METHOD FOR MANUFACTURING A BRAKING SYSTEM COMPRISING MEANS FOR ESTIMATING AT LEAST ONE USEFUL CHARACTERISTIC

Abstract

The present invention relates to a method for manufacturing a braking system comprising means for estimating at least one characteristic (Fθ) useful for controlling the braking system, comprising at least one step of developing a function for calculating said estimated characteristic from selected parameters (C1, C2 …, Cn) and at least one artificial intelligence algorithm (DL1), said step on the one hand using a set of learning data as a function of said selected parameters (C1, C2 …, Cn) and on the other hand integrating at least one physical model representative of the braking system. [Fig.2]

Description

Title of the invention: METHOD FOR MANUFACTURING A BRAKING SYSTEM COMPRISING MEANS OF ESTIMATING AT LEAST ONE USEFUL CHARACTERISTIC TECHNICAL FIELD AND PRIOR ART

[0001] The present invention relates to a method of manufacturing a braking system comprising means for estimating a characteristic useful for controlling the brake and a braking system.

[0002] A motor vehicle is equipped with a brake at each wheel. This may be a disc brake or a drum brake.

[0003] The brake can be a hydraulic brake, an electrohydraulic brake or an electromechanical brake designated EMB (“Electromechanical Brake” in English terminology).

[0004] Parking brakes are increasingly electrically activated, often with hydraulic activation for the service brake function, and it is interesting to be able to produce service brakes with fully electrical actuation. For example, a screw-nut system actuated by an electric motor causes the brake pads to be applied against the disc in the case of a disc brake, and the brake linings against the drum in the case of a drum brake.

[0005] It is desired to be able to know the clamping force exerted by each brake in order to be able to check whether the expected braking level is reached and / or to control the service brake in the case of an electromechanical service brake.

[0006] Furthermore, we wish to reduce or limit the number of sensors, in particular we do not wish to implement a force sensor to know the clamping force.

[0007] Currently, the position of the rotor in the brake is used to deduce the clamping force using predetermined curves linking the position of the rotor to the clamping force. However, it appears that the curves do not always sufficiently reflect the actual behavior of the brake. Indeed, knowledge of the position of the rotor alone is not always sufficient to estimate the clamping force with sufficient precision in certain configurations, in particular for an electromechanical service brake, because a large number of parameters are involved.

[0008] Document EP 3 853 087 describes a braking system in which the braking instructions were generated by a device equipped with a fuzzy logic processor and validated by a decision algorithm. Statement of the invention

[0009] It is an object of the present invention to provide a method of manufacturing a braking system in which the precise estimation of a characteristic useful for braking control requires a reduced number of sensors.

[0010] The aim stated above is achieved by a method for manufacturing a braking system comprising means for estimating at least one characteristic useful for controlling the braking system, comprising at least one step of developing a function for calculating said estimated characteristic from determined parameters and at least one artificial intelligence algorithm, said step on the one hand using a set of learning data as a function of said determined parameters and on the other hand integrating at least one physical model linked to the braking system.

[0011] In other words, a braking system is produced comprising a virtual sensor of the first useful characteristic by integrating into the learning of the artificial intelligence algorithm one or more physical equations representative of the braking system.

[0012] The manufacturing method according to the invention makes it possible to reduce the quantity of data necessary for training the artificial intelligence algorithm and to obtain a braking system in which the useful characteristic is estimated more precisely.

[0013] Furthermore, since the data for training is generally collected by driving a vehicle equipped with the braking system to which the virtual sensor is to be applied, reducing the amount of data required makes it possible to reduce the number of kilometers to be traveled by the vehicle.

[0014] Furthermore, the invention, by integrating one or more physical models, provides the artificial intelligence software with information on expected / realistic output values. The physical model(s) limit the space of solutions admissible by the neural network during the learning phase.

[0015] The term "artificial intelligence algorithm" means any fuzzy logic or artificial intelligence, advantageously self-learning, preferably any neural network, preferably any deep neural network. Neural networks are software and / or hardware neural networks...

[0016] For example, the useful characteristic is the clamping force applied by the brake, the temperature of the brake disc, the wear of the brake linings, the hydraulic pressure in the case of a braking system with at least part hydraulic control.

[0017] In a particularly interesting embodiment, the method according to the invention allows the manufacture of a braking system in which the physical model for calculating the clamping force takes into account the rigidity of the brake, in particular of the caliper in the case of a disc brake. Alternatively, the physical calculation model takes into account other characteristics of the brake, such as the friction properties.

[0018] For example, during the learning phase the artificial intelligence algorithm estimates a clamping force from parameter values, an error relative to the measured value is calculated. Furthermore, a physical error is calculated between the estimated clamping force and the clamping force calculated with the physical model. Then the algorithm is modified to make the sum of the errors tend towards zero.

[0019] For example, from one or more physical models, the artificial intelligence algorithm estimates in addition to the useful characteristic, at least one other characteristic of the brake, for example the stiffness of the brake is estimated.

[0020] In another example, the clamping force and stiffness and / or other characteristics of the brake are each estimated by a different artificial intelligence algorithm, each using its own parameters or characteristics. The resulting virtual sensor provides greater accuracy.

[0021] The present invention then relates to a method for manufacturing a braking system for a motor vehicle comprising at least one brake and means for estimating at least one characteristic useful for controlling said at least one brake, said brake being intended to be mounted on a given motor vehicle, said braking system comprising a control unit comprising a storage area comprising executable machine instructions, a memory and a processor for generating said estimated useful characteristic, said brake being mounted on a test vehicle or a test bench conforming to the given motor vehicle, said method comprising the steps: a) Selection of first parameters of the brake and its environment having an influence on the value of the useful characteristic, b) Selection of at least one physical model making it possible to calculate said useful characteristic, said model comprising at least one variable depending on second parameters of the brake and its environment, c) Establishing a first artificial intelligence algorithm to be trained using said first parameters as input parameters and configured to estimate the useful characteristic, d) Measurement or evaluation by means equipping the brake or said test vehicle of a training data set comprising values ​​of the first parameters and values ​​of the useful characteristic, said useful characteristic being measured by at least one sensor equipping said brake, e) Training said first artificial intelligence algorithm comprising: 1. providing the first parameters as input to the first algorithm, 2. estimating the useful characteristic by the first algorithm, 3. Establishment of a function comprising a term dependent on the measured useful characteristic and a term dependent on the estimated useful characteristic and a term dependent on the mathematical model, 4- modification of the first algorithm so as to make said function tend towards zero, 5- repetition of sub-steps 1 to 4 until said sum is less than a predetermined threshold value and the first artificial intelligence algorithm is considered satisfactory, f) development of a function for calculating the useful characteristic estimated from the first artificial intelligence algorithm considered satisfactory.

[0022] The function of sub-step 3 is for example Ldata+ Lp , with

[0023] [Math.l]

[0024] [Math.2] Lp = ][Fp - FrH

[0025] F being the measured useful characteristic, Fo being the estimated useful characteristic, Fp being the useful characteristic of the mathematical model.

[0026] In an exemplary embodiment, the at least one variable is estimated by means of a second artificial intelligence algorithm, using as input the second parameters, which are at least partly different from the first parameters, and in which said method comprises the measurement or evaluation by means equipping the brake or said vehicle with a set of training data comprising values ​​of second parameters and values ​​of the variable.

[0027] In another exemplary embodiment, the at least one variable is estimated by the first artificial intelligence algorithm.

[0028] Advantageously, the first algorithm and / or the second algorithm are implemented by artificial neural networks. Sub-step 4 then implements, for example, backpropagation.

[0029] The useful characteristic is for example the clamping force exerted by the brake.

[0030] In an exemplary embodiment, the mathematical model is of the CKcyb type, with Kc the rigidity of the brake and Kc the displacement of a part of the brake generating the braking, Kc being the variable

[0031] In another exemplary embodiment, the brake is an electric brake and the mathematical model is of the type

[0032] [Math.3] ,fe= GÀ- GC^gT^tv^J With ia the electric current supplying the motor, com the angular speed of the motor, (bm the angular acceleration of the motor, and in which C4 is the variable.

[0033] According to an additional characteristic, during step f), a polynomial function or a weighting matrix is ​​extracted from the first trained artificial intelligence algorithm considered to be satisfactory intended to be used by the processor of the estimation system to calculate the estimated clamping force or the calculation function is the first trained artificial intelligence algorithm considered to be satisfactory.

[0034] Preferably, the vehicle moves during step d), preferably until reaching a wear determined as sufficient of at least one brake lining.

[0035] The present invention also relates to a computer program medium comprising executable machine instructions for the execution by a processor of a method for estimating a characteristic useful for controlling at least one brake of a braking system, said estimation method implementing the first trained artificial intelligence algorithm considered to be satisfactory obtained by the manufacturing method according to the invention or the polynomial function or the weighting matrix extracted from the first trained artificial intelligence algorithm considered to be satisfactory obtained by the manufacturing method according to the invention.

[0036] The present invention also relates to a braking system for a motor vehicle comprising at least one brake, an electronic control unit for said brake comprising a storage area (Z) comprising executable machine instructions, a memory and a processor (P) for generating at least one useful characteristic estimated for the control of the brake by the control unit, said braking system being obtained by the method according to the invention, said braking system also comprising means for collecting the first selected parameters, said collection means comprising sensors and / or means for evaluating all or part of the first selected parameters.

[0037] In a preferred example, the collection means do not comprise or do not use a sensor for direct measurement of said useful characteristic. Said polynomial function or the weighting matrix can be recorded in the storage area, or the first trained artificial intelligence algorithm considered satisfactory can be recorded in the storage area, in particular in which the processor is a graphics processor. BRIEF DESCRIPTION OF THE FIGURES

[0038] The following description will be better understood with the aid of the attached drawings in which: - [Fig.l] is a schematic representation of a vehicle comprising a braking system capable of implementing the invention, - [Fig.2] is a schematic representation of an example of a method of manufacturing a braking system according to the invention, - [Fig.3] is a schematic representation of a variant of the manufacturing process of [Fig.2], - [Fig.4] is a schematic representation of another example of a method of manufacturing a braking system according to the invention. DETAILED DESCRIPTION OF EMBODIMENTS

[0039] In [Fig.l], we can see a vehicle V, represented schematically, comprising a braking system S comprising electric brakes FRI, FR2, FR3, FR4 each equipping a wheel.

[0040] In this example, the service braking is provided by electric brakes FRI, FR2, FR3, FR4. These advantageously incorporate an electric parking brake.

[0041] Each electric brake comprises an actuator provided with an electric motor and means for converting the rotational movement of the electric motor into a translational movement applying the brake shoes against the brake disc or the brake segments against the drum. The electric brake can be equipped with any type of electric motor, for example a direct current motor, for example a brushless direct current motor.

[0042] The braking system may or may not include an ABS and / or ESP computer.

[0043] The braking system comprises an electronic control unit ECU (Electronic control unit in English terminology) for controlling the FRI, FR2, FR3, FR4 brakes. The brakes are controlled, for example, by operating a brake pedal. Alternatively, the brakes are controlled automatically, particularly in the case of an autonomous vehicle.

[0044] The braking system also comprises means for estimating at least one characteristic useful for controlling one or more brakes. The control unit comprises a processor P, a memory area M, and a storage area Z in which executable machine instructions I and a function for calculating the estimated useful characteristic are loaded, the calculation function Fest being obtained from a trained artificial intelligence algorithm.

[0045] The invention will be described in the context of an application to one of the electric brakes, but it will be understood that the invention applies to each electric brake. Furthermore the invention also applies to braking systems which are at least partly hydraulic.

[0046] The subject of the invention is a method of manufacturing a braking system comprising means for estimating a characteristic useful for controlling braking, in particular so as not to have to resort to a real sensor of this characteristic.

[0047] In the present application, the term "useful characteristic" means any characteristic allowing the control unit of the braking system to control the braking exerted by at least one brake. This concerns, for example, the clamping force exerted by the brake, the temperature of the brake disc, the level of wear of the brake pads, the hydraulic pressure in the case of a braking system using brake fluid. The useful characteristic may relate to the brake, the braking system, or even to the vehicle or other components of the vehicle which may be involved in controlling the braking.

[0048] As a further example, the useful characteristic is selected from vehicle deceleration, vehicle speed, motor supply current, motor rotational position, motor rotational speed, and other derived characteristics such as duration of the last or current braking event, time elapsed since the last braking event, vehicle kinetic energy.

[0049] In the following description, the estimated useful characteristic is the clamping force.

[0050] The term “artificial intelligence algorithm” means automatic learning means (Machine Learning) or artificial neural network, for example a convolutional neural network.

[0051] An artificial neural network consists of several units connected to each other and between which signals are transmitted and distributed in layers. Each neuron receives an input signal which is modified by weights and biases, and generates an output signal which is transmitted to the other neurons by connections or synapses.

[0052] The steps of manufacturing a braking system comprising means for estimating the clamping force according to an exemplary embodiment will now be described.

[0053] Typically, the means for estimating the application force of a brake depends on the brake and the model of vehicle on which the brake is mounted.

[0054] To develop the braking system, the brake(s) of the braking system is or are mounted on a test vehicle or platform, corresponding to the production vehicle on which the brake is intended to be mounted, and the braking system is constructed taking into account this brake environment. The brake and the vehicle are instrumented using real sensors in order to quantify and / or qualify the different parameters. The quantification of a parameter is, for example, a measurement or a calculation or an evaluation, for example by a mathematical model, or by mapping, or even by an artificial intelligence algorithm.

[0055] Alternatively or additionally, the brake is mounted on a test bench which corresponds to the production vehicle model on which the brake is intended to be mounted.

[0056] The production method includes the phase of calibrating the brake on the platform. The term calibration is used here to designate the adaptation of the means for estimating the clamping force, which can form a subset of a broader calibration process.In this sense, this calibration phase comprises: - the selection of parameters of the brake and its environment having an influence on the value of the clamping force, - the selection of at least one physical model making it possible to calculate the clamping force, said model comprising at least one variable depending on the parameters of the brake and its environment, - the establishment of an artificial intelligence algorithm to be trained using the parameters as input parameters and configured to estimate the clamping force, - the acquisition of training data sets for estimating the clamping force and the variable of the mathematical model, - the training of the artificial intelligence algorithm until an artificial intelligence algorithm considered satisfactory is obtained, - the development of a function for calculating the estimated clamping force from the artificial intelligence algorithm considered satisfactory.

[0057] Very advantageously, the calibration phase extends over a sufficient period of time for it to take into account substantial, or even total, wear of the brake linings, which offers the advantage of calibrating the brake over a sufficiently long period and taking into account different states of the brake.

[0058] According to the invention, the design of the means for estimating the clamping force exerted by a given brake on a given platform is tailor-made.

[0059] The different stages will now be described.

[0060] The training of the artificial intelligence algorithm uses a training data set which includes parameter values ​​Cl, C2... Cn of the brake environment and the associated clamping force.

[0061] By "brake environment parameters" we mean vehicle parameters and / or parameters of the terrain on which the vehicle is moving, the weather, etc.

[0062] The parameters Cl, C2... Cn are for example and in a non-limiting manner: the acceleration of the vehicle, the deceleration of the vehicle, the position of the rotor of the motor brake or piston position, brake pad temperature, brake disc temperature, brake disc and / or lining wear that can be measured or estimated, brake pedal application force, vehicle speed, speed at brake activation and brake deactivation, braking duration, vehicle mass that is known, time interval between two braking applications, characteristics of the last braking application(s), number of braking applications, road surface condition, road gradient, brake pad thickness, vehicle type, number of pads per brake, brake pad type, friction coefficient, brake pad surface area, wear coefficient, pedal travel, brake pedal application force, current consumed by the motor, voltage at the motor terminals, vehicle speed, duration of last braking action,weather conditions, such as humidity, ambient temperature, etc.,

[0063] To acquire this data set, the test vehicle is driven for example on a test track. During the calibration phase, measurements of the clamping force are carried out using a physical sensor on the brake mounted on the test vehicle, these values ​​are linked to the parameters Cl, C2..., Cn, the values ​​of which are for some measured using real sensors, for others evaluated using a mathematical model, or mapped in a table, or even estimated using an artificial intelligence algorithm forming virtual sensors.

[0064] The measurements or evaluations of the parameters take place for example every 50 ms, as does the measurement of the clamping force.

[0065] Thus, an actual clamping force value is associated with a set of parameter values. We then have a training data set comprising actual clamping force values ​​associated with the values ​​of a certain number of parameters which have a greater or lesser influence on the clamping force.

[0066] In [Fig.2], a method of manufacturing a virtual sensor of the clamping force applied by a brake according to a first exemplary embodiment can be seen schematically represented.

[0067] We will consider the case of a floating caliper disc brake. The clamping force is proportional to the caliper stroke designated yb

[0068] The manufacturing method aims to establish a function F0(x) which takes into account the input parameters from an artificial intelligence algorithm, x denotes the set of input parameters. In addition, the configuration of the artificial intelligence algorithm makes it possible to discover in a reverse manner other characteristics of the brake, such as stiffness.

[0069] In this example, the physical model characterizing the designated clamping force Fp is as follows. This is a simplified model.

[0070] [Math.4] With yb the caliper stroke and Kc the brake stiffness.

[0071] [Math.5] 1 1 1 L is noted C

[0072] C is considered a constant and its value may not be known. Alternatively, the neural network during its training determines the value of C by inverse analysis.

[0073] Fp can be written:

[0074] [Math.6] FP = CKcyb (i)

[0075] The stiffness Kc varies depending on the stroke of the caliper and different parameters of the braking system, such as the viscoelastic nature of the brake lining depending on the temperature, braking cycles, etc. Kc is considered to be a function of x, x representing the input parameters mentioned above, we can then write Kc(x).

[0076] In the example shown, the artificial intelligence algorithm DL1 is a multi-layer neural network and is configured to provide the estimated clamping value Fo(x) and to provide the estimated brake stiffness Kc(x).

[0077] The untrained neural network DLP1 is designed by considering the selected characteristics. For this, in particular the number of layers, activation functions and connections between neurons, weights and biases are established. Some characteristics of the untrained neural network are established for example by experience and trial and error, while others are set to default values.

[0078] Ldata denotes the error between the measured force F taken from the training data set, F being called Ground Truth in English terminology, and the value of the estimated clamping force Fo:

[0079] [Math.7]

[0080] Lp denotes the error or deviation between the prediction Fo of the network and the value FP resulting from the application of the equations representing the physics of the system applied to the network, in this case formula (I). LP is designated residue.

[0081] [Math. 8] U = HFp -

[0082] When training the neural network, we seek to ensure that the sum Ldata+ Lp tends towards zero.

[0083] The estimated value of Kc extracted by the neural network DL1 and the value of C which depends for example on the rigidity, the forward efficiency of the reducer and the ball nut, the transmission ratio and the screw pitch of the ball nut, and which can be learned by an artificial intelligence algorithm, are used in the sum Ldata+ Lp. C can be determined elsewhere

[0084] The weights and biases of the DL1 neural network are then modified by means of a BR1 backpropagation loop, advantageously by backpropagation of the gradient to reduce the value of Ldata+ Lp.

[0085] A function, called the cost function, which is a weighted sum of Ldata and Lp, is used to update the weights of the neurons in the DLL network. The weighting of Ldata and LP is obtained by the trial and error method. Constant terms can be introduced into the cost function, which will be determined by the neural network during learning, especially if Ldata and LP are of close values.

[0086] The input signals are injected again into the network thus modified, we check Ldata+ Lp, The neural network is trained until Ldata+ Lp is lower than a given threshold. The neural network is then considered satisfactory.

[0087] Training continues until the neural network is satisfactory. The term “neural network considered satisfactory” and more generally “artificial intelligence algorithm considered satisfactory” means a trained artificial intelligence algorithm providing estimated clamping force values ​​located within the limits of the expected error threshold. A large number of steps of modification of the neural network can take place before obtaining the trained DL1 network considered satisfactory.

[0088] Thanks to the invention, the artificial intelligence algorithm is considered to be satisfactory when the difference between the estimated clamping force and the measured clamping force is less than 200 N. In the case of an algorithm obtained solely by training from a data set, it is considered to be satisfactory when the difference between the estimated clamping force and the measured clamping force is less than 500 N. The virtual sensor obtained thanks to the invention therefore makes it possible to increase the control precision of the brake.

[0089] According to an alternative embodiment, Kc is not estimated by the neural network but is known elsewhere, for example by means of a mapping which was carried out upstream, for example during the collection of training data or on a test bench.

[0090] In [Fig.3], we can see an alternative embodiment of the manufacturing method of [Fig.2] which differs from it in that Fu and Kc(x) are estimated by two different neural networks. This alternative has the advantage of being able to use different sets of determined input parameters for F0(x) and Kc(x), which makes it possible to increase the accuracy of the estimation. The parameters Cl, C2.. .Cn are used for the DL1 algorithm for estimating F0(x) and the parameters Cl', C2'...Cn' are used for the DL2 algorithm for estimating Kc(x'), x' designating the set of parameters Cl, C2...Cn.

[0091] As before, the errors are used to create a cost function which, by means of a BRI backpropagation loop, will modify the weight of each of the neurons. Alternatively, a different cost function is used to modify each of the neural networks.

[0092] Advantageously, provision is made to impose an additional constraint during training which is to impose that the estimated clamping force be positive.

[0093] This constraint is obtained for example by means of the ReLU activation function designated "rectified linear unit" or Rectified Linear Unit in English terminology.). If Fo is positive, the value Fo is used, if Fo is negative, Fo takes the value 0. Conversely, the additional term in the cost function takes the value 0 when Fo is positive, and the additional term takes a penalizing value in the cost function when Fo is negative.

[0094] It will be understood that this constraint can be applied during training according to the embodiment example of [Fig.2].

[0095] In [Fig.4], another example of a method of manufacturing a virtual clamping force sensor according to the invention can be seen.

[0096] In this method, we consider the complete mathematical model describing the mechanical behavior of the electric motor:

[0097] [Math.9] ( 11 / 1(L\2 M / 1 I 1 111 L f 11 / 1 i V = V™++r*

[0098] With t = K > i

[0099] Kmotor being the motor torque constant, ia being the supply current of the electric motor, com being the angular speed of the motor, (bm being the angular acceleration of the motor, sgn(com) represents the direction of rotation of the motor.

[0100] It should be noted that in a simplified manner

[0101] [Math. 11] =1 / R

[0102] We can then write:

[0103] [Math. 12] Q fa jn + + (¾.¾ + €4550. With Ci, C2 and C3 which can be considered as constants or as dependent on the input parameters.

[0104] We can then express the stiffness of the brake as a function of the other elements of the model:

[0105] [Math. 13] = Q ' Qsg» (An)

[0106] Lp is then written:

[0107] [Math. 14] A u^ sgïi — rg

[0108] Ldata is equal to F-Fu.

[0109] C4 depends on yb which is the stroke of the electric motor.

[0110] The DL3 neural network is configured to, from input parameters determined according to their influence on the clamping force, provide the output functions F0(x) and extract C4(x).

[0111] The input parameters are those stated above for the example of [Fig.2].

[0112] As in the exemplary embodiment of [Fig.2], the neural network is trained based on at least one training data set. By variation and similar to the example of [Fig.3], the implementation of an artificial intelligence algorithm for estimating the clamping force and an artificial intelligence algorithm for estimating C4(x) with their own input parameters and their own training data is applicable.

[0113] When training the DL3 neural network, we seek to ensure that the sum Ldata + Lp tends towards zero.

[0114] The weights and biases of the neural network are then modified by means of a BR2 backpropagation loop, advantageously by backpropagation of the gradient (“backpropagation” in English terminology).

[0115] The cost function, which is a weighted sum of Ldata and Lp, then makes it possible to update the weights of the neurons in the network, by backpropagation.

[0116] Training continues until the neural network is satisfactory.

[0117] It will be understood that the mathematical model may comprise several functions Ci (x) which are estimated separately or in combination by the same neural network as the clamping force or each by one or more neural networks. The cost function(s) modify each of the networks so that the sum Ldataet Lp tends towards zero.

[0118] One or more more complex physical models or a combination of physical models can be used from which the algorithm(s) will estimate the useful characteristic(s) chosen to control the braking system and extract, during learning, variables characterizing the braking system.

[0119] Furthermore, any type of constraint can be introduced into the Ldataet Lp function to improve the estimation of the useful characteristic(s).

[0120] From the trained intelligence algorithm obtained by the above examples, a clamping force calculation function, generally of the polynomial function type, is extracted. This function is loaded into the storage area of ​​the electronic control unit of the braking system, typically in the form of an implementation in an executable instruction program. This example has the advantage of only requiring a processor offering standard computing power.

[0121] In another exemplary embodiment, it is the trained algorithm, for example a trained neural network or a clone thereof, which is loaded into the computer, preferably a processor offering increased power is then implemented, for example a graphics processor.

[0122] The braking system advantageously comprises means for detecting a deviation from the estimated useful characteristic, for example a Kullback-Leibler divergence. For example, the vehicle may be equipped with a system comparing, for example, the acceleration predicted by the model and the measured acceleration. If this constantly deviates from the measured acceleration, it is considered that the current statistical distribution of the vehicle data set diverges from the entire distribution of the data set used to train the model. Such an alert would then trigger another function which would retrain or recalibrate the already existing model to take this divergence into account. The system may use a characteristic other than acceleration to detect the divergence of the estimate of the useful characteristic.

[0123] The means for estimating the clamping force thus produced are then installed during the manufacture of the vehicle marketed in series. The estimation means of the clamping force are substantially identical to those obtained at the end of the learning stage of the manufacturing process.

[0124] Furthermore, the braking system comprises means, such as sensors or evaluation and / or estimation means, making it possible to obtain the parameters selected for estimating the clamping force.

[0125] For example, the vehicle has at each brake a wheel rotation sensor, a sensor for the supply voltage of the electric brake motor and a sensor for the current consumed by the electric motor. The data which is then provided to the neural network relates to the parameters: wheel speed, electrical voltage applied to the motor and current consumed by the motor.

[0126] The parameters which are estimated are for example the temperature of the brake disc, the wear of the brake pads, the outside temperature.

[0127] It will be understood that these lists are given only as examples, and that they may vary depending on the brake, the vehicle, etc.

[0128] In one example, the same computational function or trained algorithm is implemented to estimate the application force exerted by each brake, i.e., the right front brake, the left front brake, the right rear brake, and the left rear brake. In another example, a trained function or algorithm is established for each brake or for pairs of brakes within a vehicle type, thereby providing a more accurate application force estimate.

[0129] The present invention can be implemented by means of an artificial neural network and more generally by means of an artificial intelligence algorithm, but it can also be implemented by means of a classical optimization approach, in which the sum of the losses represents the objective function and each term of the losses a constraint, which will be satisfied with a certain tolerance, and the constants and C(x) and Kc are treated as variables which will be determined when solving the optimization problem.

[0130] The system according to the invention makes it possible to precisely estimate the clamping force, and more generally any characteristic useful for controlling the brakes, in addition it makes it possible to simplify the control unit and free up space for the operation of the automatic learning means.

[0131] The manufacturing method according to the invention also makes it possible to reduce the quantity of data necessary for training the artificial intelligence algorithm and to obtain a braking system in which the useful characteristic is estimated more precisely.

[0132] Furthermore, obtaining the data for training is generally collected by driving a vehicle equipped with the braking system to which it is desired to apply the virtual sensor, reducing the amount of data required makes it possible to reduce the number of kilometers to be covered by the vehicle and therefore the duration of this stage.

[0133] Furthermore, the invention, by integrating one or more physical models, provides the artificial intelligence software with information on expected / realistic output values. The physical model(s) limit the space of solutions admissible by the neural network during the learning phase.

[0134] The method according to the invention can take into account several mathematical models and in which the estimated variables are different. In this case, either a single algorithm is implemented to estimate the useful characteristic and the variables, or several algorithms are implemented.

[0135] Furthermore, it will be understood that the rigidity

[0136] Furthermore, the method according to the invention may relate to the estimation of several useful characteristics.

[0137] The invention applies to both disc brakes and drum brakes. The invention applies to service brakes and the parking brake. REFERENCES

[0138] FRI, FR2, FR3, FR4: brakes BRI, BR2: backpropagation loop Cl, C2..., Cn: selected parameters Cl', C2'..., Cn': selected parameters DL1, DL2, DL3: neural networks ECU: Electronic Control Unit F: measured clamping force Fo: estimated clamping force FP: clamping force calculated by the mathematical model Fest: function for calculating the estimated useful characteristic I: executable machine instructions Kc(x): the stiffness of the brake Ldata: difference between F and Fo LP: difference in F and FP M: memory area P: processor S: braking system V: vehicle Z: storage area

Claims

1. Claims Method for manufacturing a braking system for a motor vehicle comprising at least one brake (FRI, FR2, FR3, FR4) and means for estimating at least one characteristic useful for controlling said at least one brake (FRI, FR2, FR3, FR4), said brake being intended to be mounted on a given motor vehicle, said braking system comprising a control unit comprising a storage area (Z) comprising executable machine instructions, a memory and a processor (P) for generating said estimated useful characteristic, said brake being mounted on a test vehicle or a test bench conforming to the given motor vehicle, said method comprising the steps: a) Selection of first parameters (Cl, C2..Cn) of the brake and its environment having an influence on the value of the useful characteristic, b) Selection of at least one physical model making it possible to calculate said useful characteristic, said model comprising at least one variable depending on second parameters of the brake and its environment, c) Establishing a first artificial intelligence algorithm (DL1) to be trained using said first parameters (Cl, C2..., Cn) as input parameters and configured to estimate the useful characteristic, d) Measurement or evaluation by means equipping the brake or said test vehicle of a training data set comprising values ​​of the first parameters and values ​​of the useful characteristic, said useful characteristic being measured by at least one sensor equipping said brake, e) Training of said first artificial intelligence algorithm (DL1) comprising:

1. providing the first input parameters of the first artificial intelligence algorithm (DL1), 2. estimation of the useful characteristic by the first artificial intelligence algorithm (DL1), 3. Establishment of a function comprising a term dependent on the measured useful characteristic and a term dependent on the estimated useful characteristic and a term dependent on the mathematical model, 4- modification of the first artificial intelligence algorithm (DL1) so as to make said function tend towards zero, 5- repetition of sub-steps 1 to 4 until said function is less than a predetermined threshold value and the first artificial intelligence algorithm (DL1) is considered satisfactory, f) development of a function for calculating the estimated useful characteristic from the first artificial intelligence algorithm (DL1) considered satisfactory.

2. Manufacturing method according to claim 1, wherein the function of sub-step 3 is Ldata+ Lp >with [Math. 15] [Math. 16] [Math. 16] F being the measured useful characteristic, Fo being the estimated useful characteristic, Fp being the useful characteristic of the mathematical model.

3. Manufacturing method according to claim 1 or 2, in which the at least one variable is estimated by means of a second artificial intelligence algorithm (DL2), using as input the second parameters (CF, C2'..., Cn'), which are at least partly different from the first parameters, and in which said method comprises the measurement or evaluation by means equipping the brake or said vehicle with a set of training data comprising values ​​of second parameters and values ​​of the variable.

4. A manufacturing method according to claim 1 or 2, wherein the at least one variable is estimated by the first artificial intelligence algorithm (DL1).

5. Manufacturing method according to one of claims 1 to 4, in which the first algorithm (DL1) and / or the second algorithm (DL2) are implemented by artificial neural networks.

6. Manufacturing method according to the preceding claim, in which sub-step 4 implements backpropagation.

7. Manufacturing method according to one of the preceding claims, in which the useful characteristic is the clamping force exerted by the brake.

8. Manufacturing method according to the preceding claim, in which said mathematical model is of the CKcyb type, with Kc the stiffness of the brake and yble displacement of a part of the brake generating the braking, Kc being the variable

9. Manufacturing method according to claim 7, in which the brake is an electric brake and the mathematical model is of the type [Math. 17] With ia the electric current supplying the motor, com the angular speed of the motor, (bm the angular acceleration of the motor, and in which C4 is the variable.

10. Manufacturing method according to one of claims 1 to 9 wherein, during step f), a polynomial function or a weighting matrix is ​​extracted from the first trained artificial intelligence algorithm considered satisfactory intended to be used by the processor of the estimation system to calculate the estimated clamping force.

11. Manufacturing method according to one of claims 1 to 9 wherein, during step f), the calculation function is the first trained artificial intelligence algorithm considered satisfactory.

12. Manufacturing method according to one of claims 1 to 11, in which the vehicle moves during step d), preferably until reaching a wear determined as sufficient of at least one brake lining.

13. Computer program medium comprising executable machine instructions for the execution by a processor of a method for estimating a characteristic useful for controlling at least one brake of a braking system, said estimation method implementing the first trained artificial intelligence algorithm considered satisfactory obtained by the method of manufacturing according to claim 11 or the polynomial function or 1 weighting matrix extracted from the first trained artificial intelligence algorithm considered satisfactory obtained by the manufacturing method according to claim 10.

14. Braking system for a motor vehicle comprising at least one brake (FRI, FR2, FR3, FR4), an electronic control unit (ECU) for said brake comprising a storage area (Z) comprising executable machine instructions, a memory and a processor (P) for generating at least one useful characteristic estimated for the control of the brake by the control unit, said braking system being obtained by the method according to one of claims 1 to 12, said braking system also comprising means for collecting the first selected parameters, said collection means comprising sensors and / or means for evaluating all or part of the first selected parameters.

15. Braking system according to claim 14, in which the collecting means do not comprise or do not use a sensor for direct measurement of said useful characteristic.

16. A braking system according to claim 14 or 15 in combination with claim 10, said polynomial function or the weighting matrix being stored in the storage area, or the first trained artificial intelligence algorithm considered to be satisfactory is stored in the storage area, in particular in which the processor is a graphics processor.

Images