A method and device for predicting terahertz scattering parameters based on neural network
By training and filtering multiple neural network models, and combining error calculations based on medium type, emission angle, and reception angle, the neural network structure was optimized, solving the problem of insufficient model applicability in terahertz communication systems. This enabled high-precision prediction of scattering parameters and reduced testing costs and time.
Patent Information
- Application Number
- CN202211019125.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-24
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-08-24
AI Technical Summary
In the existing technology, terahertz communication systems lack the support of measured data, making it difficult for neural network models to be applied to different media. Furthermore, existing neural network training algorithms are insufficient in terms of training speed and accuracy, and cannot meet the channel prediction requirements of complex scenarios.
By acquiring a preset test set, multiple neural network models are trained and tested. Predictions are made based on a three-dimensional vector composed of medium type, emission angle, and reception angle. The mean absolute error and root mean square error are calculated, and neural network models suitable for different media are selected. The model structure is optimized by combining backpropagation neural network and genetic algorithm to improve prediction accuracy.
It achieves high-precision prediction of terahertz scattering parameters under different media, solves the problem of insufficient applicability of a single model, reduces the cost and time of terahertz testing, and improves the applicability and accuracy of channel modeling.
Smart Images

Figure CN115499074B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technology, and in particular to a terahertz scattering parameter prediction method and device based on a neural network. Background Art
[0002] The 6G mobile communication system has been a hot topic of research at home and abroad in recent years. Since the 6G communication system has typical big data characteristics and machine learning is a powerful technology for big data processing and development, combining traditional channel research with neural networks for high-precision channel prediction is a hot topic in the research of the new generation of communications.
[0003] Wireless channels are the basis of radio wave propagation. In wireless mobile communications, all information must be transmitted through wireless channels. However, wireless channels are susceptible to interference during transmission and have great randomness, making it difficult to accurately analyze wireless communications. Due to the characteristics of ultra-wide spectrum, ultra-large-scale antenna array, high heterogeneity and many new application scenarios of 6G communication systems, the communication characteristics in the study of terahertz frequency band communication channels can be analyzed and processed digitally. Therefore, in the future practical application of terahertz communication, neural network technology and improved algorithms can be better applied to terahertz frequency band communication research to achieve high-precision prediction and analysis of electromagnetic parameters. Due to the high cost of terahertz test equipment, under the complex application scenario requirements of 6G in the future, the acquisition of measured data for specific scenarios has a certain threshold. The measured data sets above 100GHz are still rare and precious. Accurate and practical data sets are the basis of neural network modeling. Therefore, the research process needs to rely on more test data to support neural network training, and the researched model is also difficult to apply to the scattering parameter analysis of different media. Summary of the invention
[0004] In view of this, an embodiment of the present invention provides a terahertz scattering parameter prediction method based on a neural network to eliminate or improve one or more defects existing in the prior art.
[0005] One aspect of the present invention provides a terahertz scattering parameter prediction method based on a neural network, the method comprising the following steps:
[0006] Obtain a preset test set, input the input parameters in the test set into a plurality of pre-trained neural network models, respectively, to obtain predicted scattering parameters corresponding to the input parameters, the test set including the input parameters and real scattering parameters corresponding to the input parameters, the input parameters being a three-dimensional vector consisting of a medium type, a transmission angle, and a receiving angle, the real scattering parameters being predicted scattering parameters of a terahertz wave transmitted by a transmitting antenna and received by a receiving antenna; the transmission angle being the angle of the transmitting antenna, the receiving angle being the angle of the receiving antenna; the medium type being the type of medium between the transmitting antenna and the receiving antenna;
[0007] Dividing the predicted scattering parameters obtained by the multiple neural network models into multiple comparison groups according to preset conditions, and calculating the mean absolute error and root mean square error between the predicted scattering parameters and the true scattering parameters of each neural network model in the same comparison group, wherein the preset conditions include the medium type;
[0008] Based on the mean absolute error and the root mean square error, multiple neural network models are screened to obtain the neural network model corresponding to the medium type;
[0009] Matching a corresponding neural network model among multiple neural network models based on the medium used in the actual measurement;
[0010] The input vector consisting of the medium type, emission angle and receiving angle in the actual measurement is input into the corresponding neural network model to obtain the corresponding actual predicted scattering parameters.
[0011] Using the above scheme, since different neural network models have different effects on different media, this scheme first trains different neural network models, and tests multiple neural network models according to the test set to obtain the media that are applicable to different neural network models. In the actual measurement process, different neural network models are used according to the actual medium type. This scheme solves the problem that a single model is difficult to apply to different media. Different neural network models are matched according to the medium type to ensure the applicability of the model to the medium.
[0012] In some embodiments of the present invention, the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group comprises:
[0013] The predicted scattering parameters output by multiple neural network models in the same comparison group are marked in different two-dimensional coordinate systems according to the types of the neural network models, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters;
[0014] The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
[0015] In some embodiments of the present invention, the step of screening multiple neural network models based on the mean absolute error and the root mean square error to obtain a neural network model corresponding to the medium type includes:
[0016] Calculate the sum of the mean absolute error and root mean square error of the same neural network model in each comparison group;
[0017] The neural network model corresponding to the minimum value of the sum of the mean absolute error and the root mean square error is used as the neural network model corresponding to the medium type of the comparison group.
[0018] In some embodiments of the present invention, in the step of dividing into multiple comparison groups according to preset conditions, the preset conditions also include an emission angle or a receiving angle. If the preset conditions also include an emission angle or a receiving angle, the predicted scattering parameters are divided into multiple comparison groups based on the same medium type and the same emission angle or receiving angle.
[0019] In some embodiments of the present invention, in the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the mean absolute error is calculated according to the following formula:
[0020]
[0021] MAE stands for mean absolute error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, y r represents the value of the i-th true scattering parameter.
[0022] In some embodiments of the present invention, in the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the root mean square error is calculated according to the following formula:
[0023]
[0024] RMSE stands for root mean square error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, Represents the average value of n real scattering parameter values in the same two-dimensional coordinate system, ω i represents the weight of the i-th real scattering parameter.
[0025] In some embodiments of the present invention, in the step of inputting the input parameters in the test set into a plurality of pre-trained neural network models, the plurality of neural network models include a first network model, a second network model, and a third network model;
[0026] The structures of the first network model and the second network model adopt the model structure of the BP neural network. The first network model uses the mean absolute error or the root mean square error as the loss function and adopts the gradient descent method for training; the second network model adopts the genetic algorithm for training;
[0027] The third network model adopts the model structure of the RBF neural network. The third network model uses the mean absolute error or the root mean square error as the loss function and adopts the gradient descent method for training; the second network model adopts the genetic algorithm for training.
[0028] In some embodiments of the present invention, the structures of the first network model and the second network model are composed of an input layer, a hidden layer and an output layer, the hidden layer includes a first layer and a second layer, the first layer uses a tansig activation function, and the second layer uses a purelin linear activation function; the structure of the third network model is composed of an input layer, a hidden layer whose neuron transformation function is composed of radial basis functions, and an output layer.
[0029] In some embodiments of the present invention, the method further comprises pre-training a plurality of neural network models, and the step of pre-training a plurality of neural network models comprises:
[0030] Based on the real scattering parameters measured in advance by a vector network analyzer and the medium type, emission angle and receiving angle when the scattering parameters are measured as a data unit, a training set and a test set including multiple data units are constructed;
[0031] Pre-train multiple neural network models based on the training set.
[0032] Another aspect of the present invention provides a terahertz scattering parameter prediction device based on a neural network, which includes a computer device, wherein the computer device includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the method described above.
[0033] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention can be specifically pointed out and obtained in the specification and the accompanying drawings.
[0034] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.
[0036] Figure 1 A schematic diagram of an implementation of a terahertz scattering parameter prediction method based on a neural network according to the present invention;
[0037] Figure 2 This is a schematic diagram of the first test solution of the present invention;
[0038] Figure 3 This is a schematic diagram of a second test scheme of the present invention;
[0039] Figure 4 This is a schematic diagram of the third test scheme of the present invention;
[0040] Figure 5 This is a schematic diagram of the fourth test scheme of the present invention;
[0041] Figure 6 This is a schematic diagram of the fifth test scheme of the present invention;
[0042] Figure 7 This is a schematic diagram of the sixth test scheme of the present invention;
[0043] Figure 8 Schematic diagram of the acquisition system for acquiring test set and training set data in the present invention. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0045] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.
[0046] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.
[0047] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.
[0048] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.
[0049] Introduction to existing technology
[0050] The existing technology is based on intuitive geographical area characteristics, uses a large amount of channel data in the base station area to train the neural network, and uses different combinations of measured features at the base station to infer unknown parameters at the user, thereby predicting channel parameters for wireless channel characteristics that cannot be directly observed at the base station.
[0051] The shortcomings of the existing technology 1 are that the existing technology 1 is aimed at data analysis and prediction of communication systems below 100 GHz. The research on the terahertz communication field using neural networks lacks the support of measured data. Currently, the channel data of 200 Hz-300 Hz has not been fully measured. The application of neural networks to the channels in this frequency band still needs to be comprehensively studied in combination with measured data.
[0052] The second prior art studies multi-layer perceptron (MLP) and radial basis function (RBF) neural networks, introduces the back propagation training algorithm (BP) and the particle swarm optimization (PSO) algorithm for neural network training into the above neural networks, uses the field strength measurement value of any point for training, and adjusts the training algorithm parameters to make sufficiently accurate field strength predictions in complex indoor environments.
[0053] The disadvantages of the second existing technology are that the network structure of the data-driven channel prediction technology based on the BP neural network cannot be determined and the training speed is slow. It is easy to fall into the local minimum during training and cannot converge to the global minimum error point. It lacks targeted optimization for specific data sets. Most of the data predictions are for the main lobe data, and the error control and prediction accuracy of the side lobe data still need to be improved.
[0054] To solve the above problems, Figure 1 As shown, the present invention proposes a terahertz scattering parameter prediction method based on a neural network; the steps of the method include:
[0055] Step S100, obtaining a preset test set, inputting the input parameters in the test set into a plurality of pre-trained neural network models respectively, and obtaining predicted scattering parameters corresponding to the input parameters, wherein the test set includes the input parameters and real scattering parameters corresponding to the input parameters, wherein the input parameters are three-dimensional vectors composed of a medium type, a transmission angle, and a receiving angle, and the real scattering parameters are predicted scattering parameters of the terahertz wave transmitted by a transmitting antenna and received by a receiving antenna; the transmission angle is the angle of the transmitting antenna, and the receiving angle is the angle of the receiving antenna; and the medium type is the type of the medium between the transmitting antenna and the receiving antenna;
[0056] In some embodiments of the present invention, the test set includes real scattering parameters measured by a vector network analyzer, and the medium type, emission angle and receiving angle when the real scattering parameters are measured. The medium type is encoded, and the encoded medium type, emission angle and receiving angle are combined into a three-dimensional vector, which are input as data vectors into multiple pre-trained neural network models to obtain different predicted scattering parameters output by multiple neural network models for the same input parameters.
[0057] The scattering parameters (S parameters, Scatter parameters) include an input reflection coefficient (S11), namely an input return loss, an output reflection coefficient (S22), namely an output return loss, a reverse transmission coefficient (S12) and a forward transmission coefficient (S21), namely a gain.
[0058] The scattering parameter may be any one of the various scattering parameters mentioned above. The method may perform model training on any scattering parameter, match the corresponding model, and predict the scattering parameter.
[0059] Step S200, dividing the predicted scattering parameters obtained by the multiple neural network models into multiple comparison groups according to preset conditions, and calculating the mean absolute error and root mean square error between the predicted scattering parameters and the true scattering parameters of each neural network model in the same comparison group, wherein the preset conditions include the medium type;
[0060] In some embodiments of the present invention, when dividing into a plurality of comparison groups according to a preset condition, outputting predicted scattering parameters according to input parameters of the same medium type is divided into one comparison group according to the medium type;
[0061] The input parameters of each predicted scattering parameter are obtained in the same comparison group, and the emission angle and the receiving angle in the input parameters are obtained. The real scattering parameters are matched in the comparison group according to the emission angle and the receiving angle of the input parameters. The predicted scattering parameters output by multiple neural network models in the same comparison group are marked in different two-dimensional coordinate systems based on the type of the neural network model, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters.
[0062] The predicted scattering parameters and the real scattering parameters corresponding to each other in the two-dimensional coordinate system are in the same comparison group, and the emission angle and the receiving angle are equal. The abscissa of the two-dimensional coordinate system is the emission angle or the receiving angle, and the ordinate is the scattering parameter value.
[0063] The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
[0064] In some embodiments of the present invention, when the comparison groups are divided according to the medium type based on the preset conditions, the emission angle or the receiving angle of each real scattering parameter of the constructed two-dimensional coordinate system may be different.
[0065] Step S300, screening multiple neural network models based on mean absolute error and root mean square error to obtain a neural network model corresponding to the medium type;
[0066] Step S400, matching a corresponding neural network model among multiple neural network models based on the medium used in the actual measurement;
[0067] In some embodiments of the present invention, the mean absolute error and the root mean square error can both reflect the difference between the predicted scattering parameters and the actual scattering parameters. Under the premise of the same medium, a model with a smaller difference is more suitable for the medium, and a neural network model suitable for different media can be screened out, thus solving the problem that a single model is difficult to apply to different media.
[0068] Step S500, inputting an input vector consisting of the medium type, emission angle and receiving angle in actual measurement into a corresponding neural network model to obtain corresponding actual predicted scattering parameters.
[0069] In some embodiments of the present invention, the medium type may be air, cardboard of any specifications, glass, wood board, foam board, etc., and the arbitrary specifications include any thickness, number of sheets, and area.
[0070] In some embodiments of the present invention, the transmitting angle is the angle of a transmitting antenna, and the receiving angle is the angle of a receiving antenna.
[0071] In some embodiments of the present invention, when the transmitting antenna and the receiving antenna are in the same straight line, the transmitting angle and the receiving angle are both 0 degrees, the angle at which the transmitting antenna deviates from the straight line is the transmitting angle, and the angle at which the receiving antenna deviates from the straight line is the receiving angle.
[0072] Using the above scheme, since different neural network models have different effects on different media, this scheme first trains different neural network models, and tests multiple neural network models according to the test set to obtain the media that are applicable to different neural network models. In the actual measurement process, different neural network models are used according to the actual medium type. This scheme solves the problem that a single model is difficult to apply to different media. Different neural network models are matched according to the medium type to ensure the applicability of the model to the medium.
[0073] In some embodiments of the present invention, the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group comprises:
[0074] The predicted scattering parameters output by multiple neural network models in the same comparison group are marked in different two-dimensional coordinate systems according to the types of the neural network models, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters;
[0075] The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
[0076] In some embodiments of the present invention, input parameters of each predicted scattering parameter are obtained in the same comparison group, and the emission angle and the receiving angle in the input parameters are obtained, and the real scattering parameters in the comparison group are matched according to the emission angle and the receiving angle of the input parameters, and the predicted scattering parameters output by multiple neural network models in the same comparison group are respectively marked in different two-dimensional coordinate systems based on the type of the neural network model, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters.
[0077] The predicted scattering parameters and the real scattering parameters corresponding to each other in the two-dimensional coordinate system are in the same comparison group, and the emission angle and the receiving angle are equal. The abscissa of the two-dimensional coordinate system is the emission angle or the receiving angle, and the ordinate is the scattering parameter value.
[0078] The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
[0079] In some embodiments of the present invention, when the comparison groups are divided according to the medium type based on a preset condition, the emission angle or the receiving angle of each real scattering parameter of the constructed two-dimensional coordinate system may be different.
[0080] In some embodiments of the present invention, the step of screening multiple neural network models based on the mean absolute error and the root mean square error to obtain a neural network model corresponding to the medium type includes:
[0081] Calculate the sum of the mean absolute error and root mean square error of the same neural network model in each comparison group;
[0082] The neural network model corresponding to the minimum value of the sum of the mean absolute error and the root mean square error is used as the neural network model corresponding to the medium type of the comparison group.
[0083] In some embodiments of the present invention, in the step of screening multiple neural network models based on the mean absolute error and the root mean square error, the neural network models can be screened according to the sum of the above-mentioned mean absolute error and the root mean square error, or by the product or average of the mean absolute error and the root mean square error.
[0084] By adopting the above scheme, the mean absolute error and the root mean square error can be combined as the screening condition of the neural network model, instead of considering one error alone, thereby improving the screening accuracy of the neural network model applicable to the same medium.
[0085] In the specific implementation process, through the test steps, it can be concluded that the model suitable for a1 medium is b2 model, the model suitable for a2 medium is b3 model, and the model suitable for a3 medium is b1 model. When actually measuring, the corresponding model is matched according to different types of media, and the final actual predicted scattering parameters are output.
[0086] In some embodiments of the present invention, in the step of dividing into multiple comparison groups according to preset conditions, the preset conditions also include an emission angle or a receiving angle. If the preset conditions also include an emission angle or a receiving angle, the predicted scattering parameters are divided into multiple comparison groups based on the same medium type and the same emission angle or receiving angle.
[0087] In some embodiments of the present invention, if the preset condition further includes an emission angle or a receiving angle, the input parameter output predicted scattering parameters of the same medium type and the same emission angle or receiving angle are divided into a comparison group;
[0088] The input parameters of each predicted scattering parameter are obtained in the same comparison group, and the emission angle and the receiving angle in the input parameters are obtained. The real scattering parameters are matched in the comparison group according to the emission angle and the receiving angle of the input parameters. The predicted scattering parameters output by multiple neural network models in the same comparison group are marked in different two-dimensional coordinate systems based on the type of the neural network model, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters.
[0089] The predicted scattering parameters and the actual scattering parameters corresponding to each other in the two-dimensional coordinate system are in the same comparison group, and the emission angle and the receiving angle are equal. If the preset condition also includes the emission angle or the receiving angle, the ordinate of the two-dimensional coordinate system is the scattering parameter value, if the preset condition is the emission angle, the abscissa is the receiving angle, and if the preset condition is the receiving angle, the abscissa is the emission angle.
[0090] The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
[0091] Using the above scheme, if the preset conditions also include the emission angle or the receiving angle, then in the input parameters of multiple predicted scattering parameters and real scattering parameters in the two-dimensional coordinate system, there is the same emission angle or receiving angle. On this basis, this scheme can be further screened on the basis of the same medium type, and the neural network models corresponding to different emission angles or receiving angles under the same medium condition can be screened out, and the neural network model used for input parameter matching can be further improved. The applicability of the neural network model for different input parameters is improved.
[0092] In some embodiments of the present invention, if the preset conditions also include an emission angle or a receiving angle, then the same medium type also includes multiple comparison groups based on different emission angles or receiving angles, and if the preset conditions also include an emission angle, then in the same comparison group, the medium type and the emission angle are the same, and the receiving angles are different; if the preset conditions also include a receiving angle, then in the same comparison group, the medium type and the receiving angle are the same, and the emission angles are different.
[0093] In the specific implementation process, through the test steps, it can be concluded that the model applicable to a1 medium with an emission angle of 0° is the b1 model; the model applicable to a1 medium with an emission angle of 30° is the b2 model; the model applicable to a1 medium with an emission angle of -30° is the b3 model. When actual measurement is performed and a1 medium is used, the corresponding model is matched according to different emission angles to output the final actual predicted scattering parameters.
[0094] In some embodiments of the present invention, during the test process, if the angle range of the transmitting angle and the receiving angle is selected to be -30° to 30°, during the test data acquisition process of the same medium condition, the transmitting angle or the receiving angle may be increased or decreased by 3° each time;
[0095] If the medium during the test is a 4-layer paperboard, the preset conditions also include the emission angle, the emission angle is fixed to 0°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the third network model can be as follows: Figure 2 As shown;
[0096] If the medium during the test is a 4-layer wooden board, the preset conditions also include the emission angle, the emission angle is fixed to 0°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the first network model can be as follows: Figure 3 As shown;
[0097] If the medium during the test is a 4-layer wooden board, the preset conditions also include the emission angle, the emission angle is fixed to 0°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the second network model can be as follows: Figure 4 As shown;
[0098] If the medium during the test is a 4-layer wooden board, the preset conditions also include the emission angle, the emission angle is fixed to 30°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the third network model can be as follows: Figure 5 As shown;
[0099] If the medium during the test is a 4-layer paperboard, the preset conditions also include the emission angle, the emission angle is fixed to 30°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the first network model can be as follows: Figure 6 As shown;
[0100] If the medium during the test is a 4-layer paperboard, the preset conditions also include the emission angle, the emission angle is fixed to 30°, and the scattering parameter is the S21 parameter, the two-dimensional coordinate system image of the second network model can be as follows: Figure 7 shown.
[0101] In some embodiments of the present invention, in the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the mean absolute error is calculated according to the following formula:
[0102]
[0103] MAE stands for mean absolute error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, y r represents the value of the i-th true scattering parameter.
[0104] Using the above scheme, the mean absolute error of the real scattering parameters or the predicted scattering parameters of each two-dimensional coordinate system is calculated respectively.
[0105] In some embodiments of the present invention, in the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the root mean square error is calculated according to the following formula:
[0106]
[0107] RMSE stands for root mean square error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, Represents the average value of n real scattering parameter values in the same two-dimensional coordinate system, ω i represents the weight of the i-th real scattering parameter.
[0108] Using the above scheme, the root mean square error of the real scattering parameter or the predicted scattering parameter of each two-dimensional coordinate system is calculated respectively.
[0109] In some embodiments of the present invention, in the step of inputting the input parameters in the test set into a plurality of pre-trained neural network models, the plurality of neural network models include a first network model, a second network model, and a third network model;
[0110] The structures of the first network model and the second network model adopt the model structure of the BP neural network. The first network model uses the mean absolute error or the root mean square error as the loss function and adopts the gradient descent method for training; the second network model adopts the genetic algorithm for training;
[0111] The third network model adopts the model structure of the RBF neural network. The third network model uses the mean absolute error or the root mean square error as the loss function and adopts the gradient descent method for training; the second network model adopts the genetic algorithm for training.
[0112] By adopting the above scheme, the genetic algorithm (GA) can be used as an evolutionary algorithm to retain useful parts and remove useless parts in the optimization process of the data training process of the present invention, which has significant effects on the error control of signals collected by VNA, and can reduce the weight of data that has no learning significance in the data set. GA is used to solve practical problems by finding the solution that best meets the conditions required by the problem among all possible solutions, that is, finding an optimal solution. The genetic algorithm optimizes the bp neural network prediction algorithm, which is divided into three parts:
[0113] (1) Determine the structure of the BP neural network: Determine the topological structure of the BP neural network according to the number of fitting function input and output parameters (in the present invention, the medium type, emission angle and receiving angle, and the output is the S21 parameter value);
[0114] (2) Genetic algorithm optimization: Encode the network parameters of the BP neural network, determine the initial population, and then use the genetic algorithm to optimize the parameter values of the neural network. Through selection, crossover, and mutation, find the optimal fitness individual, that is, the optimal parameter combination of the neural network;
[0115] (3) BP neural network prediction: The network is trained based on the optimal neural network weights and thresholds of the optimal individual.
[0116] The present invention uses a genetic algorithm to obtain the optimal neural network parameters to predict the S parameters of the transmission medium, thereby improving the accuracy of bp neural network prediction.
[0117] The network parameters of the BP neural network include neural network weights and thresholds.
[0118] In some embodiments of the present invention, the structures of the first network model and the second network model are composed of an input layer, a hidden layer and an output layer, the hidden layer includes a first layer and a second layer, the first layer uses a tansig activation function, and the second layer uses a purelin linear activation function; the structure of the third network model is composed of an input layer, a hidden layer whose neuron transformation function is composed of radial basis functions, and an output layer.
[0119] In some embodiments of the present invention, the expression of the activation function R of the radial basis neural network is:
[0120]
[0121] The expression for output y is:
[0122]
[0123] Where ||dist|| represents the Euclidean distance between the input vector x and the weight vector ω, the input vector is a three-dimensional vector consisting of the medium type, the transmission angle, and the receiving angle, ω is a vector consisting of the weight parameters in the third network model, e is the Euler number, y j represents the output parameter of the output layer, h represents the number of neurons, a represents the ath neuron, ω aj Represents the weight parameter of the a-th neuron.
[0124] In step S500, y j To actually predict the scattering parameters.
[0125] In some embodiments of the present invention, the training step of the RBF neural network also includes adjusting the network parameters to obtain the expansion speed and maximum number of neurons of the optimal RBF network radial basis, specifically: in the RBF network creation process, the larger the expansion speed of the radial basis function setting, the smoother the function fitting, but at the same time the network requires a lot of neurons to adapt to the rapid changes of the function. If the expansion speed is set too small, it means that many neurons are needed to adapt to the slow changes of the function, which will affect the performance of the designed network, so it is necessary to find the optimal expansion speed during the parameter adjustment process. In the network creation process, it is necessary to continuously increase the number of neurons in the middle layer until the output error of the network meets the preset value or reaches the set maximum number of neurons, and the maximum number of neurons will affect the predicted error value.
[0126] In some embodiments of the present invention, the method further comprises pre-training a plurality of neural network models, and the step of pre-training a plurality of neural network models comprises:
[0127] Based on the real scattering parameters measured in advance by a vector network analyzer and the medium type, emission angle and receiving angle when the scattering parameters are measured as a data unit, a training set and a test set including multiple data units are constructed;
[0128] Pre-train multiple neural network models based on the training set.
[0129] In some embodiments of the present invention, in the step of measuring the real scattering parameters using a vector network analyzer, a vector network analyzer is used. Figure 8 The measurement structure includes a vector network analyzer (VNA), two spread spectrum modules connected to a transmitting end and a receiving end of the vector network analyzer, a local oscillator module connected to the two spread spectrum modules, a transmitting antenna (Tx antenna) connected to the spread spectrum module of the transmitting end, and a receiving antenna (Rx antenna) connected to the spread spectrum module of the receiving end;
[0130] The media to be tested of different media types are placed between the transmitting antenna and the receiving antenna.
[0131] The spectrum spreading module adopts a terahertz spectrum spreader for spreading the frequency to the terahertz band; the local oscillator module includes an oscillator for generating oscillator signals for the two spectrum spreading modules.
[0132] In some embodiments of the present invention, on a plane with the same viewing angle, when the transmitting antenna and the receiving antenna are in the same straight line, the transmitting angle and the receiving angle are both 0 degrees, the angle at which the transmitting antenna deviates from the straight line is the transmitting angle, the angle at which the receiving antenna deviates from the straight line is the receiving angle, and the angle on one side of the dividing line with the straight line is positive, and the angle on the other side is negative.
[0133] In some embodiments of the present invention, in the process of obtaining the training set, 1-4 pieces of cardboard (60cm*60cm*6mm), wood board (45cm*60cm*6mm), and foam board (55cm*55cm*5cm) are used as the medium to be tested without obstruction (air), and the signal frequency is selected from 220GHz to 320GHz. By changing the angle between the transmitting end and the receiving end (-30° to 30°), the VNA collects the S21 parameter values of the terahertz signal under different media in real time, and finally generates the measured terahertz channel parameter training set and test set.
[0134] Under the current situation where the cost of measuring terahertz channel data is high and other necessary hardware equipment is in short supply, this scheme provides a high-precision neural network prediction scheme for terahertz channel parameters based on partial measured data. The present invention solves the problem of incomplete measured data of terahertz channels. Based on test data sets of propagation media of different materials and combined with neural network technology, it predicts and analyzes the characteristics of electromagnetic propagation parameters in more scenarios in the terahertz frequency band.
[0135] The present invention accurately predicts the parameters of the terahertz channel. Aiming at the high cost and high test threshold of terahertz testing, an S parameter prediction method based on an improved neural network is proposed. Part of the measured data is used to predict specific target data. Through comparative analysis of multiple algorithms, the optimal model is selected to achieve the best prediction effect. The method of the present invention can save the cost and time of terahertz testing, and predict the untested data through the measured data, which is helpful for the modeling design of terahertz channels and the optimization of channel models.
[0136] An embodiment of the present invention also provides a terahertz scattering parameter prediction device based on a neural network, which includes a computer device, wherein the computer device includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the method described above.
[0137] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps implemented by the above-mentioned terahertz scattering parameter prediction method based on a neural network are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.
[0138] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.
[0139] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.
[0140] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.
[0141] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A terahertz scattering parameter prediction method based on neural network, characterized in that: The steps of the method include: Obtain a preset test set, and input the input parameters in the test set into a plurality of pre-trained neural network models, wherein the plurality of neural network models include a first network model, a second network model, and a third network model; the structures of the first network model and the second network model adopt the model structure of a bp neural network, and the first network model uses the mean absolute error or the root mean square error as a loss function and adopts a gradient descent method for training; the second network model is trained using a genetic algorithm; the third network model adopts the model structure of an RBF neural network, and the third network model uses the mean absolute error or the root mean square error as a loss function and adopts a gradient descent method for training; obtain predicted scattering parameters corresponding to the input parameters; the test set includes input parameters and real scattering parameters corresponding to the input parameters, and the input parameters are three-dimensional vectors composed of a medium type, an emission angle, and a receiving angle; the real scattering parameters are predicted scattering parameters of terahertz waves emitted by a transmitting antenna and received by a receiving antenna; the emission angle is the angle of the transmitting antenna, and the receiving angle is the angle of the receiving antenna; the medium type is the type of medium between the transmitting antenna and the receiving antenna; Dividing the predicted scattering parameters obtained by the multiple neural network models into multiple comparison groups according to preset conditions, and calculating the mean absolute error and root mean square error between the predicted scattering parameters and the true scattering parameters of each neural network model in the same comparison group, wherein the preset conditions include the medium type; Based on the mean absolute error and the root mean square error, multiple neural network models are screened to obtain the neural network model corresponding to the medium type; Matching a corresponding neural network model among multiple neural network models based on the medium used in the actual measurement; The input vector consisting of the medium type, emission angle and receiving angle in the actual measurement is input into the corresponding neural network model to obtain the corresponding actual predicted scattering parameters.
2. The terahertz scattering parameter prediction method based on neural network according to claim 1, characterized in that: The step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group comprises: The predicted scattering parameters output by multiple neural network models in the same comparison group are marked in different two-dimensional coordinate systems according to the types of the neural network models, and the corresponding real scattering parameters are marked based on the input parameters of the predicted scattering parameters; The mean absolute error and root mean square error of multiple predicted scattering parameters and true scattering parameters in each two-dimensional coordinate system are calculated.
3. The terahertz scattering parameter prediction method based on neural network according to claim 1, characterized in that: The step of screening multiple neural network models based on the mean absolute error and the root mean square error to obtain a neural network model corresponding to the medium type includes: Calculate the sum of the mean absolute error and root mean square error of the same neural network model in each comparison group; The neural network model corresponding to the minimum value of the sum of the mean absolute error and the root mean square error is used as the neural network model corresponding to the medium type of the comparison group.
4. The terahertz scattering parameter prediction method based on neural network according to claim 1, characterized in that: In the step of dividing into multiple comparison groups according to preset conditions, the preset conditions also include the emission angle or the receiving angle. If the preset conditions also include the emission angle or the receiving angle, the predicted scattering parameters are divided into multiple comparison groups based on the same medium type and the same emission angle or the receiving angle.
5. The method for predicting terahertz scattering parameters based on a neural network according to any one of claims 1 to 4, characterized in that: In the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the mean absolute error is calculated according to the following formula: MAE stands for mean absolute error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, y r represents the value of the i-th true scattering parameter.
6. The method for predicting terahertz scattering parameters based on a neural network according to any one of claims 1 to 4, characterized in that: In the step of calculating the mean absolute error and the root mean square error of the predicted scattering parameters and the true scattering parameters of each neural network model of the same comparison group, the root mean square error is calculated according to the following formula: RMSE stands for root mean square error, n stands for the number of true scattering parameters or predicted scattering parameters in the same two-dimensional coordinate system, i stands for the i-th true scattering parameter or predicted scattering parameter, y p represents the value of the i-th predicted scattering parameter, Represents the average value of n real scattering parameter values in the same two-dimensional coordinate system, ω i represents the weight of the i-th real scattering parameter.
7. The method for predicting terahertz scattering parameters based on neural network according to claim 1, characterized in that: The structures of the first network model and the second network model are composed of an input layer, a hidden layer and an output layer, the hidden layer includes a first layer and a second layer, the first layer uses a tansig activation function, and the second layer uses a purelin linear activation function; the structure of the third network model is composed of an input layer, a hidden layer whose neuron transformation function is composed of radial basis functions, and an output layer.
8. The method for predicting terahertz scattering parameters based on neural network according to claim 1, characterized in that: The method further comprises pre-training a plurality of neural network models, wherein the pre-training of a plurality of neural network models comprises: Based on the real scattering parameters measured in advance by a vector network analyzer and the medium type, emission angle and receiving angle when the scattering parameters are measured as a data unit, a training set and a test set including multiple data units are constructed; Pre-train multiple neural network models based on the training set.
9. A terahertz scattering parameter prediction device based on a neural network, characterized in that: The apparatus comprises a computer device, wherein the computer device comprises a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the apparatus implements the steps implemented by the method as described in any one of claims 1 to 8.
