A method and system for constructing a cable scattering parameter prediction model

Abstract

The application belongs to the technical field of electromagnetic simulation, and relates to a construction method and system of a cable scattering parameter prediction model. The construction method comprises the following steps: S1: constructing training data; S2: establishing a cable scattering parameter prediction model, which comprises a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer is used to optimize and fuse a multi-scale transformation through a learnable scale to obtain a high-dimensional length vector and an electric length product vector. The enhanced PINN main network takes the layout distance of the cable, the frequency, the receiving port, the excitation port, the length vector and the electric length product vector as inputs to predict the real part and the imaginary part of the cable scattering parameter; and S3: training iteration is performed to obtain the final cable scattering parameter prediction model. The cable scattering parameter prediction model with the advantages of precision, efficiency and physical consistency is constructed by combining the multi-scale Fourier feature enhancement technology and the physical constraint neural network and combining the residual feature multiplexing mechanism.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic simulation technology, and relates to a method and system for constructing a cable scattering parameter prediction model, particularly a method and system for constructing a cable scattering parameter prediction model based on a multi-scale Fourier physical information neural network. Background Technology

[0002] Electromagnetic simulation is a core and indispensable R&D tool in the electromagnetic compatibility and signal integrity design of electronic devices, automotive wiring harnesses, and aerospace cable interconnect systems. The industry's standard design process is as follows: engineers use 3D CAD software to complete the geometric design of cable length, spacing, port locations, etc., and then import the geometric model into professional electromagnetic simulation software such as HFSS, CST, and FEKO to perform full-wave electromagnetic solutions. Finally, they obtain key electromagnetic performance indicators such as S-parameters (i.e., scattering parameters) and characteristic impedance to evaluate the electromagnetic performance of the cable system.

[0003] However, traditional full-wave electromagnetic simulation technology and existing prediction model schemes have many inherent defects, the specific technical status of which is as follows: 1. Full-wave numerical simulation is extremely time-consuming. Traditional full-wave simulation based on the finite element method (FEM) and the method of moments (MoM) has the advantage of high accuracy, but whenever geometric parameters such as cable length, laying spacing, and port position are adjusted, the mesh generation and iterative solution of the equation system must be completed again. The time for a single simulation can range from several minutes to several hours, resulting in extremely low simulation efficiency.

[0004] 2. Parametric scanning involves enormous computational load. In engineering iterative optimization design scenarios, it is necessary to traverse and scan multiple parameters such as cable length, laying spacing, and operating frequency. The combination and iteration of multiple parameters will lead to an exponential increase in the amount of simulation computation, which cannot meet the needs of batch parameter optimization.

[0005] 3. Performance limitations of traditional prediction models. Currently, the industry has gradually applied prediction modeling techniques such as intrinsic orthogonal decomposition (POD), Kriging interpolation, and traditional fully connected neural networks to cable S-parameter prediction, replacing some full-wave simulation work. However, these methods are based solely on pure data-driven modeling and do not explicitly incorporate the physical laws of electromagnetic fields. They have weak fitting capabilities for high-frequency oscillation characteristics in electromagnetic responses, and their model generalization and prediction accuracy are insufficient to meet the demands of high-precision electromagnetic simulation.

[0006] Based on the above technological status quo, existing cable electromagnetic simulation and prediction model technologies suffer from the following four core defects, which severely restrict the improvement of the efficiency and accuracy of cable electromagnetic simulation: (1) Low simulation efficiency. Traditional full-wave simulation requires repeated mesh generation and solving of Maxwell's equations, which is complex and time-consuming. In engineering scenarios with massive parameter iterative optimization, the computational cost is high, which seriously prolongs the R&D cycle and cannot support rapid design iteration.

[0007] (2) Poor physical interpretability of the prediction model. Conventional data-driven prediction models do not incorporate electromagnetic field physical constraints and structured information of field equations, and cannot meet basic physical criteria such as passivity and causality. The prediction results lack physical support and have low reliability and credibility.

[0008] (3) Insufficient high-frequency electromagnetic feature learning ability. Cable S-parameters have complex high-frequency characteristics such as multi-scale oscillations and resonant spikes. Existing neural networks use fixed activation functions, which are difficult to adaptively capture refined high-frequency electromagnetic features, resulting in significantly larger prediction errors in the high-frequency band of the model.

[0009] (4) Weak parameter generalization and robustness. The existing prediction model has a single training sample working condition and parameter range, and has poor adaptability to the coupling changes of multiple parameters such as cable structure, medium, and laying conditions. It is prone to deviation and fitting failure in cross-working condition prediction. At the same time, the model has weak anti-interference ability, is sensitive to simulation noise and sample disturbance, and its engineering applicability is limited.

[0010] Therefore, given the aforementioned technical deficiencies in existing technologies, there is an urgent need for a novel method and system for constructing cable scattering parameter prediction models. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention proposes a method and system for constructing a cable scattering parameter prediction model. It integrates multi-scale Fourier feature enhancement technology with a physically constrained neural network and combines a residual feature reuse mechanism to construct a cable scattering parameter prediction model that balances accuracy, efficiency, and physical consistency, enabling rapid and accurate prediction of cable S-parameters.

[0012] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a cable scattering parameter prediction model, characterized by comprising the following steps: S1: Collect the length, spacing, frequency, receiving port, excitation port, electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths to construct training data; S2: Establish a cable scattering parameter prediction model, which includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer is used to map the low-dimensional length and electrical length product to a high-dimensional feature space through learnable scale optimization and multi-scale transformation fusion, so as to obtain a high-dimensional length vector and electrical length product vector. The enhanced PINN main network takes the cable layout spacing, frequency, receiving port, excitation port, length vector and electrical length product vector as input to predict the real and imaginary parts of the cable scattering parameters. The enhanced PINN main network adopts a residual-enhanced main network structure and has 4 sets of pre-activated style residual blocks with completely consistent structure and processing flow. The 4 sets of pre-activated style residual blocks are serially cascaded to pass features. Each set of pre-activated style residual blocks sequentially completes linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation and layer normalization. Then, the input features and output features of the block are added element-wise through the residual shortcut to achieve feature reuse and gradient optimization. S3: The cable scattering parameter prediction model is trained and iterated using the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

[0013] Preferably, in step S1, the product of the acquisition length, deployment spacing, frequency, and electric length, as well as the real and imaginary parts of the scattering parameters, are also normalized.

[0014] Preferably, the multi-scale Fourier feature embedding layer includes: The input scaling unit is used to independently scale the product of length and electrical length based on a trainable scaling vector to obtain the scaled product of length and electrical length. A multi-scale projection unit is used to map the scaled product of length and electrical length to feature subspaces of different scales based on a set of parallel trainable projection matrices to obtain projection results of different scales, wherein each of the trainable projection matrices corresponds to a preset scale factor. The nonlinear transformation and splicing unit is used to apply sine and cosine transformations to the projection results at each scale and splice the outputs of the sine and cosine transformations of the projection results at all scales along the feature dimension to form the final length vector and electrical length product vector.

[0015] Preferably, the enhanced PINN main network includes: The feature stitching layer is used to stitch together the product vectors of the deployment spacing, frequency, receiving port, excitation port, length vector and electrical length to obtain the stitched input features. The first fully connected layer is used to reduce the dimensions of the concatenated input features, thereby reducing the 516-dimensional input features to 256-dimensional features. The LayerNorm layer is used to perform layer normalization on 256-dimensional features to eliminate input distribution bias. The SiLU activation function is used to introduce smooth nonlinearity into the 256-dimensional features after layer normalization to fit the complex nonlinear mapping of cable scattering parameters, thereby obtaining the feature vector after nonlinear transformation. The four pre-activated style residual blocks, which transmit features in a serial cascade manner, are used to perform depth extraction and fine-tuning transformation of the feature vector after the nonlinear transformation block by block, so as to obtain 256-dimensional depth features. The first contraction layer is used to compress the 256-dimensional depth features into 128-dimensional depth features; The second shrinking layer is used to compress the 128-dimensional depth features into 64-dimensional depth features. An output mapping layer is used to map the 64-dimensional depth features into 2-dimensional results, which correspond to the real and imaginary parts of the cable scattering parameters, respectively.

[0016] Preferably, in step S3, before training, the training data is grouped and encoded in three levels based on the cable's geometric parameters, excitation port, and frequency, thereby assigning a unique physical constraint group identifier to each training data. During the training iteration, a group batch sampling mechanism is used to generate training batches, and the training data in the same training batch all come from several frequencies under the same geometric structure and the same excitation port.

[0017] Preferably, in step S3, the loss function for:

[0018] in, This represents the data fitting loss, which is the mean square error between the predicted values ​​of the training data and the true scattering parameters. This represents the physical loss due to energy conservation. The weight for physical loss due to energy conservation.

[0019] Preferably, in step S3,

[0020] in, This represents the total power calculated from the true scattering parameters of training data at the same frequency within the same training batch; This represents the total power calculated from the predicted values ​​of training data at the same frequency within the same training batch; This represents the total number of frequencies within the same training batch.

[0021] Furthermore, the present invention also provides a system for constructing a cable scattering parameter prediction model, characterized in that it includes: The training data construction module is used to collect the length, spacing, frequency, receiving port, excitation port, electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths in order to construct training data. A cable scattering parameter prediction model building module is used to establish a cable scattering parameter prediction model. This model includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer is used to map the low-dimensional length and electrical length product to a high-dimensional feature space through learnable scale optimization and multi-scale transformation fusion, resulting in a high-dimensional length vector and electrical length product vector. The enhanced PINN main network uses cable layout spacing, frequency, receiving port, excitation port, length vector, and electrical length product vector as parameters. The input is the quantity, which predicts the real and imaginary parts of the cable scattering parameters. The enhanced PINN main network adopts a residual enhancement main network structure and has 4 sets of pre-activated style residual blocks with completely identical structures and processing flows. The 4 sets of pre-activated style residual blocks are used to transmit features in a serial cascade manner. Each set of pre-activated style residual blocks sequentially performs linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation and layer normalization. Then, the input features and output features of the block are added element by element through the residual shortcut to achieve feature reuse and gradient optimization. The training iteration module is used to train and iterate the cable scattering parameter prediction model with the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

[0022] Furthermore, the present invention also provides a device for constructing a cable scattering parameter prediction model, characterized in that it includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing the cable scattering parameter prediction model as described above. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the steps of the method for constructing a cable scattering parameter prediction model as described above.

[0023] Compared with existing technologies, the method and system for constructing cable scattering parameter prediction models of the present invention are based on multi-scale Fourier physical information neural networks, and their main advantages are as follows: (1) Extreme simulation acceleration effect. After the prediction model is trained, the time for a single S-parameter prediction in the GPU environment is less than 10 milliseconds. Compared with the time of several minutes to tens of minutes of traditional full-wave simulation, the simulation speedup can reach more than 1,000 times, which can support large-scale parametric scanning and rapid iterative design.

[0024] (2) High-precision multi-scale feature fitting capability. Through a learnable multi-scale Fourier embedding structure, the prediction model can adaptively learn the optimal frequency coding scale and accurately capture detailed features such as peaks and resonance points of S-parameters. Experimental results show that the feature fitting accuracy is improved by 30% to 50% compared with traditional models.

[0025] (3) Excellent generalization and adaptation performance. For cable length and spacing parameter combinations that were not involved in model training, the prediction mean square error (MSE) of the prediction model of the present invention is reduced by about 40% compared with ordinary neural networks, which has a stronger adaptability to unknown parameter conditions and higher engineering practicality.

[0026] (4) Strict physical constraints guarantee. This invention introduces a power conservation physical constraint loss term, which makes the model prediction results strictly follow the principle of electromagnetic field energy conservation, and the passivity violation rate of S-parameters reaches more than 99%, which greatly improves the prediction stability and reliability of the prediction model under extreme parameter conditions.

[0027] (5) Good CAE software compatibility. The standardized input and output normalization rules and model weight saving format are designed, which can be directly embedded into mainstream CAE simulation automation scripts and platforms to achieve seamless integration of the simulation process without additional modification or adaptation. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method for constructing the cable scattering parameter prediction model of the present invention; Figure 2 This is a schematic diagram of the multi-scale Fourier feature embedding layer of the present invention; Figure 3 This is a schematic diagram of the enhanced PINN master network of the present invention; Figure 4 This is a schematic diagram of the pre-activated style residual block of the present invention; Figure 5 This is a schematic diagram of the system structure of the cable scattering parameter prediction model construction method of the present invention; Figure 6 This is a structural block diagram of the device for constructing the cable scattering parameter prediction model of the present invention. Detailed Implementation

[0029] Before detailing any embodiment of the invention, it should be understood that the invention, in its application, is not limited to the details of the construction and arrangement of the components set forth in the following description or illustrated in the following figures. The invention can have other embodiments and can be practiced or carried out in various ways. Furthermore, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered limiting. The use of “comprising” or “having” and variations thereof in this invention is intended to cover the items set forth below and their equivalents, as well as any additional items. Unless otherwise specified or limited, the terms “installation,” “connection,” “support,” and “linkage,” and variations thereof are used broadly and cover both direct and indirect installation, connection, support, and linking. Moreover, “connection” and “linkage” are not limited to physical or mechanical connections or links. Furthermore, firstly, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting the invention. Secondly, the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple. The term "a" should not be construed as a limitation on the quantity.

[0030] This invention addresses the numerous shortcomings of existing cable electromagnetic simulation technologies by proposing a method and system for constructing a cable scattering parameter prediction model. This method and system is based on a multi-scale Fourier physical information neural network (PINN) for constructing cable scattering parameter prediction models. The aim is to overcome the technical bottlenecks of low simulation efficiency, poor accuracy of conventional prediction models, lack of physical consistency, and insufficient versatility. The core improvement objectives are as follows: (1) To achieve automatic encoding of electromagnetic physics knowledge, adaptively integrating the core electromagnetic features of the electric length product of the frequency-length product into the neural network, eliminating reliance on manual feature engineering and lowering the technical application threshold; (2) To introduce a multi-scale Fourier feature embedding mechanism to enhance the model's response to S-parameters. The ability to learn multi-scale features such as high-frequency oscillation and resonance peaks greatly improves the prediction accuracy of the entire frequency band, especially the high-frequency band; (3) Design a targeted physical constraint loss function to force the S-parameters predicted by the model to strictly satisfy the basic physical laws of electromagnetic fields such as power conservation, and ensure the physical validity and stability of the prediction results; (4) Construct an end-to-end trainable prediction model, which can realize the rapid and accurate mapping and generalization reasoning of cable geometric parameters, working frequency to complex S-parameters, and efficiently replace the traditional time-consuming full-wave electromagnetic simulation calculation; (5) Establish a standardized and automated model construction, training and deployment process, which can be seamlessly integrated into mainstream CAE simulation software, and comprehensively improve the efficiency of the entire electromagnetic simulation process.

[0031] Figure 1 A flowchart illustrating the method for constructing the cable scattering parameter prediction model of the present invention is shown. Figure 1 As shown, the method for constructing the cable scattering parameter prediction model of the present invention includes the following steps: S1: Construction of training data.

[0032] Training data is constructed by collecting the length, spacing, frequency, receiving port (i.e., the receiving port number), excitation port (i.e., the excitation port number), electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths.

[0033] Specifically, relying on existing mainstream CAE electromagnetic simulation software, the length L, spacing D, and port combinations (receiving ports) of different cables are collected in batches. ref Excitation Port inc The electromagnetic data corresponding to the operating frequency f is collected, with the core tags being the real part R and the imaginary part I of the S-parameters (scattering parameters). Therefore, each training data point has six input features: frequency f, length L, deployment spacing D, and receiving port Port. ref Excitation Port incAnd the automatically constructed electric length product f×L; the output features include the real part R and the imaginary part I of the S-parameters (scattering parameters). Simultaneously, the power and sum of all S-parameters at the same excitation port in each training data set are statistically analyzed. As training labels for physical constraints, they provide a basis for subsequent physical loss calculations.

[0034] Furthermore, in this invention, the StandardScaler normalization tool can be used to normalize the training data to eliminate the impact of dimensional differences on model training. Specifically, for each set of training data, four input features are calculated based on each set of training data: frequency f, cable length L, installation spacing D, and electrical length product f×L. Their respective means and standard deviation And the two output features of the S-parameters: the real part R and the imaginary part I. Their respective means and standard deviation Based on their respective means and standard deviation For the four input features Perform normalization calculations separately And based on their respective means and standard deviation For the two output features Perform normalization calculations separately This allows for the standardization and unification of input and output features.

[0035] S2: Establishment of a prediction model for cable scattering parameters.

[0036] A cable scattering parameter prediction model is established, which includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network.

[0037] 1. Multi-scale Fourier feature embedding layer.

[0038] Traditional neural networks struggle to accurately fit the complex frequency dependence of high-frequency electromagnetic S-parameters due to their strong periodicity and multi-scale oscillation characteristics. This invention constructs a multi-scale Fourier feature embedding layer, which maps low-dimensional physical input features to a high-dimensional feature space through learnable scale optimization and multi-scale transformation fusion, thereby improving the prediction model's ability to model periodic electromagnetic responses.

[0039] According to transmission line theory, the periodicity of the S-parameters is determined by the relative relationship between the signal wavelength and the cable's geometric dimensions. Therefore, in this invention, the two core physical features directly related to periodicity—the product of length and electrical length—are selected as the input features of the multi-scale Fourier feature embedding layer. Discrete classification features such as port numbers (i.e., the receiving port Portref and the excitation port Portinc) are retained in their original form and directly participate in subsequent calculations. This preserves the physical priors while avoiding interference from irrelevant features.

[0040] Figure 2 A schematic diagram of the multi-scale Fourier feature embedding layer of the present invention is shown. For example... Figure 2 As shown, the multi-scale Fourier feature embedding layer includes: (1) Input scaling unit.

[0041] The input scaling unit is used to independently scale the product of length and electrical length based on a trainable scaling vector to obtain the scaled product of length and electrical length.

[0042] Specifically, let the input features of the multi-scale Fourier feature embedding layer be... ,in, For the selected physical feature dimension (e.g., in this invention, (corresponding to the two physical feature dimensions: the product of the length and the electric length).

[0043] The input scaling unit introduces learnable parameters, which are logarithmic scaling vectors. And obtain the scaling vector through exponential operation. : .

[0044] Input features With scaling vector Perform element-wise multiplication (broadcasting mechanism) to obtain scaled input features. : .in, This represents the Hadamard product (element-wise multiplication). When learnable parameters... The initial scaling vector when initialized to zero. It is a vector of all 1s, meaning no additional scaling is introduced.

[0045] (2) Multiscale projection unit.

[0046] The multi-scale projection unit is used to map the scaled length and electrical length product to feature subspaces of different scales based on a set of parallel trainable projection matrices to obtain projection results of different scales, wherein each trainable projection matrix corresponds to a preset scale factor.

[0047] Specifically, if the preset number of scales is ,in, In this invention, one can take ;No. The scaling factor for each scale is ,in, , The expected total embedding dimension is ,in, Even number and requires Can be Integer division; sub-embedding dimension for each scale is ,in, .

[0048] For each scale Construct a trainable projection matrix The trainable projection matrix elements Initialization starts with a mean of 0 and a standard deviation of... Random sampling from a Gaussian distribution:

[0049] Then, based on this trainable projection matrix Calculate the first Projection results at each scale :

[0050] In subsequent training iterations of the prediction model, during forward propagation, the aforementioned learnable parameters... and trainable projection matrix It participates in matrix multiplication, element-wise multiplication, and other operations, forming a complete differentiable path from input to output and updating it automatically.

[0051] (3) Nonlinear transformation and splicing unit.

[0052] The nonlinear transformation and splicing unit is used to apply sine and cosine transformations to the projection results at each scale, and splice the outputs of the sine and cosine transformations of the projection results at all scales along the feature dimension to form the final length vector and electrical length product vector.

[0053] Specifically, the projection results at each scale Apply sine and cosine transforms respectively, and concatenate them along the feature dimension to obtain Fourier feature vectors at each scale. :

[0054] All Fourier eigenvectors at each scale Concatenate along the feature dimensions to obtain the final embedding vector. That is, the final product vector of the length vector and the electric length:

[0055] The embedding vector The total dimension is .For example At that time, the output embedding vector The dimension is 512.

[0056] 2. Enhanced PINN main network.

[0057] The enhanced PINN main network takes the cable layout spacing, frequency, receiving port and excitation port, as well as the length vector and electrical length product vector output by the multi-scale Fourier feature embedding layer as inputs to predict the real and imaginary parts of the cable scattering parameters. That is, the output is the predicted real part R and imaginary part I of the cable S-parameters.

[0058] In this invention, the enhanced PINN main network adopts a residual-enhanced main network structure, which has four sets of pre-activated style residual blocks with completely identical structures and processing flows. The four sets of pre-activated style residual blocks are used to transmit features in a serial cascade manner. Each set of pre-activated style residual blocks sequentially performs linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation, and layer normalization. Then, the input vector and output vector of the block are added element-wise through the residual shortcut to achieve feature reuse and gradient optimization.

[0059] Figure 3 A schematic diagram of the enhanced PINN master network of the present invention is shown. Figure 3 As shown, the enhanced PINN main network includes: (1) Feature splicing layer.

[0060] The feature splicing layer is used to splice the layout spacing, frequency, receiving port, excitation port, and the product vector of length vector and electrical length to obtain the spliced ​​input features.

[0061] In this invention, the input of the enhanced PINN main network consists of 4-dimensional original features (layout spacing, frequency, receiving port, and excitation port) and the output features (length vector and electric length product vector) of the multi-scale Fourier feature embedding layer. The output features of the multi-scale Fourier feature embedding layer are 512-dimensional high-dimensional features. Therefore, the feature splicing layer splices the 512-dimensional high-dimensional features with the remaining 4-dimensional original features to form a spliced ​​input feature with a dimension of 516. Thus, this invention fuses high-dimensional frequency-domain electromagnetic features (512-dimensional length vector and electric length product vector) with the original layout spacing, frequency, and port features, while retaining periodic electromagnetic priors and original engineering parameter information, avoiding the loss of low-frequency fundamental information during separate embedding.

[0062] (2) First fully connected layer.

[0063] The first fully connected layer is used to compress and reduce the concatenated input features, reducing the 516-dimensional input features to 256-dimensional features. This allows the 516-dimensional mixed features to be mapped to a unified 256-dimensional underlying latent space, unifying the feature channel size and providing preprocessing for subsequent pre-activated style residual blocks.

[0064] (3) LayerNorm.

[0065] The LayerNorm layer is used to perform layer normalization on the 256-dimensional features to eliminate input distribution bias.

[0066] Specifically, the LayerNorm layer standardizes the feature dimensions and fixes the mean and variance of the input of each layer, which can eliminate feature distribution bias, stabilize the deep network training process, alleviate gradient vanishing, improve the convergence speed and training stability of the prediction model, and adapt to sampling scenarios with variable batches.

[0067] (4) SiLU activation function.

[0068] The SiLU activation function is used to introduce smooth nonlinearity into the 256-dimensional features after layer normalization, in order to fit the complex nonlinear mapping of cable scattering parameters, thereby obtaining the nonlinearly transformed feature vector. The SiLU activation function smooths gradient characteristics, and smooth gradient features avoid neuron death problems, improving the gradient flow performance of deep residual networks.

[0069] (5) Four groups of pre-activated style residual blocks that use serial cascading to transfer features.

[0070] In this invention, the structures and processing flows of the four sets of pre-activated style residual blocks that use a serial cascaded method to transfer features are completely identical, and, as... Figure 4As shown, each group of pre-activated style residual blocks includes a first linear transformation layer (first Linear layer), a first LayerNorm layer, a SiLU activation function, a random deactivation layer (Dropout layer), a second linear transformation layer (second Linear layer), a second LayerNorm layer, and a residual shortcut layer (Shortcut layer). This sequentially performs a first linear transformation (for nonlinear feature transformation, uncovering the deep coupling relationship between geometric, frequency, and other features and S-parameters), a first layer normalization (for stabilizing the input distribution of each layer), SiLU activation (for nonlinear fitting of electromagnetic multi-scale oscillations), random deactivation (for randomly shielding some neurons, suppressing model overfitting, and improving the generalization ability of unknown cable parameters), a second linear transformation (for second feature compression and purification), and a second layer normalization (for standardizing the main branch output). Then, the input features and output features of the block are added element-wise through the residual shortcut to achieve feature reuse and gradient optimization, thereby obtaining 256-dimensional deep features. Therefore, by directly passing the original input features, the gradient vanishing problem in deep networks can be solved, thus enabling the reuse of shallow basic features.

[0071] In this invention, the four sets of pre-activated style residual blocks are used to transmit features in a serial cascade manner. The final output of the previous set of pre-activated style residual blocks is directly used as the input of the next set of pre-activated style residual blocks. The feature extraction and fine transformation are completed block by block. That is, from basic transmission loss to mid-frequency fluctuations, and then to high-frequency resonant peaks and multi-parameter coupling features, the ability of the prediction model to fit complex S-parameter curves is greatly improved.

[0072] (6) First contraction layer.

[0073] The first contraction layer is used to compress the 256-dimensional depth features into 128-dimensional depth features. This filters out redundant and irrelevant features, retains high-value electromagnetic representations, and reduces subsequent computational load.

[0074] Furthermore, preferably, the first contraction layer is followed by a LayerNorm layer (normalization layer) and a SiLU activation function, so as to better maintain the nonlinear expression capability.

[0075] (7) Second contraction layer.

[0076] The second shrinking layer is used to compress the 128-dimensional depth features into 64-dimensional depth features. This completes the purification of the entire network features. The 64-dimensional depth features are a condensed implicit representation integrating all geometric parameters, frequencies, multi-scale Fourier features, and resonance information, storing the complete electromagnetic mapping relationship of the cable.

[0077] Furthermore, preferably, the second contraction layer is also followed by a LayerNorm layer (normalization layer) and a SiLU activation function, so as to better maintain the nonlinear expression capability.

[0078] (8) Output mapping layer.

[0079] The output mapping layer is used to map the 64-dimensional depth features into 2-dimensional results, which correspond to the real and imaginary parts of the cable scattering parameters, respectively. Thus, high-dimensional abstract latent features can be mapped to the complex real and imaginary parts of the S-parameters, which are applicable in engineering.

[0080] The cable scattering parameter prediction model constructed in this invention addresses the problem of poor fitting of high-frequency S-parameter oscillations in traditional networks by building a Fourier embedding branch through the multi-scale Fourier feature embedding layer. It supports deep network training through four sets of pre-activated style residual blocks that transmit features in a serial cascade manner, enabling the capture of multi-scale electromagnetic coupling features. The combination of LayerNorm layers and SiLU activation functions stabilizes the training gradient and adapts to PINN physical constraint joint optimization. Two-stage feature shrinkage through the first and second shrinking layers gradually refines effective features, balancing inference speed and prediction accuracy. The dual-channel real and imaginary part output adapts to the electromagnetic field energy conservation physical loss construction, ensuring that the prediction results conform to the electromagnetic physics of passive cables.

[0081] S3: Iterative training.

[0082] The cable scattering parameter prediction model is trained and iterated using the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

[0083] In this invention, a training mechanism that integrates data-driven and energy conservation physical constraints is adopted. During the training process, the data flows in a closed loop along the path of batch sampling → forward prediction → double loss calculation → reverse parameter update → generalization verification. Combined with a phased progressive constraint strategy and adaptive parameter configuration, the physical compliance of the prediction model output is enhanced while ensuring prediction accuracy.

[0084] 1. Training data batch generation and input flow.

[0085] Before training begins, the training data is grouped and encoded in three levels based on the cable's geometric parameters (length, spacing, etc.), excitation ports, and frequencies (i.e., geometric structure grouping, excitation port grouping, and frequency grouping). Each training data point is then assigned a unique physical constraint group identifier based on this three-level grouping (e.g., 123, where 1 is the geometric structure grouping code, 2 is the excitation port grouping code, and 3 is the frequency grouping code). During training iterations, a batch sampling mechanism is used to generate training data batches: training data within the same batch all come from several complete frequencies under the same geometric structure and excitation port (i.e., all training data within the same batch have the same geometric structure and excitation port, only different frequencies). This ensures that each batch contains as many complete port constraint groups as possible, avoiding constraint group fragmentation caused by random sampling, and providing a data foundation for the effective calculation of subsequent physical losses.

[0086] Each training data batch simultaneously carries three types of data: standardized 6-dimensional input features, S-parameter labels in real and imaginary form (i.e., the true values ​​of S-parameters), and physical constraint grouping identifiers corresponding to the training data.

[0087] 2. Forward propagation and dual-loss fusion calculation.

[0088] When training with batches of training data, the 6-dimensional input features of each training data are first fed into the cable scattering parameter prediction model. The product of length and electrical length is mapped to a high dimension by the multi-scale Fourier feature embedding layer and then concatenated with the remaining 4 original input features (layout spacing, frequency, receiving port, and excitation port) and fed into the enhanced PINN main network to complete feature extraction. Finally, a 2-dimensional prediction result is output, which corresponds to the real and imaginary parts of the S-parameters, respectively.

[0089] In this invention, two types of losses are calculated simultaneously based on the prediction results to form a fusion loss function.

[0090] (1) Data fitting loss.

[0091] Using the true values ​​of the S-parameters from the training data as a benchmark, the deviation between the predicted values ​​and the true values ​​of the S-parameters is calculated using the mean squared error. As the core constraint on the fitting accuracy of the prediction model, it drives the prediction model to learn the mapping relationship between input features and S parameters, thus ensuring basic prediction accuracy.

[0092] If the prediction result of a single training data point in a batch of training data is The actual value of the S-parameter is ,So,

[0093] Where N is the total number of training data in this batch of training data.

[0094] (2) Physical loss due to energy conservation.

[0095] First, the real and imaginary parts of the prediction results are transformed into power amplitudes. Based on pre-assigned physical constraint grouping identifiers, tensor hash summation is used to statistically analyze the power of all receiving ports under the same excitation port, obtaining the sum of power for each group. .in, This represents the complex number of scattering parameters predicted by the cable scattering parameter prediction model, which is based on the same training data (same cable geometry + same excitation port). At the same frequency, the power summation of all receiving ports; subscript It indicates that it is the first Each excitation port applies an incident excitation; subscript It indicates that it is the first Each receiving port receives a response.

[0096] After selecting the complete constraint groups within the training data batch, the fitting bias is calculated with the true power as the target. The final formula for calculating the energy conservation physical loss is as follows:

[0097] in, This represents the total power calculated from the true scattering parameters of training data at the same frequency within the same training batch; This represents the total power calculated from the predicted values ​​of training data at the same frequency within the same training batch; This represents the total number of frequencies within the same training batch.

[0098] The final total loss is the sum of the data fitting loss and the weighted energy conservation physics loss, as shown in the following formula:

[0099] in, These are the weights for physical loss and can be dynamically adjusted as training progresses. For example, in the early stages of training... Initially, the initial value is too low, allowing the model to learn basic data fitting well; in the later stages of training, the value is increased. This strengthens physical constraints and corrects predictions that violate the laws of electromagnetic fields. Therefore, through... This can balance the accuracy of data fitting with physical rationality.

[0100] 3. Training parameter configuration and iterative updates.

[0101] The training configuration uses the AdamW optimizer with an initial learning rate of 1e-3 and a weight decay factor of 1e-5; it is paired with a cosine annealing learning rate scheduling strategy, with T set to...max =100, minimum learning rate eta min =1e-5; Batch training size is 64, and the total number of training rounds is 100. During training, batch data is grouped according to the activation port, and the physical constraint loss is calculated for each group. After each training round, the validation set loss is calculated, and the model weights corresponding to the optimal validation loss are saved in real time to avoid overfitting.

[0102] The training process employs a progressive physical constraint warm-up mechanism: the initial training phase is a pure data fitting phase, with the physical loss weight set to 0, allowing the prediction model to prioritize learning the basic mapping rules at the data level and avoiding physical noise interference with the convergence direction under random initial conditions; after the warm-up phase, the physical loss weight linearly increases to the target value within a set number of rounds, gradually strengthening the role of physical constraints and achieving a smooth integration of data fitting and physical constraints.

[0103] During the backpropagation phase, the gradient is propagated backward along the network based on the total loss. At the same time, a gradient norm pruning mechanism is introduced to limit the maximum magnitude of the gradient, avoid gradient explosion that could lead to training divergence, and ensure the training stability of deep residual networks.

[0104] 4. Generalization verification and optimal model output.

[0105] After each full training iteration, the prediction model is switched to inference mode, and forward prediction is performed on an independently partitioned test set, with validation loss calculated. The test set uses geometric parameters completely different from those in the training set to verify the prediction model's ability to generalize to unknown structures. Throughout training, the historical best value of the validation loss is recorded. When the validation loss surpasses the best record, the current model weights are saved as the final output, ensuring that the final model possesses optimal out-of-domain generalization performance.

[0106] In this invention, after training, the final cable scattering parameter prediction model's structural configuration, model weight parameters, and standardized input / output mean and standard deviation can be saved to a .pth model file, achieving standardized storage of model parameters. The model file and standardized parameters are loaded into the CAE simulation automation script, and batch inference is performed on new cable geometric and frequency parameters. The true S-parameter values ​​are restored through de-standardization, ultimately outputting an S-parameter data file (.snp format) recognizable by the simulation software, completing a rapid prediction that replaces full-wave simulation.

[0107] Figure 5 A schematic diagram of the system for constructing the cable scattering parameter prediction model of the present invention is shown. Figure 5 As shown, the system for constructing the cable scattering parameter prediction model of the present invention includes: 1. Training data construction module.

[0108] The training data construction module is used to collect the length, spacing, frequency, receiving port, excitation port, electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths to construct training data.

[0109] 2. Cable scattering parameter prediction model establishment module.

[0110] The cable scattering parameter prediction model building module is used to build a cable scattering parameter prediction model. This model includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer is used to map the low-dimensional length and electrical length product to a high-dimensional feature space through learnable scale optimization and multi-scale transformation fusion, thereby obtaining a high-dimensional length vector and electrical length product vector. The enhanced PINN main network uses the cable's layout spacing, frequency, receiving port, excitation port, length vector, and electrical length product as the vectors to obtain the high-dimensional length vector and electrical length product vector. The input is the quantity, which predicts the real and imaginary parts of the cable scattering parameters. The enhanced PINN main network adopts a residual enhancement main network structure and has four sets of pre-activated style residual blocks with completely identical structures and processing flows. The four sets of pre-activated style residual blocks are serially cascaded to pass features. Each set of pre-activated style residual blocks sequentially performs linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation and layer normalization. Then, the input features and output features of the block are added element-wise through the residual shortcut to achieve feature reuse and gradient optimization.

[0111] 3. Training Iteration Module.

[0112] The training iteration module is used to train and iterate the cable scattering parameter prediction model with the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

[0113] Furthermore, this invention also provides a device for constructing a cable scattering parameter prediction model. For example... Figure 6 As shown, the device for constructing the cable scattering parameter prediction model of the present invention includes: a memory 11 for storing one or more programs; one or more processors 12; when the one or more programs are executed by the one or more processors 12, the one or more processors 12 implement the method for constructing the cable scattering parameter prediction model of the present invention. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing the cable scattering parameter prediction model of the present invention.

[0114] The computer-readable storage medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined in this invention, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0115] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0116] The steps of the methods or algorithms described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software modules executed by a processor, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention based on the concept of the present invention, without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for constructing a cable scattering parameter prediction model, characterized in that, Includes the following steps: S1: Collect the length, spacing, frequency, receiving port, excitation port, electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths to construct training data; S2: Establish a cable scattering parameter prediction model, which includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer is used to map the low-dimensional length and electrical length product to a high-dimensional feature space through learnable scale optimization and multi-scale transformation fusion, so as to obtain a high-dimensional length vector and electrical length product vector. The enhanced PINN main network takes the cable layout spacing, frequency, receiving port, excitation port, length vector, and electrical length product vector as input to predict the real and imaginary parts of the cable scattering parameters. The strong PINN main network adopts a residual-enhanced main network structure and has four sets of pre-activated style residual blocks with completely identical structures and processing flows. These four sets of pre-activated style residual blocks transmit features in a serial cascade manner. Each set of pre-activated style residual blocks sequentially performs linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation, and quadratic layer normalization. Then, the input and output features of the block are added element-wise through a residual shortcut to achieve feature reuse and gradient optimization. The enhanced PINN main network includes a feature splicing layer, which is used to deploy... The input features are concatenated by splicing the spacing, frequency, receiving port, excitation port, length vector, and electrical length product vector. A first fully connected layer is used to reduce the dimensionality of the concatenated input features, reducing the 516-dimensional input features to 256-dimensional features. A LayerNorm layer is used to normalize the 256-dimensional features to eliminate input distribution offset. A SiLU activation function is used to introduce smooth nonlinearity into the normalized 256-dimensional features to fit the complex nonlinear mapping of the cable scattering parameters, thus obtaining a nonlinear transformation. The system consists of: a feature vector; four pre-activated style residual blocks that transmit features in a serial cascade manner, used to perform depth extraction and refinement transformation of the feature vector after the nonlinear transformation block by block, to obtain 256-dimensional depth features; a first shrinking layer, used to compress the 256-dimensional depth features into 128-dimensional depth features; a second shrinking layer, used to compress the 128-dimensional depth features into 64-dimensional depth features; and an output mapping layer, used to map the 64-dimensional depth features into 2-dimensional results, the 2-dimensional results corresponding to the real and imaginary parts of the cable scattering parameters, respectively. S3: The cable scattering parameter prediction model is trained and iterated using the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

2. The method for constructing a cable scattering parameter prediction model according to claim 1, characterized in that, In step S1, the product of the acquisition length, deployment spacing, frequency, and electric length, as well as the real and imaginary parts of the scattering parameters, are also normalized.

3. The method for constructing a cable scattering parameter prediction model according to claim 1, characterized in that, The multi-scale Fourier feature embedding layer includes: The input scaling unit is used to independently scale the product of length and electrical length based on a trainable scaling vector to obtain the scaled product of length and electrical length. A multi-scale projection unit is used to map the scaled product of length and electrical length to feature subspaces of different scales based on a set of parallel trainable projection matrices to obtain projection results of different scales, wherein each of the trainable projection matrices corresponds to a preset scale factor. The nonlinear transformation and splicing unit is used to apply sine and cosine transformations to the projection results at each scale and splice the outputs of the sine and cosine transformations of the projection results at all scales along the feature dimension to form the final length vector and electrical length product vector.

4. The method for constructing a cable scattering parameter prediction model according to claim 1, characterized in that, In step S3, before training, the training data is grouped and encoded in three levels based on the cable's geometric parameters, excitation port, and frequency, thereby assigning a unique physical constraint group identifier to each training data. During the training iteration, a group batch sampling mechanism is used to generate training batches, and the training data in the same training batch all come from several frequencies under the same geometric structure and the same excitation port.

5. The method for constructing a cable scattering parameter prediction model according to claim 4, characterized in that, In step S3, the loss function for: in, This represents the data fitting loss, which is the mean square error between the predicted values ​​of the training data and the true scattering parameters. This represents the physical loss due to energy conservation. The weight for physical loss due to energy conservation.

6. The method for constructing a cable scattering parameter prediction model according to claim 5, characterized in that, In step S3 in, This represents the total power calculated from the true scattering parameters of training data at the same frequency within the same training batch; This represents the total power calculated from the predicted values ​​of training data at the same frequency within the same training batch; This represents the total number of frequencies within the same training batch.

7. A system for constructing a cable scattering parameter prediction model, characterized in that, include: The training data construction module is used to collect the length, spacing, frequency, receiving port, excitation port, electrical length product, and the real and imaginary parts of the corresponding scattering parameters of cables of different lengths in order to construct training data. A cable scattering parameter prediction model building module is used to establish a cable scattering parameter prediction model. This model includes a multi-scale Fourier feature embedding layer and an enhanced PINN main network. The multi-scale Fourier feature embedding layer uses learnable scale optimization and multi-scale transformation fusion to map the low-dimensional length and electrical length product to a high-dimensional feature space, obtaining a high-dimensional length vector and electrical length product vector. The enhanced PINN main network takes the cable's layout spacing, frequency, receiving port, excitation port, length vector, and electrical length product vector as input to predict the cable scattering parameters. The enhanced PINN main network adopts a residual-enhanced main network structure and has four sets of pre-activated style residual blocks with completely identical structures and processing flows. These four sets of pre-activated style residual blocks transmit features in a serial cascade manner. Each set of pre-activated style residual blocks sequentially performs linear transformation, layer normalization, SiLU activation, random deactivation, quadratic linear transformation, and layer normalization. Then, the input and output features of the block are added element-wise through a residual shortcut to achieve feature reuse and gradient optimization. The enhanced PINN main network includes a feature concatenation layer. It is used to concatenate the features of the deployment spacing, frequency, receiving port, excitation port, length vector, and electrical length product vector to obtain the concatenated input features; the first fully connected layer is used to reduce the dimensionality of the concatenated input features, reducing the 516-dimensional input features to 256-dimensional features; the LayerNorm layer is used to perform layer normalization on the 256-dimensional features to eliminate input distribution offset; the SiLU activation function is used to introduce smooth nonlinearity into the 256-dimensional features after layer normalization to fit the complex nonlinear mapping of the cable scattering parameters, thereby obtaining nonlinear... The system consists of: a feature vector after nonlinear transformation; four pre-activated style residual blocks that transmit features in a serial cascade manner, used to perform depth extraction and refinement transformation of the feature vector after nonlinear transformation block by block, to obtain 256-dimensional depth features; a first shrinking layer, used to compress the 256-dimensional depth features into 128-dimensional depth features; a second shrinking layer, used to compress the 128-dimensional depth features into 64-dimensional depth features; and an output mapping layer, used to map the 64-dimensional depth features into 2-dimensional results, the 2-dimensional results corresponding to the real and imaginary parts of the cable scattering parameters, respectively. The training iteration module is used to train and iterate the cable scattering parameter prediction model with the training data to obtain the final cable scattering parameter prediction model. During the training iteration, the loss function is composed of data fitting loss and energy conservation physical loss.

8. A device for constructing a cable scattering parameter prediction model, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a cable scattering parameter prediction model as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for constructing the cable scattering parameter prediction model as described in any one of claims 1-6.

Images