The invention relates to a method for configuring regeneration strategies of a particulate filter, in particular used in quality-controlled internal combustion engines.

Particulate filter systems for engines have promising potential for reducing exhaust-gas emissions. Filter systems can retain solid constituents (and substances which have accumulated thereon) in the exhaust gas. Coatings (generally catalytic) on the filter can, moreover, reactively convert gaseous constituents. Therefore, depending on -the filter efficiency, it is possible to separate solids out of the exhaust gas. The exhaust gas which remains escapes at the rear side of the filter, whereas the solids of corresponding size remain on or in the filter. The filter properties change as a result of this accumulation or incorporation of solids. Whereas a pressure drop is observed even when exhaust gas flows through empty filters, this pressure drop increases as the loading caused by deposited solids becomes greater. The degree of loading of a filter is generally described by this very variable, i.e. the pressure drop (dp) . If the pressure drop and therefore also the work which has to be applied to flow through the filter becomes too high for the particular application, it becomes necessary to clean the filter. This cleaning process may vary depending on the chemical properties of the deposited solids. In the case of solids such as diesel particulates, which mostly consist of carbon (C) , this cleaning process may, for example, comprise increasing the temperature of the medium flowing through. The increase in the temperature, while maintaining certain boundary conditions (such as for example a sufficient oxygen content) , creates conditions under which the particulates can be largely oxidized. The products which are formed as a result of this oxidation - mainly carbon dioxide (CO2) and carbon monoxide (CO) - are in gaseous form and escape during the operation, known as regeneration, at the filter outlet.

If a particulate filter is used for the exhaust-gas aftertreatment, it will inevitably be necessary to regenerate this filter. The regeneration may be triggered and operated actively or passively. It is possible to speak of passive regeneration if oxidation conditions occur without external intervention. This would arise, for example, when a vehicle is travelling at high speed on the motorway.

In the case of active regeneration, the oxidation conditions are generated by an active intervention of this nature in the overall system. In this case, by way of example, it would be possible to ignite a burner in the exhaust system.

In the majority of the operating characteristic diagram of a diesel engine, the required regeneration conditions are not achieved without any intervention. Consequently, energy has to be consumed, and this of course is to the detriment of the engine operating costs. Overall, the use of a particulate filter system has to be considered as having an adverse affect on fuel consumption, since the back-pressure of a system of this type acts on the charge cycle loop of the engine and can, as it were, have an adverse affect on the "aspiration" thereof. The regeneration per se likewise has an adverse affect on the overall fuel consumption of the engine. To enable the positive properties, namely the reduction of the particulate mass emitted, to be the dominant factor, intensive work is being carried out on minimizing the inevitable deterioration in fuel consumption.

To ensure optimum processes (loading and regeneration), process monitoring is imperative. Various measured variables which permit process monitoring of this nature are generally used. Sensors which are required to produce the measured variables are either present, i. e. they have already been fitted to monitor, control or regulate other processes in the overall system, or these sensors have to be fitted independently. Of course, additional sensors for the overall system entail costs, which always need to be minimized. In the case of the overall • engine and exhaust-gas aftertreatment system using diesel particulate filters, it is also possible to have recourse to variables which are in any case present for operation of the engine, but is not possible to do without further sensor means. There are numerous available sensor developments which allow description of the process of the particulate filter subsystem. However, these differ considerably in terms of price and functional stability. Pressure and temperature sensors have become established in this context.

In a form of process monitoring which is currently customary, as described in DE-A 42 30 180, the pressure is measured upstream (and possibly also downstream) of the particulate filter. This is obvious in particular because (as mentioned above) the back-pressure generated by the filter has a direct influence on the efficiency thereof. The pressure rise during loading phases and the pressure drop during regeneration operations can be measured and compared with limit values. The temperature measurement upstream of the filter monitors whether, in proven regeneration phases, correspondingly required thermal conditions are also present. In practice, however, purely pressure-based process monitoring of this type brings up a number of problems.

The exhaust-gas back-pressure generated by filters depends on various factors. One of the main factors of influence is the exhaust-gas volume, which changes constantly under non-steady-state engine operating conditions. Furthermore, pulsed conditions which influence the measurement result occur in an exhaust system. Overall, therefore, the pressure always provides a very fluctuating measurement variable. It is customary to define an upper and a lower limit characteristic value (oGk; uGk) , and reaching these limit values triggers the end of loading and regeneration, respectively.

Fig. 1 is intended to illustrate the problem: a differential pressure dp can be measured as a function of the exhaust-gas volumetric flow and the filter loading. Like any measurement variable, this parameter is subject to errors. Whereas the measurement is relatively uncritical at high volumetric flows, it becomes clear at low volumetric flows that both the lower and the upper limit characteristic variable can be reached independently of the loading but dependent on the random measurement error. In concrete terms, this means that, depending on the measurement error, the process control could give the command both for regeneration and to interrupt the regeneration. Consequently, for the case of the loading phase, the regeneration .would be begun too early, which can have adverse affects on the fuel consumption. If a regeneration phase is considered, the emptying would be interrupted prematurely and a residual quantity would remain on the filter. This case can likewise have a negative effect on fuel consumption. In practice, the regeneration is maintained for a certain time after the interrupt signal from the monitoring, in order to ensure reliable emptying. This required time delay in the process can only be under open-loop control, but not under closed-loop control, and therefore likewise manifests itself as increased fuel consumption.

Therefore, it is an object of the present invention to provide an improved method and an improved system for controlling the regeneration of a particulate filter arranged in the exhaust-gas purification system of a diesel engine. In particular, the increase in the fuel consumption caused by the regeneration of the particulate filter is to be reduced or minimized.

These objects are achieved by the features given in the claims .

The present invention is based on the basic idea of determining the loading of a particulate filter with the aid of an artificial neural network. According to the invention, the neural network is trained in a known state of loading by a characteristic diagram for the particulate filter, which preferably describes the pressure difference caused by the filter as a function of the exhaust-gas temperature and the mass of exhaust gas flowing through the filter. The trained neural network can then determine the current loading of the filter on the basis of measured variables which characterize the operating state of the engine. These measured variables are preferably the pressure difference caused by the filter, the exhaust-gas temperature and the mass of exhaust gas flowing through the filter.

The invention provides a method which can be used to determine the starting instant for the regeneration of a particulate filter. For this purpose, the current loading of the particulate filter is determined with the aid of the neural network described above, and this loading is compared with a predetermined limit value. In this way, the regeneration phase of the particulate filter is initiated only when the loading actually makes this necessary.

A further aspect of the invention provides a method for determining the duration of the regeneration phase required for substantially complete regeneration of a particulate filter and this phase, which causes additional fuel consumption, is reduced to the required minimum.

Furthermore, the invention provides a system for controlling the regeneration of a particulate filter.

The measured differential pressure dp at particulate filters, at a constant volumetric flow, is a function, inter alia, of the loading x of the filter. Referencing the process monitoring or control to the causal variable x in particular has the advantage that the loading x is semi-steady-state. Although the loading x rises over the course of time as a function of the volumetric flow and the concentration of particulates contained therein, this variable can be regarded as steady-state at correspondingly short time intervals.

However, it is difficult to record the filter loading while the engine is operating. In principle, various methods are conceivable, but in practice these methods entail excessive effort and/or cost arid also lack reliability. In this context, it should be noted that in extremis temperatures of between -300C and 14000C can be measured at a particulate filter.

However, if the relationship between loading x and the resulting pressure loss dp is known, it is possible to draw conclusions as to the loading from the simple (inexpensive) measured variable pressure. Although the same restrictions apply when measuring pressure, the loading is easy to evaluate as a semi-steady-state output variable.

A further advantage of the loading-based process monitoring arises during regeneration. If the regeneration conditions are known, it is possible to predict the required duration which is required for emptying of the filter on the basis of the mass x. The regeneration can also be jointly calculated online on the basis of the relevant input variables, with the result that the current loading value is available at any time. Therefore, there is no need to be dependent on the problematic pressure measurement during regeneration.

The invention will now be explained in more detail with reference to the drawings, in which:

Fig. 1 shows limit characteristic variables and measured pressure values as a function of the volumetric flow;

Fig. 2 shows types of filtration in particulate filters: depth filtration and surface filtration; [source: www.dieselnet.com]

Fig. 3 diagrammatically depicts the curve of the pressure loss dp over time (mDPF = x) ;

Fig. 4 shows, by way of example, the arrangement of pressure sensors (1,2) and temperature sensors (10,11,12) at the particulate filter;

Fig. 5 shows the projected area in the particulate filter characteristic diagram;

Fig. 6 shows the typical result of a characteristic diagram loop;

Fig. 7 shows, by way of example, the time curve of a characteristic diagram loop;

Fig. 8 shows the surface interpolation of the data points measured during the characteristic diagram loop; Fig. 9 shows, by way of example, a cut through the dp area at constant temperature;

Fig. 10 shows the pressure loss as a function of the exhaust-gas mass, the temperature and the loading (not extrapolated) ;

Fig. 11 diagrammatically depicts the data space which is actually measured and the extrapolated data space;

Fig. 12 diagrammatically depicts the system for controlling the regeneration;

Fig. 13 shows, by way of example, the typical loading in a steady-state test; and

Fig. 14 shows the loading and regeneration under realistic conditions.

To enable operation of the loading-based process monitoring, it is imperative to know the relationship between the loading mass x and the pressure loss dp. A strict distinction is drawn between the loading and regeneration, since homogeneous states which permit description and/or calculation are only present in the filter during loading.

In principle, there are various types of filtration, which are based on different mechanisms and the effects of which on the pressure loss consequently likewise differ. Fig. 2 shows the types of filtration, which may both arise in particulate filters.

It is generally the case that depth filtration (cf. phase I, Fig. 3) occurs first of all during the loading of particulate filters, which causes the pressure loss dp to rise relatively strongly. As the loading increases, the filtration merges into surface filtration, which causes the pressure loss dp to rise further (cf. phase II. Fig. 3). Fig. 3 diagrammatically depicts the curve of a loading level with the variable dp against time t.

As has already been mentioned, the filter properties change with the loading level. In addition to the change in quantity, the quality of the loading itself may change. This results, for example, from different particle size distributions and compositions at various engine operating points.

Recording, storage and provision of the relationship between dp and filter loading in loading phases

The operating range of an engine is generally described by its engine characteristic diagram. All the resulting variables can be presented as a function of the engine speed n and the output torque Md. By contrast, the volumetric flow V of the medium flowing through is the determining factor for a particulate filter or the pressure loss dp caused by the particulate filter. The volumetric flow V is in turn composed of the temperature T of the exhaust gas, the prevailing pressure p and the exhaust-gas mass mABG. Whereas the pressure and temperature measurement have to be provided separately at the particulate filter, in modern engines the variable mABG can be determined as a sum of the supplied air mass mL and the current injection quantity mK. These variables can be taken from the engine control unit (ECU) . It is then possible in tests to find a dedicated characteristic diagram for the particulate filter. The resulting pressure loss dp is recorded as a function of the mass of exhaust gas, the temperature and the loading. One (or more) artificial neural network (ANN) , which store the three- dimensional relationships in a suitable way and provide them such that they can be called up, can then be trained using the data. A problem in this context is the inability of neural networks to extrapolate. For example, if a temperature which lies outside the training range is offered to the network as an input variable, it is not possible to generate reliable output variables. Consequently, in practice it would be necessary for every engine/particulate filter combination to be tested before the method could be used. For this reason, a procedure which allows the use of known filter systems on any desired different engines has been developed. The extrapolation which is required to achieve this is made possible even before the actual network training.

At the development stage, the required data sets were generated on the engine test bench in what are known as CMS measurements. For this purpose, the engine is provided with the corresponding particulate filter in the exhaust section and the sensor means are fitted. Specifically, these are the sensors illustrated in Fig. 4, which are not all required in the vehicle. Pressure sensors 1, 2 for determining the pressure loss are arranged upstream and downstream of the particulate filter 4. The exhaust-gas temperature can be measured upstream of, in or downstream of the particulate filter 4 using temperature sensors 12, 11, 10. In addition, the exhaust-gas mass mABG is determined in the exhaust- gas stream 3.

To generate a "particulate filter characteristic diagram" using the pressure loss dp as a dependent variable of temperature T and the exhaust-gas mass mABG, the engine is operated in such a way that as far as possible all the extreme points are reached or the projected surface area illustrated in Fig. 5 is maximized.

In order, for example, to achieve the maximum exhaust- gas mass combined, at the same time, with the minimum temperature (point Pl in Fig. 5) , the engine is dragged at maximum speed n. To produce point P2, the highest load is run at maximum speed n. The exhaust-gas mass mABG is therefore dependent primarily on the engine speed, whereas the temperature of course rises with the load, i.e. the injection quantity.

To maximize the projected area A, a special driving cycle - referred to below as the characteristic diagram loop - which forms the basis of the generation of the training data has been developed in engine tests. Fig. 6 shows a typical result of a characteristic diagram loop of this type for a loading state of H g. The time curve of a characteristic diagram loop of this type can be seen in Fig. 7.

It can be seen that only 350 seconds are required to generate a characteristic diagram loop of this type. Therefore, even in the development phase the required test bench time is minimized.

In the next step, the data points are connected by means of area interpolation, as can be seen from Fig. 8. To then allow extrapolation, the area is cut at constant temperatures. Fig. 9 shows an example of one such cut.

The blue line in Fig. 9 shows the exact cut in the area shown in Fig. 8. The curve generated in this way is approximated by quadratic means (polyfit) and mathematically described, so that the curve can be extrapolated without problems:

f(x)=pθ+(pl) *xx+(p2) *x2 with f(x) = y = dp, and x = mABG

Furthermore, the cuts taken at constant temperatures offer a checking feature: if there is no flow through the filter, the pressure loss dp must be 0. For the equation, this means that the coefficient pO should always be equal to 0. If the individual coefficients of the combination of cuts for different loadings are considered, it is possible to obtain still further information. If the coefficients are plotted individually over various loading stages, it is also possible to establish a trend which permits extrapolation over the loading. In concrete terms, this means that not all the loading stages actually have to be measured in tests in order to generate the training data set. Fig. 10 shows measurements at various loading stages.

Therefore, the operations of cut, polyfit, extrapolation can be used to create, from the data set generated by the total number of measurements, an artificial data set which can cover significantly wider ranges than the range which was originally measured. Fig. 11 provides a diagrammatic illustration of this.

The correspondingly associated values are assigned to each of the data points of the data space. The assignment and storage are effected by means of the training of the neural network(s) . Once trained, the networks then provide the desired output variable as a function of various input variables that are independent of one another. For the specific application of particulate filters, the current exhaust-gas mass flow mABG, the temperature of the exhaust gas in the filter and the pressure loss dp are presented to the artificial neural network as input variables. The associated loading is recognized by classification and output.

In this context, any errors in the measured variables of the network input of course also influence the output. However, since the level of the loading x may not decrease under corresponding conditions compared to the pure pressure measurement, the output variable can be evaluated. If the particulate emission of the engine is known, the output variable can be assessed further: a rise in the loading above the level of the untreated emissions can be identified as obviously false. In addition, the output signal can be averaged on a sliding basis or in portions, since it is not necessary to update the values at the input frequency of 1 Hz. Rather, it is sufficient for the loading values to be output on a "minute-by-minute" basis, in particular when it is considered that under real conditions "Golf class" vehicles can be driven for approx. 500 km before a particulate filter has to be regenerated.

Calculation of the filter loading x in the regeneration phases

The mathematical-physical description of the relationship between filter loading x and pressure loss dp fails in regeneration phases. Even the empirically discovered relationship cannot be applied, since in this case, unlike during loading, the conditions are not homogeneous. The operation of soot oxidation, i.e. of regeneration, is in fact considerably influenced by random factors on both a time basis and a local basis. In general, soot ignition conditions are locally reached in the filter, and therefore the regeneration is initiated at these hot spots, as they are known. During the time curve of regeneration, therefore, the filter may in parts be completely burnt free, whereas further, generally cooler regions are still laden with soot. In the case of catalytically coated particulate filters, it is often the loaded particulates which are in direct contact with the catalyst material which are oxidized first. Consequently, of course, this means those particles which were deposited in the depth of the filter or close to the surface. As mentioned above, depth filtration has a considerable influence on the pressure loss. This relationship behaves in a corresponding way during regeneration, i.e. the pressure loss initially drops considerably during oxidation of the depth- filtered particles.

In general, complete regeneration is desired, and during this regeneration the information concerning the loading state of the filter is not absolutely necessary. However, it is important to have reliable information about the end of regeneration or the complete removal of loading from the filter, in order to minimize the outlay on energy required for the regeneration. For the above reasons, this information cannot be derived with sufficient accuracy from the measured variable pressure loss dp.

If the ■ desired complete regeneration needs to be interrupted prematurely (for example as a result of the driver switching off the vehicle) , it is advantageous to know the mass remaining on the filter. It is then possible to decide, on the basis of the calculated residual quantity remaining on the filter, whether it is sensible to restart the regeneration after the. vehicle or engine has been started up or whether the loading of the filter can be recommenced.

This is where the approach of the loading-based process monitoring comes to bear. The main variable factor of influence for the regeneration is known with the temperature T at the filter. According to the Arrhenius equation, which provides a very general description of chemical reactions, under the same boundary conditions the reaction rate is dependent only on the temperature. -Ea k=A-e&'T (Arrenius equation)

with : k = oxidation rate A = reaction - dependent constant R •= gas constant T = temperature Ea = activation energy Therefore, the soot conversion or the oxidation rate k can be determined as a temperature-dependent variable using variables for the activation energy Ea and the reaction-dependent constant A determined empirically in test bench tests and what is known as the DTA (differential thermal analysis) method. The chemical composition of the medium flowing through, in this case diesel exhaust gas, also has an influence but generally remains approximately constant during active regeneration or over the operating time of the engine. If different modes are provided for in the regeneration strategy, it is possible to make use of information from a corresponding regeneration characteristic diagram.

At the end of a loading phase, the value of the loading x is transmitted from the neural network to a further calculation tool. The latter uses the measured input variables temperature T, engine speed n, the current injection quantity and the air mass,, to calculate first of all the current particulate emission. This must be known, i.e. must already have been measured and stored in a suitable way. A neural network is likewise used for this purpose. The same is done using the concentration of oxygen in the exhaust gas. The oxidation rate k, which may be corrected for differing oxygen contents, is then calculated. It is thereby possible to calculate both the active regeneration and random partial regenerations originating, for example, from full-load drives. The actual loading loss dx at the particulate filter is ultimately calculated as follows:

oxidation rate k [g/h] + particulate emission [g/h] = loading loss dx[g/h]

The variable dx is an instantaneous variable which is calculated for the current prevailing conditions. Working on the basis of the starting loading of the filter during regeneration (= the final value of the loading = x) , the current loading is calculated from the loading loss dx:

Therefore, the loading x is calculated during the regeneration as described above and not calculated in the same way as during the loading, based on the pressure loss dp. In addition, the loading loss dx can be determined more accurately on the basis of exothermicity. To do this, it is necessary to insert two temperature sensors, upstream and downstream of the particulate filter.

Since the calculation of the loading x is fundamentally performed differently in the individual phases, it is imperative to exchange information concerning the respective values. During a loading phase, the loading value x is output continuously and made available to the tool which is responsible for the calculations during the regeneration. It can therefore perform its work at any instant, both in the event of the start of regeneration being initiated randomly and in the event of it being initiated actively. When the regeneration is completely concluded, the loading value x = 0 is output and is first of all compared with the current calculation of the loading tool in the subsequent loading phase. If a difference between the output values of the two tools is determined repeatedly after supposedly complete regeneration, the duration of subsequent regeneration phases is in each case increased. If differences are still detected after this measure, this phenomenon is ascribed to non- regenerative deposits and is output as a characteristic variable. The deposits may originate, for example, from ashes, which represent a constant load on a particulate filter in just the same way as pure soot particulates. The ash component is relatively low when using low-sulphur diesel fuels without additives. Whereas the assumption for the filter service life (which is generally given on the basis of the vehicle running time) is approx. 80 000 km when additives increasing the ash component are added, systems which do without additives may even attain the same service life as an engine, which is generally given as approximately 240 000 km. The regeneration strategy but also calculation of the overall loading itself is to be adapted by the pressure loss caused by these non- regenerative deposits. This circumstance is taken into account in the interaction between the two calculation tools.

Furthermore, the characteristic variable of the non- regenerative loading also provides a variable for evaluating the filter performance. It is therefore possible to signal that a filter needs to be exchanged, since the thermal purification provided is no longer sufficient to satisfy the demands for a moderate pressure loss.

To summarize, therefore, the development comprises two tools (cf. also Fig. 12), of which one (tool 1) is responsible for calculating the soot loading mass. The semi-steady-state variable of the loading mass x is output on the basis of dynamic input variables corrected by corresponding factors. The tool has the ability to subsequently calculate recorded loading levels offline. However, it may also output the current loading online, i.e. during instantaneous engine operation, which corresponds to the desired application in vehicles.

As soon as oxidation phenomena that are worthy of mention take place on the filter, the second tool (tool 2) calculates the mass loss of soot particulates in the filter on the basis of the input variables temperature (tE) , air mass flow (mL) , fuel mass flow (mK) , particulate concentration (kPM) , oxygen level in the exhaust gas (k02) and soot conversion and/or oxidation rates (k) .

The two tools operate in parallel, with the first tool only being used under loading conditions and the second working under regeneration conditions. Exchange of information takes place during a transition from loading to regeneration or from regeneration to loading.

There now follows a specific example: during loading, the variable xl for the loading mass is output by tool 1. When the difference between the upper limit characteristic variable Xo and the variable xl reaches a value of 0 (Xo-xl = 0), regeneration is initiated. The value xl is transmitted from tool 1 to tool 2 and the conversion dm/dt of soot particulates is calculated continuously. The regeneration time tR required (under current conditions) for complete emptying of the filter is calculated, and so too is the instantaneous value of the loading x2 as a function of the prevailing regeneration conditions. Therefore, the current Loading value x2 is available even in the event of incomplete regeneration and a premature transition to renewed loading, and this current loading value x2 is then taken over by tool 1 and used for evaluation and calculation until homogeneous conditions are established at the filter. Any ash formation or deposits which cannot be regenerated are recognized and incorporated in further calculations after a characteristic variable xNR has been formed. Therefore, it is always the soot mass xl or x2 which is currently deposited on the filter and the characteristic variable of the ash formation xNR which are displayed. Furthermore, the values are indicated in a suitable way in order to provide the system, as well as the driver of the vehicle, with information as to the current reliability level of the instantaneous values.

Experiments aimed at validating the virtual loading sensor were carried out. Figure 13 shows a characteristic loading test which was carried out for steady-state conditions (constant load and constant engine speed) . The test shown was carried out at an average engine speed (approximately 2000 revolutions per minute) and an average load (approximately 100 N/m) . The loading was interrupted three times and the filter was weighed, as indicated by the diamonds at 0 minutes, 120 minutes and approximately 235 minutes. The increase in loading between these points is regarded as linear, as shown by the continuous line. This was confirmed by particulate measurement. The triangles show the results of the virtual loading sensor. • ' Under real conditions, a regeneration which has been started may have to be interrupted under certain circumstances, with soot remaining behind. For regeneration strategies, it is important to be aware of the soot left behind. The performance of the virtual loading sensor was also tested under these real conditions.

Figure 14 shows the loading and regeneration of a diesel particulate filter under real conditions. The results of the virtual loading sensor, which are illustrated by diamonds, are compared with the results of a gravimetric measurement of the soot mass, which is marked by the squares and the continuous line. Overall, the highest absolute deviation in the loading mass is only 1.1 g, which occurs after the first regeneration.