Multi-parameter detection method for PEC signal grounding grid based on improved multi-polarization feature extraction aggregation network

Abstract

The application relates to a PEC signal grounding net multi-parameter detection method based on an improved multi-polarization feature extraction aggregation network, and belongs to the field of grounding net nondestructive detection. The method comprises the following steps: acquiring a time domain signal of an eddy current effect; establishing an improved multi-polarization feature extraction aggregation network for grounding net parameter estimation, wherein a multi-layer convolutional neural network is used to extract local features of an input signal waveform diagram, and a nonlinear function is introduced through an activation function; local features are fused through a feature aggregation module; all features extracted for multiple times are spliced; a plurality of parallel branches are designed for a plurality of parameters of the grounding net; each branch is fused through a mixed attention mechanism module; finally, a graph neural network is introduced to integrate the correlation information between target parameters, and regression calculation is performed through a full connection layer to output parameter estimation values. The application can simultaneously calculate three grounding net steel bar parameters while reducing the adverse effects of parameter correlation.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing of grounding grids, and relates to a multi-parameter detection method for PEC signal grounding grids based on an improved multi-polarization feature extraction and aggregation network. Background Technology

[0002] Pulsed eddy current (PEC) technology has been applied to some extent in the detection and evaluation of grounding grid structures. Determining the depth, radius, and orientation angle of the reinforcing bars in the grounding grid is one of the key factors in assessing construction quality and structural health. However, there is a strong correlation between the PEC signal detected by the pulsed eddy current detection coil and the parameters of the grounding grid reinforcing bars, making the simultaneous detection of multiple grounding grid parameters a significant challenge.

[0003] Since multiple parameters related to the grounding grid reinforcement collectively influence the waveform characteristics of the PEC signal, establishing a nonlinear relationship between the PEC signal waveform characteristics and these parameters is a possible solution. Deep learning algorithms, with their powerful feature extraction and nonlinear learning capabilities, are well-suited for this challenging task.

[0004] Current grounding grid testing methods also have the following shortcomings:

[0005] 1. Currently, the parameters of the grounding grid are often measured by direct excavation. This method is costly and time-consuming, and is not conducive to the rapid detection of the corrosion depth of the grounding grid.

[0006] 2. The main method for detecting grounding grid parameters is ground-penetrating radar (GPR), which transmits high-frequency electromagnetic waves underground and relies on the reflection of these waves between different materials to construct an image of underground objects. However, GPR has limited spatial resolution, and it may be difficult to accurately distinguish the specific location and radius of the reinforcing bars, especially in complex structures or environments with dense reinforcement.

[0007] 3. Some scholars have suggested using electromagnetic detection methods, particularly transient electromagnetic methods, to quantify the parameters of grounding grids. This method involves applying a pulsed magnetic field to generate and measure changes in the secondary magnetic field. Parameters are set using a forward model, and the signal is simulated to extract characteristic parameters, thereby estimating the position, depth, radius, and azimuth deviation of the grounding grid reinforcement. However, transient electromagnetic technology is affected by soil conductivity and environmental interference, which may reduce its accuracy and reliability.

[0008] 4. Other detection technologies, such as infrared thermal imaging and ultrasonic testing, can detect grounding grid parameters to some extent, but each has its own limitations: infrared thermal imaging is mainly used for imaging grounding grid corrosion, and its detection accuracy is low; ultrasonic testing has limited penetration performance in metal conductors, and it is difficult to achieve high-precision detection for grounding grids with complex structures. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide a multi-parameter detection method for PEC signal grounding grids based on an improved multi-polarimetric feature extraction aggregation network. The improved multi-polarimetric aggregation network, called IMPA-Net (Improved Multi-Polarimetric Aggregation Network), can simultaneously estimate multiple grounding grid rebar-related parameters from the PEC detection signal waveform, including rebar burial depth (d), radius (r), and orientation angle (φ). IMPA-Net takes the PEC received signal waveform as input, extracts the information features of the grounding grid rebar reflected signal, mitigates the adverse effects of parameter correlation, and simultaneously calculates three grounding grid rebar parameters.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A multi-parameter detection method for PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network, the method comprising:

[0012] S1. Obtain the pulsed eddy current effect time-domain signal, wherein the characteristic quantities in the pulsed eddy current effect time-domain signal include pulse width, pulse peak value, pulse valley value, peak time and second zero crossing time;

[0013] S2. Establish an improved multi-polarity feature extraction and aggregation network for grounding grid parameter estimation. This improved network includes a multi-polarity feature extraction and aggregation module and a parameter estimation module, specifically comprising the following steps:

[0014] S21. First, in the multipolar feature extraction and aggregation module, a multi-layer convolutional neural network is used to extract local features from the input signal waveform, and nonlinearity is introduced through an activation function.

[0015] S22. Fuse local features using the feature aggregation module FA;

[0016] S23. After several rounds of processing by activation functions and feature aggregation modules with different features, all extracted features are concatenated.

[0017] S24. In the parameter estimation module, several parallel branches are designed for several parameters of the grounding grid, and each branch is fused by introducing the hybrid attention mechanism module HA.

[0018] S25. Finally, a graph neural network is introduced to integrate the correlation information between the target parameters, and regression calculation is performed through a fully connected layer to output the parameter estimates.

[0019] Furthermore, in step S1, when acquiring the pulsed eddy current effect time-domain signal, a pulse generator, an electromagnetic probe, and a data acquisition system are set up sequentially; wherein, setting up the pulse generator includes setting the pulse width and frequency; setting up the electromagnetic probe includes installing the excitation coil and the detection coil, and adjusting the lift-off distance; setting up the data acquisition system includes setting its sampling rate;

[0020] After generating a transient electromagnetic field by trigger pulse excitation, the probe detects the eddy current attenuation signal in the grounding grid conductor. After pre-amplification and bandpass filtering, the time-domain waveform is recorded by a high-speed acquisition card, and finally the feature signal is extracted by digital filtering.

[0021] Furthermore, let the distance between the conductor being measured and the excitation coil be d, and the radius of the excitation coil be r. b If a low-frequency pulsed alternating current I1 is applied, a decaying magnetic field B1 is obtained that decays rapidly along the transition edge of the square wave. When the decaying magnetic field B1 comes into contact with the conductor under test, a transient eddy current I2 is generated in the conductor under test, as well as an eddy current magnetic field B2 generated by the transient eddy current I2 that is opposite in direction to the initial magnetic field. Therefore, the voltage value at the detection point can be obtained according to Faraday's law of electromagnetic induction.

[0022]

[0023] In the formula, V p Let be the coil voltage at the detection point, B be the induced magnetic field strength, ||A|| be the vector norm of the induced magnetic field, S be the cross-sectional area of ​​the detection coil, and c be the integration path. It represents the partial derivative with respect to time t, which is used to describe the instantaneous rate of change of a magnetic field or magnetic flux; This represents the unit normal vector, i.e., the normalized normal vector, whose direction is perpendicular to the surface S.

[0024] Furthermore, in step S21, the convolutional layer uses a weight-sharing mechanism to extract local features of the temporal signal;

[0025]

[0026] The activation function used is the Leaky ReLU activation function, which is expressed as:

[0027] Leaky ReLU(x) = max(αx,x)

[0028] Where x i,j To input the pixel values ​​of the PEC signal waveform, Let b be the weight value of the k-th convolutional kernel. k α is the bias value of the k-th convolutional kernel, f is the activation function Leaky ReLU, x represents the input value of the neuron in this layer, and α is the negative slope coefficient.

[0029] Furthermore, in step S22, the feature aggregation module FA combines the complementary information of the multipolar signals and uses the weighted sum operation of the convolution kernel to generate a global feature representation:

[0030]

[0031] Among them, F k It is the k-th PEC signal input feature, ω k F represents the weight coefficient of the k-th input feature. agg It is the global feature after a certain branch is aggregated.

[0032] Furthermore, in step S23, the features obtained by concatenating the corresponding features based on the pulse width, pulse peak value, pulse valley value, peak time, and second zero-crossing time are as follows:

[0033]

[0034] Among them, F global This represents the feature vector obtained by concatenating the corresponding features. Concatenate(·) is the concatenation function. This represents the global feature after the aggregation of the nth branch.

[0035] Furthermore, in step S24, the parameter estimation module combines the channel attention mechanism FA and the spatial attention mechanism SA, which analyze the importance of the feature map from the channel and spatial dimensions, respectively; the fused features are obtained through the hybrid attention mechanism HA.

[0036] f c =σ(W2·ReLU(W1·F) global ))

[0037] f s =σ(Conv 1×1 (F global ))

[0038] F HA =f c ·F global +f s ·F global

[0039] Where W1W2 is the weight matrix, σ represents the Leaky ReLU activation function, and f c Represents channel weights, Conv 1×1 f represents a 1×1 convolution operation. s F represents the spatial weight. HA Fusion characteristics of hybrid attention mechanism under dual-weighted fusion.

[0040] Furthermore, the grounding grid has multiple parameters, including the depth of the reinforcing bar burial, the radius of the reinforcing bar, and the direction angle of the reinforcing bar, and three corresponding integrated branches are designed accordingly.

[0041] Furthermore, the processing procedures of Graph Neural Network (GNN) and Fully Connected Layer (FC) are respectively represented as follows:

[0042]

[0043] y = W FC F GNN +b FC

[0044] Among them, H (l) This represents the feature matrix of the nodes in the l-th layer. The adjacency matrix A of the graph plus self-loops, i.e. Where I is the identity matrix. express The degree matrix, W (l) H represents the learnable weight matrix of the l-th layer, σ represents the activation function Leaky ReLU, and H... (l+1) This represents the feature matrix of the (l+1)th layer nodes. The feature vector of the i-th node in the L-th layer, F GNN This represents the final output of the graph neural network, where N represents the total number of nodes in the graph, and W represents the final output of the graph neural network. FC For the corresponding weight matrix, b FC For the corresponding bias vector, y is the estimated value of the burial depth, radius and direction angle of the grounding grid reinforcement, respectively, in the final output vector.

[0045] The beneficial effects of this invention are as follows:

[0046] This invention extracts the waveform characteristics of the PEC signal under different parameters of the grounding grid by pulsed eddy current detection, and proposes an improved multi-polarization aggregation network for detecting the parameters of the grounding grid reinforcement. Results show that the mean absolute errors (MAEs) for reinforcement embedment depth, radius, and orientation are 0.20 cm, 0.13 cm, and 2.54°, respectively, proving that the network can simultaneously and accurately estimate multiple grounding grid reinforcement-related parameters.

[0047] This invention improves the accuracy of multi-parameter detection for grounding grid reinforcement. It modifies the MPA-Net model by adding a spatial attention mechanism and combining it with a channel attention mechanism to form a hybrid attention mechanism. This dual-weighted strategy can more comprehensively capture the relevant features of the target parameters, improve the network's ability to analyze multi-dimensional features, and ultimately enhance the accuracy and stability of parameter estimation. Simultaneously, a graph neural network model is introduced based on the correlation between target parameters. Results show that this model improves the analytical ability of complex signal features, exhibiting stronger robustness and adaptability in highly coupled signal scenarios.

[0048] This invention can be extended to testing applications in other systems. By simply changing the specimen to be tested, other parameters of the specimen can be accurately tested in a non-excavation, non-destructive manner.

[0049] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0051] Figure 1 This is an overall flowchart of the multi-parameter detection method for PEC signal grounding grid based on an improved multi-polarization feature extraction aggregation network of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the principle of obtaining the time-domain signal of the pulsed eddy current effect according to the present invention.

[0053] Figure 3 The waveform of the pulsed eddy current effect time-domain signal obtained by this invention is shown.

[0054] Figure 4 This is a schematic diagram of the structure of the improved multipolar feature extraction aggregation network of the present invention;

[0055] Figure 5 This is a histogram showing the estimation error of relevant parameters of the grounding grid reinforcement in this invention. Detailed Implementation

[0056] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0057] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0058] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0059] Please see Figures 1-5 This is a method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction aggregation network.

[0060] This invention proposes a method for analyzing and estimating the time-domain signal waveform changes caused by the pulsed eddy current (PEC) effect in grounding grids with different parameters. Deep learning technology is used to analyze these changes and estimate the parameters. Research shows that the changes in the PEC time-domain signal waveform are closely related to its pulse width, peak value, valley value, peak time, and second zero-crossing time. Therefore, this method uses the PEC time-domain signal waveform as input to a deep learning model, and uses image recognition technology to identify features such as pulse width, peak value, valley value, peak time, and second zero-crossing time in the image to achieve accurate estimation of the rebar parameters. To this end, this invention introduces a novel deep learning model—the Improved Multipolar Feature Extraction and Aggregation Network (IMPA-Net). The IMPA-Net network architecture consists of two main modules: the Multipolar Feature Extraction-Aggregation Module (MFEM) and the Parameter Estimation Module (PEM). The MFEM module extracts features from each PEC detection signal waveform, fully utilizing the complementary information of different rebar parameters carried by different waveforms. These polarization-dependent features are represented more comprehensively and with greater information content through the multipolar aggregation method, thereby improving the robustness and accuracy of feature extraction. The PEM module designs a three-branch model for three relevant parameters of the grounding grid reinforcement (burial depth, radius, and orientation angle). Each branch enhances the features related to specific parameters by introducing a hybrid attention mechanism, thereby improving the estimation ability of these parameters. Furthermore, to further improve the estimation accuracy of each parameter, this method also introduces a graph neural network (GNN) mechanism, which reduces the correlation between parameters, enabling the model to estimate each parameter more accurately. Figure 1 As shown, the specific process is as follows:

[0061] S1. Obtain the time-domain signal of the pulsed eddy current effect;

[0062] When acquiring the time-domain signal of the pulsed eddy current effect, it is necessary to set up a pulse generator (with a pulse width of 0.1-10μs and a frequency of 25Hz-1kHz), an electromagnetic probe (with an excitation / detection coil installed and a lift-off distance of 0.1-5mm), and a data acquisition system (with a sampling rate of 10-100MS / s) in sequence. After generating a transient electromagnetic field by triggering pulse excitation, the probe detects the eddy current attenuation signal in the grounding grid conductor. After pre-amplification and bandpass filtering, the time-domain waveform is recorded by a high-speed acquisition card, and finally the feature signal is extracted by digital filtering.

[0063] Assume the distance between the conductor being measured and the excitation coil is d, and the radius is r. b When the coil is excited by a low-frequency pulsed alternating current I1, it generates a rapidly decaying magnetic field B1 along the transition edge of the square wave. When the decaying magnetic field comes into contact with the conductor, the transient eddy currents I2 generated in the conductor produce an eddy current magnetic field B with the opposite direction to the initial magnetic field. 2，However, due to the rapid weakening of the eddy current field and its interaction with the excitation magnetic field, the magnetic flux inside the coil will change, thereby altering the initial magnetic field. As the secondary magnetic field decays, a time-varying instantaneous voltage signal can be induced in the detection coil. The schematic diagram of pulsed eddy current detection is shown below. Figure 2 As shown.

[0064] The voltage value at the detection point is obtained according to Faraday's law of electromagnetic induction:

[0065]

[0066] In the formula, V p Let B be the coil voltage at the detection point, B be the induced magnetic field strength, ||A|| be the vector norm of the induced magnetic field, and S be the cross-sectional area of ​​the detection coil. The partial derivative with respect to time t is used to describe the instantaneous rate of change of a magnetic field or magnetic flux. t Indicates time, Let represent the unit normal vector (normalized normal vector), with its direction perpendicular to surface S, and c represent the integration path. Changes in the grounding grid parameters will cause a change in the induced magnetic field strength B, ultimately resulting in a change in the transient induced voltage on the detection coil. Because pulses contain a wide spectrum, and pulsed eddy currents decay more slowly than single-frequency sinusoidal eddy currents, the transient induced voltage signal contains important information about the grounding grid parameters.

[0067] The PEC signal waveform is obtained by acquiring the voltage signal from the detection coil and performing noise reduction processing, as shown below. Figure 3 As shown, the PEC signal waveform diagram can reflect the signal's pulse width, pulse peak value, pulse valley value, peak time, and second zero-crossing time, among other characteristic quantities.

[0068] S2. An improved multipolar feature extraction and aggregation network is established for grounding grid parameter estimation. IMPA-Net includes a multipolar feature extraction and aggregation module (MFEM) and a parameter estimation module (PEM). Specifically, as follows... Figure 4 As shown, it includes the following specific steps:

[0069] S21. First, in the multi-polarization feature extraction and aggregation module, a multi-layer convolutional neural network (CNN) is used to extract local features from the input PEC signal waveform. The extracted features are nonlinearly introduced through an activation function. Specifically, the convolutional layer uses a weight-sharing mechanism to extract local features of the time-domain signal.

[0070]

[0071] The activation function used is the Leaky ReLU activation function, which is expressed as:

[0072] Leaky ReLU(x) = max(αx,x)

[0073] Where x i,j To input the pixel values ​​of the PEC signal waveform, Let b be the weight value of the k-th convolutional kernel. k α is the bias value of the k-th convolutional kernel, f is the activation function Leaky ReLU, x represents the input value of the neuron in this layer, and α is the negative slope coefficient, which is usually set to 0.01 to prevent negative features from being completely ignored during propagation.

[0074] S22. Then, the local features are fused using the feature aggregation module FA. This module combines the complementary information of the multipolar signals and uses the weighted sum operation of the convolution kernel to generate a global feature representation:

[0075]

[0076] Among them, F k It is the k-th PEC signal input feature, ω k F represents the weight coefficient of the k-th input feature. agg It is the global feature after a certain branch is aggregated.

[0077] S23. After several rounds of processing using activation functions and feature aggregation modules with different features, all extracted features are concatenated. The extracted features include corresponding features obtained based on pulse width, pulse peak value, pulse valley value, peak time, and second zero-crossing time. The final concatenated features are as follows:

[0078]

[0079] Among them, F global This represents the feature vector obtained by concatenating the corresponding features.

[0080] S24. In the parameter estimation module, several parallel branches are designed for several parameters of the grounding grid. Each branch is fused by introducing a hybrid attention mechanism (HA) module. This module combines channel attention mechanism (FA) and spatial attention mechanism (SA), which analyze the importance of feature maps from the channel and spatial dimensions, respectively, to improve the resolution quality of the target parameters. The fused features are obtained through the hybrid attention mechanism HA.

[0081] f c =σ(W2·ReLU(W1·F) global ))

[0082] f s =σ(Conv 1×1 (F global ))

[0083] F HA =f c ·F global +f s ·F global

[0084] Where W1W2 is the weight matrix, σ represents the Leaky ReLU activation function, and f c Represents channel weights, Conv 1×1 f represents a 1×1 convolution operation. s F represents the spatial weight. HA Fusion characteristics of hybrid attention mechanism under dual-weighted fusion.

[0085] The channel attention mechanism extracts the overall characteristics of each channel by compressing the spatial information of the input feature map, and then performs nonlinear transformation and weighting to generate a channel importance score, thereby amplifying relevant channel features and suppressing irrelevant channel features. The spatial attention mechanism aggregates the feature map along the channel dimension, extracts global contextual information for each spatial location, generates a spatial weight map, enhances key region features, and weakens background or irrelevant region features. This dual-weighting strategy can more comprehensively capture the highly relevant features of the target parameters, improve the network's multi-dimensional feature parsing ability, and ultimately enhance the accuracy and stability of parameter estimation. Since there are usually certain correlations between the target parameters, capturing these correlations during parameter estimation helps the network to perform more reasonable feature allocation and information sharing.

[0086] In this embodiment, the grounding grid has multiple parameters including the depth of the reinforcing bar burial, the radius of the reinforcing bar, and the direction angle of the reinforcing bar, thereby designing three corresponding fusion branches.

[0087] S25. Finally, a graph neural network (GNN) is introduced to effectively integrate the correlation information between target parameters, and regression calculation is performed through a fully connected layer (FC) to output parameter estimates, thereby achieving collaborative estimation of each parameter.

[0088]

[0089] y = W FC F GNN +b FC ,

[0090] Among them, H (l) This represents the feature matrix of the nodes in the l-th layer. The adjacency matrix A of the graph plus self-loops (i.e., (where I is the identity matrix) express The degree matrix, W (l)H represents the learnable weight matrix of the l-th layer, σ represents the activation function Leaky ReLU, and H... (l+1) This represents the feature matrix of the (l+1)th layer nodes. The feature vector of the i-th node in the L-th (last) layer, F GNN This represents the final output of the graph neural network, where N represents the total number of nodes in the graph, and W represents the final output of the graph neural network. FC For the corresponding weight matrix, b FC For the corresponding bias vector, y is the estimated value of the burial depth, radius and direction angle of the grounding grid reinforcement, respectively, in the final output vector.

[0091] In this embodiment, the trained IMPA-Net model was used to estimate the parameters of the grounding grid reinforcement in 500 experimental datasets. The estimation error histograms for the three relevant parameters of the grounding grid reinforcement are shown below. Figure 5 As shown, the mean absolute errors (MAEs) for rebar embedment depth, radius, and orientation angle in 500 test data sets were 0.20 cm, 0.13 cm, and 2.54°, respectively. The low MAEs in the test data demonstrate that the network can accurately estimate multiple grounding grid rebar-related parameters simultaneously.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network, characterized in that: The method includes: S1. Obtain the pulsed eddy current effect time-domain signal, wherein the characteristic quantities in the pulsed eddy current effect time-domain signal PEC include pulse width, pulse peak value, pulse valley value, peak time and second zero crossing time. In step S1, when acquiring the pulsed eddy current effect time-domain signal, a pulse generator, an electromagnetic probe, and a data acquisition system are set up sequentially. Setting up the pulse generator includes setting the pulse width and frequency; setting up the electromagnetic probe includes installing the excitation coil and detection coil, and adjusting the lift-off distance; setting up the data acquisition system includes setting its sampling rate. After generating a transient electromagnetic field by trigger pulse excitation, the probe detects the eddy current attenuation signal in the grounding grid conductor. After pre-amplification and bandpass filtering, the time-domain waveform is recorded by a high-speed acquisition card, and finally the feature signal is extracted by digital filtering. S2. Establish an improved multi-polarity feature extraction and aggregation network for grounding grid parameter estimation. This improved network includes a multi-polarity feature extraction and aggregation module and a parameter estimation module, specifically comprising the following steps: S21. First, in the multi-polarization feature extraction and aggregation module, a multi-layer convolutional neural network is used to extract local features from the input pulse eddy current effect time-domain signal waveform, and nonlinearity is introduced through the activation function. S22. Fuse local features using the feature aggregation module; S23, pulse width, pulse peak value, pulse valley value, peak time and second zero crossing time are processed by activation function and feature aggregation module several times, and then the corresponding features obtained from pulse width, pulse peak value, pulse valley value, peak time and second zero crossing time are spliced ​​together. S24. In the parameter estimation module, three parallel branches are designed for the multiple parameters of the grounding grid, and a hybrid attention mechanism module is introduced into each branch; the multiple parameters of the grounding grid include the depth of the reinforcing bar burial, the radius of the reinforcing bar, and the direction angle of the reinforcing bar. S25. Finally, a graph neural network is introduced to integrate the correlation information between the target parameters, and regression calculation is performed through a fully connected layer to output the parameter estimates.

2. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 1, characterized in that: Let the distance between the conductor being measured and the excitation coil be... d The radius of the excitation coil is r b If a low-frequency pulsed alternating current I1 is applied, a decaying magnetic field B1 is obtained that decays rapidly along the transition edge of the square wave. When the decaying magnetic field B1 comes into contact with the conductor under test, a transient eddy current I2 is generated in the conductor under test, as well as an eddy current magnetic field B2 generated by the transient eddy current I2 that is opposite in direction to the initial magnetic field. Therefore, the voltage value at the detection point can be obtained according to Faraday's law of electromagnetic induction. In the formula, V p Let B be the coil voltage at the detection point, and B be the induced magnetic field strength. A || represents the vector norm of the induced magnetic field. S To detect the cross-sectional area of ​​the coil, Indicates the integration path; Indicates time t The partial derivative of is used to describe the instantaneous rate of change of a magnetic field or magnetic flux; This represents the unit normal vector, i.e., the normalized normal vector, whose direction is perpendicular to the surface. S .

3. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 1, characterized in that: In step S21, the convolutional layer uses a weight-sharing mechanism to extract local features of the temporal signal; The activation function used is the Leaky ReLU activation function, which is expressed as: Among them, x i,j To input the pixel values ​​of the PEC signal waveform, For the first k The weight values ​​of each convolutional kernel. b k For the first k The bias value of each convolution kernel. f The activation function is Leaky ReLU. This represents the input value of the neurons in this layer. α It is a negative slope coefficient.

4. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 3, characterized in that: In step S22, the feature aggregation module combines the complementary information of the multipolar signals and generates a global feature representation using the weighted sum operation of the convolution kernel: in, F k It is the first k PEC signal input characteristics, ω k Indicates the first k The weight coefficients of each input feature. F agg It is the global feature after a certain branch is aggregated.

5. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 4, characterized in that: In step S23, the features obtained by splicing the corresponding features based on the pulse width, pulse peak value, pulse valley value, peak time, and second zero-crossing time are as follows: in, F global This represents the feature vector obtained by concatenating the corresponding features. For concatenation functions, This represents the global feature after the aggregation of the nth branch.

6. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 5, characterized in that: In step S24, the parameter estimation module combines channel attention and spatial attention (SA) mechanisms, which analyze the importance of feature maps from the channel and spatial dimensions, respectively; fused features are obtained through the hybrid attention mechanism. in, W 1 W 2 represents the weight matrix, and σ represents the Leaky ReLU activation function. f c Indicates channel weight, This represents a 1×1 convolution operation. f s Indicates spatial weights, F HA Fusion characteristics of hybrid attention mechanism under dual-weighted fusion.

7. The method for multi-parameter detection of PEC signal grounding grid based on an improved multi-polarization feature extraction and aggregation network according to claim 6, characterized in that: The processing steps of Graph Neural Network (GNN) and Fully Connected Layer (FC) are represented as follows: in, Indicates the first l The node feature matrix of the layer, Adjacency matrix of a graph Adding a self-loop, i.e. ,in It is the identity matrix. express The degree matrix, Indicates the first l Layer learnable weight matrix, This represents the activation function LeakyReLU. Indicates the first l +1 layer node feature matrix No. L The first layer i The feature vector of each node This represents the final output of the graph neural network. N This represents the total number of nodes in the graph. For the corresponding weight matrix, For the corresponding bias vector The final output vectors correspond to the estimated values ​​of the burial depth, radius, and orientation angle of the grounding grid reinforcement, respectively.

Images