Intelligent prediction method for sub-terahertz propagation coefficient of dual-algorithm optimized DNN

Abstract

The invention relates to an intelligent prediction method for a sub-terahertz propagation coefficient of a dual-algorithm optimized DNN, and belongs to the field of communication technologies and deep learning. The technical problems that a traditional propagation model depends on actual measurement and is poor in generalization ability are solved. High-frequency electromagnetic waves and materials are complex in action; artificial setting of the neural network is easy to fall into local optimum. The method specifically comprises the following steps: constructing a DNN model to capture a nonlinear relationship between electromagnetic waves and a material; actually measuring the reflection / transmission coefficient of the sub-terahertz frequency band material to construct a data set; fusing a simulated annealing algorithm to optimize a network structure; optimizing a weight initial value by a genetic algorithm; and training the model and evaluating the performance based on MAE, RMSE, MAPE and error probability. According to the invention, the prediction precision and the model generalization ability are significantly improved through a dual optimization mechanism; limitation of a traditional simplified model is broken through; complicated electromagnetic characteristics are autonomously learned; and environment modeling and beam management of the 6G communication perception integrated system are supported.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of communication technology and deep learning, and relates to a dual-algorithm-optimized intelligent prediction method for a sub-terahertz propagation coefficient of a deep neural network (DNN). BACKGROUND

[0002] With the rapid development of the sixth generation mobile communication technology 6G, the communication system not only faces the challenges of higher data rates and lower delays, but also needs to undertake new tasks of environmental perception and intelligent interaction. The ISAC system integrating perception and communication as a core direction of 6G aims to realize the dual functions of efficient communication and accurate perception through the same hardware and spectrum resources. In this framework, the propagation characteristics generated by the interaction between electromagnetic waves and environmental objects become a key information source for perceiving the environment. As a core parameter describing the interaction characteristics of electromagnetic waves and materials, the propagation coefficient directly determines the communication link quality and the perception accuracy, and has important value for the environmental modeling, beam management and dynamic link adaptation of the ISAC system.

[0003] In the 6G sub-terahertz frequency band, the interaction between electromagnetic waves and materials presents a high degree of complexity. Especially in a multipath environment such as an indoor environment, the material density, surface roughness, incident angle, frequency and polarization mode and other characteristics significantly affect the propagation coefficient, and the traditional prediction method based on a simplified physical model faces serious challenges:

[0004] Strong dependence on actual measurement: the existing technology needs to measure the reflection / transmission coefficients of different materials through a vector network analyzer and other devices, which has high experimental cost and long cycle;

[0005] Insufficient generalization ability: the empirical formula based on limited measured data fitting is difficult to adapt to the changing actual building environment, and the prediction error significantly increases for new materials or complex surface characteristics that have not been seen before;

[0006] Poor adaptability to high frequencies: the short wavelength characteristics of the sub-terahertz frequency band exacerbate the nonlinear effects of electromagnetic scattering, and traditional linear or semi-empirical models cannot accurately capture the complex coupling relationship between high-frequency electromagnetic waves and materials.

[0007] Although current research attempts to introduce machine learning methods, the unoptimized deep neural network is prone to local optimal solution, and the network structure and initial parameters depend on artificial experience setting, resulting in unstable prediction accuracy. Therefore, an intelligent prediction method for the propagation coefficient is needed, which can automatically learn the complex nonlinear relationship, adapt to the characteristics of the high frequency band, and have strong generalization ability, to support the actual deployment and application of the 6G communication and perception integrated (ISAC) system. SUMMARY

[0008] In view of this, the purpose of the present application is to provide a dual-algorithm optimized DNN sub-terahertz propagation coefficient intelligent prediction method for accurately predicting the reflection and transmission coefficients of building materials in a 6G communication environment. The proposed method is based on ISAC, and combines simulated annealing algorithm and genetic algorithm to optimize the deep neural network, so as to improve the prediction accuracy and adaptability in complex material electromagnetic characteristic modeling tasks.

[0009] To achieve the above purpose, the present application provides the following technical solutions:

[0010] A dual-algorithm optimized DNN sub-terahertz propagation coefficient intelligent prediction method, comprising the following steps:

[0011] S1: Construct a DNN-based propagation coefficient prediction model, which includes an input layer, a hidden layer and an output layer, for capturing the nonlinear relationship between electromagnetic waves and material scatterers;

[0012] S2: Construct a propagation coefficient measured data set, and measure the reflection and transmission coefficients of different materials based on the frequency domain measurement method of a vector network analyzer (VNA);

[0013] S3: Optimize the model structure and initial parameters, including:

[0014] (a) Use the simulated annealing algorithm to optimize the number of hidden layers, the number of neurons and the learning rate of the DNN model;

[0015] (b) Use the genetic algorithm to optimize the initial values of the weights and biases of the DNN model;

[0016] S4: Train the optimized DNN model, and evaluate the prediction performance based on the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), prediction probability P err5 and prediction probability P err1 with a relative error of less than 1%.

[0017] Further, in the S1, the DNN training process comprises:

[0018] S11: Initialize the number of neurons in the input layer, hidden layer and output layer, and the number of hidden layers, and perform forward propagation calculation;

[0019] S12: Calculate the error of the output layer according to the loss function, and propagate the error back to the hidden layer, and calculate the neuron gradient layer by layer;

[0020] S13: Update the network parameters based on the gradient direction, and the update formula is:

[0021]

[0022] Among them, W l is the weight matrix of the lth layer, b l is the bias vector of the lth layer, η is the learning rate, and J is the loss function.

[0023] Furthermore, in the S11, the output of the neurons in the first layer is a (l) Calculated as:

[0024] a (l) =σ(W (l) a (l-1) +b (l) ) (1)

[0025] in, is the weight matrix, n l is the number of neurons in this layer, n l-1 is the number of neurons in the upper layer, a (l) is the output vector of the lth layer; is the bias vector output by this layer; σ(·) is the activation function, and ReLU is selected as the activation function between connecting layers;

[0026] Furthermore, the S2 specifically includes:

[0027] S21: Build a measurement platform consisting of a transmitter, receiver, RF front end, signal generator, and VNA;

[0028] S22: Fix the material to be tested, adjust the angle of the transmitting and receiving antennas, record the average value of five scans, and save the reflection coefficient and transmission coefficient

[0029] S23: Switch the antenna polarization mode and repeat the measurement;

[0030] S24: Record air interface measurement data To eliminate the impact of path loss and antenna gain;

[0031] S25: Yes and Perform time domain threshold filtering, the formula is:

[0032] S 21_process =FFT[IFFT(S 21_initial )·W k ]

[0033] Among them, S 21_process is the processed frequency domain data, S 21_initial is the original measurement data, FFT(·) is the fast Fourier transform, IFFT(·) is the frequency domain data after inverse fast Fourier transform, W ka time domain window function;

[0034] S26: Calculate the reflection and transmission coefficients based on the double measurement method:

[0035] S27: Construct a dataset with material density, surface correlation length, surface height standard deviation, incident angle, frequency and polarization mode as input, and reflection / transmission coefficient as output.

[0036] Further, in the S25, the filtered The reflection coefficient R is calculated by the following formula mea and the transmission coefficient T mea :

[0037]

[0038] wherein, is the processed reflection measurement data, is the processed transmission measurement data, ω is the angular frequency, ω = 2πf i , f i is the signal frequency, and i represents the frequency index.

[0039] Further, the S3 comprises:

[0040] S31: Set the initial temperature T max , the termination temperature T min and the cooling rate of the simulated annealing algorithm, search for the optimal number of hidden layers, number of neurons and learning rate in the hyperparameter space;

[0041] S32: Generate a new hyperparameter combination to accept a better solution with the loss function as the evaluation index;

[0042] S33: Flatten the DNN parameters into a one-dimensional array, and generate new initial values through single-point crossover and Gaussian mutation of the genetic algorithm;

[0043] S34: When the maximum number of iterations or early stopping conditions are reached, output the optimal initial weight and bias.

[0044] Further, the S4 further comprises:

[0045] S41: Input the network structure optimized by S3 and the initial parameters into the DNN model;

[0046] S42: Normalize and divide the dataset into 60% training set, 20% validation set and 20% test set;

[0047] S43: Train the model using the Adam optimizer, and save the optimal model based on the validation set early stopping.

[0048] Further, the data preprocessing of the S42 comprises:

[0049] (1) flatten the raw data and normalize;

[0050] (2) save the normalizer for inverse transform of prediction results.

[0051] Further, in the S4, the evaluation index calculation formula is:

[0052]

[0053] Where, MAE is the mean absolute error, RMSE is the root mean square error, MAPE is the mean absolute percentage error, err is the relative error, N is the total number of test samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample;

[0054] 1(·) is an indicator function, which is 1 when the condition is true, and 0 otherwise; err i is the relative error of the i-th sample.

[0055] The beneficial effects of the present application are:

[0056] (1) Abandoning the dependence on simplified physical models or high-frequency assumptions, the deep neural network is driven based on the measured data to autonomously learn the complex nonlinear action mechanism of electromagnetic waves and materials, which significantly improves the prediction accuracy and model generalization ability, and adapts to the changing actual communication sensing scene.

[0057] (2) Innovatively combining simulated annealing algorithm and genetic algorithm:

[0058] Simulated annealing globally searches for the optimal network structure, and adaptively determines the number of hidden layers, neuron size and learning rate;

[0059] Genetic algorithm intelligently optimizes the initial values of weights and biases to avoid training into local optimal solution;

[0060] The double algorithm cooperation significantly enhances the model robustness and convergence efficiency.

[0061] (3) High-precision prediction of reflection / transmission coefficients provides key environmental information for ISAC system:

[0062] Enabling dynamic beamforming and link adaptation to improve communication reliability;

[0063] Building a precise environmental perception database to enhance target recognition and positioning ability;

[0064] Realize the cooperative optimization of communication performance and sensing accuracy.

[0065] (4) Based on the normalization data preprocessing and automatic parameter optimization mechanism, the model training process does not need manual repeated parameter adjustment;

[0066] The lightweight network structure and the automatic process greatly reduce the consumption of computing resources and accelerate the integration of practical system applications.

[0067] Other advantages, objects, and features of the application will be set forth in part in the following specification taken in conjunction with the accompanying drawings, and in part will become apparent to those skilled in the art from a consideration of the following specification and drawings. The objects and other advantages of the application will be realized and attained by means of the instrumentalities and combinations pointed out in the following specification. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to make the objects, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0069] Figure 1 The flowchart of the present application;

[0070] Figure 2 The schematic diagram of the principle of the present application;

[0071] Figure 3 The schematic diagram of the DNN network structure;

[0072] Figure 4 The performance result diagram of the prediction model on the verification set;

[0073] Figure 5 The comparison diagram of the propagation coefficient prediction results of the optimized DNN network and the traditional DNN network. DETAILED DESCRIPTION

[0074] The embodiments of the present application are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0075] The drawings are only used for illustrative explanation, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation of the present application; in order to better illustrate the embodiments of the present application, some components in the drawings are omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some known structures and their descriptions in the drawings can be omitted.

[0076] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the orientations or positional relationships indicated by the terms "upper", "lower", "left", "right", "front", "back", etc. are based on the orientations or positional relationships shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationships in the drawings are only used for exemplary illustration and cannot be understood as a limitation on the present application, for those skilled in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0077] Please refer to Figures 1-5 , a sub-terahertz propagation coefficient intelligent prediction method for optimizing DNN by a double algorithm, the method comprising the following steps:

[0078] Step 1: Considering the complex interaction between electromagnetic waves and material scatterers, a material propagation coefficient prediction model based on DNN network is constructed, including input layer, hidden layer and output layer.

[0079] Step 2: Construct the propagation coefficient measured data set. Based on the frequency domain measurement method of the vector network analyzer, the reflection coefficient and transmission coefficient of different materials are measured to construct the propagation coefficient data set for subsequent model training.

[0080] Step 3: Network structure optimization and optimal initial value determination. First, the structure and hyperparameters of the model are optimized based on the simulated annealing algorithm; second, the genetic algorithm is used to determine the optimal initial value of the model to avoid the model falling into a local optimal solution.

[0081] Step 4: Performance evaluation. The optimized DNN model is used for training and compared with the traditional DNN model to verify the accuracy of the proposed method for intelligent prediction of material propagation coefficient from the aspects of mean absolute error, root mean square error, mean absolute percentage error and relative error.

[0082] In step 1 of the present embodiment, it specifically comprises the following processes:

[0083] Step 11: The DNN training process is divided into forward propagation, backward propagation and parameter update stages. Before the forward propagation starts, the number of input layer, hidden layer and output layer neurons and the number of hidden layers are initialized. For the lth layer of neurons, its output z (l) is calculated as follows:

[0084] z (l) = σ(W (l) a (l-1) +b (l) ) (1)

[0085] wherein, is the weight matrix nl is the number of neurons in the current layer, nl -1 is the number of neurons in the upper layer; is the bias vector of the current layer output; σ(·) is an activation function, and ReLU is selected as the activation function between the connection layer and the layer in order to capture the nonlinear relationship between the electromagnetic wave and the material scatterer. The ReLU function is used to introduce a nonlinear transformation, so that the neural network can learn more complex patterns and features.

[0086] Step 12: Back propagation. Starting from the output layer, the error of the output layer is calculated according to the loss function, and then the error information is propagated back to the hidden layer, and the error gradient of each neuron is calculated layer by layer. When the mean square error is selected as the loss function, the expectation is to minimize: where x and y are the input and output of the training sample, respectively. The gradient of the output layer W and b can be solved:

[0087]

[0088] The gradient of the lth layer can be calculated by the output layer:

[0089]

[0090] According to the chain rule, the value of can be calculated:

[0091]

[0092] Through the above process, the gradient value of each layer parameter can be calculated.

[0093] Step 13: After obtaining the gradient value of the previous layer, the network parameters that can make the loss function of this layer decrease the fastest in this round of training can be obtained by updating along the gradient direction. The parameter update method is:

[0094]

[0095] Through multiple iterations and layer-by-layer updates, the loss function value gradually decreases, thereby accurately predicting the target.

[0096] In step 2 of the embodiment, the following processes are specifically included:

[0097] Step 21: The sub-terahertz frequency domain measurement platform is composed of a transmitter (Tx), a receiver (Rx), two radio frequency front ends (RF), a signal generator and a vector network analyzer. Before starting the measurement, the relevant configurations need to be completed, including the measurement frequency, signal power level, sweep interval and intermediate frequency bandwidth, etc. Before connecting the IF signal to the RF front end with a cable, the VNA needs to be calibrated using the short circuit, open circuit, load and direct connection calibration methods to eliminate the influence of the frequency response of the cable, adapter and VNA internal on the measurement results. After calibration, the RF front end is connected in turn for DC bias, local oscillator signal and intermediate frequency signal, and the platform is built.

[0098] Step 22: After selecting the measurement material and antenna polarization mode, fix the measured material on the center horizontal axis of the measurement graph. Use the horizontal laser instrument to adjust the measurement angle of the transmitter and receiver, and ensure that the transceiver is always close to the measurement arc during adjustment to ensure that the measurement radius is always consistent. In order to reduce measurement error, for each measurement angle, the average value of five scan data is taken as the measurement result, which is saved. The reflection measurement result of the material is recorded as The transmission measurement result is recorded as After the measurement of the material at all angles is completed, replace the measured material and repeat the above steps until the measurement of all materials under the antenna polarization is completed.

[0099] Step 23: Change the antenna polarization mode and repeat Step 2 until the measurement of all materials at all angles under horizontal and vertical polarization is completed.

[0100] Step 24: After all materials are measured, keep the system configuration of the measurement platform unchanged, separate the transceiver antennas by a distance of twice the measurement radius, and align the center of the horn antenna. Record a set of open mouth measurement data as a control group, recorded as The purpose of adding the control group measurement is to facilitate the elimination of the effects of free space path loss, antenna gain and RF front end frequency response in subsequent data processing.

[0101] Step 25: After the measurement is completed, the measurement data needs to be processed to obtain the sub-terahertz reflection and transmission coefficients of the building materials. During the measurement, due to the reflection of electromagnetic waves between the antenna and the building materials, as well as the scattering of other objects in the measurement environment, other multipath signals will be mixed in the received signal. In order to eliminate the influence of multipath interference signals on the frequency domain measurement data, time domain threshold technology is needed to filter them out. The steps of time domain threshold technology are as follows: first, the original measurement data S 21The inverse fast Fourier transform is applied to convert the time domain data into frequency domain data. Then, a time domain window function is applied to filter the time domain data. Finally, the filtered data is converted back to frequency domain data by fast Fourier transform. This process can be represented as:

[0102] S 21_process = FFT [IFFT (S 21_initial · W k ) ] (9)

[0103] where S 21_initial represents the original measurement data and S 21_process represents the processed data and FFT(·) and IFFT(·) represent fast Fourier transform and inverse fast Fourier transform, respectively, and W k represents a time domain window function.

[0104] To obtain the reflection and transmission coefficients of building materials, the frequency response of the radio frequency front end, the antenna gain, and the path loss need to be removed from the measurement data. Since the frequency response of the radio frequency front end cannot be accurately obtained, the data processing steps based on the double measurement method are selected in this paper. and are the ratios of the received IF signal electric field strength to the transmitted IF signal electric field strength when measuring the building materials. Compared with the control group measurement data , they contain more reflection or transmission loss caused by the measured materials. Therefore, the reflection and transmission coefficients can be represented as the ratio of the measurement data of the building materials to the air interface measurement data, i.e.:

[0105]

[0106] In the formula, ω = 2πf i , i represents the frequency index.

[0107] Step 26: To determine the surface statistical properties of the selected material sample, i.e. the standard deviation and correlation length of the surface height distribution, the Leeb beam splitting roughness meter is used to measure the surface roughness and characterize the rough surface. The Leeb handheld roughness meter estimates the hardness and roughness of an object by measuring the rebound speed of the surface of the object, has a 320 μm high-precision electric sensor, and the Leeb handheld roughness meter displays the measurement results on its screen.

[0108] Step 27: Construct the data set required for training the intelligent prediction model of the propagation coefficient. The input features of the data set are: material density, material surface correlation length, material surface height distribution standard deviation, incident angle, system frequency, and polarization mode; the output of the data set is the reflection coefficient and the transmission coefficient.

[0109] In step 3 of the embodiment, it specifically includes the following process:

[0110] Step 31: Global search of hyperparameters using simulated annealing algorithm. Set the initial temperature T max , the minimum temperature T min , and the cooling rate. Set the hyperparameter space, which includes the number of hidden layers, learning rate, number of neurons, etc.

[0111] Step 32: Randomly generate a set of combined hyperparameter combinations in the hyperparameter space, and record the training loss of the set of hyperparameters. Randomly select a hyperparameter according to the neighborhood generating function, replace it with another possible value in the candidate set, and generate a new combination. Calculate the loss of the new solution, if the new solution is better, accept the new solution, otherwise accept the new solution with a certain probability. If the new solution is better than the previous global optimal solution, update the global optimal solution. After each round, cool down according to the cooling rate, until the minimum or early stopping condition is reached.

[0112] Step 33: Optimize the best weight and bias initial value using genetic algorithm. First, define the fitness function to evaluate the performance of the model on the test set, output the inverse normalized MSE as the fitness score; second, flatten the initial parameters of DNN into a one-dimensional array, and randomly generate the initial value in a certain range to complete the initialization; then, select parents according to the fitness distribution probability and use single-point crossover and Gaussian disturbance random mutation method to quickly combine parents, generate new structures, fine-tune individuals, enhance diversity and local search ability.

[0113] Step 34: Return the optimal initial weight and bias when the maximum number of iterations is reached or the early stopping condition is met, complete the genetic algorithm parameter optimization.

[0114] In step 4 of the embodiment, it specifically includes the following process:

[0115] Step 41: Input the network structure and optimal initial value optimized by simulated annealing algorithm and genetic algorithm into the initial DNN network.

[0116] Step 42: Data loading and preprocessing

[0117] (1) Read the dataset through the loading function, and get the training data by grouping (divided into X train , Y train ).

[0118] (2) Flatten the original data of X train and Y train , normalize it, and reshape it back to its original shape, while saving the normalizer.

[0119] (3) Divide 60% training set, 20% test set and 20% validation set, convert to tensor and use data loader to encapsulate into iterable batch data.

[0120] Step 43: Propagation coefficient prediction model training

[0121] (1) Set training parameters, loss function and optimizer (Adam).

[0122] (2) Define a function to train for a single training round (epoch) and return the training error.

[0123] (3) Calculate the parameter gradient according to the training error and update the parameters according to the gradient direction.

[0124] (4) Perform training in the defined training and validation loop, and set early stopping mechanism (stop if validation loss does not improve within a certain round).

[0125] (5) Save the current model whenever a better validation loss is obtained;

[0126] (6) Load the saved best model.

[0127] (7) Use the validation set for prediction inference, collect and splice the output and target value.

[0128] (8) Perform inverse normalization operation on the prediction result and target result.

[0129] (9) Calculate and output evaluation indexes such as mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), prediction probability of relative error ≤5% (P err5 ) and prediction probability of relative error ≤1% (P err1 ) and record them to the log, and the evaluation index expressions are as follows:

[0130]

[0131] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs, characterized by: The following steps are involved: S1: Construct a DNN-based propagation coefficient prediction model, which includes an input layer, a hidden layer, and an output layer to capture the nonlinear relationship between the interaction between electromagnetic waves and material scatterers; S2: Construct a propagation coefficient measurement dataset and measure the reflection and transmission coefficients of different materials using the frequency domain measurement method using a vector network analyzer (VNA). S3: Optimize model structure and initial parameters, including: (a) Using simulated annealing algorithm to optimize the number of hidden layers, number of neurons and learning rate of the DNN model; (b) optimizing the initial values ​​of weights and biases of the DNN model using a genetic algorithm; S4: Train the optimized DNN model and calculate the prediction probability P based on the mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE, and relative error ≤ 5%. err5 and the prediction probability P with relative error ≤ 1% err1 Evaluate prediction performance.

2. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 1, characterized in that: In S1, the DNN training process includes: S11: Initialize the number of neurons in the input layer, hidden layer, and output layer, as well as the number of hidden layers, and perform forward propagation calculations; S12: Calculate the output layer error based on the loss function, and backpropagate the error to the hidden layer, calculating the neuron gradient layer by layer; S13: Update network parameters based on the gradient direction. The update formula is: Among them, W l is the weight matrix of the lth layer, b l is the bias vector of the lth layer, η is the learning rate, and J is the loss function.

3. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 2, characterized in that: In the S11, the output of the neurons in the first layer is a (l) Calculated as: a (l) =σ(W (l) a (l-1) +b (l) ) (1) in, is the weight matrix, n l is the number of neurons in this layer, n l-1 is the number of neurons in the upper layer, a (l) is the output vector of the lth layer; is the bias vector output by this layer; σ(·) is the activation function, and ReLU is selected as the activation function between connecting layers.

4. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 1, characterized in that: The S2 specifically includes: S21: Build a measurement platform consisting of a transmitter, receiver, RF front end, signal generator, and VNA; S22: Fix the material to be tested, adjust the angle of the transmitting and receiving antennas, record the average value of five scans, and save the reflection coefficient and transmission coefficient S23: Switch the antenna polarization mode and repeat the measurement; S24: Record air interface measurement data To eliminate the impact of path loss and antenna gain; S25: Yes and Perform time domain threshold filtering, the formula is: S 21_process =FFT[IFFT(S 21_initial )·W k ] Among them, S 21_process is the processed frequency domain data, S 21_initial is the original measurement data, FFT(·) is the fast Fourier transform, IFFT(·) is the frequency domain data after inverse fast Fourier transform, W k is the time domain window function; S26: Calculate the reflection coefficient and transmission coefficient based on the dual measurement method: S27: Construct a data set with material density, surface correlation length, surface height standard deviation, incident angle, frequency, and polarization as input and reflection / transmission coefficient as output.

5. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 4, characterized in that: In the S25, the filtered The reflection coefficient R is calculated by the following formula mea and the transmission coefficient T mea : in, is the processed reflection measurement data, is the processed transmission measurement data, ω is the angular frequency, ω=2πf i , f i is the signal frequency, and i represents the frequency index.

6. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 1, characterized in that: The S3 includes: S31: Set the initial temperature T of the simulated annealing algorithm max , termination temperature T min and cooling rate, searching for the optimal number of hidden layers, number of neurons, and learning rate in the hyperparameter space; S32: Generate new hyperparameter combinations and accept the better solution using the loss function as the evaluation indicator; S33: Flatten the DNN parameters into a one-dimensional array and generate new initial values ​​through single-point crossover and Gaussian mutation of the genetic algorithm; S34: When the maximum number of iterations or the early stopping condition is reached, the optimal initial weights and biases are output.

7. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 6, characterized in that: Said S4 further comprises: S41: Input the network structure and initial parameters optimized in S3 into the DNN model; S42: Normalize the dataset and divide it into 60% training set, 20% validation set and 20% test set; S43: Use the Adam optimizer to train the model and save the optimal model based on the validation set early stopping.

8. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 7, characterized in that: The data preprocessing of S42 includes: (1) Flatten the original data and normalize it; (2) Save the normalizer for inverse transformation of the prediction results.

9. The method for intelligently predicting sub-THz propagation coefficients using dual-algorithm optimized DNNs according to claim 1, characterized in that: In S4, the evaluation index calculation formula is: Among them, MAE is the mean absolute error, RMSE is the root mean square error, MAPE is the mean absolute percentage error, err is the relative error, N is the total number of test samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample; 1(·) is an indicator function, which is 1 when the condition is met and 0 otherwise; err i is the relative error of the i-th sample.

Images