FR3090036A1 - METHOD FOR CORRECTING AN ESTIMATE OF NITROGEN OXIDES IN AN EXHAUST LINE - Google Patents

Abstract

The invention relates to a method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine, the estimate being made by an estimation model (14) previously calibrated, the model estimate (14) being adjusted according to a difference (E) periodically noted between the estimate by the model (14) and a measurement (NOxmes) of a nitrogen oxide probe (15) present in line d 'exhaust. A correction (Cor) of the deviation (E) is carried out periodically by a neural network as a function of at least two input parameters relating respectively to an engine speed and to an active engine torque. Figure 1

Description

Description

Title of the invention: METHOD FOR CORRECTING AN ESTIMATE OF NITROGEN OXIDES IN A LINE

EXHAUST

Technical field of the invention

The present invention relates to a method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine, the estimate being made by an estimation model previously calibrated during a design of the type of thermal engine on a nominal calibration engine.

This applies as well to a spark ignition engine with petrol or petrol fuel as for a compression ignition engine, in particular an engine running on diesel.

Prior art

For the depollution of nitrogen oxides or NOx, in particular for a compression ignition engine, a selective catalytic reduction system is often used otherwise known by the French abbreviation of RCS corresponding to the English abbreviation of SCR for "selective catalytic reduction". Subsequently in the present application, the selective catalytic reduction system may also be cited by its abbreviation RCS, as well as nitrogen oxides may be cited under their abbreviation NOx.

In an RCS system, a liquid reducing agent is used which is intended to be introduced in predefined quantities and by consecutive injections into an exhaust line of a motor vehicle. The addition of this depollution reducing agent performs the treatment of NOx present in the exhaust line of the heat engine of the motor vehicle.

[0005] In addition, there may exist a second system for depolluting nitrogen oxides by using a NOx trap, for example a poor-rich adsorption nitrogen oxide trap system known by the abbreviation of LNT or a passive nitrogen oxide trap system known by the abbreviation of PNA.

These traps allow the retention of NOx in non-favorable engine operating conditions for pollution control, these NOx traps being able to release the trapped nitrogen oxides in other conditions more favorable for their destruction. These different types of NOx adsorption trap will be grouped below under the name of NOx trap.

Finally, for a spark-ignition engine, a three-way catalyst can be used which simultaneously ensures depollution of nitrogen oxides by reduction of nitrogen oxides to dinitrogen, oxidation of carbon monoxide to carbon dioxide and oxidation of unburnt hydrocarbons to carbon dioxide and water.

For an SCR system as a nonlimiting example for the present invention, the redox reactions take place in a specific catalyst and are controlled by a probe, called a NOx probe, placed downstream of this catalyst . Indeed, this NOx probe makes it possible to correct the quantities to be injected by seeking the best compromise between the efficiency of the system and the desorption of ammonia.

In parallel, the quantities of reducing agent to be injected are calculated continuously as a function of the quantity of nitrogen oxides determined at the outlet of the engine, that is to say upstream of the catalytic reduction system. For cost reasons, this quantity of nitrogen oxides is calculated via an emission model.

The quantities can not be permanently corrected by the NOx probe, the estimation model must be precise enough not to cause the system to diverge.

In fact, in the case where the quantity of nitrogen oxides at the engine output is underestimated by the model, the computer controls a sub-injection of reducer, which results in an insufficient efficiency of the post- treatment to comply with pollution control standards.

In the case where the quantity of nitrogen oxides at the engine outlet is overestimated by the model, the engine ECU controls an over-injection of the reducer, which causes a release of ammonia at the exhaust outlet.

The estimation model is calibrated on a nominal engine and does not take into account the particularities of the heat engine specific to the motor vehicle and in particular its aging or its imbalance. However, with the dispersion of the engines produced and their aging, the estimation error can be significant.

The estimation model is adjusted according to a periodically noted difference between the estimation by the model and a measurement of a nitrogen oxide probe present in the exhaust line.

Document US-A-2003/216855 describes a process for controlling NOx emissions at the outlet of a heat engine. To do this, a model is used reflecting a predetermined relationship between the control parameters and NOx emissions, the control parameters including ambient humidity, engine inlet pressure, temperature at the exhaust manifold, fuel flow and engine speed.

The method also provides for an adjustment of the model as a function of a relationship between predicted NOx values and actual NOx values measured by a NOx probe. The adjusted model can be stored in a memory associated with the engine, so that NOx emissions discharged from the engine can be reduced based on virtual NOx emission values determined from the adjusted model.

If the estimation of this model brings a deviation from the measurement of the NOx probe, there is then learning of the model, in the form of a neural network, in order to better stick with the measurement.

This document requires the presence of a model comprising a neural network which is not the basic model fitted to a motor vehicle for the estimation of nitrogen oxide emissions in the exhaust line and implies an additional cost as well as a replacement of all existing models in operation.

Therefore, the problem underlying the invention is to correct a basic model for estimating the nitrogen oxide emissions discharged from a heat engine in the exhaust line, this model being developed for a nominal engine of the same type as the heat engine and not taking into account the drifts in operation of the heat engine, so that the nitrogen oxide emissions are estimated as accurately as possible.

Summary of the invention

To this end, the present invention relates to a method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine, the estimate being made by a previously calibrated estimation model. , the estimation model being adjusted according to a periodically noted difference between the estimation by the model and a measurement of a nitrogen oxide probe present in the exhaust line, in which a correction of the difference is performed periodically by a neural network as a function of at least two input parameters relating respectively to an engine speed and a torque of the engine in force.

The technical effect is to correct the estimate provided by an emission model of basic nitrogen oxides without changing the design of this model which is mapped and is calibrated on a standard model taking as input the parameters of usual engine control.

The basic model provides an estimate of nitrogen oxide emissions in the form of a basic value. In accordance with the invention, this basic value is then corrected by a corrective neural network, with two inputs relating respectively to the speed and to the torque of the heat engine in force. The basic value is first compared with a measured value of nitrogen oxide emissions supplied by a NOx probe to give a difference. It is this difference that is corrected by a fix provided by the neural network.

It follows that the basic model can be kept, which represents a great saving compared to the solution proposed by the closest prior art which proposed a new type of starting model integrating a network of neurons taking into account the multiple parameters previously cited with regard to document US-A-2003/216855.

The present invention therefore uses a simplified neural network with two input values that are the torque and the engine speed and independent of the basic model whereas a basic model takes into account more parameters, which would require a much more complex neural network and therefore a higher cost if the neural network were integrated into the model.

This allows to have a robust estimate on all of the engine operating points, whatever the engine considered, by learning the difference observed between the measurement of the NOx probe, advantageously downstream of the turbine d 'A turbocharger for a turbocharged thermal engine, which is not limiting in the context of the invention, and the nominal NOx estimate produced by the model.

The present invention makes it possible to improve the performance of the current on-board NOx models on board which do not need to be updated, the correction being made outside and independently of these models.

The present invention makes it possible to rely on a corrected model to carry out a diagnosis of the NOx probe. When the NOx emissions estimated by the model corrected by the method according to the invention corresponds to the NOx emissions measured by the NOx probe, it is that both the estimate is correct and the NOx probe is reliable.

Advantageously, the input parameter relating to the torque is an engine torque setpoint in force.

Advantageously, learning is carried out by the neural network in real time.

Advantageously, learning is suspended during the transient phases of operation of the heat engine implying a variation of at least a torque gradient or an engine speed gradient taken with respect to the torque or the engine speed greater than a calibratable threshold. of respective gradient variation.

Indeed, the principle underlying the present invention is that the neural network makes learning while driving the vehicle. It is therefore wise to determine windows of driving conditions on which learning is authorized, and others on which a temporary inhibition of the correction process is carried out.

The transient phases are not retained for learning because of the point effects on nitrogen oxides with NOx peaks not representative of regular NOx emissions.

Advantageously, learning is only suspended during the transient phases of operation of the heat engine implying a simultaneous variation of the torque gradient and of the engine speed gradient taken with respect to the torque or the engine speed greater than a calibrable threshold of variation. of respective gradient.

This makes it possible to limit and characterize the suspensive phases.

Advantageously, the neurons are positioned in a two-dimensional table with a dimension relating to the engine speed values and a dimension relating to the torque values by being integrated into rectangles or squares of pairs of an engine speed value and a torque value, each rectangle or square being defined by the pairs integrated in the rectangle or square plotting a respective portion of nitrogen oxide emission curves with respect to the iso-regime torque and with respect to the regime at iso-torque, the portion being a linear function having a relatively constant slope, varying for all the integrated pairs by less than a percentage of predetermined slope variation, a neuron being positioned in the middle of each predetermined square or rectangle.

This allows the positioning of the neurons as a function of the speed and torque pairs of the samples corresponding to the engine operating points.

Advantageously, for an engine operating point identified by a pair of engine speed and torque values, as many neurons are taken as there are squares or rectangles framing this operating point, a Gaussian activation function being developed. then normalized according to a normalized engine speed and torque for the neuron of the square or rectangle or each neuron of the squares or rectangles to give a correction, the neuron correction or a weighted or unweighted average of the neuron corrections being applied to the periodically noted difference between the estimation by the model and the measurement of a nitrogen oxide probe.

Advantageously, the normalized speed or torque are modified by specific weights.

Advantageously, a bias with a specific weight is taken into account for the calculation of the correction of the neuron or of each neuron.

This allows an additional calibration of the neural network due to the weights associated with the inputs that are the engine speed and the torque and the specific bias of each neuron.

The present invention relates to a method of depolluting nitrogen oxides from the gases discharged from a heat engine in an exhaust line, at least one depollution element being supervised in its loading or in a quantity of reducing agent introduced into the depollution line as a function of an estimate of nitrogen oxides in the exhaust line at the outlet of a heat engine by a previously calibrated estimation model, remarkable in that the estimate is corrected in accordance to such a correction process.

Brief description of the figures

Other characteristics, aims and advantages of the present invention will appear on reading the detailed description which follows and with regard to the appended drawings given by way of nonlimiting examples and in which:

[Fig.l]

FIG. 1 is a schematic representation of a flow diagram of a method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine according to an embodiment of the present invention,

[Fig.2]

FIG. 2 is a diagrammatic representation of a logic diagram for calculating a correction in a neural network, this correction being intended to be added to a difference between a value estimated by the model and a value measured by a NOx probe, [Fig.3]

FIG. 3 is a two-dimensional table according to an engine speed and a torque showing the positioning of neurons of the neural network used in the method according to the invention, each neuron being positioned in a square or a rectangle, [0045] [ fig.4]

- Figure 4 shows various NOx emission curves according to a torque of the engine with iso-engine speed, the substantially linear parts of the curves with a constant slope can partially define the rectangles and squares shown in Figure 3.

It should be borne in mind that the figures are given by way of examples and are not limitative of the invention. They constitute schematic representations of principle intended to facilitate the understanding of the invention and are not necessarily at the scale of practical applications. In particular, the dimensions of the various elements illustrated are not representative of reality.

Detailed description of the invention

In what follows, reference is made to all the figures taken in combination. When reference is made to one or more specific figures, these figures are to be taken in combination with the other figures for the recognition of the designated numerical references.

Referring to all the figures and mainly to Figure 1, the present invention relates to a method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine, advantageously for a vehicle and preferably for a motor vehicle.

The estimate is made by a previously calibrated estimation model 14 performed on a nominal engine of the type of the heat engine on which the method according to the invention is implemented but without taking into account the specifics of the engine due to its aging or its deviations from the nominal motor.

The estimation model 14 is in the form of a map and has several input values as usual engine control parameters, for example air flow, engine speed, torque, temperature, etc. .

The estimation model 14 is adjusted according to a difference E periodically noted between the estimation by the model 14 and a NOxmes measurement of a nitrogen oxide probe 15 present in the exhaust line.

According to the present invention, in the correction method, a correction Cor of the difference E is carried out periodically by a neural network as a function of at least two input parameters relating respectively to an engine speed and a torque of the current engine. This correction is made in the correction device 16 comprising the neural network.

The nitrogen oxide emissions estimated by model 14 are corrected. A comparison can also be made between the emissions of nitrogen oxides measured NOxmes by the NOx probe and the emissions of nitrogen oxides NOxcor which have been corrected.

A final value of NOxfin nitrogen oxide emissions can be obtained at the output in FIG. 1. It is then possible to deduce that both the model 14 is precise and that the NOx probe 15 gives reliable measurements when the NOxcor corrected nitrogen oxide emissions and the measured NOxmes nitrogen oxide emissions are the same.

The input parameter relating to the torque can be an engine torque setpoint in force, known to and controlled by the engine control unit for optimal operation of the heat engine.

In the correction device 16, learning is carried out by the neural network in real time and, if necessary during driving for a motor vehicle, since the information on the engine speed and the torque allow the network of adaptive neurons to calculate the deviation E observed and to calibrate the neural network in relation to this observation.

The engine speed and torque variables indicated will make it possible to define learning conditions.

To define the learning authorization, that is to say eliminate the driving phases for which the NOx emissions are not representative of the operation of the heat engine, it is possible to use the speed gradient which will be lower. to a calibratable threshold or the set torque gradient which will be lower than a calibratable threshold.

The learning can be suspended during the transient phases of operation of the heat engine implying a variation of at least a torque gradient or an engine speed gradient taken with respect to the torque or the engine speed greater than a respective calibratable threshold. . The percentage of the torque gradient to torque ratio or engine speed to engine speed gradient forming the calibratable threshold can be 5% but this is not mandatory.

In a less exclusive mode, learning can only be suspended during the transient phases of operation of the heat engine involving both a variation of the torque gradient and a variation of the engine speed gradient taken with respect to the torque or the engine speed greater than a calibrated threshold for varying the respective gradient.

The percentage of the torque to torque gradient ratio and of the engine speed to engine speed gradient forming the calibratable threshold can also be 5% but this is not compulsory. A suspension, in this case, implies that the two percentages are greater than 5%.

Figure 3 shows the positioning of neurons in a two-dimensional space in the form of a table.

In this table, the abscissa of the table shows an engine speed scale R ranging from 1,000 to 3,500 revolutions per minute. The ordinate of the table shows a torque scale C in N.m or Newton per meter ranging from 0 to 400.

Inside each box of the table are referenced nitrogen oxide emission values for each of the pairs of an engine speed value R and a torque value C.

The 13 white boxes numbered from 1 to 13 are the neurons placed respectively in a square or a rectangle cut out on the table. The number of neurons is not limiting.

Thus, neurons 1 to 13 can be positioned in a two-dimensional table with a dimension relating to the engine speed values R and a dimension relating to the torque values C. Neurons 1 to 13 can be individually integrated into rectangles or squares of pairs with an engine speed value R and a torque value C.

Each rectangle or square can be defined by the pairs integrated in the rectangle or square, therefore by the pairs of engine speed R and torque C values that the square or rectangle contains. These values of engine speed R and of torque C correspond to a respective portion of nitrogen oxide emission curves with respect to the torque C at iso-speed and with respect to the speed with iso-torque C.

These portions form for the engine speed R and torque C values integrated in each rectangle a linear function having a relatively constant slope, varying for all the integrated pairs by less than a percentage of predetermined slope variation, this percentage variation being for example less than 5%. A neuron can be positioned in the middle of each predetermined square or rectangle.

It follows that neurons 1 to 13 can be positioned in a 2D plane for value pairs of engine speed R and of torque C which accept a relatively constant slope of the NOx emissions at iso-torque and iso respectively. -diet. It is indeed judicious to cut out the motor field compared to this technical particularity for a better correction of the function.

If it has been recognized by experience that on certain types of heat engine, the optimal number of neurons allowing a good performance / calculation time adequacy is 13 neurons, this number is not limiting and may be other.

In general, the higher the number of neurons and the higher the accuracy of the correction for the entire operating field of the engine, but the more complex the calculation with frequent overlaps of the squares or rectangles unitarily dedicated to a neuron individual.

It would be possible to use another type of neural network for this correction function.

To improve the robustness of the function, one or more inputs in addition to the engine speed R and the torque C could be added, but the real-time calculation would become heavy enough to be performed on an engine computer.

FIG. 4 shows emission curves for nitrogen oxides NOx expressed in ppm as a function of torque values C expressed in Nm or Newton meters at iso-engine speed, each curve being related to a specific engine speed between 1500 and 3750 and ten different engine speeds.

It can be seen on each of these curves linear portions of the same slope. The top of most curves has a break due to the end of exhaust gas recirculation at the intake of the engine.

Such linear portions of the same slope can be used to develop two parallel sides bounding the torque C values of a square or of a respective rectangle shown in FIG. 3, these parallel sides being vertical in FIG. 3.

The same process can be applied for the definition of two sides perpendicular to the two parallel sides already drawn, this for speed values, these sides being horizontal in FIG. 3. The grouping of the two respective delimitations of the values of torque C and regime values R form the squares and rectangles shown in FIG. 3 and receiving a respective neuron referenced from 1 to 13.

Thus, for an engine operating point Pfonc identified by a pair of engine speed R and torque C values, it can be taken as many neurons as squares or rectangles framing this operating point Pfonc.

This Pfonc operating point can indeed be in a square or rectangle or simultaneously in several squares or rectangles having superimposed portions.

The number of squares or rectangles is higher when the precision on the linearity of the portions and their constant slope is increased.

Referring to FIG. 2, in the correction device referenced 16 in FIG. 1, which shows a processing of calculation with two neurons but which can also be done for more than two neurons as sketched by dashed dashed lines engine speed R and torque C measured.

References 1 and 2 are taken for the two neurons which are not necessarily neurons 1 and 2 in FIG. 3 and which can be any neuron.

In successive samples of pairs of engine speed R and of torque C, the engine speed R is, for each neuron, processed in a CalRl or CalR2 calculation unit and the torque C is, for each neuron, processed in a CalCl or CalC2 calculation unit.

This gives for each neuron a Gaussian activation function Fl act, F2act grouping together several samples. The calculation can be sampled at a time step of 20 milliseconds for example, without this being limiting. The reference Y. indicates the whole of the sampling at the end of a calculation and 1 / E indicates the normalization.

The calculation is done according to a calculation dedicated to a neural network which is known to the skilled person and that the skilled person can choose between several neural network calculations available to him.

These activation functions Fl act, F2act, with a number equal to the number of neurons are individually normalized norml, norm2. Then, a recalculation is carried out as a function of an engine speed R and of a torque C normalized for the neuron of the square or rectangle or each neuron of the squares or rectangles in which the operating point referenced Pfonc is located in FIG. 3.

The normalized recalculation, carried out in Calnorml or Calnorm2 standardized calculation units is carried out similarly to what had been implemented for the calculation of the Flact and F2act activation functions, that is to say with blocks respective calculation for the normalized engine speed R and for the normalized torque C, this for each neuron.

The calculation gives a corl correction, cor2 for each neuron, the correction of a corl neuron when unique or an average, weighted or not, of corl corrections, cor2 of the neurons being applied to the difference E recorded periodically between l estimation by the model and the NOxmes measurement of a nitrogen oxide probe 15, as previously shown in Figure L

In FIG. 2, an average correction cor after average, weighted or not of the corl corrections, cor2 is shown.

The normalized R or torque C regime can be modified by specific weights wlr, wlc; w2r, w2c, this for each neuron.

Likewise, a bias Bl, B2, specific to each neuron and recognized by experience, advantageously associated with a specific weight wlb, w2b, can be taken into account for the calculation of the corrective corl, cor2 of the neuron or of each neuron .

The method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine as previously described can find a particularly advantageous application in a method for depolluting nitrogen oxides from gases evacuated from a heat engine in an exhaust line.

In this depollution process, at least one depollution element is supervised in its loading, for example by being a three-way catalyst or a nitrogen oxide trap, or in a quantity of reducing agent introduced into the line. of pollution control, for example for a selective catalytic reduction system with calculation of the quantity of reducing agent injected into the exhaust line.

This supervision is done as a function of an estimate of nitrogen oxides in the exhaust line at the outlet of a heat engine by an estimation model previously calibrated with an estimate corrected in accordance with a method such as described previously.

The invention is in no way limited to the embodiments described and illustrated which have been given only by way of examples.

Claims (1)

Claims [Claim 1] Method for correcting an estimate of nitrogen oxides in an exhaust line at the outlet of a heat engine, the estimate being made by an estimation model (14) previously calibrated, the estimation model (14 ) being adjusted according to a difference (E) recorded periodically between the estimate by the model (14) and a measurement (NOxmes) of a nitrogen oxide probe (15) present in the exhaust line, characterized in that a correction (Cor) of the deviation (E) is carried out periodically by a neural network (1 to 13) as a function of at least two input parameters relating respectively to an engine speed (R) and a torque (C) of the current engine. [Claim 2] Method according to the preceding claim, in which the input parameter relating to the torque (C) is an engine torque setpoint in force. [Claim 3] Method according to any one of the preceding claims, in which learning is carried out by the neural network (1 to 13) in real time. [Claim 4] Method according to the preceding claim, in which learning is suspended during the transient phases of operation of the heat engine implying a variation of at least one torque gradient (C) or an engine speed gradient (R) taken with respect to the torque. (C) or at engine speed (R) greater than a calibrated threshold for the respective gradient variation. [Claim 5] Method according to the preceding claim, in which learning is only suspended during the transient phases of operation of the thermal engine implying a simultaneous variation of the torque gradient (C) and of the engine speed gradient (R) taken with respect to the torque (C ) or at engine speed (R) greater than a calibrated threshold for the respective gradient variation. [Claim 6] Method according to any one of the preceding claims, in which the neurons (1 to 13) are positioned in a two-dimensional table with a dimension relating to the engine speed values (R) and a dimension relating to the torque values (C) by being integrated into rectangles or squares of pairs with a torque value (C) and an engine speed value (R), each rectangle or square being defined by the pairs integrated in the rectangle or square plotting a respective portion nitrogen oxide emission curves with respect to to the iso-regime torque (C) and compared to the iso-torque regime (C), the portion being a linear function having a relatively constant slope, varying for all the integrated pairs by less than a percentage of slope variation predetermined, a neuron being positioned in the middle of each predetermined square or rectangle. [Claim 7] Method according to the preceding claim, in which, for an engine operating point (Pfonc) identified by a pair of torque (C) and engine speed (R) values, as many neurons (there 13) are taken as there are squares or rectangles surrounding this operating point (Pfonc), a Gaussian activation function (Flact, F2act) being developed and then normalized (norml, norm2) according to an engine speed (R) and a torque (C) normalized for the neuron of the square or rectangle or each neuron of the squares or rectangles to give a correction (corl, cor2), the correction of the neuron or an average (horn), weighted or not, of the corrections (corl, cor2) of the neurons ( 1 to 13) being applied to the difference (E) recorded periodically between the estimate by the model (14) and the measurement (NOxmes) of a nitrogen oxide probe (15). [Claim 8] Method according to the preceding claim, in which the normalized speed (R) or torque (C) are modified by specific weights (wlr, wlc; w2r, w2c). [Claim 9] Method according to the preceding claim, in which a bias (Bl, B2) with a specific weight (wlb, w2b) is taken into account for the calculation of the correction of the neuron or of each neuron. [Claim 10] Method for depolluting nitrogen oxides from gases discharged from a heat engine in an exhaust line, at least one depollution element being supervised in its loading or in a quantity of reducing agent introduced into the depollution line as a function of an estimate of nitrogen oxides in the exhaust line at the outlet of a heat engine by a previously calibrated estimation model (14), characterized in that the estimate is corrected according to a method according to l 'any of the preceding claims. 1/2