Method for governing an internal combustion engine of a motor vehicle

Abstract

The present invention relates to a method for governing an internal combustion engine (1) of a motor vehicle, said governing method comprising: - a step (E1) of receiving a plurality of first operating parameters (Rm, Car, a, b, d, e) of said internal combustion engine (1) and receiving a first reference parameter, called measurement reference parameter (P Col mes ); - a step (E2) of estimating, using a neural network (711), at least one second operating parameter (F Col ) of said internal combustion engine (1) and estimating a second reference parameter, called estimated reference parameter (P Col est ), based on all or some of the plurality of first operating parameters (Rm, Car, a, b, d, e); - a step (E3) of correcting at least one second operating parameter (F Col ) using an extended Kalman filter (712) to obtain at least one corrected second operating parameter (F Col cor ), said correction step comprising comparing the measurement reference parameter (P Col mes ) and the estimated reference parameter (P Col est ) in order to determine a correction gain (k), the correction of at least one second operating parameter (F Col ) depending on said correction gain (k).

Description

TECHNICAL FIELD

[0001] The present invention relates to a method for governing an internal combustion engine of a motor vehicle, a governing device for implementing this governing method, an internal combustion engine comprising said device and a motor vehicle comprising such an internal combustion engine. BACKGROUND

[0002] In order to improve the overall performance of an internal combustion engine, it is a well-known practice to provide such an internal combustion engine with an EGR (Exhaust Gas Recirculation) valve. This EGR valve allows combustion gases from the exhaust pipe of said internal combustion engine to be recirculated to the intake manifold of the same internal combustion engine. Introducing combustion gases that did not react during combustion into the combustion chamber of the engine makes it possible to reduce the overall temperature of combustion and limit the appearance of a phenomenon known as knock. This phenomenon corresponds to an unwanted self-ignition of the air / fuel mixture in the combustion chamber, which can seriously damage said combustion chamber after a certain period of time.

[0003] The use of an EGR valve has a significant impact on the overall operation of the internal combustion engine. In particular, in a configuration with an EGR valve, the mass of air trapped in each combustion chamber is lower, since combustion gases replace fresh air in said chamber. In order to obtain a stoichiometric air / fuel mixture, i.e. a mixture in which the fuel will be completely consumed in the chamber during combustion, it is necessary to know very precisely the mass fraction of combustion gases (or the mass fraction of fresh gases) in the intake manifold of the internal combustion engine. The mass fraction of combustion gases is understood to be the proportion of combustion gases in the air / fuel mixture. This parameter is then used to optimally control the EGR valve. However, it is difficult to measure the mass fraction of combustion gases using a conventional physical sensor.

[0004] The document FR 2981404 describes a method for governing an internal combustion engine based on an estimation of the mass fraction of burned gases in the intake manifold. More specifically, the mass fraction of burned gases is determined from a mass-based estimation of the burned gases present in a certain volume of mixture upstream of the intake manifold and based on a transport delay. The transport delay corresponds to the time taken to transport a certain volume of mixture to the intake manifold. This corresponds to a dynamic function in the delivery of the air / fuel mixture inside the internal combustion engine. This transport delay is estimated by dividing the intake line according to the thermodynamic conditions and applying the law of laminar flow to each segment. Thus, the document FR 2981404 makes the assumption that the thermodynamic conditions within each zone are stable or that these thermodynamic conditions are linearly variable. Now, the internal combustion engine can be compared to a nonlinear system for which the output parameters are not proportional to the input parameters. Thus, the estimation of the mass fraction of burned gases made by the document FR 2981404 contains a certain inaccuracy which can prove to be incompatible with future European emission standards for gasoline engines, for example the future Euro 7 standard.

[0005] Thus, there remains a need to propose a method for governing an internal combustion engine of a motor vehicle which is better performed, in particular by determining with a higher level of precision the mass fraction of burned gases in the intake manifold of said internal combustion engine. SUMMARY

[0006] The present invention seeks to meet at least partly this need.

[0007] More specifically, the present invention seeks to improve the estimation of the mass fraction of burned gases in the intake manifold of an internal combustion engine.

[0008] A first subject of the present invention relates to a method for governing an internal combustion engine of a motor vehicle, said internal combustion engine comprising a plurality of components, such as an intake manifold, at least one combustion chamber, an EGR valve, said control method comprising:

[0009] - a step of receiving a plurality of first operating parameters and a first reference parameter, called a measurement reference parameter, relating to said internal combustion engine;

[0010] - a step of estimating, using a neural network, at least a second operating parameter of said internal combustion engine and a second reference parameter, called an estimation reference parameter, based on all or some of the plurality of first operating parameters;

[0011] - a step of correcting at least the second operating parameter using an extended Kalman filter to obtain at least a corrected second operating parameter, said correction step comprising a comparison between the measurement reference parameter and the estimation reference parameter in order to determine a correction gain, the correction of at least the second operating parameter being a function of said correction gain;

[0012] - based on the at least corrected second operating parameter:

[0013] - a step of controlling at least one of the components of the internal combustion engine, such as the intake valve at the inlet of the intake manifold and / or the EGR valve; and / or

[0014] - a step of estimating at least a third operating parameter, such as the level of emission of pollutants or the temperature of the gases leaving the combustion chamber.

[0015] Thus, the estimation of the mass fraction of the combustion gases in the intake manifold is performed by a neural network. By definition, a neural network is an architecture whose design is inspired by the operation of biological neurons. A neural network is an architecture capable of learning by applying the inductive principle, that is to say by learning from experience. A neural network generally consists of a succession of layers, each of these layers taking its input from the output of the preceding layer. Each layer (i) consists of N i neurons, so that the N i-1 neurons of the preceding layer take their inputs. By appropriate learning, the neural network will be able to induce the non-linear operation of the internal combustion engine in order to extract a linearized operation therefrom. The extended Kalman filter is an infinite impulse response filter capable of correcting measurement errors that are incomplete or noisy. It is a filter whose operation is optimal on a range of values of a non-linear system that can appear to be linear. Thus, the extended Kalman filter is particularly suitable for the parameters estimated by the neural network. This filter uses a correction gain to correct all or some of the parameters estimated by the neural network. This correction gain is determined by comparing the reference parameters measured and the reference parameters estimated by the neural network. To determine this correction gain, the extended Kalman filter loops back to a controlled parameter that can be measured by a physical sensor, such as the pressure in the intake manifold. With the combined use of the neural network and the extended Kalman filter, the estimation of the mass fraction of the combustion gases in the intake manifold is improved, and more generally, the management of the internal combustion engine is improved. Thus, the present invention makes it possible to achieve a level of precision of + / - 2 points in the estimation of the mass fraction of the combustion gases, as opposed to + / - 5 points in the prior art.

[0016] In one particular embodiment, the neural network is a recurrent neural network.

[0017] The recurrent neural network is a neural network with excellent data compression properties. Thus, it is capable of storing a long list of earlier values captured for a given parameter and of capturing its dynamic phenomena, thus making it possible to estimate future values thereof.

[0018] In one particular embodiment, the recurrent neural network comprises a long short-term memory layer.

[0019] Short-term memory layers, referred to as LSTMs, which stands for "Long Short-Term Memory", are a specific architecture of neurons that are especially used in the field of deep learning. It is an architecture that is conventionally used in regression or classification on time-based sequences, such as speech or video. Typically, a short-term memory layer comprises a plurality of "feature" units. Each feature unit comprises sub-elements, such as one or more input gates, one or more output gates and one or more forget gates. Using these various gates, the short-term memory layer is able to best capture the delayed dynamic effects of air being transferred to the intake manifold of the internal combustion engine. More specifically, the short-term memory layer is defined by a plurality of dynamic (variable) internal states X. This LSTM layer is thus able to capture the nonlinear dynamic phenomena of the engine. These phenomena are largely responsible for the production of pollutants.

[0020] In one particular embodiment, the neural network comprises at least one fully connected layer connected to said short-term memory layer.

[0021] The fully connected layer of the neural network is a layer comprising a plurality of neurons, each connected to a neuron of the preceding layer, that is to say in the present case the short-term memory layer. This makes it possible to create numerous connections to which the weights to be calibrated can be applied in order to best capture the operation of the internal combustion engine. These numerous connections thus make it possible to best capture the nonlinearities of the operation of said internal combustion engine in order to extract therefrom a linearized operation.

[0022] In another particular embodiment, the first operating parameter is chosen from a first list of operating parameters comprising:

[0023] - the speed of the internal combustion engine;

[0024] - the amount of fuel injected into said internal combustion engine;

[0025] - the position of the fresh air intake butterfly valve;

[0026] - the position of the EGR valve;

[0027] - the position of the intake valve at the inlet of the intake manifold:

[0028] - the position of the blades of the turbine of the turbocharger.

[0029] It is thus possible to introduce into the neural network a set of parameter characteristics of the operation of the internal combustion engine, each of these parameters being able to be measured using a physical sensor in said internal combustion engine.

[0030] In another embodiment, the at least one second operating parameter is chosen from a second list of operating parameters comprising:

[0031] - the pressure in the intake manifold;

[0032] - the temperature in the intake manifold;

[0033] - the mass fraction of combustion gases or the mass fraction of fresh gases in the intake manifold;

[0034] - the fresh air flow rate;

[0035] - the flow rate through the EGR valve;

[0036] - the pressure in the exhaust manifold;

[0037] - the temperature in the exhaust manifold;

[0038] - the pressure in a volume of pipe before the compressor of the turbocharger;

[0039] - the mass fraction of combustion gases in the volume before the compressor of the turbocharger.

[0040] Thus, another set of parameters can be obtained at the output of the neural network. This other set of parameters is estimated. In particular, these parameters are difficult to measure using conventional physical sensors in an internal combustion engine.

[0041] In another embodiment, the first reference parameter is a measured pressure in the intake manifold and the second reference parameter is an estimated pressure in the intake manifold.

[0042] Thus, based on the measured pressure or the estimated pressure in the intake manifold, a correction gain can be determined in the correction step of the piloting method.

[0043] Another subject of the application is a piloting device for piloting an internal combustion engine of a motor vehicle, said internal combustion engine comprising a plurality of components, such as an intake manifold, at least one combustion chamber, an EGR valve, said piloting device comprising:

[0044] - means for receiving a plurality of first operating parameters relating to the internal combustion engine and a first reference parameter, referred to as a measured reference parameter;

[0045] - a neural network for estimating, based on all or some of the plurality of first operating parameters, at least a second operating parameter relating to the internal combustion engine and a second reference parameter, referred to as an estimated reference parameter;

[0046] - an extended Kalman filter for correcting the at least second operating parameter so as to obtain at least a corrected second operating parameter, said filter being designed to make a comparison between the measured reference parameter and the estimated reference parameter so as to determine a correction gain, the correction of the at least second operating parameter being a function of said correction gain;

[0047] - a control device for controlling at least one of the components of the internal combustion engine, such as the intake valve at the inlet of the intake manifold and / or the EGR valve, based on the at least corrected second operating parameter; and / or

[0048] - a device for estimating at least a third operating parameter, such as the level of emission of pollutants or the temperature of the gases leaving the combustion chamber, based on the at least corrected second operating parameter.

[0049] The processing of a system operating in a non-linear way, such as the operation of an internal combustion engine, is not very demanding in terms of computing power. Moreover, neural networks have excellent data compression properties and therefore occupy very little memory. As a result, the device for governing an internal combustion engine containing such a neural network and an extended Kalman filter is simplified and its space occupation in a motor vehicle is limited. Moreover, the extended Kalman filter is able to take non-linear models since the algorithm of the extended Kalman filter linearizes these non-linear models at each step of the calculation. This linearization of the neural network is improved when the neural network is a recurrent neural network. In one particular embodiment, the recurrent neural network comprises a short-term memory LSTM layer. Advantageously, the recurrent neural network also comprises at least one fully connected layer connected to the short-term memory layer. The combination of a short-term memory layer with at least one fully connected layer forms a non-linear neural network model which can then be linearized by the extended Kalman filter. The linearization of such a neural network is made possible by the properties of the activation functions preserved in the short-term memory LSTM layer and in the at least one fully connected layer.

[0050] Another subject of the application is an internal combustion engine comprising an EGR valve and a governing device for governing said internal combustion engine according to the aforementioned other subject of the application.

[0051] In another embodiment, the internal combustion engine is a gasoline engine, especially compatible with the European emission standards starting from Euro 7.

[0052] In another embodiment, the EGR valve of the internal combustion engine is a low-pressure EGR valve.

[0053] Another subject of the application is a motor vehicle comprising an internal combustion engine according to one of the aforementioned subjects. BRIEF DESCRIPTION OF DRAWINGS

[0054] The application will be better understood by reading the detailed description of some embodiments, illustrated by the attached drawings and made in a completely non-limiting way, in which:

[0055] Figure 1 is a schematic view of an internal combustion engine according to the application;

[0056] Figure 2 is a control device for controlling Figure 1 a control device for controlling

[0057] Figure 3 depicts steps of a control method for controlling Figure 2 an internal combustion engine by a control device. Figure 1 DETAILED DESCRIPTION

[0058] The present invention is not limited to the embodiments and variants presented, and other embodiments and variants will be obvious to those skilled in the art.

[0059] In the various figures, identical or similar elements have the same reference signs.

[0060] Figure 1 An internal combustion engine 1 according to the invention is schematically depicted.

[0061] In the present description, the terms "upstream" and "downstream" will be used according to the direction of the gas flow through the engine from the point where fresh air is captured to the point where combustion gases exit the motor vehicle.

[0062] Figure 1 More particularly, an internal combustion engine 1 of a motor vehicle is depicted, the internal combustion engine comprising an engine block 10 provided here with four combustion chambers 101. Each combustion chamber 101 is associated with two inlet valves 102 and two outlet valves 103. The inlet valves 102 and the outlet valves 103 control the entry of air and the exit of combustion gases. These valves 102, 103 are connected to a first crankshaft 104 and to a second crankshaft 105 respectively, these crankshafts being intended to control the reciprocating movement of said valves 102, 103 in the combustion chambers 101. It will also be noted that each crankshaft 104, 105 is connected to a variable-lift intake valve system 106 and to a variable-lift exhaust valve system 107 respectively. The internal combustion engine 1 also comprises a sensor 108 for sensing the speed of the internal combustion engine. This sensor 108 is able to send data relating to the speed Rm of said internal combustion engine 1.

[0063] The engine is in the present example a controlled ignition (gasoline) engine.

[0064] Upstream of the combustion chambers 101, the internal combustion engine 1 comprises an intake line 20 which takes fresh air from the atmosphere and which leads to an intake manifold 21 designed to distribute intake gases to each of the four combustion chambers 101 of the engine block 10. The intake manifold 21 comprises an intake sensor 210 designed to deliver data relating to the pressure P Col mes measured in the intake manifold 21. ​

[0065] The intake line 20 comprises, in the direction of flow of the gas: an air filter 22 filtering fresh air taken from the atmosphere; a fresh air intake butterfly valve 23 able to adjust the flow rate of fresh air in the intake line 20; a compressor 24 compressing the gas; a boost pressure sensor 25 designed to measure the air intake pressure in order to regulate the amount of fuel injected into the combustion chamber 101 ; a main air cooler 26 cooling the gas compressed by the compressor 24; an intake valve 27 or high pressure valve able to supply air to the intake manifold 21. The fresh air intake butterfly valve 23 is designed to deliver data relating to the position a of said valve 23. The intake valve 27 is designed to deliver data relating to the position d of said intake valve 27.

[0066] At the outlet of the combustion chamber 101, the internal combustion engine 1 comprises an exhaust line 30 extending from an exhaust manifold 31 to which the gases combusted previously in the combustion chamber 101 are directed, to a particulate filter 32 able to filter particulate matter from the combustion gases before they are discharged into the atmosphere.

[0067] The exhaust line 30 comprises, in the direction of flow of the combustion gases, a turbine 33 driven in rotation by the flow of combustion gases leaving the exhaust manifold 31, and a broadband UEGO ("universal exhaust gas oxygen") probe 34 making it possible to monitor the quality of the combustion and in particular whether any fuel remains in the combustion gases leaving the combustion chamber 101.

[0068] Here, the turbine 33 is coupled to the compressor 24 by a mechanical coupling such as a transmission shaft, so that the compressor 24 forms a turbocharger with the turbine 33.

[0069] The turbine 33 has a variable geometry in the present example since it comprises blades mounted with the ability to pivot about their longitudinal axis, thus exhibiting a variable pitch. The pitch of the blades is adjusted continuously in the present example by an actuator 35 such as a stepper motor, thus allowing the compressor 24 to be driven at greater or lesser rotational speed. The actuator 35 is also designed to deliver data relating to the position e of the blades in the turbine 33.

[0070] The internal combustion engine 1 further comprises a low pressure combustion gas recirculation line 40. This recirculation line is commonly referred to as an EGR-LP ("exhaust gas recirculation low pressure") line 40. This recirculation line 40 starts at the exhaust line 30 after the particulate filter 32 and leads to the intake line 20 between the fresh air intake butterfly valve 23 and the compressor 24. This EGR-LP line 40 allows some of the combustion gases circulating in the exhaust line 30 to be discharged and re-injected into the combustion chamber 101 in order to reduce the pollutant emissions of the engine, in particular the nitrogen oxide emissions.

[0071] The EGR-LP line 40 comprises a secondary air cooler 41 for cooling the split portion of the combustion gases, a low pressure EGR valve 42 for adjusting the flow rate of the combustion gases to the intake manifold 21. The low pressure EGR valve 42 is designed to deliver data related to the position beta of said EGR valve 42.

[0072] The internal combustion engine 1 also comprises an injection line 60 for injecting fuel into the combustion chambers 101. This injection line 60 comprises a fuel reservoir 61, an injection pump 62 designed to take fuel from the reservoir 61 to compress the fuel, and a distribution line set 63 allowing the distribution of this fuel to four injectors (not depicted) respectively leading to the four combustion chambers 101. The injection pump 62 is also designed to supply data related to the quantity of fuel Car injected into said internal combustion engine 1.

[0073] To control the various components of the internal combustion engine 1, an internal combustion engine governing device 70 is provided. This governing device 70 is a computer comprising a processor (CPU) 71, a random access memory (RAM) 72, a read only memory (ROM) 73, an analog-digital converter 74, and various input and output interfaces.

[0074] By means of these input and output interfaces and various sensors integrated into the internal combustion engine (internal combustion engine speed sensor 108, injection pump 62, fresh air intake butterfly valve 23, low pressure EGR valve 42, intake valve 27, actuator 35, intake sensor 210). These various integrated sensors 108, 62, 23, 42, 27, 35, 210 are designed to deliver data related to the first operating parameters Rm, Car, a, beta, delta, epsilon, P Col mes to the governing device 70. The transmission of these various parameters is indicated by the dashed arrows pointing to said governing device 70. The governing device is also designed to transmit control data to certain components of the internal combustion engine 1, such as the intake valve 27 and / or the low pressure EGR valve 42. The transmission of these control data is indicated by the dashed arrows starting from the governing device 70 and pointing to said controlled components 27, 42.

[0075] Figure 2 The processor 71 of the governing device 70 is illustrated. Here, this processor 71 comprises:

[0076] - a receiving means 710;

[0077] - a neural network 711;

[0078] - an extended Kalman filter 712;

[0079] - a control means 713;

[0080] - an estimation device 714.

[0081] The receiving device 710 is designed to receive a plurality of first operating parameters Rm, Car, a, b, d, e related to the internal combustion engine 1, and a first reference parameter, which is referred to as a measurement reference parameter P Col mes corresponding to a measured pressure in the intake manifold 21 measured by the intake sensor 210. The receiving device 710 is designed to transmit the plurality of first operating parameters Rm, Car, a, b, d, e to the neural network 711. The receiving device 710 is also designed to transmit the measurement reference parameter P Col mes to the extended Kalman filter.

[0082] The neural network 711 is designed to estimate at least a second operating parameter related to the internal combustion engine 1 and a second reference parameter, which is referred to as an estimation reference parameter, based on the first operating parameters Rm, Car, a, b, d, e. Here the second operating parameter corresponds to a mass fraction F Col of combustion gases in the intake manifold 21. The second reference parameter corresponds to an estimated pressure P Col est in the intake manifold 21. The neural network 711 is in the present example a recurrent neural network. More specifically, this neural network 711 comprises a short-term memory layer and at least one fully connected layer connected to the short-term memory LSTM layer. The short-term memory layer allows to predict intermediate (non-physical) values that will be consumed later by one or more fully connected layers. The short-term memory layer allows to capture the dynamics of the system, in particular the dynamics in the production of nitrogen oxides NOx. The one or more fully connected layers feed the output of the short-term memory layer, i.e. the non-physical intermediate values. If several fully connected layers are stacked together, these layers exchange non-physical values with each other. These fully connected layers allow to capture the non-linearities in the production of nitrogen oxides by combining the effects of the inputs. The output from the last fully connected layer will allow to model the emissions of nitrogen oxides.

[0083] Hence, the neural network 711 is perceived as a state model that is similar to a regular but non-linear one. First, the internal state X of the predictive neural network is predicted, and then second the output values Y are predicted. This then gives the following system of equations:

[0084]

[0085] Y = g(X, u)

[0086] where X is the internal state, u is the input of the system, and Y is the output. is the derivative of the state X with respect to time. The function f and the function g are non-linear functions which are a model of the phenomenon to be modeled. The neural network 711 will predict the state X and the output Y from the input u.

[0087] The combination of these layers of neurons makes it possible to take into account delayed phenomena and quite non-linear phenomena, thanks in particular to the functioning of the short-term memory layer and the forget gate. Thus, it is easy to extract a linearization from this neural network 711 at each instant. In particular, the neural network 711 forms a non-linear model, but its coefficients at the operating point can be put back into the form of a linear state model, where: where A and B are matrices extracted from the model of the neural network 711 with h t representing the hidden state at instant t. Moreover, the output Y can be written according to the equation Y = C * X + D * u, where C and D are also matrices. Thanks to the characteristics of the neural network 711, the linearization is performed dynamically.

[0088] In one particular embodiment, the neural network 711 comprises a first short-term memory layer comprising 33 neurons. This layer of neurons is designed to process data sequences of 5 to 10 seconds. In particular, each gate has a parameter allowing it to open and close in order to perform a gain-type operation on the signal. The slowest dynamic effect is defined by the length of the longest sequence. In this embodiment, the neural network 711 also comprises a second fully connected layer and a third fully connected layer. The second and third layers each comprise 33 neurons.

[0089] As regards the learning of the neural network 711, the short-term memory layer is designed to learn either from a physical model or directly from tests. In the case of so-called offline learning, the learning method can be performed in three steps. In a first step, a physical model of 0D type is constructed and identified from tests on an engine test bench or on a vehicle. The 0D model is a model with zero spatial dimension, considering only time. Such a physical model is produced for example using the AMESim tool. This model then becomes the reference and is validated in steady and transient tests such as the WLTC or RDE driving cycles.

[0090] In a second step, a dynamic numerical experimental plan is performed on this 0D model. This allows a large number of combinations to be tested, permitting a wider learning base to be acquired than would be achieved using physical tests alone. Thus, it is possible to test at the limits of engine operation, but which cannot be done on a real engine test bench because of the risk of breaking. Based on these virtual tests at the limits, it is possible to make the neural network learn not to exceed these limits. For example, the neural network can learn not to exceed the turbocharger maximum speed constraint. Moreover, in the 0D model, it is possible to add the rules of the art, which will then be learned by the neural network.

[0091] In a third step, the short-term memory layer of the neural network is made to learn after the input / output of the 0D model has been normalized or standardized. This learning is performed by decomposing the sequences and defining the architecture of the short-term memory layer. The real or virtual tests are driving wheels, maneuvers, cycles set in standards (NEDC, WLTC, RDE, etc. types). These cycles are decomposed into sequences lasting approximately 5 to 10 seconds. To perform these sequences, the solver of ADAM can be used.

[0092] This proposed 3-step method involves using a physical model at the start in order to pre-calibrate the neural network 711. Then, the operating domain is more extensively covered by providing virtual tests to the learning algorithm of the neural network. Using real tests alone would be too expensive, with the risk of damaging the engine.

[0093] The neural network 711 then transmits at least a second operating parameter and a second reference parameter called estimated reference parameter to the extended Kalman filter. This second operating parameter is chosen from a list of second operating parameters comprising:

[0094] - the temperature T in the intake manifold 21 Col ;

[0095] - the mass fraction F of combustion gases in the intake manifold 21 Col ;

[0096] - the fresh air flow rate Q 空气 ;

[0097] - the flow rate Q through the EGR valve EGR ;

[0098] - the pressure P in the exhaust manifold Echap ;

[0099] - the temperature T in the exhaust manifold Echap ;

[0100] - the pressure P in a volume of pipe before the compressor 24 of the turbocharger C ;

[0101] - the mass fraction F of combustion gases in the volume before the compressor 24 of the turbocharger C .

[0102] In the embodiment of Figure 2 , the second operating parameter transmitted via the neural network 711 is the mass fraction F of combustion gases in the intake manifold 21 Col , and the second reference parameter transmitted is the estimated pressure P in the intake manifold 21 Colest The extended Kalman filter is able to correct unmeasurable quantities (such as the mass fraction of burned gases in intake manifold 21) by looping the correction back onto the controlled measurements (such as the pressure P Col mes measurements in intake manifold 21). The extended Kalman filter 712 operates in two stages. The first stage is the prediction stage, which involves solving the following equations:

[0103] Xp = A * X + B * u (estimate of the predicted state)

[0104] Pp = A * P * A' + Q (estimate of the predicted covariance)

[0105] In the second stage of correction, the extended Kalman filter 712 involves solving the following equations:

[0106] Y = Z - H * Xp (innovation or residual measurement)

[0107] S = H * Pp * H' + R (innovation or covariance measurement)

[0108] K = Pp * H' * S"1 (approximate optimal Kalman gain)

[0109] X = Xp + K * Y (update estimate of the state)

[0110] P = (I - K * Y) * Pp (update estimate of the covariance)

[0111] In these equations:

[0112] - A and B are matrices of the model extracted from the neural network 711;

[0113] - Xp and X are the states of the extended Kalman filter at time t+1 and time t;

[0114] - Pp and P are the covariance matrices at time t+1 and time t;

[0115] - Z is the loopback measurement (the pressure P Col mes measurements in intake manifold 21);

[0116] - H is the sensor model based on the state X;

[0117] - Q and R are tuning matrices that contain the concept of measurement noise and modeling noise;

[0118] - I is the identity matrix, which is all ones along the diagonal and zero values in the rest of the matrix.

[0119] The only parameters to be calibrated in the extended Kalman filter 712 are the matrices Q and R. R is a matrix related to the modeling noise. This matrix is calibrated according to an estimation of the model accuracy. Assuming that the model is normalized or standardized, that is to say that all the values are between -1 and 1, the matrix R is limited to a diagonal matrix with multiplication coefficients to be calibrated. This coefficient of the matrix R is determined empirically. The starting point is therefore the initial value of the first application of the same estimator and it is optimized to maximize the convergence with respect to the actual tests. It will also be noted that R reflects the measurement accuracy on the internal combustion engine in the same way as the measured pressure Pcol in the intake manifold ejected in the extended Kalman filter. Q is a matrix related to the measurement noise and is calibrated according to the level of accuracy of the sensors. For example, for the intake sensor 210 measuring the intake pressure P Col mes in the intake manifold 21, an accuracy of + / - 5% can be assumed.

[0120] In other words, the extended Kalman filter 712 comprises a comparison device 7121 and a correction device 7122. The comparison device 7121 is designed to compare the measured reference parameter P Col mes transmitted by the reception device 710 with the estimated reference parameter P Col est transmitted by the neural network 711. This comparison makes it possible to determine a correction gain k. Based on this correction gain k, the correction device 7122 is designed to correct the second operating parameter F Col in order to obtain a corrected second operating parameter F Col cor . The corrected second operating parameter F Col cor is then transmitted to the control device 713. This control device is designed to transmit instructions to the intake valve 27 and / or to the EGR valve 42 as a function of the corrected second parameter. In Figure 2 an embodiment, the processor 71 also comprises estimation devices 714. These estimation devices 714 are designed to estimate a third operating parameter, such as the level of emissions of pollutants or the temperature of the gases leaving the combustion chamber, as a function of the corrected second parameter in order to prevent overheating at the turbine 33.

[0121] Figure 3 A method for governing an internal combustion engine 1 of a motor vehicle according to the application is illustrated, said internal combustion engine 1 comprising a plurality of components, such as an intake manifold 21, at least one combustion chamber 101 and an EGR valve 42. This control method comprises:

[0122] - receiving a plurality of first operating parameters Rm, Car, a, b, d, e related to said internal combustion engine 1, called measured reference parameters P Colmes Step E1 of the first reference parameter;

[0123] -Based on all or some of a plurality of first operating parameters Rm, Car, α, β, δ, ε, a neural network 711 is used to estimate at least a second operating parameter F of the internal combustion engine 1. Col and is called the estimated reference parameter P Col est Step E2 of the second reference parameter;

[0124] - Use an extended Kalman filter 712 to correct at least the second operating parameter F. Col To obtain at least the corrected second operating parameter F Col cor Step E3, the correction step includes measuring the reference parameter P Col mes With the estimated reference parameter P Col est The comparisons are made between the two parameters to determine the correction gain k, which is at least the second operating parameter F. Col The correction is a function of the correction gain k;

[0125] -Based on the at least modified second operating parameter F Col cor :

[0126] - Step E4, which controls at least one of the multiple components of the internal combustion engine (such as the intake valve 27 at the inlet of the intake manifold 21 and / or the EGR valve 42); and / or

[0127] - Step E5: Estimate at least a third operating parameter (such as the level of pollutant emissions or the temperature of the gas leaving the combustion chamber).

[0128] In a preferred embodiment, the neural network 711 is a recurrent neural network. Advantageously, the recurrent neural network 711 includes a short-term memory layer.

[0129] In a preferred embodiment, the neural network 711 includes at least one fully connected layer connected to the short-term memory layer.

[0130] Correction step E3 is performed by the extended Kalman filter 712.

[0131] In a preferred embodiment, the first operating parameter is selected from a list of first operating parameters including the following:

[0132] - Internal combustion engine speed Rm;

[0133] - The amount of fuel injected into the internal combustion engine 1, Car;

[0134] - position a of the fresh air intake butterfly valve;

[0135] - position b of the EGR valve;

[0136] - position d of the intake valve at the inlet of the intake manifold;

[0137] - position e of the blades of the turbine of the turbocharger.

[0138] In a preferred embodiment, the at least one second operating parameter is selected from the list of second operating parameters comprising:

[0139] - temperature T in the intake manifold 21 Col ;

[0140] - mass fraction F of combustion gases in the intake manifold 21 Col ;

[0141] - fresh air flow rate Q 空气 ;

[0142] - flow rate Q through the EGR valve EGR ;

[0143] - pressure P in the exhaust manifold Echap ;

[0144] - temperature T in the exhaust manifold Echap ;

[0145] - pressure P in a volume of duct before the compressor 24 of the turbocharger C ;

[0146] - mass fraction F of combustion gases in the volume before the compressor 24 of the turbocharger C .

[0147] As a preference, the first reference parameter is the measured pressure P in the intake manifold 21 Col mes , and the second reference parameter is the estimated pressure P Col est in the intake manifold 21.

[0148] Another subject of the application is an internal combustion engine 1 comprising an EGR valve 42 and a control device 70 for controlling said internal combustion engine 1.

[0149] As a preference, the internal combustion engine 1 is a gasoline engine, in particular an engine compatible with the Euro 7 European emission standard.

[0150] As a preference, the EGR valve 42 of said internal combustion engine 1 is a low-pressure EGR valve.

[0151] Another subject of the application is a motor vehicle comprising the internal combustion engine 1.

[0152] The application is not limited to the presented embodiments and variants, and other embodiments and variants will become apparent to those skilled in the art.

[0153] Thus, the loopback parameter of the extended Kalman filter can be different from the pressure in the manifold. The loopback parameter can be any parameter in the first list of operating parameters.

[0154] Thus, the correction parameter can be different from the mass fraction of the combustion gases in a certain volume before the compressor 24. The correction parameter can be any parameter in the second list of operating parameters.

[0155] Thus, the learning step of the neural network 711 is not offline, but is performed directly on the motor vehicle.

[0156] Thus, the internal combustion engine is a diesel engine comprising a low-pressure EGR valve and a high-pressure EGR valve.

Claims

1. A governing method for governing an internal combustion engine (1) of a motor vehicle, said internal combustion engine (1) comprising a plurality of components, including a turbocharger having a turbine (33), said governing method comprising: - a step (E1) of receiving a plurality of first operating parameters (Rm, Car, a, b, d, e) related to the internal combustion engine (1) and a first reference parameter, called measurement reference parameter (P Colmes ), - a step (E2) of estimating, using a neural network (711), at least a second operating parameter (F Col ) of said internal combustion engine (1) and a second reference parameter, referred to as estimated reference parameter (P Colest ), based on all or some of said plurality of first operating parameters (Rm, Car, a, b, d, e); - a step (E3) of correction of the at least second operating parameter (F Col ) using an extended Kalman filter (712) to obtain at least a corrected second operating parameter (F Colcor ), said step of correction comprising a comparison between the measured reference parameter (P Colmes ) and the estimated reference parameter (P Colest ) in order to determine a correction gain (k), the correction of the at least second operating parameter (F Col ) being a function of said correction gain (k); - based on the at least corrected second operating parameter (F Colcor ): - a step (E4) of controlling at least one component of the plurality of components of the internal combustion engine; and - a step (E5) of estimating at least a third operating parameter in order to prevent overheating at the turbine (33).

2. The management method according to claim 1, wherein The neural network (711) is a recurrent neural network.

3. The method of claim 2, wherein, The recurrent neural network (711) comprises a long short-term memory (LSTM) layer.

4. The method of claim 3, wherein, The neural network (711) comprises at least one fully connected layer connected to said long short-term memory layer.

5. The method of managing according to any one of claims 1 to 4, wherein, The internal combustion engine (1) comprises an intake manifold (21), at least one combustion chamber (104), an EGR valve (42).

6. The method of claim 5, wherein, The plurality of first operating parameters is selected from a list of first operating parameters comprising: - an internal combustion engine speed (Rm); - an amount of fuel injected into the internal combustion engine (1) (Car); - a position of a fresh air intake butterfly valve (a); - a position of the EGR valve (b); - a position of an intake valve at an inlet of the intake manifold (d); - a position of blades of a turbine of a turbocharger (e).

7. The method of claim 6, wherein, The at least one second operating parameter is selected from a list of second operating parameters comprising: - the temperature (T Col ) in the intake manifold (21); - the mass fraction (F Col ) of the combustion gases in the intake manifold (21); - fresh air flow rate (Q 空气 ); - the flow rate (Q EGR ) through the EGR valve; - the pressure (P Echap ) in the exhaust manifold; - the temperature (T Echap ) in the exhaust manifold; - the pressure (P) in a volume of the duct preceding the compressor (24) of the turbocharger C ); - the mass fraction (F C ) of the combustion gases in the volume preceding the compressor (24) of the turbocharger.

8. The method of claim 5, wherein, The first reference parameter is a measured pressure (P Colmes ) in the intake manifold (21), and the second reference parameter is an estimated pressure (P Colest ) in the intake manifold (21).

9. The method of claim 5, wherein, The step (E4) of controlling at least one component of the plurality of components of the internal combustion engine comprises controlling an intake valve (27) at an inlet of the intake manifold (21) and said EGR valve (42).

10. The method of claim 5, wherein, The third operating parameter is a temperature of gases leaving the combustion chamber.

11. A governing device for governing an internal combustion engine of a motor vehicle, said internal combustion engine (1) comprising a plurality of components, including a turbocharger having a turbine (33), said governing device comprising: - means (710) for receiving a plurality of first operating parameters (Rm, Car, a, b, d, e) related to the internal combustion engine (1) and a first reference parameter, called measurement reference parameter (P Colmes ), - a neural network (711) for estimating, on the basis of all or some of the plurality of first operating parameters (Rm, Car, a, b, d, e), at least a second operating parameter (F Col ) related to said internal combustion engine (1) and a second reference parameter, referred to as estimated reference parameter (P Colest ) ; - an extended Kalman filter (712) for correcting the at least second operating parameter (F Col ) so as to obtain at least a corrected second operating parameter (F Colcor ), said filter (712) being designed to make a comparison between the measured reference parameter (P Colmes ) and the estimated reference parameter (P Colest ) so as to determine a correction gain (k), the correction of the at least second operating parameter (F Col ) being a function of said correction gain (k); - a control device (713) for controlling at least one component of the components of the internal combustion engine on the basis of the at least modified second operating parameter (F Colcor ). and - means (714) for estimating at least a third operating parameter based on the at least corrected second operating parameter (F Colcor ) in order to prevent overheating at the turbine (33).

12. The management device of claim 11, wherein, The neural network (711) is a recurrent neural network.

13. The management device of claim 12, wherein, The recurrent neural network (711) comprises a long short-term memory (LSTM) layer.

14. The management device of claim 13, wherein, The neural network (711) comprises at least one fully connected layer connected to said long short-term memory (LSTM) layer.

15. The management device of any one of claims 11 to 14, wherein, The internal combustion engine (1) comprises an intake manifold (21), at least one combustion chamber (101), an EGR valve (42).

16. The management device of claim 15, wherein, The control means are for controlling an intake valve (27) at an inlet of the intake manifold (21) and said EGR valve (42).

17. The management device of claim 15, wherein, The third operating parameter is a temperature of gases leaving the combustion chamber (101).

18. An internal combustion engine comprising an EGR valve (42) and a governing device for governing said internal combustion engine (1) and as claimed in any one of claims 11 to 17.

19. The internal combustion engine of claim 18, wherein, The internal combustion engine (1) is a gasoline engine.

20. The internal combustion engine of any one of claims 18 or 19, wherein, The EGR valve (42) of the internal combustion engine is a low pressure EGR valve.

21. A motor vehicle comprising an internal combustion engine (1) as claimed in any one of claims 19 or 20.

Images