METHOD FOR RECONFIGURING A LEARNING FUNCTION FOR FUEL RICHNESS REGULATION FOR ENGINE CONTROL

Abstract

The present invention relates to a method for reconfiguring a learning function (22) of an adaptive (AD) correction of a fuel mixture control function (21) of an engine control system, a new adaptive (AD) value to be stored being composed of a first part resulting from the calculation of the last correction and a second part resulting from the already stored adaptive (AD), the method comprising a first step of learning the fuel mixture during which the value of the first part is lower than the second part when storing a new adaptive value, a second step of monitoring for a failure of the fuel mixture control and, in the event of detection of said failure,a reconfiguration phase including a third step of parameterizing the calibration of a learning speed (CV) so as to configure the first part to a value that is higher than the second part when memorizing the new adaptive value. Figure 1.

Description

Title of the invention: METHOD FOR RECONFIGURING A RICHNESS REGULATION LEARNING FUNCTION FOR ENGINE CONTROL

[0001] The field of the invention relates to a method for reconfiguring a learning function in motor control.

[0002] The behavior of vehicle thermal engine actuators may differ from the behavior models integrated into the engine control unit, due to manufacturing variations, wear, fouling and the quality of representation of the engine actuators. In the case of actuator models relating to the intake branch and the injection branch, this discrepancy may lead to drifts in the richness of the mixture. This situation causes overconsumption and / or an increase in pollutant emissions. This may also have an impact on driving pleasure. It is necessary to correct these richness drifts throughout the life of the vehicle.

[0003] Conventionally, this correction is carried out by engine control models, carrying out the richness regulation function, which permanently corrects the injector control time based on the richness measurement provided by the richness sensor present in the exhaust. In addition, to optimize this correction, the engine control unit executes in parallel an engine control learning function responsible for memorizing the correction necessary for engine operating fields. More precisely, on a stabilized point of engine operating field, as soon as the learning conditions are respected, the learning function memorizes for this point the correction value determined by the engine control model in charge of richness regulation. This value is designated by the term adaptive correction or corrective.For example, the learning process can be carried out using a self-adaptive neural network. At the same time, the control unit communicates the value being stored, which has the effect of converging the richness regulation correction value. Thus, the regulation work is minimized and the correction dynamics are improved.

[0004] The learning function plays a major role when the richness regulation model cannot correct the richness, in particular when the richness sensor is not available, and significantly improves engine control during load transients by minimizing the work of the richness regulation function. Document FR3057031A1 is known, describing a method for correcting the richness filed by the applicant. The method described is a least squares minimization type optimization technique based on the analysis of errors present on the system in order to determine terms of the correction function and the aim of which is to correct the errors by improving the models. Document FR3088965A1 is also known describing a method for correcting the control of a heat engine in which correction adaptives are calculated according to two calculation methods, direct and indirect, aimed at applying these adaptives in specific engine operating zones, hot and cold. In addition, the method is configured to gradually transition from the application of an adaptive calculated according to the first method to the adaptive calculated according to the second method.

[0005] Furthermore, the regulations governing on-board diagnostics require manufacturers to monitor learning functions to alert the driver in the event of a fault occurring that could cause the tolerable thresholds for pollutant emissions to be exceeded. Conventionally, the diagnosis of richness deviation consists of checking the values ​​taken by the correction adaptors in comparison with limit values. Furthermore, certain regulations require a diagnosis to be supplemented by an end-of-fault validation mechanism.

[0006] A first known solution for resolving a failure consists of deactivating the learning function so as not to modify the adaptives that have been previously learned. A second solution consists of continuing the learning in the event of failure detection. The aim is then to try to absorb the new richness drift induced by the failure with the adaptives already memorized.

[0007] However, the disadvantage of the first solution is that the adaptors are no longer updated. Thus, if the failure was a one-off and disappears by itself, the system instantly regains its performance in terms of richness control from before the failure. On the other hand, if the failure is proven, the driver suffers the consequences of not controlling the richness, i.e. excess pollutant emissions, excess consumption, or even impacts on driving pleasure until intervention on the vehicle is carried out to resolve the failure.

[0008] Concerning the second solution, the main problem is that in the event of a sudden failure, for example a rupture of an injector supply hose, the richness drift will be instantaneous. However, the learning system is generally calibrated to learn a slow richness drift which corresponds to the slow effects on the richness of fouling and aging of the engine.

[0009] Indeed, certain sophisticated learning functions implement a learning speed calibration technique which provides that a new adaptive value to be memorized is made up of only one part resulting from the calculation of the last richness correction dependent on this speed calibration learning. To illustrate this principle, at the beginning of the vehicle's life, the adaptives have not yet been calculated, so their confidence level is zero. This share is initialized at 100%, meaning that if the strategy observes a need to correct the fuel mass by 15% (15% richness deviation measured by the exhaust richness sensor), the learning function will consider 100% of this 15% to update the adaptives. On the other hand, after several learning phases, the calibration of the learning speed will reduce this share over time by taking into account, for example, the vehicle's mileage or the number of learning cycles already carried out, for example. If this share is configured at 25%, the learning function will then consider only 25% of the richness deviation to update the adaptives.

[0010] Consequently, due to the calibration of the learning speed, the learning strategy will therefore generally not be able to quickly and completely absorb a drift in richness due to this type of failure. In addition, in the event of a one-off failure, if it were to disappear by itself, for example due to the temporary obstruction of an injector by a foreign body, the system finds itself shifted again and must relearn this new state of richness to slowly return to its initial state. During the failure, the system is therefore shifted in richness and the customer suffers the consequences of a lack of control of the richness.

[0011] There is therefore a need to overcome the aforementioned problems. One objective of the invention is to propose an improved engine control solution which makes it possible to reconfigure the learning strategy of the richness adaptors in the event of a failure impacting the richness control. One objective of the invention is to reduce pollutant gas emissions by optimized control of the richness regulation.

[0012] More specifically, the invention relates to a method for reconfiguring a learning function of an adaptive correction of a richness regulation function of an engine control, in which the richness regulation function calculates corrections from measurements of the deviation of the richness and the learning function stores adaptive values ​​in a set of correction models, said adaptive values ​​being derived from the corrections and learned for an operating domain of the engine, a new adaptive value to be stored in each of the models of said set being made up of a first part derived from the calculation of the last correction and a second part derived from the adaptive already stored in the model, and where said first and second parts are dependent on a calibration parameter of the learning speed.

[0013] According to the invention, the method comprises the following steps:

[0014] - a first stage of learning about wealth during which the value of the first part is less than the second part when storing a new adaptive value in at least one first model,

[0015] - a second step of monitoring a failure of the wealth control consisting of detecting whether at least two successive corrections of the first model in the engine operating domain are greater than a first predetermined threshold,

[0016] - and, in the event of detection of said failure, a reconfiguration phase comprising a third step of setting the learning rate calibration parameter so as to configure the first part to a value which is greater than the second part when storing the new adaptive value.

[0017] The method according to the invention may include the following additional characteristics, alone or in combination:

[0018] - Monitoring the failure of the richness control consists of detecting in furthermore if prior to detection, a robustness parameter representative of the number of learning cycles carried out for said adaptive is greater than a second threshold.

[0019] - The second threshold is equal to a value representative of at least fifteen cycles learning performed for said adaptive.

[0020] - The third step includes a command to decrease the ro parameter bust so as to increase the value of the first share.

[0021] - During the reconfiguration phase, a new adaptive memory is created. of the entire first part resulting from the last calculated correction.

[0022] - The reconfiguration phase is ordered for the first correction model only among the set of models during failure detection and updating of other correction models of the set of models is inhibited.

[0023] - The reconfiguration phase further comprises a fourth memo step rization of the new adaptive value in a second model, specific to the failure detection period and distinct from the first model used before the failure detection time, so that the richness regulation of the engine control is controlled by an adaptive of the second model during the failure detection.

[0024] - The reconfiguration phase further comprises a fifth inhibition step of a function for limiting the variation of the new adaptive value to be stored in relation to the adaptive value already stored.

[0025] According to the invention, there is provided a control unit for an engine control for an internal combustion engine comprising a richness regulation function and a learning function for a richness correction adaptor, comprising means for implementing the method for reconfiguring the learning function according to any one of the preceding embodiments.

[0026] The invention further provides a motor vehicle comprising an internal combustion engine, an exhaust line and such a control unit.

[0027] The invention further provides a computer program comprising instructions which, when the program is executed by a control unit of an internal combustion engine, cause the latter to implement the reconfiguration method according to any one of the preceding embodiments.

[0028] The method according to the invention is advantageous because it proposes an engine control solution which makes it possible to improve the corrective response of a richness adaptive learning function in the event of a fault occurring which impacts the richness control. This solution makes it possible to accelerate the taking into account of the sudden richness deviation whilst avoiding affecting the adaptives previously learned before the fault. The method limits the risks of overconsumption, excess pollutant emissions and impacts on the driving pleasure felt by the driver.

[0029] Other characteristics and advantages of the present invention will appear more clearly on reading the detailed description which follows, comprising embodiments of the invention given as non-limiting examples and illustrated by the appended drawings, in which:

[0030] [Fig. 1] schematically represents an engine control architecture of an internal combustion engine configured to implement the method of reconfiguring a richness learning function according to the invention.

[0031] [Fig.2] illustrates the principle of a richness regulation learning function implemented within the framework of the invention.

[0032] [Fig.3] represents an embodiment of the algorithm of the reconfi method configuration of a wealth learning function according to the invention.

[0033] The invention finds an application for internal combustion engine actuator control models and in particular for monitoring a learning function for such control models. The invention relates more specifically to a learning function of a model for regulating the richness of an engine control of the actuators relating to the intake branch and the injection branch.

[0034] In the present application, reference is made to an injection richness. It will be noted in this regard that a richness equal to 1 represents an air-fuel mixture with a stoichiometric dosage. A richness greater than 1 represents an air-fuel mixture with an excess of fuel. This is then referred to as a rich mixture. On the other hand, a richness less than 1 represents an air-fuel mixture with an excess of air. In the latter case, this is then referred to as a lean mixture. A richness error or drift is a non-zero difference between an actual richness and a richness setpoint. Such a richness error can be positive or negative.

[0035] In [Fig.l], there is shown schematically in the form of a block diagram an engine control system of an internal combustion engine 1 allowing the implementation of the method of reconfiguring a learning function of an adaptive richness correction according to the invention. The system comprises a control unit 2 of the engine 1 connected on the one hand to this engine 1 and on the other hand to an exhaust system 3 of combustion gases of this engine 1. The engine 1 comprises an intake 11 and injection 12 device which uses, to control this engine, a CSR richness setpoint value from the control unit 2, in particular. The exhaust system 3 comprises a device 31 for depolluting the exhaust gases and a unit 32 for measuring the richness in the exhaust gases. This richness measurement MR can typically be obtained by means of an oxygen sensor, also called a Lambda sensor.

[0036] The control unit 2 implements a richness regulation function 21 of the engine control coupled with a learning function 22 making it possible to adjust an engine control function 23 from richness correction adaptives to control the richness regulation of the engine 1.

[0037] More precisely, the richness regulation function 21 is responsible for calculating correction adaptives AD of the engine actuators making it possible to regulate the richness. The value of an adaptive is estimated from the instantaneous measurements of physical quantities from the engine sensor and the exhaust system 3, for example an engine manifold pressure and the richness measurement MR. The adaptives AD adjust the actuators of the intake devices 11 and / or the injection device 12 of the thermal engine 1 making it possible to control the richness, i.e. the models of the gas intake valve opening control, and / or the fuel injection duration control, and / or the position of the camshaft phase shifters, in particular.The engine control function 23 can take into account either an adaptive AD delivered directly by the richness regulation 21, or an adaptive AD delivered by the learning function 22, in particular as a function of the state of a selector SL whose state can depend on the availability of the richness measurement MR or the instantaneous operating conditions of the engine 1.

[0038] More precisely, the learning function 22 is responsible for storing the adaptives AD for correcting the richness in the engine operating domain calculated by the richness regulation function 21, in a set of correction models MAT comprising at least one model MAT1. The MAT models are for example models depending on the variables of speed, engine load, gas suction flow rate of the cylinders, in particular. In addition, the learning function 22 implements the updating of the adaptives AD for an operating domain and the function of restoring said adaptives later in a similar situation where the engine finds itself again in an operating domain for which an adaptive AD has already been stored. A learning model can be provided for each actuator adaptive. The learning model MAT1 is for example a matrix of adaptives of the injector model.

[0039] Furthermore, in the context of the reconfiguration method according to the invention, the learning function preferably comprises, stored in the memory of the control unit 2, a model MAT2 of richness adaptives activated specifically and used only during a richness deviation failure detection period. This model MAT2 makes it possible to immediately apply the entire correction required from the moment the failure is detected. The use of this second model also has the advantage of not affecting the adaptives stored prior to the moment the failure is detected.In the event of a one-off injector failure, this allows the MAT1 model adaptives to be used again without the adaptives already stored before the failure having been modified during the failure.

[0040] Furthermore, the engine control function 23 includes the MOD actuator models of the engine 1 recorded in non-volatile memory of the engine control 2 making it possible to control the richness, in particular the injector model, the air filling model of the cylinders and the camshaft position model. The engine control function 23 adjusts a richness setpoint CSR according to the adaptives delivered AD by the richness regulation function 21 or the learning function 22.

[0041] To return more precisely to the principle of parameterization of the learning function 22, the latter can be configured by a set of parameters and functions that the reconfiguration method according to the invention can modify when a failure is detected, in particular the parameters and functions relating to the learning speed and the MAT correction models used in order to specifically and immediately treat a high deviation in richness even though the learning operates in slow learning mode.

[0042] More precisely, a stored adaptive value of the learning function 22 is the value of the correction which is recorded in the memory of the control unit of the thermal engine. An updated or to be stored adaptive value of the learning function 22 may comprise the entire value, or only a PT part of the value, of the last calculated correction not yet applied by the regulation function 21. This PT part is calibrated according to a CV parameter for calibrating the learning speed.

[0043] Thus, the learning function provides that the updated adaptive value is a value consisting of a first part resulting from the calculation of the last richness correction and a second part from the adaptive already stored in the MAT model. The first and second parts for each richness adaptive in an operating domain are dependent on this CV learning speed calibration parameter. And in accordance with the invention, the method is able to modulate the value of these PT parts in the event of detection of a failure in order to quickly refocus the richness. It should be noted that the CV parameter can be a parameter, algorithm or function making it possible to modulate the learning speed, i.e. the value of the first part taken into account when updating an adaptive.

[0044] To detect a failure in the context of the method according to the invention, the parameterization of the learning function further provides a failure indicator DF, which can take two states, for example Boolean, the value of which is representative of a failure detection state. The failure detection function is based on the one hand on the deviation value of the instantaneous corrections calculated by the richness regulation function 21 with respect to a first deviation threshold SI, and on the other hand, on a level of robustness or confidence in the adaptives already stored which is provided by means of a so-called robustness parameter NR. The deviation value around a target setpoint is included in a range of values ​​expressed between [-1,1] or in value points [-100%,100%] for example. The first deviation threshold S1 which triggers an alert can be set for example at a value of 0.25 or 25% for example.

[0045] Furthermore, a preferred variant of the method provides that a failure is detected or confirmed when several successive corrections calculated by the regulation function 21 are greater than the first threshold, for example at least four successive corrections.

[0046] Furthermore, the robustness parameter NR is for example representative of a number of learning cycles already executed for an adaptive prior to the detection of a failure. In the context of the invention, the method provides that a failure is detected or confirmed for an adaptive if in addition the parameter NR associated with this adaptive is greater than a second threshold S2, for example a predetermined number of learning cycles, for example fifteen learning cycles. The parameter NR may alternatively be a number of kilometers traveled by the vehicle.

[0047] More precisely, the robustness parameter NR associated with a MAT learning model is a value ranging between a low robustness level (no learning cycles performed) corresponding to a high updating share with the value of the last calculated correction and a high robustness level (several learning cycles performed) corresponding to a low updating share with the value of the last calculated correction. The NR parameter can be a value ranging in a range of values ​​of [0,1] for example. Any robustness value scale is possible. Thus, the NR robustness parameter directly impacts the value of the CV speed calibration parameter and consequently the value of the PT shares, which allows, within the framework of the reconfiguration process, to increase the update speed of an adaptive when a failure is detected.

[0048] Furthermore, it should be noted that each MAT learning model can be associated with a CV speed calibration parameter and an NR robustness parameter. This thus makes it possible to trigger a reconfiguration specifically for one of the learning models among the set of MAT models.

[0049] Furthermore, the learning function 22 may comprise a function for limiting the variation of the adaptives AD stored in a correction model, at least the first model MAT1 and the second model MAT2. This function has the effect of controlling the evolution of the adaptives and thus avoiding excessive variations which are not normally observed for corrections corresponding to a “slow” calibration of the learning speed, that is to say corresponding to the compensation of the phenomena of wear and aging of the engine.

[0050] Furthermore, the learning function 22 may comprise FID identification functions of the actuator model(s) among the set of actuator models that are the source of a richness failure. Typically, these identification functions, known to those skilled in the art, are provided for observing engine parameters, in particular the richness measurement or the air pressure, to determine the actuator(s) that are the source of a failure. In the case of a sudden deviation of the richness measurement, an identification function is capable of determining that the fuel injection actuators are faulty in order to focus the action of the reconfiguration specifically on the adaptives of this actuator model.

[0051] In [Fig.2], the principle of the richness learning function is illustrated before describing in more detail the reconfiguration method according to the invention.

[0052] Curves C1 to C5 are represented on a time abscissa axis during which two learning cycles A1 and A2 intervene to determine respectively a correction adaptive which will be applied by the engine control. On curve C1, in the upper part, the number of measurements of the richness over a stabilized engine operating domain is represented. Each learning cycle A1, A2 comprises the calculation of a series of temporary corrections, represented by curve C2, i.e. intermediate values ​​not applied by the engine control during the piloting of the actuators and which are calculated from the measurements of the deviation of the richness. The values ​​of the corrections are established between [-1,1] or [-100%,100] to lean or enrich the mixture. The value of each correction adaptive resulting from a learning cycle is illustrated by curve C3. Curve C4 represents two instants tl and t2 of carrying out a richness deviation diagnosis. Once the learning function has sufficient confidence in the temporary corrections with regard to, for example, a minimum number of measurements, or a convergence criterion on the corrections, the richness correction is applied from the last correction adaptive in the actuator behavior models. Curve C5 represents the values ​​of the richness deviation.

[0053] In [Fig. 3], an algorithm of the reconfiguration method according to the invention is represented in the form of a block diagram. The engine control unit 2 is provided with an integrated circuit computer and electronic memories, the computer and the memories being configured to execute the reconfiguration method according to the invention. But this is not obligatory. Indeed, the computer could be external to the control unit 2, while being coupled to the latter 2. In the latter case, it can itself be arranged in the form of a dedicated computer comprising a possible dedicated program, for example. Consequently, the control unit, according to the invention, can be produced in the form of software modules (or computer software), or electronic circuits (or hardware), or a combination of electronic circuits and software modules.

[0054] In a first step E1 of the method, the richness regulation function calculates corrections from measurements of the richness deviation and the learning function stores adaptive values ​​learned for an operating domain of the engine in at least a first correction model MAT1. During nominal operation of the richness regulation, conventionally the engine control uses several correction models from the set of MAT models for controlling the engine actuators.

[0055] It is assumed that the engine control is established in a configuration for which the learning function has already carried out several learning cycles for all the operating domains of the engine. The robustness parameter NR of the learning function is therefore high, that is to say that the adaptives are deemed reliable by the engine control. In this situation, the learning CV speed calibration is set to a slow learning level, that is to say that it aims to correct the effects of wear and fouling of the injectors. Consequently, the PT share of a newly calculated correction is lower than that of the correction already stored when updating the adaptives stored by the learning function, in particular for the first injector control model MAT1.

[0056] Furthermore, in step E1, the failure indicator DF is in a first state representative of a normal operating state of the richness regulation.

[0057] Furthermore, the reconfiguration method comprises a second step E2 of the sur monitoring of a failure of the richness control. Monitoring is carried out periodically, for example at each diagnostic moment. More precisely, monitoring consists of detecting whether at least two successive corrections, preferably at least four, for an engine operating point are greater than a first predetermined threshold SI, and whether, prior to detection, the robustness parameter representing the number of learning cycles carried out for the richness adaptive is greater than a second threshold S2. For example, SI is configured at a value of 25% deviation, SI may be greater than 25%, and S2 may be equal to 15 learning cycles or more, for example.

[0058] If no failure is detected, the continuous monitoring loop and the motor control and the learning function remain set identically to the first step E1.

[0059] If a failure is detected, the reconfiguration method comprises a reconfiguration phase E3 where the failure indicator DF is controlled in a second state representative of the detection of a failure. For this purpose, the reconfiguration phase comprising a third step E31 of setting the calibration parameter of the learning speed CV so as to configure the first part to a value which is greater than the second part when storing the new adaptive value. For example, the value of the first part is configured at 100% and the value of the second part is configured at 0%. This configuration implies that a new adaptive value to be stored is equal to the entire value of the last calculated correction. The reconfiguration phase can apply for the first model MAT1 or for all the models of the set MAT.

[0060] Other values ​​are possible, for example 90% and 10%, or 80% and 20% for the first part and second part respectively. This modification of the speed calibration makes it possible to correct the richness deviance immediately and more quickly than in the case of a slow learning calibration.

[0061] It should be noted that other learning speed calibration actions are possible depending on the type of learning function implemented. The principle of step E31 consists of increasing the portion taken into account of the last calculated richness correction to converge quickly towards a correction taking into account the failure generating a high and sudden deviation.

[0062] In a variant of the reconfiguration phase E3, the third step E31 comprises a command to decrease the robustness parameter NR so as to increase the value of the first part. Thus, it is possible to temporarily accelerate the learning of the correction adaptives by decreasing the level of robustness or confidence in one of the correction models.

[0063] In a first preferred mode, the reconfiguration phase E3 applies only to the parameterization of the first correction model MAT1 of the injection device among the set of MAT models. In this way, it is possible to specifically correct the source of the richness problem by directly adapting the mass and quantity of fuel injected. This preferred embodiment improves the richness correction performance. According to this first variant, the other correction models of the MAT set are no longer updated during failure detection.

[0064] A second mode is provided, where, if a failure is detected, the reconfiguration phase E3 comprises a fourth step E32 of recording the correction adaptive by the learning function in a second correction model MAT2, specific to the failure detection period and dedicated to a failure situation, distinct from the first model MAT1 used before the failure detection time. Like the other models, the model MAT2 is used for the storage, updating and restitution of adaptives with the difference that it is used only in the event of a failure. Thus, the richness regulation does not affect the values ​​of the adaptives previously stored before the failure. Furthermore, in the event of a one-off failure, the learning function again uses the first correction model having a high level of robustness used during step E1.A second, fault-specific correction model can be implemented for each richness correction adaptive. The second MAT2 model is specific to the correction of the injector model. According to this second variant, the other correction models of the MAT set are no longer updated during fault detection.

[0065] It should be noted that the learning speed calibration steps, in particular the CV parameter and the NR parameter, allowing the first part of the adaptives to be increased during their updating apply to the second MAT2 model.

[0066] Furthermore, if a failure is detected, the reconfiguration phase E3 may comprise a fifth step E33 of inhibiting the variation limitation function of the adaptives stored in a correction model. The limitation function may apply to the first correction model MAT1 and to the second correction model MAT2. The learning function applies a correction completely and without limitation so as to accelerate the richness correction resulting from the failure.

[0067] Phase E3 can activate in a first reconfiguration mode step E31 of modifying the calibration of the learning speed only, or in a second reconfiguration mode activate only step E31 and E32 jointly, or activate in a third reconfiguration mode steps E31, E32 and E33 jointly.

[0068] A step E4 consists of monitoring the state of the failure indicator DF. The reconfiguration phase E3 is controlled by the engine control as long as the failure is detected. If the failure is no longer detected, the reconfiguration method returns to step El.

[0069] The reconfiguration method applies to motor vehicles equipped with an engine control system comprising an internal combustion engine and an exhaust gas system comprising a pollution control device and a richness measuring device.

[0070] The invention is described in the above by way of example. It is understood that the person skilled in the art is able to produce different variant embodiments of the invention by associating, for example, the different characteristics above taken alone or in combination, without departing from the scope of the invention.

Claims

Claims

1. Method for reconfiguring a learning function (22) of an adaptive (AD) for correcting a richness regulation function (21) of an engine control, in which the richness regulation function (21) calculates corrections from measurements of the richness deviation (MR) and the learning function (22) stores adaptive values ​​(AD) in a set of correction models (MAT), said adaptive values ​​(AD) being derived from the corrections and learned for an operating domain of the engine, a new adaptive value (AD) to be stored in each of the models of said set (MAT) being made up of a first part derived from the calculation of the last correction and a second part derived from the adaptive (AD) already stored in the model (MAT), and where said first and second parts (PT) are dependent on a calibration parameter of the learning speed (CV),the method being characterized in that it comprises the following steps: - a first step (El) of learning the richness (El) during which the value of the first part is lower than the second part when storing a new adaptive value (AD) in at least a first model (MAT1), - a second step of monitoring (E2) a failure of the richness control consisting in detecting whether at least two successive corrections of the first model (MAT1) in the engine operating domain are greater than a first predetermined threshold (SI), - and, in the event of detection of said failure, a reconfiguration phase (E3) comprising a third step (E31) of parameterizing the calibration parameter of the learning speed (CV) so as to configure the first part to a value which is greater than the second part when storing the new adaptive value.

2. Method according to claim 1 in which the monitoring (E2) of the failure of the richness control further consists of detecting whether, prior to the detection, a robustness parameter (NR) representative of the number of learning cycles carried out for said adaptive is greater than a second threshold (S2).

3. Method according to claim 2 in which the third step (E31) further comprises a command to reduce the robustness parameter (NR) so as to increase the value of the first part.

4. Method according to any one of claims 1 to 3 in which during the reconfiguration phase (E3) a new adaptive part (AD) to be stored is made up of the entire first part resulting from the last calculated correction.

5. Method according to any one of claims 1 to 4 in which the reconfiguration phase (E3) is controlled for the first correction model (MAT1) only among the set of models (MAT) during the detection of the failure and in that the updating of the other correction models of the set of models (MAT) is inhibited.

6. Method according to any one of claims 1 to 4 in which the reconfiguration phase (E3) further comprises a fourth step (E32) of storing the new adaptive value (AD) in a second model (MAT2), specific to the period of detection of the failure and distinct from the first model (MAT1) used before the instant of detection of the failure, so that the richness regulation of the engine control is controlled by an adaptive (AD) of the second model (MAT2) during the detection of the failure.

7. Method according to any one of claims 1 to 6 in which the reconfiguration phase (E3) further comprises a fifth step of inhibiting (E33) a function (FL) for limiting the variation of the new adaptive value (AD) to be stored in relation to the adaptive value already stored.

8. Control unit (2) of an engine control for an internal combustion engine (1) comprising a richness regulation function (21) and a learning function (22) of an adaptive (AD) of richness correction, characterized in that it comprises means for implementing the method of reconfiguring the learning function according to any one of claims 1 to 7.

9. Motor vehicle comprising an internal combustion engine (1), an exhaust line (3) and a control unit (2) according to claim 8.

10. A computer program comprising instructions which, when the program is executed by a control unit of an internal combustion engine, cause the latter to implement the reconfiguration method according to any one of claims 1 to 7.

Images