Random cable bundle broadband common-mode current fast prediction method and system

Abstract

The application relates to a random cable bundle broadband common-mode current fast prediction method and system. The method comprises the following steps: acquiring a three-dimensional geometric sample of a random cable bundle to be analyzed; acquiring source end common-mode current and load end common-mode current at each discrete frequency point through joint solution of a moment method and a multi-conductor transmission line, and combining the two to obtain a broadband common-mode current reference label; mapping the three-dimensional geometric sample to a top view and a side view geometric pixel map using the same axial sampling reference; constructing a double-view convolutional neural network, including a top view encoding branch, a side view encoding branch, a feature fusion module and a double-output regression module, training the network by taking double-view images as input and the broadband common-mode current reference label as supervision; and for a new sample, inputting the double-view into the trained network to obtain a broadband common-mode current prediction curve of the source end and the load end. The method can reduce repeated calculation in random cable bundle broadband response prediction and improve prediction efficiency without repeatedly performing a complete physical solution process when predicting a new sample.

Description

Technical Field

[0001] This application relates to the field of electromagnetic compatibility analysis technology, and in particular to a method and system for rapid prediction of broadband common-mode current in random cable bundles. Background Technology

[0002] In complex electronic and electrical systems, cable bundles are crucial for connecting power supplies, control units, sensors, and actuators, and are widely used in aerospace, rail transportation, automotive electronics, marine equipment, power electronic systems, and industrial control equipment. As systems evolve towards higher integration, higher power density, and miniaturization, the number of cables is constantly increasing, and deployment space is becoming increasingly limited, leading to more prominent electromagnetic coupling issues between cables and between cables and surrounding metal structures. Common-mode current in cable bundles not only affects signal transmission quality and control reliability but can also trigger conducted interference, radiated interference, and abnormal port responses. Therefore, rapid and accurate prediction of broadband common-mode current in random cable bundles is a critical technical issue in the electromagnetic compatibility analysis and design of complex electrical systems.

[0003] Existing research on predicting the broadband electromagnetic response of random cable bundles mainly focuses on traditional electromagnetic modeling methods, data-driven modeling methods, and geometric input representation methods. Traditional electromagnetic modeling methods include transmission line theory, multi-conductor transmission line models, field-line coupling models, the method of moments, the finite element method, and the time-domain finite difference method. These methods have strong physical foundations and good solution accuracy, enabling the analysis of the electromagnetic response of cable bundles under specific geometric shapes, termination conditions, and frequency ranges. However, for the large number of random cable bundle samples present in practical engineering, when the prediction object covers a wide frequency band, it is usually necessary to repeatedly perform unit length parameter extraction and port response solving for each geometric sample. This results in a large computational load and low solution efficiency, making it difficult to meet the application requirements of rapid screening, batch evaluation, and iterative optimization in the engineering design stage.

[0004] In recent years, some studies have attempted to apply data-driven methods, such as neural networks, to model electromagnetic problems in cables. For example, some methods use neural networks to predict the unit-length parameter matrix of multi-conductor transmission lines, and then combine this with analytical models to complete subsequent response calculations; other methods directly establish a mapping relationship between geometry and electromagnetic response for crosstalk prediction, defect identification, or port response estimation. These methods can reduce the computational cost of repetitive simulations to some extent, but existing research mainly focuses on unit-length parameter prediction or estimation of specific response quantities, and it is still difficult to directly achieve rapid prediction of broadband common-mode current curves at the source and load ends. At the same time, existing data-driven methods are not tightly integrated with the physical reference solution process, and there are still shortcomings in balancing the physical consistency of supervised labels and the efficiency of prediction models.

[0005] For spatially random cable bundles, the cables typically exhibit characteristics such as bending, local undulations, random bundling, conductor constraints, and limited space placement. The conductor spacing, proximity relationships, axial positional variations, and termination conditions all affect the port common-mode current response. Existing data-driven methods sometimes use only cross-sectional images as input, which can reflect the lateral arrangement of conductors but struggles to describe the axial bending and undulations of the cables. Others use only a single projected view as input, which, while reflecting some spatial orientation information, easily overlooks proximity relationships and height variations in another spatial direction. Therefore, existing input representation methods do not fully utilize the three-dimensional geometric features of random cable bundles and cannot comprehensively characterize their spatial structural properties.

[0006] In summary, existing technologies still have the following shortcomings in predicting broadband common-mode current of random cable bundles: On the one hand, traditional electromagnetic solution methods have high repetitive calculation costs and limited prediction efficiency under conditions of large-scale random samples and wide frequency points; on the other hand, existing data-driven methods do not make sufficient use of the three-dimensional geometric information of random cable bundles and lack a unified and clear collaborative modeling process between physical reference label generation and fast neural network prediction, making it difficult to achieve fast and direct prediction of broadband common-mode current at the source and load ends while ensuring physical consistency. Therefore, it is difficult to meet the practical application needs of rapid screening and engineering evaluation of electromagnetic compatibility of complex cable bundles.

[0007] Therefore, there is an urgent need in related technologies for a method to achieve fast and direct prediction of broadband common-mode current at the source and load ends under the conditions of large-scale random samples and wide frequency points, while ensuring physical reference consistency. Summary of the Invention

[0008] Therefore, it is necessary to provide a method and system for rapid prediction of broadband common-mode current of random cable bundles to address the above-mentioned technical problems.

[0009] Firstly, this application provides a method for rapid prediction of broadband common-mode current in random cable bundles. The method includes: Obtain a three-dimensional geometric sample of a random cable bundle to be analyzed. The three-dimensional geometric sample of the random cable bundle includes three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions. The source-side common-mode current and load-side common-mode current at each discrete frequency point are obtained by jointly solving the method of moments and multi-conductor transmission lines, and the source-side common-mode current and load-side common-mode current at each discrete frequency point are combined into a broadband common-mode current reference label. The three-dimensional geometric samples of the random cable bundle are mapped into a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference; A dual-view convolutional neural network is constructed, comprising a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel map and the side-view geometric pixel map are respectively input into the top-view encoding branch and the side-view encoding branch to extract top-view features and side-view features. The feature fusion module fuses the top-view features and the side-view features and then inputs them into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point. The dual-view convolutional neural network is trained using the wideband common-mode current reference label as supervision. The top-view and side-view geometric pixel maps of the new random cable bundle sample to be predicted are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves of the source and load ends of the new random cable bundle sample.

[0010] Optionally, in one embodiment of this application, the top view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the first transverse direction, and the side view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the second transverse direction, wherein the first transverse direction and the second transverse direction are orthogonal to each other.

[0011] Optionally, in one embodiment of this application, mapping the three-dimensional geometric samples of the random cable bundle into a top-view geometric pixel map and a side-view geometric pixel map further includes: Each conductor is assigned to an independent image channel, and the channel corresponding to the excitation conductor is fixed at a preset channel position.

[0012] Optionally, in one embodiment of this application, the top view encoding branch and the side view encoding branch adopt convolutional encoders with independent parameters and identical structures. Each convolutional encoder includes an initial convolutional layer, multiple cascaded residual downsampling modules and an adaptive average pooling layer. The residual downsampling module includes a convolution operation unit, a normalization processing unit, an activation function unit, a channel attention mechanism unit and a residual skip connection unit.

[0013] Optionally, in one embodiment of this application, the feature fusion module fuses the top view features and the side view features, which includes at least constructing original feature terms, difference terms, element-wise product terms and mean terms between the top view features and the side view features, and combining the original feature terms, the difference terms, the element-wise product terms and the mean terms into fused features.

[0014] Optionally, in one embodiment of this application, the dual-output regression module includes a shared mapping layer, a load-side regression head, and a source-side regression head. After the fused features are processed by the shared mapping layer, they are respectively input to the load-side regression head and the source-side regression head. The load-side regression head outputs the load-side common-mode current prediction curve at each discrete frequency point, and the source-side regression head outputs the source-side common-mode current prediction curve at each discrete frequency point.

[0015] Optionally, in one embodiment of this application, training the dual-view convolutional neural network using the wideband common-mode current reference label as supervision includes: The wideband common-mode current reference label is subjected to logarithmic transformation and standardization. The dual-view convolutional neural network outputs the standardized logarithmic domain prediction value, which is then restored to the amplitude domain common-mode current prediction curve through inverse standardization and inverse logarithmic transformation.

[0016] Optionally, in one embodiment of this application, training the dual-view convolutional neural network using the wideband common-mode current reference label as supervision further includes: The dual-view convolutional neural network is trained using an enhanced combined loss function, which includes a logarithmic domain robust loss term, an amplitude domain relative error loss term, a spectral slope consistency loss term, a peak region loss term, and a low current region loss term.

[0017] Secondly, this application also provides a fast prediction system for broadband common-mode current of random cable bundles. The system includes: The geometric sample acquisition module is used to acquire three-dimensional geometric samples of random cable bundles to be analyzed. The three-dimensional geometric samples of random cable bundles include three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions. The physical reference label generation module is used to obtain the source-side common-mode current and load-side common-mode current at each discrete frequency point by jointly solving the method of moments and multi-conductor transmission lines, and to combine the source-side common-mode current and load-side common-mode current at each discrete frequency point into a broadband common-mode current reference label. The dual-view mapping module is used to map the three-dimensional geometric samples of the random cable bundle into a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference. The model building module is used to construct a dual-view convolutional neural network. The dual-view convolutional neural network includes a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel map and the side-view geometric pixel map are respectively input into the top-view encoding branch and the side-view encoding branch to extract top-view features and side-view features. The feature fusion module fuses the top-view features and the side-view features and then inputs them into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point. The model training module is used to train the dual-view convolutional neural network using the wideband common-mode current reference label as supervision. The fast prediction module is used to acquire the three-dimensional geometric data of the new random cable bundle sample to be predicted, and construct its top view geometric pixel map and side view geometric pixel map according to the mapping method consistent with the training phase. The top view geometric pixel map and the side view geometric pixel map are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves of the source end and load end of the new random cable bundle sample.

[0018] The above-mentioned method and system for rapid prediction of broadband common-mode current in random cable bundles has the following advantages compared with the prior art: First, this invention reduces the repetitive electromagnetic solution process in broadband response prediction of random cable bundles, thereby improving the efficiency of broadband response prediction. Existing technologies for broadband electromagnetic response analysis of random cable bundles typically require repeated parameter extraction using the method of moments (MoM) and segmented propagation solutions for each geometric sample, leading to high computational complexity and low analysis efficiency. This invention proposes a rapid prediction method based on collaborative modeling using physical reference labels and a dual-view convolutional neural network. During the training phase, this invention utilizes the joint solution process of MoM and multi-conductor transmission lines to generate broadband common-mode current physical reference labels for the source and load ends. During the prediction phase, the trained neural network directly outputs the broadband common-mode current response curve corresponding to the new random cable bundle sample, eliminating the need to repeat the complete electromagnetic parameter extraction and propagation solution process for each new sample. Therefore, this invention reduces the complex solution process caused by repeated calculations per sample and per frequency band in traditional methods, improving the speed of broadband response prediction for random cable bundles. It is suitable for rapid screening and response evaluation of large-scale random samples in complex electrical systems.

[0019] Secondly, this invention improves the representation capability of three-dimensional spatial information of random cable bundles through a dual-view geometric input method. Addressing the problem that existing technologies using a single cross-sectional image or a single projected view as input cannot fully reflect the three-dimensional spatial structural characteristics of random cable bundles, this invention proposes a dual-view geometric representation method that combines a top view and a side view as input. The top view characterizes the lateral arrangement, proximity relationships, and in-plane spatial distribution characteristics of the conductors, while the side view characterizes the undulations, bending, and height variations of the conductors along the axial direction. By inputting both view information into a convolutional neural network, the model can simultaneously acquire information on the lateral arrangement and axial geometric changes of the cable bundle. Compared to existing methods that rely solely on a single view input, this invention can more completely describe the three-dimensional spatial geometric characteristics of random cable bundles, thereby improving the model's ability to identify differences in different geometric samples and enhancing the sensitivity and accuracy of broadband common-mode current prediction results to spatial structural changes.

[0020] Third, this invention balances physical reference consistency with data-driven prediction efficiency, improving the engineering reliability of the model results. Existing data-driven methods often directly use empirical samples or single numerical results as supervision, failing to fully reflect the impact of conductor axial relative position changes on unit length parameters and port responses, resulting in insufficient physical consistency of the model. This invention employs a joint solution process using the method of moments (MoM) and multi-conductor transmission lines to construct supervision labels. The MoM extracts the unit length parameters of each cross-section of a random cable bundle, and combines this with a segmented propagation model of a non-uniform multi-conductor transmission line to solve for the broadband common-mode current response at the source and load ends. This allows the supervision labels to reflect the actual impact of axial geometric changes in the random cable bundle on the port response. The neural network surrogate model trained on this basis inherits the physical consistency of the physical reference solution while possessing the speed of data-driven models in predicting new samples. Therefore, this invention technically achieves effective synergy between physical modeling methods and neural network surrogate prediction, contributing to improved engineering applicability and reliability of the prediction results.

[0021] Fourth, this invention enables direct prediction of key engineering parameters, simplifying existing multi-stage post-processing procedures. Addressing the issue that some existing methods primarily predict intermediate parameters such as the inductance and capacitance matrices per unit length, requiring subsequent analytical models or numerical solutions to obtain port response results, this invention directly uses the wideband common-mode current sequences at the source and load ends as supervisory outputs and employs a dual-output regression structure to complete the prediction. Through this technical solution, this invention eliminates the need to first predict the distributed parameter matrix before subsequent propagation calculations, directly outputting the common-mode current curves at the source and load ends that are of engineering interest. Therefore, this invention reduces intermediate modeling steps and post-processing in the technical process, making the prediction results more readily applicable for rapid electromagnetic compatibility screening of cable bundles, risk identification, and response comparison analysis between different wiring schemes.

[0022] Fifth, it is applicable to wide-band, multi-frequency, and large-sample random cable bundle tasks, improving the method's adaptability. This invention constructs a complete technical process for predicting wide-band common-mode current in random cable bundles, from acquiring existing random geometric samples and calculating physical reference labels to training and predicting dual-view neural networks. In this process, the frequency range, number of frequency points, image resolution, and sample size can all be set according to engineering requirements. Therefore, this invention can not only output source and load common-mode current response curves within a specified wide-band range, but is also suitable for batch modeling and prediction tasks under multi-frequency and large-sample conditions. Compared to existing methods that analyze single samples, single frequency bands, or limited frequency points, this invention has greater flexibility and scalability in terms of task scale and frequency band adaptation.

[0023] Sixth, it improves the uniformity and completeness of the broadband response modeling process for complex random cable bundles. In existing technologies, the geometric input, electromagnetic solution, and data-driven prediction of random cable bundles are often separate, lacking a unified collaborative process, resulting in insufficient connection between model training and engineering applications. This invention organically combines acquired random geometric samples, the method of moments (MoM) and multi-conductor transmission line joint solution for physical reference label construction, and dual-view convolutional neural network training and prediction, forming a unified modeling process from geometric input to target response output. This unified process improves the consistency between geometric input, physical solution, and surrogate model prediction in the broadband common-mode current prediction problem of random cable bundles, providing a clear and easily implemented technical solution for electromagnetic compatibility analysis of cable bundles in complex electrical systems. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for rapid prediction of broadband common-mode current in a random cable bundle, as described in one embodiment. Figure 2 This is a schematic diagram of a three-dimensional geometric sample of a random cable bundle obtained in one embodiment; Figure 3 This is a schematic diagram of the construction of a dual-view geometric pixel map in one embodiment; Figure 4 This is a schematic diagram of the structure of a dual-view convolutional neural network in one embodiment; Figure 5 This is a detailed schematic diagram of the residual downsampling module in one embodiment; Figure 6 This is a block diagram of a fast prediction system for broadband common-mode current of random cable bundles in one embodiment; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In one embodiment, such as Figure 1 As shown, a method for fast prediction of broadband common-mode current in random cable bundles is provided, including the following steps: In this embodiment of the application, firstly, a three-dimensional geometric sample of the random cable bundle to be analyzed is obtained. For example... Figure 2 As shown, the three-dimensional geometric sample of the random cable bundle includes three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions, and may also include information such as conductor number, conductor radius, outer radius of insulation layer, dielectric parameters, cable length, termination conditions and excitation conditions.

[0027] The three-dimensional geometric sample of the random cable bundle can be obtained through electromagnetic simulation modeling, computer-aided design, three-dimensional scanning, engineering measurement, existing geometric databases or other geometric modeling methods. This application does not limit the specific generation method of the three-dimensional geometric sample of the random cable bundle.

[0028] For the i-th random cable bundle sample, its three-dimensional geometry can be represented as a set of three-dimensional coordinates of the center lines of each conductor at multiple axial positions. After obtaining the three-dimensional geometry sample, it is used for physical reference label generation and dual-view geometric pixel map construction, respectively.

[0029] In this embodiment, the source-side common-mode current and load-side common-mode current at each discrete frequency point are then obtained by jointly solving the method of moments and multi-conductor transmission lines. The source-side common-mode current and load-side common-mode current at each discrete frequency point are then combined into a broadband common-mode current reference label. For each random cable bundle geometry sample, it is discretized along the axial direction. A locally uniform small segment. In the first... Within a small segment, based on the conductor positions, conductor radii, outer radius of the insulation layer, and dielectric parameters on the corresponding cross-section, the local conductor arrangement is transformed into a two-dimensional quasi-static cross-sectional problem. The method of moments (MoM) parameter extraction program is then used to calculate the capacitance matrix per unit length for that segment. and unit length inductance matrix Therefore, we obtain the first... Piecewise parameter field of a random geometric sample:

[0030] in, For the first The length of a local segment.

[0031] After obtaining the segmented parameter field, a non-uniform multi-conductor transmission line model is constructed based on the port excitation and termination conditions. For each discrete frequency point... Propagation matrices are constructed for each locally uniform segment and cascaded according to the cable axial direction to solve for the conductor current vectors at the source and load ends. The conductor current vectors at the source and load ends can be denoted as follows: and .

[0032] Calculate the common-mode current for the current vectors at the source and load ends respectively. The common-mode current is defined as the average of the current vectors and amplitudes of all conductors.

[0033]

[0034] in, Indicates all elements are column vectors, For the number of conductors. For those containing A broadband prediction task at discrete frequency points combines the common-mode current sequences at the load and source ends into a supervision label:

[0035] The above steps enable the construction of a physically consistent broadband common-mode current reference tag for each random cable bundle geometry sample. This tag contains both the load-side and source-side responses, making it suitable for subsequent dual-output neural network regression training.

[0036] In this embodiment of the application, the three-dimensional geometric samples of the random cable bundle are then mapped to a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference. For example... Figure 3 As shown, in the dual-view construction process, the same random cable bundle 3D geometric sample is projected into a top view and a side view. The top view represents the conductor's lateral arrangement, local proximity relationships, and in-plane spatial distribution; the side view represents the conductor's undulations, bends, and height changes along the axial direction. To ensure spatial correspondence between the two types of views during feature fusion, the top view and side view use the same axial sampling reference, ensuring that pixels in the same column correspond to the same axial position of the cable bundle in both views. Subsequently, the geometric pixel map is converted into a multi-channel encoded form, where each conductor corresponds to an independent channel, and the channel corresponding to the excitation wire is adjusted to a fixed channel position to maintain consistency in the input meaning between different samples. The random cable bundle geometric sample... According to the preset image resolution Normalization mapping is performed, and two types of two-dimensional geometric pixel maps are constructed, namely top view and side view. The top view uses... - Planar projection is used to characterize the positional changes of conductors along the axial direction, their lateral arrangement, and local proximity relationships; the side view adopts... - Planar projection is used to characterize the conductor's bending, undulation, and height changes along the axial direction. During image mapping, the trajectory of the conductor's centerline on the corresponding projection plane is mapped to pixels or pixel segments, and normalization is applied to ensure that different random cable bundle samples can be mapped to an image space of a uniform size. The top view and side view use the same axial sampling reference, ensuring that pixels in the same column in both views correspond to the same axial position of the cable bundle, thereby guaranteeing the spatial correspondence of features from both views in the subsequent network fusion process.

[0037] Let the first The three-dimensional geometric set of a random cable bundle sample is Then its dual-view image input can be represented as:

[0038] in, Operators representing the mapping from geometry to image space. and These represent top view input and side view input, respectively. Both types of views correspond to the same broadband common-mode current label. .

[0039] In one embodiment of this application, the top view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the first transverse direction, and the side view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the second transverse direction, wherein the first transverse direction and the second transverse direction are orthogonal to each other.

[0040] In one embodiment of this application, when mapping the three-dimensional geometric samples of the random cable bundle to the top view geometric pixel map and the side view geometric pixel map, each conductor is assigned to an independent image channel, and the channel corresponding to the excitation conductor is fixed at a preset channel position to ensure that the meaning of the input channel is consistent between different samples, and to avoid the neural network from learning the mapping relationship between geometric features and common-mode current response due to changes in conductor number or channel order.

[0041] In this embodiment of the application, a dual-view convolutional neural network is then constructed, such as... Figure 4As shown, the dual-view convolutional neural network includes a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel image and the side-view geometric pixel image are input into the top-view encoding branch and the side-view encoding branch, respectively, to extract top-view features and side-view features. The feature fusion module fuses the top-view features and side-view features and then inputs the fused features into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point. Through this dual-branch structure, the network can learn the geometric features under different projection perspectives, rather than simply mixing the two types of views into a single image input. The dual-output structure enables the network to directly predict the broadband common-mode current curves at the source and load sides, avoiding the complex process of first predicting the distributed parameter matrix and then performing transmission line post-processing.

[0042] In one embodiment of this application, the top-view encoding branch and the side-view encoding branch employ convolutional encoders with independent parameters but identical structures. Each convolutional encoder includes an initial convolutional layer, multiple cascaded residual downsampling modules, and an adaptive average pooling layer, such as... Figure 5 As shown, the residual downsampling module includes a convolution operation unit, a normalization processing unit, an activation function unit, a channel attention mechanism unit, and a residual skip connection unit. The convolution operation is used to extract local geometric texture and conductor trajectory features; the normalization processing is used to improve training stability under different batch and resolution conditions; the channel attention mechanism is used to enhance the importance expression between different conductor channels and feature channels; and the residual skip connection is used to alleviate gradient decay and feature degradation problems in the deep feature extraction process. Through this structure, the network can extract multi-scale geometric features from local conductor proximity relationships to the overall cable bundle spatial morphology during the step-by-step downsampling process.

[0043] In one embodiment of this application, the feature fusion module fuses the top-view features and the side-view features, including at least constructing original feature terms, difference terms, element-wise product terms, and mean terms between the top-view features and the side-view features, and combining the original feature terms, the difference terms, the element-wise product terms, and the mean terms into fused features. The original feature terms are used to retain the independent information of the top-view and side-view features, the difference terms are used to characterize the complementary differences between the geometric features under the two perspectives, the element-wise product terms are used to characterize the correlation between the features of the two views, and the mean term is used to characterize the common geometric features reflected by the two types of views. Through the above fusion method, the network can simultaneously utilize the independent information, complementary information, and related information of the two views, thereby enhancing its ability to represent the three-dimensional spatial structure of random cable bundles.

[0044] In one embodiment of this application, the dual-output regression module includes a shared mapping layer, a load-side regression head, and a source-side regression head. After the fused features are processed by the shared mapping layer, they are respectively input to the load-side regression head and the source-side regression head. The load-side regression head outputs the load-side common-mode current prediction curve at each discrete frequency point, and the source-side regression head outputs the source-side common-mode current prediction curve at each discrete frequency point.

[0045] In this embodiment, the wideband common-mode current reference label is used as supervision to train the dual-view convolutional neural network. For the first... A random cable bundle sample is used, with its top-view and side-view inputs combined as the network input. The load-source broadband common-mode current sequences obtained by the method of moments and multi-conductor transmission lines are used as supervision labels. If the prediction task includes... For each discrete frequency point, the supervisory label for a single sample includes the load end. Common-mode current response and source at each frequency point The common-mode current response at each frequency point, the network output dimension is After training is complete, save the network parameters that best perform on the validation set.

[0046] In one embodiment of this application, when training the dual-view convolutional neural network using the wideband common-mode current reference label as supervision, the wideband common-mode current reference label undergoes logarithmic transformation and standardization. The dual-view convolutional neural network outputs a standardized logarithmic domain predicted value, which is then recovered as an amplitude domain common-mode current predicted curve through inverse standardization and inverse logarithmic transformation. This processing method can reduce the training instability caused by the large amplitude range of the wideband common-mode current across different frequency bands and different samples.

[0047] The training dataset can be divided into training, validation, and test sets according to a preset ratio. For example, in one specific embodiment, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used to update network parameters, the validation set is used for model selection and monitoring the training process, and the test set is used to evaluate the broadband common-mode current prediction performance of the final model. For comparative experiments with different input methods or different image resolutions, the sample numbers of the training, validation, and test sets are kept consistent to ensure the comparability of comparison results between different models.

[0048] The model training employs a dual-output regression training method. After forward propagation, the load-side common-mode current prediction sequence and the source-side common-mode current prediction sequence are obtained, and their errors are calculated against the corresponding MoM-MTL reference labels. In one embodiment of this application, when training the dual-view convolutional neural network using the broadband common-mode current reference label as supervision, an enhanced combined loss function is used to train the dual-view convolutional neural network. The enhanced combined loss function has the following form:

[0049] in, Logarithmic domain robust loss is used to constrain the overall magnitude regression error; This is the relative error loss in the amplitude domain, used to improve the model's sensitivity to relative deviations; This is the spectral slope consistency loss, used to constrain the variation trend between adjacent frequency points; This is the peak region loss, used to enhance the model's ability to fit local high-response frequency bands; The loss term for the low-current region is used to improve the prediction stability in the weak response range. The loss term is calculated at the load and source outputs respectively, and averaged over the frequency and sample dimensions. For loss weighting, weights are set based on frequency location and the amplitude distribution of the true common-mode current curve. Frequency weights are used to increase the contribution of key frequency bands during training, while local weights are set based on peak and low-current regions in the true curve, allowing the model to simultaneously consider the overall response trend, local peak-valley variations, and prediction accuracy in the weak response range.

[0050] In one embodiment of this application, the model training employs a two-stage training strategy. The first stage uses the full set of training samples for basic training, enabling the network to learn the overall mapping relationship between the dual-view geometry of a random cable bundle and the broadband common-mode current at the source and load ends. The second stage filters difficult samples based on the prediction errors of the model in the first stage on the training or validation set, and further fine-tunes the model parameters on this subset of difficult samples to enhance the model's fitting ability to high-error samples, local peak-valley responses, and low-current ranges. The difficult sample selection criteria comprehensively consider relative error, peak error, spectral slope error, and low-current range error. During the specific training process, the AdamW optimizer is used to update the network parameters, and a learning rate scheduling strategy is employed to adjust the learning rate during training. To improve training stability, gradient pruning, mixed-precision training, early stopping mechanisms, and exponential moving average parameter updates are also used. For inputs with different image resolutions, gradient accumulation is used to unify the equivalent batch size, reducing the impact of memory limitations and batch size differences on the model training results. Since Group Normalization can be used instead of Batch Normalization in dual-view convolutional neural networks, its normalization process does not depend on batch dimension statistics, making it more suitable for training processes under multi-resolution and small-batch conditions.

[0051] In this embodiment, the top-view and side-view geometric pixel maps of the new random cable bundle sample to be predicted are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves for the source and load ends of the new random cable bundle sample. For the new random cable bundle sample to be predicted, the three-dimensional geometric data of the sample is first acquired, and the top-view and side-view geometric pixel maps are constructed according to the same mapping method as in the training phase. Then, the dual-view input is fed into the trained convolutional neural network surrogate model, which directly outputs the broadband common-mode current prediction curves for the load and source ends at preset frequency points. In the prediction phase, it is not necessary to repeatedly perform the complete method of moments parameter extraction and multi-conductor transmission line segment propagation solution for each new sample. Only the dual-view geometric construction and neural network forward propagation are needed to obtain the prediction results, thereby achieving rapid evaluation of the broadband common-mode current of the random cable bundle.

[0052] In one specific embodiment, a 7-conductor random cable bundle model was used for verification. Both the source and load ends were terminated at 50Ω. The excitation condition was set to apply a 1V single-ended voltage excitation to the first conductor, with no independent excitation for the remaining conductors. The frequency bands included 10kHz–30MHz, 0.03GHz–2GHz, 0.8GHz–2GHz, and 10kHz–2GHz, with 30 discrete frequency points selected within each band. For each random cable bundle sample, the output label consisted of the common-mode current response at the load end and the common-mode current response at the source end at 30 frequency points, with a 60-dimensional output label for a single sample. Approximately several random cable bundle samples containing three-dimensional geometric coordinate data were obtained, and corresponding broadband common-mode current reference labels for the source and load ends were generated using a joint solution process combining the method of moments and multi-conductor transmission lines. To verify the effectiveness of the dual-view structure, different input methods, including TopOnly, SideOnly, FullView, and DualView, were compared. In this diagram, TopOnly indicates input using only the top view, SideOnly indicates input using only the side view, FullView indicates input using a complete stitched view, and DualView indicates input using the dual-view branch input and feature fusion method of this invention. Experimental results show that, under the conditions of 0.8GHz to 2GHz frequency band and 192×192 image resolution, the DualView model achieves a comprehensive coefficient of determination R² of 0.9150 and a mean absolute percentage error (MAPE) of 19.87%, indicating that using a combination of top and side views as input and performing dual-branch feature fusion can improve the model's ability to utilize the three-dimensional geometric information of random cable bundles and enhance the prediction effect of broadband common-mode current curves at the source and load ends.

[0053] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0054] Based on the same inventive concept, this application also provides a system for rapidly predicting broadband common-mode current of random cable bundles to implement the aforementioned method for rapidly predicting broadband common-mode current of random cable bundles. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for rapidly predicting broadband common-mode current of random cable bundles provided below can be found in the limitations of the method for rapidly predicting broadband common-mode current of random cable bundles described above, and will not be repeated here.

[0055] In one embodiment, such as Figure 6 As shown, a fast prediction system 600 for broadband common-mode current of random cable bundles is provided, including: a geometric sample acquisition module 601, a physical reference label generation module 602, a dual-view mapping module 603, a model building module 604, a model training module 605, and a fast prediction module 606, wherein: The geometric sample acquisition module 601 is used to acquire three-dimensional geometric samples of random cable bundles to be analyzed. The three-dimensional geometric samples of random cable bundles include three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions.

[0056] The physical reference tag generation module 602 is used to obtain the source-side common-mode current and load-side common-mode current at each discrete frequency point by jointly solving the method of moments and multi-conductor transmission lines, and to combine the source-side common-mode current and load-side common-mode current at each discrete frequency point into a broadband common-mode current reference tag.

[0057] The dual-view mapping module 603 is used to map the three-dimensional geometric sample of the random cable bundle into a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference.

[0058] The model building module 604 is used to construct a dual-view convolutional neural network. The dual-view convolutional neural network includes a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel map and the side-view geometric pixel map are respectively input into the top-view encoding branch and the side-view encoding branch to extract top-view features and side-view features. The feature fusion module fuses the top-view features and the side-view features and then inputs them into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point.

[0059] The model training module 605 is used to train the dual-view convolutional neural network using the wideband common-mode current reference label as supervision.

[0060] The fast prediction module 606 is used to acquire the three-dimensional geometric data of the new random cable bundle sample to be predicted, and construct its top view geometric pixel map and side view geometric pixel map according to the mapping method consistent with the training phase. The top view geometric pixel map and the side view geometric pixel map are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves of the source end and load end of the new random cable bundle sample.

[0061] The modules in the aforementioned broadband common-mode current rapid prediction system for random cable bundles can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0062] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for rapid prediction of broadband common-mode current in random cable bundles. The display screen can be an LCD screen or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0063] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0064] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0065] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0066] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0067] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for rapid prediction of broadband common-mode current in random cable bundles, characterized in that, The method includes: Obtain a three-dimensional geometric sample of a random cable bundle to be analyzed. The three-dimensional geometric sample of the random cable bundle includes three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions. The source-side common-mode current and load-side common-mode current at each discrete frequency point are obtained by jointly solving the method of moments and multi-conductor transmission lines, and the source-side common-mode current and load-side common-mode current at each discrete frequency point are combined into a broadband common-mode current reference label. The three-dimensional geometric samples of the random cable bundle are mapped into a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference; A dual-view convolutional neural network is constructed, comprising a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel map and the side-view geometric pixel map are respectively input into the top-view encoding branch and the side-view encoding branch to extract top-view features and side-view features. The feature fusion module fuses the top-view features and the side-view features and then inputs them into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point. The dual-view convolutional neural network is trained using the wideband common-mode current reference label as supervision. The three-dimensional geometric data of the new random cable bundle sample to be predicted is obtained, and its top view geometric pixel map and side view geometric pixel map are constructed according to the mapping method consistent with the training phase. The top view geometric pixel map and the side view geometric pixel map are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves of the source end and load end of the new random cable bundle sample.

2. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The top view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the first transverse direction, and the side view geometric pixel image is a projection image of the random cable bundle three-dimensional geometric sample on a plane defined by the axial direction and the second transverse direction, wherein the first transverse direction and the second transverse direction are orthogonal to each other.

3. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The step of mapping the three-dimensional geometric samples of the random cable bundle into top view geometric pixel maps and side view geometric pixel maps further includes: Each conductor is assigned to an independent image channel, and the channel corresponding to the excitation conductor is fixed at a preset channel position.

4. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The top view encoding branch and the side view encoding branch adopt convolutional encoders with independent parameters and identical structures. Each convolutional encoder includes an initial convolutional layer, multiple cascaded residual downsampling modules and an adaptive average pooling layer. The residual downsampling module includes a convolution operation unit, a normalization processing unit, an activation function unit, a channel attention mechanism unit and a residual skip connection unit.

5. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The feature fusion module fuses the top view features and the side view features, including at least constructing the original feature terms, difference terms, element-wise product terms and mean terms between the top view features and the side view features, and combining the original feature terms, the difference terms, the element-wise product terms and the mean terms into the fused features.

6. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The dual-output regression module includes a shared mapping layer, a load-side regression head, and a source-side regression head. After fusion, the features are processed by the shared mapping layer and then input to the load-side regression head and the source-side regression head, respectively. The load-side regression head outputs the load-side common-mode current prediction curve at each discrete frequency point, and the source-side regression head outputs the source-side common-mode current prediction curve at each discrete frequency point.

7. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The step of training the dual-view convolutional neural network using the broadband common-mode current reference label as supervision includes: The wideband common-mode current reference label is subjected to logarithmic transformation and standardization. The dual-view convolutional neural network outputs the standardized logarithmic domain prediction value, which is then restored to the amplitude domain common-mode current prediction curve through inverse standardization and inverse logarithmic transformation.

8. The method for rapid prediction of broadband common-mode current in random cable bundles according to claim 1, characterized in that, The step of training the dual-view convolutional neural network using the broadband common-mode current reference label as supervision further includes: The dual-view convolutional neural network is trained using an enhanced combined loss function, which includes a logarithmic domain robust loss term, an amplitude domain relative error loss term, a spectral slope consistency loss term, a peak region loss term, and a low current region loss term.

9. A fast prediction system for broadband common-mode current of random cable bundles, characterized in that, The system includes: The geometric sample acquisition module is used to acquire three-dimensional geometric samples of random cable bundles to be analyzed. The three-dimensional geometric samples of random cable bundles include three-dimensional coordinate data of the center lines of multiple conductors at multiple axial positions. The physical reference label generation module is used to obtain the source-side common-mode current and load-side common-mode current at each discrete frequency point by jointly solving the method of moments and multi-conductor transmission lines, and to combine the source-side common-mode current and load-side common-mode current at each discrete frequency point into a broadband common-mode current reference label. The dual-view mapping module is used to map the three-dimensional geometric samples of the random cable bundle into a top view geometric pixel map and a side view geometric pixel map, wherein the top view geometric pixel map and the side view geometric pixel map use the same axial sampling reference. The model building module is used to construct a dual-view convolutional neural network. The dual-view convolutional neural network includes a top-view encoding branch, a side-view encoding branch, a feature fusion module, and a dual-output regression module. The top-view geometric pixel map and the side-view geometric pixel map are respectively input into the top-view encoding branch and the side-view encoding branch to extract top-view features and side-view features. The feature fusion module fuses the top-view features and the side-view features and then inputs them into the dual-output regression module. The dual-output regression module outputs the source-side common-mode current prediction curve and the load-side common-mode current prediction curve at each discrete frequency point. The model training module is used to train the dual-view convolutional neural network using the wideband common-mode current reference label as supervision. The fast prediction module is used to acquire the three-dimensional geometric data of the new random cable bundle sample to be predicted, and construct its top view geometric pixel map and side view geometric pixel map according to the mapping method consistent with the training phase. The top view geometric pixel map and the side view geometric pixel map are input into the trained dual-view convolutional neural network to obtain the broadband common-mode current prediction curves of the source end and load end of the new random cable bundle sample.

Images