Method for manufacturing a polyphase electric current rotating electromagnetic machine intended to be fitted to an electric or hybrid motor vehicle

Abstract

The invention relates to a method for manufacturing a polyphase electric current rotating electromagnetic machine intended to be fitted to an electric or hybrid motor vehicle, which method is implemented by an assembly (2) comprising at least one device (4) for manufacturing a rotating electromagnetic machine, and a computer system (6) comprising computing means (8) and memory means (10), the memory means (10) storing an optimiser (12) and a neural network (14) configured at least as a physics-informed neural network, the neural network (14) being configured to predict, from a set of instantiated values of physical parameters relating to the rotating electromagnetic machine and two vector control currents, a two-dimensional model relating to two flux obtained from the vector control currents. The invention further relates to an assembly (2) for carrying out such a method.

Description

[0001] METHOD FOR MANUFACTURED A ROTATING ELECTROMAGNETIC MACHINE WITH MULTIPHASE ELECTRIC CURRENT INTENDED TO EQUIP A

[0002] ELECTRIC OR HYBRID MOTOR VEHICLE

[0003] DESCRIPTION

[0004]

[0001] The present invention relates to the manufacture of a rotating electromagnetic machine with polyphase electric current intended for use in an electric or hybrid motor vehicle. The rotating electromagnetic machine is typically an axial magnetic flux machine, although this is not a limitation within the scope of the present invention.

[0005]

[0002] An electric or hybrid motor vehicle is typically equipped with a rotating electromagnetic machine of the permanent magnet electric motor type. This type of motor is a specific type of axial magnetic flux machine, which offers several advantages over radial magnetic flux machines, although they are currently less common in the automotive industry. The manufacturing of such rotating electromagnetic machines requires rapid and efficient design optimization.

[0006]

[0003] As with many design optimization problems, it all begins with an assumption and a reduction order model. A parametric model representing the rotating electromagnetic machine can then be obtained. Next, a simulation tool is used to estimate key performance indicators (such as manufacturing cost, weight, power, torque, energy density, etc.) that help to fairly compare several rotating electromagnetic machine design models. During this process, a two-dimensional model, classically called a "DQ" model, must be implemented to solve the control strategy and to obtain characterization maps representing the rotating electromagnetic machine, thus allowing all the necessary information to be deduced to estimate the design performance.Such a two-dimensional “DQ” model relates to two vector control currents, the first of which allows adjustment of the magnetic flux of the rotating electromagnetic machine, and the second of which allows adjustment of the torque of the machine.

[0007]

[0004] It is known that manufacturing processes for rotating electromagnetic machines involve computational steps consisting of iterating over several machine design models using a computer simulation system. However, the computation time generated by such a simulation tool is particularly high when used in an optimization process aimed at identifying the best possible design model. Indeed, in such an optimization process, which requires numerous iterations (typically thousands or more), the time resources needed to run a simulation can lead to an intractable optimization. This is particularly true for rotating electromagnetic machines with axial magnetic flux due to their electromagnetic flux paths, which require three-dimensional physical simulations.

[0008]

[0005] An artificial intelligence solution implementing a neural network can be used as a complement to simulation when a sufficiently large amount of data is available. However, this is infrequent due to the complexity and computational difficulty inherent in collecting the data for training the neural network. The idea is to use this data so that the neural network can learn to predict performance on a new design model, thereby accelerating the process. However, training such a neural network requires collecting certain data with the simulation tool. Therefore, such a solution does not completely resolve the time-consuming problem introduced by simulation tools; it requires a time-consuming step during the data collection phase.Furthermore, in such processes implementing artificial intelligence solutions, the more free design parameters there are, the more data needs to be collected, due to an empirical observation commonly known as the "curse of dimensionality." Consequently, it is often impossible to perform effective artificial intelligence substitutions due to a lack of available data.

[0006] There is therefore a need for a manufacturing process for a rotating electromagnetic machine with polyphase electric current intended for use in an electric or hybrid motor vehicle, some steps of which are implemented by a computer system including a neural network, that reduces the computation time of these steps and / or the amount of training data required for the neural network, while accelerating the design and therefore the manufacturing time of the machine.

[0009]

[0007] To this end, the invention relates to a method for manufacturing a rotating electromagnetic machine with polyphase electric current intended to equip an electric or hybrid motor vehicle, the method being implemented by an assembly comprising at least one device for manufacturing a rotating electromagnetic machine and a computer system connected to said at least one manufacturing device, the computer system comprising computing means and memory means connected to the computing means, said memory means storing an optimizer and a neural network configured at least as a physics-informed neural network, the optimizer being connected to the neural network and being configured to define distinct design models for the electromagnetic machine, each design model corresponding to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine,the neural network being configured to predict, from a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine and two vector control currents, a first current allowing adjustment of the magnetic flux of the rotating electromagnetic machine, and a second current allowing adjustment of the torque of the rotating electromagnetic machine, a two-dimensional model relating to two fluxes obtained from said first and second currents, the process comprising the following steps iteratively looped until a predetermined criterion is satisfied, said predetermined criterion including at least one metric:,

[0010] - selection, by the optimizer, of a current design model for the rotating electromagnetic machine; - prediction by the neural network, from the selected current design model, of a current two-dimensional model;

[0011] - calculation by the computing means of the computer system, from the predicted current two-dimensional model, of a vector control strategy including current instructions relating to said two flows;

[0012] - calculation, using the computer system's computing resources, based on the calculated vector control strategy, of at least one metric related to the current design model; and the method further comprising the following steps, when said predetermined criterion is met:

[0013] - selection, by the computer system, of the current design model as the design model to be manufactured for the rotating electromagnetic machine;

[0014] - manufacture, by said at least one manufacturing device, of the rotating electromagnetic machine on the basis of the design model to be manufactured.

[0015]

[0008] In the manufacturing process according to the invention, using a physics-informed neural network in the computer system reduces reliance on complex and computationally intensive simulation tools during the process optimization phase, thereby accelerating computation times and significantly reducing the time required to select an optimal design model for the rotating electromagnetic machine. This, in turn, accelerates the design and manufacturing time of such a machine. The use of such a neural network also significantly reduces the time required during the data collection phase.

[0016]

[0009] Furthermore, by providing the geometry and physical equations behind the characterization of the rotating electromagnetic machine, such a physics-informed neural network is able to solve the various equations in order to calculate different key performance indicators. Consequently, this constrains the neural network and leads to less data being needed for good generalization of the neural network.

[0017]

[0010] According to a preferred embodiment of the invention, the rotating electromagnetic machine is an axial magnetic flux machine.

[0018]

[0011] According to one variant, the neural network is a hybrid neural network, which is both a Bayesian neural network and a physics-informed neural network.

[0019]

[0012] According to one embodiment, the neural network is trained using data collected by the computer system, said data being measured or simulated data relating to physical parameters of the rotating electromagnetic machine. This allows the neural network to converge rapidly towards a solution capable of performing good interpolation and extrapolation, since the laws of physics play a regularizing role.

[0020]

[0013] According to one variant, the neural network is retrained regularly for each new data collection performed by the computer system. This further improves the performance of the neural network in terms of computation time and the accuracy of optimizing a design model.

[0021]

[0014] According to one embodiment, the method further comprises an iterative step of calculating, using the computing means of the computer system, from the predicted current two-dimensional model, a quantification of the uncertainty of the two-dimensional model, and a step of comparing, using the computing means of the computer system, the calculated quantification of uncertainty to a threshold quantification of uncertainty. This makes it possible to estimate the level of confidence that can be placed in the prediction of a current two-dimensional model performed by the neural network. When the level of confidence is estimated to be insufficient, a further step of simulating a current two-dimensional model can, for example, be carried out.Furthermore, when such a step of predicting an uncertainty quantification of the two-dimensional model is performed by linear approximation of the magnetic flux as a function of the two vector control currents, this helps the neural network in its generalization and reduces the time and complexity of the computation.

[0015] According to one embodiment, if the calculated uncertainty quantification is less than or equal to the threshold uncertainty quantification, the step of calculating a vector control strategy is performed, and if the calculated uncertainty quantification is greater than the threshold uncertainty quantification, the method further includes an intermediate step of simulating, using the computing means of the computer system, a current two-dimensional model from the selected current design model.

[0022]

[0016] According to one variant, the optimizer is chosen from the group consisting of: a Bayesian optimizer, an optimizer implementing a genetic algorithm, and an optimizer implementing a learning algorithm using a Parzen estimator.

[0023]

[0017] The invention also relates to an assembly comprising at least one device for manufacturing a rotating electromagnetic machine and a computer system connected to said at least one manufacturing device, the computer system comprising computing means and memory means connected to the computing means, said memory means storing an optimizer and a neural network configured at least as a physics-informed neural network, the optimizer being connected to the neural network and being configured to define distinct design models for the electromagnetic machine, each design model corresponding to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine, the neural network being configured to predict,from a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine and two vector control currents, the first current allowing adjustment of the magnetic flux of the rotating electromagnetic machine, and the second current allowing adjustment of the torque of the rotating electromagnetic machine, a two-dimensional model relating to two fluxes obtained from said first and second currents, the whole being configured to implement the steps of the process as described above.

[0024]

[0018] By way of non-limiting examples, embodiments of the present invention will be described below, with reference to the accompanying figures in which: - [Fig. 1] schematically illustrates an assembly according to the invention, the assembly comprising a device for manufacturing a rotating electromagnetic machine and a computer system; and

[0025] - [Fig.2] is a flowchart representing a manufacturing process for a rotating electromagnetic machine with polyphase electric current, implemented by the assembly of figure 1, according to the present invention.

[0026]

[0019] For the purposes of this invention, a "physics-informed neural network" means any neural network that combines machine learning and physical laws. In such a neural network, also called a PINN (Physics-Informed Neural Network), the physical laws that can be integrated are very diverse, ranging from fluid mechanics (e.g., Navier-Stokes equations) to electromagnetism (Maxwell's equations), including thermodynamics (Fourier equations), mechanics and powder theory (Dirichlet equations), quantum mechanics (Schrödinger equation), and even wave modeling on free surfaces (KdV equations for Korteweg-de Vries). More generally, any physical law in the form of differential equations can be translated into use within a physics-informed neural network.All areas of physics are affected because the vast majority of physical laws are represented by partial differential equations. A neural network informed by physics thus makes it possible to encode the laws of physics into learning algorithms, making them more robust and efficient, and also allowing for a more accurate representation of the reality of physical phenomena.

[0027]

[0020] With reference to Figure 2, the present invention relates to a method, implemented by an assembly 2 (visible in Figure 1), for manufacturing a rotating electromagnetic machine with polyphase electric current intended for use in an electric or hybrid motor vehicle. Preferably, but not limitingly within the scope of the present invention, the rotating electromagnetic machine with polyphase electric current is an axial magnetic flux machine, typically a permanent magnet electric motor.

[0021] The assembly 2 comprises a device 4 for manufacturing the rotating electromagnetic machine and a computer system 6 connected to the manufacturing device 4. The computer system 6 includes computing means 8 and memory means 10 connected to the computing means 8. The memory means 10 store an optimizer 12 and a neural network 14.

[0028]

[0022] The optimizer 12 is connected to the neural network 14 and is configured to define distinct design models for the electromagnetic machine, each design model corresponding to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine. Such physical parameters are, for example, the thickness of the machine's magnetic poles, the stack length, or the machine radius, etc. The optimizer 12 is, for example, a Bayesian optimizer, an optimizer implementing a genetic algorithm, or an optimizer implementing a learning algorithm using a Parzen estimator. When the optimizer 12 is an optimizer implementing a genetic algorithm, the algorithm in question is, for example, an NSGA-II type algorithm (from the English acronym "Non-Sorted Genetic Algorithm").When optimizer 12 is an optimizer implementing a learning algorithm using a Parzen estimator, the algorithm in question is, for example, a TPE type algorithm (from the English acronym "Tree-Structured Parzen Estimator >>).

[0029]

[0023] The neural network 14 is configured at least as a physics-informed neural network. According to a particular embodiment of the invention, the neural network 14 is a hybrid neural network, which is both a Bayesian neural network and a physics-informed neural network. The neural network 14 is configured to predict a two-dimensional model (also called the "DQ" model) from a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine, and from two vector control currents, the first of which allows adjustment of the magnetic flux of the rotating electromagnetic machine, and the second of which allows adjustment of the torque of the rotating electromagnetic machine. The two-dimensional model relates to two fluxes obtained from the first and second vector control currents.Each point in such a two-dimensional model corresponds to an operating point of the rotating electromagnetic machine associated with a torque (rotational speed; torque) of the machine.

[0030]

[0024] With reference to figure 2, an embodiment of the manufacturing process for a rotating electromagnetic machine with polyphase electric current intended to equip an electric or hybrid motor vehicle according to the invention will now be described, implemented by an assembly 2 as described previously.

[0031]

[0025] It is assumed that initially the neural network 14 is trained using data collected by the computer system 6, such data being measured or simulated data relating to physical parameters of the rotating electromagnetic machine. When the neural network 14 is a hybrid neural network, which is both a Bayesian neural network and a physics-informed neural network, the neural network 14 is preferably retrained regularly for each new data collection performed by the computer system 6. The manufacturing process comprises steps 20 to 30 which are iteratively repeated until a predetermined criterion is met. This predetermined criterion includes at least one metric and consists, for example, of a predefined number of iterations or the optimization of a metric such as the price of the rotating electromagnetic machine.

[0032]

[0026] The process thus includes an initial step 20 during which the optimizer 12 selects a current design model for the rotating electromagnetic machine. As explained previously, such a current design model corresponds to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine.

[0033]

[0027] The method includes a subsequent step 22 in which the neural network 14 predicts a current two-dimensional model from the selected current design model. The current two-dimensional model relates to two fluxes obtained from the first and second vector control currents. To do this, the neural network 14 performs a linear approximation of the magnetic flux of the rotating electromagnetic machine at every point of the two-dimensional model (in other words, a linear or pseudo-linear function assumption for the flux is taken as input to the neural network 14, and the loss function of the latter is represented by a linear equation with a predefined number of coefficients). The linear equation is a polynomial equation that allows the two fluxes supplied as output by the neural network 14 to be related to the first and second vector control currents taken as input to the latter.

[0034]

[0028] Preferably, as illustrated in Figure 2, the method includes a subsequent step 24 in which the computing means 8 of the computer system 6 calculate, from the predicted current two-dimensional model, a quantification of the uncertainty of the two-dimensional model. To do this, the computing means 8 of the computer system 6 reuse the linear equation used by the neural network 14 during the prediction step 22, and verify this equation a posteriori with estimated coefficients, in order to verify the quantified uncertainty. It should be noted that there are at least two ways to quantify the uncertainty: either with the neural network informed by the physics 14 directly by checking whether the equation is satisfied, or by using a Bayesian (which is configured to provide distributions and therefore standard deviations), or even via a combination of these two methods.The process includes a subsequent step 25 in which the computing means 8 of the computer system 6 compare the calculated uncertainty quantification to a threshold uncertainty quantification.

[0035]

[0029] At the end of the comparison step 25, if the calculated uncertainty quantification is less than or equal to the threshold uncertainty quantification, the method includes a subsequent step 28 in which the computing means 8 of the computer system 6 calculate a vector control strategy from the predicted current two-dimensional model. The vector control strategy includes current setpoints relating to the two flows of the predicted current two-dimensional model.

[0036]

[0030] At the end of the comparison step 25, if the calculated uncertainty quantification is greater than the threshold uncertainty quantification, the process includes an intermediate step 26 during which the computing means 8 of the computer system 6 simulate a current two-dimensional model, based on the selected current design model. The calculation step 28 is then implemented.

[0037]

[0031] Following calculation step 28, the process includes a subsequent step 30 in which the computing means 8 of the computer system 6 calculate at least one metric related to the current design model. Such a metric makes it possible, in particular, to characterize the rotating electromagnetic machine in terms of physical and / or economic performance, and to verify whether or not the predetermined criterion mentioned above is met. The calculation of such a metric may also involve calculating one or more key performance indicators (such as, for example, manufacturing cost, weight, power, torque, energy density, etc.).

[0038]

[0032] If the predetermined criterion is not satisfied, steps 20 to 30 are repeated, with a current design model selected by the optimizer 12 during the initial step 20.

[0039]

[0033] If the predetermined criterion is met, the manufacturing process comprises the following steps 32 to 34.

[0040]

[0034] In a subsequent step 32, the computer system 6 selects the current design pattern as the design pattern to be manufactured for the rotating electromagnetic machine.

[0041]

[0035] During a final step 34, the manufacturing device 4 manufactures the rotating electromagnetic machine, based on the design model to be manufactured.

[0042]

[0036] The method according to the invention thus makes it possible to reduce the computation time of the steps implemented by the computer system 6 and / or the amount of training data required for the neural network 14, while accelerating the design and therefore the manufacturing time of the machine, compared to simulation tools or prior art artificial intelligence solutions. Furthermore, the use of a physics-informed neural network 14 makes it possible to considerably reduce the time required during the data collection phase.

Claims

DEMANDS 1. A method for manufacturing a rotating electromagnetic machine with polyphase electric current intended for use in an electric or hybrid motor vehicle, the method being implemented by an assembly (2) comprising at least one device (4) for manufacturing a rotating electromagnetic machine and a computer system (6) connected to said at least one manufacturing device (4), the computer system (6) comprising computing means (8) and memory means (10) connected to the computing means (8), said memory means (10) storing an optimizer (12) and a neural network (14) configured at least as a physics-informed neural network, the optimizer (12) being connected to the neural network (14) and being configured to define distinct design models for the electromagnetic machine, each design model corresponding to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine,the neural network (14) being configured to predict, from a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine and two vector control currents, a first current allowing adjustment of the magnetic flux of the rotating electromagnetic machine, and a second current allowing adjustment of the torque of the rotating electromagnetic machine, a two-dimensional model relating to two fluxes obtained from said first and second currents, the process being characterized in that it comprises the following steps iteratively looped until a predetermined criterion is satisfied, said predetermined criterion including at least one metric:, - selection (20), by the optimizer (12), of a current design model for the rotating electromagnetic machine; - prediction (22) by the neural network (14), from the selected current design model, of a current two-dimensional model; - calculation (28) by the computing means (8) of the computer system (6), from the predicted current two-dimensional model, of a strategy of vector control including current instructions relating to said two flows; - calculation (30) by the computing means (8) of the computer system (6), based on the calculated vector control strategy, of at least one metric relating to the current design model; and in that the method further comprises the following steps, when said predetermined criterion is met: - selection (32), by the computer system (6), of the current design model as the design model to be manufactured for the rotating electromagnetic machine; - manufacture (34), by said at least one manufacturing device (4), of the rotating electromagnetic machine on the basis of the design model to be manufactured.

2. Method according to claim 1, wherein the rotating electromagnetic machine is an axial magnetic flux machine.

3. Method according to claim 1 or 2, wherein the neural network (14) is a hybrid neural network, which is both a Bayesian neural network and a physics-informed neural network.

4. A method according to any one of claims 1 to 3, wherein the neural network (14) is trained using data collected by the computer system (6), said data being measured or simulated data relating to physical parameters relating to the rotating electromagnetic machine.

5. A method according to claim 4 when it depends on claim 3, wherein the neural network (14) is retrained regularly, for each new data collection carried out by the computer system (6).

6. A method according to any one of claims 1 to 5, wherein the method further comprises an iterative calculation step (24), by the computing means (8) of the computer system (6), from the predicted current two-dimensional model, of an uncertainty quantification of the two-dimensional model, and a comparison step (25), by the computing means (8) of the computer system (6), of the calculated uncertainty quantification to a threshold uncertainty quantification.

7. A method according to claim 6, wherein, if the calculated uncertainty quantification is less than or equal to the threshold uncertainty quantification, the step (28) of calculating a vector control strategy is carried out and, if the calculated uncertainty quantification is greater than the threshold uncertainty quantification, the method further comprises an intermediate step (26) of simulating, by the computing means (8) of the computer system (6), from the selected current design model, a current two-dimensional model.

8. A method according to any one of claims 1 to 7, wherein the optimizer is chosen from the group consisting of: a Bayesian optimizer, an optimizer implementing a genetic algorithm, and an optimizer implementing a learning algorithm using a Parzen estimator.

9. Assembly (2) comprising at least one device (4) for manufacturing a rotating electromagnetic machine and a computer system (6) connected to said at least one manufacturing device (4), the computer system (6) comprising computing means (8) and memory means (10) connected to the computing means (8), said memory means (10) storing an optimizer (12) and a neural network (14) configured at least as a physics-informed neural network, the optimizer (12) being connected to the neural network (14) and being configured to define distinct design models for the electromagnetic machine, each design model corresponding to a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine, the neural network (14) being configured to predict,from a set of instantiated values ​​of physical parameters relating to the rotating electromagnetic machine and two vector control currents, a first current allowing adjustment of the magnetic flux of the rotating electromagnetic machine, and a second current allowing adjustment of the torque of the rotating electromagnetic machine, a two-dimensional model relating to two fluxes obtained from said first and second currents, in which the assembly (2) is configured to implement the steps of the process according to any one of claims 1 to 8.

Images