Method for improving the reliability of a learning function of a model of combustion engine actuators

Abstract

The present invention relates to a method for improving the reliability of a learning function of a model for controlling actuators of a motor vehicle combustion engine, wherein the method comprises the following steps of calculating (25) a difference between a first value stored by the learning function for the term of the model and a second current value of the term of the model, which value is estimated from instantaneous measurements, of determining (26) a confidence index assigned to the first stored value of the term of the model, of determining (27) the value of a reliability-improving parameter dependent on the difference and on the confidence index, of updating (28) the term of the model by a third value made up of the sum of the parts of the first value and of the second value, during which one of the parts or each of the parts is dependent on the reliability-improving parameter.

Description

[0001] The field of the invention relates to a method for securing a learning function of a control model for actuators of a motor vehicle internal combustion engine.

[0002] Internal combustion engine actuators in vehicles can exhibit behavior that differs from the behavior models integrated into the engine control unit (ECU) due to manufacturing variations, wear, fouling, and the quality of the actuator modeling. In the case of actuator models related to the intake and injection ports, this discrepancy can lead to fuel mixture deviations. This situation results in increased fuel consumption and / or higher emissions. It can also negatively impact driving experience. It is necessary to correct these mixture deviations throughout the vehicle's lifespan.

[0003] Typically, this correction is performed by engine control models, which implement the fuel mixture regulation function. This function continuously adjusts the injector timing based on the fuel mixture measurement provided by the exhaust gas mixture sensor. Furthermore, to optimize this correction, the engine control unit (ECU) simultaneously executes an engine control learning function responsible for memorizing the necessary correction for specific engine operating conditions. More precisely, at a stabilized point within the engine operating conditions, as soon as the learning conditions are met, the learning function memorizes the correction value determined by the engine control model responsible for fuel mixture regulation. For example, the learning process can be implemented using a self-adaptive neural network.Simultaneously, the control unit communicates the value being stored, which causes the fuel mixture correction value to converge. This minimizes the regulation workload and improves the correction dynamics.

[0004] The learning function plays a major role when the fuel mixture control model cannot correct the mixture, particularly when the mixture sensor is not available, and significantly improves engine control during load transients by minimizing the work of the fuel mixture control function.

[0005] We are familiar with document FR3057031A1, which describes a wealth correction method filed by the applicant. The described method is a least-squares minimization optimization technique based on analyzing the errors present in the system to determine terms of the correction function. The goal is to correct the errors by improving the models. This method describes tests on the quality of the correction terms during calculation to guide the convergence of the optimization loop. These tests are included in the loop for calculating the correction terms and are applied to the corrections during convergence. We are also familiar with document FR2979390A1, which describes an optimization technique based on a Kalman filter. Documents FR 3 033 840 A1 and FR 3 055 667 A1 are other important documents.

[0006] When a new correction value is calculated, some learning strategies involve fully learning the value and overwriting the value stored in the learning model's memory. However, a drawback arises when the current value is calculated from measurements taken under atypical engine operating conditions, for example, if a specific event occurs and temporarily shifts the fuel mixture control system. The function then learns an incorrect value, unrepresentative of the operating range, which causes a fuel mixture shift across the entire allocated operating range once the atypical event has disappeared. Other strategies involve learning only a portion of the current correction to limit the negative effects caused by a one-off atypical event. However, the trade-off is that if the current value is correct, the learning speed is reduced.

[0007] Therefore, there is a need to address the aforementioned problems and improve the learning process of a motor control function. One objective of the invention is to secure the learning models in order to limit the impact of a value that is temporarily shifted due to an atypical event.

[0008] More specifically, the invention relates to a method for securing a learning function of a control model for actuators of a thermal engine for a motor vehicle, the learning function being based on a model consisting of control terms stored in the memory of a control unit of the thermal engine and being adapted to operate a learning cycle consisting of loops for updating the control terms.

[0009] According to the invention, the process comprises the following steps: Calculation of a difference between a first value stored by the learning function for said model term and a second current value of said model term estimated from instantaneous measurements; Determination of a confidence index assigned to the first stored value of said model term; Determination of the value of a safety parameter dependent on said difference and the confidence index, said safety parameter being a weighting coefficient delivered by a predetermined matrix having as input parameters said difference and the confidence index; and during the update step of the model term, the share of the second current value is calculated as a function of the value of said weighting coefficient; Update of the model term by a third value consisting of shares of the first value and the second value, during which one or each of said shares is dependent on said safety parameter. and in that it further comprises, at each update loop, a step of updating the value of the confidence index according to said deviation and when said deviation is greater than a third threshold, the confidence index is decremented by a predetermined value and when said deviation is less than said third threshold the confidence index is incremented by said predetermined value.

[0010] According to one variant, the process also includes, prior to the cycle of update loops, a step of initializing the confidence index of the model term to an initial confidence level.

[0011] According to one variant, when updating the model term, if the confidence index is greater than or equal to the initial confidence level and: If, in addition, the difference is greater than or equal to a first threshold and less than or equal to a second threshold, the share of the second value is strictly less than the share of the first value stored. If, in addition, the difference is greater than the second threshold, the share of the second value is zero so as to retain in full the share of the first value stored.

[0012] According to one variant, when updating the model term, if the confidence index is strictly less than the initial confidence level and greater than a first minimum limit, the share of the second value is zero so as to retain in full the share of the first stored value.

[0013] According to one variant, when updating the model term, if the confidence index is equal to a second minimum limit, the share of the second value is at least equal to or greater than the share of the first stored value.

[0014] According to one variant, the first minimum limit is equal to the second minimum limit, or greater than the second minimum limit.

[0015] According to a second embodiment of the process, the learning function further includes an internal configuration parameter representing a confidence level of the stored value of the model term, and in that said internal parameter is configured according to the value of the safety parameter.

[0016] According to one variant, the control module is a richness regulation function intended to calculate adaptive correction terms and the learning function is adapted to memorize said adaptive correction terms according to a matrix of engine speed and load variables.

[0017] The invention provides a control unit for an internal combustion engine of a motor vehicle, comprising a learning function for an actuator control module of said internal combustion engine. The learning function is based on a model consisting of control terms stored in the control unit's memory and is adapted to perform a cycle of updating the control terms. According to the invention, the control unit is configured to implement the method for securing the learning function according to any one of the preceding embodiments.

[0018] The invention also relates to a motor vehicle comprising a thermal engine controlled by such a control unit.

[0019] The invention relates to a control unit for a heat engine implementing a learning function for an actuator control module of said heat engine, the learning function being based on a model consisting of control terms stored in the memory of the control unit and being adapted to perform a cycle of updating the control terms, further comprising: a means of calculating a difference between a first value stored by the learning function for said term of the model and a second current value of said term of the model estimated from instantaneous measurements, a means of determining a confidence index attributed to the first stored value of said term of the model, a means of determining the value of a safety parameter dependent on said difference and the confidence index, a means of updating the term of the model by a third value consisting of the sum of the parts of the first value and the second value, where one or each of said parts is dependent on said safety parameter.

[0020] Thanks to the invention, the method and device for securing a learning function make it possible to diagnose the reliability of the stored values ​​of a motor actuator control model. Updating a learning function is generally subject to a native convergence mechanism designed to gradually update the stored values ​​with only a portion of the newly measured values. When the diagnosis determines that a value of a term in the model is unreliable, the securing method accelerates convergence to the new measured values. This accelerated convergence is further subject to additional checks to prevent updates based on temporarily erroneous values.

[0021] The invention finds preferential application in the context of fuel mixture control models and therefore actively contributes to rapid and robust control of fuel mixture errors throughout the vehicle's life. More generally, the invention helps improve the performance of learning strategies for engine actuator control models and thus contributes to an improvement in their performance.

[0022] Other features and advantages of the present invention will become more apparent upon reading the following detailed description, which includes embodiments of the invention given by way of non-limiting examples and illustrated by the accompanying drawings, in which: [ Fig.1 ] represents a device for securing a learning function of a control model for internal combustion engine actuators according to the invention; [ Fig. 2] represents a preferred embodiment of the algorithm for the security process according to the invention; [ Fig.3 ] represents an example of a calculation matrix for the safety parameter operating the weighting of an adaptive value for wealth correction for the preferred embodiment of the safety process according to the invention; [ Fig. 4 ] represents a second variant of the device for securing a learning function according to the invention.

[0023] The invention finds application in internal combustion engine actuator control models, and in particular in securing a learning function for such control models. To this end, in a first preferred embodiment, the invention provides an additional securing function for a native learning function, enabling a secure diagnostic for both the stored and current adaptive parameters before the new value is recorded in the learning matrix. In a second embodiment, the additional securing function cooperates with the learning function to modulate the model noise of the model terms based on the result of the diagnostic performed by the securing function. Model noise refers to any model configuration parameter representative of a confidence level in the model.According to the result, the safety function accelerates the native convergence process towards the value of the new adaptives by decreasing the confidence level of the model in this variant.

[0024] In this description, the term actuator control model refers to any actuator control function of the engine. The learning function implements the function of storing the terms calculated by the control model, the function of updating the model terms, and the function of retrieving said terms. The term stored value of the learning function refers to the value of the control model that is stored in the memory of the internal combustion engine control unit. The term current value of the learning function refers to the value estimated from instantaneous measurements of physical quantities from engine sensors, for example, the exhaust air-fuel ratio sensor.

[0025] The embodiments of the invention will be described in more detail by way of non-limiting example for the case of a learning function for a fuel mixture control model. The control model is a fuel mixture control function responsible for calculating correction terms for the engine actuators used to regulate the fuel mixture. In this example, a control term designates the adaptive fuel mixture correction parameters, and the learning function is adapted to memorize said adaptive fuel mixture correction parameters, based, for example, on a matrix of engine speed and load variables. The current value is estimated from measurements of the engine's manifold pressure and the oxygen content of the engine's exhaust gas.The adaptive correction systems adjust the actuators of the intake branch and / or the injection branch of the internal combustion engine to control the mixture, either the control of the opening of the gas intake valve and / or the control of the duration of fuel injection.

[0026] There figure 1 represents the first preferred embodiment of the functional modules of a thermal engine control unit enabling the implementation of the safety process according to the invention.

[0027] The control unit is equipped with an integrated circuit computer and electronic memory, the computer and memory being configured to execute the safety procedure. However, this is not mandatory. The computer could be external to the engine control unit, while still being coupled to it. In this latter case, it could itself be configured as a dedicated computer including, for example, a dedicated program. Consequently, the control unit, according to the invention, can be implemented as software modules, electronic circuits, or hardware, or a combination of electronic circuits and software modules.

[0028] More specifically, the control unit includes a learning function 11 for the control model of the mixture regulation actuators and a safety module 1 for the value of the term to be learned by the learning function 11. The learning function 11 is configured to store an adaptive correction under pre-established learning conditions, specifically at a stabilized operating point in terms of engine load and speed, or within a stabilized operating range in terms of engine speed and load. The learning process continuously executes update loops of the terms in the learning matrix during engine operation when the learning conditions are met.In other words, the learning function iteratively, or recursively, calculates the values ​​of the model's terms based on instantaneous measurements of physical quantities recorded by one or more sensors of the engine. These values ​​are then evaluated against a stored value from the previous update loop. Depending on the optimization technique used, the new value stored by an update loop includes all or part of the current value, or retains the entire value stored in the previous loop.

[0029] For example, the values ​​updated by the learning function can be obtained using a least-squares minimization optimization technique, as illustrated in document FR3057031 A1, or a Kalman filter as illustrated in application FR2979390A1. These two techniques are cited by way of illustration, and those skilled in the art will be able to adapt the implementation of the invention to any learning function used for optimizing motor actuator control models.

[0030] According to the invention, the safety module 1 includes a calculation means 2 capable of receiving as input a first stored value 10 from the learning function 11 for a term of the model and a second current value 12 of said term estimated from instantaneous measurements. The calculation means 2 is configured to calculate the difference between these two values. The safety module 1 can be used to evaluate a point, several points, an area, or each motor operating area defined by the learning matrix. The calculation of the difference can be performed in each update loop of the learning function. The difference value 6 is output from module 2.

[0031] The security module 1 further includes a means for configuring and determining a confidence index 7 assigned to the first stored value 10 of the model term in the training matrix. The calculation means 3 is capable of configuring an initial predetermined confidence level and modifying the value of the confidence index based on the evaluation of the gap 6, provided by means 2, in comparison with a predetermined threshold. The confidence index 7 is a parameter used to develop the security diagnosis relating to the stored values ​​of the model.

[0032] The confidence index 7 acts as an evaluation scale, for example, as a percentage between 0% and 100%. However, this is not mandatory. The purpose of this confidence index 7 is to quantify the degree of confidence in the reliability of the stored value and to provide an indication of what happened in previous update loops. The value of the initial predetermined level, configurable at the beginning of the learning cycle for the first stored value of a term in the learning matrix, is, in this example, the average intermediate value of 50%. The configuration and determination means 3 is capable of incrementing and decrementing the index value in predetermined steps of constant value, between 1% and 10%, for example. Within the framework of the invention, the step determines the number of update loops before triggering a re-evaluation of the stored value.

[0033] The security module 1 further includes a means 4 for determining the value of a security parameter 8 taking The input parameters are the value of the deviation 6 and the value of the confidence index.7 allows for the establishment of a security diagnosis in the stored values ​​10 and the current values ​​12 of the control model during an update loop. The determination means 4, in this embodiment, provides a weighting coefficient value 8 that allows for adjusting the proportion of the current value 12 of the model term calculated from the instantaneous measurements for updating the learning model 13. More precisely, depending on the confidence level 7 and the error 6, the weighting coefficient allows for storing only a portion of the current value 12 during the update, ignoring the current value 12 in the case of a very large error if the confidence level is high, a significant portion if the confidence level of the stored value 10 is low, i.e., has reached a predetermined minimum limit, and the full value if the error is zero.A method for calculating the safety parameter 8 will be described as an example later in this document. The calculation of the safety parameter 8, in the form of a weighting parameter, can take the form of a matrix, providing predetermined values ​​for two input variables: the deviation 6 and the confidence index 7. The values ​​of the weighting parameter 8 are, for example, between 0 and 1.

[0034] The safety module 1 further includes a determination means 5 whose function is to update a term of the model with a third value to be stored, consisting of the sum of parts of the first value 10 and the second value 12. The part of the second value 12 is calculated from the value of the weighting parameter 8 assigned to the current value of the model term. The determination means 5 outputs the new value 14 corresponding to the current value of the richness correction to which the weighting parameter 8 is assigned for its update by the learning function 11. The new value stored in the learning matrix 13 by the learning function consists of the sum of the parts of the first value 10 and the second value 12, which are calculated according to the weighting coefficient 8. The new stored value is obtained using a centroid calculation.More specifically, the share of the second value 12 corresponds to the product of this second value 12 with the weighting coefficient 8 and the share of the first value 10 corresponds to the product of this first value 10 with a weighting term equal to one minus the value of the weighting parameter 8. The updated values ​​to be stored are recorded in the memory of the motor control unit, in a storage means with volatile and non-volatile memory and programmable so as to be reused later, in particular when the motor stops and resumes operation.

[0035] We now describe through the figure 2 an example of an algorithm of the first preferred embodiment of the securing process according to the invention in which the securing parameter 8 is a weighting parameter determining the part of the current value 12 of the term of the model to be memorized in the learning matrix.

[0036] A learning cycle according to the invention provides, for a point, each operating point, a zone, or each motor operating zone, an initialization phase 200 of the learning model. At this stage, the learning matrix contains no control term values ​​for the operating point or operating zone subject to control.

[0037] For initialization purposes, in a first step, the process performs a check 20 of the learning conditions specific to the function, during which it verifies, in particular, whether the engine is operating at a stable speed and load. This stabilization condition is not limiting. Other conditions may trigger the calculation of a correction term for the model's fuel mixture. If stable engine operation is detected, the process stores, in step 21, the entire current value of the control term derived from the instantaneous measurements; the weighting coefficient 8 is therefore set to 1. In step 22, the process sets the confidence index assigned to this term to the initial value, for example, 50%.The initial confidence level represents an average level, that is, the level below which the security diagnostic system doubts the value and the level above which it considers the value reliable. This confidence level will be used in conjunction with the gap value during the learning cycle to determine at which update loop the convergence of the stored value to the current value can be accelerated.

[0038] Next, the safety process enters phase 201, which updates the learning matrix for the control term or each control term for which an initial value is recorded. This update phase is recursive and is triggered by the detection of optimal learning conditions, through operation under stable operating conditions and load.

[0039] At a monitoring stage 23, when the learning function 11 detects for a term of the model that the learning conditions are met to perform an update of the model terms, the learning function at a stage 24 calculates a new current value of the model richness correction term from the instantaneous measurements.

[0040] The securing process then includes a step 25 of calculating the difference between the stored value of the correction term by the learning function 11, from the previous loop, and the current value of said term estimated from instantaneous measurements by the learning function 11.

[0041] The process includes a step 26, during which the safety function 1 determines the confidence level assigned to the stored value of the model term in the training matrix. During the second update loop of a term in the training matrix, the safety function reads the value 50% in this example. For a subsequent update loop, the confidence level will depend on the progress of the training.

[0042] Next, the process includes a step 27 for determining the value of a weighting coefficient to be assigned to the current value for updating the training matrix. The weighting coefficient is the result of a safety diagnostic evaluation that depends on the deviation calculated during this update loop and the confidence level associated with the stored value of the model term. Based on the result of the safety diagnostic, the safety function 1 defines the value of the weighting coefficient, which will be multiplied by the current value of the model's richness correction term.

[0043] The determination of the weighting coefficient is carried out in this example case by a diagnostic matrix having as input parameters the said gap and the confidence index.

[0044] In figure 3We have represented a non-limiting example of the matrix. The first column represents confidence index values, graduated between 0% and 100%, of the stored value, and the first row represents the value of the difference between the stored value and the current value of the model term, between 0 and 0.1.

[0045] The security diagnosis is carried out according to the following strategy to calculate the weighting coefficient.

[0046] In the case where the difference between the stored value and the current value is zero, the weighting coefficient is equal to 1. Not showing any difference between the learning model and the actual measurements, the learning strategy records the entire current value.

[0047] In the case where the confidence index is greater than or equal to the initial confidence level, in this example 50% and: Furthermore, if the discrepancy is greater than or equal to a first threshold S1 (0.01 in this example), and less than or equal to a second threshold S2 (0.06 in this example), the weighting coefficient is graduated in a decreasing manner, between the values ​​0.3 and 0.05 for example, proportionally to the discrepancy so that the proportion of the current value in the correction term is strictly less than the proportion of the stored value in the learning matrix. The new current value for the wealth correction is only partially, or minimally, taken into account for the update, primarily because the stored value is, at this stage of the learning cycle, still considered reliable. The objective of the security diagnostic is for the update to occur according to an initial convergence rate graduated according to the value of the discrepancy. The larger the discrepancy, and as long as the confidence index remains reliable, the slower the convergence rate.Furthermore, if the difference exceeds the second threshold S2 (0.06), the weighting coefficient is set to 0 to fully retain the portion of the first stored value. The new current value of the wealth correction term is therefore disregarded. In this situation, particularly because the stored value is considered reliable at this stage of the learning cycle, the convergence of the correction term towards the current value is halted as long as the confidence level remains above the initial 50%.

[0048] However, when the confidence index falls strictly below the initial confidence level (50% in this example) but remains above a first minimum limit (Lim, 0% for example), the weighting coefficient is set to 0. This implies that the current value portion of the wealth correction term is zero and is not included. The safety diagnostic function acts to fully preserve the initial stored value for this uncertainty zone based on the index value. The diagnostic strategy involves temporarily inhibiting changes to the stored value during a number of update loops. At this stage of the learning cycle, there is a phase during which the safety diagnostic function questions both the stored value and the current value.

[0049] However, if this measured discrepancy persists over several update loops, the safety function is configured to update the stored value and, to this end, triggers an accelerated convergence process towards the measured current values. More specifically, to accelerate convergence, if the confidence level is equal to the minimum limit Lim, here 0%, the weighting coefficient is set to 0.5. It is envisaged that other higher values ​​between 0.5 and 1 could be configured so that the proportion of the current value in the correction term is at least equal to or greater than the proportion of the stored value, in order to accelerate convergence towards the current value. The weighting coefficient value is specifically designed to accelerate the convergence cycle compared to the native learning mechanism implemented by the learning function 11.For this purpose, in this situation the weighting coefficient has a value driving a convergence speed towards the current value which is greater than that of the native convergence mechanism of the learning function 11.

[0050] It should be noted that the number of update loops required to reach the minimum limit Lim can be calibrated by the increment and decrement threshold value of the confidence index.

[0051] Then, at a step 28, the security function 1 calculates the new value to be stored in the learning matrix, where the current value of the wealth correction term is weighted by the weighting coefficient from the diagnosis of the previous step.

[0052] Furthermore, the safety process includes a step 29 that updates the confidence index value based on the difference calculated in step 25 between the measured value and the current value of an update loop. When this difference exceeds a predetermined threshold (e.g., S2 equal to 0.06), the confidence index is decremented by the predetermined threshold value. In this case, the safety function 1 detects a significant inconsistency between the stored value and the current value of the richness correction term. The function therefore decreases the confidence level in the stored value. Conversely, when the difference is less than the threshold S2, the confidence index is incremented by the threshold value because it detects little difference between the stored and current values. The threshold value is considered to be different, for example, greater than S2, or between S1 and S2, or even equal to S1.

[0053] Finally, once the new wealth correction value is stored in the training matrix and the associated confidence index is updated, the security process moves to monitoring step 23 of a subsequent update loop in phase 201 of the training. The training consists of running a cycle of update loops, in which each loop is triggered when the training conditions are met to calculate, in this example, a wealth correction value.

[0054] We now describe in figure 4The second embodiment of the security function 11. In this embodiment, the security function 1 uses the same functional blocks 2, 3, and 4 as those implemented in the first embodiment. The references remain unchanged. This embodiment differs, however, in that the security function 1 does not execute function 5, which calculates the value of the wealth correction term to be stored and is described in the first embodiment.

[0055] The safety parameter 8 calculated by module 4 is used to configure an internal parameter of the learning function 11, in which the internal parameter represents a confidence level in the stored value of the model term. A calculation module 51 of the learning function 11 uses this internal parameter to define model noise associated with the stored model terms and to natively drive the convergence process to the current richness correction value. In a non-limiting example, this internal parameter depends on the model's modification history during a training cycle. Within the scope of the invention, the learning function 11 is adapted to modify said internal parameter of the learning function 11 according to the value of the safety parameter 8. This internal parameter is, for example, used by Kalman filter-type learning functions.

[0056] Similar to the first embodiment, module 4 establishes the security diagnosis of the stored values ​​10 and the current values ​​12 based on the difference 6 between the values ​​10 and 12 and the value of the confidence index 7 calculated by module 3. Depending on the result of the diagnosis, the determination method 4 increases or decreases the confidence level of the internal parameter of the learning function, which has the effect, during the update process, of slowing down or accelerating the convergence towards the current value calculated by the learning function 11. According to this embodiment, the modification of the internal parameter during an update loop "n" determines the proportion of the model's stored value used to calculate the new value to be recorded 141 in the next update loop "n+1".

[0057] Moreover, unlike the first embodiment, the safety process does not directly modify the value of the stored correction term, but retains the native management mode of the convergence of the control terms of the learning function 11 and the value of this internal parameter representing the model noise to drive the speed of convergence towards current values.

[0058] Thus, when several update loops are performed and the security function 1 detects a significant constant deviation, the invention makes it possible to accelerate convergence towards the new current value.

[0059] Within the framework of the invention, the security method further includes a function for selecting the convergence mode of a stored value to a current value of the model. This selection is configurable and allows the user to choose, for the learning cycle, either the implementation of the first embodiment or the implementation of the second embodiment. In the first embodiment, the security parameter 8 directly affects the calculation of the proportion of the current value constituting the new stored value, while in the second embodiment, the security parameter 8 directly affects the native internal parameter used by the learning function to determine the proportion of the stored value constituting the new value to be recorded.

[0060] In this second embodiment, the securing process differs from the sequence of the figure 2in that step 27 calculates a safety parameter designed to modify the value of the internal parameter of the learning function representing the confidence level of the model, and in that step 28 consists of modifying this value of the internal parameter to slow down or accelerate the convergence process towards the current value, by increasing or decreasing the proportion of the stored value in the calculation of the new value to be stored. In each update loop, the proportion of the stored value depends on the value of the safety parameter calculated in the previous loop.

[0061] The invention applies to any learning function for models of internal combustion engine actuators, particularly for motor vehicle engines.

Claims

1. Method for securing a learning function (11) of an actuator control model of an internal combustion engine for a motor vehicle, the learning function being based on a model (13) consisting of control terms stored in a memory of an engine control unit and being adapted to operate a learning cycle consisting of update loops of the control terms, the method comprising the following steps: - Calculation (25) of a deviation (6) between a first value (10) stored by the learning function (11) for said term of the model (13) and a second current value (12) of said term of the model (13) estimated from instantaneous measurements, - Determination (26) of a confidence index (7) assigned to the first stored value (10) of said term of the model (13), - Determination (27) of the value of a securing parameter (8) depending on said deviation and on the confidence index, said securing parameter (8) being a weighting coefficient delivered by a predetermined matrix having said deviation (6) and the confidence index (7) as input parameters, and during the step of updating the term of the model (13), the share of the second current value (12) is calculated as a function of the value of said weighting coefficient (8), - Updating (28) the term of the model (13) with a third value consisting of the sum of the shares of the first value (10) and of the second value (12), wherein one of said shares or each of said shares depends on said securing parameter (8),in that it further comprises, at each update loop, a step of updating (29) the value of the confidence index (7) as a function of said deviation (6), and in that when said deviation (6) is greater than a third threshold, the confidence index is decremented by a predetermined value, and when said deviation is lower than said third threshold, the confidence index is incremented by said predetermined value.

2. Method for securing according to claim 1, characterized in that it further comprises, prior to the update loop cycle, an initialization step (22) of the confidence index (7) of the term of the model (13) to an initial confidence level.

3. Method for securing according to claim 2, characterized in that, during the updating of the term of the model (13), if the confidence index is greater than or equal to the initial confidence level and: - if, moreover, the deviation (6) is greater than or equal to a first threshold (S1) and less than or equal to a second threshold (S2), the share of the second value (12) is strictly less than the share of the first stored value (10), - if, moreover, the deviation is greater than the second threshold (S2), the share of the second value (12) is null so as to fully retain the share of the first stored value (10).

4. Method for securing according to claim 2 or 3, characterized in that, during the updating of the term of the model, if the confidence index (7) is strictly less than the initial confidence level and greater than a first minimum limit (Lim), the share of the second value (12) is null so as to fully retain the share of the first stored value (10).

5. Method for securing according to any one of claims 1 to 4, characterized in that during the updating of the term of the model, if the confidence index is equal to a second minimum limit (Lim), the share of the second value (12) is at least equal to or greater than the share of the first stored value (10).

6. Method for securing according to any one of claims 1 to 5, characterized in that the learning function (11) further comprises an internal configuration parameter representative of a confidence level of the stored value of the term of the model (13), and in that said internal parameter is configured as a function of the value of the securing parameter (8).

7. Method for securing according to any one of claims 1 to 6, characterized in that the control module (13) is an air-fuel ratio control function provided for calculating adaptive correction terms, and the learning function (11) is adapted to store, according to a matrix of engine speed and engine load variables, said adaptive correction terms.

8. Engine control unit of an internal combustion engine of a motor vehicle comprising a learning function (11) of a control module of actuators of said internal combustion engine, the learning function (11) being based on a model consisting of control terms stored in a memory of the control unit and being adapted to operate a cycle of updating the control terms, characterized in that it is configured to implement the method for securing the learning function according to any one of claims 1 to 7.

Images