System and method for determining a quantity in a motor vehicle

Abstract

A method and system for iteratively determining a quantity of interest of at least one component of a motor plant used to drive a motor vehicle, the system comprising: a supervised learning estimation module, such as a neural network, taking a first set of input data as input and providing at least one intermediate output quantity (Y) as output; an authenticity verification module (2, MK) taking the first input data set and the second input data set as inputs and providing an authenticity index (IK) as an output; a downstream processing module (3) taking the authenticity index and the intermediate output quantity as inputs and providing a final output quantity (Z) as an output, and when the authenticity index is good, the final output quantity comes from the intermediate output quantity and otherwise from the physical representation model.

Description

Technical Field

[0001] The invention relates to a method for determining a quantity of at least one component of a motor assembly for driving a motor vehicle, in particular when the quantity is not directly measured by a sensor, in which case the quantity should be estimated based on other acquired information and / or other parameters.

[0002] Throughout the text, a (thermal or electric) motor set designates the collection of components that participate directly or indirectly in the traction of the vehicle and may include auxiliary processing components, such as pollution abatement components.

[0003] The quantity indicates at least one physical parameter. It advantageously concerns a temperature, for example the temperature of a rotor of an electric motor, or the temperature of a battery powering the electric motor, or a temperature of a power electronics component, or a temperature of the fuel upstream of the injection in the cylinder, etc. However, the invention is not limited to the case in which the quantity of interest is a temperature. In any case, the quantity of interest relates to at least one parameter of at least one component of the motor unit whose value is used by at least one computer to drive the motor unit.

[0004] In a mass vehicle configuration, it is not possible to envisage sensors capable of measuring these quantities directly. Background Art

[0005] Estimation systems using models based on physical laws are known to determine estimates of quantities of interest. The models use measurable quantities. With the use of algorithmic functions, it is possible to provide as output estimates of quantities of interest that are not directly measurable.

[0006] It is also known to use supervised learning neural networks ("RN" for short). The learning domain is limited to supervised learning operations. Therefore, there are deficiencies in the coverage of cases learned by the RN. In practice, this coverage has therefore proven to be incomplete. If a case is encountered that is not covered by the experimental planning of the learning case of the neural network, there is therefore a risk that the output volume diverges.

[0007] The inventors have sought to improve this situation. Summary of the invention

[0008] To this end, according to the present disclosure, a system is proposed for iteratively determining at least one quantity of interest in a motorization system of a vehicle, said quantity of interest not being directly measured by a sensor, said system comprising:

[0009] a supervised learning estimation module taking as input a first set of input data Xi and providing as output at least one intermediate output quantity Y,

[0010] an authenticity verification module MK which takes as input a first set of input data Xi and a second set of input data Cj and provides as output an authenticity index IK;

[0011] a downstream processing module which takes as input the authenticity index IK and the intermediate output quantity Y and provides as output a final output quantity Z corresponding to said quantity of interest,

[0012] - The downstream processing module is configured such that:

[0013] When the authenticity index IK is good, the final output quantity Z represents the intermediate output quantity Y, preferably equal to the intermediate output quantity Y;

[0014] When the reality index IK is bad, the final output quantity comes in particular from a physical representation model (or "physical model" for short).

[0015] Note that the supervised learning estimation module can be, for example, a neural network module. However, other types of supervised learning modules can be considered, such as, for example, decision trees, kernel regression, large margin separation. Neural network solutions are now well known and easy to implement.

[0016] The authenticity index classified in the "good" category is basically related to the conditions covered by the learning plan of the supervised learning estimation module (e.g., the neural network module). The authenticity index classified in the "bad" category is basically related to the conditions not covered by the learning plan of the supervised learning estimation module.

[0017] "Good" or "bad" thus designates two categories or classes for the authenticity index (IK), which are separated, for example, by one or more predetermined thresholds (for example according to the authenticity index being a vector of one or more dimensions). At least one threshold is advantageously predetermined taking into account the known learning plan of the supervised learning estimation module and is recorded in the memory of the downstream processing module.

[0018] In other words:

[0019] - when the authenticity index (IK) is good, obtaining the final output quantity Z with the help of the model based on artificial intelligence, the model being obtained at the end of the preliminary learning step; and

[0020] When the authenticity index (IK) is bad, the final output quantity Z is obtained, in particular, with the aid of a physical representation model based on physical laws.

[0021] Switching from a good authenticity index to a bad authenticity index thus represents a switch from determining the final quantity by artificial intelligence with the aid of a supervised learning estimation module to determining the final quantity with the aid of a physical model representing the evolution of said quantity according to predetermined physical parameters.

[0022] By virtue of this, it is possible to avoid having to use the abnormal output of the supervised learning estimation module in cases where the specific situation conditions are far from the domain covered by the supervised learning. In other words, the open-loop physical model is selected when the current operating conditions are substantially outside the domain in which the supervised learning is carried out. This open-loop physical model, even if it is not completely accurate or in any case less accurate than the output of the supervised learning estimation module, allows to avoid encountering abnormal values ​​that cause the supervised learning estimation module (e.g. neural network module) to diverge.

[0023] In the most common case, the final output quantity Z is determined with the aid of a supervised learning estimation module. It then has a high degree of accuracy and only limited computing power is required to obtain it.

[0024] There are still few cases where the final output quantity is determined with the aid of a physical representation model. The physical representation model therefore does not need to have a very high accuracy, so that again, only limited computing power is required to determine the final output quantity. The invention thus allows providing a system for iteratively determining at least one quantity of interest, which system requires limited computing power (and therefore has reduced manufacturing costs and / or high computing speed) while providing high accuracy on average.

[0025] The quantity of interest is related to at least one parameter, the value of the at least one parameter being used by at least one computer to drive at least one component of the motor train according to the value derived from the quantity of interest.

[0026] It should be noted that the quantities Xi, Cj, Y and Z are time series of values. Xi and Ci come from the sampling process, and Y and Z are iteratively calculated by the modules mentioned above. Therefore, the symbol Xi actually represents the time series Xi(t k ), the symbol Cj represents the time series Cj(t k ), the symbol Y represents the time series Y(t k ), symbol Z represents the time series Z(t k ). The variable k is here the recursive index of the sampling and calculation. The expression "the final output quantity Z represents the intermediate output quantity Y" is meant to be understood that the final output quantity Z may be directly equal to Y, or indeed corresponds to the value Y after applying a filter (e.g. a smoothing filter). In other words, Z(t k )=F3(Y(t k )), where function F3 is, for example, a digital filter with smoothing and limiting.

[0027] Preferably:

[0028] Z=F(Ci,Xj,Y,t)+CT;

[0029] in:

[0030] Ci is at least a portion of the input data belonging to the second input data set, and preferably the entire second input data set;

[0031] Xj is at least a portion of the input data belonging to the first input data set, preferably the entire first input data set;

[0032] Y is the intermediate output;

[0033] t is the time measured from the time of the most recent switch from a good reality index to a bad reality index (switch to the physical model); and

[0034] CT is a constant.

[0035] Preferably, the constant CT makes it possible to establish continuity of the value taken by the final output quantity Z during the switch to the physical model.

[0036] In various embodiments of the invention, it is possible to also rely on one or more of the following arrangements taken alone or in combination.

[0037] According to one aspect, it can be provided that CT frees the output Z from discontinuity at time t1 when the authenticity index IK changes from good to bad at time t1. Advantageously, even in case of a switch to an open-loop model solution, there is no discontinuity for the drive function using the quantity of interest.

[0038] According to one aspect, the function F is or uses an open-loop model function (YR). Such a function is easy to implement and requires very little memory. The open-loop model function YR represents a good reliable fallback solution, even though it may be slightly less accurate than the output of the neural network.

[0039] According to the present invention, in the authenticity verification module, the authenticity index represents the distance measurement between the first input data set (Xi) and the first reference input data domain and / or between the second input data set (Ci) and the second reference input data domain, and the larger the value of the distance measurement (MD), the "bad" the authenticity index is.

[0040] Conversely, the smaller the value of the distance metric, the “better” the authenticity index.

[0041] When the value of the distance metric exceeds a first predetermined threshold (MD1), a switch to the physical model may thus occur. When the value of the distance metric becomes below a second predetermined threshold (MD2), a return to normal, ie using the output of the neural network, may occur.

[0042] According to one aspect, the quantity of interest is the temperature of a component connected to the motorization system of the vehicle, such as the temperature of the rotor of an electric motor, or the temperature of a battery cell, or the temperature of a power electronics component, or the temperature of the fuel upstream of the injection in the cylinder. The inventors have noted that temperature measurements are particularly well suited for the application of the method and system upgraded herein.

[0043] According to one aspect, the quantity of interest is the temperature of a fluid of the motorization system of the vehicle, for example the temperature of the fuel. In a vehicle with a thermal motor, the temperature of a particular fluid may substantially influence the calibration of the actuation of the fuel injection and the ignition.

[0044] According to yet another variant, the quantity of interest may be the flow rate of a fluid (for example, air in a thermal motor) or a liquid (for example, a fuel for a thermal motor or a coolant for a component of an electric motor).

[0045] According to one aspect, the quantities Y and Z may be multi-dimensional quantities. For example, they may involve multiple temperatures estimated at different locations; for example, for a rotor of an electric traction motor, the temperature of the shaft, the temperature of the windings or permanent magnets, the temperature gradient between one axial side of the rotor and the other axial side may be considered.

[0046] According to one aspect, the downstream processing module comprises a digital filtering function and / or a limiting function. The limiting function allows to exclude any values ​​that are abnormal or outside the normally expected interval, and the filtering function allows to avoid using sudden changes in the drive function of the quantity of interest.

[0047] According to another aspect, it can be provided that three modules, namely the supervised learning estimation module, the authenticity verification module and the downstream processing module, are contained in a single computing unit. This can involve a computing unit different from or aggregated in at least one computer, which drives at least one component of the motor unit as a function of the value obtained by the quantity of interest. Thus, in the case of a thermal motor motorization system, the three modules are, for example, hosted directly in the motor control computer or in the motor unit supervisor. The same is true for the case of electric motorization in which the three modules can be implemented directly in the traction control computer or in the operation supervisor. As a variant, the three modules can be hosted in an auxiliary computing unit connected to at least one auxiliary computer that uses the quantity of interest to drive at least one component of the motor unit.

[0048] According to other implementations, the modules may be in different units.

[0049] According to another aspect, the calculation of the final output quantity (Z) is performed at a frequency at least equal to once per second. Thus, the calculation iterations are fast enough and the provision of information on the quantity of interest is done and updated in real time.

[0050] According to another aspect, the invention also relates to a method for iteratively determining at least one quantity of interest of at least one component of a motor assembly used to drive a motor vehicle, the quantity of interest not being directly measured by a sensor, the method comprising:

[0051] - providing a first input data set X,

[0052] - inject the first input dataset into the supervised learning estimation module,

[0053] - obtaining at least one intermediate output quantity Y as output,

[0054] - providing a second input data set C,

[0055] - injecting the first input data set X and the second input data set C into the authenticity verification module and obtaining an authenticity index (IK) as output,

[0056] - injecting the intermediate output quantity Y and the authenticity index IK into a downstream processing module which provides as output a final output quantity Z corresponding to said quantity of interest,

[0057] - when the authenticity index IK is good, the value representing the intermediate output quantity Y is selected for the final output quantity Z,

[0058] When the reality index IK is bad, values ​​are chosen, in particular from the physical representation model, for the final output quantity.

[0059] In particular, the invention relates to a method for iteratively determining at least one variable of interest in a motorization system of a vehicle.

[0060] As previously detailed, in an advantageous manner, when the authenticity index (IK) is bad, the final output quantity comes from the function F defined by:

[0061] Z=F(Ci,Xj,Y,t)+CT;

[0062] in:

[0063] Ci is at least a portion of the data of the second input data set;

[0064] Wj is at least a portion of the data in the first input data set;

[0065] Y is the intermediate output;

[0066] t is the time from the switch from the most recent good authenticity index to the bad authenticity index; and

[0067] CT is a constant.

[0068] According to the present invention, in the authenticity verification module, the authenticity index represents the distance measurement between the first input data set (Xi) and the first reference input data domain and / or between the second input data set (Ci) and the second reference input data domain, and the larger the value of the distance (MD), the "bad" the authenticity index is.

[0069] On the contrary, the smaller the distance, the “better” the authenticity index is.

[0070] When the value of the distance metric exceeds a first predetermined threshold (MD1), a switch to the physical model may thus occur. When the value of the distance metric becomes below a second predetermined threshold (MD2), a return to normal, ie using the output of the neural network, may occur.

[0071] According to one aspect, CT frees the output Z from discontinuity at time t1 when the authenticity index IK changes from good to bad at time t1.

[0072] And when the authenticity index IK changes from bad to good at time t2, the output Z switches to the output of the neural network, possibly with a discontinuity at time t2.

[0073] Furthermore, by means of a gradual switching system (crossed relative weights W1 W2), output discontinuities in both switching situations can be avoided. In particular, if the predetermined thresholds MD1 and MD2 described above are different, the switching can also be stabilized by a hysteresis mechanism.

[0074] The invention also relates to a computer, characterized in that it comprises a system as described above or that it implements a method as described above. It can directly relate to a computer dedicated to controlling at least one component of a motor unit of a motor vehicle, such as a motor control computer, or to an auxiliary computer connected to at least one computer for controlling at least one component of a motor unit of a motor vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Other aspects, objects and advantages of the invention will become apparent upon reading the following description of embodiments of the invention given by way of non-limiting example. The invention will also be better understood with reference to the accompanying drawings, in which:

[0076] - Figure 1 A block diagram illustrating an example of a system for iteratively determining at least one quantity of interest according to the present invention,

[0077] - Figure 2 represents an illustrative example regarding the timing of outputs of a system for making an iterative determination,

[0078] - Figure 3 Similar to Figure 1and shows a block diagram for a normal situation in which a neural network is fully used,

[0079] - Figure 4 Similar to Figure 1 and shows a block diagram for a situation in which the authenticity verification module takes control and in which the neural network is short-circuited,

[0080] - Figure 5 Similar to Figure 1 And illustrate the elements of change,

[0081] - Figure 6 Schematically represents an example of a supervised learning neural network,

[0082] - Figure 7 Illustration of the distance measurement metric used in the calculation of the authenticity index. DETAILED DESCRIPTION

[0083] In the various drawings, the same reference numerals designate the same or similar elements. For reasons of clarity of the disclosure, some elements are not necessarily shown to scale.

[0084] In this document, the term "module" indicates a collection of electronic circuits capable of performing a given function, which may be programmable and support the execution of software code instructions. By extension, the term module may indicate a complete functional block or even a well-identified physical block.

[0085] Of interest here are the motor assemblies of motor vehicles and in particular the motorization systems of the vehicles, whether they are powered by a fuel-burning internal combustion engine or an electric traction motor or both.

[0086] There are certain quantities of interest that cannot be directly measured in a mass vehicle configuration. According to an example, the quantity of interest may be the temperature of the rotor of the electric motor. According to another example, the quantity of interest may be the fuel temperature upstream of the injection in the cylinder.

[0087] According to yet another example, in the case of a hybrid vehicle or a 100% electric vehicle, the quantity of interest may be the temperature of the battery cells.

[0088] According to yet another example, the quantity of interest may be, for example, the flow rate of a fluid in gas phase (eg intake air in a thermal motor); or the flow rate of a fluid in liquid phase, such as the flow rate of fuel or the flow rate of coolant.

[0089] The variables of interest set out above are used in at least one computer which drives the operation of a thermal or electric motor assembly of a motor vehicle, and in particular the operation of a thermal motor or an electric traction motor of the vehicle, as the case may be.

[0090] During the development phase of a vehicle or motor assembly, engineers use laboratory devices that allow these quantities of interest to be acquired. However, in mass-produced vehicles or motor assemblies, it is not practical to install such sensors and therefore the measurement of the quantity of interest is not directly available. Systems for estimating the quantity of interest should then be used instead.

[0091] In this regard, the present invention proposes a system 10 for providing an estimate of the current value of a quantity of interest when it is not possible to directly measure the current value of the quantity of interest.

[0092] The system 10 includes a neural network module, denoted 1 or RN, see Figure 1 The neural network module 1 comprises: a computer; and a memory storing the prediction model constructed during a previous learning phase. Said phase of constructing the prediction model implements a learning algorithm of the “machine learning” type, which is why in practice a “neural network” module is mentioned in the illustrated example. The prediction model is constructed by using measurements obtained with the aid of sensors that are only implemented during this learning phase and that will no longer be available subsequently when the system 10 is used under real conditions of use.

[0093] It should be noted that a neural network module is only an example among other types of supervised learning modules such as, for example, decision trees, kernel regression, large margin separation. As such, it can be said that the system 10 according to the present invention comprises a "supervised learning module" in a general manner.

[0094] The neural network module 1 takes as input a first set 41 of input data (X1-Xm) and is configured to provide at least one intermediate output quantity Y as output.

[0095] The intermediate output quantity Y is an estimate of the current value of the quantity of interest obtained with the aid of the prediction model stored in the neural network module 1 .

[0096] The first input data Xi may include the volumetric flow rate of the liquid, at least one fluid temperature, and / or at least one fluid pressure, and / or at least one (multiple) mechanical component temperature, and / or at least one speed, and / or the state of at least one actuator, and / or at least one data about the surrounding conditions.

[0097] In the specific case of fuel temperature, the first input data Xi may include the temperature of the motor coolant, the outside air temperature, the average speed of the vehicle over the last minute, the motor load, the fuel flow rate request, the flow rate or amount of fuel consumed during the last 10 seconds, the air temperature under the motor cover, etc.

[0098] In the specific case of the temperature of the rotor of the electric traction motor, the first input data Xi may include the outside temperature, the average speed of the vehicle in the last minute, the load required for the motor in the last 10 seconds, the average speed of the vehicle, the temperature of the stator, one or more temperatures of one or more electric batteries, the air temperature under the motor cover, etc.

[0099] It is important to note that the number m of first input data should not be too large to be able to perform real-time iterative calculations by limiting the amount of memory and CPU resources (computing power) consumed. In other words, the neural network should be of reasonable or even minimal size.

[0100] According to an example, m may be between 5 and 20.

[0101] The system 10 according to the invention also comprises an authenticity verification module, referenced 2 or MK, which takes as input a first set 41 of input data (X1-Xm) and a second set 42 of input data (C1-Cn).

[0102] The authenticity verification module 2 is configured to provide as output an authenticity index identification IK.

[0103] According to an example, n may be between 10 and 30 for the second input data set.

[0104] As in Figure 7 As shown schematically in , the authenticity index IK advantageously characterizes a distance measure MD between the first input data set Xi and the first reference input data domain Eapp.

[0105] The multidimensional space taken into account may also cover the second input data Ci, so that the distance measure MD may also take into account the distance between the second input data set and the second reference input data domain.

[0106] If for each dimension corresponding to the input information (e.g., X1), a coverage interval of supervised learning expressed in the form of [X1min–X1max] is arranged, and if the current value is outside the interval, the distance X1d separating the current X1 from the closest boundary (X1min or X1max) is determined. This calculation is performed m times, i.e., for all variables / dimensions, and then the distance Xid is added. This is an example of a distance metric applied to a multidimensional space. The sum of the squares of the individual distances or any other combination known to those skilled in the art can be selected for the metric.

[0107] As indicated above, the authenticity index IK may also be taken into account according to the same logic with the second set of input data (variable Cj) to calculate the distance Cjd. The distance metric MD may be the sum of Xid and Cjd.

[0108] Preferably, the calculation of the distance metric MD is performed in a multi-dimensional space.

[0109] For example, IK=1 / MD can be chosen. Therefore, if MD is small or zero, the authenticity index IK is high, or in other words good. In this case, and as described in detail below, it can be considered that the intermediate output Y of the neural network RN can be given credibility.

[0110] On the contrary, if MD is large, the authenticity index IK is small, in other words bad. In this case, and as described in detail below, it can be considered that the intermediate output Y of the neural network RN may not be given credibility. As described in detail below, the system according to the invention can then switch to an output representing a physical model, preferably an open-loop model.

[0111] For example, in Figure 7 In , the first current operating condition represented by point P1 is inside the domain Eapp adopted for supervised learning purposes. In this case, the authenticity index IK is good and it can be considered that credibility can be given to the intermediate output Y of the neural network RN.

[0112] In contrast, the second operating condition represented by the point P2 is at a distance MD from the domain Eapp and it may then be decided to switch to an output representing a physical model, preferably an open-loop model.

[0113] As in Figure 7 As shown in the figure, when the value of the distance metric MD exceeds a first predetermined threshold MD1, a switch to the physical model can occur. When the value of the distance metric becomes below a second predetermined threshold MD2, a return to normal, i.e. using the output of the neural network (hysteresis), can occur.

[0114] In the case of large differences or inconsistencies in the set of data Xi, Cj, the authenticity index IK may also give a low value. The authenticity verification module 2 may therefore include an algorithm for verifying the consistency between specific data, which allows detecting an obvious failure of one of the information sources. As a supplement to the distance measurement mentioned above or in lieu of the distance measurement, the authenticity index may represent an inconsistency index of the input data.

[0115] According to the invention, the system 10 further comprises a computer configured to calculate a fallback value denoted YR. The fallback value YR is calculated with the aid of a physical representation model or an open-loop physical model. More precisely, the time series YR(t k ).

[0116] Here, the fallback values ​​YR are calculated within the authenticity verification module 2. As a variant, they can be calculated in a dedicated module or in a downstream processing module described below.

[0117] For example, YR=FR(Ci,Xj), the fallback function FR has been characterized by previous tests on the development system or vehicle.

[0118] Finally, the system 10 according to the invention also comprises a downstream processing module 3 which takes as input the authenticity index IK and the intermediate output quantity Y.

[0119] The downstream processing module 3 also takes into account a backoff value, denoted as YR, which may be directly calculated by the downstream processing module 3 or received as an input by the downstream processing module 3, as in Figure 1 The downstream processing module 3 is configured to provide as output a final output quantity Z corresponding to the quantity of interest.

[0120] According to the present invention, the downstream processing module 3 is configured to determine the value of the final output quantity Z based on the following logic:

[0121] - when the authenticity index IK is good, the final output quantity Z represents the intermediate output quantity Y; and

[0122] When the authenticity index IK is bad, the final output quantity Z comes in particular from the physical representation model. In other words, the final output quantity is then a function of the above-mentioned fallback value YR.

[0123] In the steady state, Figure 3 In the case of normal function shown in the figure, the output essentially comes from the intermediate output quantity Y.

[0124] On the contrary, Figure 4 In the case shown in the figure, the output Z is essentially a function of the fallback value YR.

[0125] To ensure the continuity of the output, the switch between the intermediate value Y and the fallback value YR may be gradual within a time interval of several seconds.

[0126] For example, the following logic can be applied:

[0127] Z=W1×Y+W2×YR,

[0128] Where W1+W2=1, W1 and W2 are the respective weights of the intermediate output value Y and the fallback value YR. This configuration is illustrated in Figure 5 .

[0129] The switch from W1=1 to W1=0 may be gradual, for example within a time window of 5 to 10 seconds. In a similar manner, the switch from W1=0 to W1=1 may be gradual, for example within a time window of 5 to 10 seconds.

[0130] exist Figure 2In FIG. 1 , the evolution as a function of time of a final output quantity as provided as output by the system 10 according to the invention is illustrated.

[0131] Until the instant t1 , the downstream processing module 3 functions in a nominal mode in which the final output quantity essentially results from the intermediate quantity Y coming from the neural network module 1 .

[0132] At time t1 , the switch to the open loop function occurs without discontinuity, with a fallback value YR.

[0133] At time t2, a new switch occurs to return to the situation in which the final output quantity essentially comes from the intermediate quantity Y from the neural network module 1.

[0134] In any case, it is possible to define a function F that determines the value taken by the final output quantity when the authenticity index IK is bad.

[0135] This function F depends at least on a fallback value YR which depends on Ci and Xj (see the fallback function FR mentioned above).

[0136] The function F can also depend on:

[0137] The intermediate value Y is used to ensure the continuity of the output.

[0138] The time t from the switch from the most recent good authenticity index to the bad authenticity index only needs to ensure the continuity of the output within a predetermined time interval from the switch.

[0139] The function F can also be a function of a constant CT for optimal continuity of the output.

[0140] As mentioned above, in steady state, the output essentially comes from the intermediate output quantity Y or the fallback value YR, depending on the situation.

[0141] "Essentially derived from" means that the output may be exactly equal to said intermediate output quantity Y, respectively equal to the back-off value YR, or the output may be derived from filtering and / or limiting said intermediate output quantity Y, respectively said back-off value YR.

[0142] The downstream processing module 3 may therefore comprise a module 5 performing a digital filtering function and / or a limiting function.

[0143] Next refer to Figure 6 The neural network implemented in the neural network module 1 is described.

[0144] The first set of input data X1 to Xm is supplied to the input layer 11 of the neural network. The intermediate layers of the neural network are labeled 12, 13, 14, 15, and the output layer delivers a value, which is an intermediate output variable Y in this case.

[0145] Figure 6 The dotted arrows in illustra te the learning process, in which the coefficients and weights of the nodes of the neural network are adjusted using a cost function calculated on the output for the entirety of the domain in which the supervised learning is performed. Once the learning is performed, the coefficients and weights of the nodes are fixed in the neural network.

[0146] In an advantageous manner, the three modules 1 , 2 , 3 of the system 10 according to the invention are contained in a computing unit which may or may not constitute an integral part of a control unit for driving at least one component of a motor assembly of a motor vehicle.

[0147] Of course, the calculation of the intermediate quantity Y and the calculation of the final output quantity Z are iterative. Preferably, the calculation of the final output quantity Z is performed with a frequency at least equal to once per second. This proves to be sufficient if the quantity of interest is the temperature. According to a specific example, the calculation of the final output quantity Z is performed with a frequency at least equal to once per second or even 4 times per second.

[0148] In the case where the quantity of interest is the fluid flow rate, the frequency of calculations may be at least equal to 10 times per second. It should be noted here that the neural network module performs the intended calculations in a very short time, and it is the upstream sampling process that is the limiting factor. Of course, for specific applications, the frequency may be increased to greater than 10 times per second.

Claims

1. A system for iteratively determining at least one quantity of interest of at least one component of a motor unit used to drive a motor vehicle, said quantity of interest not being directly measured by a sensor, The system comprises: - a supervised learning estimation module (1) which takes as input a first set of input data (X1-Xm) and provides as output at least one intermediate output quantity (Y), - an authenticity verification module (2, MK) which takes as input a first input data set (X1-Xm) and a second input data set (C1-Cn) and provides as output an authenticity index (IK), the authenticity index characterizing a distance measure between the first input data set (Xi) and a first reference input data domain and / or between the second input data set (Ci) and a second reference input data domain, and the greater the value of the distance measure (MD), the more "bad" the authenticity index is, a downstream processing module (3) which takes as input the authenticity index (IK) and the intermediate output quantity (Y) and provides as output a final output quantity (Z) corresponding to said quantity of interest, - The downstream processing module is configured such that: When the authenticity index (IK) is good, the final output quantity Z represents the intermediate output quantity Y, the final output quantity Z is then directly equal to Y or corresponds to the value Y after applying the filter, When the realism index (IK) is bad, the final output is especially from the physically represented model.

2. The system of claim 1, wherein the supervised learning estimation module is a neural network (RN) module.

3. A system as claimed in claim 1 , wherein when the authenticity index (IK) is bad, the final output is derived from a function F defined as follows: Z=F(Ci,Xj,Y,t)+CT; in: Ci is at least a portion of the data of the second input data set; Xj is at least a portion of the data in the first input data set; Y is the intermediate output; t is the time from the switch from the most recent good authenticity index to the bad authenticity index; and CT is a constant.

4. The system of claim 3, wherein CT frees the output Z from discontinuity at time t1 when the authenticity index (IK) changes from good to bad at time t1.

5. A system as claimed in one of claims 3 or 4, wherein the function F is or uses an open loop model function (YR).

6. System as claimed in one of claims 1 to 5, wherein the quantity of interest is the temperature of a component of a motor unit of a motor vehicle.

7. A system as claimed in any one of claims 1 to 5, wherein the quantity of interest is the temperature of a fluid in a motor unit of a motor vehicle.

8. The system of any one of claims 1 to 7, wherein the downstream processing module comprises a digital filtering function and / or a limiting function for forming the final output quantity Z from the intermediate output quantity Y.

9. The system as claimed in any one of claims 1 to 8, wherein three modules, namely the supervised learning estimation module (1), the authenticity verification module (2) and the downstream processing module (3), are contained in a single control unit.

10. A method for iteratively determining at least one quantity of interest of at least one component of a motor assembly used to drive a motor vehicle, the quantity of interest not being directly measured by a sensor, the method comprising: - providing a first input data set (Xi), - injecting the first input data set into the supervised learning estimation module (1), - obtaining at least one intermediate output quantity Y as output, - providing a second set of input data (Cj), - injecting the first input data set and the second input data set into the authenticity verification module and obtaining an authenticity index (IK) as output, the authenticity index characterizing the distance measure between the first input data set (Xi) and the first reference input data domain and / or between the second input data set (Ci) and the second reference input data domain, and the greater the value of the distance measure (MD), the more "bad" the authenticity index is, - injecting the intermediate output quantity (Y) and the authenticity index (IK) into a downstream processing module which provides as output a final output quantity (Z) corresponding to said quantity of interest, - when the authenticity index (IK) is good, a value representing the intermediate output quantity (Y) is chosen for the final output quantity Z, which is then directly equal to Y or corresponds to the value Y after application of the filter, When the reality index IK is bad, values ​​are chosen, in particular from the physically representative model, for the final output quantity.

11. The method of claim 10, wherein when the authenticity index (IK) is bad, the final output quantity is derived from a function F defined as follows: Z=F(Ci,Xj,Y,t)+CT; in: Ci is at least a portion of the data of the second input data set; Xj is at least a portion of the data in the first input data set; Y is the intermediate output; t is the time from the switch from the most recent good authenticity index to the bad authenticity index; and CT is a constant.

12. The method of claim 10 or 11, wherein CT frees the output Z from discontinuity at time t1 when the authenticity index IK changes from good to bad at time t1.

13. A computer (10) for controlling at least one component of a motor unit of a motor vehicle, characterized in that It comprises a system as claimed in one of claims 1 to 9 or it implements a method as claimed in one of claims 10 to 12.

Images