Method for predicting performance of multiple-input multiple-output antenna decoupling based on deep learning

Abstract

The application discloses a multiple-input multiple-output antenna decoupling performance prediction method based on deep learning, belongs to the technical field of computer-aided electromagnetic system design and prediction, is used for antenna decoupling performance prediction, and comprises the following steps: determining key size parameters of an antenna to be optimized; performing parameter scanning on the variation range of the key size parameters to obtain full-band S parameter simulation data and form an original data set; the data amount of the original data set is expanded by using an SMOTE algorithm to generate an enhanced data set; an antenna decoupling performance prediction model is constructed and trained, the antenna to be optimized is input, and corresponding full-band S parameter curve final prediction results are output. The application effectively expands high-dimensional data by introducing the SMOTE algorithm, solves the problem that small sample data in the electromagnetic simulation field cannot support deep learning model training, constructs a DNN model to establish a nonlinear mapping of a continuous S parameter curve, and can comprehensively evaluate the S parameters of the antenna, instead of being limited to discrete frequency points.

Description

Technical Field

[0001] This invention discloses a method for predicting the decoupling performance of multi-input multi-output antennas based on deep learning, belonging to the field of computer-aided electromagnetic system design and prediction technology. Background Technology

[0002] Multiple-input multiple-output (MIMO) technology is a key technology for improving the capacity and reliability of modern wireless communication systems. In MIMO antenna design, the mutual coupling effect between antenna elements can severely affect system performance, especially in compact devices. Therefore, it is crucial to introduce decoupling structures (such as coupling neutral lines) and optimize their dimensions during the design process.

[0003] The industry has attempted to use surrogate models to establish an approximate mapping relationship between antenna structural parameters and electromagnetic performance, in order to replace some simulation calculations. However, traditional surrogate models, such as polynomial response surfaces and Kriging models, often lack sufficient fitting ability and accuracy when dealing with the mapping relationship from low-dimensional parameters to high-dimensional, strongly nonlinear full-band continuous S-parameter curves, making it difficult to meet the engineering requirements of high-precision antenna design.

[0004] While deep neural networks (DNNs) possess powerful potential for nonlinear fitting, their application in antenna design faces a dual challenge: Firstly, training requires massive amounts of data, but acquiring full-wave simulation data is extremely time-consuming, with single simulations typically taking tens of minutes to hours, and large-scale parameter scans requiring continuous computation for tens of hours or even days. This creates a sharp contradiction between the available data (usually only a few dozen sets) and the model requirements (usually hundreds to thousands of sets). Secondly, existing DNN structures present the following technical difficulties for the specific needs of antenna prediction: ① Directly applying general-purpose DNNs to small-sample antenna data easily leads to overfitting and poor model generalization ability; ② A specially designed network structure suitable for full-band continuous response prediction is needed; ③ A neural network output layer capable of outputting continuous real values ​​is required to accurately fit the normalized S-parameter data. Therefore, there is an urgent need in this field for a new method capable of achieving high-precision, second-level prediction of MIMO antenna decoupling performance across the entire frequency band under small-sample conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for predicting the decoupling performance of multi-input multi-output antennas, in order to solve the problems in the prior art, where directly applying a general DNN to small sample antenna data is prone to overfitting, the model has poor generalization ability, and a network structure suitable for predicting continuous response across the entire frequency band needs to be specially designed. It also requires designing a neural network output layer that can output continuous real values ​​to accurately fit the normalized S-parameter data.

[0006] Deep learning-based methods for predicting the decoupling performance of multi-input multi-output antennas include: S1. Determine the key dimensional parameters of the antenna to be optimized; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned to obtain full-band S-parameter simulation data under different parameter combinations, which constitutes the original dataset. S3. Use a synthetic minority class oversampling algorithm to expand the data volume of the original dataset and generate an enhanced dataset; S4. Construct an antenna decoupling performance prediction model. Use the augmented dataset as input to train the neural network. Set a loss round threshold and use an early stopping mechanism to constrain the neural network model. After terminating the training, the trained antenna decoupling performance prediction model is obtained. S5. Generate an enhanced dataset from the antenna to be optimized through steps S1, S2 and S3. Input the enhanced dataset into the trained antenna decoupling performance prediction model and output the final prediction result of the corresponding full-band S-parameter curve.

[0007] S1 includes selecting no fewer than three adjustable geometric dimension parameters as key dimension parameters from among the adjustable geometric dimension parameters that can cause changes in the coupling strength between antenna elements when changed, and setting other dimension parameters to fixed values ​​except for the key dimension parameters.

[0008] S2 includes, S2.1, setting the preset frequency band as... arrive Set a minimum sample point threshold. By using full-wave electromagnetic simulation software to scan the preset parameter variation range of key dimensional parameters, the following results were obtained. The simulation data of full-band S-parameters under different parameter combinations constitute the original dataset; Each set of data includes key dimension parameter values ​​and S-parameter curves of the key dimension parameter values ​​within a preset frequency band. The S-parameter curves include S11 parameters, which characterize the antenna operating frequency band, and S21 parameters, which characterize the antenna port isolation.

[0009] S2 includes S2.2, standardizing the original dataset, including removing parameter combinations with missing values; Set the normalization interval for key dimension parameters and normalize the key dimension parameters. Set the normalization interval for full-band S-parameters and normalize the full-band S-parameters.

[0010] S3 includes setting a minimum threshold for the expansion factor. and the maximum threshold of expansion factor A synthetic minority class oversampling algorithm is used to expand the data volume of the original dataset. times, .

[0011] The synthetic minority class oversampling algorithm includes feature space clustering, hierarchical generation of synthetic samples, and boundary protection and random shuffling; Feature space clustering involves calculating the Euclidean distance between each sample point in the original dataset and all other sample points. , and These are all indices of the original dataset. ; Set the nearest neighbor sample threshold Each sample point is selected from the nearest... One sample point; Stratified sampling includes, for each original sample point that needs to be augmented. ,from Randomly select a sample point from the nearest neighbor sample points. ; Linear interpolation for generating synthetic samples includes, in and On the connection between, based on New synthetic samples are generated through linear interpolation. and the corresponding S-parameter output : ; ; In the formula, for Random numbers within the interval for The corresponding S-parameter output; Boundary protection includes, if Any dimension exceeding the preset parameter variation range will be removed. ; Random shuffling involves merging sample points from the original dataset with the generated synthetic samples and randomly shuffling their order to create an augmented dataset.

[0012] S4 includes an antenna decoupling performance prediction model that is a feedforward neural network, comprising an input layer, three processing units, and an output layer. The number of nodes in the input layer corresponds to the number of key size parameters. The output layer is a fully connected layer, and the number of nodes in the output layer is set using a linear activation function. ; Each processing unit consists of a fully connected layer, a batch normalized layer, and a LeakyReLU activation function.

[0013] Batch standardization layers perform implicit regularization by suppressing covariates; Setting the slope of the LeakyReLU activation function in the negative interval preserves signal characteristics.

[0014] The antenna decoupling performance prediction model employs a dynamic learning rate decay strategy, setting a training epoch threshold. When the training epoch reaches the threshold, the learning rate decreases. Attenuation to ; Set the threshold for the number of loss rounds. The antenna decoupling performance prediction model employs an early stopping mechanism, whereby the output loss of the antenna decoupling performance prediction model is continuous. Training will automatically terminate if the number of rounds does not increase.

[0015] The Adam optimization algorithm is used to forward calculate and predict the full-band S-parameter curve. The mean square error function is used to calculate the deviation between the predicted and true values ​​of the full-band S-parameter curve. The gradient is calculated and the network weights are updated through backpropagation.

[0016] Compared to existing technologies, this invention offers the following advantages: By using the SMOTE algorithm for intelligent interpolation in the feature space, synthetic samples consistent with the original data distribution are generated, effectively expanding a small amount of simulation data to the scale required for model training. This overcomes the data barrier for deep learning applications in the electromagnetic field, making it possible to build high-precision models with limited cost. By directly setting the number of nodes in the DNN output layer to the number of sampling points for the full-band S-parameters, an end-to-end mapping from low-dimensional size parameters to high-dimensional continuous frequency response curves is established. This allows the model to output complete S-parameter curves at once, achieving comprehensive and continuous prediction of antenna broadband performance, with evaluation capabilities far exceeding traditional methods. The introduction of a batch normalization layer in the network stabilizes the inter-layer input distribution and accelerates training convergence. The use of the LeakyReLU activation function instead of the traditional ReLU ensures that small gradients in the negative interval can still propagate. This design ensures the stability and efficiency of model training on small sample data and improves its ability to accurately fit complex electromagnetic mapping relationships; it shortens the single performance evaluation from simulation time of "minutes" to prediction time of "seconds", making rapid and large-scale parameter scanning possible, and shortening the decoupling design cycle of MIMO antennas from tens of hours to less than twenty minutes. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the S11 feature spatial distribution map; Figure 3 This is a comparison chart of S11 parameter distribution; Figure 4 This is a comparison image of the S11 enhanced sample and the original sample; Figure 5 This is the spatial distribution map of S21 features; Figure 6 This is a comparison chart of S21 parameter distribution; Figure 7 This is a comparison image of the S21 enhanced sample and the original sample; Figure 8 This is the DNN network structure of the present invention; Figure 9 This is a comparison chart of the prediction results of the model of this invention and the simulation results of CST; Figure 10 Frequency domain error and Distribution map; Figure 11 It is a correlation prediction graph. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Deep learning-based methods for predicting the decoupling performance of multi-input multi-output antennas include: S1. Determine the key dimensional parameters of the antenna to be optimized; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned to obtain full-band S-parameter simulation data under different parameter combinations, which constitutes the original dataset. S3. Use a synthetic minority class oversampling algorithm to expand the data volume of the original dataset and generate an enhanced dataset; S4. Construct an antenna decoupling performance prediction model. Use the augmented dataset as input to train the neural network. Set a loss round threshold and use an early stopping mechanism to constrain the neural network model. After terminating the training, the trained antenna decoupling performance prediction model is obtained. S5. Generate an enhanced dataset from the antenna to be optimized through steps S1, S2 and S3. Input the enhanced dataset into the trained antenna decoupling performance prediction model and output the final prediction result of the corresponding full-band S-parameter curve.

[0020] S1 includes selecting no fewer than three adjustable geometric dimension parameters as key dimension parameters from among the adjustable geometric dimension parameters that can cause changes in the coupling strength between antenna elements when changed, and setting other dimension parameters to fixed values ​​except for the key dimension parameters.

[0021] S2 includes, S2.1, setting the preset frequency band as... arrive Set a minimum sample point threshold. By using full-wave electromagnetic simulation software to scan the preset parameter variation range of key dimensional parameters, the following results were obtained. The simulation data of full-band S-parameters under different parameter combinations constitute the original dataset; Each set of data includes key dimension parameter values ​​and S-parameter curves of the key dimension parameter values ​​within a preset frequency band. The S-parameter curves include S11 parameters, which characterize the antenna operating frequency band, and S21 parameters, which characterize the antenna port isolation.

[0022] S2 includes S2.2, standardizing the original dataset, including removing parameter combinations with missing values; Set the normalization interval for key dimension parameters and normalize the key dimension parameters. Set the normalization interval for full-band S-parameters and normalize the full-band S-parameters.

[0023] S3 includes setting a minimum threshold for the expansion factor. and the maximum threshold of expansion factor A synthetic minority class oversampling algorithm is used to expand the data volume of the original dataset. times, .

[0024] The synthetic minority class oversampling algorithm includes feature space clustering, hierarchical generation of synthetic samples, and boundary protection and random shuffling; Feature space clustering involves calculating the Euclidean distance between each sample point in the original dataset and all other sample points. , and These are all indices of the original dataset. ; Set the nearest neighbor sample threshold Each sample point is selected from the nearest... One sample point; Stratified sampling includes, for each original sample point that needs to be augmented. ,from Randomly select a sample point from the nearest neighbor sample points. ; Linear interpolation for generating synthetic samples includes, in and On the connection between, based on New synthetic samples are generated through linear interpolation. and the corresponding S-parameter output : ; ; In the formula, for Random numbers within the interval for The corresponding S-parameter output; Boundary protection includes, if Any dimension exceeding the preset parameter variation range will be removed. ; Random shuffling involves merging sample points from the original dataset with the generated synthetic samples and randomly shuffling their order to create an augmented dataset.

[0025] S4 includes an antenna decoupling performance prediction model that is a feedforward neural network, comprising an input layer, three processing units, and an output layer. The number of nodes in the input layer corresponds to the number of key size parameters. The output layer is a fully connected layer, and the number of nodes in the output layer is set using a linear activation function. ; Each processing unit consists of a fully connected layer, a batch normalized layer, and a LeakyReLU activation function.

[0026] Batch standardization layers perform implicit regularization by suppressing covariates; Setting the slope of the LeakyReLU activation function in the negative interval preserves signal characteristics.

[0027] The antenna decoupling performance prediction model employs a dynamic learning rate decay strategy, setting a training epoch threshold. When the training epoch reaches the threshold, the learning rate decreases. Attenuation to ; Set the threshold for the number of loss rounds. The antenna decoupling performance prediction model employs an early stopping mechanism, whereby the output loss of the antenna decoupling performance prediction model is continuous. Training will automatically terminate if the number of rounds does not increase.

[0028] The Adam optimization algorithm is used to forward calculate and predict the full-band S-parameter curve. The mean square error function is used to calculate the deviation between the predicted and true values ​​of the full-band S-parameter curve. The gradient is calculated and the network weights are updated through backpropagation.

[0029] The following description, in conjunction with the accompanying drawings and embodiments, provides further details. The antenna used in this embodiment is a performance prediction of an improved parasitic stub-fed dual-frequency inverted-F antenna array with a decoupling structure. The antenna array includes two back-to-back improved parasitic stub-fed inverted-F antenna elements and a U-shaped coupling neutralization line connecting the two antenna elements. The full-wave electromagnetic simulation software used is CST Microwave Studio; the preset frequency band is [missing information]. , Set key dimensional parameters, including the distance between the short arm and the open arm. Length of short-circuit stubs in coupled neutral lines and antenna horizontal segment length For key parameters, the length b of the coupled feeder arm, the outer length h of the U-shaped stub of the coupled neutralization line, and the height L1 of the open arm of the improved inverted F array element are set to fixed values. The preset range of variation for these three key dimensions is as follows: Fix other parameters as , Take 8 equally spaced values ​​within the range of 3mm to 7mm. Six equally spaced values ​​were selected within the range of 1.3mm to 4.3mm. Six equally spaced values ​​were selected within the range of 11mm to 16mm, resulting in a total of 48 parameter combinations. Set the normalization interval for key dimension parameters to: The normalization range for the full-band S-parameters is set to -6dB to the minimum value of the full-band S-parameters, and the expansion factor threshold is [value missing]. , The training round threshold is 100 rounds. .

[0030] The process of this invention is as follows Figure 1 As shown, the process is divided into two stages: offline model building (stage 1) and online rapid prediction (stage 2). In stage 1, the antenna and decoupling structure are first defined parametrically, and a finite number of CST simulation scans are performed to obtain an initial dataset (size, full-band response). Then, SMOTE enhancement is performed to generate an enhanced dataset (data volume increased k times). A DNN model is then built and trained, and validated and optimized. If the model accuracy is insufficient, validation and optimization continue; if the model accuracy is sufficient, a high-precision surrogate model is obtained. In stage 2, the antenna size parameters to be evaluated are input, the model performs second-level calculations, and outputs the full-band S-parameter prediction curve, completing the performance evaluation.

[0031] The simulation took approximately 24 hours to complete. The SMOTE oversampling algorithm was used to augment the 48 sets of original data. First, cluster analysis was performed on the multidimensional feature space containing dimensional parameters and S-parameter curves, including calculating the Euclidean distance between each sample point in the original dataset and all other sample points. : ; In the formula, and These are all indices of the original dataset. , For the first sample points , For the first sample points , For the first sample points , For the first sample points , For the first sample points , For the first sample points ; Then, new synthetic samples are generated within each category using linear interpolation. With an augmentation factor of 5, a final augmented dataset containing 240 samples is generated. Figures 2 to 7 The distribution comparison of the S-parameter curves before and after data augmentation is shown, indicating that the distribution of the augmented data is basically consistent with the original data. Among them, Figure 2 The distribution of the original S11 parameter samples and the SMOTE enhanced samples in the dimensionality reduction space of principal component analysis is shown. The enhanced samples are distributed around the original samples, filling the gaps between the original samples, and no obvious outliers are found. Figure 3 The numerical distribution of the S11 parameter in the original sample and the enhanced sample at different frequencies is compared. The value range of the enhanced sample at each frequency point is basically the same as that of the original sample, and the coverage range has not been significantly expanded. Figure 4 It is a curve comparison between some enhanced samples in the S11 parameters and the original sample with the closest Euclidean distance. The curve trend of the enhanced sample is consistent with the overall trend of the corresponding original sample. Figure 5 The distribution of the original and enhanced samples of parameter S21 in the dimensionality reduction space of principal component analysis shows that the enhanced samples are evenly distributed around the original samples, with no distribution shift or abnormal clustering. Figure 6 The numerical distribution of the S21 parameter in the original sample and the enhanced sample at different frequencies is compared. The value range of the enhanced sample at each frequency point is highly consistent with that of the original sample, and no outliers exceeding the original range are found. Figure 7 The curves of some enhanced samples in the S21 parameters are compared with the original samples whose Euclidean distance is closest. The variation trends of the enhanced samples and the original samples in the frequency band are basically consistent, and the amplitude differences at key frequency points are small.

[0032] Construct a deep neural network model. For example... Figure 8As shown, the model includes an input layer, three processing units, and an output layer. The input layer has 3 nodes, corresponding to three size parameters. The first processing unit expands the 3-dimensionality to 128 feature channels through a fully connected layer; the second processing unit expands the 128-dimensionality to 256-dimensionality through a fully connected layer; and the third processing unit expands the 256-dimensionality to 512-dimensionality through a fully connected layer. The first processing unit includes a fully connected layer (output dimension 128), a batch normalization layer, and a LeakyReLU activation function (slope 0.2 in the negative interval). The second processing unit includes a fully connected layer (output dimension 256), a batch normalization layer, and a LeakyReLU activation function. The third processing unit includes a fully connected layer (output dimension 512), a batch normalization layer, and a LeakyReLU activation function. The output layer is a fully connected layer with an output dimension of 1000, corresponding to 1000 S-parameter values ​​at different frequency points. A linear activation function is used to preserve the positive and negative characteristics of the S-parameters.

[0033] A deep neural network model was trained using an augmented dataset. 240 samples were randomly divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer was used, with an initial learning rate of 0.0005, which decreased by 50% every 100 training epochs. The batch size was set to 16. Mean squared error was used as the loss function. An early stopping mechanism was implemented, automatically terminating training when the validation set loss showed no improvement for 20 consecutive epochs. The model converged after approximately 150 epochs.

[0034] Use the trained model for prediction. Given any set of size parameters... By inputting the model, the corresponding S11 and S21 full-band curves can be output within seconds.

[0035] To verify the model's performance, it was evaluated on the test set. Figure 9 The comparison between the predicted curve and the CST simulation curve for a certain sample is shown, and the two have a high degree of agreement. Figure 10 The prediction error distribution and coefficient of determination are shown for the entire frequency band (3 GHz to 8 GHz). Distribution, average More than 0.900. Figure 11 A correlation plot between predicted and true values, showing the training, validation, and test sets. The values ​​were 0.892, 0.828, and 0.826, respectively, indicating that the model has good generalization ability and no overfitting occurred.

[0036] Other alternatives to this invention include, in this embodiment, the preferred method for data augmentation is the SMOTE oversampling algorithm. It is understood that any method capable of generating reasonable synthetic data from small samples that is consistent with the original data distribution can be used as an alternative to the SMOTE algorithm. Examples include ADASYN (an adaptive oversampling method that generates more synthetic data based on the learning difficulty of the samples), Borderline-SMOTE (an improved SMOTE algorithm that focuses on generating synthetic samples near class boundaries), and data augmentation methods based on generative adversarial networks (generating high-quality synthetic simulation data through adversarial training). The core objective of these alternative methods is consistent with SMOTE, namely, to effectively expand the training dataset without increasing the actual simulation cost.

[0037] This invention employs a specific DNN structure with multiple fully connected layers, batch normalization, and the LeakyReLU activation function. It is understood that any deep neural network architecture capable of establishing a complex nonlinear mapping from low-dimensional input to high-dimensional continuous output can achieve the objectives of this invention. For example, fully connected neural networks with different depths (number of layers) and widths (number of neurons per layer), fully connected networks incorporating residual connections (to mitigate gradient problems that deeper networks may introduce), using activation functions with self-normalization properties such as SELU and Mish instead of LeakyReLU, and introducing appropriate Dropout layers into the network as a supplementary or alternative means to prevent overfitting. The core of these alternative architectures lies in ensuring that the number of nodes in the output layer matches the number of sampling points on the target full-band S-parameter curve to achieve end-to-end continuous frequency response prediction.

[0038] This invention focuses on predicting the isolation S21 curve. It can be understood that the same methodology is fully applicable to the full-band prediction of other key antenna performance indicators, such as Voltage Standing Wave Ratio (VSWR), gain, efficiency, and ECC. Simply replace the output data with the corresponding target curve during the data preparation phase and retrain the network.

[0039] This invention uses the specific structural combination of an inverted-F antenna and a coupled neutralization line as an example to verify the effectiveness of the prediction method. It is understood that the prediction method system based on data augmentation and deep neural networks proposed in this invention is not limited to this application. This method has universality and can be widely applied to various antenna and decoupling structure design scenarios requiring rapid performance evaluation and optimization, including other types of antenna elements such as patch antennas, monopole antennas, PIFA antennas, and dipole antennas; other forms of MIMO decoupling structures such as defective ground structures, parasitic elements, decoupling networks, and electromagnetic bandgap structures; and more complex antenna arrays, such as large-scale MIMO arrays and phased arrays, where key structural parameters of their elements or arrays can be used as input. The core of the above alternative application scenarios lies in using key adjustable structural parameters affecting antenna performance as model input and the desired full-band electromagnetic response as model output. The prediction method of this invention provides a general and efficient solution for this wide range of structure and performance mapping problems.

[0040] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that, in addition to the parameters of the above embodiments, other geometric dimensional parameters that can affect the coupling strength are also applicable to this method. They can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for predicting the decoupling performance of multi-input multi-output antennas, characterized in that, include: S1. Determine the key dimensional parameters of the antenna to be optimized; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned to obtain full-band S-parameter simulation data under different parameter combinations, which constitutes the original dataset. S3. Use a synthetic minority class oversampling algorithm to expand the data volume of the original dataset and generate an enhanced dataset; S4. Construct an antenna decoupling performance prediction model. Use the augmented dataset as input to train the neural network. Set a loss round threshold and use an early stopping mechanism to constrain the neural network model. After terminating the training, the trained antenna decoupling performance prediction model is obtained. S5. Generate an enhanced dataset from the antenna to be optimized through steps S1, S2 and S3. Input the enhanced dataset into the trained antenna decoupling performance prediction model and output the corresponding full-band S-parameter curve final prediction result. S1 includes selecting no less than 3 adjustable geometric dimension parameters as key dimension parameters from the adjustable geometric dimension parameters that can cause changes in the coupling strength between antenna elements when changed, and setting other dimension parameters to fixed values ​​except for the key dimension parameters; S2 includes, S2.1, setting the preset frequency band as... arrive Set a minimum sample point threshold. By using full-wave electromagnetic simulation software to scan the preset parameter variation range of key dimensional parameters, the following results were obtained. The simulation data of full-band S-parameters under different parameter combinations constitute the original dataset; Each set of data includes key dimension parameter values ​​and S-parameter curves of key dimension parameter values ​​within a preset frequency band. The S-parameter curves include S11 parameters characterizing the antenna operating frequency band and S21 parameters characterizing the antenna port isolation. S2 includes S2.2, standardizing the original dataset, including removing parameter combinations with missing values; Set the normalization interval for key dimension parameters and normalize the key dimension parameters; set the normalization interval for full-band S-parameters and normalize the full-band S-parameters. S3 includes setting a minimum threshold for the expansion factor. and the maximum threshold of expansion factor A synthetic minority class oversampling algorithm is used to expand the data volume of the original dataset. times, ; The synthetic minority class oversampling algorithm includes feature space clustering, hierarchical generation of synthetic samples, and boundary protection and random shuffling; Feature space clustering This includes calculating the Euclidean distance between each sample point in the original dataset and all other sample points. , and These are all indices of the original dataset. ; Set the nearest neighbor sample threshold Each sample point is selected from the nearest... One sample point; Stratified sampling includes, for each original sample point that needs to be augmented. ,from Randomly select a sample point from the nearest neighbor sample points. ; Linear interpolation for generating synthetic samples includes, in and On the connection between, based on New synthetic samples are generated through linear interpolation. and the corresponding S-parameter output : ； ； In the formula, for Random numbers within the interval for The corresponding S-parameter output; Boundary protection includes, if Any dimension exceeding the preset parameter variation range will be removed. ; Random shuffling involves merging sample points from the original dataset with the generated synthetic samples and randomly shuffling their order to generate an augmented dataset. S4 includes an antenna decoupling performance prediction model that is a feedforward neural network, comprising an input layer, three processing units, and an output layer. The number of nodes in the input layer corresponds to the number of key size parameters. The output layer is a fully connected layer, and the number of nodes in the output layer is set using a linear activation function. ; Each processing unit consists of a fully connected layer, a batch normalized layer, and a LeakyReLU activation function. Batch standardization layers perform implicit regularization by suppressing covariates; Set the slope of the LeakyReLU activation function in the negative interval to preserve signal characteristics; The antenna decoupling performance prediction model employs a dynamic learning rate decay strategy, setting a training epoch threshold. When the training epoch reaches the threshold, the learning rate decreases. Attenuation to ; Set the threshold for the number of loss rounds. The antenna decoupling performance prediction model employs an early stopping mechanism, whereby the output loss of the antenna decoupling performance prediction model is continuous. Training will automatically terminate if the number of rounds does not improve. The Adam optimization algorithm is used to forward calculate and predict the full-band S-parameter curve. The mean square error function is used to calculate the deviation between the predicted and true values ​​of the full-band S-parameter curve. The gradient is calculated and the network weights are updated through backpropagation.

Images