A method for predicting antenna performance using an MLP based on an adaptive naked mole-mole algorithm.

Abstract

This invention discloses an antenna performance prediction method based on an adaptive naked mole-mole algorithm (MLP), belonging to the field of antenna performance prediction technology. This invention utilizes an MLP neural network to construct a surrogate model to predict antenna performance and introduces an improved adaptive naked mole-mole algorithm to replace the gradient descent algorithm. It can analyze the system importance of different parameters, adaptively and dynamically adjust the algorithm's search strategy to obtain the optimal MLP weights and biases. After model construction, it can fit the nonlinear relationship between antenna parameters and performance, effectively overcoming the time and computational costs of traditional electromagnetic simulation for complex antenna structures, improving the speed and accuracy of weight and bias optimization, avoiding getting trapped in local optima, and greatly improving the efficiency of antenna optimization.

Description

Technical Field

[0001] This invention relates to the field of antenna performance prediction technology, and in particular to an antenna performance prediction method based on an adaptive naked mole-mole algorithm-based MLP. Background Technology

[0002] As the front end of a wireless communication system, the performance of an antenna directly determines communication quality, transmission rate, and system reliability. Therefore, rapid and accurate prediction of antenna performance parameters (such as S-parameters and gain) during the antenna design phase has become the cornerstone of modern antenna design and optimization. Currently, antenna performance prediction methods mainly rely on two types of techniques: full-wave electromagnetic simulation based on physical laws and data-driven neural network construction. However, both of these methods have significant limitations when dealing with the high efficiency and high precision required by modern antenna design.

[0003] Full-wave electromagnetic simulation software, such as HFSS and CST, based on Maxwell's equations and employing numerical methods like the finite element method and the method of moments, can provide high-precision antenna performance analysis. However, as antenna structures become increasingly complex, full-wave simulation requires fine meshing of the computational domain and solving large systems of linear or nonlinear equations, resulting in extremely high computational costs and consuming significant computational resources. Furthermore, antenna design optimization often involves combined analysis and optimization of multiple geometric parameters. The excessively long single-point calculation time in full-wave simulation leads to exceptionally slow parameter scanning and global optimization processes, significantly increasing the optimization cycle.

[0004] To overcome the efficiency bottleneck of full-wave simulation, data-driven surrogate model methods have emerged. Among them, the Multilayer Perceptron (MLP), a classic feedforward neural network, has been attempted to establish a mapping relationship from antenna geometric parameters to performance parameters due to its powerful nonlinear fitting ability. Currently, most MLP models use gradient descent-based algorithms to optimize the network's weights and bias parameters. However, since gradient descent iterates along the negative gradient direction of the error surface with small steps, it may require thousands of iterations to converge when training complex networks, resulting in lengthy training times. Furthermore, gradient descent is prone to getting trapped in a non-ideal local optimum, leading to poor model generalization ability and prediction accuracy that fails to meet engineering requirements.

[0005] There is a need in this field for a new technical solution that can maintain the advantage of fast prediction by neural networks while solving the problems of slow convergence and getting stuck in local optima during the training process, thereby achieving fast and accurate prediction of antenna performance. Summary of the Invention

[0006] In view of this, this invention proposes an antenna performance prediction method based on an adaptive naked mole-mole algorithm-based MLP. This method effectively overcomes the time and computational costs of traditional electromagnetic simulation for complex antenna structures. Furthermore, it uses an adaptive naked mole-mole algorithm to replace the optimization process of weights and biases in neural networks, improving the speed and accuracy of weight and bias optimization and avoiding getting trapped in local optima. Thus, it utilizes the inference and computational capabilities of neural networks to solve the time-consuming and labor-intensive problem in antenna optimization design, greatly improving the efficiency of antenna optimization.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An antenna performance prediction method for MLP based on the adaptive naked mole-mole algorithm includes the following steps:

[0009] Step 1: Determine the D antenna design parameter variables used for antenna performance prediction, use the Latin hypercube sampling method to select values ​​for each antenna design parameter, and then use electromagnetic simulation software to obtain the corresponding antenna frequency-performance index simulation curves to further obtain the training dataset.

[0010] Step 2: Construct an MLP network for predicting antenna performance indicators. The input of the MLP network is D antenna design parameter values ​​and one frequency point value, and the output is the predicted value of the antenna performance indicators.

[0011] Step 3: Use the improved naked mole-rat algorithm to iteratively train the weights and biases of neurons in the MLP network. After training, the final MLP network structure is obtained.

[0012] Step 4: In practical applications, based on the final MLP network structure, input the actual values ​​of D antenna design parameters and the current operating frequency, and output the predicted values ​​of antenna performance indicators to complete the antenna performance prediction based on the adaptive naked mole-mole algorithm MLP.

[0013] Furthermore, the specific method in step 1 is as follows:

[0014] Step 101, for the first Antenna design parameters , Let its range of values ​​be denoted as Then, the value range is divided into intervals. , The preset number of sampling points;

[0015] Step 102, regarding the first There are D sampling points, each containing D antenna design parameters, with values ​​as shown below:

[0016]

[0017] Indicates the first The first sampling point in the nth sampling point The values ​​of each antenna design parameter, ; Representing an E-dimensional matrix The first in Each element value Indicates that for the first Antenna design parameters generated from 1 to A random permutation matrix; A random number uniformly distributed in the interval [0,1].

[0018] Step 103, based on the first The D antenna design parameters in the sampling points are used to obtain the corresponding antenna frequency-performance index simulation curve using electromagnetic simulation software. The antenna frequency-performance index simulation curve is the simulated value of the antenna performance index in the entire operating frequency band under the current D antenna design parameter values. Uniform frequency sampling is performed on the entire operating frequency band to obtain F frequency point sampling values, and the corresponding simulated performance index values ​​are recorded.

[0019] Further, a training dataset of size N is obtained, where N = E × F; each training data in the training dataset is the D antenna design parameter values ​​and a frequency sampling value of a sampling point, and the simulated performance index value corresponding to the training data is the corresponding ground truth label.

[0020] Furthermore, the MLP network used to predict antenna performance metrics in step 2 includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0021] The input layer comprises D+1 neurons. ;

[0022] The first hidden layer comprises I neurons, wherein the first hidden layer has I neurons. Weighted input of each neuron for:

[0023]

[0024] in, Represents the first input layer The first neuron is connected to the first hidden layer. The weights of each neuron, It is the first hidden layer. Bias of each neuron ;

[0025] The first hidden layer The output of each neuron for:

[0026]

[0027] The second hidden layer comprises K neurons, wherein the second hidden layer has K neurons. Weighted input of each neuron for:

[0028]

[0029] in, Represents the first hidden layer's... The first neuron is connected to the second hidden layer. The weights of each neuron, It is the second hidden layer. Bias of each neuron ;

[0030] The second hidden layer The output of each neuron for:

[0031]

[0032] The output layer consists of one neuron, and the weighted inputs of the output layer neuron are... for:

[0033]

[0034] in It is the second hidden layer. The weights from each neuron to the output layer neurons. It refers to the bias of the output layer neurons. This is the predicted performance index of the output antenna.

[0035] Furthermore, the specific method for step 3 is as follows:

[0036] Step 301: Initialize the reproduction probability pb, the mutation probability pm, and the naked mole-rat population, where the number of breeding rats B = R / 5 and the number of working rats W = RB; where R is the naked mole-rat population size.

[0037] Step 302: Set the maximum number of iterations (max-iteration) for the naked mole-mole algorithm and start the optimization iteration.

[0038] Furthermore, the formula for initializing the naked mole-rat population in step 301 is as follows:

[0039]

[0040] in, , , This represents the number of all weights and biases in the MLP network. This indicates the first [number] species in the naked mole-rat population. The first individual A dimension variable, that is, the value of one of the weights or biases in an MLP network. and They are the first Lower and upper bounds for dimensional variables, for as well as Its lower and upper bounds take values ​​of respectively and ;against as well as Its lower and upper bounds take values ​​of respectively and ;against as well as Its lower and upper bounds take values ​​of respectively and , This indicates rounding up to the nearest integer.

[0041] Furthermore, the specific steps of step 302 are as follows:

[0042] Step 3021: For each individual in the current naked mole-rat population, calculate its corresponding SHAP value;

[0043] Step 3022: Calculate the fitness function value corresponding to the current naked mole-rat population, update and record the global optimal solution;

[0044] Step 3023: Sort the fitness function values ​​of each individual in the current naked mole rat population, and record the B individuals with the smallest fitness function values ​​as breeding rats in the current naked mole rat population, and the remaining W individuals as working rats in the current naked mole rat population;

[0045] Step 3024: Enter the worker stage, determine the weights based on the SHAP value, and update the worker mouse;

[0046] Step 3025: Enter the breeding stage. Determine whether to move to the optimal position based on the breeding probability pb, and update the breeding mice.

[0047] Step 3026: Enter the mutation phase and perform SHAP adaptive mutation operation on the updated working mice and breeding mice;

[0048] Step 3027: Determine whether the maximum number of iterations (max-iteration) has been reached. If it has, calculate the fitness function value corresponding to the current naked mole rat population, and output the global optimal solution after updating and recording it. If it has not been reached, repeat steps 3021 to 3027.

[0049] Furthermore, the specific steps of step 3021 are as follows:

[0050]

[0051]

[0052]

[0053] in This indicates the number of naked mole rats in the current population. individual The complete set of features consisting of the values ​​of each dimension. Indicates the total set of features Delete from middle The resulting set for Any feature subset, including the empty set, The symbol '!' indicates the number of elements in the set, and '!' indicates factorial. for marginal contribution For feature subset The predicted contribution value is given by the random forest model; and This indicates the current naked mole-rat population's [number]. individual Each dimension The maximum and minimum values ​​among the values; That is, the current naked mole-rat population The first individual The SHAP value of the solution.

[0054] Furthermore, the specific steps of step 3022 are as follows:

[0055] For each individual in the current naked mole-rat population, its corresponding weights and biases are assigned to the MLP network, and then the data from the training dataset are used... Each training data point is sequentially input into the MLP network to obtain the predicted performance index of the corresponding antenna output, and then the corresponding loss value is calculated. Let be the fitness function value corresponding to the current individual:

[0056]

[0057] in, For the first After training data is input into the MLP network, the predicted performance metrics of the output antenna are... For the first The true labels of the training data; the global optimal solution is the minimum fitness function value in all iterations.

[0058] Furthermore, the specific steps of step 3024 are as follows:

[0059]

[0060]

[0061] in, In the t-th iteration, the first... The first working mouse The SHAP value corresponding to the dimensional solution. In the t-th iteration, the first... The first working mouse Solution, This indicates its corresponding updated value. and This represents the first iteration of two random working mice in the t-th iteration. Solution, It is a scaling factor based on LEVY flight, ranging from... Control the update step size. .

[0062] Furthermore, the specific process of step 3025 is as follows:

[0063]

[0064]

[0065] in, In the t-th iteration, the first... The first breeding mouse The SHAP value corresponding to the dimensional solution. In the t-th iteration, the first... The first breeding mouse Solution, This indicates its corresponding updated value. The first individual corresponding to the current global optimal solution Solution, For the t-th iteration The first breeding mouse The random probability of generating a dimensional solution. Let be a random variable that follows a uniform distribution and has a range of [-1, 1]. It is a scaling factor based on LEVY flight, ranging from... Control the update step size. .

[0066] Furthermore, the SHAP adaptive mutation probability in step 3026 is as follows:

[0067] .

[0068] Due to the adoption of the above technical solution, the beneficial effects of this invention compared with the prior art are as follows:

[0069] 1. This invention uses the naked mole-rat algorithm to replace the original gradient descent method to optimize the weights and biases of the MLP multilayer perceptron neural network, which solves the problems of slow convergence speed and easy getting trapped in local optima in the training process of traditional MLP neural networks.

[0070] 2. This invention designs an adaptive naked mole-mole algorithm based on the traditional naked mole-mole algorithm. The algorithm introduces a machine learning interpretation method, which dynamically adjusts the search strategy of the algorithm by measuring the feature contribution in the prediction results, thereby further improving the optimization speed and accuracy of the weights and biases of the MLP neural network.

[0071] 3. This invention proposes a method for predicting antenna performance in the field of antennas, which can effectively reduce the computational load of electromagnetic simulation software, save computational resources, and improve antenna optimization efficiency. Attached Figure Description

[0072] Figure 1 This is an overall flowchart of an antenna performance prediction method based on the adaptive naked mole-mole algorithm in an embodiment of the present invention.

[0073] Figure 2 This is a structural diagram of an MLP neural network.

[0074] Figure 3 A comparison of the errors in weights and biases when using the adaptive naked mole-mole algorithm and when using gradient descent to optimize the MLP.

[0075] Figure 4 A comparison of training time for weights and biases optimized using the adaptive naked mole-mole algorithm and gradient descent is presented.

[0076] Figure 5 This is a schematic diagram comparing the S11 curve predicted by the surrogate model with the actual value obtained by electromagnetic simulation software under the same parameters, given fixed antenna structure parameters in this embodiment of the invention.

[0077] Figure 6This is a schematic diagram comparing the gain curve predicted by the surrogate model with the actual value obtained by electromagnetic simulation software under the same parameters, given fixed antenna structure parameters in an embodiment of the present invention. Detailed Implementation

[0078] The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0079] An antenna performance prediction method based on an adaptive naked mole-mole algorithm (MLP) is proposed, which obtains a surrogate model to facilitate subsequent optimization. While traditional electromagnetic simulation offers high accuracy, it is computationally expensive, hindering large-scale parameter analysis and rapid iteration. To address this, we utilize an MLP neural network to construct a surrogate model of antenna performance, fitting the nonlinear relationship between antenna parameters and performance. In model construction, traditional multilayer perceptron models employ gradient descent to optimize weights and biases, which suffers from slow convergence and susceptibility to local optima. Therefore, we replace this with an improved adaptive naked mole-mole algorithm. This improved algorithm analyzes the system importance of different parameters, dynamically adjusting the optimization strategy to improve overall optimization efficiency and accuracy. Figure 1 As shown, the specific steps include:

[0080] Step 1: Determine the D antenna design parameter variables used for antenna performance prediction, use the Latin hypercube sampling method to select values ​​for each antenna design parameter, and then use electromagnetic simulation software to obtain the corresponding antenna frequency-performance index simulation curves to further obtain the training dataset.

[0081] Step 2: Construct an MLP network for predicting antenna performance indicators. The input of the MLP network is D antenna design parameter values ​​and one frequency point value, and the output is the predicted value of the antenna performance indicators.

[0082] Step 3: Use the improved naked mole-rat algorithm to iteratively train the weights and biases of neurons in the MLP network. After training, the final MLP network structure is obtained.

[0083] Step 4: In practical applications, based on the final MLP network structure, input the actual values ​​of D antenna design parameters and the current operating frequency, and output the predicted values ​​of antenna performance indicators to complete the antenna performance prediction based on the adaptive naked mole-mole algorithm MLP.

[0084] Furthermore, the specific method of step 1 is as follows:

[0085] Step 101, for the first Antenna design parameters , Let its range of values ​​be denoted as Then, the value range is divided into intervals. , Given the preset number of sampling points, the k-th value interval of this variable is:

[0086] , where K=1,2,…,E.

[0087] Step 102, regarding the first There are D sampling points, each containing D antenna design parameters, with values ​​as shown below:

[0088]

[0089] Indicates the first The first sampling point in the nth sampling point The values ​​of each antenna design parameter, ; Representing an E-dimensional matrix The first in Each element value Indicates that for the first Antenna design parameters generated from 1 to A random permutation matrix; A random number uniformly distributed in the interval [0,1].

[0090] Step 103, based on the first The D antenna design parameters in the sampling points are used to obtain the corresponding antenna frequency-performance index simulation curve using electromagnetic simulation software. The antenna frequency-performance index simulation curve is the simulated value of the antenna performance index in the entire operating frequency band under the current D antenna design parameter values. Uniform frequency sampling is performed on the entire operating frequency band to obtain F frequency point sampling values, and the corresponding simulated performance index values ​​are recorded.

[0091] Further, a training dataset of size N is obtained, where N = E × F; each training data in the training dataset is the D antenna design parameter values ​​and a frequency sampling value of a sampling point, and the simulated performance index value corresponding to the training data is the corresponding ground truth label.

[0092] Furthermore, the MLP network used to predict antenna performance metrics in step 2 includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0093] The input layer comprises D+1 neurons. ;

[0094] Specifically, in this embodiment, the number of antenna design parameters D is 3;

[0095] The first hidden layer comprises I neurons, wherein the first hidden layer has I neurons. Weighted input of each neuron for:

[0096]

[0097] in, Represents the first input layer The first neuron is connected to the first hidden layer. The weights of each neuron, It is the first hidden layer. Bias of each neuron ;

[0098] Specifically, in this embodiment, the first hidden layer includes 128 neurons;

[0099] The first hidden layer The output of each neuron for:

[0100]

[0101] The second hidden layer comprises K neurons, wherein the second hidden layer has K neurons. Weighted input of each neuron for:

[0102]

[0103] in, Represents the first hidden layer's... The first neuron is connected to the second hidden layer. The weights of each neuron, It is the second hidden layer. Bias of each neuron ;

[0104] Specifically, in this embodiment, the second hidden layer includes 64 neurons;

[0105] The second hidden layer The output of each neuron for:

[0106]

[0107] The output layer consists of one neuron, and the weighted inputs of the output layer neuron are... for:

[0108]

[0109] in It is the second hidden layer. The weights from each neuron to the output layer neurons. It refers to the bias of the output layer neurons. This is the predicted performance index of the output antenna.

[0110] Furthermore, the specific method for step 3 is as follows:

[0111] Step 301: Initialize the reproduction probability pb, the mutation probability pm, and the naked mole rat population, where the number of breeding rats B = R / 5 and the number of working rats W = RB; where R is the naked mole rat population size and R is an integer multiple of 5.

[0112] Step 302: Set the maximum number of iterations (max-iteration) for the naked mole-mole algorithm and start the optimization iteration.

[0113] Furthermore, the formula for initializing the naked mole-rat population in step 301 is as follows:

[0114]

[0115] in, , , This represents the number of all weights and biases in the MLP network. This indicates the first [number] species in the naked mole-rat population. The first individual A dimension variable, that is, the value of one of the weights or biases in an MLP network. and They are the first Lower and upper bounds for dimensional variables, for as well as Its lower and upper bounds take values ​​of respectively and ;against as well as Its lower and upper bounds take values ​​of respectively and ;against as well as Its lower and upper bounds take values ​​of respectively and , This indicates rounding up to the nearest integer.

[0116] Furthermore, the specific steps of step 302 are as follows:

[0117] Step 3021: For each individual in the current naked mole-rat population, calculate its corresponding SHAP value;

[0118] Step 3022: Calculate the fitness function value corresponding to the current naked mole-rat population, update and record the global optimal solution;

[0119] Step 3023: Sort the fitness function values ​​of each individual in the current naked mole rat population, and record the B individuals with the smallest fitness function values ​​as breeding rats in the current naked mole rat population, and the remaining W individuals as working rats in the current naked mole rat population;

[0120] Step 3024: Enter the worker stage, determine the weights based on the SHAP value, and update the worker mouse;

[0121] Step 3025: Enter the breeding stage. Determine whether to move to the optimal position based on the breeding probability pb, and update the breeding mice.

[0122] Step 3026: Enter the mutation phase and perform SHAP adaptive mutation operation on the updated working mice and breeding mice;

[0123] Step 3027: Determine whether the maximum number of iterations (max-iteration) has been reached. If it has, calculate the fitness function value corresponding to the current naked mole rat population, and output the global optimal solution after updating and recording it. If it has not been reached, repeat steps 3021 to 3027.

[0124] Furthermore, the specific steps of step 3021 are as follows:

[0125]

[0126]

[0127]

[0128] in This indicates the number of naked mole rats in the current population. individual The complete set of features consisting of the values ​​of each dimension. Indicates the total set of features Delete from middle The resulting set for Any feature subset, including the empty set, The symbol '!' indicates the number of elements in the set, and '!' indicates factorial. for marginal contribution For feature subset The predicted contribution value is given by the random forest model; and This indicates the current naked mole-rat population's [number]. individual Each dimension The maximum and minimum values ​​among the values; That is, the current naked mole-rat population The first individual The SHAP value of the solution.

[0129] Furthermore, the specific steps of step 3022 are as follows:

[0130] For each individual in the current naked mole-rat population, its corresponding weights and biases are assigned to the MLP network, and then the data from the training dataset are used... Each training data point is sequentially input into the MLP network to obtain the predicted performance index of the corresponding antenna output, and then the corresponding loss value is calculated. Let be the fitness function value corresponding to the current individual:

[0131]

[0132] in, For the first After training data is input into the MLP network, the predicted performance metrics of the output antenna are... For the first The true labels of the training data; the global optimal solution is the minimum fitness function value in all iterations.

[0133] Furthermore, the specific steps of step 3024 are as follows:

[0134]

[0135]

[0136] in, In the t-th iteration, the first... The first working mouse The SHAP value corresponding to the dimensional solution. In the t-th iteration, the first... The first working mouse Solution, This indicates its corresponding updated value. and This represents the first iteration of two random working mice in the t-th iteration. Solution, It is a scaling factor based on LEVY flight, ranging from... Control the update step size. .

[0137] Furthermore, the specific process of step 3025 is as follows:

[0138]

[0139]

[0140] in, In the t-th iteration, the first... The first breeding mouse The SHAP value corresponding to the dimensional solution. In the t-th iteration, the first... The first breeding mouse Solution, This indicates its corresponding updated value. The first individual corresponding to the current global optimal solution Solution, For the t-th iteration The first breeding mouse The random probability of generating a dimensional solution. Let be a random variable that follows a uniform distribution and has a range of [-1, 1]. It is a scaling factor based on LEVY flight, ranging from... Control the update step size. .

[0141] Furthermore, the SHAP adaptive mutation probability in step 3026 is as follows:

[0142] .

[0143] The SHAP adaptive mutation operation is the same as the mutation operation in the naked mole-mole algorithm, targeting the updated working and breeding mice, and the mutation probability is... .

[0144] To comprehensively evaluate the algorithm's performance, this embodiment selects a filter antenna and constructs a surrogate model for it. Since the bandwidth and gain of the filter antenna are key design considerations, predictions are made on the antenna's S11 curve and gain curve, and the nonlinear relationship between its three structural parameters and antenna performance is fitted. Figure 2 The diagram shows the architecture of the MLP neural network constructed to predict the antenna performance. It has four input parameters: three structural parameters and one operating frequency. It also has two hidden layers and one output layer. The output is the predicted value. It will be compared with the truth label and the weights and biases will be updated using the adaptive naked mole-mole algorithm. Figure 3The results demonstrate the differences in error between using the adaptive naked mole-mole algorithm and gradient descent to optimize weights and thresholds during the construction of an MLP neural network. It can be seen that the error optimized using the adaptive naked mole-mole algorithm is slightly smaller than that using gradient descent, indicating that the MLP neural network model optimized using the adaptive naked mole-mole algorithm has higher prediction accuracy.

[0145] MSE (mean squared error)

[0146] RMSE (Root Mean Square Error)

[0147] MAE (Mean Absolute Error)

[0148] coefficient of determination , For the sample true value, For predicted values, The coefficient of determination is approximately 1, which is the average of the true values ​​of the sample. This indicates that the model fits the sample better.

[0149] Figure 4 The comparison of training time using the adaptive naked mole-mole algorithm and the gradient descent method shows that the adaptive naked mole-mole algorithm takes less time, indicating that it can obtain the surrogate model faster and improve efficiency. Figure 5 The comparison between the predicted S11 curve and the actual S11 curve obtained through electromagnetic simulation software is shown. Figure 6 The comparison between the predicted gain curve and the actual gain curve obtained through electromagnetic simulation software is shown. Figure 5 Figure 6 It can be seen that the surrogate model can predict the antenna's S11 curve and gain curve very well.

[0150] Those skilled in the art will recognize that the described embodiments are intended to help readers understand the principles of the invention and should be understood as not limiting the scope of protection of the invention to the described embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for predicting the performance of an antenna based on an adaptive naked mole-rat algorithm-based MLP, characterized by, The method comprises the following steps: Step 1, determining D antenna design parameter variables for antenna performance prediction, using Latin hypercube sampling method to value each antenna design parameter, and then using electromagnetic simulation software to obtain the corresponding antenna frequency-performance index simulation curve, and further obtaining a training data set; Step 2, constructing an MLP network for predicting the antenna performance index, the input of the MLP network being D antenna design parameter values and a frequency point value, and the output being an antenna performance index prediction value; Step 3, using an improved naked mole rat algorithm to iteratively train the weights and biases of neurons in the MLP network, and obtaining a final MLP network structure after training; specifically: Step 301, initializing a breeding probability pb, a mutation probability pm, and initializing a naked mole rat population, wherein the number of breeding mice in the population B = R / 5, and the number of working mice in the population W = R-B; wherein R is the naked mole rat population size; Step 302, setting the maximum number of iterations max-iteration of the naked mole rat algorithm, and starting optimization iteration: Step 3021: for each individual in the current naked mole rat population, calculate its corresponding SHAP value; Step 3022: calculate the fitness function value corresponding to the current naked mole rat population, update and record the global optimal solution; Step 3023: sort the fitness function values corresponding to each individual in the current naked mole rat population, and record the B individuals with the smallest fitness function values as the breeding mice in the current naked mole rat population, and the remaining W individuals as the working mice in the current naked mole rat population; Step 3024: enter the worker stage, determine the weights according to the SHAP values, and update the working mice; Step 3025: enter the breeding stage, determine whether to move towards the optimal solution according to the breeding probability pb, and update the breeding mice; Step 3026: enter the mutation stage, and perform SHAP adaptive mutation operation on the updated working mice and breeding mice; Step 3027: determine whether the maximum number of iterations max-iteration has been reached, if so, calculate the fitness function value corresponding to the current naked mole rat population, update and record the global optimal solution, and then output the global optimal solution, if not, repeat steps 3021 to 3027; Step 4, in actual application, based on the final MLP network structure, input D antenna design parameter actual values and the current working frequency point value, output the antenna performance index prediction value, and complete the MLP antenna performance prediction based on the adaptive naked mole rat algorithm.

2. The method of claim 1, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The specific method in step 1 is: Step 101, for the first antenna design parameter , , its value range is recorded as , then the value interval division interval , is the preset sampling point number; Step 102, for the first sample point, which contains D antenna design parameters, the values are as follows: ； represents the value of the i-th antenna design parameter in the j-th sample point, represents the value of the i-th element in the E-dimensional matrix represents the value of the i-th element in the E-dimensional matrix represents the random permutation matrix generated for the i-th antenna design parameter from 1 to is a random number uniformly distributed in the interval [0, 1].​​​​​​ Step 103, based on the first The D antenna design parameters in the sampling points are used to obtain the corresponding antenna frequency-performance index simulation curve using electromagnetic simulation software. The antenna frequency-performance index simulation curve is the simulated value of the antenna performance index in the entire operating frequency band under the current D antenna design parameter values. Uniform frequency sampling is performed on the entire operating frequency band to obtain F frequency point sampling values, and the corresponding simulated performance index values ​​are recorded. Further obtain a training data set with a size of N, N = E x F; Each training data in the training data set is D antenna design parameter values and a frequency point sampling value of a sampling point, and the performance index simulation value corresponding to the training data is the true value label.

3. The method of claim 1, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The MLP network for predicting the antenna performance index in step 2 comprises an input layer, a first hidden layer, a second hidden layer, and an output layer; The input layer comprises D+1 neurons ; The first hidden layer comprises I neurons, wherein the weighted input of the i-th neuron of the first hidden layer is: ​​ ； wherein, represents a weight of the i-th neuron of the input layer to the j-th neuron of the first hidden layer, is a bias of the j-th neuron of the first hidden layer, ;​​​ The output of the first hidden layer's first neuron is: ​ ； The second hidden layer comprises K neurons, wherein the weighted input to the j-th neuron of the second hidden layer is: K = 2.​ ； wherein, represents a weight of the i-th neuron of the first hidden layer to the j-th neuron of the second hidden layer, is a bias of the i-th neuron of the second hidden layer, ;​​​ The output of the first neuron of the second hidden layer is: :​ ； The output layer includes 1 neuron, the weighted input of the output layer neuron is is: ； in It is the second hidden layer. The weights from each neuron to the output layer neurons. It refers to the bias of the output layer neurons. This is the predicted performance index of the output antenna.

4. The method of claim 3, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The formula for initializing the naked mole rat population in step 301 is as follows: ； wherein, , , is the number of all weights and biases in the MLP network, represents the value of the th variable of the th individual in the naked mole-rat population, i.e. one of the weights or biases of the MLP network, and are the lower and upper bounds of the th variable, respectively, for and , whose lower and upper bounds are and , respectively; for and , whose lower and upper bounds are and , respectively; for and , whose lower and upper bounds are and , represents the ceiling function.

5. The method of claim 4, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The specific steps of step 3021 are as follows: ； ； ； where represents the feature set composed of the values of the dimensions of the th individual in the current naked mole-rat population, represents the set obtained by deleting from the feature set , is an arbitrary feature subset of , including the empty set, represents the number of elements in the set, and represents the factorial; is the marginal contribution of is the prediction contribution value of the feature subset , given by the random forest model; is the SHAP value of the th individual in the current naked mole-rat population, th individual in the current naked mole-rat population, th individual in the current naked mole-rat population, th individual in the current naked mole-rat population, th individual in the current naked mole-rat population, th individual in the current naked mole-rat population, th individual in the current naked mole-rat population.

6. The method of claim 3, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The specific steps of step 3022 are as follows: For each individual in the current naked mole-rat population, its corresponding weight and bias value is assigned to the MLP network, and then the training data set The individual training data is input into the MLP network in turn, and the performance index prediction value of the corresponding output antenna is obtained, and then the corresponding loss value is calculated, which is the fitness function value corresponding to the current individual: ； wherein, is the performance index prediction value of the antenna output by the MLP network after the first training data input into the MLP network, is the true value label of the first training data; the global optimal solution is the minimum fitness function value in all iterations.

7. The method of claim 5, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The specific steps of the step 3024 are as follows: ； ； wherein, represents the SHAP value of the t-th iteration of the j-th working mouse for the i-th feature, dimensional solution, represents the i-th feature of the t-th iteration of the j-th working mouse, dimensional solution, represents its corresponding updated value, and represents the i-th feature of the t-th iteration of the j-th working mouse, is a scaling factor based on LEVY flight, ranging from , controlling the update step size, .​​​ 8. The method of claim 5, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The specific process of the step 3025 is as follows: ； ； wherein, represents the SHAP value of the t-th iteration of the j-th breeding mouse, represents the t-th iteration of the j-th breeding mouse, represents the corresponding updated value thereof, represents the t-th iteration of the j-th breeding mouse, represents the random probability generated for the t-th iteration of the j-th breeding mouse, is a random variable subject to a uniform distribution, ranging from [-1, 1], is a scaling factor based on LEVY flight, ranging from , controls the update step size, .​​​​​​​ 9. The method of claim 5, wherein the method is based on an adaptive naked mole-rat algorithm (ANMA) for MLP. The SHAP adaptive mutation probability in the step 3026 is as follows: 。

Images