A ground penetrating radar profile reconstruction and imaging method and system

Abstract

The application discloses a ground penetrating radar profile reconstruction and imaging method and system, and relates to the technical field of data processing. The method comprises the following steps: acquiring original profile data of a ground penetrating radar, performing abnormality detection on the original profile data of the ground penetrating radar, and obtaining processed profile data of the ground penetrating radar; reconstructing the processed profile data of the ground penetrating radar based on a preset deep neural network model, and obtaining reconstructed profile data of the ground penetrating radar; and imaging the reconstructed profile data of the ground penetrating radar by using a reverse time migration imaging method based on wave field decomposition, and obtaining a final imaging result of a target body. The reverse time migration imaging method based on wave field decomposition is obtained by improving a mutual correlation imaging condition based on wave field decomposition. The application maximally retains details of useful signals, and improves imaging resolution and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a ground penetrating radar profile reconstruction and imaging method and system. BACKGROUND

[0002] Ground penetrating radar (GPR) has been widely used in the field of engineering quality detection due to its high detection efficiency, high resolution and convenient operation, especially in the quality detection of hydraulic structures and other concrete lining structures in the water conservancy and hydropower industry.

[0003] During the data acquisition process of ground penetrating radar, due to the influence of the device itself and external environmental factors, a large amount of interference signals such as clutter, bad channels and sharp pulses are often mixed in the original echo profile, which causes the effective signal to be affected or even covered. It will affect the identification of effective signals, and may even completely cover the effective signals reflecting the characteristics of underground targets. At present, when processing ground penetrating radar data containing the above interference, it is largely dependent on manual interpretation and manual processing by the operator. However, this processing method is low in efficiency and the accuracy of the interpretation result is highly dependent on the technical level and practical experience of the operator, which is prone to errors due to subjective judgment.

[0004] In order to more accurately position and image the underground target, the reverse time migration (RTM) algorithm has become a preferred technology for ground penetrating radar data imaging. However, the conventional reverse time migration imaging method generally has a low-frequency noise problem when processing the full wave field, and these noises will form artifacts, which will interfere with the clarity and accuracy of the final imaging result.

[0005] In view of this, the present application is proposed. SUMMARY

[0006] The technical problem to be solved by the present application is how to effectively suppress the complex interference in the ground penetrating radar profile data while maximizing the preservation of the details of the useful signals, and how to overcome the problem of low-frequency noise forming artifacts and affecting the accuracy of imaging in existing imaging. The purpose of the present application is to provide a ground penetrating radar profile reconstruction and imaging method and system, which reconstructs the ground penetrating radar profile data based on the RCAN deep learning network, and improves the imaging resolution and accuracy by combining the reverse time migration imaging method based on wave field decomposition.

[0007] The present application is realized by the following technical scheme:

[0008] In a first aspect, the present application provides a ground penetrating radar profile reconstruction and imaging method, which comprises:

[0009] Obtaining ground penetrating radar original profile data, and performing anomaly detection on the ground penetrating radar original profile data to obtain processed ground penetrating radar profile data;

[0010] Based on a pre-defined deep neural network model, the processed ground-penetrating radar profile data is reconstructed to obtain the reconstructed ground-penetrating radar data profile.

[0011] The inverse time migration imaging method based on wave field decomposition is used to image the reconstructed ground penetrating radar data profile to obtain the final imaging result of the target. The inverse time migration imaging method based on wave field decomposition is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition.

[0012] Furthermore, anomaly detection is performed on the raw ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data, including:

[0013] Based on the original ground-penetrating radar profile data, multiple statistical features of each signal channel are extracted;

[0014] Based on statistical features, machine learning algorithms are used to perform cluster analysis on the signal channels, identify and classify normal and abnormal signal channels; and the data of abnormal signal channels are set to null values ​​to obtain processed ground penetrating radar profile data.

[0015] Furthermore, based on statistical features, machine learning algorithms are used to perform cluster analysis on the signal channels, identifying and classifying normal and abnormal signal channels, including:

[0016] Based on statistical features, the feature vector of each signal channel is calculated;

[0017] Based on the eigenvectors of each signal channel, a feature matrix is ​​constructed;

[0018] Principal component analysis was used to reduce the dimensionality of the feature matrix, and the principal component features of each signal channel after dimensionality reduction were obtained.

[0019] K-means clustering analysis was performed on the principal component features after dimensionality reduction to classify each signal channel into normal signal channels and at least one abnormal signal channel; the abnormal signal channels included noise signal channels and missing signal channels.

[0020] Furthermore, statistical characteristics include mean absolute deviation, standard deviation, root mean square value, and interquartile range.

[0021] Furthermore, the preset deep neural network model is a residual channel attention network model containing a residual-residual nested structure;

[0022] The residual-residual nested structure includes several residual group modules. Each residual group module includes multiple residual groups with long jump connections, and each residual group contains multiple residual channel attention modules with short jump connections.

[0023] Multi-scale features (features of different levels or types) are learned by dividing the attention modules of different residual channels, and multi-scale features are fused by skip connections to complete the data of the removed anomalous signal channels, thus obtaining the reconstructed ground-penetrating radar data profile; including:

[0024] By utilizing multi-level long-hop connections and short-hop connections to bypass missing information in the ground-penetrating radar profile, the remaining effective waveform features are learned to complete the data of the eliminated anomalous signal channels.

[0025] By utilizing the built-in channel attention mechanism in the residual channel attention module, global information extraction, nonlinear mapping, and weight generation are performed on the feature channels extracted during processing. This enhances the expression of feature channels containing key ground-penetrating radar profile information and suppresses interference, resulting in a reconstructed ground-penetrating radar data profile.

[0026] Furthermore, the training process of the pre-defined deep neural network model includes:

[0027] The loss between the model's predicted value and the actual output sample is calculated based on a preset hybrid loss function.

[0028] A preset optimizer with adaptive learning rate adjustment capability is used to iteratively optimize the parameters of the preset deep neural network model based on the loss.

[0029] The preset hybrid loss function is Loss due to mean square error Loss of the Pyramid of Laplace Weighted composition:

[0030] ;

[0031] in, These are the weighting coefficients; ,in, For the sample size, For the actual output sample, These are the model's predicted values; ,in, and Represent the predicted data and the actual output sample, respectively. Features of the Laplace Pyramid Based on the scale of the Laplace pyramid;

[0032] The optimizer is the RMSprop optimizer, and its parameter update process is as follows: ,in, For model parameters, Let be the model parameters for the t-th iteration; These are the model parameters for the (t+1)th iteration; The current gradient; It is an exponentially weighted average of the cumulative squared gradient. To prevent extremely small positive numbers with a denominator of 0; This is the learning rate.

[0033] Furthermore, the inverse time migration imaging method based on wavefield decomposition includes:

[0034] Substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field; the first electromagnetic wave field is the source wave field propagating forward from the source point.

[0035] The reconstructed ground-penetrating radar data profile is used as the source input at the corresponding receiving point location to calculate the second electromagnetic wave field; the second electromagnetic wave field is the received wave field propagating backward from the receiving point.

[0036] The first and second electromagnetic wave fields were decomposed directionally using the Poynting vector decomposition method, resulting in four sets of wave field component combinations.

[0037] False imaging conditions are removed from the four sets of wave field component combinations to obtain the effective wave field component combinations; the effective wave field component combinations are the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field.

[0038] Based on the combination of effective wavefield components, the improved cross-correlation imaging conditions are applied to generate the final imaging result.

[0039] Furthermore, the improved cross-correlation imaging conditions are as follows:

[0040] ;

[0041] In the formula, For the final imaging result; This represents the down-current wave field component of the first electromagnetic wave field; This represents the upward wave field component of the second electromagnetic wave field; The maximum time range considered for electromagnetic wave field propagation.

[0042] Secondly, the present invention provides a ground-penetrating radar profile reconstruction and imaging system, the system comprising:

[0043] The acquisition unit is used to acquire raw profile data from ground-penetrating radar.

[0044] Anomaly detection unit is used to detect anomalies in the raw ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data.

[0045] The reconstruction unit is used to reconstruct the processed ground-penetrating radar profile data based on a preset deep neural network model, so as to obtain the reconstructed ground-penetrating radar data profile.

[0046] An improved imaging unit is used to image the reconstructed ground-penetrating radar data profile using a wave field decomposition-based inverse time migration imaging method to obtain the final imaging result of the target. The wave field decomposition-based inverse time migration imaging method is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition.

[0047] Furthermore, the improved imaging unit includes:

[0048] The first calculation subunit is used to substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field; the first electromagnetic wave field is the source wave field propagating forward from the source point.

[0049] The second calculation subunit uses the reconstructed ground-penetrating radar data profile as the source input at the corresponding receiving point location to calculate the second electromagnetic wave field; the second electromagnetic wave field is the received wave field propagating backward from the receiving point.

[0050] The wave field decomposition sub-unit is used to perform directional decomposition of the first electromagnetic wave field and the second electromagnetic wave field using the Poynting vector decomposition method, respectively, to obtain four sets of wave field component combinations.

[0051] The wave field component optimization subunit is used to remove false imaging conditions from the four sets of wave field component combinations to obtain the effective wave field component combination; the effective wave field component combination is the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field.

[0052] The imaging subunit is used to generate the final imaging result by applying improved cross-correlation imaging conditions based on the combination of effective wavefield components.

[0053] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0054] 1. This invention provides a ground-penetrating radar profile reconstruction and imaging method and system. It utilizes an automated anomaly detection method based on principal component analysis (PCA) and K-means clustering algorithm to extract and cluster the statistical features of the signal, thereby achieving accurate identification of anomaly signal channels. The identified anomaly signal channels are removed from the original ground-penetrating radar profile data, while normal signal channel data is retained. Noisy signal channels and missing signal channel data are removed to ensure the purity of subsequent processing.

[0055] 2. The present invention provides a ground-penetrating radar profile reconstruction and imaging method and system, which proposes a GPR profile image reconstruction method based on a residual channel attention network model. By introducing residual connections, channel attention mechanisms and multi-scale hybrid loss functions, a high-fidelity GPR image is reconstructed, i.e., the reconstructed ground-penetrating radar data profile.

[0056] 3. This invention provides a ground-penetrating radar (GPR) profile reconstruction and imaging method and system. The reconstructed GPR data profile is then imaged with high precision using a reverse time migration imaging method based on wavefield decomposition. This method introduces wavefield decomposition into the imaging conditions and combines it with Poynting vector decomposition to achieve uplink and downlink wavefield decomposition, retaining only the effective imaging conditions. This significantly reduces low-frequency artifacts and improves imaging resolution, ultimately generating a high-quality imaging profile of underground structures, thus enhancing the imaging resolution and accuracy of underground targets. Attached Figure Description

[0057] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a ground-penetrating radar profile reconstruction and imaging method according to the present invention;

[0059] Figure 2 This is a diagram of the pre-defined deep neural network model (RCAN network) structure of this invention;

[0060] Figure 3 This is a schematic diagram of the laboratory setup for a simulation experiment according to Embodiment 1 of the present invention;

[0061] Figure 4 This is Example 1 of the present invention. Figure 3 Raw data obtained from data collection in a sand pit where pipes were buried; Figure 5 In Embodiment 1 of the present invention, the data that has been detected for anomalies and the abnormal signal channel has been set to null values ​​is the processed ground-penetrating radar profile data.

[0062] Figure 6 This is the ground-penetrating radar data profile after reconstruction in step 2 of Embodiment 1 of the present invention;

[0063] Figure 7 This is the final imaging result of Embodiment 1 of the present invention after step 3 of the inverse time migration imaging method based on wave field decomposition;

[0064] Figure 8 This is a block diagram of a ground-penetrating radar profile reconstruction and imaging system according to the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0066] To effectively suppress complex interference in ground-penetrating radar (GPR) profile data while preserving the details of useful signals to the maximum extent, and to overcome the problem that low-frequency noise in existing imaging systems easily forms artifacts and affects the accuracy of imaging, this invention proposes a GPR profile reconstruction and imaging method and system, including the following core steps:

[0067] First, the GPR profile is imported as the original input data (i.e., the original ground-penetrating radar profile data). For any potential anomalous signal channels, this invention utilizes an automated anomaly detection method based on Principal Component Analysis (PCA) and K-means clustering to extract and cluster the statistical features of the signals, achieving accurate identification of anomalous signal channels. Identified anomalous signal channels are then removed from the original ground-penetrating radar profile data, retaining normal signal channel data while eliminating noisy and missing signal channels to ensure the purity of subsequent processing. Second, in cases of missing data, to restore image integrity, this invention proposes a GPR profile image reconstruction method based on Residual Channel Attention Network (RCAN). By introducing residual connections, channel attention mechanisms, and a multi-scale hybrid loss function, a high-fidelity GPR image is reconstructed, i.e., the reconstructed ground-penetrating radar data profile. Third, for the reconstructed ground-penetrating radar data profile, high-precision imaging is performed using a wavefield decomposition-based inverse time-shifting imaging method. This method achieves up-and-down decomposition of the wave field through Poynting vector decomposition, retaining only the effective imaging conditions, significantly reducing low-frequency artifacts and improving imaging resolution, and ultimately generating a high-quality imaging profile of the subsurface target.

[0068] Example 1

[0069] like Figure 1 As shown, the present invention provides a ground-penetrating radar profile reconstruction and imaging method, the method comprising:

[0070] Step 1: Obtain the raw ground-penetrating radar profile data and perform anomaly detection on the raw ground-penetrating radar profile data to obtain the processed ground-penetrating radar profile data.

[0071] In this embodiment, the raw ground-penetrating radar profile data refers to the unprocessed two-dimensional data map directly recorded by the ground-penetrating radar equipment during on-site measurement, consisting of multiple signal channels arranged along the survey line. Each signal channel represents a series of radar echo signals that vary with time (or depth) received at a measurement point.

[0072] In this embodiment, step 1 specifically includes:

[0073] Step 11: Acquire raw ground-penetrating radar profile data, including signal data from several channels. i The signal data of each signal channel is ,in,n This represents the number of sampling points for each signal channel; T is the transpose.

[0074] Step 12: Extract multiple statistical features from each signal channel based on the original ground-penetrating radar profile data;

[0075] This invention uses mathematical calculations to transform each raw, unstructured signal waveform into a structured feature vector composed of multiple numerical values. The statistical features include: mean absolute deviation, standard deviation, root mean square value, and interquartile range.

[0076] Among them, the mean absolute deviation and standard deviation are mainly used to measure the dispersion or volatility of signal data points relative to their mean; the root mean square value mainly reflects the effective strength or overall energy of the signal over a period of time; and the interquartile range describes the dispersion of the data by measuring the distribution range of the middle 50% of the data.

[0077] Specifically, for each signal data Extract the following four features:

[0078] Mean absolute deviation characteristic ,in For the first i The average value of the signal data from each signal channel reflects the average fluctuation amplitude of the signal; For the first i The first in the signal channel j Data from each sampling point;

[0079] Standard deviation characteristic This measures the degree of dispersion of the data.

[0080] Root mean square value eigenvalue It is related to the power of the signal and characterizes the overall energy intensity of the signal.

[0081] The interquartile range characteristic is: ,in, and The first i Signal channel signal data The values ​​at the 25th and 75th percentiles after the data is sorted from smallest to largest describe the dispersion of the middle part of the data.

[0082] This step uses mathematical statistical methods to quantify and analyze each signal, calculating numerical values ​​that reflect its signal strength, dispersion, waveform morphology, and other characteristics, i.e., statistical feature quantities.

[0083] Step 13: Based on statistical features, a machine learning algorithm is used to perform cluster analysis on the signal channels to identify and classify normal and abnormal signal channels; and the data of abnormal signal channels are set to null values ​​to obtain the processed ground-penetrating radar profile data.

[0084] In this step, machine learning algorithms are techniques that enable computers to automatically learn patterns from data and make decisions. Cluster analysis, as an unsupervised learning method, can automatically group signal channels with similar characteristics into one category without pre-defined labels. For example, signal channels with extremely low or completely absent signal strength (missing channels), and signal channels with violent and irregular fluctuations (noise channels), will have statistically distinct characteristics compared to normal signal channels with clear waveforms and regular reflections. Cluster analysis enables the automatic detection and separation of abnormal signal channels.

[0085] Step 13 of this invention can be divided into two stages: First, the data is dimensionality reduced by principal component analysis (PCA) to extract key features and reduce computational complexity; then, based on the dimensionality-reduced features, the signal channels are automatically classified using the K-means clustering algorithm.

[0086] In this embodiment, step 13 includes:

[0087] Step 131: Calculate the feature vector for each signal channel based on statistical features;

[0088] Step 132: Construct a feature matrix based on the feature vectors of each signal channel;

[0089] Step 133: Principal component analysis is used to reduce the dimensionality of the feature matrix to obtain the principal component features of each signal channel after dimensionality reduction.

[0090] Step 134: Perform K-means clustering analysis on the dimensionality-reduced principal component features to classify each signal channel into normal signal channels and at least one abnormal signal channel; abnormal signal channels include noise signal channels and missing signal channels.

[0091] Specifically, step 132 includes:

[0092] The first i Mean absolute deviation characteristic of signal data in each channel Standard deviation characteristic Root mean square value eigenvalue interquartile range feature Constructing the feature vector of this signal channel : ;

[0093] Will The feature vector of each signal channel Combine them to form a feature matrix : ,in This represents the total number of signal channels; each row represents a signal channel (i.e., a sample), and each column represents a specific statistical characteristic (e.g., the first column is the MAD value of all signal channels, the second column is the STD value of all signal channels, etc.).

[0094] Specifically, step 133 includes:

[0095] To reduce cluster dimensionality and enhance feature discriminability, the feature matrix... Each column is standardized to obtain a standardized feature matrix. The standardized calculation formula is: ,in, Characteristic matrix The Line number Column eigenvalues ​​(Note: =1 indicates MAD. =2 represents STD. =3 represents RMS. =4 is for IQR), For the first The mean of the column features, For the first The standard deviation of the column features; after this processing, the standardized feature matrix is ​​obtained. Each column will have a zero mean and a unit standard deviation;

[0096] Calculate the normalized feature matrix covariance matrix , ,in The total number of signal channels; the covariance matrix is ​​a square matrix, where the elements on the diagonal represent the variance of each feature itself, and the elements off the diagonal represent the covariance (i.e., correlation) between different features.

[0097] Calculate the covariance matrix eigenvalues and the corresponding feature vectors ;

[0098] Select the eigenvectors corresponding to the two largest eigenvalues Constructing the projection matrix And through the formula Standardize the feature matrix Dimensionality reduction to a two-dimensional principal component space yields the dimensionality-reduced principal component features of each signal channel. , For the first Two-dimensional principal components of the signal channel.

[0099] Specifically, step 134 includes:

[0100] The number of clusters is determined based on the total number of normal and abnormal signal channels. And randomly select or generate Initial cluster centroids ; The value is determined by the total number of signal channel categories to be distinguished as needed. In this implementation, abnormal signal channels are set as noise channels and missing channels, i.e. ;

[0101] Repeat the following allocation and update steps until a preset termination condition is met. The preset termination condition is: , The preset threshold;

[0102] Allocation steps: in the first t During each iteration, the algorithm traverses every principal component feature point. For each principal component feature... According to the formula Calculate its Euclidean distance to all current centroids and assign it to the nearest centroid. In the corresponding cluster, For the first t The v-th cluster centroid of the next iteration; that is, calculating its relationship with all current clusters... Cluster centroid The Euclidean distance between them. Then, according to the principle of minimum distance, the feature point is... It is assigned to the cluster represented by the nearest centroid.

[0103] Update steps: After all feature points have been assigned, the algorithm needs to recalculate the centroid for each cluster; for the th... l The one in the first t Clusters under round iteration According to the formula Calculating the mean of all principal component features within the cluster is equivalent to moving the centroid to the center of its member points, and using this centroid as the cluster centroid for the next iteration. ;

[0104] After the iteration terminates, all signal channels are classified into normal signal channels and abnormal signal channels based on the final cluster assignment results. The loop stops when the iteration process meets the termination condition. At this point, each feature point belongs to a fixed, final cluster. Since the characteristics of normal signal channels are usually more consistent and their numbers are greater, they will form a large, concentrated main cluster in the feature space; while various abnormal signal channels will form one or more smaller, more remote secondary clusters.

[0105] In this embodiment, step 13 further includes:

[0106] Step 135: Set the data of abnormal signal channels to null values. For example, assign null values ​​to the data of noise signal channels and missing signal channels to obtain the processed ground penetrating radar profile data.

[0107] Step 2: Based on the preset deep neural network model, the processed ground-penetrating radar profile data is reconstructed to obtain the reconstructed ground-penetrating radar data profile.

[0108] This step, following the processing of the ground-penetrating radar (GPR) profile data obtained in step 1, removes detected anomalous signal data, leading to data gaps in the GPR profile and affecting subsequent data processing. Therefore, this invention, after eliminating interference from anomalous signal channels, utilizes a trained deep learning network model to repair and enhance the processed GPR profile data, i.e., reconstructing the processed GPR profile data. This network model can understand the normal waveform structure (such as hyperbolic morphology) in the GPR profile, fill in the data gaps caused by the removal of anomalous signal channels, and output a more complete and clearly structured reconstructed GPR profile data, i.e., the reconstructed GPR data profile.

[0109] This invention employs a Residual Channel Attention Network (RCAN) model for ground-penetrating radar profile reconstruction. The RCAN network consists of multiple residual groups with long-skip connections, each containing multiple residual blocks with short-skip connections, forming a Residual Nesting (RIR) structure. The RIR structure allows multiple skip connections to bypass a large amount of missing information, enabling the main network to focus on learning waveform information. Furthermore, this network introduces a channel attention mechanism, adaptively recalibrating channel features by considering the interdependencies between them. Generally, attention can be viewed as a form of guidance, allocating available processing resources to the most information-rich parts of the input.

[0110] In this embodiment, the preset deep neural network model is a residual channel attention network model containing a residual-residual nested structure;

[0111] The residual-residual nested structure includes several residual group modules. Each residual group module includes multiple residual groups with long jump connections, and each residual group contains multiple residual channel attention modules with short jump connections.

[0112] Multi-scale features (features of different levels or types) are learned by dividing the attention modules of different residual channels, and multi-scale features are fused by skip connections to complete the data of the removed anomalous signal channels, thus obtaining the reconstructed ground-penetrating radar data profile; including:

[0113] (1) By using multi-level long-hop connections and short-hop connections to bypass the missing information in the ground-penetrating radar profile, the remaining effective waveform features are learned to complete the data of the eliminated abnormal signal channels.

[0114] In the above technical solutions, this multi-level skip connection design, especially the long skip connections which can span multiple residual groups and the short skip connections which operate between attention blocks in the residual channels within a residual group, allows the network to bypass a large amount of missing information. In the processed ground-penetrating radar profile data, due to the detection of abnormal trajectories leading to data gaps, the multi-level skip connections ensure the effective transmission of information, allowing the main network to focus on learning the remaining valuable waveform information without being excessively interfered with by the missing parts, thus providing more accurate basic information for reconstruction.

[0115] In traditional processing, treating missing parts equally hinders the expressive power of neural networks. However, the RCAN network, through its unique structure, can distinguish between missing parts at different locations and perform differentiated processing on different feature information. This allows the network to more flexibly recover missing parts based on the actual data during reconstruction, improving the accuracy and reliability of the reconstruction.

[0116] (2) By using the channel attention mechanism built into the residual channel attention module, global information extraction, nonlinear mapping and weight generation are performed on the feature channels extracted during the processing, the expression of the feature channels containing key ground-penetrating radar profile information is enhanced and interference is suppressed, and the reconstructed ground-penetrating radar data profile is obtained.

[0117] In the above technical solutions, the residual channel attention blocks and the overall residual structure in the RCAN network enable the network to learn more complex feature representations without increasing computational cost. Residual connections allow the network to directly learn the differences between input and output, which helps train deeper networks and thus uncover more subtle and complex features in GPR profile data, playing a crucial role in reconstructing fine GPR profile details.

[0118] like Figure 2 As shown, Figure 2 This is a diagram of the pre-defined deep neural network model structure. The pre-defined deep neural network model (i.e., the RCAN network) forms a hierarchical and powerful network architecture through the organic combination of the aforementioned modules. It can extract and analyze features from ground-penetrating radar data from multiple angles and levels, fully mining the potential information within the data. When recovering missing trace data, it can utilize both local detailed features and grasp the global structure and correlation of the data, thereby generating recovery results highly similar to the original data, effectively improving the integrity and usability of ground-penetrating radar data in cases of missing traces.

[0119] In this embodiment, the training process of the preset deep neural network model includes:

[0120] Step A, Dataset Construction

[0121] The RCAN network takes a GPR profile with missing signal channels as input (i.e., the processed ground-penetrating radar profile data obtained in step 1) and outputs the corresponding reconstructed image (i.e., the reconstructed ground-penetrating radar data profile). To train this network, this embodiment constructs a dataset containing 2000 samples, specifically including:

[0122] 1600 sets of data: GPR images with missing signal channels are generated through numerical simulation (simulating common hydraulic structures).

[0123] 400 sets of measured data: Images with anomalies were selected from real GPR data and the missing parts were manually marked (measured data of hydraulic structures).

[0124] All samples were divided into training, validation, and test sets in a 4:2:2 ratio to ensure that the model training process has good generalization ability.

[0125] Step B, Model Training

[0126] The loss between the model's predicted value and the actual output sample is calculated based on a preset hybrid loss function.

[0127] To enhance the model's ability to recover image details while maintaining the spatiotemporal continuity of GPR data, a pre-defined hybrid loss function is constructed using a combination of Mean Squared Error (MSE) and Laplacian Pyramid Loss.

[0128] The preset hybrid loss function is Loss due to mean square error Loss of the Pyramid of Laplace Weighted composition:

[0129] ;

[0130] in, This is the weighting coefficient, set to 0.1, used to balance the influence of the two loss components;

[0131] ,in, For the sample size, For the actual output sample, The model predicts the values; using mean squared error as the loss function ensures more stable model predictions that are consistent with the characteristics of ground-penetrating radar profile data.

[0132] ,in, and Represent the predicted data and the actual output sample, respectively. Features of the Laplace Pyramid Based on the scale of the Laplace pyramid; The value is set to 5, meaning the proposed loss function is computed on images at five different scales. Using a Laplacian kernel to process image features at different scales yields the Laplacian pyramid loss function, extracting more image texture and holding potential significance for multi-scale image denoising and enhancement.

[0133] Step C, Optimize Training

[0134] A preset optimizer with adaptive learning rate adjustment capability is adopted to iteratively optimize the parameters of a preset deep neural network model based on the loss; this optimizer has adaptive learning rate adjustment capability and can effectively deal with gradient oscillation problem.

[0135] The optimizer is the RMSprop optimizer, and its parameter update process is as follows: ,in, For model parameters, Let be the model parameters for the t-th iteration; These are the model parameters for the (t+1)th iteration; The current gradient; This is an exponentially weighted average of the cumulative squared gradient;

[0136] ,in, For the first t Model parameters for each iteration; The current gradient; The learning rate; This is the momentum coefficient (usually 0.9). To prevent extremely small positive numbers with a denominator of 0; This is the learning rate.

[0137] During model training, the following hyperparameters were set: Batch Size: 32; Maximum Training Epochs: 500; After training, a model with robustness and reconstruction accuracy was obtained.

[0138] By constructing an RCAN network with multi-scale residual connections and channel attention mechanisms, high-fidelity reconstruction of GPR images after anomaly removal is achieved, which has the following advantages:

[0139] (1) It can stably recover the signal structure under different missing modes; (2) The network automatically focuses on the high information channel and suppresses noise interference; (3) It combines MSE and multi-scale loss function to significantly improve the ability to recover details; (4) It supports end-to-end image recovery process, which is efficient and convenient, and can perform subsequent imaging processing on the reconstructed image.

[0140] Step 3: The reconstructed ground-penetrating radar data profile is imaged using the wave field decomposition-based inverse time migration imaging method to obtain the final imaging result of the target. The wave field decomposition-based inverse time migration imaging method is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition.

[0141] Specifically, the ground-penetrating radar data profile reconstructed in step 2 is subjected to reverse time migration (RTM) imaging. To address the low-frequency noise and artifact problems in traditional RTM algorithms, this invention proposes a reverse time migration imaging method based on wavefield decomposition. By introducing wavefield decomposition into the imaging conditions, the imaging resolution and accuracy of underground targets are improved.

[0142] Traditional reverse-time migration algorithms rely on zero-delay cross-correlation imaging conditions and utilize the principle of explosion reflection. Based on the principle of time consistency, the electromagnetic waves emitted from each source point propagate backward at the receiver point. At each time step, the forward wavefield from the source point is multiplied point-by-point with the backward-propagating wavefield from the receiver point and accumulated to obtain the final imaging profile. The expression for this process is as follows:

[0143] ;

[0144] In the formula, I ( x , z This indicates the imaging results of the underground space. S ( x , z , t () represents the first electromagnetic wave field propagating forward from the source point. R ( x , z , t () represents the second electromagnetic wave field propagating backward from the receiving point. However, conventional imaging conditions often contain low-frequency noise and spurious images caused by erroneous wave paths. To eliminate low-frequency noise and improve the quality of the migrated images, this invention proposes a reverse-time migration imaging method based on wavefield decomposition.

[0145] In this embodiment, the inverse time migration imaging method based on wavefield decomposition proposed in this invention includes:

[0146] Step 31: Substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field. S ( x ,z , t ), where x and z represent positions, t Indicates the recording time; the first electromagnetic wave field is the source wave field propagating forward from the source point;

[0147] The Maxwell equations for two-dimensional transverse magnetic waves (TM waves) are as follows:

[0148] ;

[0149] in, H This represents the magnetic field strength (unit: V / m). This represents the component of the magnetic field strength in the x-direction. This represents the component of the magnetic field strength in the y-direction; E This represents the electric field strength (unit: V / m). This represents the component of the electric field intensity in the z-direction; The dielectric constant of the medium (unit: F / m). μ The magnetic permeability of the medium (unit: H / m). σ Indicates electrical conductivity (unit: S / m). t Indicates time (unit: seconds). Jz The excitation source is represented by x; x represents the x-direction of the spatial coordinate system; y represents the y-direction of the spatial coordinate system; and z represents the z-direction of the spatial coordinate system.

[0150] Step 32: Using the reconstructed ground-penetrating radar data profile as the source input at the corresponding receiving point location, calculate the second electromagnetic wave field. R ( x , z , t The second electromagnetic wave field is the received wave field that propagates in the reverse direction from the receiving point.

[0151] Step 33: The first electromagnetic wave field and the second electromagnetic wave field are decomposed directionally using the Poynting vector decomposition method to obtain four sets of wave field component combinations.

[0152] Specifically, during electromagnetic wave propagation, it can be divided into an ascending wave field and a descending wave field. The forward-propagating source wave field (i.e., the first electromagnetic wave field) and the backward-propagating receiving wave field (i.e., the second electromagnetic wave field) can be represented as follows:

[0153] ;

[0154] Therefore, for conventional imaging conditions, the summation of the four wave fields is expressed as follows: , , and .in, and Considered reliable imaging conditions, while and Considered as spurious imaging conditions, this could lead to incorrect wave paths, and Low-frequency artifacts will be generated. All three of these situations can interfere with the reverse-time migration profile, leading to misjudgment by the operator.

[0155] Specifically, the Poynting vector decomposition is used to separate the entire wavefield into ascending and descending waves. For the ground-penetrating radar wavefield, the Poynting vector decomposition method can be expressed as:

[0156] ;

[0157] in, P Let represent the wave field to be separated, and ∇ represent the spatial derivative operator. Based on this formula, when the wave field... P When the peak value is reached, the calculation of the Poynting vector faces a stability problem, causing the Poynting vector amplitude to become 0, thus losing its directionality. To enhance the robustness of the algorithm, the Poynting vector defined by local least squares average is used, and the following is applied: After processing the formula accordingly, we obtain the following formula:

[0158] ;

[0159] in, For stable local Poynting vectors; This is the instantaneous Poynting vector calculated at each sampling point within the neighborhood; For local spatiotemporal neighborhood; This is to sum the values ​​of all sampling points in the neighborhood.

[0160] Based on this formula, the propagation direction of electromagnetic waves can be calculated using the Poynting vector decomposition method. The Poynting vector allows the entire wave field to be separated into components in different directions.

[0161] Step 34: Remove false imaging conditions from the four sets of wave field component combinations to obtain the effective wave field component combinations; the effective wave field component combinations are the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field.

[0162] The four sets of wave field components obtained from the decomposition in step 33 are combined , and Removed, leaving only reliable imaging conditions. ,in, This represents the downward wave field component of the first electromagnetic wave field. This represents the down-current wave field component of the first electromagnetic wave field.

[0163] Step 35: Based on the effective wave field component combination, apply the improved cross-correlation imaging conditions, multiply the first electromagnetic wave field and the second electromagnetic wave field point by point and sum them to generate the final imaging result.

[0164] Specifically, the improved cross-correlation imaging conditions are as follows:

[0165] ;

[0166] In the formula, For the final imaging result; This represents the down-current wave field component of the first electromagnetic wave field; This represents the upward wave field component of the second electromagnetic wave field; The maximum time range considered for electromagnetic wave field propagation.

[0167] In the above technical solution, both the first and second electromagnetic wave fields are decomposed into Poynting vectors in the z direction, and the improved cross-correlation imaging conditions are applied to perform reverse time migration imaging to generate the final imaging result.

[0168] The reverse time migration imaging method based on wavefield decomposition proposed in this invention has the following advantages:

[0169] (1) Significantly reduce low-frequency noise: By filtering out erroneous imaging terms, artifacts and false reflection signals are effectively suppressed;

[0170] (2) Enhance imaging resolution: Emphasize the energy superposition of effective reflected waves to improve the quality of detailed imaging;

[0171] (3) Improve the reliability of interpretation: reduce the risk of misjudgment in geological interpretation and are suitable for high-precision identification of complex underground structures.

[0172] In practice, Figure 3 A schematic diagram of the layout for a laboratory simulation experiment shows five pipes buried in a sandpit in the laboratory, illustrating their relative positions. Figure 4 To Figure 3 Raw data obtained from data collection in a sand pit where pipes were buried;

[0173] Figure 5 The data that has been detected for anomalies and has anomaly signal channels set is empty, i.e., the processed ground-penetrating radar profile data; Figure 6 This is the ground-penetrating radar data profile after reconstruction in step 2; Figure 7 The final imaging result is obtained after step 3, which uses the inverse time migration imaging method based on wave field decomposition.

[0174] Example 2

[0175] The difference between this embodiment and Embodiment 1 is that, based on the final imaging results of Embodiment 1, geological interpretation and structural analysis are performed to identify hydraulic structure anomalies, thereby achieving precise positioning of the target anomaly.

[0176] Example 3

[0177] like Figure 8 As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a ground-penetrating radar profile reconstruction and imaging system, which corresponds one-to-one with the ground-penetrating radar profile reconstruction and imaging method in Embodiment 1; the system includes:

[0178] The acquisition unit is used to acquire raw profile data from ground-penetrating radar.

[0179] Anomaly detection unit is used to detect anomalies in the raw ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data.

[0180] The reconstruction unit is used to reconstruct the processed ground-penetrating radar profile data based on a preset deep neural network model, so as to obtain the reconstructed ground-penetrating radar data profile.

[0181] An improved imaging unit is used to image the reconstructed ground-penetrating radar data profile using a wave field decomposition-based inverse time migration imaging method to obtain the final imaging result of the target. The wave field decomposition-based inverse time migration imaging method is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition.

[0182] As a further implementation, the improved imaging unit includes:

[0183] The first calculation subunit is used to substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field; the first electromagnetic wave field is the source wave field propagating forward from the source point.

[0184] The second calculation subunit uses the reconstructed ground-penetrating radar data profile as the source input at the corresponding receiving point location to calculate the second electromagnetic wave field; the second electromagnetic wave field is the received wave field propagating backward from the receiving point.

[0185] The wave field decomposition sub-unit is used to perform directional decomposition of the first electromagnetic wave field and the second electromagnetic wave field using the Poynting vector decomposition method, respectively, to obtain four sets of wave field component combinations.

[0186] The wave field component optimization subunit is used to remove false imaging conditions from the four sets of wave field component combinations to obtain the effective wave field component combination; the effective wave field component combination is the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field.

[0187] The imaging subunit is used to generate the final imaging result by applying improved cross-correlation imaging conditions based on the combination of effective wavefield components.

[0188] The execution process of each unit can be carried out according to the steps of the ground-penetrating radar profile reconstruction and imaging method in Example 1, and will not be described in detail in this example.

[0189] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 Figure 1 The steps of the function specified in one or more boxes.

[0193] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A ground-penetrating radar profile reconstruction and imaging method, characterized in that, The method includes: The raw ground-penetrating radar profile data is acquired, and anomaly detection is performed on the raw ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data. Based on a preset deep neural network model, the processed ground-penetrating radar profile data is reconstructed to obtain a reconstructed ground-penetrating radar data profile. The reconstructed ground-penetrating radar data profile is imaged using a wave field decomposition-based inverse time migration imaging method to obtain the final imaging result of the target. The wave field decomposition-based inverse time migration imaging method is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition. The training process of the preset deep neural network model includes: The loss between the model's predicted value and the actual output sample is calculated based on a preset hybrid loss function. A preset optimizer with adaptive learning rate adjustment capability is used to iteratively optimize the parameters of the preset deep neural network model based on the loss. The preset hybrid loss function is Loss due to mean square error Loss of the Pyramid of Laplace Weighted composition: ； in, These are the weighting coefficients; ,in, For the sample size, For the actual output sample, These are the model's predicted values; ,in, and Represent the predicted data and the actual output sample, respectively. Features of the Laplace Pyramid Based on the scale of the Laplace pyramid; The optimizer is the RMSprop optimizer, and its parameter update process is as follows: ,in, For model parameters, Let be the model parameters for the t-th iteration; These are the model parameters for the (t+1)th iteration; This is the current gradient; It is an exponentially weighted average of the cumulative squared gradient. To prevent extremely small positive numbers with a denominator of 0; This is the learning rate.

2. The ground-penetrating radar profile reconstruction and imaging method according to claim 1, characterized in that, Anomaly detection is performed on the raw ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data, including: Based on the original ground-penetrating radar profile data, multiple statistical feature quantities of each signal channel are extracted; Based on the statistical features, a machine learning algorithm is used to perform cluster analysis on the signal channels to identify and classify normal and abnormal signal channels; and the data of abnormal signal channels are set to null values ​​to obtain the processed ground-penetrating radar profile data.

3. The ground-penetrating radar profile reconstruction and imaging method according to claim 2, characterized in that, Based on the aforementioned statistical features, a machine learning algorithm is used to perform cluster analysis on the signal channels, identifying and classifying normal and abnormal signal channels, including: Based on the aforementioned statistical features, the feature vector of each signal channel is calculated. Based on the eigenvectors of each signal channel, a feature matrix is ​​constructed; Principal component analysis was used to reduce the dimensionality of the feature matrix to obtain the dimensionality-reduced principal component features of each signal channel. K-means clustering analysis was performed on the principal component features after dimensionality reduction to classify each signal channel into normal signal channels and at least one abnormal signal channel; the abnormal signal channels include noise signal channels and missing signal channels.

4. The ground-penetrating radar profile reconstruction and imaging method according to claim 2, characterized in that, The statistical characteristics include mean absolute deviation, standard deviation, root mean square value, and interquartile range.

5. The ground-penetrating radar profile reconstruction and imaging method according to claim 1, characterized in that, The preset deep neural network model is a residual channel attention network model containing a residual-residual nested structure; The residual-residual nested structure includes several residual group modules, each residual group module includes multiple residual groups with long jump connections, and each residual group contains multiple residual channel attention modules with short jump connections. Multi-scale features are learned by dividing the attention modules of different residual channels, and the multi-scale features are fused by skip connections to complete the data of the removed anomalous signal channels, thus obtaining the reconstructed ground-penetrating radar data profile; including: By utilizing multi-level long-hop connections and short-hop connections to bypass missing information in the ground-penetrating radar profile, the remaining effective waveform features are learned to complete the data of the eliminated anomalous signal channels. By using the channel attention mechanism built into the residual channel attention module, global information extraction, nonlinear mapping and weight generation are performed on the feature channels extracted during the processing, thereby enhancing the expression of feature channels containing key ground-penetrating radar profile information and obtaining the reconstructed ground-penetrating radar data profile.

6. The ground-penetrating radar profile reconstruction and imaging method according to claim 1, characterized in that, Reverse time migration imaging based on wavefield decomposition includes: Substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field; the first electromagnetic wave field is the source wave field propagating forward from the source point. The reconstructed ground-penetrating radar data profile is used as the source input at the corresponding receiving point location to calculate the second electromagnetic wave field; the second electromagnetic wave field is the receiving wave field propagating backward from the receiving point. The first electromagnetic wave field and the second electromagnetic wave field were decomposed directionally using the Poynting vector decomposition method, resulting in four sets of wave field component combinations. False imaging conditions are removed from the four sets of wave field component combinations to obtain the effective wave field component combinations; the effective wave field component combinations are the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field. Based on the combination of effective wavefield components, the improved cross-correlation imaging conditions are applied to generate the final imaging result.

7. The ground-penetrating radar profile reconstruction and imaging method according to claim 6, characterized in that, The improved cross-correlation imaging conditions are as follows: ； In the formula, For the final imaging result; This represents the down-current wave field component of the first electromagnetic wave field; This represents the upward wave field component of the second electromagnetic wave field; The maximum time range considered for electromagnetic wave field propagation.

8. A ground-penetrating radar profile reconstruction and imaging system, characterized in that, The system includes: The acquisition unit is used to acquire raw profile data from ground-penetrating radar. An anomaly detection unit is used to perform anomaly detection on the original ground-penetrating radar profile data to obtain processed ground-penetrating radar profile data. The reconstruction unit is used to reconstruct the processed ground-penetrating radar profile data based on a preset deep neural network model to obtain the reconstructed ground-penetrating radar data profile. An improved imaging unit is used to image the reconstructed ground-penetrating radar data profile using a wave field decomposition-based inverse time migration imaging method to obtain the final imaging result of the target; the wave field decomposition-based inverse time migration imaging method is an inverse time migration method obtained by improving the cross-correlation imaging conditions based on wave field decomposition. The training process of the preset deep neural network model includes: The loss between the model's predicted value and the actual output sample is calculated based on a preset hybrid loss function. A preset optimizer with adaptive learning rate adjustment capability is used to iteratively optimize the parameters of the preset deep neural network model based on the loss. The preset hybrid loss function is Loss due to mean square error Loss of the Pyramid of Laplace Weighted composition: ； in, These are the weighting coefficients; ,in, For the sample size, For the actual output sample, These are the model's predicted values; ,in, and Represent the predicted data and the actual output sample, respectively. Features of the Laplace Pyramid Based on the scale of the Laplace pyramid; The optimizer is the RMSprop optimizer, and its parameter update process is as follows: ,in, For model parameters, Let be the model parameters for the t-th iteration; These are the model parameters for the (t+1)th iteration; This is the current gradient; It is an exponentially weighted average of the cumulative squared gradient. To prevent extremely small positive numbers with a denominator of 0; This is the learning rate.

9. A ground-penetrating radar profile reconstruction and imaging system according to claim 8, characterized in that, The improved imaging unit includes: The first calculation subunit is used to substitute the source function into Maxwell's equations to calculate the first electromagnetic wave field; the first electromagnetic wave field is the source wave field propagating forward from the source point. The second calculation subunit uses the reconstructed ground-penetrating radar data profile as the source input at the corresponding receiving point location to calculate the second electromagnetic wave field; the second electromagnetic wave field is the receiving wave field propagating backward from the receiving point. The wave field decomposition subunit is used to perform directional decomposition of the first electromagnetic wave field and the second electromagnetic wave field respectively using the Poynting vector decomposition method to obtain four sets of wave field component combinations. The wave field component optimization subunit is used to remove false imaging conditions from four sets of wave field component combinations to obtain effective wave field component combinations; the effective wave field component combinations are the down-going wave field component of the first electromagnetic wave field and the up-going wave field component of the second electromagnetic wave field. An imaging subunit is used to generate the final imaging result by applying improved cross-correlation imaging conditions based on the combination of the effective wavefield components.

Images