An electromagnetic device optimization design method based on a multi-view semi-supervised learning assisted particle swarm optimization algorithm

Abstract

The application belongs to the field of microwave technology and electromagnetic optimization design, and particularly discloses an electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning. The method disclosed by the application comprises the following steps: defining design variables and optimization objectives, and generating initial samples; updating particle positions by using an improved social learning particle swarm algorithm, selecting a prediction model to perform real simulation on potential particles, and storing the remaining particles in an unlabeled sample pool; screening high-confidence samples by using a multi-view model, generating pseudo-labels by using convex combination optimization, and expanding a training set; constructing a semi-supervised learning model, calculating a prediction error, and setting a tolerance interval; selecting a prediction model according to a three-mode decision rule, iteratively updating the model and the particle swarm, and stopping until the design objectives are met. The application can efficiently optimize the performance of electromagnetic devices with a small number of simulation times. The application is suitable for electromagnetic design problems with high-cost simulation evaluation.

Description

Technical Field

[0001] This invention belongs to the field of microwave technology and electromagnetic optimization design, specifically relating to an electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning. Background Technology

[0002] Currently, in the field of electromagnetic device design, optimization typically employs a combination of full-wave electromagnetic simulation software and global optimization algorithms. However, electromagnetic simulation itself is computationally expensive, and optimization algorithms require tens of thousands of simulation iterations, resulting in lengthy design cycles. To accelerate optimization, researchers have introduced surrogate model techniques to fit the nonlinear relationship between electromagnetic device structural parameters and performance indicators, replacing some of the simulation. However, most existing surrogate model-assisted methods are based on supervised learning paradigms, and their modeling accuracy is highly dependent on the support of a large amount of high-quality labeled data. With the increase in the size and topological complexity of electromagnetic devices, the time cost of a single full-wave simulation increases exponentially, making it extremely expensive and difficult to obtain sufficient labeled samples within a limited design cycle.

[0003] Semi-supervised learning offers a solution to the problem of scarce labeled data. Currently, it has been preliminarily applied to static modeling tasks such as antenna resonant frequency prediction. However, existing methods still suffer from insufficient robustness and optimization efficiency in deeply integrating semi-supervised learning models with dynamic optimization algorithms to automatically find the optimal size parameters of electromagnetic devices such as antennas. Summary of the Invention

[0004] The purpose of this invention is to provide an electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning, which mines the potential information in unlabeled design parameters and achieves efficient optimization of electromagnetic device performance under a low-cost electromagnetic simulation budget.

[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] An electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning, the method comprising the following steps performed sequentially:

[0007] S1. Define the design variables and optimization objectives of electromagnetic devices, use Latin hypercube sampling to generate the initial population and training samples, perform initial simulation and determine the individual optimum and global optimum;

[0008] S2. Determine whether the best particle in the training samples meets the design objective;

[0009] If so, proceed to step S9;

[0010] If not, a supervised learning model is built based on the current training samples as the prediction model;

[0011] S3. The improved social learning particle swarm algorithm is used to update the particle position and velocity, and the current prediction model is selected to predict the updated particles.

[0012] S4. Determine whether the predicted value is less than the individual optimal value;

[0013] If so, perform a real simulation evaluation, update the global optimum, and then proceed to step S5;

[0014] If not, store the particles in the unlabeled sample archive pool and proceed to step S7;

[0015] S5. Determine whether the updated global optimum satisfies the design objective;

[0016] If so, proceed to step S9;

[0017] If not, proceed to step S6;

[0018] S6. Determine if the current iteration count is the maximum iteration count;

[0019] If so, proceed to step S9;

[0020] If not, proceed to step S7;

[0021] S7. Construct a multi-view learning model using all current labeled data, filter unlabeled samples in the archive pool, and use convex combination optimization technology to generate pseudo-labels for them. Add the samples with pseudo-labels to the training samples, and construct and update the semi-supervised learning model based on the latest training samples.

[0022] S8. Calculate the average relative error between the real simulation evaluation sample and the semi-supervised learning prediction value in step S4, and the average relative error between the real simulation evaluation sample and the supervised learning prediction value in step S4, and update the two average relative errors.

[0023] After selecting the latest prediction model using the three-mode decision rule, return to step S3;

[0024] S9. Output the optimal particle as the final optimal physical size of the electromagnetic device, and end.

[0025] As a limitation, the improved social learning particle swarm optimization algorithm in step S3 includes:

[0026]

[0027]

[0028]

[0029] In the formula, For the first The particle in the first Position at the next iteration For the first The particle in the first Speed ​​at the next iteration , and A random number between [0,1] For demonstrator k, the individual historical best position is where k is from a position with better fitness than particle k. Randomly selected index from the set

[0030] As a further limitation, the formula for calculating the average relative error in step S8 is:

[0031]

[0032] In the formula, The average relative error, For the sample size, For the predicted values ​​of the surrogate model, For realistic simulation values

[0033] To further define this, the three-mode decision-making rule is as follows:

[0034] Set the range of error tolerance and calculate the relative error ratio;

[0035] If the relative error ratio is less than the minimum error tolerance, then a semi-supervised learning model should be selected.

[0036] If the relative error ratio is greater than the maximum value of the error tolerance, then a supervised learning model should be selected.

[0037] Otherwise, use both models simultaneously for collaborative prediction.

[0038] Preferably, the formula for calculating the relative error ratio is:

[0039]

[0040] In the formula, This is the relative error ratio. To evaluate the average relative error between real-world simulation samples and semi-supervised learning predictions, This is used to evaluate the average relative error between the real simulation samples and the supervised learning predictions.

[0041] As a second limitation, the multi-view learning model employs three Gaussian process models with kernel functions of RBF kernel, Matérn 32 kernel, and Matérn 52 kernel, respectively.

[0042] The calculation formulas are as follows:

[0043]

[0044]

[0045]

[0046] In the formula, The standard deviation of the signal. For Euclidean distance.

[0047] As a further limitation, the multi-view learning model filters unlabeled samples by calculating the sum of variances of the predicted values ​​of each view, and selects the sample with the smallest variance as the unlabeled sample.

[0048] As a further limitation, the formula for the convex combination optimization technique includes:

[0049]

[0050]

[0051]

[0052] In the formula, For each view in the multi-view learning model, the initial prediction value is... for Weight values ​​under the view for The optimal weight value under the view. For the number of views, For view Security pseudo-labels from different perspectives This is the final pseudo-tag.

[0053] The present invention, by adopting the above-described technical solution, achieves the following technical advancements compared to existing technologies:

[0054] (1) The method of the present invention uses a dual mechanism of “selective simulation” and “pseudo-label data augmentation” to use expensive full-wave simulation only to verify high-potential solutions, while continuously enhancing the model with a large amount of low-cost pseudo-label data, thereby completing the optimization in a very small number of simulations;

[0055] (2) The “three-mode decision rule” proposed by the method of the present invention can adaptively switch or fuse prediction models according to the real-time accuracy of the semi-supervised model, avoiding the risk of the search going astray when a single model is inaccurate, and ensuring the stable convergence of the optimization process.

[0056] (3) The method of the present invention employs an innovative "multi-view" Figure 1 The "consistency screening" and "convex combination optimization labeling" processes ensure that the pseudo-label data added to the training set has high confidence and high accuracy, effectively avoiding the contamination of the semi-supervised learning model by noisy data;

[0057] (4) The method of the present invention has high versatility, does not depend on the structure of specific electromagnetic devices or simulation software, and is applicable to various electromagnetic designs that require high-cost evaluation function optimization.

[0058] This invention belongs to the field of microwave technology and electromagnetic optimization design, and can achieve efficient optimization of electromagnetic device performance under low-cost electromagnetic simulation budget. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0060] In the attached diagram:

[0061] Figure 1 This is a flowchart illustrating the processing of an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram showing the dimensional parameters of the slotted hybrid structure broadband antenna in an embodiment of the present invention;

[0063] Figure 3 This is a comparison of the return loss curves of the antenna before and after optimization in an embodiment of the present invention.

[0064] Figure 4 This is a convergence curve showing the fitness value increasing with the number of simulations during the optimization process. Detailed Implementation

[0065] The preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0066] Example

[0067] like Figure 1 As shown, this embodiment is an electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning, which includes the following steps performed sequentially:

[0068] S1. Define antenna design variables and optimization objectives, use Latin hypercube sampling to generate initial population and training samples, perform initial simulation and determine individual optimality and global optimality.

[0069] This embodiment takes the optimized design of a slotted hybrid structure broadband antenna operating in the 4.35-6.55 GHz frequency band as an example. The antenna structure to be optimized is as follows: Figure 2 As shown, the strip-slot hybrid broadband antenna consists of four strips: (a) the top layer of the patch; (b) the middle layer of the patch and ground plane; (c) the bottom layer of the Y-fed microstrip line; and (d) a side view of the overall structure. These strips are separated by three narrow slots. By controlling the dimensions of the strips and slots, dual-mode operation, namely TM10 mode and inverted TM20 mode, is excited and coupled to increase the operating bandwidth. The strip-slot hybrid structure can be excited using aperture-coupled Y-fed microstrip lines through optimized impedance matching. The dimensions to be optimized include nine key dimensional parameters, the value ranges of which are shown in Table 1.

[0070]

[0071] The optimization objective was to reduce the maximum return loss S11 of the antenna in the 4.35–6.55 GHz band to below -10 dB. Fifty initial size samples were generated using Latin hypercube sampling, and full-wave simulation was performed using HFSS software with a frequency scan set to 4–7 GHz and a step size of 0.01 to obtain the initial labeled dataset.

[0072] S2. Determine whether the best particle in the training samples meets the design objective;

[0073] If so, proceed to step S9;

[0074] If not, a supervised learning model is built based on the current training samples as the prediction model.

[0075] Based on 50 sets of initial data, such as Figure 3 As shown in the figure, the S11 curve before optimization, indicated by the dashed line, has a relatively small bandwidth. It is determined that the best particle in the existing training samples does not meet the design goal.

[0076] S3. The particle position and velocity are updated using an improved social learning particle swarm algorithm, and the current prediction model is selected to predict the updated particles.

[0077] Specifically, the improved social learning particle swarm optimization algorithm includes:

[0078]

[0079]

[0080]

[0081] In the formula, For the first The particle in the first Position at the next iteration For the first The particle in the first Speed ​​at the next iteration , and A random number between [0,1] For demonstrator k, the individual historical best position is where k is from a position with better fitness than particle k. The index is randomly selected from the set.

[0082] S4. Determine whether the predicted value is less than the individual optimal value;

[0083] If so, perform a real simulation evaluation, update the global optimum, and then proceed to step S5;

[0084] If not, store the particles in the unlabeled sample archive pool and proceed to step S7.

[0085] Only when particle predictions demonstrate the potential to surpass their historical best are expensive full-wave simulations used for realistic evaluation, and the optimal solution is updated accordingly. The remaining particles are stored as unlabeled data in an archive pool. This step ensures that simulation resources are precisely allocated to the most promising search regions.

[0086] S5. Determine whether the updated global optimum satisfies the design objective;

[0087] If so, proceed to step S9;

[0088] If not, proceed to step S6.

[0089] S6. Determine if the current iteration count is the maximum iteration count;

[0090] If so, proceed to step S9;

[0091] If not, proceed to step S7.

[0092] S7. Construct a multi-view learning model using all current labeled data, filter unlabeled samples in the archive pool, and use convex combination optimization technology to generate pseudo-labels for them. Add the samples with pseudo-labels to the training samples, and construct and update the semi-supervised learning model based on the latest training samples.

[0093] S8. Calculate the average relative error between the real simulation evaluation sample and the semi-supervised learning prediction value in step S4, and the average relative error between the real simulation evaluation sample and the supervised learning prediction value in step S4, and update the two average relative errors.

[0094] After selecting the latest prediction model using the three-mode decision rule, return to step S3.

[0095] A multi-view learning model is trained using all currently labeled data, employing three Gaussian process models with kernel functions of RBF kernel, Matérn 32 kernel, and Matérn 52 kernel, respectively.

[0096] The calculation formulas are as follows:

[0097]

[0098]

[0099]

[0100] In the formula, The standard deviation of the signal. For characteristic length scales, For Euclidean distance.

[0101] The multi-view learning model predicts unlabeled samples in the archive pool and uses the consistency of predictions across views as a criterion to select a subset of unlabeled samples with high confidence. The model calculates the sum of variances of the predictions from each view to filter unlabeled samples, selecting the sample with the smallest variance.

[0102] To improve the quality of semi-supervised learning data, convex combination optimization techniques are used to generate pseudo-labels for the selected high-confidence unlabeled samples. By having each view act as a committee to cross-validate, the optimal weighted combination is solved, ultimately resulting in robust and accurate pseudo-labels. The formula for convex combination optimization includes:

[0103]

[0104]

[0105]

[0106] In the formula, For each view in the multi-view learning model, the initial prediction value is... for Weight values ​​under the view for The optimal weight value under the view. For the number of views, For view Security pseudo-labels from different perspectives This is the final pseudo-tag.

[0107] After obtaining the optimal weights, we in the first... Security pseudo-labels from a different perspective Defined as a weighted sum of the committee’s predicted values, in order to integrate information from all views and further eliminate bias, the final pseudo-label assigned to the sample is calculated as the arithmetic mean of the safe pseudo-labels from all view perspectives.

[0108] Then, the average relative error between the real simulation evaluation sample and the semi-supervised learning prediction value, and the average relative error between the real simulation evaluation sample and the supervised learning prediction value are calculated using the following formulas:

[0109]

[0110] In the formula, The average relative error, For the sample size, For the predicted values ​​of the surrogate model, These are numerical values ​​used in realistic simulations.

[0111] Based on the calculated average relative error, a range of error tolerance is set, and the relative error ratio is calculated.

[0112] If the relative error ratio is less than the minimum error tolerance, a semi-supervised learning model is selected; if the relative error ratio is greater than the maximum error tolerance, a supervised learning model is selected; otherwise, both models are used simultaneously for collaborative prediction. In this embodiment, the minimum and maximum error tolerance values ​​are 0.9 and 1.1, respectively.

[0113] The formula for calculating the relative error ratio is:

[0114]

[0115] In the formula, This is the relative error ratio. To evaluate the average relative error between real-world simulation samples and semi-supervised learning predictions, This is used to evaluate the average relative error between the real simulation samples and the supervised learning predictions.

[0116] S9. Output the optimal particle as the final antenna physical size, then end.

[0117] After iteration, the final physical dimensions of the antenna are obtained.

[0118] The optimized physical dimensions of the final slotted hybrid structure broadband antenna are shown in Table 2. The S11 curve obtained from the simulation of the optimal antenna physical dimensions in HFSS is shown in Table 2. Figure 3 As shown by the solid line.

[0119]

[0120] pass Figure 3 It can be seen that the optimized bandwidth covers 4.22-6.66GHz, with a relative bandwidth of 44.85%, which is about 5% wider than the preset optimization target of operating in the 4.35-6.55GHz range. Figure 4 The convergence curve of the maximum S11 value in the frequency band as the number of simulations increases is shown throughout the optimization process. Within the specified 300 simulations, this invention only used 153 simulations to optimize the slotted hybrid structure broadband antenna, which greatly improves the optimization efficiency.

[0121] In summary, this embodiment achieves efficient antenna performance optimization within a low-cost electromagnetic simulation budget. Furthermore, the above implementation examples are only for antenna devices among electromagnetic devices; the method proposed in this invention can be applied to the optimization of passive filters, inductors, capacitors, and other electromagnetic devices.

Claims

1. A method for optimizing the design of electromagnetic devices based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning, characterized in that, The method includes the following steps performed sequentially: S1. Define the design variables and optimization objectives of electromagnetic devices, use Latin hypercube sampling to generate the initial population and training samples, perform initial simulation and determine the individual optimum and global optimum; S2. Determine whether the best particle in the training samples meets the design objective; If so, proceed to step S9; If not, a supervised learning model is built based on the current training samples as the prediction model; S3. The improved social learning particle swarm algorithm is used to update the particle position and velocity, and the current prediction model is selected to predict the updated particles. S4. Determine whether the predicted value is less than the individual optimal value; If so, perform a real simulation evaluation, update the global optimum, and then proceed to step S5; If not, store the particles in the unlabeled sample archive pool and proceed to step S7; S5. Determine whether the updated global optimum satisfies the design objective; If so, proceed to step S9; If not, proceed to step S6; S6. Determine if the current iteration count is the maximum iteration count; If so, proceed to step S9; If not, proceed to step S7; S7. Construct a multi-view learning model using all current labeled data, filter unlabeled samples in the archive pool, and use convex combination optimization technology to generate pseudo-labels for them. Add the samples with pseudo-labels to the training samples, and construct and update the semi-supervised learning model based on the latest training samples. S8. Calculate the average relative error between the real simulation evaluation sample and the semi-supervised learning prediction value in step S4, and the average relative error between the real simulation evaluation sample and the supervised learning prediction value in step S4, and update the two average relative errors. After selecting the latest prediction model using the three-mode decision rule, return to step S3; S9. Output the optimal particle as the final optimal physical size of the electromagnetic device, and end.

2. The electromagnetic device optimization design method based on a multi-view semi-supervised learning-assisted particle swarm optimization algorithm according to claim 1, characterized in that, The improved social learning particle swarm optimization algorithm in step S3 includes: In the formula, For the first The particle in the first Position at the next iteration For the first The particle in the first Speed ​​at the next iteration , and A random number between [0,1] For demonstrator k, the individual historical best position is where k is from a position with better fitness than particle k. The index is randomly selected from the set.

3. The electromagnetic device optimization design method based on a multi-view semi-supervised learning-assisted particle swarm optimization algorithm according to claim 2, characterized in that, The formula for calculating the average relative error in step S8 is as follows: In the formula, The average relative error, For the sample size, For the predicted values ​​of the surrogate model, These are numerical values ​​used in realistic simulations.

4. The electromagnetic device optimization design method based on a multi-view semi-supervised learning-assisted particle swarm optimization algorithm according to claim 3, characterized in that, The three-mode decision-making rule is as follows: Set the range of error tolerance and calculate the relative error ratio; If the relative error ratio is less than the minimum error tolerance, then a semi-supervised learning model should be selected. If the relative error ratio is greater than the maximum value of the error tolerance, then a supervised learning model should be selected. Otherwise, use both models simultaneously for collaborative prediction.

5. The electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning, as described in claim 4, is characterized in that... The formula for calculating the relative error ratio is: In the formula, This is the relative error ratio. To evaluate the average relative error between real-world simulation samples and semi-supervised learning predictions, This is used to evaluate the average relative error between the real simulation samples and the supervised learning predictions.

6. The electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning according to claim 1, characterized in that, The multi-view learning model employs three Gaussian process models with kernel functions of RBF kernel, Matérn 32 kernel, and Matérn 52 kernel, respectively. The calculation formulas are as follows: In the formula, The standard deviation of the signal. For characteristic length scales, For Euclidean distance.

7. The electromagnetic device optimization design method based on a multi-view semi-supervised learning-assisted particle swarm optimization algorithm according to claim 6, characterized in that, The multi-view learning model filters unlabeled samples by calculating the sum of variances of the predicted values ​​for each view, and selects the sample with the smallest variance as the unlabeled sample.

8. The electromagnetic device optimization design method based on a particle swarm optimization algorithm assisted by multi-view semi-supervised learning according to claim 6, characterized in that, The formula for the convex combination optimization technique includes: In the formula, For each view in the multi-view learning model, the initial prediction value is... for Weight values ​​under the view for The optimal weight value under the view. For the number of views, For view Security pseudo-labels from different perspectives This is the final pseudo-tag.

Images