Dual-frequency MIMO antenna optimization method based on weighted KNN

Abstract

The application discloses a double-frequency MIMO antenna optimization method based on weighted KNN, belongs to the technical field of machine learning aided engineering design, and is used for double-frequency MIMO antenna optimization, and comprises the following steps: acquiring key size parameters, obtaining S11 and S21 parameter curves through full-wave electromagnetic simulation software and constructing original samples; screening effective samples to construct a training data set; constructing a weighted KNN proxy model to calculate S parameter prediction values of to-be-predicted samples, calculating a determination coefficient to adjust the total number of nearest neighbors; generating candidate samples, inputting the trained weighted KNN proxy model, calculating evaluation index weights by using a coefficient of variation method; and calculating a comprehensive score function to screen an optimal antenna design scheme. The application replaces full-wave electromagnetic simulation with the weighted KNN proxy model to realize rapid prediction, reduces the S parameter evaluation time of the candidate structure from the order of tens of minutes to the order of milliseconds, avoids the subjectivity of artificial weighting, and realizes efficient and automatic optimization and screening of antenna structure parameters.

Description

Technical Field

[0001] This invention discloses a dual-frequency MIMO antenna optimization method based on weighted KNN, belonging to the field of machine learning-aided engineering design technology. Background Technology

[0002] With the rapid development of wireless communication technology, Multiple-Input Multiple-Output (MIMO) technology has become a key technology in modern communication systems due to its ability to significantly improve system capacity and spectrum utilization. In MIMO antenna design, inter-port coupling is one of the core issues affecting system performance, and decoupling techniques such as neutral lines and defective ground structures are commonly used to improve isolation. However, the mapping relationship between antenna structural parameters and electromagnetic performance is complex. Traditional design optimization methods heavily rely on the designer's experience, using full-wave electromagnetic simulation software to perform point-by-point parameter scanning to obtain the optimal structural parameters. Since a single full-wave simulation takes several minutes to tens of minutes, the total time for traversing multiple parameter combinations often reaches tens of hours, resulting in long design cycles, high computational resource consumption, and difficulty in exploring a broad design space within a limited timeframe.

[0003] In recent years, researchers have introduced machine learning methods into the field of antenna design, using models such as neural networks and Gaussian processes to establish the mapping relationship between structural parameters and S-parameters to replace full-wave simulation. However, existing methods still have the following shortcomings: (1) Neural network models require a large number of training samples, while antenna simulation data acquisition is costly, and overfitting is prone to occur under small sample conditions, making it difficult to guarantee prediction accuracy; (2) Antenna optimization involves multiple mutually constraining evaluation indicators such as isolation, return loss, and size compactness. Existing methods usually rely on manually setting weights in multi-objective evaluation, which is highly subjective and difficult to take into account the discrete characteristics of different indicators; (3) When the number of parameters to be optimized is large, the model complexity increases sharply, limiting its application in practical engineering. In response to the above problems, this invention provides a dual-frequency MIMO antenna optimization method based on a weighted K-nearest neighbor proxy model, which can achieve high-precision prediction under small sample conditions without iterative training, and automatically determines the weight of each evaluation indicator through the coefficient of variation method, realizing objective and efficient multi-objective structural parameter optimization and screening. Summary of the Invention

[0004] The purpose of this invention is to provide a dual-frequency MIMO antenna optimization method based on weighted KNN, so as to solve the problems in the prior art, such as large computational load of full-wave simulation, insufficient prediction accuracy of machine learning models under small sample conditions, and strong subjectivity of multi-objective evaluation relying on manual weighting.

[0005] The dual-band MIMO antenna optimization method based on weighted KNN includes: S1. Obtain the key dimensional parameters of the inverted L-shaped and monopole composite dual-band antenna; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned using combined parameters to obtain the parameter curves of S11 and S21, and the original sample is constructed. S3. Calculate the average value of the S21 parameter curve for each group of samples within the target working frequency band, set an effective sample threshold to filter effective samples and construct a training dataset, and standardize the key size parameters in the training dataset to zero mean. S4. Construct a weighted KNN surrogate model. First, calculate the Euclidean distance between the sample to be predicted and the effective sample. Set the total number of nearest neighbors to filter the nearest neighbor samples. Calculate the weight of the nearest neighbor sample based on the Euclidean distance and normalize it. Combine the true value of the S-parameter of the effective sample to calculate the predicted value of the S-parameter of the sample to be predicted. S5. Calculate the determination coefficient of the predicted S-parameter values ​​of the sample to be predicted, adjust the total number of nearest neighbors, select the total number of nearest neighbors with the highest determination coefficient, and obtain the trained weighted KNN surrogate model. S6. Within the range of values ​​for the four key size parameters, generate several sets of candidate key size parameters according to a preset step size as candidate samples; input each set of candidate samples into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves; set the worst value of return loss in the target frequency band, the worst value of isolation in the target frequency band, and the sum of the key size parameters as evaluation indicators, and calculate the weight of the evaluation indicators using the coefficient of variation method; S7. Set the worst-case threshold values ​​for return loss and isolation within the target frequency band as hard constraints. Construct a comprehensive scoring function using the weights and normalized values ​​of the evaluation indicators. Sort the comprehensive scores from smallest to largest, and determine the smallest score as the optimal antenna design scheme.

[0006] S1 includes key dimensional parameters such as the feed gap between the inverted L antenna and the monopole antenna, the spacing between the inverted L antenna and the monopole antenna, the straight arm length of the inverted L antenna, and the height of the neutral line stub. Except for the critical dimensional parameters, all other parameters remain at fixed values.

[0007] S2 includes discretizing each key dimension parameter within its value range by step size, and then combining them into a four-dimensional Cartesian product to obtain all parameter combinations. For each parameter combination, a full-wave electromagnetic simulation was performed using full-wave electromagnetic simulation software to obtain the S11 and S21 parameter curves in the range of 2GHz to 8GHz for the current parameters. Each set of original samples includes four key size parameter values, one S11 parameter curve, and one S21 parameter curve; Each S-parameter curve contains Data from each frequency point.

[0008] S3 includes S3.1, calculating the average value of S21 for each group of samples within the target operating frequency band. : ; In the formula, This refers to the number of frequency points included within the target frequency band. For the target operating frequency band, For frequency, For absolute values, For frequency The amplitude of S21 at point S21; Set effective sample threshold Within the target operating frequency band, The current sample is a valid sample; all valid samples will be used to construct the training dataset. , , For the number of valid samples, For valid sample index; S3 includes, S3.2, zero-mean standardization of each key size parameter in the training dataset: ; In the formula, The standardized value. For the first The first sample The original values ​​of the key dimensional parameters, For the first The average of the key dimensional parameters, For the first The standard deviation of each key dimensional parameter .

[0009] S4 includes S4.1, constructing a weighted KNN surrogate model, first calculating the distance to the sample to be predicted. Calculation and the first Sample Euclidean distance: ; In the formula, for To the Euclidean distance of each sample For the first Key dimensional parameters; Set the total number of nearest neighbors. All samples are arranged according to Sort the Euclidean distances in ascending order and select the one with the closest Euclidean distance. The group of samples is used as the nearest neighbor sample set.

[0010] S4 includes, S4.2, for the first The nearest neighbor samples, and the weights of the nearest neighbor samples. for: ; In the formula, for index, For the first The weights of the nearest neighbor samples, for To the The Euclidean distance of the nearest neighbor samples, To prevent division by zero; Normalize the weights: ; In the formula, These are the weights after normalization; S4 includes, S4.3, calculating the frequency of the sample to be predicted. S-parameter prediction values ​​at [location] : ; In the formula, For the first The nearest neighbor samples have the same frequency The actual value of the S-parameter at that location.

[0011] S5 includes constructing the coefficient of determination. : ; In the formula, The total number of data points participating in the evaluation. For indexing data points, For the first The true value of each data point For the first Predicted values ​​for each data point This is the average of all true values.

[0012] S6 includes, S6.1, generating according to a preset step size within the value range of the four key dimensional parameters. The candidate key size parameters are used as candidate samples; each group of candidate samples is input into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves.

[0013] S6 includes S6.2, setting the worst-case return loss value within the target frequency band. The worst value of isolation within the target frequency band and the sum of key dimension parameters As an evaluation indicator, the coefficient of variation method is used for automatic weighting; No. Coefficient of variation of each evaluation indicator for: ; In the formula, For evaluation index, For the first The standard deviation of each indicator across all valid simulation samples For the first The mean of each indicator across all valid simulation samples; The coefficients of variation of each indicator are normalized to obtain the weights of the evaluation indicators. : .

[0014] S7 includes, S7.1, setting hard constraints for candidate structures: ; ; In the formula, The worst-case threshold for return loss within the target frequency band. This is the worst-case threshold for isolation within the target frequency band; S7 includes, in S7.2, constructing a comprehensive scoring function: ; In the formula, For candidate sample index, , For the first The overall score of the candidate samples. For the first The first candidate sample Normalized values ​​of each evaluation indicator; Sort by comprehensive score from smallest to largest, and output the top rankings. Among the candidate antenna structures, the one with the lowest score is determined as the optimal antenna design; the rest... One antenna structure is an alternative.

[0015] Compared with existing technologies, this invention has the following advantages: Based on a weighted K-nearest neighbor surrogate model, this invention constructs a nonlinear mapping relationship between antenna structural parameters and S-parameters, achieving high-precision prediction under small sample conditions without iterative training. The determination coefficients for the S11 and S21 test sets reach 0.9826 and 0.9722, respectively. By combining the coefficient of variation method, it automatically assigns weights to the three evaluation indicators of isolation, return loss, and size compactness, avoiding the subjectivity of manual weight setting in multi-objective optimization. Through hard constraint screening and a comprehensive scoring function, it quickly selects the optimal solution that meets the dual-band design indicators from a large number of candidate structures. Compared to traditional full-wave simulation point-by-point scanning, generating 372 sets of simulation data takes approximately 20 hours, while the trained model can predict 600 candidate structures in just seconds, significantly reducing the computational cost in the antenna optimization process. This effectively solves the technical problems of long simulation time, insufficient model accuracy under small sample conditions, and reliance on subjective experience in multi-objective evaluation in traditional methods, providing reliable technical support for the efficient and automated optimization design of dual-band MIMO antennas. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the accuracy result of the weighted KNN surrogate model of this invention; Figure 3 The frequency domain error and coefficient of determination distribution of S11; Figure 4 The frequency domain error and coefficient of determination distribution of S21; Figure 5 It is the S11 predictive correlation; Figure 6 It is the S21 predictive correlation; Figure 7 This is a comparison chart of the S11 prediction results for all samples in the test set; Figure 8 This is a comparison chart of the S21 prediction results for all samples in the test set; Figure 9 It is the training result of model S11 for a certain set of samples in the test set; Figure 10 This is the training result of model S21 on a certain set of test samples; Figure 11 This is a comparison chart of the S11 curves of the first 20 antennas; Figure 12 This is a comparison chart of the S21 curves of the first 20 antennas; Figure 13 These are the actual S11 and S21 curves of the optimal antenna. Detailed Implementation

[0017] 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.

[0018] The dual-band MIMO antenna optimization method based on weighted KNN includes: S1. Obtain the key dimensional parameters of the inverted L-shaped and monopole composite dual-band antenna; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned using combined parameters to obtain the parameter curves of S11 and S21, and the original sample is constructed. S3. Calculate the average value of the S21 parameter curve for each group of samples within the target working frequency band, set an effective sample threshold to filter effective samples and construct a training dataset, and standardize the key size parameters in the training dataset to zero mean. S4. Construct a weighted KNN surrogate model. First, calculate the Euclidean distance between the sample to be predicted and the effective sample. Set the total number of nearest neighbors to filter the nearest neighbor samples. Calculate the weight of the nearest neighbor sample based on the Euclidean distance and normalize it. Combine the true value of the S-parameter of the effective sample to calculate the predicted value of the S-parameter of the sample to be predicted. S5. Calculate the determination coefficient of the predicted S-parameter values ​​of the sample to be predicted, adjust the total number of nearest neighbors, select the total number of nearest neighbors with the highest determination coefficient, and obtain the trained weighted KNN surrogate model. S6. Within the range of values ​​for the four key size parameters, generate several sets of candidate key size parameters according to a preset step size as candidate samples; input each set of candidate samples into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves; set the worst value of return loss in the target frequency band, the worst value of isolation in the target frequency band, and the sum of the key size parameters as evaluation indicators, and calculate the weight of the evaluation indicators using the coefficient of variation method; S7. Set the worst-case threshold values ​​for return loss and isolation within the target frequency band as hard constraints. Construct a comprehensive scoring function using the weights and normalized values ​​of the evaluation indicators. Sort the comprehensive scores from smallest to largest, and determine the smallest score as the optimal antenna design scheme.

[0019] S1 includes key dimensional parameters such as the feed gap between the inverted L antenna and the monopole antenna, the spacing between the inverted L antenna and the monopole antenna, the straight arm length of the inverted L antenna, and the height of the neutral line stub. Except for the critical dimensional parameters, all other parameters remain at fixed values.

[0020] S2 includes discretizing each key dimension parameter within its value range by step size, and then combining them into a four-dimensional Cartesian product to obtain all parameter combinations. For each parameter combination, a full-wave electromagnetic simulation was performed using full-wave electromagnetic simulation software to obtain the S11 and S21 parameter curves in the range of 2GHz to 8GHz for the current parameters. Each set of original samples includes four key size parameter values, one S11 parameter curve, and one S21 parameter curve; Each S-parameter curve contains Data from each frequency point.

[0021] S3 includes S3.1, calculating the average value of S21 for each group of samples within the target operating frequency band. : ; In the formula, This refers to the number of frequency points included within the target frequency band. For the target operating frequency band, For frequency, For absolute values, For frequency The amplitude of S21 at point S21; Set effective sample threshold Within the target operating frequency band, The current sample is a valid sample; all valid samples will be used to construct the training dataset. , , For the number of valid samples, For valid sample index; S3 includes, S3.2, zero-mean standardization of each key size parameter in the training dataset: ; In the formula, The standardized value. For the first The first sample The original values ​​of the key dimensional parameters, For the first The average of the key dimensional parameters, For the first The standard deviation of each key dimensional parameter .

[0022] S4 includes S4.1, constructing a weighted KNN surrogate model, first calculating the distance to the sample to be predicted. Calculation and the first Sample Euclidean distance: ; In the formula, for To the Euclidean distance of each sample For the first Key dimensional parameters; Set the total number of nearest neighbors. All samples are arranged according to Sort the Euclidean distances in ascending order and select the one with the closest Euclidean distance. The group of samples is used as the nearest neighbor sample set.

[0023] S4 includes, S4.2, for the first The nearest neighbor samples, and the weights of the nearest neighbor samples. for: ; In the formula, for index, For the first The weights of the nearest neighbor samples, for To the The Euclidean distance of the nearest neighbor samples, To prevent division by zero; Normalize the weights: ; In the formula, These are the weights after normalization; S4 includes, S4.3, calculating the frequency of the sample to be predicted. S-parameter prediction values ​​at [location] : ; In the formula, For the first The nearest neighbor samples have the same frequency The actual value of the S-parameter at that location.

[0024] S5 includes constructing the coefficient of determination. : ; In the formula, The total number of data points participating in the evaluation. For indexing data points, For the first The true value of each data point For the first Predicted values ​​for each data point This is the average of all true values.

[0025] S6 includes, S6.1, generating according to a preset step size within the value range of the four key dimensional parameters. The candidate key size parameters are used as candidate samples; each group of candidate samples is input into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves.

[0026] S6 includes S6.2, setting the worst-case return loss value within the target frequency band. The worst value of isolation within the target frequency band and the sum of key dimension parameters As an evaluation indicator, the coefficient of variation method is used for automatic weighting; No. Coefficient of variation of each evaluation indicator for: ; In the formula, For evaluation index, For the first The standard deviation of each indicator across all valid simulation samples For the first The mean of each indicator across all valid simulation samples; The coefficients of variation of each indicator are normalized to obtain the weights of the evaluation indicators. : .

[0027] S7 includes, S7.1, setting hard constraints for candidate structures: ; ; In the formula, The worst-case threshold for return loss within the target frequency band. This is the worst-case threshold for isolation within the target frequency band; S7 includes, in S7.2, constructing a comprehensive scoring function: ; In the formula, For candidate sample index, , For the first The overall score of the candidate samples. For the first The first candidate sample Normalized values ​​of each evaluation indicator; Sort by comprehensive score from smallest to largest, and output the top rankings. Among the candidate antenna structures, the one with the lowest score is determined as the optimal antenna design; the rest... One antenna structure is an alternative.

[0028] The following description, in conjunction with the accompanying drawings, further illustrates the process of this invention. Figure 1 As shown, the process begins with data reading and parameter setting, followed by effective sample screening (average S21 ≤ -20dB), calculation of the coefficient of variation weights (S11, S21, size), Z-score standardization, construction of a KNN prediction model for training, batch prediction of parameters (600 groups), output of the TOP20 selection, plotting of the results and CST verification, and finally output to a visualization prediction platform.

[0029] The following is a further explanation with reference to the embodiments. This embodiment of the invention adjusts antenna structural parameters and neutralization line dimensions, and analyzes the variation law of S-parameters obtained from full-wave simulation. Four structural parameters that significantly affect the operating frequency band performance of the inverted-L antenna and monopole antenna are selected as optimization variables: feed gap (gap), neutralization line stub height (h4), inverted-L antenna straight arm length (l), and the distance between the two antennas in the x-axis direction (min2). Among these, gap mainly affects the coupling strength and port impedance matching state between the two antennas; h4 mainly changes the coupling compensation effect of the neutralization line, thus affecting the isolation; l is used to adjust the resonant frequency position of the inverted-L antenna; and min2 affects the spatial coupling strength and overall surface current distribution between the two antennas. These four parameters together affect the antenna's matching and isolation characteristics; therefore, they are used as input variables for subsequent surrogate modeling and multi-objective optimization.

[0030] The scanning range and step size for the four key parameters are as follows: feed gap gap: 0.5 to 0.6 mm, step size 0.02 mm; x-axis spacing min2: 8 to 10 mm, step size 1 mm; inverted L antenna straight arm length l: 11.5 to 13 mm, step size 0.5 mm; neutral line stub height h4: 1.2 to 1.4 mm, step size 0.05 mm. Combined scans were performed within the above parameter ranges, and S-parameter curves corresponding to each set of structural parameters were obtained using electromagnetic simulation software. Each sample set has four structural dimension parameters, ranging from 2 to 8 GHz. Data for S11 and S21 at each frequency point.

[0031] This embodiment focuses on the 2.4GHz and 5GHz frequency bands. The 2.4GHz band is defined as 2.4 to 2.484GHz, and the 5GHz band as 4.4 to 5.0GHz. Considering the importance of isolation in high-isolation MIMO antenna design, physical constraints are further filtered based on the original simulation data. First, the average S21 value of each sample group within the target operating frequency band is calculated, and a threshold for valid samples is set. By selecting effective samples, the training data can be more concentrated in high-isolation structural regions, thereby improving the surrogate model's ability to learn effective antenna structures.

[0032] In this embodiment, a total of 372 sets of original simulation samples were obtained. All 372 sets of original samples met the isolation constraint and were therefore retained as valid samples. The valid samples were randomly divided into a training set and a test set in an 8:2 ratio, with the training set containing 298 sets of samples and the test set containing 74 sets of samples. The training set was used to build a weighted KNN surrogate model, and the test set was used to evaluate the model's predictive ability for unknown structural parameters. It should be noted that the screening conditions at this stage are used to ensure that the training data is generally in a region with a relatively good isolation. In the subsequent candidate structure optimization process, this embodiment further uses the worst values ​​of S11 and S21 within the target frequency band as hard constraints to ensure that the optimization results meet the design indicators throughout the entire target frequency band.

[0033] In this embodiment, due to the differences in the dimensions and value ranges of the four structural parameters, directly calculating the Euclidean distance may lead to variables with larger value ranges having a stronger influence on the distance metric. To eliminate the differences in dimensions, prediction models are established for S11 and S21 respectively, thus obtaining complete frequency response curves for S11 and S21.

[0034] In this embodiment, the coefficient of determination is used to evaluate the overall prediction accuracy of the model. Since each test sample contains S-parameter data at 1000 frequency points, this embodiment calculates the coefficient of determination by expanding the true and predicted values ​​of all test samples at all frequency points, and evaluates the overall fitting ability of the surrogate model to the frequency response curve.

[0035] The model's test set prediction accuracy results are as follows: Figure 2 As shown, the S11 prediction model The S21 prediction model has a value of 0.9826. The value of 0.9722 indicates that the weighted KNN surrogate model can learn the mapping relationship between structural parameters and S-parameter curves relatively well.

[0036] Depend on Figure 3 It can be seen that the model has good predictive ability for the S11 parameters across the entire frequency band from 2 to 8 GHz. (See figure) The curves generally remain at a high level, indicating that the model can fit the overall trend of S11 with frequency variation well. Meanwhile, the MSE and MAE curves show relatively small fluctuations, indicating that the prediction error is within a relatively stable range, and the model's prediction results for the S11 frequency response have high reliability. The appearance of localized increases indicates that the prediction error is relatively large at these frequency points. This is mainly because the antenna is more sensitive to changes in structural parameters at the resonant point or in regions of rapid curve change; even small dimensional changes can cause significant shifts in the S-parameter curve, making predictions relatively more difficult in these areas. However, overall... The curve did not show a significant drop, indicating that local error fluctuations did not significantly affect the overall predictive performance of the model.

[0037] Depend on Figure 4 It can be seen that the model also has a good predictive effect on the S21 parameter. The curves remain at a high level across most frequency ranges, indicating that the model can accurately predict the trend of port isolation as a function of frequency. Although the MSE and MAE curves exhibit some fluctuations in certain frequency bands, the overall error remains within an acceptable range, demonstrating the model's good stability in predicting S21. Compared to S11, S21 is slightly more difficult to predict. This is because S21 primarily reflects the isolation characteristics between antenna ports, is more significantly affected by the coupling effects of multiple structural parameters, and has a relatively more complex nonlinear relationship. Therefore, in frequency regions where isolation characteristics change rapidly, the model error may increase. However, overall, S21... The fact that it remains at a high level indicates that the model can effectively learn the mapping relationship between structural parameters and port isolation performance.

[0038] Depend on Figure 5 It can be seen that most of the sample points in both the training and test sets are distributed near the ideal line, indicating a strong consistency between the model's predictions and the actual values. (S11 training set) The value is 0.9831, for the test set. The S11 value is 0.9826, very close to the training set value, indicating that the model not only has high fitting accuracy on the training set but also maintains good predictive ability on the test set. Looking at the S11 scatter distribution, neither the training nor the test set shows significant deviations, indicating that the model does not suffer from obvious overfitting. The scatter points in some deeply negative regions are relatively dispersed, suggesting that the prediction error may increase slightly when the S11 value is low or the curve is complex. However, overall, the sample points are still distributed near the ideal line, indicating that the model can predict the S11 parameter relatively accurately.

[0039] Depend on Figure 6It can be seen that the model's predicted values ​​and the actual values ​​generally exhibit a strong linear correlation, with the scatter points of both the training and test sets concentrated near the ideal line. (S21 training set) The value is 0.9796, for the test set. The value is 0.9722, indicating that the model also has high prediction accuracy and good generalization ability for the S21 parameters. Compared with S11, the S21 test set... The slightly lower value indicates that the prediction complexity of S21 is relatively higher. This may be because S21 reflects the coupling and isolation characteristics between ports, and its changes are affected by multiple structural parameters, thus the model will produce a relatively larger error when predicting S21. However, the test set... The results still show a high level, indicating that the model can predict antenna isolation performance well.

[0040] Depend on Figure 7 , Figure 8 , Figure 9 and Figure 10 It can be seen that within the target frequency bands of 2.4 GHz and 5 GHz, the trends of the predicted curves and the actual curves are basically consistent, indicating that the surrogate model can reflect the changes in antenna matching performance and isolation performance with frequency quite well. Therefore, this weighted KNN model is used as a surrogate model for subsequent batch parameter prediction and multi-objective optimization screening.

[0041] High-isolation MIMO antenna optimization typically involves multiple performance metrics. For the design object in this embodiment, the optimization objectives mainly include three aspects: First, minimizing S21 within the target frequency band to improve the isolation between the two ports; second, minimizing S11 within the target frequency band to improve port impedance matching; and third, achieving the most compact structural dimensions while meeting electromagnetic performance constraints.

[0042] Therefore, this embodiment selects the worst values ​​of S21 and S11 within the target frequency band, along with the structural dimensions, as evaluation indicators. Since the S-parameters are expressed in dB and are usually negative, the maximum value represents the position with the worst performance within the target frequency band. For example, when the maximum S11 of a certain antenna group within the target frequency band is still less than -6 dB, it indicates that the structure meets the return loss requirement throughout the entire target frequency band; when the maximum S21 is less than -20 dB, it indicates that it meets the isolation requirement throughout the entire target frequency band. The structural dimensions indicator is defined as the sum of the four optimization parameters. Although this indicator is not completely equivalent to the overall antenna area, it can reflect the compactness of the locally adjustable dimensions under a fixed antenna topology. In multi-target evaluation, different indicators have different dimensions and variation amplitudes. If weights are directly set manually, subjectivity can easily be introduced. Therefore, this embodiment uses the coefficient of variation method to automatically assign weights to the three evaluation indicators. The coefficient of variation can reflect the dispersion of a certain indicator in the sample set. The more obvious the fluctuation of the indicator, the higher its discrimination of the sample and the greater its contribution to the comprehensive evaluation. Therefore, the coefficient of variation is calculated and normalized to obtain the corresponding weights. In this embodiment, the coefficient of variation weights are calculated based on real, valid simulation samples, rather than on candidate prediction samples generated by the subsequent surrogate model. This is because real simulation samples can more reliably reflect the actual changes in antenna structure parameters and performance indicators, avoiding interference from surrogate model prediction errors or changes in candidate sample distribution on the weight calculation. The calculation results include a coefficient of variation of... , The size is 0.0443; the automatic weight is... , The size is 0.0393.

[0043] After obtaining the weights, this embodiment performs a comprehensive score on the candidate samples that meet the hard constraints. Since smaller values ​​are better for the three indicators S21, S11, and size, the range of each indicator in the qualified samples is first normalized: ; Then construct the comprehensive scoring function: ; In the formula, For candidate sample index, For evaluation index, For the first The candidate sample at the th The original values ​​of each evaluation indicator Among all qualified candidate samples, the first The minimum value of each evaluation indicator. Among all qualified candidate samples, the first The maximum value of each evaluation indicator. For the first The candidate sample at the th The normalized range value for each evaluation indicator For the first The normalized value of the S21 index for each candidate sample. For the first The normalized value of the S11 index for each candidate sample. For the first The normalized value of the Size index for each candidate sample. The weight of indicator S21 in the overall score. The weight of the S11 indicator in the overall score. The weight of the Size indicator in the overall score. For the first The overall score of each candidate sample; The lower the overall score, the better the candidate structure performs in terms of isolation, return loss, and size compactness.

[0044] After completing the training and evaluation weight calculation of the weighted KNN surrogate model, this embodiment further performs batch sampling and rapid prediction of the parameter space to screen out antenna structure parameter combinations with superior overall performance. First, the candidate parameter range is determined based on the minimum and maximum values ​​of the four structural parameters in the valid samples. To ensure a relatively uniform distribution of candidate samples within the parameter space, this embodiment uniformly discretizes the four parameters gap, h4, l, and min2, and generates candidate structures through a four-dimensional grid combination. The program sets the number of candidate parameter groups to 600, taking approximately the same number of discrete points in each dimension, ultimately generating 600 candidate parameter combinations.

[0045] For each set of candidate structural parameters, the same standardization method as in the training phase is used for processing, and the parameters are input into a weighted KNN surrogate model to quickly predict the S11 and S21 curves at 1000 frequency points in the 2 to 8 GHz range. Compared with traditional full-wave simulation and group-by-group calculation, the surrogate model prediction only requires distance calculation and weighted summation of neighboring samples, thus significantly reducing the time cost of the parameter optimization process.

[0046] To ensure that the final optimization result meets the basic electromagnetic performance requirements, this embodiment first performs hard constraint screening on the candidate structures and sets the worst-case threshold for return loss within the target frequency band. Set the worst-case threshold for isolation within the target frequency band. Of the 600 candidate structures, 84 met the hard constraints. Using a fixed coefficient of variation weight, all qualified candidate structures were comprehensively scored and sorted from lowest to highest score, ultimately outputting the top 20 candidate antenna structures. The top-ranked candidate structures all simultaneously met the design requirements of S11 < -6dB and S21 < -20dB. The structure ranked first had the lowest comprehensive score, indicating that it achieved a good balance between isolation performance, matching performance, and structural size within the target frequency band. Therefore, this embodiment selected this set of structural parameters as the optimal antenna design scheme obtained based on the weighted KNN surrogate model optimization.

[0047] Figure 11 and Figure 12 In the diagram, the bold blue curve represents the optimal structure ranked first in overall score, while the remaining curves represent the other Top 20 candidate structures. The gray areas represent the target operating frequency bands of 2.4 GHz and 5.0 GHz, respectively, and the red dashed lines represent the corresponding performance thresholds. The S11 curves of the Top 20 candidate structures within the target operating frequency bands are all below -6 dB, indicating that these structures possess a certain impedance matching capability in both the 2.4 GHz and 5.0 GHz bands. The optimal structure does not exhibit any mismatch points significantly higher than the threshold within the target frequency band, indicating that it can meet the return loss requirements under dual-frequency operating conditions. The S21 curves of the Top 20 candidate structures within the target frequency bands are all below -20 dB, indicating that the optimized structure can effectively suppress coupling between the two antenna elements. Especially in the 2.4 GHz and 5.0 GHz bands, the isolation of the optimal structure meets the design specifications, demonstrating that the coordinated adjustment of the neutralization line structure and key size parameters has a significant effect on improving the isolation performance of MIMO antennas.

[0048] Depend on Figure 13 As can be seen, the overall trend of the actual simulation curve and the model prediction curve is basically consistent, indicating that the established machine learning model can accurately predict the electromagnetic response of the antenna in the target frequency band. Meanwhile, the optimal structure meets the impedance matching and isolation design requirements in both the 2.4GHz and 5.0GHz target frequency bands, verifying the effectiveness of the algorithm's selection results.

[0049] The optimal structure obtained by the algorithm was compared with the antenna structure obtained from the initial parameter scanning. The results show that, compared with the S-parameters of the initial parameter scanning structure, the optimized antenna has an S11 of less than -6dB in the target frequency band, an S21 of -5dB in the low-frequency band, and an S21 of around -28dB in the high-frequency band.

[0050] In summary, the optimization method based on the weighted KNN surrogate model can quickly predict the S-parameter responses of a large number of candidate structures based on existing simulation data. Combined with hard constraint screening and a comprehensive scoring method based on the coefficient of variation, it selects the antenna structure with superior performance from the parameter space. This method reduces the computational burden of traditional point-by-point full-wave simulation, improves the efficiency of antenna structure parameter optimization, and provides a feasible data-driven optimization approach for the rapid design of high-isolation MIMO antennas.

[0051] 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 modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such 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 dual-frequency MIMO antenna optimization method based on weighted KNN, characterized in that, include: S1. Obtain the key dimensional parameters of the inverted L-shaped and monopole composite dual-band antenna; S2. By using full-wave electromagnetic simulation software, the variation range of key size parameters is scanned using combined parameters to obtain the parameter curves of S11 and S21, and the original sample is constructed. S3. Calculate the average value of the S21 parameter curve for each group of samples within the target working frequency band, set an effective sample threshold to filter effective samples and construct a training dataset, and standardize the key size parameters in the training dataset to zero mean. S4. Construct a weighted KNN surrogate model. First, calculate the Euclidean distance between the sample to be predicted and the effective sample. Set the total number of nearest neighbors to filter the nearest neighbor samples. Calculate the weight of the nearest neighbor sample based on the Euclidean distance and normalize it. Combine the true value of the S-parameter of the effective sample to calculate the predicted value of the S-parameter of the sample to be predicted. S5. Calculate the determination coefficient of the predicted S-parameter values ​​of the sample to be predicted, adjust the total number of nearest neighbors, select the total number of nearest neighbors with the highest determination coefficient, and obtain the trained weighted KNN surrogate model. S6. Within the range of values ​​for the four key size parameters, generate several sets of candidate key size parameters according to a preset step size as candidate samples; input each set of candidate samples into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves. The worst-case return loss value, the worst-case isolation value, and the sum of key size parameters within the target frequency band are set as evaluation indicators, and the weights of the evaluation indicators are calculated using the coefficient of variation method. S7. Set the worst-case threshold values ​​for return loss and isolation within the target frequency band as hard constraints. Construct a comprehensive scoring function using the weights and normalized values ​​of the evaluation indicators. Sort the comprehensive scores from smallest to largest, and determine the smallest score as the optimal antenna design scheme.

2. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 1, characterized in that, S1 includes key dimensional parameters such as the feed gap between the inverted L antenna and the monopole antenna, the spacing between the inverted L antenna and the monopole antenna, the straight arm length of the inverted L antenna, and the height of the neutral line stub. Except for the critical dimensional parameters, all other parameters remain at fixed values.

3. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 2, characterized in that, S2 includes discretizing each key dimension parameter within its value range by step size, and then combining them using a four-dimensional Cartesian product to obtain all parameter combinations. For each parameter combination, a full-wave electromagnetic simulation was performed using full-wave electromagnetic simulation software to obtain the S11 and S21 parameter curves in the range of 2GHz to 8GHz for the current parameters. Each set of original samples includes four key size parameter values, one S11 parameter curve, and one S21 parameter curve; Each S-parameter curve contains Data from each frequency point.

4. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 3, characterized in that, S3 includes S3.1, calculating the average value of S21 for each group of samples within the target operating frequency band. : ； In the formula, This refers to the number of frequency points included within the target frequency band. For the target operating frequency band, For frequency, For absolute values, For frequency The amplitude of S21 at point S21; Set effective sample threshold Within the target operating frequency band, The current sample is a valid sample; all valid samples will be used to construct the training dataset. , , For the number of valid samples, For valid sample index; S3 includes, S3.2, zero-mean standardization of each key size parameter in the training dataset: ； In the formula, The standardized value. For the first The first sample The original values ​​of the key dimensional parameters, For the first The average of the key dimensional parameters, For the first The standard deviation of each key dimensional parameter .

5. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 4, characterized in that, S4 includes S4.1, constructing a weighted KNN surrogate model, first calculating the distance to the sample to be predicted. Calculation and the first Sample Euclidean distance: ； In the formula, for To the Euclidean distance of each sample For the first Key dimensional parameters; Set the total number of nearest neighbors. All samples are arranged according to Sort the Euclidean distances in ascending order and select the one with the closest Euclidean distance. The group of samples is used as the nearest neighbor sample set.

6. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 5, characterized in that, S4 includes, S4.2, for the first The nearest neighbor samples, and the weights of the nearest neighbor samples. for: ； In the formula, for index, For the first The weights of the nearest neighbor samples, for To the The Euclidean distance of the nearest neighbor samples, To prevent division by zero; Normalize the weights: ； In the formula, These are the weights after normalization; S4 includes, S4.3, calculating the frequency of the sample to be predicted. S-parameter prediction values ​​at [location] : ； In the formula, For the first The nearest neighbor samples have the same frequency The actual value of the S-parameter at that location.

7. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 6, characterized in that, S5 includes constructing the coefficient of determination. : ； In the formula, The total number of data points participating in the evaluation. For indexing data points, For the first The true value of each data point For the first Predicted values ​​for each data point This is the average of all true values.

8. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 7, characterized in that, S6 includes, S6.1, generating according to a preset step size within the value range of the four key dimensional parameters. The candidate key size parameters are used as candidate samples; each group of candidate samples is input into the trained weighted KNN surrogate model to obtain the corresponding S11 and S21 prediction curves.

9. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 8, characterized in that, S6 includes S6.2, setting the worst-case return loss value within the target frequency band. The worst value of isolation within the target frequency band and the sum of key dimension parameters As an evaluation indicator, the coefficient of variation method is used for automatic weighting; No. Coefficient of variation of each evaluation indicator for: ； In the formula, For evaluation index, For the first The standard deviation of each indicator across all valid simulation samples For the first The mean of each indicator across all valid simulation samples; The coefficients of variation of each indicator are normalized to obtain the weights of the evaluation indicators. : 。 10. The dual-frequency MIMO antenna optimization method based on weighted KNN according to claim 9, characterized in that, S7 includes, S7.1, setting hard constraints for candidate structures: ； ； In the formula, The worst-case threshold for return loss within the target frequency band. This is the worst-case threshold for isolation within the target frequency band; S7 includes, in S7.2, constructing a comprehensive scoring function: ； In the formula, For candidate sample index, , For the first The overall score of the candidate samples. For the first The first candidate sample Normalized values ​​of each evaluation indicator; Sort by comprehensive score from smallest to largest, and output the top rankings. Among the candidate antenna structures, the one with the lowest score is determined as the optimal antenna design; the rest... One antenna structure is an alternative.

Images