Sensor signal processing method and system based on metasurface and complex-valued neural network

Abstract

The present invention relates to a sensor signal processing method and system based on a super surface and a complex-valued neural network, wherein the method is based on a complex-valued neural network after training and dynamic optimization, processes the S11 coefficient and the S21 coefficient, and outputs the dielectric constant or solution concentration of the sample; the complex-valued neural network includes an input layer, a preprocessing layer, a plurality of hidden layers and an output layer; the S11 coefficient and the S21 coefficient are obtained by microwave signals passing through the sample and the super surface in sequence; the system includes a transmitting antenna, a receiving antenna, a vector network analyzer, a super surface and a sensor signal processing module, which are used to implement the above method. Compared with the prior art, the present invention has the advantages of being able to realize the measurement of multiple types of substances and having high measurement accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of dielectric constant measurement and solution concentration measurement, and in particular to a sensor signal processing method and system based on a metasurface and a complex-valued neural network. Background Art

[0002] Microwave sensing technology has unique advantages in material property detection and solution concentration measurement. By receiving, processing and analyzing microwave signals, non-contact detection and identification of target materials can be achieved. Different technologies can be used to measure dielectric constants according to the properties of the sample and the frequency range of interest. In recent years, sensors based on structures such as electromagnetic bandgap (EBG), high-impedance surface (HIS), split ring resonator (SRR) and complementary split ring resonator (CSRR) have gradually attracted attention. These sensors adjust the reflection phase or resonant frequency by utilizing the change in the dielectric properties of the material to detect different substances. However, the above-mentioned sensing structures cannot accurately predict small differences in dielectric constants, and the detection accuracy and range are limited. In addition, the traditional linear fitting method can only roughly estimate the range of change of the dielectric constant and cannot achieve high-precision prediction. In order to improve the detection accuracy, some studies have proposed methods combining neural networks, such as a non-invasive sub-terahertz glucose concentration measurement system based on the Levenberg-Marquardt algorithm (LM) combined with a back propagation neural network (BPNN). Among them, the Chinese patent application "CN116087625A" discloses a dielectric constant measurement method based on an open coaxial probe and a neural network. The open coaxial probe is combined with a deep learning network to measure the dielectric constant, and the Debye formula and the admittance formula are used to generate data sets to train the neural network. Although the dielectric constant measurement is achieved and the test accuracy is ensured, it still has the following problems: 1) The detection range is limited and it is only applicable to liquids; 2) During the neural network training process, the network parameters need to be manually adjusted to achieve the best performance of the network, which has great limitations and low training efficiency, thereby affecting the neural network processing accuracy and efficiency.

[0003] Therefore, providing a method that has a wide detection range and can measure information efficiently and accurately is a technical problem that needs to be solved. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to provide a sensor signal processing method and system based on a metasurface and a complex-valued neural network. The combination of the metasurface and the complex-valued neural network improves the sample detection performance. The metasurface enhances the offset sensitivity of the resonant frequency to the dielectric constant of the sample or the change of solution concentration, and the complex-valued neural network can improve the detection accuracy and range by extracting amplitude and phase information.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to a first aspect of the present invention, a sensor signal processing method based on a metasurface and a complex-valued neural network is provided, wherein the method processes the S11 coefficient and the S21 coefficient based on a complex-valued neural network after training and dynamic optimization, and outputs the dielectric constant of the sample or the solution concentration; the complex-valued neural network comprises an input layer, a preprocessing layer, a plurality of hidden layers and an output layer; the S11 coefficient and the S21 coefficient are obtained by microwave signals passing through the sample and the metasurface in sequence; wherein the dynamic optimization performs the following steps until the optimization target is reached:

[0007] After each training cycle, the ratio of the number of samples correctly predicted by the complex-valued neural network in the test set to the total number of samples is obtained. If the ratio is less than a preset value, multiple new weight nodes are added to the hidden layer that is closest to the input layer and meets the total number of weight nodes condition, and the next round of training is continued; the total number of weight nodes condition is that the total number of weight nodes of the hidden layer after the new weight nodes are added is less than or equal to the preset maximum value corresponding to the layer; the initial preset maximum value of each hidden layer is the same, and the preset maximum value of the hidden layer with the new weight node is dynamically updated after each round of training;

[0008] If the total number of weight nodes of all hidden layers of the complex-valued neural network does not meet the total number of weight nodes condition, a new hidden layer is added after the last hidden layer to continue the next round of training;

[0009] If the ratio is greater than or equal to the preset value, the dynamic optimization is terminated.

[0010] As a preferred technical solution, the dynamic update is that each time a training cycle is executed, the preset maximum value is increased by a preset amount, and the preset amount is less than the number of new weight nodes added.

[0011] As a preferred technical solution, the dynamic optimization further includes, when the total number of added hidden layers and initial hidden layers is greater than or equal to a layer number threshold, expanding the range of values ​​of the preset maximum value.

[0012] As a preferred technical solution, the training includes:

[0013] Obtain the S11 coefficient and S21 coefficient of the sample with known dielectric constant or solution concentration, use the S11 coefficient and S21 coefficient in the preset frequency band as a data set, and divide it into a training set and a test set; use the known dielectric constant or solution concentration as a training label;

[0014] Initialize the weight parameters of the complex-valued neural network, the expression is:

[0015] ,

[0016] in, represents the weight parameter, represents the number of neurons in the input layer, represents the number of output neurons, represents normal distribution;

[0017] Repeat the following steps until you reach your training goal:

[0018] The data in the training set is processed by the preprocessing layer and then input into the hidden layer for forward propagation. The activation function during the forward propagation is:

[0019] ,

[0020] in, represents the complex vector of the i-th neuron in the current layer of the complex-valued neural network, The modulus of the complex vector representing the i-th neuron;

[0021] The loss function is calculated, and the gradient of the loss function of each layer parameter is calculated in turn starting from the output layer using the back propagation algorithm, and the weight parameters are updated based on the gradient.

[0022] As a preferred technical solution, the method for outputting the dielectric constant of the sample or the solution concentration includes:

[0023] The input layer extracts a plurality of frequency point data from the S11 coefficient and the S21 coefficient within a preset frequency band, wherein the frequency point data includes the amplitude and phase of the frequency point;

[0024] The preprocessing layer normalizes the amplitude and phase of each frequency point data, and then forms a complex value vector with the amplitude and phase of each frequency point data after normalization.

[0025] The complex-valued vector is input into a hidden layer, processed by the hidden layer and the processing result is transmitted to the next hidden layer;

[0026] The output layer receives the processing results of the last hidden layer and outputs the dielectric constant of the sample or the complex mapping value of the solution concentration.

[0027] As a preferred technical solution, the normalization processing includes: linearly normalizing the amplitude to 0-1, and linearly normalizing the phase to 0°-180°.

[0028] According to a second aspect of the present invention, a sensor signal processing system based on a metasurface and a complex-valued neural network is provided, characterized in that the system comprises a transmitting antenna, a receiving antenna, a vector network analyzer, a metasurface and a sensor signal processing module, wherein the transmitting antenna is used to transmit a microwave signal to a sample, and the microwave signal is received by a receiving antenna after passing through the sample and the metasurface in sequence;

[0029] The receiving antenna is used to receive microwave signals and transmit them to the vector network analyzer;

[0030] The vector network analyzer processes the microwave signal sent by the receiving antenna to obtain the S11 coefficient and the S21 coefficient;

[0031] The metasurface introduces a surface current at the contact surface with the sample to control the microwave signal, so as to enhance the offset sensitivity of the resonance frequency of the scattering parameter to the change of the parameter to be measured;

[0032] The sensor signal processing module executes the above method, processes the S11 coefficient and the S21 coefficient using a complex-valued neural network, and outputs the dielectric constant of the sample or the solution concentration.

[0033] As a preferred technical solution, the metasurface comprises an array of N×N units periodically arranged in both horizontal and vertical directions, wherein N represents the number of units in each row or column of the array.

[0034] As a preferred technical solution, adjacent units are not in direct contact, and each of the units includes a metal layer and a dielectric layer; the metal layer includes a square ring metal patch and a square metal patch, and the square metal patch is located inside the ring of the square ring metal patch and does not contact the square ring metal patch.

[0035] As a preferred technical solution, the square ring metal patch and the square metal patch are etched on the dielectric layer symmetrically about the unit center.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1) The present invention provides a sensor signal processing method based on a metasurface and a complex-valued neural network. The metasurface is used to regulate the microwave signal passing through the sample, and then the regulated microwave signal is processed by a complex-valued neural network after the network structure is dynamically optimized. This method fits the nonlinear relationship between the dielectric constant and the scattering parameter, and expands the types of detectable substances. It can be applied to the dielectric constant or solution concentration detection of various substances including solids or liquids.

[0038] 2) The present invention provides a dynamic optimization training method for a complex-valued neural network, which adaptively adjusts the model structure according to the performance of the model on the test set, improves the robustness and generalization ability of the model, avoids overfitting, and enables the neural network to more accurately extract the amplitude and phase information in the data, and can more flexibly and efficiently process different types of data, so as to be suitable for a wider range of material detection.

[0039] 3) The present invention attaches the metasurface to the side of the sample to be detected facing the receiving antenna. The metal pattern of the metasurface unit is equivalent to a specific characteristic impedance boundary condition. These boundary conditions are related to the tangential components of the surface electric field and magnetic field on the contact surface between the sample and the metasurface, that is, the metasurface introduces a surface current between the contact surface with the sample, and finally realizes that the metasurface provides a unique field manipulation capability at microwave frequencies and affects the propagation and reflection characteristics of electromagnetic waves in space, that is, directly affects the reflection (S11) and transmission (S21) coefficients. These scattering parameter coefficients are very sensitive to the dielectric constant of the material, so material characterization can be performed. At the same time, the metasurface array ensures enhanced electromagnetic coupling through a periodic structure, further amplifying the sensitivity of the S11 coefficient and the S21 coefficient to changes in the dielectric constant, thereby achieving wide-range and high-precision material detection, that is, expanding the detection range, and is suitable for dielectric constant or solution concentration detection of various substances including solids or liquids. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the method of the present invention;

[0041] Figure 2 A flowchart of the complex-valued neural network training of the present invention;

[0042] Figure 3 A flow chart of dynamic optimization of the present invention;

[0043] Figure 4 A schematic diagram of a training label for a glucose solution material of the present invention;

[0044] Figure 5 is the prediction result of the glucose solution material test set of the present invention;

[0045] Figure 6A schematic diagram of training labels for the ethanol solution material of the present invention;

[0046] Figure 7 is the prediction result of the ethanol solution material test set of the present invention;

[0047] Figure 8 A schematic diagram of a training label for a glass material of the present invention;

[0048] Fig. 9 is the prediction result of the glass material test set of the present invention;

[0049] Fig.10 A schematic diagram of the arrangement of a periodic array of a metasurface of the present invention;

[0050] Fig.11 Schematic diagram of a unit of the super surface of the present invention

[0051] Fig.12 Schematic diagram of the test scenario of the present invention. DETAILED DESCRIPTION

[0052] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0053] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantitative limitation, and may represent the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0054] In order to solve the shortcomings of traditional material measurement methods in the prior art, the present application provides a sensor signal processing method and a corresponding system based on a metasurface and a complex-valued neural network, which uses a metasurface to introduce a surface current between the metasurface and the sample contact surface to achieve the regulation of electromagnetic waves, so as to improve the offset sensitivity of the resonant frequency to the change of the dielectric constant, expand the detection range, and facilitate the detection of the dielectric constant or solution concentration of various substances including solids or liquids. Due to the outstanding signal processing ability of complex-valued neural networks, the present application adopts complex-valued neural networks to extract the amplitude and phase information in microwave signals for more accurate processing, so as to achieve high-precision and multi-category detection. The above-mentioned method and system are described in detail through the following embodiments.

[0055] Example 1

[0056] This embodiment provides a sensor signal processing method based on a metasurface and a complex-valued neural network. The method processes the S11 coefficient and the S21 coefficient based on a trained and dynamically optimized complex-valued neural network, and finally outputs the dielectric constant of the sample or the solution concentration. The specific S11 coefficient and the S21 coefficient are obtained by the microwave signal passing through the sample and the metasurface in sequence. The complex-valued neural network includes an input layer, a preprocessing layer, multiple hidden layers, and an output layer. The process of complex-valued neural network signal processing is as follows: Figure 1 As shown, the detailed steps include:

[0057] Step 1: The input layer extracts multiple frequency point data from the S11 coefficient and the S21 coefficient within a preset frequency band, and the frequency point data includes the amplitude and phase of the frequency point.

[0058] Step 2: The preprocessing layer normalizes the amplitude and phase of each frequency point data, and then forms a complex-valued vector from the normalized amplitude and phase of each frequency point data. The specific normalization refers to linearly normalizing the amplitude to 0~1 and the phase to 0°~180°.

[0059] Step 3: Input the complex-valued vector into the hidden layer, which processes it and transmits the processing result to the next hidden layer.

[0060] Step 4: The output layer receives the processing result of the last hidden layer and outputs the dielectric constant of the sample or the complex mapping value of the solution concentration.

[0061] This embodiment also provides a method for training a complex-valued neural network and dynamically optimizing the network structure to achieve model adaptive adjustment, thereby improving the robustness and generalization ability of the model and avoiding overfitting, so that the neural network can process different types of data more flexibly and efficiently to meet different detection requirements. The specific training process is as follows: Figure 2 As shown, after each training cycle, the following is executed Figure 3The dynamic optimization shown performs adaptive structural adjustment of the complex-valued neural network, and selects glucose solution and ethanol solution with known solution concentration and glass with known dielectric constant for model training.

[0062] Scenario 1: Select a glucose solution with known solution concentration for model training.

[0063] A1. Obtain the S11 coefficient and S21 coefficient of multiple glucose solution samples with known solution concentrations, where the concentration range of the glucose solution samples is 0-300 mg / dL, with a gradient of 30 mg / dL, and there are 11 different concentrations in total. The relative dielectric constants of the above 11 different concentrations of glucose solutions range from 79 to 81, and the loss tangent is 0.4.

[0064] 20 S11 coefficients and 20 S21 coefficients are evenly extracted from 2 to 8 GHz as the data set and divided into a training set and a test set.

[0065] The known dielectric constant or solution concentration is used as the training label, such as Figure 4 As shown in the figure, the radius of the glucose solution class circle is defined as 1, and 11 glucose solutions with different concentrations ranging from 0 to 300 mg / dL are mapped to the circle with a radius of 1. Then the experimental data of glucose solution are divided into training set and test set. The experimental data of 300 mg / dL is used as the test set, and the rest is used as the training set. The ratio of training set to test set is 10:1.

[0066] A2. Use the complex Xavier initialization method to initialize the weight parameters of the complex-valued neural network. The expression is:

[0067] ,

[0068] in, represents the weight parameter, represents the number of neurons in the input layer, represents the number of output neurons, Represents a normal distribution.

[0069] A3. Repeat the following steps until the training goal is reached:

[0070] A31. After the data in the training set is processed by the preprocessing layer, it is input into the hidden layer for forward propagation, and the activation function during the forward propagation process is:

[0071] ,

[0072] in, represents the complex vector of the i-th neuron in the current layer of the complex-valued neural network, The magnitude of the complex vector representing the i-th neuron.

[0073] A32. Calculate the loss function and use the back propagation algorithm to calculate the gradient of each layer parameter to the loss function starting from the output layer, and update the weight parameters based on the gradient.

[0074] In detail, the calculation method of the loss function is:

[0075] ,

[0076] in, Represents the number of training set samples in the dataset, represents the current sample, Represents the complex-valued neural network prediction The predicted value of samples, Indicates The true value of the samples.

[0077] The method for updating the weight parameter gradient is:

[0078] ,

[0079] ,

[0080] Among them, K is the learning rate; Indicates that the l The i-th neuron in layer -1 is connected to the l The weight of the jth neuron in the layer; Weight the magnitude of Weight The phase of is the loss function value; r represents the number of iterations in the training process, and r+1 represents a new weight update based on the current step (r).

[0081] If the current training cycle has not ended, training will continue according to the unupdated structure; otherwise, dynamic optimization of the complex-valued neural network structure will be performed, including:

[0082] B1. Set the optimization goal to make the ratio of the number of correctly predicted samples to the total number of samples greater than or equal to 90%. The initial number of hidden layers is 2, and the maximum number of hidden layers allowed is 4. The preset maximum value of the initial weight node of each hidden layer is 20, and the range of the preset maximum value is .

[0083] B2. After each training cycle, obtain the ratio of the number of samples correctly predicted by the complex-valued neural network in the test set to the total number of samples. If the ratio is less than 90%, execute steps B3 to B4; if the ratio is greater than or equal to 90%, execute step B5.

[0084] B3. Add multiple (e.g., 2) new weight nodes to the hidden layer that is closest to the input layer and meets the total number of weight nodes condition, and continue the next round of training, where the total number of weight nodes condition means that the total number of weight nodes of the hidden layer after adding the new weight nodes is less than or equal to the preset maximum value corresponding to the layer. The preset maximum value of the hidden layer with the newly added new weight nodes is dynamically updated after each round of training, that is, each time a training cycle is executed, the preset maximum value increases by a preset number, and the preset number is less than the number of new weight nodes added.

[0085] B4. If the total number of weight nodes of all hidden layers of the complex-valued neural network does not meet the total number of weight nodes condition, a new hidden layer is added after the last hidden layer and the next round of training continues.

[0086] B5, then the dynamic optimization ends.

[0087] Furthermore, when executing steps B3 to B4, when the total number of added hidden layers and the initial hidden layers is greater than or equal to 3 layers, the preset maximum value range is expanded to When the number of dynamic optimization executions reaches 4,000 and the prediction accuracy does not meet the optimization goal, the dynamic optimization will be stopped and the current optimal model will be selected as the final complex-valued neural network.

[0088] The specific network parameters of the optimal model for predicting the concentration of glucose solution were obtained through training and dynamic optimization, as shown in Table 1.

[0089] Table 1 Network structure parameters for predicting glucose solution

[0090]

[0091] The above parameters are verified, and an accuracy index is defined for the prediction of 300 mg / dL glucose solution. When the error between the output value and the phase angle of the label is within ±7.5° and When the error is within ±0.04, the prediction is considered accurate. The verification results are as follows: Figure 5 As shown in the figure, it can be observed that the predicted values ​​of 300 mg / dL glucose solution are distributed near the true value, with an accuracy of 88% and a root mean square error (RMSE) of 0.039.

[0092] Furthermore, in order to verify the superiority of the model performance provided by Scenario 1, this embodiment also uses the K-fold cross-validation method to evaluate the performance of the trained model. For glucose solution, the experimental data of 0~300 mg / dL are divided into 11 groups according to the gradient of 30 mg / dL. The first fold uses the data of 30~300 mg / dL glucose solution as the training set, and the data of 0 mg / dL glucose solution is used as the test set alone. The second fold uses the data of 30 mg / dL glucose solution as the test set, and the remaining ten groups of sample data are used as training sets. According to this rule, 10 concentrations of experimental data are used for training each time, and the remaining concentration of experimental data is used for verification. This cycle is repeated 11 times to ensure that the model has good prediction performance. As shown in Table 2, the prediction results of the K-fold cross-validation method are shown in detail, that is, the average accuracy and RMSE of 11 glucose solutions with a concentration range of 0~300 mg / dL. It can be seen from the table that in the different concentrations of glucose solutions, the average accuracy and RMSE of the 11 glucose solutions with a concentration range of 0~300 mg / dL are good. In high-precision predictions, the accuracy rate can reach up to 96%.

[0093] Table 2 Scenario 1 model performance verification results

[0094]

[0095] Scenario 2: Select an ethanol solution with known concentration for model training.

[0096] The S11 coefficient and S21 coefficient of multiple ethanol solution samples with known solution concentrations are obtained, wherein the ethanol content of the ethanol solution samples is 97%~100%, with a gradient change of 1%, and the relative dielectric constant of the above four ethanol solutions of different concentrations ranges from 16 to 18, and the loss tangent is 0.12.

[0097] 20 S11 coefficients and 20 S21 coefficients are evenly extracted from 2 to 8 GHz as the data set and divided into a training set and a test set.

[0098] The known dielectric constant or solution concentration is used as the training label, such as Figure 6 As shown in the figure, the radius of the ethanol solution class circle is defined as 2, and four ethanol solutions with different concentrations and relative dielectric constants ranging from 16 to 18 are mapped to circles with a radius of 2 respectively. Then, the experimental data of ethanol solutions are divided into training sets and test sets. 100% of the experimental data is used as the test set, and the rest is used as the training set. The ratio of training set to test set is 3:1.

[0099] After training in steps A2 to A3 and dynamic optimization in steps B1 to B5, the specific network parameters of the optimal model for predicting the concentration of glucose solution are obtained, as shown in Table 3.

[0100] Table 3 Network structure parameters for predicting ethanol solution

[0101]

[0102] And verify the above parameters, such as Figure 7 As shown, it can be seen that the predicted value of 100% ethanol solution is distributed near the true value, with an accuracy of 92% and a root mean square error (RMSE) of 0.034.

[0103] Furthermore, in order to verify the superiority of the model performance provided by Scenario 2, this embodiment also uses the K-fold cross-validation method to evaluate the performance of the trained model. For ethanol solutions with a content of 97% to 100%, they are divided into 4 samples according to the content. Each time, one of the samples is used as a test set, and the remaining samples are used as training sets. For example, ethanol solutions with a content of 97% are used as a test set, and ethanol solutions with a content of 98% to 100% are used for training. The cycle is repeated four times to ensure that each sample is verified. The results are shown in Table 4, which shows in detail the prediction results of the K-fold cross-validation method, namely the average accuracy and RMSE of the four ethanol solutions with a content of 97% to 100%. It can be seen from the table that in the ethanol solutions with different contents, In high-precision predictions, the accuracy rate can reach up to 96%.

[0104] Table 4 Performance verification results of scenario 2 model

[0105]

[0106] Scenario 3: Select glass with known dielectric constant for model training.

[0107] The S11 coefficient and S21 coefficient of multiple glasses with known dielectric constants are obtained. The test samples are four different glasses, including ultra-white glass, ordinary white glass, soda-lime glass and potassium-lime glass. Their relative dielectric constants are They are 6.5, 7, 7.8 and 8.5 respectively, and the loss angle The tangent is 0.002.

[0108] 20 S11 coefficients and 20 S21 coefficients are uniformly extracted from 10 to 18 GHz as the data set and divided into a training set and a test set.

[0109] The known dielectric constant or solution concentration is used as the training label, such as Figure 8 As shown, the radius of the glass class circle is defined as 3, and the experimental data of glass is divided into a training set and a test set. The experimental data of tempered glass is used as the test set, and the rest is used as the training set. The ratio of the training set to the test set is 3:1.

[0110] After training in steps A2 to A3 and dynamic optimization in steps B1 to B5, the specific network parameters of the optimal model for predicting the concentration of glucose solution are obtained, as shown in Table 5.

[0111] Table 5 Prediction of glass network structure parameters

[0112]

[0113] And verify the above parameters, such as Fig. 9 As shown in the figure, it can be observed that the predicted values ​​of tempered glass are distributed near the true values, with an accuracy of 90% and a root mean square error (RMSE) of 0.038.

[0114] Furthermore, in order to verify the superiority of the model performance provided by scenario three, this embodiment also uses the K-fold cross-validation method to evaluate the performance of the trained model. For four different types of glass, one glass sample is used as the test set each time, and the remaining glass samples are used as the training set. As shown in Table 6, the prediction results of the K-fold cross-validation method are shown in detail, that is, the average accuracy and RMSE of the four types of glass with relative dielectric constants ranging from 6.5 to 8.5. It can be seen from the table that In high-precision predictions, the accuracy rate can reach up to 96%.

[0115] Table 6 Performance verification results of scenario 3 model

[0116]

[0117] Based on the experimental results of scenarios 1 to 3, it can be seen that the training method and dynamic optimization method provided in this embodiment are feasible and can ensure the accuracy of the complex-valued neural network after training and dynamic optimization.

[0118] Example 2

[0119] This embodiment provides a sensor signal processing system based on a metasurface and a complex-valued neural network. The system includes a transmitting antenna, a receiving antenna, a vector network analyzer (VNA), a metasurface, and a sensor signal processing module.

[0120] Among them, the transmitting antenna is used to transmit microwave signals to the sample, and the microwave signals are received by the receiving antenna after passing through the sample and the metasurface in turn.

[0121] The receiving antenna is used to receive the microwave signal and transmit it to the vector network analyzer.

[0122] The vector network analyzer processes the microwave signal sent by the receiving antenna to obtain the S11 coefficient and the S21 coefficient. The S11 coefficient can reflect the ratio of the microwave signal power reflected by the system to the microwave signal power input to the system; the S21 coefficient can reflect the ratio of the microwave signal power input to the system to the microwave signal power obtained by the receiving antenna and transmitted through the transmitting antenna and through the system. Both include the amplitude and phase of multiple frequency points.

[0123] The metasurface introduces a surface current at the contact surface with the sample to control the microwave signal, so as to enhance the sensitivity of the resonance frequency of the scattering parameter to the deviation of the measured parameter. In detail, the metasurface includes an array of N×N units arranged periodically in both the horizontal and vertical directions, where N represents the number of units in each row or column of the array, such as Fig.10 As shown, each adjacent unit is not in direct contact, and each unit includes a metal layer and a dielectric layer. The metal layer includes a square ring metal patch and a square metal patch, the square metal patch is located in the ring of the square ring metal patch and does not contact the square ring metal patch, and the square ring metal patch and the square metal patch are etched on the dielectric layer symmetrically about the unit center, as shown in FIG. Fig.11 As shown, the thickness of the metal layer is 30 microns, and the dielectric layer is a 0.762 mm thick Rogers FR4 board.

[0124] The sensor signal processing module executes the sensor signal processing method based on the metasurface and complex-valued neural network provided in Example 1, processes the S11 coefficient and the S21 coefficient using the complex-valued neural network, and outputs the dielectric constant or solution concentration of the sample.

[0125] The transmitting antenna, receiving antenna, vector network analyzer (VNA) and metasurface are connected as follows: Fig.12 arrangement for sample detection. In addition, in order to fix the sample, the present application fixes the sample on a metal plate, and sets the distance between the receiving antenna and the metasurface and the distance between the transmitting antenna and the metal plate to 20 cm. When performing detection, the transmitting antenna transmits a microwave signal to the sample, and the microwave signal passes through the metal plate, the sample and the metasurface in turn, is received by the receiving antenna, and is input into the sensor signal processing module for comprehensive signal analysis after being processed by the VNA.

[0126] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A sensor signal processing method based on a metasurface and a complex-valued neural network, characterized in that: The method is based on a complex-valued neural network after training and dynamic optimization, processes the S11 coefficient and the S21 coefficient, and outputs the dielectric constant of the sample or the solution concentration; the complex-valued neural network includes an input layer, a preprocessing layer, a plurality of hidden layers and an output layer; the S11 coefficient and the S21 coefficient are obtained by microwave signals passing through the sample and the metasurface in sequence; wherein the dynamic optimization performs the following steps until the optimization target is reached: After each training cycle, the ratio of the number of samples correctly predicted by the complex-valued neural network in the test set to the total number of samples is obtained. If the ratio is less than a preset value, multiple new weight nodes are added to the hidden layer that is closest to the input layer and meets the total number of weight nodes condition, and the next round of training is continued; the total number of weight nodes condition is that the total number of weight nodes of the hidden layer after the new weight nodes are added is less than or equal to the preset maximum value corresponding to the layer; the initial preset maximum value of each hidden layer is the same, and the preset maximum value of the hidden layer with the new weight node is dynamically updated after each round of training; If the total number of weight nodes of all hidden layers of the complex-valued neural network does not meet the total number of weight nodes condition, a new hidden layer is added after the last hidden layer to continue the next round of training; If the ratio is greater than or equal to the preset value, the dynamic optimization is terminated.

2. The sensor signal processing method based on a metasurface and a complex-valued neural network according to claim 1, characterized in that: The dynamic update is that each time a training cycle is executed, the preset maximum value is increased by a preset amount, and the preset amount is less than the number of new weight nodes added.

3. The sensor signal processing method based on metasurface and complex-valued neural network according to claim 1 is characterized in that: The dynamic optimization also includes, when the total number of added hidden layers and initial hidden layers is greater than or equal to a layer number threshold, expanding a preset maximum value range.

4. The sensor signal processing method based on metasurface and complex-valued neural network according to claim 1 is characterized in that: The training includes: Obtain the S11 coefficient and S21 coefficient of the sample with known dielectric constant or solution concentration, use the S11 coefficient and S21 coefficient in the preset frequency band as a data set, and divide it into a training set and a test set; use the known dielectric constant or solution concentration as a training label; Initialize the weight parameters of the complex-valued neural network, the expression is: ， in, represents the weight parameter, represents the number of neurons in the input layer, represents the number of output neurons, represents normal distribution; Repeat the following steps until you reach your training goal: The data in the training set is processed by the preprocessing layer and then input into the hidden layer for forward propagation. The activation function during the forward propagation is: ， in, represents the complex vector of the i-th neuron in the current layer of the complex-valued neural network, The modulus of the complex vector representing the i-th neuron; The loss function is calculated, and the gradient of the loss function of each layer parameter is calculated in turn starting from the output layer using the back propagation algorithm, and the weight parameters are updated based on the gradient.

5. The sensor signal processing method based on metasurface and complex-valued neural network according to claim 1 is characterized in that: The method for outputting the dielectric constant of the sample or the solution concentration comprises: The input layer extracts a plurality of frequency point data from the S11 coefficient and the S21 coefficient within a preset frequency band, wherein the frequency point data includes the amplitude and phase of the frequency point; The preprocessing layer normalizes the amplitude and phase of each frequency point data, and then forms a complex value vector with the amplitude and phase of each frequency point data after normalization. The complex-valued vector is input into a hidden layer, processed by the hidden layer and the processing result is transmitted to the next hidden layer; The output layer receives the processing results of the last hidden layer and outputs the dielectric constant of the sample or the complex mapping value of the solution concentration.

6. The sensor signal processing method based on metasurface and complex-valued neural network according to claim 5 is characterized in that: The normalization process includes: linearly normalizing the amplitude to 0-1, and linearly normalizing the phase to 0°-180°.

7. A sensor signal processing system based on a metasurface and a complex-valued neural network, characterized in that: The system includes a transmitting antenna, a receiving antenna, a vector network analyzer, a metasurface and a sensor signal processing module, wherein the transmitting antenna is used to transmit a microwave signal to the sample, and the microwave signal is received by the receiving antenna after passing through the sample and the metasurface in sequence; The receiving antenna is used to receive microwave signals and transmit them to the vector network analyzer; The vector network analyzer processes the microwave signal sent by the receiving antenna to obtain the S11 coefficient and the S21 coefficient; The metasurface introduces a surface current at the contact surface with the sample to control the microwave signal, so as to enhance the offset sensitivity of the resonance frequency of the scattering parameter to the change of the parameter to be measured; The sensor signal processing module executes the method described in any one of claims 1 to 6, processes the S11 coefficient and the S21 coefficient using a complex-valued neural network, and outputs the dielectric constant of the sample or the solution concentration.

8. The sensor signal processing system based on metasurface and complex-valued neural network according to claim 7, characterized in that: The metasurface comprises an array of N×N units arranged periodically in both horizontal and vertical directions, wherein N represents the number of units in each row or column of the array.

9. The sensor signal processing system based on metasurface and complex-valued neural network according to claim 8, characterized in that: Adjacent units are not in direct contact, and each of the units includes a metal layer and a dielectric layer; the metal layer includes a square ring metal patch and a square metal patch, and the square metal patch is located inside the ring of the square ring metal patch and does not contact the square ring metal patch.

10. The sensor signal processing system based on metasurface and complex-valued neural network according to claim 9, characterized in that: The square ring metal patch and the square metal patch are etched on the dielectric layer symmetrically about the unit center.

Images