Method for making learning function of actuator model of heat engine safe

Abstract

The invention relates to a method for securing a learning function of an actuator control model of a thermal engine of a motor vehicle, wherein the method comprises the steps of: - calculating (25) a difference between a first stored value stored by the learning function for a term of the model and a second instant value estimated from an instantaneous measurement for the term of the model, - determining (26) a confidence index assigned to the first stored value for the term of the model, - determining (27) a value of a securing parameter depending on the difference and on the confidence index, - updating (28) the term of the model with a third value constituted by a sum of a part of the first value and a part of the second value, during the update, one of the parts or each of the parts depending on the securing parameter.

Description

[0001] This invention claims priority to French application No. 2100573, filed on January 21, 2021, the contents of which (text, drawings and claims) are incorporated herein by reference. Technical Field

[0002] The present invention relates to a method for making the learning function of an actuator control model for a thermal engine of a motor vehicle safe. Background Technology

[0003] Due to manufacturing variations, wear, fouling, and performance quality, the performance of a vehicle's thermal engine actuators may differ from the performance model integrated into the engine control unit. In the case of actuator models related to the intake and injection branches, this deviation can lead to a variation in the richness of the fuel mixture. This results in excessive consumption and / or increased emissions. It can also affect driving pleasure. It is necessary to correct these richness deviations throughout the vehicle's lifespan.

[0004] Typically, this correction is implemented by an engine control model that performs abundance adjustment, which persistently corrects the injector's timing based on abundance measurements provided by an abundance detector located at the exhaust pipe. Furthermore, to optimize this correction, the engine control unit executes an engine control learning function in parallel, which stores the corrections required for the engine's operating range. More specifically, at a stable point within the engine's operating range, the learning function stores the correction value determined by the engine control model responsible for abundance adjustment, provided the learning conditions are met. For example, the learning process can be implemented using an adaptive type of neural network. Simultaneously, the control unit transmits the stored values ​​in real time, which has the effect of converging the abundance adjustment correction values. Therefore, the adjustment effort is minimized and the correction strength is improved.

[0005] This learning function plays an important role when the abundance adjustment model is unable to correct the abundance (especially when the abundance detector is unavailable), and significantly improves engine control during load transients by minimizing the workload of the abundance adjustment function.

[0006] Document FR3057031A1 describes an abundance correction method proposed by the applicant. The described method is an optimization technique based on least-squares methods with errors present in the system, used to determine terms that correct for errors while improving the model. The method describes testing the quality of the on-the-fly computed correction terms to orient the convergence of the optimization loop. These tests are included in the computation loop of the correction terms and applied to on-the-fly convergence correction. Document FR2979390A1 also describes a Kalman filter-based optimization technique.

[0007] When a new correction value is calculated, the learning strategy aims to learn that value completely and to overwrite the values ​​recorded in the memory of the learning model. However, a drawback arises when the instantaneous value is calculated based on measurements obtained under atypical engine operating conditions (e.g., if a specific event occurs and causes the abundance adjustment system to temporarily deviate). The function thus learns an error value that does not represent the operating range, which can lead to abundance deviations across the entire assigned operating range once the atypical event disappears. Other strategies aim to learn only a portion of the instantaneous correction to limit the negative impact of atypical point events. However, on the other hand, the learning rate decreases if the instantaneous value is correct. Summary of the Invention

[0008] Therefore, there is a need for a learning process to overcome the aforementioned problems and improve engine control functionality. The purpose of this invention is to make the learning model safer in order to limit the impact of point deviations caused by atypical events.

[0009] More specifically, the present invention relates to a safety method for making the learning function of an actuator control model of a motor vehicle's thermal engine safe, the learning function being based on a model consisting of control terms recorded in the memory of the thermal engine's control unit and adapted to execute a learning cycle consisting of an update loop for updating the control terms.

[0010] According to the present invention, the security method includes the following steps:

[0011] - Calculate the difference between a first stored value stored in the model for the term by the learning function and a second instantaneous value estimated by the model for the term based on instantaneous measurement.

[0012] - Determine a confidence index, which is assigned to a first stored value for the term in the model.

[0013] - Determine the value of the security parameter, which depends on the difference and the confidence index.

[0014] - Update the item of the model with a third value, which is composed of a portion of the first stored value and a portion of the second instantaneous value, during the update, one or each of the portions depends on the security parameter.

[0015] According to a preferred embodiment of the present invention, the security parameter is a weighting coefficient given by a predetermined matrix having the difference and the confidence index as input parameters, and during the step of updating the terms of the model, the portion of the second instantaneous value is calculated based on the value of the weighting coefficient.

[0016] According to one variant, the security method further includes an initialization step prior to the period of the update cycle, the initialization step initializing the confidence index of the item in the model to an initial confidence level.

[0017] According to one variation, during the updating of the terms of the model, if the confidence index is greater than or equal to the initial confidence level, and:

[0018] If, furthermore, the difference is greater than or equal to a first threshold and less than or equal to a second threshold, then the portion of the second instantaneous value is strictly less than the portion of the first stored value.

[0019] - If, in addition, the difference is greater than the second threshold, the portion of the second instantaneous value is zero so that the portion of the first stored value is fully preserved.

[0020] According to one variation, during the updating of the terms of the model, if the confidence index is strictly less than the initial confidence level and greater than a first minimum limit, the portion of the second instantaneous value is zero so as to completely preserve the portion of the first stored value.

[0021] According to one variation, during the updating of the terms of the model, if the confidence index is equal to the second minimum limit, then the portion of the second instantaneous value is at least equal to or greater than the portion of the first stored value.

[0022] According to one variation, the first minimum limit is equal to or greater than the second minimum limit.

[0023] According to a second embodiment of the method, the learning function further includes an internal configuration parameter, which represents the confidence level of the stored value of the item in the model, and the internal configuration parameter is configured according to the value of the security parameter.

[0024] According to one variant, the security method further includes a correction step in each update cycle, the correction step correcting the value of the confidence index based on the difference, and wherein the confidence index is reduced by a predetermined value when the difference is greater than a third threshold, and the confidence index is increased by the predetermined value when the difference is less than the third threshold.

[0025] According to one variant, the control module is an abundance adjustment function, which is configured to calculate adaptation correction terms, and the learning function is adapted to store the adaptation correction terms according to a matrix of variables of engine speed and load.

[0026] The present invention provides a control unit for a thermal engine of a motor vehicle, the control unit including a learning function for an actuator control module of the thermal engine, the learning function being based on a model composed of control terms recorded in the memory of the control unit and adapted to perform a cycle for updating the control terms. According to the present invention, the control unit is configured to implement a method for making the learning function safe according to any of the above embodiments.

[0027] The present invention also relates to a motor vehicle comprising a thermal engine controlled by such a control unit.

[0028] This invention relates to a control unit for a thermal engine, the control unit implementing a learning function of an actuator control module of the thermal engine, the learning function being based on a model composed of control terms recorded in the memory of the control unit and adapted to perform a cycle for updating the control terms, the control unit further comprising:

[0029] - A calculation component that calculates the difference between a first stored value stored in the model for the item by the learning function and a second instantaneous value estimated by the model for the item based on instantaneous measurement.

[0030] - A determining component that determines a confidence index, which is assigned a first stored value to the term in the model.

[0031] - A determining component that determines the value of a security parameter, the security parameter depending on the difference and the confidence index.

[0032] - An update component, which is used to update the item of the model with a third value, the third value being the sum of a portion of the first stored value and a portion of the second instantaneous value, wherein during the update, one or each of the portions depends on the security parameter.

[0033] Because of this invention, the method and apparatus for safety-enhancing the learning function can establish diagnostics relating to the reliability of stored values ​​of an engine actuator control model. Updates to the learning function are typically subject to a natural convergence mechanism, the purpose of which is to gradually update the stored values ​​with only a subset of new measurements. When the diagnostics establish that the values ​​of terms in the model are unreliable, the safety-enhancing method can accelerate convergence toward these new measurements. This acceleration is also subject to additional controls to prevent updates at point error values.

[0034] The preferred application of this invention is within the context of abundance control models and thus actively participates in the rapid and robust control of abundance errors throughout the vehicle's lifespan. Generally, this invention helps improve the performance of the learning strategies of engine actuator control models and thus contributes to these performance improvements. Attached Figure Description

[0035] Other features and advantages of the invention will become clearer from the following detailed description of embodiments of the invention, which are described by way of non-limiting example only, and with reference to the accompanying drawings, in which:

[0036] - Figure 1 An apparatus according to the invention is shown for making the learning function of the actuator control model of a thermal engine safe;

[0037] - Figure 2 A preferred embodiment of the algorithm for the security method according to the present invention is shown;

[0038] - Figure 3 An example of a safetyization parameter calculation matrix for this preferred embodiment of the safetyization method according to the present invention is shown, wherein the safetyization parameters are weighted according to abundance correction fit values;

[0039] - Figure 4 A second variation of the device according to the invention for making the learning function safe is shown. Detailed Implementation

[0040] The present invention applies to actuator control models for thermal engines, and particularly to safety-enhancing the learning function of such control models. To this end, in a first preferred embodiment, the invention provides an additional function to safety-enhance the natural learning function, which is capable of establishing safety-enhancing diagnostics related to stored fit values ​​and immediate fit values ​​before recording new values ​​in the learning matrix. In a second embodiment, this additional safety-enhancing function works in conjunction with the learning function to modulate the model noise of the model's terms based on the results of diagnostics performed by the safety-enhancing function. Model noise refers to any configuration parameter in the model that represents the confidence level. Based on the results, the safety-enhancing function accelerates the natural convergence process toward new fit values ​​by reducing the model's confidence level in this variant.

[0041] In this specification, the term "actuator control model" refers to any engine actuator control function. The learning function implements: a storage function that stores items calculated by the control model; an update function that updates the items of the model; and a reconstruction function that reconstructs the items. The term "stored value of the learning function" refers to the value of the control model recorded in the memory of the operating unit of the thermal engine. The term "instantaneous value of the learning function" refers to a value estimated based on instantaneous measurements of a physical quantity from engine sensors (e.g., an abundance detector at the exhaust pipe).

[0042] Embodiments of the present invention will be described more precisely as a non-limiting example of the learning function of an abundance adjustment model. The control model is an abundance adjustment function responsible for calculating correction terms for engine actuators that can adjust the abundance. In this example, the control terms refer to abundance correction adaptation values, and the learning function adapts to storing these abundance correction adaptation values, for example, according to a matrix of variables such as engine speed and load. The instantaneous values ​​are estimated based on engine manifold pressure measurements and the oxygen content of the engine exhaust gas. The correction adaptation values ​​adjust the actuators of the thermal engine's intake and / or injection branches capable of controlling the abundance, i.e., the manipulation of the opening of the gas intake valves and / or the manipulation of the fuel injection duration.

[0043] Figure 1 A first preferred embodiment of the functional modules of a thermal engine control unit is shown, which is capable of implementing the safety method according to the invention.

[0044] The control unit is equipped with an integrated circuit computer and electronic memory, which are configured to execute the safety measures. However, this is not mandatory. In practice, the computer can be external to the engine control unit, while being coupled to it. In the latter case, the computer can be configured itself to have a dedicated computer, which includes, for example, optional dedicated programs. Therefore, the control unit according to the invention can be implemented as having software (or information (or "software")) modules, or circuits (or "hardware"), or a combination of circuits and software modules.

[0045] More specifically, the manipulation unit includes a learning function 11 for the control model of the abundance adjustment actuator, and a safetyization module 1 for making the values ​​of the items to be learned by the learning function 11 safe. The learning function 11 is configured to store correction adaptation values ​​under learning conditions, which are specifically pre-established at stable operating points of engine load and speed or within stable operating regions of engine load and speed. While the learning conditions are established, the learning process continuously executes a loop to update the items of the learning matrix during engine operation. That is, the learning function iteratively or recursively performs calculations of the values ​​of the items of the model (derived from instantaneous measurements of physical quantities obtained by one or more sensors of the engine) to perform evaluation by comparing them with stored values ​​from the previous update loop. Depending on the optimization technique used, the new values ​​stored through the update loop may include all or part of the instantaneous values, or may completely retain the values ​​stored during the previous loop.

[0046] For example, the updated values ​​obtained through the learning function can be obtained using optimization techniques of the least squares type as shown in document FR3057031A1 or a Kalman filter as shown in application FR2979390A1. These two techniques are listed as illustrative examples, and those skilled in the art can adapt implementations of the invention to any learning function used to optimize the control model of an engine actuator.

[0047] According to the present invention, the safety module 1 includes a calculation unit 2 adapted to receive at its input a first stored value 10 and a second instantaneous value 12, the first stored value being stored by the learning function 11 for items of the model, and the second instantaneous value 12 being estimated based on instantaneous measurements. The calculation unit 2 is configured to calculate the difference between the two values. The safety module 1 can be used to evaluate points, multiple points, regions, or each region of engine operation defined by the learning matrix. The calculation of the difference is adapted to be performed in each loop of updating the learning function. The value of the difference 10 is given at the output of the module 2.

[0048] The security module 1 further includes a configuration and determination component 3 for configuring and determining a confidence index 7, which is assigned to a first stored value 10 of the item in the model in the learning matrix. The calculation component 3 is adapted to configure an initial predetermined confidence level and to modify the value of the confidence index by comparing it with a predetermined threshold based on an evaluation of the difference 6 given by component 2. The confidence index 7 is a parameter capable of creating security diagnostics related to the stored values ​​of the model.

[0049] A confidence index of 7 is used as an evaluation level, for example, between 0% and 100%. However, this is not mandatory. The purpose of the confidence index 7 is to quantify the level of confidence related to the reliability of the stored value, and to obtain the index that has appeared in previous update cycles. The initial predetermined level that can be configured for the first stored value of the item in the learning matrix at the start of the learning cycle is, in this example, the average median value of 50%. The configuration and determination component 3 is adapted to increase and decrease the value of the index in predetermined steps (palier) with a constant value (e.g., between 1% and 10%). In the framework of the invention, the step size determines the number of update cycles before triggering a challenge to the stored value.

[0050] The safety module 1 also includes a determination component 4, which uses the value of the difference 6 and the value of the confidence index 7 as input parameters to determine the value of the safety parameter 8. This enables the establishment of safety diagnostics in the stored value 10 and instantaneous value 12 of the control model during the update cycle. In this embodiment, the determination component 4 provides a weighting coefficient value 8, which can adjust a portion of the instantaneous value 12 calculated based on instantaneous measurements for the terms of the model to update the learning model 13. More specifically, based on the confidence index 7 and the difference 6, the weighting coefficient can store only a portion of the instantaneous value 12 during the update, ignoring the instantaneous value 12 if the difference is very large when the confidence index is high, storing a large portion of the instantaneous value 12 if the confidence index of the stored value 10 is small (i.e., has reached a predetermined minimum limit), and storing the complete value of the instantaneous value 12 if the difference is zero. The calculation mode for the safety parameter 8 will be described as an example in the following description. The calculation of the safety parameter 8, which is in the form of a weighted parameter, can be performed in matrix form, which gives predetermined values ​​for the two input variables (i.e., the difference 6 and the confidence index 7). The value of the weighted parameter 8 is, for example, between 0 and 1.

[0051] The security module 1 also includes a determining component 5, which updates the terms of the model with a third value to be stored, the third value being the sum of a portion of a first value 10 and a portion of a second value 12. During this update, the portion of the second value 12 is calculated based on the value of a weighting parameter 8 assigned to the instantaneous value of the term in the model. The determining component 5 provides a new value 14 corresponding to the instantaneous value of abundance correction for updating via the learning function 11, to which the weighting parameter 8 is assigned. The new stored value stored in the learning matrix 13 via the learning function is the sum of a portion of the first value 10 and a portion of the second value 12, these portions being calculated based on the weighting coefficient 8. The new stored value is obtained by centroid calculation. More specifically, the portion of the second value 12 corresponds to the product of the second value 12 and the weighting coefficient 8, and the portion of the first value 10 corresponds to the product of the first value 10 and a weighted term, the value of which is 1 minus the value of the weighting parameter 8. The corrected value to be stored is recorded in the memory of the engine control unit, in a storage component of volatile and non-volatile memory, and can be programmed for later use (especially when the engine is stopped and restarted).

[0052] Now through Figure 2 An algorithmic example of a first preferred embodiment of the security method according to the present invention is described, wherein the security parameter 8 is a weighting parameter used to determine a portion of the instantaneous value 12 of the item of the model to be stored in the learning matrix.

[0053] According to the present invention, the learning cycle includes an initialization phase 200 for initializing the learning model at each engine operating point, each operating point, region, or each engine operating region. In this phase, the learning matrix does not contain any control term values ​​for the controlled operating point or operating region.

[0054] To achieve initialization, in the first step, the method performs control 20 on learning conditions, which function specifically to verify, during this step, whether the engine is operating at a stable speed and load. This stabilization condition is not restrictive. Other conditions may determine the triggering of the calculation of the abundance correction term for the model. Upon detection of stable engine operation, in step 21, the method stores all instantaneous values ​​of the control term derived from instantaneous measurements, with the weighting factor 8 thus configured to 1, and in step 22, the method configures the value of a confidence index (having, for example, 50% of the initial value) assigned to the term. The initial confidence level represents an average level below which the safety diagnostics suspects the value, and above which the safety diagnostics estimate the value to be reliable. During the learning cycle, the confidence index is used in combination with the value of the difference to determine which update cycle can be implemented to accelerate the convergence of the stored values ​​toward the instantaneous value.

[0055] The safetyization method then proceeds to update phase 201, which updates the learning matrix for each control item (with the first value recorded). This update phase is recursive and is triggered by stable rotational speed and load operation when optimal learning conditions are detected.

[0056] In monitoring step 23, when learning function 11 detects the learning condition clustering for the items of the model to perform an update of these items of the model, the learning function calculates a new instantaneous value of the abundance correction item of the model based on instantaneous measurements in step 24.

[0057] The securitying method then includes a calculation step 25, which calculates the difference between the stored value of the correction term from the previous cycle stored by the learning function 11 and the instantaneous value of the term estimated by the learning function 11 based on instantaneous measurements.

[0058] The method includes step 26, during which the safetyization function 1 determines a confidence index value (of the stored values ​​assigned to the items in the model in the learning matrix). During a second update cycle of the items in the learning matrix, the safetyization function reads this value as 50% in this example. For the next update cycle, the confidence index will depend on the changes in the learning.

[0059] The method then includes a determination step 27, which determines the values ​​of weighting coefficients to be assigned to the instantaneous values ​​to update the learning matrix. The weighting coefficients are the result of a safetyization diagnostic assessment, which depends on the difference calculated during the update cycle and on a confidence index associated with the stored values ​​of the terms in the model. Based on the result of the safetyization diagnostic, safetyization function 1 defines the values ​​of the weighting coefficients, which will be multiplied by the instantaneous values ​​of the abundance correction terms of the model.

[0060] In this example, the determination of the weighting coefficients is performed by a diagnostic matrix that takes the difference and the confidence index as input parameters.

[0061] exist Figure 3 The diagram shows a non-limiting example of the matrix. The first column shows the confidence index value of the stored value, graded between 0% and 100%, and the first row shows the value of the difference between the stored value and the instantaneous value of the item in the model, which is between 0 and 0.1.

[0062] The safety diagnostics are performed according to the following strategy to calculate the weighting coefficients.

[0063] When the difference between the stored value and the instantaneous value is zero, the weighting coefficient is equal to 1. The difference between the learning model and the actual measurement is not displayed; the learning strategy records the entire instantaneous value.

[0064] If the confidence index is greater than or equal to the initial confidence level (50% in this example):

[0065] If, furthermore, the difference 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 decreases proportionally to the difference, for example, in increments between values ​​of 0.3 and 0.05, so that the portion of the immediate value of the correction term is strictly less than the portion of the stored value stored in the learning matrix. The new abundance correction immediate value is partially or rarely considered for updates, especially since the stored value is judged to still be reliable at this stage of the learning cycle. The purpose of the safety diagnostics is to ensure that the updates are performed according to a first convergence rhythm, which is graded based on the value of the difference. The larger the difference and the more reliable the confidence index, the slower the convergence rate.

[0066] If, furthermore, the difference is greater than the second threshold S2 (0.06), then the weighting coefficient is equal to 0 so that a portion of the first stored value is fully preserved. Therefore, the instantaneous value of the new abundance correction term is not considered. In this case, especially since the stored value is judged to be reliable at this stage of the learning cycle, the convergence of the correction term toward the instantaneous value is stopped as long as the confidence index is greater than the initialization level of 50%.

[0067] However, when the confidence index becomes significantly less than the initial confidence level (50% in this example) and remains greater than the first minimum limit Lim (e.g., 0%), the weighting coefficient equals 0. This means that a portion of the instantaneous value of the abundance correction term is zero and is not considered. The safety diagnostics aims to completely preserve a portion of the first stored value for this uncertain region based on the index value. The diagnostic strategy is designed to temporarily prevent modifications to the stored value during a number of update cycles. During this phase of the learning cycle, there is a stage where the safety diagnostics question both the stored value and the instantaneous value.

[0068] However, if the measurement difference persists across multiple update cycles, the safety feature is configured to update the stored value and, for this purpose, trigger an accelerated convergence process toward the immediate measurement value. More specifically, to accelerate convergence, the weighting coefficient is configured to a value of 0.5 if the confidence index equals the minimum limit Lim (0% in this case). Other larger values ​​between 0.5 and 1 can be considered to ensure that the portion of the immediate value of the correction term is at least equal to or greater than the portion of the stored value, thereby accelerating convergence toward the immediate value. The value of the weighting coefficient is specifically set to accelerate the convergence cycle relative to the natural learning mechanism implemented by learning function 11. Therefore, in this case, the value of the weighting coefficient manipulates the convergence rate toward the immediate value, which is greater than the convergence rate of the natural convergence mechanism of learning function 11.

[0069] It should be clarified that the number of update cycles required to reach the minimum limit Lim can be calibrated by the step size values ​​of the increase and decrease of the confidence index.

[0070] Then, in step 28, the security function 1 calculates the new value to be stored in the learning matrix, during which the instantaneous value of the abundance correction term is weighted by the weighting coefficients from the diagnosis in the previous step.

[0071] Furthermore, the safetyization method includes a correction step 29, which corrects the value of the confidence index based on the difference between the measured value and the instantaneous value in the update cycle calculated in step 25. When the difference is greater than a predetermined threshold (e.g., S2 equal to 0.06), the confidence index reduces the predetermined value of the step size. In this case, the safetyization function 1 detects a significant inconsistency between the stored value and the instantaneous value of the abundance correction term. This function therefore reduces the confidence level in the stored value. And, when the difference is less than the threshold S2, the confidence index increases the value of the step size because it detects almost no difference between the stored value and the instantaneous value. It is possible to consider making the value of this threshold different, for example, greater than S2, or between S1 and S2, or equal to S1.

[0072] Finally, once the new abundance correction values ​​are stored in the learning matrix and the associated confidence index is corrected, the safetyization method proceeds to monitoring step 23, which monitors the update cycle following learning phase 201. This learning is designed to perform a periodic update cycle during which each cycle is triggered when learning conditions are aggregated to compute the abundance correction values ​​in this example.

[0073] Now through Figure 4A second embodiment of the safetyization function 11 is described. In this embodiment, the safetyization function 1 uses the same operating units 2, 3, and 4 as in the first embodiment. Reference numerals remain the same. However, the difference in this embodiment is that the safetyization function 1 does not perform the calculation function 5 described in the first embodiment for calculating the value of the abundance correction term to be stored.

[0074] The safety parameter 8, calculated by module 4, is used to configure the internal parameters of learning function 11, in which the internal parameters represent the confidence level of the stored values ​​of the items in the model. The computation module 51 of learning function 11 uses these internal parameters to limit model noise associated with the stored items and naturally manipulates the convergence process toward the instantaneous values ​​of abundance correction. In a non-limiting example, these internal parameters depend on the modification history of the model during the learning cycle. Within the framework of the invention, learning function 11 is adapted to modify the internal parameters of learning function 11 according to the value of the safety parameter 8. These internal parameters are used, for example, by a Kalman filter-type learning function.

[0075] Similar to the first embodiment, module 4 is capable of establishing a safety assessment of the stored value 10 and the instantaneous value 12 based on the difference 6 between values ​​10 and 12 and the value of the confidence index 7 calculated by module 3. Based on the result of the assessment, component 4 determines whether to increase or decrease the confidence level of the internal parameters of the learning function, the effect of which is to slow down or accelerate the convergence towards the instantaneous value calculated by the learning function 11 during the update process. According to this embodiment, the modification of the internal parameters determines a portion of the stored value of the model during update loop "n" for calculating the new recordable value 141 for the next update loop "n+1".

[0076] Furthermore, unlike the first embodiment, the security method does not directly modify the stored value of the correction term, but instead retains the natural convergence management mode of the control term of the learning function 11 and the value of the internal parameter representing the model noise, in order to manipulate the convergence speed toward the instantaneous value.

[0077] Therefore, when multiple update cycles are implemented and when the safety function 1 detects a large, unchanging difference, the present invention can accelerate the convergence toward the new instantaneous value.

[0078] Within the framework of this invention, the security method also incorporates a selection function that chooses a pattern in which the stored values ​​of the model converge toward the instantaneous values. This selection is configurable and can be implemented for either the first embodiment or the second embodiment during the learning cycle. In the first embodiment, the security parameter 8 directly affects the calculation of the portion of the instantaneous value constituting the new stored value, while in the second embodiment, the security parameter 8 directly affects the natural intrinsic parameters used by the learning function to determine the portion of the stored value constituting the new value to be recorded.

[0079] In this second embodiment, the security method is similar to... Figure 2 The difference in the sequence lies in that step 27 calculates a safety parameter, which sets the value of an internal parameter representing the confidence level of the model for the learning function, and step 28 aims to modify this value of the internal parameter to slow down or speed up the convergence process toward the instantaneous value, while increasing or decreasing the portion of the stored value stored in the calculation of the new value to be stored. In each update loop, the portion of the stored value depends on the value of the safety parameter calculated during the previous loop.

[0080] This invention applies to any learning function of actuator models for thermal engines (especially engines of motor vehicles).

Claims

1. A method of securing a learning function (11) of an actuator control model of a thermal engine of a motor vehicle, said learning function being based on a model (13) consisting of control terms recorded in a memory of a handling unit of said thermal engine and adapted to perform a learning cycle consisting of an update loop for updating said control terms, characterized in that, The security method includes the following steps: - Calculate (25) the difference (6) between the first stored value (10) of the control term of the model (13) stored by the learning function (11) and the second instantaneous value (12) of the control term of the model (13) estimated based on instantaneous measurement. - Determine (26) the confidence index (7), which is assigned to the first stored value (10) of the control item of the model (13). - Determine the value of the safety parameter (8) (27), which depends on the difference and the confidence index. The safety parameter (8) is a weighting coefficient given by a predetermined matrix, which has the difference (6) and the confidence index (7) as input parameters. During the step of updating the control terms of the model (13), a portion of the second instantaneous value (12) is calculated based on the value of the weighting coefficient (8). - Update the control item of the model (13) with a third value (28), the third value being the sum of a portion of the first stored value (10) and a portion of the second instantaneous value (12), during the update, one or each of the portions depends on the security parameter (8). Furthermore, the security method also includes a correction step (29) in each update cycle, the correction step correcting the value of the confidence index (7) according to the difference (6), and when the difference (6) is greater than a third threshold, the confidence index is reduced by a predetermined value, and when the difference is less than the third threshold, the confidence index is increased by the predetermined value.

2. The method of claim 1, wherein, The security method further includes an initialization step (22) prior to the period of the update cycle, which initializes the confidence index (7) of the control item of the model (13) to an initial confidence level.

3. The method of claim 2, wherein, During the update of the control terms of the model (13), if the confidence index is greater than or equal to the initial confidence level, and: - If, in addition, the difference (6) is greater than or equal to the first threshold (S1) and less than or equal to the second threshold (S2), then the portion of the second instantaneous value (12) is strictly less than the portion of the first stored value (10). - If, in addition, the difference is greater than the second threshold (S2), the portion of the second instantaneous value (12) is zero so that the portion of the first stored value (10) is fully preserved.

4. The security method according to claim 2 or 3, characterized in that, During the updating of the control terms of the model, if the confidence index (7) is strictly less than the initial confidence level and greater than the first minimum limit, the portion of the second instantaneous value (12) is zero so as to fully preserve the portion of the first stored value (10).

5. The safety method according to any one of claims 1 to 3, characterized in that, During the updating of the control terms of the model, if the confidence index is equal to the second minimum limit, then the portion of the second instantaneous value (12) is at least equal to or greater than the portion of the first stored value (10).

6. The safety method according to any one of claims 1 to 3, characterized in that, The learning function (11) also includes internal configuration parameters, which represent the confidence level of the stored values ​​of the control items of the model (13), and the internal configuration parameters are configured according to the value of the security parameter (8).

7. The safety method according to any one of claims 1 to 3, characterized in that, The control model is an abundance adjustment function, which is set to calculate the adaptation correction term, and the learning function (11) is adapted to store the adaptation correction term according to the matrix of variables of engine speed and load.

8. A control unit for a thermal engine of a motor vehicle, the control unit comprising a learning function (11) for an actuator control model of the thermal engine, the learning function (11) being based on a model composed of control terms recorded in a memory of the control unit and adapted to perform a cycle for updating the control terms, characterized in that, The control unit is configured to implement a safety method for making the learning function of the actuator control model of the thermal engine of a motor vehicle safe, according to any one of claims 1 to 7.

Images