Intelligent method for rock mass fracture distribution form identification and parameter inversion based on GPR

Abstract

The invention provides a GPR-based intelligent method for rock mass fracture distribution form identification and parameter inversion. The method comprises the following steps: a first stage: intelligently classifying fracture types; acquiring a ground penetrating radar B-scan image of the rock mass; data sets of different types of fractures are constructed, and intelligent classification of fracture types is realized by improving convolutional neural network (CNN) model training; a second stage: intelligently extracting fracture parameters; and based on the identified fracture distribution form, feature points on the B-scan image are extracted through deep learning, and then the position coordinates of the fracture are automatically calculated through a parameter inversion formula. The method is high in detection precision, strong in anti-interference performance and wide in engineering applicability.

Description

Technical Field

[0001] The present invention relates to an intelligent method for identifying the distribution form of rock mass fractures based on GPR and parameter inversion, and is applicable to the field of rock mass fracture detection. Background Art

[0002] The fracture distribution inside the slope rock mass is crucial for its stability and safety. The existence of fractures will affect the integrity and mechanical properties of the rock mass, and it is easy to become a channel for rainwater infiltration in the case of rainfall, which is an important factor leading to collapses, slides, etc. of rock slopes. Therefore, accurately depicting the distribution, development and expansion law of rock mass fractures can provide a scientific basis for slope stability evaluation and prevention and control engineering design.

[0003] At present, common rock mass fracture detection methods include photogrammetry, drilling method, adit method, etc. These methods all have certain limitations. For example, photogrammetry can only obtain information on outcrop fractures and cannot directly obtain the parameters of internal rock mass fractures; the drilling method has a large workload, low efficiency, high risk, and it is difficult to obtain the fracture distribution of the entire slope; the adit method is extremely costly and can only count the fracture information in a certain area near the adit.

[0004] As a non-destructive detection technology, GPR can obtain image information of the internal structure of rock masses based on the principle of electromagnetic wave reflection. At present, the most common method for manually identifying the fracture distribution form and parameter inversion relies on the experience and professional knowledge of geological engineers. They identify the fracture distribution form and geometric parameters by observing the morphological characteristics of GPR images, such as the shape, orientation and amplitude of hyperbolas. However, manual identification has the following deficiencies: (1) Manual identification requires a large amount of time and effort, and the efficiency is much lower than that of intelligent identification methods; (2) The results of manual identification are limited by the experience and professional knowledge of geological engineers and are easily affected by subjective factors, resulting in inconsistent results; (3) The results of manual identification are difficult to standardize, have poor repeatability, and are not conducive to data sharing and analysis.

[0005] With the continuous development of machine learning and deep learning technologies, using algorithms such as convolutional neural networks to automatically extract feature images and feature points from GPR images can improve the identification efficiency and accuracy of rock mass fractures. However, in the prior art, although the classification method based on convolutional neural network (CNN) can achieve partial automation, it does not combine target detection technology, cannot simultaneously complete the determination of fracture types and the accurate extraction of multiple parameters, and lacks identification and inversion algorithms for complex fracture morphologies (such as V-shaped and orthogonal fractures), resulting in insufficient fracture classification and positioning accuracy. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: in view of the above problems, to provide an intelligent method for identifying the distribution form of rock mass fractures and inverse parameter inversion based on GPR.

[0007] The technical solution adopted by the present invention is: an intelligent method for identifying the distribution form of rock mass fractures and inverse parameter inversion based on GPR, including:

[0008] The first stage: intelligent classification of fracture types;

[0009] Obtain the ground penetrating radar (GPR) B-scan image of the rock mass, construct a data set of different types of fractures, and realize the intelligent classification of fracture types through training an improved convolutional neural network CNN model, and identify the fractures as one of nine distribution forms;

[0010] The nine types of fracture distribution forms include: single horizontal fracture, single vertical fracture, single inclined fracture, two horizontal fractures, two vertical fractures, two inclined fractures, orthogonal fractures, V-shaped fracture with upward opening, V-shaped fracture with downward opening;

[0011] The CNN model is constructed based on the improved YOLOv8 framework, trained with the simulation image data set generated by GPRMax, and data augmentation is performed using rotation, translation, shearing and mirroring;

[0012] Input the GPR B-scan image into the trained image recognition model to identify the fracture distribution form corresponding to the fractures in the B-scan image.

[0013] Specifically, it includes the following steps:

[0014] (1) Data set construction and augmentation:

[0015] Use the electromagnetic simulation software (GPRMax) to batch generate a B-scan image data set containing nine types of fractures; expand the sample size through data augmentation techniques such as rotation, translation, shearing and mirroring, and optimize the input data using image preprocessing techniques; divide the data set into a training set, a validation set and a test set according to a ratio.

[0016] (2) Improved CNN model training:

[0017] Construct a convolutional neural network based on the YOLOv8 framework and optimize the network structure as follows:

[0018] Backbone network: Adopt the CBS module, C2f module and SPPF module to enhance the multi-scale feature extraction ability;

[0019] Neck network: Combine the feature pyramid (FPN) and the path aggregation network (PAN) to fuse shallow details and deep semantic information;

[0020] Decoupled Detection Head: Separating classification and regression tasks to improve the model convergence speed and accuracy.

[0021] (3) Fracture type determination:

[0022] The input dataset is used to train the network model. During training, forward calculations are continuously performed and the network parameters are updated until the specified number of iterations is reached, at which point the network converges, and the recognition results of the fracture types are saved.

[0023] The second stage: Intelligent extraction of fracture parameters;

[0024] Based on the distribution form of the fractures, determine the corresponding geometric parameter types of this distribution form, and invert the geometric parameters of the corresponding fracture types in combination with the image features of the ground penetrating radar B-scan image. Specifically, it includes:

[0025] Step 1: Generate the target detection dataset;

[0026] Use gprMax to batch generate radar B-scan images as the target detection dataset, and divide it into a training set, a test set, and a validation set according to the ratio of 8:1:1;

[0027] Step 2: Label production and model training

[0028] Annotate the B-scan images of different fracture types: label two straight segments for parallel fractures, label the straight segment and the vertex of the hyperbola for orthogonal fractures, and label three vertices of the hyperbola for V-shaped fractures;

[0029] Save as a file in xml format;

[0030] Optimize the YOLOv8 model structure: Introduce the C2f module in the Backbone to enhance feature extraction, fuse the FPN+PAN structure in the Neck network to reduce false detections, and design an adaptive anchor box to match the geometric characteristics of the fractures;

[0031] Build the YOLOv8 model, including a feature extraction network, a region proposal network, and a target extraction network;

[0032] Step 3: Parameter inversion and coordinate transformation

[0033] Send the dataset into the network for training until the specified number of iterations;

[0034] After training, save the model parameters and generate the target detection result map;

[0035] Extract the vertex pixel coordinates (u, v) of the rectangle in the detection result map and save them in text format;

[0036] Convert the pixel coordinates (u, v) to image coordinates (n, t) using the conversion relationship between pixel coordinates and image coordinates;

[0037] Use the parameter inversion method to convert the image coordinates (n, t) to fracture parameter coordinates (x, y), and develop an automated script to batch process the detection results to achieve efficient parameter output.

[0038] The vertical coordinate parameter inversion method is as follows:

[0039]

[0040] where t is the time difference between the ground direct wave and the top of the hyperbolic echo, c is the propagation speed of electromagnetic waves in air (3×10 8 m / s), and ε r is the relative dielectric constant of the rock matrix.

[0041] The horizontal coordinate parameter inversion method is as follows:

[0042] x = n×△l + x i (2)

[0043] where n is the acquisition trace number corresponding to the image feature point, Δl is the trace interval (m), and x i is the initial position of the transmitting antenna (m).

[0044] Develop a Python script to batch output structured parameter reports (in.xlsx format) and labeled images.

[0045] In the label making:

[0046] For parallel fractures, label the upper straight section as "p" and the lower straight section as "q"; for orthogonal fractures, label the horizontal straight section as "o" and the vertical hyperbola vertex as "r"; for V-shaped fractures, label the left, right, and lower hyperbola vertices as "a", "b", and "c".

[0047] The loss function of the YOLOv8 model includes: the classification loss uses VFL Loss, and the regression loss uses DFLLoss + CIoU Loss; for the similarity of V-shaped fracture endpoints, add a class similarity penalty term with weight coefficients λ1 = 0.5, λ2 = 1.5, and λ3 = 2.0.

[0048] A device for inverting the internal fracture parameters of a rock mass based on a ground penetrating radar image, comprising:

[0049] Image acquisition module: Obtain GPR B-scan images;

[0050] Intelligent classification module: Identify the fracture distribution form through the YOLOv8 model;

[0051] Target detection module: Identify feature points of B-scan images through the YOLOv8 model;

[0052] Intelligent inversion module: Call the inversion formula according to the fracture type and output geometric parameters;

[0053] Result generation module: Automatically generate parameter reports and visualization images, and support manual review.

[0054] A storage medium, on which a computer program is stored, and when the program is executed, the steps of the method are implemented.

[0055] A data processing device, having a memory and a processor, and a computer program capable of being executed by the processor is stored on the memory, and when the computer program is executed, the steps of the method are implemented.

[0056] The detection of the present invention has high accuracy, strong anti-interference ability, and wide engineering applicability. Description of the Drawings

[0057] Fig. 1(a) is a flow chart for intelligent identification of fracture types;

[0058] Fig. 1(b) is a flow chart for intelligent identification of fracture parameters;

[0059] Figure 2 It is a structural diagram for improving the YOLOv8 network.

[0060] Fig. 3 is an example diagram of a fracture type data set;

[0061] Fig. 3(a) is an example diagram of a horizontal fracture type data set;

[0062] Fig. 3(b) is an example diagram of a vertical fracture type data set;

[0063] Fig. 3(c) is an example diagram of an inclined fracture type data set;

[0064] Fig. 3(d) is an example diagram of a pair of horizontal fracture type data sets;

[0065] Fig. 3(e) is an example diagram of a pair of vertical fracture type data sets;

[0066] Fig. 3(f) is an example diagram of a pair of inclined fracture type data sets;

[0067] Fig. 3(g) is an example diagram of an orthogonal fracture type data set;

[0068] Fig. 3(h) is an example diagram of an upward-opening V-shaped fracture type data set;

[0069] Fig. 3(i) is an example diagram of a downward-opening V-shaped fracture type data set;

[0070] Figure 4It is the classification effect diagram of fracture types;

[0071] Figure 5 It is the confusion matrix of fracture type classification;

[0072] Figure 6(a) is one of the schematic diagrams of target detection annotation;

[0073] Figure 6(b) is the second schematic diagram of target detection annotation;

[0074] Figure 6(c) is the third schematic diagram of target detection annotation;

[0075] Figure 7 It is the target detection effect diagram;

[0076] Figure 8 It is the target detection confusion matrix;

[0077] Figure 9(a) is the parameter inversion effect diagram of the x coordinate (m) of the left end point of the first fracture;

[0078] Figure 9(b) is the parameter inversion effect diagram of the y coordinate (m) of the first fracture;

[0079] Figure 9(c) is the parameter inversion effect diagram of the x coordinate (m) of the right end point of the first fracture;

[0080] Figure 9(d) is the parameter inversion effect diagram of the x coordinate (m) of the left end point of the second fracture;

[0081] Figure 9(e) is the parameter inversion effect diagram of the y coordinate (m) of the second fracture;

[0082] Figure 9(f) is the parameter inversion effect diagram of the x coordinate (m) of the right end point of the second fracture. Specific implementation manner

[0083] The specific technical solution of the present invention will be described with reference to the accompanying drawings.

[0084] An intelligent method for identifying the distribution form and parameter inversion of rock mass fractures based on GPR includes the following parts:

[0085] 1. Intelligent classification of fracture types. As shown in Figure 1(a), it includes the following steps:

[0086] (1) Dataset construction and enhancement:

[0087] Generate a dataset through the electromagnetic simulation software GPRMax. In the dataset of ground penetrating radar images, each fracture model is placed in a rectangular shale area with a total area of 1.6 m × 0.68 m, the minimum grid length of FDTD is 0.01 m, and the time window is 22 ns. The center frequency of the antenna is 800 MHz, and the number of scanning channels is 68. The relative dielectric constants of each medium in the model are: ε 页岩 = 8, ε空气 = 1. Nine types of samples were established, namely single horizontal crack, single vertical crack, single inclined crack, two horizontal cracks, two vertical cracks, two inclined cracks, orthogonal cracks, "V"-shaped crack with upward opening, and "V"-shaped crack with downward opening. Each type of crack contains 200 samples.

[0088] To address the problem of insufficient data volume, the data augmentation module (ImageDataGenerator) in Keras was used for data expansion. As a Python generator, this module retrieves data from the original dataset and performs augmentation during training to achieve dynamic data expansion during training. In data augmentation, the module parameters of ImageDataGenerator were set as follows: image rotation angle of 10°, horizontal image translation of 0.1, vertical image translation of 0.1, image shearing of 0.3°, image scaling of 0.2 times, and random horizontal mirroring of the image. In this way, each category was expanded to 400 samples, for a total of 3600 samples.

[0089] The samples were divided into a training set, a test set, and a validation set in the ratio of 8:1:1. The B-scan images of different types of cracks in the dataset are as Figures 3(a) to 3(i) shown.

[0090] (2) Improvement of CNN model training:

[0091] Programming was carried out based on the Python language, with Pycharm as the programming platform. A convolutional neural network based on the YOLOv8 framework was constructed to optimize the network structure. The experimental system selected the Windows operating system and used an NVIDIA graphics card to accelerate the tuning of the model. For each round of tuning, some adjustments were made to the Ultraft parameters.

[0092] (3) Crack type determination:

[0093] The recognition effect of crack types is as Figure 4 shown. The meanings of the numbers are as follows: 0 - a pair of "V"-shaped cracks with upward opening; 1 - a pair of "V"-shaped cracks with downward opening; 2 - a pair of parallel horizontal cracks; 3 - a pair of parallel vertical cracks; 4 - a pair of parallel inclined cracks; 5 - a pair of orthogonal cracks; 6 - a single horizontal crack; 7 - a single inclined crack; 8 - a single vertical crack. The normalized confusion matrix (Confusion matrix) of the recognition effect is as Figure 5 shown. It can be seen from the figure that most of the values are concentrated on the main diagonal and the values are 1.00 or close to 1.00, which means that the model has achieved a high accuracy rate for most categories, with only a very small number of classification errors.

[0094] 2. Intelligent identification of fracture parameters, as shown in Fig. 1(b), includes the following steps:

[0095] (1) Dataset construction;

[0096] The radar image dataset is simulated by gprMax software and batch-generated through self-developed python scripts. The total size of the concrete area of each model in this dataset is 1.6m × 0.68m, the minimum grid length is 0.01m, the time window is 22ns, the center frequency of the antenna is 800MHz, and the number of scan channels is 68. A large number of B-scan images generated by three types of fracture forms, namely a pair of horizontal fractures, orthogonal fractures, and V-shaped fractures with the opening upward, are selected as the dataset to implement the intelligent method of parameter inversion.

[0097] For each of the three fracture forms, namely a pair of horizontal fractures, a pair of orthogonal fractures, and a pair of V-shaped fractures with the opening upward, 250 original txt files are made, and then simulation images are batch-generated through python scripts, with a total of 750 images as the dataset. The dataset is divided into a training set, a test set, and a validation set according to 8:1:1, including 600 images in the training set, 75 images in the test set, and 75 images in the validation set. Among them, the training set, the test set, and the validation set respectively contain images generated by the above three types of fractures.

[0098] (2) Label production;

[0099] For each target area contained in each image in the training set, border calibration and class identification are carried out, and there are multiple labels in one image. For the radar image of a pair of horizontal fractures, the upper straight section is marked as p, and the lower straight section is marked as q; for a pair of orthogonal fractures, the straight section is marked as o, and the top of the hyperbola is marked as r; for a pair of V-shaped fractures with the opening upward, the top of the left hyperbola is marked as a, the top of the right hyperbola is marked as b, and the top of the lower hyperbola is marked as c. The specific process is as follows: Enter labelimg in the Anaconda terminal to enter the image annotation interface, set the annotation format to Pascal VOC format, click the "Open Dir" button, select the folder of the training set images to be annotated, after selecting an image, click the "Create RectBox" button in the upper left corner of the interface, draw a rectangular box to enclose the target area, and set a label for this target area, such as Figures 6(a) to 6(c) . After the annotation is completed, click the "Save" button to save the annotation data. LabelImg will generate the corresponding xml file under the specified save path, which contains the annotation object information (category, coordinates, etc.).

[0100] (3) Improved YOLOv8 model training;

[0101] Model optimization:

[0102] Feature extraction module: Introduce the C2f module into the Backbone, and enhance the ability to capture deep crack features through multi-level Bottleneck stacking;

[0103] Anti-interference design: Integrate the FPN+PAN structure in the Neck network, strengthen the interaction between shallow edge information and deep semantic features, and reduce false detections caused by multiple reflections of electromagnetic waves;

[0104] Adaptive anchor box: Dynamically adjust the anchor box size according to the distribution law of crack feature points, and give priority to matching the geometric characteristics of the straight section and the top of the hyperbola. The improved YOLOv8 model structure is as Figure 2 shown.

[0105] Loss function design:

[0106] The classification loss uses VFL Loss, and the regression loss uses DFL Loss+CIoU Loss. The multi-task learning is balanced through the weight coefficients (λ1 = 0.5, λ2 = 1.5, λ3 = 2.0); For the problem of the similarity of the endpoints of V-shaped cracks, a category similarity penalty term is added to reduce the confusion probability between a and b labels.

[0107] The effect of crack target detection based on YOLOv8 through B-scan images is as Figure 7 shown. Figure 8 The normalized confusion matrix of the recognition results of this model is shown. It can be found that the recognition effect of this model is relatively good, especially for the 'o' and 'r' categories, which are classified correctly.

[0108] (4) Pixel coordinate to parameter coordinate conversion;

[0109] Save the pixel coordinates of the target areas detected in each image in the test set and the validation set as.txt files and store them in the specified folder. Each.txt file contains the pixel coordinates of the four vertices of each target rectangle in the detected image. Among them, the pixel coordinates we need are:

[0110] For horizontal cracks, extract the vertex coordinates of the rectangles corresponding to the straight sections of the two bowl-shaped hyperbolas. Specifically, the abscissa of the upper left vertex of the rectangle corresponding to the straight section of the upper bowl-shaped hyperbola is denoted as u1, the abscissa of the upper right vertex is denoted as u2, and the ordinate is denoted as v1; The abscissa of the upper left vertex of the rectangle corresponding to the straight section of the lower bowl-shaped hyperbola is denoted as u3, the abscissa of the upper right vertex is denoted as u4, and the ordinate is denoted as v2.

[0111] For orthogonal fractures, extract the coordinates of the midpoint of the upper side of the rectangular frame corresponding to the top of the high and narrow hyperbola, as well as the coordinates of the upper left and upper right corners of the rectangular frame corresponding to the straight section of the bowl-shaped hyperbola. The abscissa of the midpoint of the upper side of the rectangular frame is the average of the abscissas of the left and right vertices of the upper side, and the ordinate is the abscissa of the upper left vertex. Among them, the abscissa of the midpoint of the upper side of the rectangular frame corresponding to the top of the high and narrow hyperbola is denoted as u5, and the ordinate is denoted as v3; the abscissa of the upper left corner of the rectangular frame corresponding to the straight section of the bowl-shaped hyperbola is denoted as u6, the abscissa of the upper right corner is denoted as u7, and the ordinate is denoted as v4.

[0112] V-shaped fractures with an upward opening: Extract the coordinates of the midpoints of the upper sides of the rectangular frames corresponding to the tops of the three hyperbolas. The abscissa is the average of the abscissas of the left and right vertices of the upper side, and the ordinate is the abscissa of the upper left vertex. Among them, the coordinates of the midpoint of the upper side of the rectangular frame corresponding to the top of the upper left hyperbola are denoted as (u8, v5), the coordinates of the midpoint of the upper side of the rectangular frame corresponding to the top of the upper right hyperbola are denoted as (u9, v6), and the coordinates of the midpoint of the upper side of the rectangular frame corresponding to the top of the lower hyperbola are denoted as (u 10 , v7).

[0113] Subsequently, based on the pre-established mapping relationship, convert these pixel coordinates we need into image coordinates, and further combine with the fracture parameter inversion method to convert the image coordinates into actual fracture position coordinates. The mapping relationship between pixel coordinates (u, v) and image coordinates (n, t) is:

[0114]

[0115] The relational expression for inversely calculating the actual position coordinates (x, y) of the fracture using image coordinates (n, t) is:

[0116] y = 0.001296t - 0.0648 (5)

[0117] x = 0.02n + 0.1 (6)

[0118] Substitute Equation (3) into Equation (5) and Equation (4) into Equation (6), and the actual position coordinates can be directly obtained from the pixel coordinates.

[0119] (5) Batch processing;

[0120] Develop a Python program to batch parse YOLOv8 output files (in.txt format), automatically perform coordinate conversion and parameter calculation, and generate a structured report (in.xlsx format); output B-scan images with annotation frames and parameter distribution maps, supporting manual review and model iteration optimization. Conduct a quantitative comparative analysis of the intelligent recognition results (position parameters) and the original true position parameters for three types of test sets of horizontal fractures, orthogonal fractures, and V-shaped fractures (50 images for each type) and 50 images randomly selected from the training set, and calculate the correlation coefficient. The results of intelligent recognition and the comparison effects are asFigures 9(a) to 9(f) As shown. The correlation coefficient R between the recognition results of each parameter and the true value exceeds 0.7, and most of the correlation coefficients are higher than 0.8, indicating that there is a significant statistical correlation between the intelligent recognition results and the true value.

[0121] Technical effects:

[0122] Detection accuracy: The mAP@0.5 of target detection reaches 97.8%;

[0123] Anti-interference ability: Through FPN+PAN fusion and post-processing algorithm, the false detection rate is reduced to less than 3%;

[0124] Engineering applicability: The processing time of a single image is <0.5s, supporting real-time on-site survey data analysis.

[0125] Embodiment 2: This embodiment is a storage medium on which a computer program executable by a processor is stored. When the computer program is executed, the steps of the method in Embodiment 1 are implemented.

[0126] Embodiment 3: A data processing device has a memory and a processor. A computer program executable by the processor is stored on the memory. When the computer program is executed, the steps of the method in Embodiment 1 are implemented.

Claims

1. An intelligent method for identifying the distribution form of rock mass fractures and inverse parameter inversion based on GPR, characterized in that Including: The first stage: intelligent classification of fracture types; Obtain the ground penetrating radar B-scan image of the rock mass; Construct a dataset of different types of fractures, and through training by improving the convolutional neural network CNN model, realize the intelligent classification of fracture types; The second stage: intelligent extraction of fracture parameters; Based on the identified fracture distribution form, extract the feature points on the B-scan image through deep learning, and then automatically calculate the position coordinates of the fractures through the parameter inversion formula.

2. The intelligent method for identifying the distribution form of rock mass fissures and inverse parameter inversion based on GPR according to claim 1, characterized in that, In the first stage, the fractures are identified into nine distribution forms, and the nine fracture distribution forms include: single horizontal fracture, single vertical fracture, single inclined fracture, two horizontal fractures, two vertical fractures, two inclined fractures, orthogonal fractures, V-shaped fractures with upward opening, and V-shaped fractures with downward opening.

3. The intelligent method for identifying the distribution form of rock mass fissures and inverse parameter inversion based on GPR according to claim 1, characterized in that, The first stage specifically includes the following steps: (1) Dataset construction and enhancement: Use electromagnetic simulation software to batch generate a B-scan image dataset containing nine types of fractures; expand the sample size through data enhancement technology, and use image preprocessing technology to optimize the input data; divide the dataset into training set, validation set and test set according to a certain proportion; (2) Improved CNN model training: Construct a convolutional neural network based on the YOLOv8 framework, and optimize the network structure as follows: Backbone network: Adopt CBS module, C2f module and SPPF module to enhance the multi-scale feature extraction ability; Neck network: Combine the Feature Pyramid Network FPN and the Path Aggregation Network PAN to fuse shallow details and deep semantic information; Decoupled detection head: Separate the classification and regression tasks to improve the model convergence speed and accuracy; (3) Fracture type determination: Input the dataset to train the network model. During the training, continuously perform forward calculation and update the network parameters until the specified number of iterations is reached and the network converges, and save the recognition results of the fracture types.

4. The intelligent method for identifying the distribution form of rock mass fissures and inverse parameter inversion based on GPR according to claim 1, characterized in that The second stage includes the following steps: Step 1: Generate the target detection dataset; Use GPRMax simulation software to batch generate B-scan images, and divide them into training set, test set and validation set according to 8:1:1; Step 2: Label production and model training Annotate the B-scan images of different fracture types: mark two straight segments for parallel fractures, mark the straight segment and the vertex of the hyperbola for orthogonal fractures, and mark three vertices of the hyperbola for V-shaped fractures; Optimize the YOLOv8 model structure: Introduce the C2f module in the Backbone to enhance feature extraction, fuse the FPN+PAN structure in the Neck network to reduce false detections, and design an adaptive anchor box to match the geometric characteristics of the fractures; Step 3: Parameter inversion and coordinate transformation Extract the pixel coordinates (u, v) in the detection results, convert them into image coordinates (n, t) through the mapping relationship, and then call the following inversion formula to calculate the actual position coordinates (x, y): x = n×△l + x i Develop a Python script to batch output structured parameter reports and annotated images; t is the time difference between the direct ground wave and the top of the hyperbolic echo, c is the propagation speed of electromagnetic waves in air, ε r is the relative dielectric constant of the rock matrix, n is the acquisition trace number corresponding to the image feature point, Δl is the trace interval, x i is the initial position of the transmitting antenna.

5. The intelligent method for identifying the distribution form of rock mass fissures and inverse parameter inversion based on GPR according to claim 4, characterized in that, In Step 2 of label production: for parallel fractures, the straight section above the parallel fracture annotation is labeled as "p", and the straight section below is labeled as "q"; for orthogonal fractures, the horizontal straight section is labeled as "o", and the vertex of the vertical hyperbola is labeled as "r"; for V-shaped fractures, the vertices of the left, right, and lower hyperbolas are labeled as "a", "b", and "c".

6. The intelligent method for identifying the distribution form of rock mass fissures and inverse parameter inversion based on GPR according to claim 4, characterized in that In Step 2, the loss function of the YOLOv8 model includes: the classification loss uses VFL Loss, and the regression loss uses DFLLoss + CIoU Loss; for the similarity of the endpoints of V-shaped fractures, a category similarity penalty term is added, and the weight coefficients are λ1 = 0.5, λ2 = 1.5, and λ3 = 2.

0.

7. Apparatus for inverting internal fracture parameters of rock mass based on ground penetrating radar images, characterized in that, Including: Image acquisition module: Obtain GPR B-scan images; Intelligent classification module: Identify the fracture distribution pattern through the YOLOv8 model; Target detection module: Identify the feature points of the B-scan image through the YOLOv8 model; Intelligent inversion module: Call the inversion formula according to the fracture type and output geometric parameters; Result generation module: Automatically generate parameter reports and visualization images, and support manual review.

8. A storage medium, on which a computer program is stored, characterized in that, When the program is executed, it implements the steps of the method described in any one of claims 1-6.

9. A data processing device, having a memory and a processor, with a computer program stored on the memory and executable by the processor, characterized in that: When the computer program is executed, it implements the steps of the method described in any one of claims 1-7.

Images