Vehicle parts development process

Abstract

A method for developing vehicle parts, said method comprising the following steps: defining (EA) at least one technical characteristic of a state parameter cycle, generating (EB), using a driven generator, and from said at least one technical characteristic, at least one synthetic state parameter cycle representative of an actual state parameter cycle, said at least one state parameter cycle representing the evolution over time of a vehicle state parameter, and developing (EC) the vehicle parts from said at least one synthetic state parameter cycle. Abstract figure: Figure 2.

Description

Title of the invention: Method for developing vehicle parts

[0001] This document relates to a method of developing vehicle parts from a generated synthetic state parameter cycle. State of the art

[0002] In a motor vehicle, it is known to verify a control law concerning, for example, the charging, discharging, and temperature management of a battery.

[0003] It is also known to optimize the aging of the vehicle (which we therefore seek to minimize).

[0004] It is the energy consumption on a given or planned journey. Such optimization can be carried out on fuel, on electrical energy, on hydrogen consumption (in the case of a vehicle equipped with a fuel cell) or on two or three of these criteria at the same time.

[0005] It may also prove useful to have to dimension certain elements of the vehicle in order to carry out these optimizations or verifications, such elements being able for example to consist of a fuel or hydrogen tank, an electric battery and / or super-capacitors capable of supplying electrical energy, a thermal engine powered by the fuel tank or a fuel cell powered by the hydrogen tank, or even an electric machine powered by electrical energy supplied by the battery and / or the super-capacitors.

[0006] In the case of vehicles equipped with fuel cells, the parameters relating to the elements of the vehicle are, for example, the number of cells in the electric battery, the size of the electric battery, the number of stacking layers in the fuel cell, the number of cells for each stacking layer in the fuel cell, the threshold dead zone of the fuel cell, etc. The associated optimal control then consists, for example, in determining, on a given or planned journey, the distribution of electrical power between the fuel cell and the electric battery at any time.

[0007] Furthermore, it may be useful to design certain elements of the vehicle in a certain way in order to carry out these optimizations or verifications so that the elements are calibrated or valid according to certain systems with simulated scenarios.

[0008] It is, for example, known from the state of the art to size an engine by considering a single state parameter cycle approved according to the WLTC standard (“Worldwide Harmonized Light Vehicles Test Cycle” in English).

[0009] A state parameter cycle is a predefined sequence that represents the evolution over time of a state parameter of a motor vehicle.

[0010] A state parameter cycle may for example be a driving cycle that simulates real driving conditions, this is generally used to evaluate the performance, fuel consumption and emissions of vehicles. It should be noted that different driving cycles can be defined for different driving conditions, for example, urban, extra-urban, highway, etc.

[0011] A state parameter cycle is for example carried out on a laboratory test bench to ensure reproducible conditions. This cycle may comprise several phases, which may be repeated several times during testing.

[0012] Vehicle testing procedures are typically carried out using the WLTC standard. This is an international standard for testing emissions and fuel consumption for light vehicles such as cars, vans and light commercial vehicles. The WLTC aims to provide accurate and representative real-life driving test data to assess the environmental and energy performance of vehicles.

[0013] Indeed, the WLTC aims to be the best compromise in relation to the need to be representative of the real use of the vehicle over its entire lifespan and the necessary computing power.

[0014] However, the WLTC test data are not representative of certain state parameter cycle scenarios: they are represented by fixed state parameter cycles.

[0015] There is therefore a need to improve the representativeness of the actual use of the vehicle during its life cycle.

[0016] This document proposes to exploit the computing power of generative artificial intelligence models to improve the representativeness of state parameter cycles, and therefore to improve vehicle optimization in general or to test control laws. Statement of the invention

[0017] For this purpose, the present document relates to a method for developing parts of a vehicle, said method comprising the following steps: a. define at least one technical characteristic of a state parameter cycle, b. generate, using a driven generator, and from said at least one technical characteristic, at least one synthetic state parameter cycle representative of a real state parameter cycle, said at least one synthetic state parameter cycle representing the evolution over time of a state parameter of a vehicle, and c. developing the vehicle parts from said at least one synthetic state parameter cycle.

[0018] The technical characteristic of a state parameter cycle may be a characteristic representative of one or more of the following: - the type of road: the nature of the road surface (asphalt, gravel, cobblestone) and its condition (damaged, smooth) can influence the vehicle's performance, - topography: the presence of ascending, descending slopes or flat terrain can affect fuel consumption and emissions, - traffic density: dense or fluid traffic can influence the acceleration, deceleration and stopping phases of the vehicle, - the vehicle load: the presence of passengers, luggage or a trailer can affect the vehicle's performance, - the condition of the vehicle: the state of maintenance of the vehicle (new, old, regularly maintained) and the type of fuel used can influence the results, - active auxiliary systems: the use of air conditioning, heating, headlights or other systems can have an impact on energy consumption, - specific maneuvers: actions such as overtaking, lane changes or parking can be included to make the cycle more realistic, - altitude: driving at different altitudes can affect air density and, therefore, combustion in the engine, - maximum and average speed: the speed at which the vehicle is driven for most of the cycle and the maximum speed reached can be determining factors, - total cycle duration: cycle duration may vary, with some cycles being longer to simulate longer journeys, - weather conditions: rain, snow, wind and other weather conditions can be simulated, - the type of driving: aggressive with frequent accelerations and decelerations, or gentle with smoother driving.

[0019] The technical characteristic of a state parameter cycle may also reflect the distribution of driving zones within a state parameter cycle. Such a distribution may be defined in terms of a percentage of the total cycle time or the total distance traveled. The driving zones describe different driving environments and scenarios that a vehicle may encounter.

[0020] The driving zones may be one or more of the following zones: - an urban area: it simulates driving in an urban environment, and is characterized by frequent stops due to traffic lights, intersections, pedestrian crossings and traffic density. Such an area may be characterized by limited speed (e.g., up to 50 km / h), frequent accelerations and decelerations, and short distances between stops, - an extra-urban area: this area simulates driving outside dense urban areas, but not on motorways. This may include national or departmental roads. Such an area may be characterized by moderate to high speeds (e.g., 50 to 90 km / h), fewer stops than in urban areas, and with the possibility of turns and intersections, - a motorway zone: this zone simulates driving on motorways or expressways, designed for long distances at high speeds. Such an area may be characterized by high speed (e.g., 90 to 130 km / h or more), access and exit via interchanges, few stops or obstructions, multiple lanes, - a mountain area: this area simulates driving in mountainous or hilly areas. Such an area may be characterized by the presence of ascending and descending slopes, tight turns, potentially changing conditions due to altitude (such as air density), variable speed depending on the slope and road conditions, - a rural area: this area simulates driving in less populated areas, for example, single-lane roads. Such an area may be characterized by moderate speed, less traffic, the possibility of varied terrain (such as dirt roads), - a construction zone: this zone simulates the presence of roadworks where traffic may be restricted. Such a zone may be characterized by reduced speed, lane changes, the presence of construction equipment and personnel, and temporary signs.

[0021] Each of these zones presents unique challenges and characteristics for driving. State parameter cycles can take these different zones into account to evaluate a vehicle's performance under varying conditions.

[0022] The state parameter can be one of the following parameters: - the vehicle's travel speed, - vehicle acceleration, - the position of the pedals, i.e. the position of the vehicle's accelerator, brake and clutch pedals, - engine speed: the number of revolutions per minute (rpm) of the engine.

[0023]

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] - fuel consumption over a given period. - the vehicle engine temperature represented by the engine coolant temperature, - the oil pressure in the vehicle's engine, - engine load: the percentage of the total engine capacity used at a given time, - gas and particle emissions emitted by the vehicle, - the position of the steering wheel, i.e. the angle and / or direction of the steering wheel. - brake pressure, - the state of the vehicle battery, i.e. the measurement of the remaining battery charge, - the temperature of the intake air entering the engine, - the condition of auxiliary systems such as air conditioning, power steering, and other systems that may affect vehicle performance, and - the position of the gearbox at a given moment. The state parameter can also be one of the following hybrid, electric, or hydrogen vehicle state parameters: - the power requested or received by the battery, - the power required from the electric machine, and - the power requested or received by the fuel cell (in the case of hydrogen vehicles). The development of the vehicle parts of step (c) of the method may include sizing the vehicle parts. By dimensioning we mean determining the appropriate dimensions of a structure, component or material so that they are suitable / optimized for their intended use or choosing electrical components so that they have the appropriate values ​​for the operation of an electrical circuit. The development of the vehicle parts of step (c) of the method may include a design of the vehicle parts. The development of the vehicle parts of step (c) of the method may include a verification of the control laws of the vehicle parts. A vehicle's propulsion system refers to the various systems and technologies used to generate the power needed to move the vehicle. Propulsion systems can include an internal combustion engine, an electric motor, or a fuel cell. Dimensioning can also include power sources such as a battery in an electric motor, or additional functions. A synthetic state parameter cycle is characterized as a time series. In a time series, measurements or observations are collected or recorded at a regular or irregular time interval, allowing the analysis of the evolution of a given parameter over time. In the context of a synthetic state parameter cycle, state parameters can be recorded at regular intervals, creating a time series that reflects the vehicle dynamics over a given period.

[0030] This method saves time and resources thanks to the computing power of the trained generator.

[0031] Furthermore, it also makes it possible to improve the overall precision of the dimensioning of the propulsion means with respect to the consumption of the vehicle thanks to a synthetic state parameter cycle.

[0032] In particular, carrying out step (b) from said at least one technical characteristic makes it possible to specify the synthetic state parameter cycle that one wishes to generate, the technical characteristic possibly being different from the state parameter cycles usually approved.

[0033] In this way, the synthetic state parameter cycle improves the representativeness for the dimensioning of the vehicle's propulsion means.

[0034] Of course, this same method can be used in other technical fields.

[0035] Step (b) can comprise the sub-steps consisting of: (bl) generating, using a driven generator, and from said at least one technical characteristic, a plurality of intermediate state parameter cycles, (b2) concatenating the plurality of intermediate state parameter cycles with an interpolation called spline interpolation, and (b3) obtaining said at least one synthetic state parameter cycle which is the result of the concatenation.

[0036] Spline interpolation is an interpolation in which smooth curves (splines) are fitted to known data points, allowing more flexible estimation of missing values.

[0037] The intermediate state parameter cycles of the plurality of intermediate state parameter cycles may be 15 min in duration.

[0038] The intermediate state parameter cycles of the plurality of intermediate state parameter cycles may be between 15 and 120 min in duration.

[0039] These sub-steps make it possible to overcome the limitation of the driven generator: it may not be suitable for generating long state parameter cycles, hence the need to generate a plurality of intermediate state parameter cycles. By long state parameter cycles, we mean state parameter cycles lasting 14 hours or 1440 min.

[0040] These sub-steps also make it possible to split into a plurality of cycles of intermediate state parameter a state parameter cycle with technical characteristics that vary over time. For example, if a state parameter cycle has a first half cycle in a mountain area and a second half cycle in a motorway area, it is possible to generate a first and second intermediate cycle corresponding to the first and second half of the cycle respectively.

[0041] Furthermore, the use of a plurality of short-duration intermediate state parameter cycles provides flexibility, since the synthetic state parameter cycle is the result of the concatenation.

[0042] Step (c) may consist of developing parts of the vehicle from said at least one synthetic state parameter cycle and / or at least one real state parameter cycle.

[0043] The trained generator may comprise a generator portion of a conditional trained generative adversarial network.

[0044] A generative adversarial network, or GAN, is a class of machine learning models designed to generate fictitious data that resembles real data.

[0045] A GAN is composed of two distinct parts: a generator part (also called generator part) and a discriminator part (also called discriminator part), which are trained simultaneously and in competition.

[0046] The generator part attempts to produce dummy data from random noise. The discriminator part attempts to distinguish real data from dummy data generated by the generator part.

[0047] When training a GAN, the generator part creates a dummy data from random noise. Then, the discriminator part evaluates this generated data and gives a score on its resemblance to real data. Both parts are then updated according to their performance: the generator part tries to fool the discriminator part by producing better quality data, while the discriminator part tries to improve itself to better distinguish real data from dummy data.

[0048] This process is repeated many times until the generator produces data of good enough quality that the discriminator has difficulty distinguishing it from real data.

[0049] Once the GAN is properly trained, the generator part can be used alone to generate new dummy data. For example, if the GAN has been trained on real state parameter cycles, the generator part can produce new (synthetic) state parameter cycles that, although artificially generated, will look like real cycles.

[0050] A Conditional Generative Adversarial Network, or cGAN, is an extension of the standard GAN model. The particularity of cGAN is that it generates data based on a certain condition (conditional information).

[0051] Unlike the standard GAN where the generator receives only random noise as input, in a cGAN the generator receives both the random noise and conditional information (e.g. data in the form of a tensor or a value, representing the at least one technical characteristic). The generator part uses this conditional information to produce dummy data that corresponds to the conditional information. The discriminator part, just like in the standard GAN, tries to distinguish between the real data and the generated dummy data. However, in the cGAN the discriminator part also receives the conditional information to make its decision.

[0052] Once the cGAN is correctly trained, the generator part can generate new fictitious data while continuing to take into account the conditional information.

[0053] This provides more precise control over the type of data generated, thus improving the diversity and control of generation.

[0054] In this document, the conditional information corresponds to said at least one technical characteristic defined in step (a).

[0055] It may be envisaged to distinguish two cases:

[0056] In a case called discrete case, the conditional information is one or more control vectors, each element of the vector corresponding to a class.

[0057] For example, one of the control vectors may be a vector called a speed vector of size equal to a number of speed intervals of the vehicle. For 5 speed intervals (and therefore a vector of size 5): - the element of index 0 of the speed vector corresponds to class 0 and to a speed V in km / h such that 0 < V < 20, - the element of index 1 of the speed vector corresponds to class 1 and to a speed V in km / h included in the interval 20 < V < 40, - the element of index 2 of the speed vector corresponds to class 2 and to a speed V in km / h included in the interval 40 < V < 60, - the element of index 3 of the speed vector corresponds to class 3 and to a speed V in km / h included in the interval 60 < V < 80, and - the element of index 4 of the speed vector corresponds to class 4 and to a speed V in km / h included in the interval 80 < V.

[0058] Again for example, one of the control vectors can be a vector called acceleration vector which is of size 2: the index 0 of the acceleration vector corresponds to a acceleration phase of the vehicle and index 1 of the acceleration vector corresponds to a deceleration phase of the vehicle.

[0059] Of course, it can be planned to give as input to the cGAN one or more control vectors with a desired number of classes to control the output.

[0060] In a case called continuous case, the conditional information is a scalar. For example, for a scalar S between 0 and 1, S can correspond to a normalized value of the vehicle speed.

[0061] Of course, it may be provided to provide a plurality of scalars as conditional information as input to the cGAN to control several characteristics.

[0062] Ding et al. (2021) “CcGAN: Continuous Conditional Generative Adversarial Networks for Image Generation” describes examples of cGANs.

[0063] The driven generator may comprise a decoder portion of a conditional variational autoencoder.

[0064] An autoencoder (or autoencoder or AE) is a type of artificial neural network used in the field of machine learning and pattern recognition.

[0065] The basic principle of an autoencoder is to compress the input data into a reduced-dimensional internal representation (or code), and then decompress this representation to reconstruct the original data. The objective is to minimize the reconstruction error between the input data and the output data. This forces the network to learn a compact and meaningful representation of the data.

[0066] An autoencoder is composed of two main parts: - the encoder (also called the encoder-forming part): this part takes the input data and transforms it into a reduced-dimensional internal representation. The encoder generally consists of layers of neurons that progressively reduce the dimension of the data. - the decoder (also called the decoder-forming part): the decoder takes the internal representation (the code) and decompresses it to reconstruct the original data. It has layers of neurons that gradually increase the dimension of the data until it matches the input data.

[0067] The autoencoder is trained using input and output data pairs, so that the network learns to correctly reconstruct the data.

[0068] Once trained, the encoder can be used for various tasks, such as dimension reduction, anomaly detection, similar data generation, and pre-training neural networks for other deep learning tasks.

[0069] The document “Autoencoders” (Bank et al.) describes examples of AE.

[0070] A variational autoencoder (VAE) is a probabilistic variant of the standard autoencoder.

[0071] Just like the AE, the VAE includes in particular an encoder part and a decoder part.

[0072] The encoder part takes an input data and transforms it into a latent representation, generally of reduced dimension, in the latent space. Unlike a standard autoencoder, the encoder part of a VAE produces two vectors for each data: a vector of means and a vector of standard deviations. These vectors define a probabilistic distribution in the latent space.

[0073] Instead of obtaining a fixed latent representation for each data, an entry is associated with a probability distribution in this space. A sample is then taken from this distribution for the decoding phase.

[0074] The decoder part takes a sample of the latent space and reconstructs it to produce data that should resemble the original input.

[0075] During training, the VAE learns to encode the input data into probabilistic distributions in the latent space and to decode these distributions to reconstruct the original data. The objective is to minimize the difference between the original data and the reconstructions, while ensuring that the distributions in the latent space follow a standard normal distribution. This is achieved through a cost function composed of a reconstruction term and a regularization term.

[0076] Once trained, the decoder portion of the VAE is used to generate data by sampling the latent space and passing these samples through the decoder portion, resulting in new dummy data in the original data space.

[0077] A conditional VAE (Conditional Variational Autoencode or cVAE) is an extension of the standard VAE model that allows data to be generated based on certain specific conditions.

[0078] Like the VAE, the cVAE learns to encode and decode data, but it does so by taking into account conditional information. During training, the conditional information is provided to both the encoder and the decoder. This information guides the generation process. Once the cVAE is trained, it can be used to generate data that matches a specific condition.

[0079] This allows for the generation of dummy data that not only resembles actual training data, but also matches specific conditions or criteria, thus providing more precise control over the type of data generated.

[0080] The paper “Learning Structured Output Representation using Deep Conditional Generative Models” (Sohn et al.) describes conditional VAEs.

[0081] The trained generator may comprise a probabilistic conditional denoising diffusion model.

[0082] A Denoising Diffusion Probabilistic Model (DDPM) is an approach to data generation that uses a diffusion process to progressively convert a random noisy image into a target image.

[0083] The central concept behind DDPM is to reverse the process by which data is corrupted by noise. Starting with noisy data, the model attempts to eliminate this noise progressively through a series of steps, relying on probabilistic techniques to guide this transformation. Each step in the process is designed to reduce the noise incrementally, in order to get ever closer to the target data.

[0084] A conditional probabilistic diffusion denoising model (or cDDPM) is an extension of DDPM that incorporates conditional information to guide the generation process.

[0085] In the context of denoising diffusion, the conditional approach works by introducing a condition during the transformation process. This condition serves to direct the diffusion process towards a specific result.

[0086] Zhang et al. (2023) “ShiftDDPMs: Exploring Conditional Diffusion Models by Shifting Diffusion Trajectories” describes examples of cDDPMs.

[0087] The driven generator can be driven during a training phase comprising the following steps: (i) obtaining a plurality of actual state parameter cycles representing the evolution over time of a state parameter of the vehicle, said state parameter being measured via at least one sensor fitted to the vehicle, (iii) driving said generator from the plurality of actual state parameter cycles.

[0088] The number of actual state parameter cycles can be between 3000 and 7000.

[0089] The number of actual state parameter cycles can be equal to 5000.

[0090] The training phase may further comprise a step (ii) carried out between the steps (i) and (iii), said step (ii) comprising preprocessing the plurality of actual state parameter cycles.

[0091] After preprocessing, the actual state parameter cycles of the plurality of actual state parameter cycles may be between 3000 and 4200 s in duration.

[0092] After preprocessing, the actual state parameter cycles of the plurality of actual state parameter cycles may be of a duration equal to 3600 s.

[0093] Preprocessing step (ii) may comprise, for at least a portion of the plurality of actual state parameter cycles, identifying at least one time interval data interruption in a cycle, called an incomplete cycle, and reconstruct the given time interval.

[0094] The data outage time interval may be a period during which data is unavailable or is lost. This may occur for various reasons such as connection problems, system errors, or deficiencies in data collection.

[0095] It may be provided to reconstruct the given time interval if the duration of said time interval is between a first threshold value and a second threshold value.

[0096] If the duration of said time interval is less than the first threshold value, reconstruct the time interval by interpolation.

[0097] The first threshold value can be between 5 and 15 s.

[0098] The first threshold value may be equal to 10 s.

[0099] Data reconstruction by interpolation is a technique used to estimate missing values ​​in a data set. This technique is particularly useful when there are interruptions or gaps in the collected data.

[0100] Interpolation involves using known values ​​before and after the interruption to estimate missing values. There are several interpolation methods: - linear interpolation: a straight line is drawn between two known data points and missing values ​​are estimated based on this line, - polynomial interpolation: a polynomial curve is fitted to the known data points, and missing values ​​are estimated from this curve. - spline interpolation, - nearest neighbor interpolation: missing values ​​are estimated based on the values ​​of the nearest data points, - statistical or stochastic interpolation: it exploits the statistical distribution of the data.

[0101] If the data interruption time interval is greater than a second threshold value, provision may be made to remove the corresponding incomplete cycle from the plurality of actual state parameter cycles used to drive said generator.

[0102] The second threshold value can be between 500 and 700 s.

[0103] The second threshold value can be equal to 600s.

[0104] The reconstruction of the given time interval can be carried out using a trained reconstructor using the incomplete cycle as input data.

[0105] The reconstructor can be trained with a convolutional autoencoder, (also called convolutional autoencoder or CAE for “Convolutional Autoencoder” in English).

[0106] In this particular case, the encoder is the first part of the network and is composed of convolution layers, pooling layers and normalization layers. It takes the input signal and transforms it into a lower-dimensional representation called the latent code. This step allows important features of the signal to be extracted while reducing its dimensionality.

[0107] The latent code is the compressed representation of the input signal. It contains essential information for reconstructing the original signal.

[0108] The decoder is the second part of the network and is also composed of convolution layers, deconvolution (or transpose) layers and normalization layers. It takes the latent code as input and transforms it to reconstruct the original signal as faithfully as possible.

[0109] The reconstructor training data may be training state parameter cycles comprising data interruption time intervals less than the first threshold value to which additional data interruption time intervals are added.

[0110] Reconstructor training is self-supervised (or supervised) training, in which the reconstructor learns to reconstruct incomplete cycles by minimizing the difference with the training state parameter cycles.

[0111] Supervised training includes the following steps: - A dataset that contains examples of inputs (features) and corresponding outputs (labels or targets) is gathered. This data is collected or annotated by hand in the case of supervised training. - A learning algorithm (such as logistic regression, decision trees, neural networks, etc.) is chosen, which will be responsible for creating the model. The algorithm uses the training dataset to learn relationships between features (inputs) and labels (outputs). - The learning algorithm adjusts the model parameters to minimize the error between the model predictions and the actual labels in the training dataset, for example through optimization techniques, such as gradient descent. - Once the model has been trained, it is evaluated on a separate dataset called the validation set to check its performance and generalization ability. Then, it is tested on a test dataset to evaluate its performance on unknown data.

[0112] The trained reconstructor may comprise a 1D Convolution layer followed by an Average Pooling layer.

[0113] The trained reconstructor may comprise a decoder portion of an autoencoder.

[0114] The present document also relates to a computer program comprising instructions for implementing the method according to the aforementioned type, when this program is executed by a processor.

[0115] This document may also relate to a non-transitory computer-readable recording medium on which a program is recorded for implementing the method according to the aforementioned type, when this program is executed by a processor.

[0116] The present document also relates to a computer system comprising: - an input interface for receiving a plurality of actual state parameter cycles, - a memory for storing at least the instructions of a computer program as presented previously, - a processor accessing the memory to read said instructions and then execute the method according to the aforementioned type, - an output interface for providing said at least one generated synthetic state parameter cycle.

[0117] The present document also relates to a method for training a vehicle state parameter cycle generator, said method comprising the following steps: - obtaining a plurality of actual state parameter cycles representing the evolution over time of a state parameter of the vehicle, said state parameter being measured via at least one sensor fitted to the vehicle; - preprocessing the plurality of actual state parameter cycles to obtain a plurality of preprocessed actual state parameter cycles; - training at least one generator with the plurality of preprocessed real state parameter cycles, capable of generating a synthetic state parameter cycle representative of a real state parameter cycle. Brief description of the drawings

[0118] Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the attached drawing showing various figures, in which: - [Fig.l] schematically represents an example of a computer system according to this document, - [Fig.2] illustrates the different stages of the process according to one embodiment of this document, - [Fig.3] illustrates the different stages of the generator training phase according to an embodiment of this document, and - [Fig.4] is a graph of vehicle speed versus time for a real state parameter cycle. Detailed description

[0119] [Fig.l] schematically represents a computer system 1 comprising

[0120] - an input interface 2 for receiving data from at least one sensor of a technical system and representative of a technical system,

[0121] - a memory 3 for storing at least the instructions of a program computer,

[0122] - a processor 4 accessing memory 3 to read said instructions and execute then the process illustrated in [Fig.2],

[0123] - an output interface 5.

[0124] Reference is now made to [Fig.2] which illustrates the different stages of the method according to an embodiment of the present document.

[0125] This method aims to develop parts of a vehicle from at least one synthetic state parameter cycle.

[0126] This method comprises a step EA during which at least one technical characteristic of a state parameter cycle is defined. In the illustrated example, the defined technical characteristics are the distribution of the driving zones in the state parameter cycle.

[0127] Then, during a step EB, at least one synthetic state parameter cycle representative of a real state parameter cycle is generated, using a trained generator, and from said at least one technical characteristic.

[0128] Finally, during an EC step, parts of the vehicle are developed from said at least one generated synthetic state parameter cycle.

[0129] [Fig.3] illustrates the training phase of the generator according to an embodiment of the present document.

[0130] During a step Ei, a plurality of real state parameter cycles representing the evolution over time of a state parameter of the vehicle is obtained, said state parameter being measured by means of at least one sensor equipping the vehicle.

[0131] Then, the plurality of actual state parameter cycles are preprocessed and the preprocessing is distinguished according to three cases for each cycle among the plurality of cycles. of actual state parameters: - if we identify a data interruption time interval in a cycle, called an incomplete cycle, which is less than a first threshold value, here 10s, (result Interruption < 10s) then we reconstruct the given time interval by interpolation (step Eii), - if we identify a data interruption time interval in the incomplete cycle between the first threshold value and the second threshold value, here 600s, (result 10s < Interruption < 600s) then we reconstruct the given time interval using a trained reconstructor (step Eii'). - if a data interruption time interval in the incomplete cycle is identified that is greater than the second threshold value (Interruption result > 600s), then the corresponding incomplete cycle is removed from the plurality of actual state parameter cycles (step Eii”).

[0132] [Fig.4] is a graph of vehicle speed versus time.

[0133] A curve Cl represents the vehicle speed as a function of time during a state parameter cycle. It is observed that the curve Cl includes a data interruption time interval Interrupt equal to 300s and therefore between 10s and 600s.

[0134] Curve C2 represents the Cl curve reconstructed in Intenuption using the trained reconstructor.

[0135] Reference is again made to [Fig.3]. In a step Eiii, the generator is trained from the plurality of actual state parameter cycles.

Claims

Claims

1. Method for developing parts of a vehicle, implemented by a processor, said method comprising the following steps: a. defining (EA) at least one technical characteristic of a state parameter cycle, b. generating (EB), using a trained generator, and from said at least one technical characteristic, at least one synthetic state parameter cycle representative of a real state parameter cycle, said at least one synthetic state parameter cycle representing the evolution over time of a state parameter of a vehicle, and c. developing (EC) the parts of the vehicle from said at least one synthetic state parameter cycle.

2. A method according to the preceding claim, wherein the trained generator comprises a generator portion of a conditional trained generative adversarial network.

3. The method of claim 1, wherein the trained generator comprises a decoder portion of a conditional variational autoencoder.

4. The method of claim 1, wherein the trained generator comprises a probabilistic conditional denoising diffusion model.

5. Method according to one of the preceding claims, in which said driven generator is driven during a training phase comprising the following steps: (i) obtaining (Ei) a plurality of real state parameter cycles representing the evolution over time of a state parameter of the vehicle, said state parameter being measured via at least one sensor fitted to the vehicle, (iii) driving (Eiii) said generator from the plurality of real state parameter cycles.

6. Method according to the preceding claim, wherein the training phase further comprises a step (ii) performed between steps (i) and (iii), said step (ii) consisting of preprocessing (Eii, Eii', Eii”) the plurality of actual state parameter cycles.

7. Method according to the preceding claim, in which the preprocessing step (ii) consists of, for at least a part of the plurality of actual state parameter cycles, identifying at least one data interruption time interval (Iinterruption) in a cycle, called incomplete cycle, and reconstructing (Eii, Eii') the given time interval.

8. Method according to the preceding claim, wherein if said data interrupt time interval (Iinterrupt) is greater than a second threshold value, removing (Eii”) the corresponding incomplete cycle from the plurality of actual state parameter cycles used to drive said generator.

9. Method according to the preceding claim, in which the reconstruction of the given time interval is carried out using a reconstructor trained using the incomplete cycle as input data.

10. A method according to the preceding claim, wherein the trained reconstructor comprises a decoder portion of an autoencoder.

11. A method according to any preceding claim, wherein the state parameter is one of the vehicle travel speed, the vehicle acceleration, the pedal position, the engine speed, the fuel consumption over a given period, the vehicle engine temperature, the oil pressure in the vehicle engine, the engine load, the gas and particulate emissions emitted by the vehicle, the vehicle steering wheel position, the brake pressure, the state of the vehicle battery, the temperature of the intake air entering the engine, the state of the auxiliary systems, the position of the gearbox at a given time, the power requested or received by the battery, the power requested from the electric machine and the power requested or received by the vehicle fuel cell.

12. A method according to any preceding claim, wherein said at least one technical characteristic is at least one of road type, topography, traffic density, vehicle load, vehicle condition, active auxiliary systems, specific maneuvers, altitude, maximum and average speed, total cycle time, weather conditions, driving type, and driving zone distribution.

13. A method according to any preceding claim, wherein the development of the vehicle parts of step (c) of the method comprises dimensioning the vehicle parts.

14. Computer program comprising instructions for implementing implementation of the method according to one of the preceding claims, when this program is executed by a processor.

15. Computer system (1) comprising: - an input interface (2) for receiving a plurality of actual state parameter cycles, - a memory (3) for storing at least the instructions of a computer program according to the preceding claim, - a processor (4) accessing the memory (3) to read said instructions and then execute the method according to one of claims 1 to 13, - an output interface (5) for providing said at least one generated synthetic state parameter cycle.

16. A method of training a vehicle state parameter cycle generator, said method comprising the following steps: - obtaining a plurality of actual state parameter cycles representing the evolution over time of a state parameter of the vehicle, said state parameter being measured via at least one sensor fitted to the vehicle; - preprocessing the plurality of actual state parameter cycles to obtain a plurality of preprocessed actual state parameter cycles; - training at least one generator with the plurality of preprocessed real state parameter cycles, capable of generating a synthetic state parameter cycle representative of a real state parameter cycle.

Images