (Equation 44) the vector A (χ (t)) is restricted in all its components.

This leads to a monotonically decreasing function | Φ (t) |, which means lim _{t → ∞} (t) = ϑ .

In the present exemplary embodiments, the learning ability shown is the THEN used for the synthesis of a control law for fuel injection by at Learning the error between the target mixture ratio and the actual mixture ratio to zero is made.

The above exemplary embodiment is now to be demonstrated in a simulation: the ability of the approach is to be shown that a mapping can be learned so that applies

NL _a (p _m, n) = λ _should NL _f [NN (p _m, n), n] (49)

For the demonstration let λ _be equal to 1. For the purpose of illustration, the static steady non-linear relationship becomes for NL _a (p _m , n)

NL _a (p _m , n) = [1-exp (-4p _m )] exp (-n) (50)

suppose for NL _f (a _f , n) we take

an, a nonlinear function that increases monotonically with a _f . These two non-linearities assumed for the air path and the fuel path are shown in FIGS. 10 and 11.

If these dependencies are known, the mapping NN _target can be calculated analytically, so that equation 49 with NN = NN target _is fulfilled. With the relationships according to equations 50 and 51, NN _is obtained for the mapping

which can be learned through controller synthesis.

Speed n is normalized to the maximum speed, as is the intake manifold pressure to ambient pressure, these variables therefore both fluctuate between 0 and 1. For the learning simulation, the input space n ∈ [0.2,0.8] and p _m ∈ [0.2,0.8] are traversed throughout, in Fig . 16 this is clear from the learned control surface. In FIGS. 12, 13 and 14 here is to learn the control flow according to Equation 52 and the online learned from the network history of a _f displayed during the learning process. These figures each show the optimal course of the control variable (dashed line) and the course learned from the network (solid line). FIG. 12 shows the beginning of learning (with no previous knowledge being started), FIG. 13 shows the course of the control variables during learning, and in FIG. 14 the learning is practically complete. Fig. 15 shows the mixture error that quickly decreases during learning.

The case learned in the neural network connection, FIG. 16 to be trained dependence on. The figure shows the relationship to be learned and the relationship learned, in each case the control surface. In the context of the goal, only the area above the entrance area to be traversed during learning is shown in order to enable a simple comparison. Of course, learning was only done in the entrance hall, so the knowledge is only meaningful there.

This simulation demonstrates that without any prior knowledge in the order of 10,000 cycles (distributed over the various operating points), a whole map for the injection control with very high accuracy "learned" is.

Claims (26)

1. A method for controlling the mixture in an internal combustion engine, comprising the following steps:

• a) measuring at least one variable with which the air mass entering a combustion chamber of the internal combustion engine is related (so-called input variable);
• b) determining at least one output variable controlling the fuel quantity to be supplied as a function of at least the input variable (s) measured in a) with the aid of stored imaging information;
• c) supplying the amount of fuel according to the initial variable from b);
• d) measuring a quantity that bears information about the resulting mixture (so-called actual size);
• e) Determining a deviation of the actual variable measured in d) from a target value for this variable:
• f) changing the stored image information as a function of the deviation determined in e) for the operating state measured in a), so that when steps a) to e) are carried out in the future, the deviation becomes smaller in the same operating state;
  and that realizes a learning process.

2. The method of claim 1, wherein steps a) to e) are carried out in cycles are, steps d) and e) being assigned to the cycle in which the Steps a) to c) are carried out. 3. The method of claim 1 or 2, wherein the internal combustion engine is an engine with or without charging and the input variable or one of the Input variables is the pressure in the intake manifold of the internal combustion engine.   4. The method according to any one of claims 1 to 3, in which the internal combustion engine is equipped with a variable intake system, and one of its position characterizing variable is the input variable or one of the input variables. 5. The method according to any one of claims 1 to 4, in which the internal combustion engine has an intake valve control with adjustable valve timing and the Input variable or one of the input variables one or more valve tax time parameters are. 6. The method according to any one of claims 1 to 5, in which the internal combustion engine Freely actuated inlet valves and the input variable or one of the Input variables the duty cycle and / or the closing and / or opening time of the intake valves. 7. The method according to any one of claims 1 to 6, wherein the input variable or one of the input variables is the speed of the internal combustion engine. 8. The method according to any one of claims 1 to 7, in which the internal combustion engine is controlled with a variable mixture ratio, and therefore the target Mixing ratio is one of the input variables. 9. The method according to claim 1 to 8, wherein the output variable one or more of the following sizes is: injection duration, duty cycle of Injector opening, injection pressure, degree of opening of the injector. 10. The method according to any one of claims 1 to 9, in which as an internal combustion engine Engine is used, which has essentially no fuel storage effect has, in particular a direct injection engine. 11. The method according to claim 10, wherein the determination of the output variable / n in Step b) is based only on quantities that correspond to the current working cycle belong.   12. The method according to any one of claims 1 to 9, in which an internal combustion engine Fuel storage effect is used and in the determination of the Output size / s in step b) also sizes from one or more previous work cycles are included. 13. The method according to any one of claims 1 to 12, wherein the stored Image information has the shape of a map, which the Output variable / n contains directly or indirectly, and in which in step f) an or several saved values can be changed. 14. The method according to any one of claims 1 to 13, wherein the mapping of Input variable / s to the output variable / s through a neural network Algorithm is done. 15. The method according to claim 14, in which the output variable is formed by linking a base value vector ( Θ ) representing the stored image information and an activation vector ( A ) dependent on the input variable (s). 16. The method of claim 15, wherein the link is linear, and in particular in the form of a dot product. 17. The method according to claim 15 or 16, in which the activation vector ( A ) is normalized and depends only on the distance of the input variable / n to the support points on which the vector representation is based. 18. The method according to any one of claims 14 to 17, wherein the mapping of Input variable / n is essentially local to the output variable / n. 19. The method as claimed in claim 18, in which only that component (s) of the activation vector ( A ) receives / receive appreciably large values which are / are located at a short distance from the input variable (s), while components at a larger distance are negligibly small are or disappear. 20. The method according to any one of claims 15 to 19, wherein the components A _{i of} the activation vector ( A ) from the distance d _{i of} the input variable / n to the associated support point i according to a center function, in particular
depend, where σ is a width parameter. 21. The method according to any one of claims 14 to 20, wherein the change in Mapping information substantially in the range of one Deviation point takes place, which is also in the figure at this point horizontal input variable (s) would be included in the output variable (s). 22. The method according to any one of claims 14 to 21, wherein the change in Mapping information takes place essentially locally to the point of deviation. 23. The method as claimed in one of claims 15 to 22, in which the mapping information is changed by adding a base value correction vector (Δ Θ ) to the base value vector ( Θ ) which links it, in particular to the product of the deviation value (e) the activation vector (A) is proportional. 24. The method according to any one of claims 14 to 23, wherein the adaptation of the stored image information or the learning process with a Adaptation law happens with stability and parameter convergence. 25. Device for controlling the mixture in an internal combustion engine, comprising:

• a) at least one device for measuring at least one variable with which the air mass entering a combustion chamber of the internal combustion engine is related (so-called input variable);
• b) at least one control device for supplying fuel;
• c) at least one device for measuring a size, which carries information about the resulting mixture or its combustion (so-called actual size);
• d) at least one memory for receiving the variable image information;
• e) and a programmed and / or hard-wired computer for carrying out the method according to claim 1.

26. The apparatus of claim 25, wherein the computer for executing one or several embodiments of the method according to claims 2 to 24 is programmed or wired.