A method for optimizing a magnetic coupling mechanism of a wireless power transfer device

Abstract

The application discloses a kind of magnetic coupling mechanism optimization design methods of wireless energy transmission device, belong to wireless power transmission technical field.The method includes: determining the initialization parameter of magnetic coupler;Physical information neural network model is constructed, and the magnetic vector potential distribution of magnetic coupler is output;Geometric parameter optimization loss function is constructed;Gradient of geometric parameter optimization loss function to structure parameter is calculated using automatic differentiation technique;According to the gradient, the structure parameter is iteratively updated using optimization algorithm;Judge whether the optimization process converges, if convergence then output optimal structure parameter.The present application combines electromagnetic physical law and data-driven model through physical information neural network, realizes the differentiable optimization of magnetic coupler parameter, solves the problems such as high calculation cost, low efficiency, insufficient optimization precision and poor flexibility of existing magnetic coupler design method.

Description

Technical Field

[0001] This invention relates to the field of wireless power transmission technology, and more specifically to an optimized design method for the magnetic coupling mechanism of a wireless power transmission device. Background Technology

[0002] Wireless power transfer (WPT) technology, as a contactless power transmission method, has been widely used in recent years in fields such as implantable medical devices, smart homes, electric vehicle charging, and consumer electronics. Among them, magnetic field coupled power transfer (MCPT) has become the mainstream technology for wireless charging systems for electric vehicles due to its high transmission efficiency and relatively controllable electromagnetic exposure risks. Figure 1 As the core component for energy transfer in MCPT systems, the magnetic coupler's structural design and parameter optimization directly determine the system's transmission power, efficiency, and offset resistance, making it a key constraint on system performance. Currently, modeling and optimization methods for magnetic couplers mainly fall into two categories: physics-based numerical simulation and data-driven alternative models. However, many technical bottlenecks still exist in practical engineering applications. Representative solutions in existing technologies include the following: (1) An optimization design method based on an equivalent circuit model. Patent document (CN107425611A) discloses an optimization design method for the magnetic coupling mechanism of a wireless power transmission device. This method establishes an equivalent circuit model of the magnetic coupling mechanism, derives the relationship curve between system efficiency and size parameters, and then determines parameters such as the number of coil turns and size. Although this method simplifies the design process, its prediction accuracy is limited because the equivalent circuit model idealizes and simplifies complex physical phenomena such as edge effects and material nonlinearity. In particular, it is difficult to accurately reflect the real electromagnetic field distribution in the design of new or complex structures, resulting in the optimization results deviating from the optimal solution of actual performance.

[0003] (2) Combining intelligent algorithms with multi-objective optimization methods, patent document (CN120217775A) discloses a multi-objective optimization design method for wireless charging systems. By introducing K-means clustering and non-dominated sorting to improve the particle swarm optimization algorithm, it achieves collaborative optimization of multiple parameters such as number of turns, size, and shape. Although this method supports decision-maker preference settings, each iteration requires calling high-fidelity finite element simulation (FEM) to calculate the objective function (such as mutual inductance and self-inductance), resulting in hundreds or thousands of FEM solutions required for the optimization process. This consumes huge computational resources and the design cycle can last for days or even weeks, making it difficult to meet the needs of rapid iteration.

[0004] (3) Hardware structure integration design method. Patent document (CN116131475B) discloses an IPT system with a highly integrated magnetic coupler and an IPT system integration method. By introducing hardware structures such as a reverse coil and integrated inductors on the primary / secondary sides, the anti-offset characteristics and zero-voltage switching performance of the system are improved. Although this solution has excellent performance in specific scenarios, it relies on additional hardware, which leads to a complex magnetic coupler structure, increased manufacturing costs, and a design process that is specific to a particular topology, lacking universality and making it difficult to adapt to the flexible design requirements of multiple scenarios and multiple constraints.

[0005] (4) The method combining parameter scanning and Pareto optimization: Patent document (CN111428337B) discloses a design method and system for a magnetic coupler used in a wireless charging system for maglev trains. This method generates a set of solution points by traversing the parameter value range, and then combines constraint screening and multi-objective optimization to determine the optimal design scheme. While this method guarantees the existence of solutions, its accuracy heavily depends on the parameter step size setting: too small a step size leads to combinatorial explosion, while too large a step size easily misses the globally optimal solution. Simultaneously, a large amount of computational resources are consumed in evaluating non-optimal solutions, resulting in low optimization efficiency.

[0006] In summary, existing magnetic coupler optimization methods generally suffer from high computational cost, low efficiency, insufficient optimization accuracy, and poor flexibility. Especially in high-precision, multi-parameter, and fast-iteration applications (such as dynamic wireless charging for electric vehicles), traditional methods struggle to balance efficiency and accuracy. Therefore, a novel optimization design method based on magnetic coupling mechanisms is urgently needed to address these issues. Summary of the Invention

[0007] The purpose of this invention is to provide an optimization design method for the magnetic coupling mechanism of a wireless power transmission device. By combining electromagnetic physical laws with a data-driven model through a Physical Information Neural Network (PINN), the differentiable optimization of the magnetic coupler parameters can be achieved, thereby solving the problems of high computational cost, low efficiency, insufficient optimization accuracy, and poor flexibility of existing magnetic coupler design methods.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for optimizing the magnetic coupling mechanism of a wireless power transfer device includes the following steps: S1: Determine the initialization parameters of the magnetic coupler and generate the corresponding vectorized structure parameters; S2: Construct a physical information neural network model, take the parameter vector and spatial coordinates as input, output the magnetic vector potential distribution of the magnetic coupler, and train the model to obtain a surrogate model that can predict the electrical parameters of the magnetic coupler; S3: Construct a geometric parameter optimization loss function, which includes a relative error term based on the target electrical parameters and the predicted electrical parameters, and a penalty function term based on geometric constraints; S4: Using automatic differentiation techniques, calculate the gradient of the geometric parameter optimization loss function with respect to the structural parameters; S5: Based on the gradient, the structural parameters are iteratively updated using an optimization algorithm; S6: Determine whether the optimization process has converged. If it has converged, output the optimal structure parameters.

[0009] Furthermore, in S1: The initialization parameters include three categories: coil geometric parameters, magnetic material parameters, and shielding layer parameters; The structural parameters are an ordered vector formed by combining initialization parameters in a predetermined order.

[0010] Furthermore, the physical information neural network model is constructed in S2, including: The three-dimensional model of the magnetic coupler is reduced to a 2.5-dimensional model, and multiple material subdomains and corresponding boundary conditions are defined. Based on the vector magnetic potential method, the partial differential equations of magnetic potential and boundary conditions for each material subdomain are derived, and these are embedded as physical constraints into the loss function of the model.

[0011] Furthermore, the total loss function of the physical information neural network model is composed of a weighted sum of multiple losses, including partial differential equation residual loss, boundary condition loss, material interface continuity loss, and finite element simulation data calibration loss.

[0012] Further, in S3: the relative error term is used to measure the deviation between the predicted electrical parameter value and the respective target electrical parameter value, wherein the electrical parameters include self-inductance, mutual inductance and coupling coefficient.

[0013] Furthermore, the penalty function term is implemented using a smooth penalty function, including turn spacing constraints, coil size envelope constraints, material layer size constraints, working gap constraints, and turn number discrete constraints.

[0014] Furthermore, the gradient calculation described in S4 is performed by using the chain rule and automatic differentiation techniques to calculate the sum of the partial derivatives of the loss function with respect to the structural parameters.

[0015] Furthermore, in S5: The optimization algorithms include the Adaptive Moment Estimation Optimization Algorithm ADAM and the Finite Memory Boundary Constraint Quasi-Newton Algorithm L-BFGS-B. The iterative update constrains the updated parameters within a preset feasible region using a projection operator.

[0016] Another object of the present invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for optimizing the magnetic coupling mechanism of a wireless power transmission device.

[0017] Another object of the present invention is to provide an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the magnetic coupling mechanism optimization design method of the wireless power transmission device.

[0018] This invention achieves a significant improvement in technical efficiency through an optimized design method for magnetic coupling mechanisms based on Physical Information Neural Network (PINN). Specific beneficial effects include: 1. High-precision and high-fidelity modeling: By embedding electromagnetic physical laws such as Maxwell's equations into the neural network in the form of partial differential equation constraints, the model can accurately capture complex electromagnetic field distributions with the support of a small amount of limited metadata. This overcomes the accuracy limitations caused by the simplification assumptions of traditional equivalent circuit models and significantly improves prediction reliability.

[0019] 2. Breakthrough improvement in computational efficiency: The constructed differentiable surrogate model enables instantaneous prediction of electrical parameters, transforming the traditional optimization process that relies on hundreds or thousands of FEM simulations into a single forward inference and gradient backpropagation. The design cycle is shortened from several days to hours, greatly reducing the consumption of computing resources and computational costs.

[0020] 3. It flexibly adapts to the needs of multiple scenarios, adopts a continuous relaxation strategy to handle discrete variables such as the number of turns, and combines differentiable penalty functions to embed complex constraints such as geometry and thermal-EMI. High-performance designs that meet different application scenarios can be quickly generated through software optimization on a fixed hardware platform, and it has strong portability.

[0021] 4. The innovative implementation of reverse optimization capability, relying on an end-to-end differentiable architecture, has for the first time realized gradient-driven reverse optimization from system-level performance objectives (such as L, M, k) to geometric parameters, avoiding the blindness of traditional trial and error methods and making the design process more targeted and globally optimal. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a typical block diagram of a wireless power transfer system for electric vehicles. Figure 3 This is a front view of the magnetic energy coil structure of the present invention; Figure 4 This is a side view of the magnetic energy coil structure of the present invention; Figure 5This is a schematic diagram of the three-dimensional magnetic coupler structure of the present invention; Figure 6 This is a schematic diagram of a 2.5-dimensional reduced-order model of the three-dimensional magnetic coupler of the present invention, wherein (a) is a schematic diagram of the 2.5-dimensional reduced-order model with the upper half of the model hidden, and (b) is a front view of the 2.5-dimensional reduced-order model; Figure 7 This is a model architecture diagram of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] This embodiment provides an optimized design method for the magnetic coupling mechanism of a wireless power transfer device, such as... Figure 1 As shown, the process includes: determining the initialization parameters of the magnetic coupler and generating corresponding vectorized structural parameters; constructing a physical information neural network model, taking the parameter vectors and spatial coordinates as input, outputting the magnetic vector potential distribution of the magnetic coupler, and training the model to obtain a surrogate model capable of predicting the electrical parameters of the magnetic coupler; constructing a geometric parameter optimization loss function, which includes a relative error term based on the target electrical parameters and the predicted electrical parameters, and a penalty function term based on geometric constraints; using automatic differentiation techniques to calculate the gradient of the geometric parameter optimization loss function with respect to the structural parameters; iteratively updating the structural parameters using an optimization algorithm based on the gradient; and determining whether the optimization process has converged, and if it has converged, outputting the optimal structural parameters.

[0025] Wireless charging systems for electric vehicles are based on magnetic field-coupled wireless power transfer technology, which uses high-frequency magnetic fields to achieve contactless power transfer. For example... Figure 2 As shown, the core of the system comprises four units: the primary-side transmitting unit converts the input power into high-frequency AC power via an inverter; the secondary-side receiving unit outputs DC power adapted to the battery via a rectifier and DC / DC converter; the wireless power transmission unit consists of a magnetic coupler and a compensation network, transferring energy through magnetic field coupling between coils; and the control and communication unit adjusts system parameters in real time to ensure stable operation. In this system, the magnetic coupler, as the core component for energy transmission, involves multi-physics coupling issues involving electromagnetic fields, thermodynamics, and structural mechanics. Its design directly determines the system's transmission efficiency, power capacity, and anti-offset capability, making it a key challenge for technical optimization.

[0026] The following describes the specific process of an optimized design method for the magnetic coupling mechanism of a wireless power transmission device, using examples.

[0027] Step 1: Determine the initialization parameters of the magnetic coupler and generate the corresponding vectorized structure parameters.

[0028] This embodiment uses the dual-D coil structure, widely adopted in wireless charging systems for electric vehicles, as a typical example. This structure optimizes the magnetic field distribution and improves resistance to deflection through a symmetrical coil layout. Its complete parametric geometric model comprises three layers: a Litz wire winding layer, a ferrite magnetic layer, and an aluminum plate shielding layer. All parameters must be determined collaboratively based on electromagnetic physics laws and engineering constraints. Figures 3-4 As shown.

[0029] For ease of practical operation, the parameters are divided into three categories: coil geometric parameters, magnetic material parameters, and shielding layer parameters, as follows: 1. Coil geometry parameters (determine electromagnetic coupling performance) Litz wire line width w: the equivalent diameter determined by the number of strands and the diameter of each strand. Based on the target current RMS value. Select the standard specification. For example: If =20A, optional 1000 strands of 0.1mm wire, w≈3.5mm.

[0030] Number of turns N: The total number of turns of the Litz wire in a single-sided coil, assuming symmetry between the primary and secondary sides. Based on an initial estimate of the target inductance value Ltarget, a modified Wheeler formula can be used: in The average area of ​​the coil, , This is the shape factor (1.2~1.5 for dual-D coils). The initial value can be set to 8~12 turns, and then refined through optimization algorithms.

[0031] Turn spacing s: The center distance between adjacent Litz lines, which must satisfy s≥w to prevent inter-turn short circuits. The initial value is taken as s=1.5w-2w, balancing compactness and insulation requirements. It is treated as a continuous variable in the optimization.

[0032] The fillet radius r is the inner diameter or radius of curvature of the Litz wire when it bends at the corner of the coil. r must meet the minimum bending radius constraint. This minimum value is determined by the specifications of the selected Litz wire and is usually proportional to the wire width w. For example, r is usually ≥ 4w-5w.

[0033] Working gap g: The vertical distance between the primary and secondary coils, which is a core constraint. In the static charging scenario of electric vehicles, g = 150mm-200mm.

[0034] The outline dimensions (length) of the Lids line region ,width ): Overall outer envelope dimensions and profile dimensions of the coil winding portion (including all turns). , The outline length is a secondary parameter derived from the basic parameters (N, s, w), and is strictly constrained by the upper limit of the installation space. For square or double D-type coils, the outline dimensions can be directly calculated using more fundamental parameters such as the number of turns N, line width w, and turn spacing s. For a simple square coil, its outline length... It can be approximated as: Due to symmetry, width It also follows the same relationship.

[0035] 2. Magnetic materials and shielding layer parameters (optimizing magnetic field path and suppressing leakage) Contouring magnetic material manganese zinc ferrite layer profile dimensions (length) ,width ,high ): Guides magnetic flux and increases coupling coefficient. Material can be DMR95 (μr≈2000). Dimensional relationship: , The remaining amount Take 10%~20%, thickness Take a diameter of 4mm to 8mm.

[0036] The outline dimensions (length) of the aluminum plate shielding layer ,width ,high Function: To limit magnetic field dissipation and meet EMC requirements. Dimensional relationships: , margin Take 5%~10%, thickness It needs to be greater than the skin depth (the skin depth of aluminum is about 0.3 mm at 85 kHz, and is actually taken as 1 mm to 3 mm).

[0037] Therefore, based on the target design values ​​of the electrical parameters of the magnetic coupling mechanism in the wireless power transmission system, the initial values ​​of each parameter of the magnetic coupler mechanism are determined, and the value range of each parameter to be optimized in the magnetic coupling mechanism is defined. The parameters to be optimized include the number of Litz wire turns N, the turn spacing s, the working gap g, and the length and width of the ferrite core. and .

[0038] Based on the above parameters, vectorized structural parameters p are generated that can be directly used in the PINN physical information neural network model: .

[0039] Step 2: Construct a physical information neural network model, take the parameter vector and spatial coordinates as input, output the magnetic vector potential distribution of the magnetic coupler, and train the model to obtain a surrogate model that can predict the electrical parameters of the magnetic coupler.

[0040] To simplify the computational complexity, for example... Figure 5 The 3D magnetic coupler model shown is subjected to dimensionality reduction processing. By using techniques such as layer overlay, shading, and perspective to simulate a three-dimensional effect on a plane, the 3D model of the magnetic coupler is flattened into a 2.5D model based on the 2D model, resulting in the model shown. Figure 6 The 2.5D structural model of the magnetic coupler in Cartesian coordinates is shown below. (a) is a schematic diagram of the 2.5D reduced-order model with the upper half hidden, and (b) is the front view of the 2.5D reduced-order model. The simplified two-dimensional model consists of seven regions and eight boundaries. The seven regions are as follows: secondary side aluminum plate shielding layer region I, secondary side ferrite layer region II, secondary side coil Litz line region III, air gap region IV, primary side coil Litz line region V, primary side ferrite layer region VI, and primary side aluminum plate shielding layer region VII. The eight boundaries are as follows: z1, the outer boundary of the secondary side aluminum plate shielding layer area; z2, the overlapping boundary between the secondary side aluminum plate shielding layer area and the secondary side ferrite layer area; z3, the overlapping boundary between the secondary side ferrite layer area and the secondary side coil Litz line area; z4, the overlapping boundary between the secondary side coil Litz line area and the air gap area; z5, the overlapping boundary between the air gap area and the primary side coil Litz line area; z6, the overlapping boundary between the primary side coil Litz line area and the primary side ferrite layer area; z7, the overlapping boundary between the primary side ferrite layer area and the primary side aluminum plate shielding layer area; z8, the outer boundary of the primary side aluminum plate shielding layer area.

[0041] Based on the 2.5D structural model of the magnetic coupler, the partial differential equation of the magnetic vector potential and the boundary conditions of the magnetic vector potential are obtained using the vector magnetic potential method, as follows: In constructing the vector partial differential equations for the magnetic coupler, the constitutive relations of magnetic induction and magnetic field strength are used as a starting point, based on Ampere's law and the definition of magnetic vector potential. The frequency domain is represented by time harmonics. In non-axisymmetric geometry, Fourier expansion is no longer used; instead, the magnetic vector potential components can be directly established in three-dimensional space. .

[0042] The frequency domain form of the resulting magnetic-quasi-static equation is: If time harmonics are used, it can be set as follows: in Permeability, Where J is the equivalent conductivity, and J is the volume current density in each region.

[0043] And in locally uniform regions, it can be written in a form similar to Helmholtz: The equations for the components in each direction are identical (i.e., Ax, Ay, and Az all satisfy the same equation). (The operator), which is precisely the result of "local homogeneity" and "isotropy". Three special cases also need to be considered: 1) Non-conductive region (air / ferrite): The Laplace equation is obtained. (The region without a source of life degenerates into) ); 2) Approximation of a good conductor: , Skin depth; the field inside the conductor decays exponentially with the normal. 3) Anisotropic media: If μ and σ are tensors, then they should be retained. The tensor form cannot be simplified to a scalar in this case. .

[0044] Based on this, the vector partial differential equations (magnetic vector potential partial differential equations) corresponding to the following regions in the magnetic coupler are derived: secondary side aluminum plate shielding layer region I, secondary side ferrite layer region II, secondary side coil Litz wire region III, air gap region IV, primary side coil Litz wire region V, primary side ferrite layer region VI, and primary side aluminum plate shielding layer region VII. The partial derivatives in these equations can be obtained through automatic differentiation. The specific equations are as follows: Where A1~A7 are dependent variables, A1 is the magnetic vector potential of the secondary side aluminum plate shielding layer region, A2 is the magnetic vector potential of the secondary side ferrite layer region, A3 is the magnetic vector potential of the secondary side coil Litz wire region, A4 is the magnetic vector potential of the air gap region, A5 is the magnetic vector potential of the primary side coil Litz wire region, A6 is the magnetic vector potential of the primary side ferrite layer region, A7 is the magnetic vector potential of the primary side aluminum plate shielding layer region, z is the sum of the heights of the primary and secondary side aluminum plates, primary and secondary side ferrite, primary and secondary side Litz wires and air gap, μ0 is the vacuum permeability, σ Al σ Cu The conductivity of the aluminum plate and ferrite.

[0045] Furthermore, when constructing the magnetic vector potential boundary conditions for the magnetic coupler, due to the continuity of the magnetic field strength and magnetic flux density, the following relationship exists at the boundaries of each region. Where μ Al μ is the equivalent permeability of the aluminum plate. Fe Let x, y, and z be the equivalent permeability of the ferrite, and A1 to A7 be the dependent variables. Based on the above process, the PINN model, a physical information neural network model proposed in this invention, is based on the PyTorch deep learning framework. It utilizes the Python programming language to construct a fully connected neural network on the Spyder platform, comprising an input layer, hidden layers, and an output layer. The parameters to be optimized are... Using spatial variables x, y, and z as the input layer of a neural network for physical information, a neural network framework is constructed, outputting the self-inductance of the magnetic coupler coil. Mutual induction Output layer.

[0046] In the Physical Information Neural Network (PINN) model, x, y, and z are not continuous but participate in the calculation as discrete sampling points. These sampling points are generated within a certain region of the two-dimensional plane defined above. This region needs to be large enough to encompass the entire magnetic coupler and approximate an open boundary (or satisfy the set boundary conditions). Inside the region (such as inside the air gap, coil, ferrite, etc.), random sampling (such as Latin hypercube sampling) or uniform grid sampling is often used. Random sampling helps avoid overfitting and allows the model to better learn the physical laws. At the boundaries of the region (such as z1, z8) and at the interfaces of different materials (such as z2-z7), dense sampling is required to ensure that the boundary conditions are accurately satisfied. For each point that the neural network needs to predict, we assign it a specific pair of (x, y, z) coordinate values. The set of all these coordinate values, together with the structural parameters p of the magnetic coupler at this time (such as the number of turns, working gap, etc.), constitutes the features of the input layer of the PINN model.

[0047] With 5 parameters to be optimized The physical information neural network consists of the number of turns, turn spacing, working gap, and ferrite length and width, along with spatial variables x, y, and z. It has 5 hidden layers and 128 neurons per layer, and outputs the coil self-inductance. Mutual induction This is the output layer. The output of each layer of the neural network is shown in the following formula: In the formula, , The first Layer weight matrix and bias vector, For the first Layer output, As an activation function, each layer in a neural network has weights. and bias Correspondingly, this example selects the Tanh function, which is suitable for this model, as the activation function. It has the advantage of low computational cost in expressing the partial differential equation of the magnetic vector potential in the model, resulting in faster convergence of the neural network and providing more stable gradient flow. Its expression is: .

[0048] Calibration was then performed using lightweight FEM data, utilizing k sets of data from finite element simulations based on the Maxwell module of Ansys Electronics Suite 2023 R1 software. The specific steps for finite element data calibration are as follows: a. Each set of data is treated as one input: , in It includes complete geometric and material parameters (number of turns, turn spacing, wire diameter, coil shape, working gap, ferrite / aluminum plate size and thickness, etc.). b. Use PINN to solve for the structural parameters of these K groups under unit current to obtain the "original" inductance. For each group The forward solution in the frequency domain, A (considering PDE partial differential equations, boundary conditions, and interface continuity), is calculated using the following formula: in For magnetic field energy, The magnetic flux passing through the receiving coil window (both are obtained by automatic differentiation and numerical integration from the PINN output A or B); c. Using least squares to fit the "calibration mapping", a common and stable approach is linear calibration: The analytical solution is: , ; d. will It is saved together with the PINN weights and can be directly called in subsequent predictions.

[0049] Output and These two weight values ​​can be used to correct and calibrate the values ​​predicted by the PINN model. They are used as a term in the loss function of the PINN model and weighted to improve the overall prediction accuracy.

[0050] Then, the total loss function of the PINN model is calculated. Based on the vector magnetic potential method, the partial differential equations of magnetic potentials and boundary conditions for each material subdomain are derived and embedded as physical constraints into the loss function of the model. Combining the previous seven regions and eight boundary conditions, the computational domain is divided into seven material subdomains: The sets are formed by the intersection of each pair of interfaces. outer boundary The dependent variable is the complex magnetic vector potential A. Material properties are defined as follows: , Electrical conductivity; For the remanence spectrum, apply a volume current density. (Only the region containing the Lids line has a non-zero value). Electromagnetic parameters are expressed as follows: , (Quasi-static) .

[0051] In addition to obtaining the scaling coefficients through fitting learning, the introduced finite element data should also be weighted and incorporated into the total loss function for calculating the magnetic vector potential. All four terms in the total loss function are obtained using the mean square error (MSE) metric, aiming to ensure that the model's predictions strictly adhere to the laws of electromagnetic physics and accurately match the calibration data of the finite element simulation (FEM). The expression for the total loss function is: The following section provides a detailed explanation of the four components of the loss function and their corresponding sampling strategies.

[0052] (1) Residual loss term of partial differential equation PDE This term is the core of the physical constraints, ensuring that the magnetic vector potential Az predicted by the neural network satisfies the governing equations (such as the Helmholtz equation in the frequency domain) within the computational domain. Its goal is to make the PDE residuals approach zero at all sampling points across the entire domain.

[0053] Set of sampling points (interior points): ,in ; Governing equations: ; Interior point MSE: ; All first / second derivatives are obtained by automatic differentiation.

[0054] (2) Boundary condition loss term This ensures that the solution satisfies the defined physical conditions at the boundaries of the computational domain. A unified boundary operator B allows for the flexible representation of various boundary condition types, such as Dirichlet (fixed potential), Neumann (fixed normal derivative, e.g., magnetic insulation), and Robin (mixed conditions) for simulating open boundaries. The "boundary operator" uses… This indicates that the losses are itemized and added separately.

[0055] Set of sampling points (boundary points): ; The three most commonly used: a. Outer field truncation (Dirichlet, default): b. Symmetrical / Magnetic Insulation (Neumann): c. Far-field absorption (Robin): ( (Given by extensional mapping or experience) After merging, the total boundary condition loss term is: .

[0056] (3) Loss of continuity at the material interface At the interfaces between different materials in a magnetic coupler (such as the interface between a coil and ferrite, or between ferrite and air), the electromagnetic field solution must satisfy the continuity condition. This loss is achieved by applying constraints to the sampling points on both sides of the interface, forcing the magnetic vector potential Az and its normal derivative (related to the magnetic flux density) to be continuous.

[0057] Sampling point set (paired points on the interface): (+ and - are located on opposite sides of the interface, and the normal unit vector n points to the "+" side); On each interface The above requires continuity of potential and continuity of normal magnetic flux: , in Used to accommodate interface sources (such as surface currents) (or remanent tangential component), without interface source .

[0058] Paired-point MSE: (4) Calibration loss of finite element simulation (FEM) data This is a data-driven constraint that "calibrates" the prediction results of the PINN model using a small amount of high-fidelity FEM simulation data. By introducing a learnable scaling factor 'a', the PINN output is linearly scaled to match the absolute magnitude of the FEM data. This effectively compensates for potential overall biases in purely physics-driven models, significantly improving prediction accuracy and reliability.

[0059] Sampling point set (FEM calibration): ; The raw quantities obtained from post-processing of the network field are represented by a small number of FEM points. Align to absolute magnitude. Use linear scaling (no intercept by default): .

[0060] MSE: It can be replaced by relative error MSE ( (etc.) or Huber to enhance robustness; intercepts can also be added. Perform affine calibration.

[0061] In summary, parameterized PINN Approximation solution ( (Network weights). The initial default is: ; (Avoid overly strong boundaries); (When considering interface continuity, the smaller weight is used). (Perform a light calibration).

[0062] Specifically, after obtaining the loss function, this approach selects the Adam optimizer suitable for this model for training. The set of parameters to be optimized is trained 2500 times. During the optimization process, the loss function value gradually decreases, while the weight matrix and bias vector of the neural network are updated. When the loss function value decreases to a minimum, the PINN framework satisfies both the partial differential equation of the magnetic vector potential and the given boundary conditions, accurately simulating the electromagnetic field process of the magnetic coupler. The training result of the PINN model is considered the solution to the magnetic vector potential of the partial differential equation system, from which the self-inductance L and mutual inductance M can be calculated.

[0063] An adaptive moment estimation (Adam) optimizer is employed, which combines the adaptive learning rate characteristics of first-order and second-order momentum, making it suitable for efficient convergence in non-convex optimization problems. The initial learning rate η = 0.001 (which can be dynamically adjusted based on loss reduction), the number of training iterations: T = 2500 (for a set of structural parameters to be optimized), and the batch typically covers all sampling points (PDE points, boundary points, interface points, and FEM calibration points).

[0064] In each training iteration (t = 1 to T), the following steps are performed: Forward propagation: Input the current structural parameters p and the spatial sampling points (x, y, z) into the PINN network; calculate the predicted magnetic vector potential Az and the derived electrical parameters.

[0065] Then, the loss function is used to calculate and comprehensively evaluate the four losses; Backpropagation and parameter update: Calculating the gradient of the total loss with respect to the network weights θ using automatic differentiation. The network weights and bias parameters are updated according to the Adam algorithm.

[0066] During training, the loss value needs to be monitored in real time, and training should be terminated if any of the following conditions are met: Loss convergence threshold: (like = 1e-6); Gradient convergence threshold: (For example = 1e-5); Maximum number of iterations: up to the preset T = 2500 iterations.

[0067] Once the model converges, the electromagnetic field is solved using the trained PINN network: Magnetic vector potential analysis: Inputting global sampling points into the trained network yields a magnetic vector potential distribution A(x, y,z) that satisfies the PDE and boundary conditions; Inductance parameter calculation: self-inductance =2Wm / I² (where Wm is the magnetic field energy, calculated by integrating Az with the magnetic field strength H), mutual inductance = Ψ secondary / I primary (where Ψ is the magnetic flux, obtained by integrating Az over the coil region), and the coupling coefficient k is calculated based on the electromagnetic correlation formula.

[0068] Step 3: Construct a geometric parameter optimization loss function, which includes a relative error term based on the target electrical parameters and the predicted electrical parameters, and a penalty function term based on geometric constraints.

[0069] An optimizable function is established that connects electrical performance indicators with geometric parameters. By constructing a composite loss function that includes performance errors and geometric constraints, the magnetic coupler design problem is transformed into a mathematical optimization problem.

[0070] Based on the given target (the target value obtained from circuit simulation, the self-inductance value of the magnetic coupling mechanism) Mutual inductance value Coupling coefficient ) and the predicted value output from step 2 (self-induction) Mutual induction The coupling coefficient k) is used to construct a relative error objective function, where the objective can be one or more of the following: The relative error measures the deviation between the predicted electrical parameter values ​​and their respective target electrical parameter values. These electrical parameters include self-inductance, mutual inductance, and the coupling coefficient. The function is as follows: in For each weight, The penalty function is a geometric constraint. The penalty function method is a classic approach for solving constrained optimization problems by transforming the constrained problem into an unconstrained one. The penalty function term is used to constrain the structural parameters to satisfy physical realizability conditions.

[0071] The specific constraint in this invention employs a smooth penalty function, which is a smooth function based on the exact penalty function, exhibiting better convergence and stability. Geometric constraint penalty function. It mainly consists of 7 parts, and the expression is as follows: These seven parts are: (1) Turn spacing constraint To ensure that the turn spacing s is not less than the wire diameter w, and to prevent short circuits, the expression is: .

[0072] (2) Coil size envelope constraint ( and The expression for limiting the coil's external dimensions to not exceed the installation space is: (3) Material layer size constraints and ), indicating that the ferrite layer and the aluminum plate layer have a relative margin. , The outward space under the given conditions is expressed as: (4) Working interval constraints : Indicates the spacing between the primary and secondary coils, limiting the working gap to a mechanically feasible range. The expression is: (5) Turns number discrete constraint : Indicates that the number of turns is continuously relaxed in half-turns, and is added as a regularization term during training, a slack variable. The expression is: This example selects the softplus function, which is suitable for this model, as the activation function. Its expression is: Softplus is a smooth approximation of the ReLU activation function, which solves the problem that ReLU is not differentiable at x=0. It has the advantages of being differentiable at all real points and having a continuous derivative, thus alleviating the gradient vanishing problem (better than Sigmoid and Tanh).

[0073] This process relies on the output of the PINN model trained in step 2. The constructed objective function provides the object for gradient calculation in step 4.

[0074] Step 4: Using automatic differentiation, calculate the gradient of the geometric parameter optimization loss function with respect to the structural parameters.

[0075] In scalar objective function After construction, the geometric parameter optimization loss function, i.e., the objective function, is calculated using automatic differentiation and the chain rule with respect to the structural parameter p. The gradient indicates the direction and magnitude in which the geometric parameters should be adjusted to minimize the loss function. Based on the parameterized PINN model, where p is the network input / mask generator input, the second term is obtained by flux integration and automatic differentiation; the third term is obtained by backpropagation of the framework. The core formula for gradient calculation is as follows: in: (1) Performance sensitivity item This parameter characterizes the sensitivity of the final design objective to the predicted electrical parameters. The specific calculation is based on the objective function form defined in step 3. (2) Electromagnetic field-electrical parameter conversion item This section establishes a mathematical relationship between electromagnetic field distribution and macroscopic electrical parameters. Self-induction gradient calculation: The magnetic field energy Wm is expressed by the quadratic form of the magnetic vector potential A; Calculate the gradient of mutual inductance: The magnetic flux linkage Ψ is calculated using the area integral ∮Adl; Coupling coefficient gradient: This can be transformed into a combination of gradients of basic parameters using the quotient rule.

[0076] (3) Geometric-electromagnetic field mapping term This is the core capability of the PINN model: establishing the sensitivity of geometric parameter changes to electromagnetic field distribution. The geometric parameter p is used as an additional input feature of the neural network, propagating forward together with the spatial coordinates (x, y, z). An automatic differentiation framework records the complete computational graph from the input parameters to the output field quantity A, enabling gradient calculation of p. This process fully utilizes the differentiability of neural networks, avoiding the computationally intensive problems of traditional numerical differentiation.

[0077] (4) Geometric constraint gradient term This ensures that the optimization process satisfies all physical realizability constraints. Based on the smoothing properties of the Softplus function, the gradient of the constraint penalty function can be analytically calculated. Each constraint sub-item generates a corresponding gradient component, which together guides the optimization direction toward the feasible region.

[0078] Modern deep learning frameworks (such as PyTorch) provide technical support for this step. During forward propagation, the framework automatically constructs a complete computation graph from input p to output J, recording all intermediate variables and computational relationships. During backpropagation, the framework's `backward()` function is called to automatically compute the gradient of each node along the computation graph, starting from the final scalar loss J. Finally, gradient aggregation automatically aggregates the gradient products along the path to obtain the total gradient of the loss function with respect to the original geometric parameters p. Gradient checking, in actual implementation, can be compared and verified with the numerical differentiation results to ensure the correctness of the automatic differentiation process.

[0079] Step 5: Based on the gradient, use an optimization algorithm to iteratively update the structural parameters.

[0080] The goal of this step is to automatically adjust the geometric structure parameter p using a gradient optimization algorithm to minimize the loss function J(p) defined in step 3. A suitable non-convex optimization algorithm, such as the Adaptive Moments Estimation (ADAM) algorithm or the Finite Memory Boundary Constraint Quasi-Newton (L-BFGS-B) algorithm, is selected to iterate over the geometric structure parameter p with a projection operator, directly and continuously updating it. The ADAM algorithm is suitable for the initial search phase, exhibits good adaptability to noisy gradients, and can quickly locate potential optimal regions. The L-BFGS-B algorithm has a superlinear convergence speed when approaching the optimal solution, making it suitable for fine-grained optimization. The optimization process is described below.

[0081] (1) ADAM algorithm, which achieves adaptive parameter update by calculating the first and second moments of the gradient. The parameter update rule is as follows: Specific implementation process: Gradient calculation: In the t-th iteration, calculate the gradient of the loss function at the current parameters. Moment estimation update: First moment: Second moment: Bias correction: Parameter update and projection: Project the updated parameters into the feasible region B.

[0082] Hyperparameter settings: The learning rate (t) adopts an exponential decay strategy, with an initial value of 0.001; the moment estimation decay factors β1=0.9, β2=0.999; and the numerical stability constant... =10-8.

[0083] (2) L-BFGS-B algorithm: This algorithm achieves more efficient parameter updates by constructing an inverse Hessian matrix approximation. The parameter update rules are as follows: Gradient calculation: Obtain the gradient at the current point Inverse Hessian approximation: Constructed using the finite-memory BFGS method Only save the curvature information of the most recent m times. Line search: Use strong Wolfe conditions to determine the step size η(t) to ensure sufficient descent conditions. Projection operation: via Ensure that the updated parameters satisfy all geometric constraints.

[0084] in, It's a projection operator, which constrains the updated parameters within a preset feasible region. It projects the variables onto the box constraint. Perform hard-boundary clipping (e.g.) (e.g., upper / lower boundaries of dimensions). Every few steps, slack variables are... Forcefully round to the nearest half-integer and write back to the material field. Effective.

[0085] In the initial exploration phase, the ADAM optimizer is used due to its strong global search capability. When the solution is close to the optimum, the algorithm switches to the L-BFGS-B optimizer to improve convergence accuracy. This hybrid strategy balances convergence speed and solution quality. Step 6: Determine whether the optimization process has converged. If it has converged, output the optimal structure parameters.

[0086] The convergence criterion for determining whether a geometry is optimal is: Relative changes in function values ; Gradient norm condition: ; Reaching the maximum number of iterations If set to 2500 Meeting any one of the above conditions is sufficient. During training, output a set of results each time the convergence condition is met. This can be saved as a set of optimal solutions.

[0087] Then, all local optima that satisfy the convergence condition are automatically saved. Record the complete optimization trajectory Save the optimizer state for easy hot restart analysis.

[0088] In summary, this method constructs a PINN model by making the coil's geometric parameters (such as the number of turns, turn spacing, and working gap) differentiable and embedding electromagnetic physical laws as constraints. Figure 7 The model is trained using a small amount of finite element method (FEM) simulation data to achieve rapid prediction of electrical parameters such as self-inductance and mutual inductance of magnetic couplers. Utilizing the differentiability of the model, structural parameters are optimized through gradient backpropagation, thereby addressing the problems of high computational cost, long design cycle, and limited optimization accuracy in existing technologies.

[0089] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for optimizing the magnetic coupling mechanism of a wireless power transmission device.

[0090] This embodiment also provides an electronic device, including a memory and a processor. The memory is used to store computer programs, and the processor runs the computer programs to enable the electronic device to execute the magnetic coupling mechanism optimization design method of a wireless power transmission device.

[0091] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for optimizing the magnetic coupling mechanism of a wireless power transfer device, characterized in that, Includes the following steps: S1: Determine the initialization parameters of the magnetic coupler and generate the corresponding vectorized structure parameters; S2: Construct a physical information neural network model, taking the parameter vector and spatial coordinates as input, outputting the magnetic vector potential distribution of the magnetic coupler, and training the model to obtain a surrogate model capable of predicting the electrical parameters of the magnetic coupler; including: The three-dimensional model of the magnetic coupler is reduced to a 2.5-dimensional model, and multiple material subdomains and corresponding boundary conditions are defined. With 5 parameters to be optimized The parameters are as follows: number of turns of the Litz wire, turn spacing, working gap, ferrite length and width, spatial variables x, y, z are the input layer of the physical information neural network, with 5 hidden layers and 128 neurons in each layer, and the output coil self-inductance. Mutual induction This is the output layer; the output of each layer of the neural network is shown in the following formula: ； In the formula, , The first Layer weight matrix and bias vector, For the first Layer output, As an activation function, each layer in a neural network has weights. and bias Correspondingly, the Tanh function is chosen as the activation function, and its expression is: ; Then, lightweight FEM data was used for calibration, as follows: a. Each set of data is treated as one input: ， ； in It contains complete geometric and material parameters; b. Use PINN to solve for the structural parameters of these K groups under unit current to obtain the original inductance; for each group The forward frequency domain solution A is calculated using the following formula: in For magnetic field energy, For the magnetic flux passing through the receiving coil window, both are obtained by automatic differentiation and numerical integration from the PINN output A or B; c. Calibrate the mapping using least squares fitting: Where 'a' is the learnable proportional coefficient; ； The analytical solution is: ， ； d. will Saved together with PINN weights; Output and Two weight values ​​are used to correct and calibrate the values ​​predicted by the PINN model, which are then used as a term in the loss function of the PINN model. Then, the total loss function of the PINN model is calculated. Based on the vector magnetic potential method, the partial differential equations of magnetic potentials and boundary conditions for each material subdomain are derived and embedded as physical constraints into the loss function of the model, dividing the computational domain into seven material subdomains: The sets are formed by the intersection of each pair of interfaces. ; The boundary between the r-th and s-th material layers; outer boundary The dependent variable is the complex magnetic vector potential A. Material properties are defined as follows: , Electrical conductivity; For the remanence spectrum, apply a volume current density. Only the region containing the Lidz line is not zero; electromagnetic parameters are expressed as follows: , Quasi-static, ; In addition to obtaining the scaling coefficients through fitting learning, the introduced finite metadata is also weighted and incorporated into the total loss function for calculating the magnetic vector potential. All four terms in the total loss function are obtained using the mean square error method. The expression for the total loss function is: ； in, Let be the parameter to be optimized, and p be the structural parameter. , , , , This is the residual loss term in the partial differential equation (PDE). For boundary condition loss terms, For the loss of material interface continuity, Calibration loss for finite element simulation (FEM) data S3: Construct a geometric parameter optimization loss function, which includes a relative error term based on the target electrical parameters and the predicted electrical parameters, and a penalty function term based on geometric constraints. The relative error term measures the deviation between the predicted electrical parameter values ​​and their respective target electrical parameter values. The electrical parameters include self-inductance, mutual inductance, and coupling coefficient. The penalty function term is implemented using a smoothing penalty function, including turn spacing constraints, coil size envelope constraints, material layer size constraints, working gap constraints, and turn number discrete constraints. S4: Using automatic differentiation techniques, calculate the gradient of the geometric parameter optimization loss function with respect to the structural parameters; in the scalar objective function After construction, the gradient of the geometric parameter optimization loss function (i.e., the objective function) with respect to the structural parameter p is calculated using automatic differentiation and the chain rule. ; S5: Based on the gradient, the structural parameters are iteratively updated using an optimization algorithm; S6: Determine whether the optimization process has converged. If it has converged, output the optimal structure parameters.

2. The method for optimizing the magnetic coupling mechanism of a wireless power transfer device according to claim 1, characterized in that, In S1: The initialization parameters include three categories: coil geometric parameters, magnetic material parameters, and shielding layer parameters; The structural parameters are an ordered vector formed by combining initialization parameters in a predetermined order.

3. The method for optimizing the magnetic coupling mechanism of a wireless power transfer device according to claim 2, characterized in that: The total loss function of the physical information neural network model consists of a weighted sum of multiple losses, including partial differential equation residual loss, boundary condition loss, material interface continuity loss, and finite element simulation data calibration loss.

4. The method for optimizing the magnetic coupling mechanism of a wireless power transfer device according to claim 1, characterized in that, The gradient calculation described in S4 is based on the chain rule and uses automatic differentiation to calculate the sum of the partial derivatives of the loss function with respect to the structural parameters.

5. The method for optimizing the magnetic coupling mechanism of a wireless power transfer device according to claim 1, characterized in that, In S5: The optimization algorithms include the Adaptive Moment Estimation Optimization Algorithm ADAM and the Finite Memory Boundary Constraint Quasi-Newton Algorithm L-BFGS-B. The iterative update constrains the updated parameters within a preset feasible region using a projection operator.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the magnetic coupling mechanism optimization design method of a wireless power transmission device according to any one of claims 1-5.

7. An electronic device, characterized in that, It includes a memory and a processor, the memory being used to store computer programs, and the processor running the computer programs to cause the electronic device to perform the magnetic coupling mechanism optimization design method of any one of claims 1-5.

Images