Drilling image intelligent pickup method and system based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence and geophysical prospecting, and discloses a drilling image intelligent pickup method and system based on deep learning. The method comprises the following steps: firstly, making a data set of a drilling image by using drilling data collected by a drilling probe, and automatically picking up different types of structural surfaces such as rock vein filling and cracks on a hole wall image by using a multi-classification task constructed by an instance segmentation module in a deep learning network model YOLOv8; according to a polynomial fitting formula, mask coordinates obtained after instance segmentation are subjected to fitting to obtain a sine curve of a structural plane, extreme points of the curve after fitting are calculated through mathematical transformation, spatial transformation and other modes, corresponding tendencies, dip angles and other occurrence of the structural plane are obtained, and finally a drilled hole image obtained after automatic analysis and drawing of an algorithm model is output.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of artificial intelligence and geophysical exploration, and in particular to a deep learning-based borehole image intelligent picking method and system thereof. Background Art

[0002] The rock mass structural surface is an important part of the rock mass. Its distribution and combination characteristics significantly affect the strength, deformation and seepage characteristics of the rock mass. It often plays an important role in geological exploration, engineering and resource development. In actual engineering, borehole camera technology is often used to analyze the information of the borehole wall after collecting the information of the borehole wall and the information on the digital core image to obtain the interpretation of the cracks or discontinuities. These interpretations include the buried depth, inclination, dip, width, roughness of the crack surface and the properties of the filling. The distribution of cracks obtained from the borehole image can provide intuitive data for data processing personnel to evaluate indicators such as stratum characteristics, rock types and rock stability. However, with the continuous development of borehole television technology and the deepening of drilling depth, more and more image data need to be processed by borehole television. At the same time, since most of the cracks and discontinuities on the borehole wall images are identified based on the experience of data processing personnel, the efficiency and accuracy of manual picking of borehole wall image data are particularly prominent. How to realize automatic identification of structural surfaces in drilling images through various algorithms has gradually become a research hotspot to solve this problem.

[0003] The automatic detection method of borehole images initially used a series of feature extraction algorithms based on Hough transform, which can automatically identify and extract various line features in the borehole wall image data, and use polynomial fitting and other methods to obtain accurate structural surface information. Based on this idea, many researchers have used edge detection algorithms such as Canny operator or Sobel operator to further extract structural surface information in borehole images, and improved the corresponding algorithms to make them more in line with the requirements of feature curve extraction of borehole images. However, the edge detection algorithm has problems such as high noise and unclear edge features when extracting edges from borehole images.

[0004] With the rapid development of deep learning in the field of image processing, more and more scholars have chosen to use algorithms such as CNN (Convolutional Neural Networks) to perform operations such as target detection and image segmentation on borehole images to analyze and calculate the relevant parameters of the rock structure surface in the image. However, there is a more serious problem in using the CNN model for target detection on borehole images, that is, the target detection algorithm can only identify the location of the structure surface, but cannot identify the specific shape of the structure surface like the edge detection algorithm. If the specific shape of the structure surface cannot be identified, then the parameters such as the inclination and dip angle of the structure surface cannot be calculated. However, the image segmentation task based on CNN can well identify the specific location and shape of the structure surface, thus providing a basis for the next step of calculating the occurrence of the structure surface.

[0005] However, there is a problem in the application of both edge detection algorithms and target detection algorithms on borehole images: there is no classification of structural surfaces. Because in the actual data processing process, data processing personnel must not only identify the location and occurrence of the structural surface, but also distinguish whether the structural surface here is a joint, a crack or a vein filling, and whether the structural surface shape is open or closed. Because in most cases, closed joints and vein fillings will not have much impact on the properties of the rock mass at this location, and data processing personnel will not pick up these stable structural surfaces when picking up images. However, the edge detection algorithm and image segmentation algorithm will identify all structural surfaces that need to be picked up and those that do not need to be picked up. The picking results obtained in this way will cause misjudgment in the design and construction of subsequent projects.

[0006] The deep learning method uses artificial neural networks to form an instance segmentation network, and realizes the automatic picking process of borehole images by extracting structural surface features such as vein filling and cracks on the borehole wall image. Compared with traditional manual picking, the automatic borehole image picking algorithm based on deep learning completely breaks away from the manual participation process in the stage of processing borehole exploration data, has a high degree of automation, and has greatly improved the picking accuracy. Secondly, the automatic borehole image picking algorithm based on deep learning can adapt to work area scenes of different complexities after being trained with a large number of borehole image data sets, and has strong generalization ability and robustness. Summary of the invention

[0007] The purpose of the present invention is to provide a method and system for intelligent picking of borehole images based on deep learning. The YOLOv8_seg instance segmentation model is used to perform multi-classification picking, and feature extraction and mask coordinate acquisition are performed for structural surfaces such as vein filling and fractures in the borehole wall image. Subsequently, the coordinates are input into a polynomial fitting to calculate the curve and extreme points of the structural surface. Finally, based on the fitting results, the inclination, dip angle and other attributes of the structural surface are automatically calculated, and the analyzed and drawn borehole image is output.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] A drilling image intelligent picking method based on deep learning, comprising the following steps:

[0010] Step 1: Collect drilling data to obtain a dataset of drilling images, and divide the collected dataset into a training set, a validation set, and a test set;

[0011] Step 2: Through the multi-classification picking task constructed by the instance segmentation model YOLOv8_seg, feature extraction is performed on different types of structural surfaces such as vein fillings and fractures on the hole wall image, and the corresponding mask coordinates of each structural surface are obtained as the basis for the next calculation of the algorithm model;

[0012] Step 3: Input the mask coordinates obtained in step 2 into the polynomial fitting formula, and calculate the structural curve of the structural surface in the image according to the mask coordinates obtained by the model, so as to obtain the extreme value points of the structural surface fitting curve;

[0013] Step 4: Then, based on the extreme value of the curve fitted in step 3 and the corresponding parameters, the inclination and dip angle of the structural surface are calculated, and finally the drilling image automatically analyzed and drawn by the algorithm model is output.

[0014] As a further solution of the present invention: in step 1, the data set is divided into a training set, a validation set and a test set in the form of 8:1:1.

[0015] As a further solution of the present invention: in step 2, the instance segmentation model YOLOv8_seg picks up unconformities of different shapes in the hole wall image;

[0016] The labels for the structural surface include Close, Open, and Dike Filling.

[0017] As a further solution of the present invention: the picking rule of Close in the structural surface picking is the structural surface area in the image that is in a closed state;

[0018] The picking rule of Open in the structural surface picking is the structural surface area in the open state;

[0019] The picking rule of Dike Filling in structural surface picking is to find structural surface areas with calcite or mica that are obviously different in color from the rock mass.

[0020] As a further solution of the present invention: in step three, when the drilling image automatic picking algorithm model BIAR is training the data set, the input data set format is COCO format, the backbone network uses the backbone network of YOLOv8, the optimizer selects AdamW, the learning rate is designed to be 1e-4, and the GPU is used to accelerate the training process, and no frozen training is performed.

[0021] As a further solution of the present invention: the process of instance segmentation by the instance segmentation model YOLOv8_seg is:

[0022] The input module preprocesses the input image (drilling image), including scaling and normalization, to ensure that the image fits the input size and format of the model. The target features are extracted from the input image through Backbone, thereby converting the image into a feature representation with rich semantic information. Then, the FPN (Feature Pyramid Networks) module is used to perform multi-scale feature fusion and construct a multi-scale feature map to enhance the model's detection ability for targets of different sizes. The PANet (Path Aggregation Network) is used to further integrate feature information at different levels to improve the expressiveness of the feature map. Next, the Detection Head is responsible for target classification and bounding box regression, generating category predictions and bounding box regression values ​​(Bbox) for each candidate region. After detecting the target, the Instance Segmentation Head performs pixel-level segmentation on the image and outputs the precise outline of each target. During the training of the model, the Loss Function module calculates the difference between the model output and the true annotation, and adjusts the model parameters through back propagation to minimize the loss value. The post-processing module removes redundant overlapping boxes and filters out overlapping detection boxes through NMS (Non-Maximum Suppression) to obtain the precise target location. Finally, the detection and segmentation results of the algorithm are output, including the target category, confidence, bounding box coordinates, and segmentation mask.

[0023] As a further solution of the present invention: the structural surface mask obtained in the previous image segmentation step is fitted by polynomial fitting to obtain the corresponding curve parameters. Before fitting the data using the polynomial function, it is first necessary to define the polynomial function f(x)=a according to the fitting requirements. n x n +a n-1 x n-1+...+a1x+a0. Then, we use the least squares method and other optimization algorithms to minimize the residual sum of squares formula. (i.e., the sum of the squares of the differences between the actual observed values ​​and the predicted values ​​of the fitted curve) to solve the polynomial coefficients. Finally, once the polynomial coefficients are determined, the fitted polynomial function can be used to perform further data analysis on the mask coordinates;

[0024] Among them, a in the formula n , a n-1 ,…, a1, a0 are the coefficients to be calculated, n is the degree of the polynomial, y i is the i-th observed value of the dependent variable, that is, the value actually recorded in the experiment or in the data, f(x i ) is based on the model and the independent variable x i The predicted value is calculated from the value of .

[0025] As a further solution of the present invention: the extreme points A(MinX, MinY) and B(MaxX, MaxY) of the structural surface calculated in the fitting formula are connected by the formula The coordinates on the y-axis are interpolated to calculate the absolute value of the difference between the extreme point coordinates X val and Y val , and convert the difference into Calculate the inclination angle of each structural surface.

[0026] Among them, Y in the formula cal is the actual diameter of the drilled hole, C degrees is the calculated inclination angle.

[0027] As a further solution of the present invention: the azimuth corresponding to the lowest point of the structural surface curve on the y-axis is mapped between 0° and 360°. i Between 0 and x mid (the minimum value of the drilling image on the x-axis), its inclination can be averagely mapped to between 180° and 360°; if the coordinate x of the maximum point is i In x mid to x max (the maximum value of the drilling image on the x-axis), then its inclination can be averagely mapped to between 0° and 180°. The mapping rule formula is as follows:

[0028]

[0029] Among them, α in the formula Angle It is the structural surface inclination obtained by piecewise function calculation.

[0030] A drilling image intelligent picking system based on deep learning, characterized by comprising:

[0031] A data acquisition module, wherein the data acquisition module is used to acquire drilling data to obtain a data set of drilling images, and divide the acquired data set into a training set, a validation set, and a test set;

[0032] An instance segmentation module, which extracts features of different types of structural surfaces such as vein fillings and fractures in the borehole image through a multi-classification picking task constructed by the instance segmentation model YOLOv8_seg, and obtains the corresponding mask coordinates of each structural surface as the basis for the next step of calculation of the algorithm model;

[0033] The occurrence calculation module is used to input the initially obtained structural surface mask coordinates into a polynomial function to fit a structural surface curve that is similar to a sine function, and calculate the occurrence of the structural surface such as the inclination and dip angle through the extreme value of the fitted structural surface curve.

[0034] The borehole image automatic picking module uses a trained model to automatically and quickly pick up structural surfaces such as cracks and discontinuities in the interference of complex information on the borehole wall image and calculate their occurrence, and finally outputs the borehole image automatically picked up by the model.

[0035] Beneficial effects of the invention: The invention not only completes the complete process from image recognition to parameter calculation, but also realizes intuitive visualization of the results (automatically marking and drawing the calculated occurrence information on the borehole image). At the same time, it provides geological engineers with an intuitive and convenient analysis tool, which can greatly improve the efficiency and accuracy of geological exploration and engineering evaluation, and can provide new ideas and solutions for the development of the field of borehole image analysis in the direction of intelligence and automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will be further described below in conjunction with the accompanying drawings.

[0037] Figure 1 The overall structural framework of the BIAR algorithm model is shown;

[0038] Figure 2 This is the YOLOv8_seg model structure diagram;

[0039] Figure 3 is the polynomial fitting result;

[0040] Figure 4 is the structural diagram of the inclination angle of the structural surface in space;

[0041] Figure 5 It is the structural diagram of the inclination of the structural surface in space;

[0042] Figure 6Automatically pick up results for the BIAR algorithm model. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Example 1

[0045] See also Figure 1 As shown, the present invention is a method and system for intelligently picking up drilling images based on deep learning, comprising the following steps:

[0046] Step 1: Build a dataset

[0047] To verify the model's ability to automatically pick up various complex structural surfaces in drilling images and to improve the accuracy and generalization of the model as much as possible during training;

[0048] In this implementation, a total of 6 well data located in different locations in a certain area in southern China were selected, and these wells have a total of more than 1,000 meters of borehole image data;

[0049] The strata in the area where the borehole data is located are mostly composed of dense igneous rocks and metamorphic rocks, and the structural surfaces in the borehole images are relatively clear and regular;

[0050] The borehole data was processed according to the rule of one image for every 50 cm of actual depth, and a data set containing 2,000 borehole images was obtained.

[0051] The obtained borehole image dataset was divided into training set, validation set and test set in the form of 8:1:1, and the velocity spectrum images were labeled by Labelimg software;

[0052] Step 2: Instance Segmentation

[0053] During instance segmentation, structural surface picking includes but is not limited to Close, Open, and Dike Filling labels, which transforms a single classification task into a multi-classification task;

[0054] Among them, the picking rule of Close in the structural surface picking is the structural surface area in the image that is in a closed state;

[0055] The picking rule of Open in the structural surface picking is the structural surface area in the open state;

[0056] The picking rule of Dike Filling in structural surface picking is to find structural surface areas with calcite or mica that are obviously different in color from the rock mass.

[0057] Step 3: Polynomial fitting

[0058] In order to accurately depict the structural surface mask identified in the previous image segmentation step, a polynomial fitting method is used. This process starts with setting the basic form of the polynomial function according to the specific requirements of the fitting.

[0059] Then, optimization techniques such as the least squares method are used to minimize the sum of squares of the deviations between the actual observations and the polynomial predictions by adjusting the coefficients of the polynomial, that is, to find the best fitting curve.

[0060] When these polynomial coefficients are accurately solved, the optimized polynomial function can be used to perform more in-depth data analysis on the mask coordinates, thereby achieving an accurate fit to the structural surface characteristics.

[0061] Step 4: Automatically pick up drilling images

[0062] After training, the borehole image automatic picking algorithm model BIAR can recognize all kinds of complex structural surface information. When the borehole image to be picked is input, the borehole image automatic picking algorithm model BIAR will automatically pick it up and calculate the occurrence of the corresponding structural surface.

[0063] The borehole image automatic picking algorithm model BIAR constructed in this embodiment can accurately identify a large number of borehole images in a short time and output instance segmentation images. It can accurately identify the relevant information of structural surfaces such as vein filling and fractures on the borehole image, and can accurately and quickly calculate the corresponding occurrence information of each structural surface. Compared with the traditional manual picking method, it greatly improves the efficiency and picking accuracy of velocity spectrum picking, and does not require manual participation.

[0064] Example 2

[0065] like Figure 2As shown in the figure, the instance segmentation algorithm model YOLOv8_seg can complete the instance segmentation task of the drilling image through a series of modules, and finally output the structural surface picking result. First, the input module preprocesses the input image (drilling image), including scaling and normalization, to ensure that the image fits the input size and format of the model. Then, Backbone extracts the target features from the input image and converts the image into a feature representation with rich semantic information. Then, FPN (FeaturePyramid Networks) constructs a multi-scale feature map through multi-scale feature fusion to enhance the model's detection ability for targets of different sizes. PANet (Path Aggregation Network) further integrates feature information at different levels to improve the expression ability of the feature map. Next, Detection Head is responsible for target classification and bounding box regression, generating category predictions and bounding box regression values ​​(Bbox) for each candidate region. After detecting the target, Instance Segmentation Head will segment the image at the pixel level and output the precise outline of each target. During the training process of the model, the LossFunction module calculates the difference between the model output and the true annotation, and adjusts the model parameters through back propagation to minimize the loss value. The post-processing module removes redundant overlapping frames through NMS (Non-Maximum Suppression) and filters out overlapping detection frames to obtain accurate target locations. Finally, the detection and segmentation results of the algorithm are output, including the target category, confidence, bounding box coordinates, and segmentation mask. Through the collaborative work of these modules, YOLOv8 can efficiently and accurately perform instance segmentation tasks of structural surfaces in drilling images.

[0066] Example 3

[0067] The overall process of polynomial fitting is as follows: first, define the form of the polynomial function according to the fitting requirements, then use the least squares method and other optimization algorithms to solve the polynomial coefficients by minimizing the residual sum of squares formula (that is, the sum of the squares of the difference between the actual observed value and the predicted value of the fitted curve). Finally, when the polynomial coefficients are determined, the fitted polynomial function can be used to perform further data analysis on the mask coordinates.

[0068] like Figure 3 As shown in a, the image shows a drill hole with an actual length of 80 cm and a diameter of 15 cm. The blue area represents the mask area obtained in the previous image segmentation process, which is surrounded by hundreds of mask coordinates; Figure 3b is an image of these hundreds of mask coordinates plotted in a rectangular coordinate system. The coordinates of the enlarged points in the figure are the coordinate points of the mask in the coordinate system. The coordinate system is established based on the upper left corner of each drilling image as the coordinate origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis. At the same time, the coordinates of each picked point are set as the coordinates of the pixel point on the image (without units). Figure 3 c is the comparison between the fitting curve after fitting by the polynomial fitting formula and the structural surface in the original image. It can be seen that the curve retains the sinusoidal characteristics of the rock mass structural surface as much as possible.

[0069] Specific:

[0070] f(x)=a n x n +a n-1 x n-1 +...+a1x+a0

[0071]

[0072] Among them, a n , a n-1 ,…,a1,a0 are the coefficients to be found, and n is the degree of the polynomial. i is the i-th observed value of the dependent variable, that is, the value actually recorded in the experiment or in the data, f(x i ) is based on the model and the independent variable x i The predicted value is calculated from the value of .

[0073] Example 4

[0074] In order to further calculate the attitude of the structural surface in the borehole image, the intersection line of the rock layer and any imaginary horizontal plane is called the strike line, that is, a straight line connecting two equal height points on the same layer. The directions indicated by its two ends are the strike of the rock layer, which can be represented by two azimuths 180° apart, which indicate the horizontal extension direction of the rock layer in space; the straight line on the rock layer that is perpendicular to the strike line and extends downward along the slope is called the dip line, and the direction indicated by the projection of the dip line on the horizontal plane is called the dip of the rock layer, which indicates the direction in which the rock layer tilts; the angle between the dip line and its projection on the horizontal plane is called the dip angle, which indicates the intersection angle between the rock layer and the horizontal plane, and reflects the degree of inclination of the rock layer.

[0075] like Figure 4 As shown in a, point A is the highest point where the structural surface intersects the rock core, point B is the lowest point where the structural surface intersects the rock core, point C is the intersection of point A extending downward and the horizontal plane where point B is located, straight line AB is the inclined line, the direction indicated by straight line CB is the dip of the rock formation, and the angle α between straight line AB and straight line CB is the dip angle of the rock formation.

[0076] At the same time, after calculation of the fitting formula in the previous step, the coordinate values ​​"MaxX", "MinX", "MaxY" and "MinY" of the extreme points of the fitting curve of each structural surface in the coordinate system can be obtained (the coordinate value is the position of the pixel point where the extreme point is located in the coordinate system).

[0077] When two right-angled sides are known, the angle between a right-angled side and the hypotenuse can be calculated by the inverse tangent function. Therefore, to know the inclination angle α of the structural surface, you only need to know the distance AC between the highest point A where the structural surface intersects with the core and the intersection point C on the horizontal plane where the lowest point B where the structural surface intersects with the core is located, and the diameter BC of the borehole. Figure 4 As shown in b, because the diameter BC of the borehole is known in actual work, if you want to calculate the inclination of the structural surface through the coordinates of the extreme points in the borehole image, you only need to know the coordinate difference between the extreme points A (MinX, MinY) and B (MaxX, MaxY) of the structural surface on the y-axis, and convert the difference according to the standard of 1500 pixels per meter in the borehole image to calculate the actual length of the AC segment in the core, and then you can calculate the inclination of the structural surface based on the inverse tangent function.

[0078] like Figure 5 As shown in a, if the x-coordinate of the maximum point A is XEast (the borehole image is at one quarter of the x-coordinate axis), then its corresponding orientation is due east. At the same time, according to the definition of the two-dimensional unfolded diagram of the structural surface as a sine function, the orientation corresponding to the minimum point B is due west, so the inclination of the inclination line of the structural surface is 270°; Figure 5 As shown in Figure 2, if the coordinate of the maximum point A in the x direction is XWest (the borehole image is at three-quarters of the x coordinate axis), then its corresponding orientation is due west, and the orientation of the minimum point B is due east, so the inclination line of the structural surface indicates an inclination of 90°. Figure 5 As shown in Figure c, if the coordinate of the maximum point A in the x direction is XNE (the eighth of the borehole image on the x coordinate axis), then its corresponding orientation is northeast, and its corresponding orientation of the minimum point B is southwest, so the inclination line of the structural surface indicates an inclination of 225°. Similarly, if the coordinate xi of the maximum point is between 0 and xmid, then its inclination can be averagely mapped to between 180°-360°; if the coordinate xi of the maximum point is between xmid and xmax (the maximum value of the borehole image on the x coordinate axis), then its inclination can be averagely mapped to between 0°-180°.

[0079] like Figure 6 The drilling image automatic picking result diagram shown. Figure 6 a is two overlapping and intersecting structural planes. It can be seen that even if the two structural planes are mixed together, the algorithm can still distinguish the differences between different structural planes well; Figure 6 b and c are the picking results of structural surfaces with different opening degrees; Figure 6 d is two structural surfaces of different categories that are very close to each other; Figure 6 e is a number of structural surfaces of the same category that are very close to each other; Figure 6 f is a structural surface with irregular structural surface curve. Figure 6 It can be observed that the algorithm model BIAR is capable of automatically and accurately picking up structural surfaces and calculating their occurrence in various complex drilling images.

[0080] Specific:

[0081]

[0082] Among them, Y cal is the actual diameter of the drilled hole, C degrees is the calculated inclination angle, α Angle It is the structural surface inclination obtained by piecewise function calculation.

[0083] Example 5

[0084] A drilling image intelligent picking system based on deep learning, comprising:

[0085] A data acquisition module, wherein the data acquisition module is used to acquire drilling data to obtain a data set of drilling images, and divide the acquired data set into a training set, a validation set, and a test set;

[0086] The instance segmentation module, the target picking module extracts features of different types of structural surfaces such as vein fillings and fractures in the borehole image through the multi-classification picking task constructed by the instance segmentation model YOLOv8_seg, and obtains the corresponding mask coordinates of each structural surface as the basis for the next calculation of the algorithm model;

[0087] The occurrence calculation module is used to input the initially obtained structural surface mask coordinates into a polynomial function to fit a structural surface curve that is similar to a sine function, and calculate the occurrence of the structural surface such as the inclination and dip angle through the extreme value of the fitted structural surface curve.

[0088] The borehole image automatic picking module uses a trained model to automatically and quickly pick up structural surfaces such as cracks and discontinuities in the interference of complex information on the borehole wall image and calculate their occurrence, and finally outputs the borehole image automatically picked up by the model.

[0089] The present invention not only completes the complete process from image recognition to parameter calculation, but also realizes intuitive visualization of the results (automatically marking and drawing the calculated occurrence information on the borehole image). At the same time, it provides geological engineers with an intuitive and convenient analysis tool, which can greatly improve the efficiency and accuracy of geological exploration and engineering evaluation, and can provide new ideas and solutions for the development of the field of borehole image analysis in the direction of intelligence and automation.

[0090] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A drilling image intelligent picking method based on deep learning, characterized in that: The following steps are involved: Step 1: Collect drilling data to obtain a dataset of drilling images, and divide the collected dataset into a training set, a validation set, and a test set; Step 2: Through the multi-classification picking task constructed by the instance segmentation model YOLOv8_seg, feature extraction is performed on different types of structural surfaces such as vein fillings and fractures on the hole wall image, and the corresponding mask coordinates of each structural surface are obtained as the basis for the next calculation of the algorithm model; Step 3: Input the mask coordinates obtained in step 2 into the polynomial fitting formula, and calculate the structural curve of the structural surface in the image according to the mask coordinates obtained by the model, so as to obtain the extreme value points of the structural surface fitting curve; Step 4: Then, based on the extreme value of the curve fitted in step 3 and the corresponding parameters, the inclination and dip angle of the structural surface are calculated, and finally the drilling image automatically analyzed and drawn by the algorithm model is output.

2. According to the method of intelligent picking of drilling images based on deep learning in claim 1, it is characterized in that: In step 1, the dataset is divided into training set, validation set and test set in the form of 8:1:

1.

3. According to the method of intelligent picking of drilling images based on deep learning in claim 1, it is characterized in that: In step 2, the instance segmentation model YOLOv8_seg picks up unconformities of different shapes in the hole wall image; The labels for the structural surface include Close, Open, and Dike Filling.

4. According to the method of intelligent picking of drilling images based on deep learning in claim 3, it is characterized in that: The picking rule of Close in structural surface picking is the structural surface area in a closed state in the image; The picking rule of Open in the structural surface picking is the structural surface area in the open state; The picking rule of Dike Filling in structural surface picking is to find structural surface areas with calcite or mica that are obviously different in color from the rock mass.

5. According to the method of intelligent picking of drilling images based on deep learning in claim 1, it is characterized in that: In step three, when the drilling image automatic picking algorithm model BIAR is training the data set, the input data set format is COCO format, the backbone network uses the backbone network that comes with YOLOv8, the optimizer selects AdamW, the learning rate is designed to be 1e-4, and the GPU is used to accelerate the training process without freezing training.

6. The method for intelligently picking up drilling images based on deep learning according to claim 1 is characterized in that: The instance segmentation model YOLOv8_seg performs instance segmentation as follows: The input module preprocesses the input image (drilling image), including scaling and normalization, to ensure that the image fits the input size and format of the model, and extracts target features from the input image through Backbone, thereby converting the image into a feature representation with rich semantic information. Then, the FPN (Feature Pyramid Networks) module performs multi-scale feature fusion and constructs a multi-scale feature map to enhance the model's detection ability for targets of different sizes. The PANet (Path Aggregation Network) further integrates feature information at different levels to improve the expressiveness of the feature map. Next, the Detection Head is responsible for target classification and bounding box regression, generating category predictions and bounding box regression values ​​(Bbox) for each candidate area. After detecting the target, the Instance Segmentation Head performs pixel-level segmentation on the image and outputs the precise outline of each target. During the training of the model, the Loss Function module calculates the difference between the model output and the true annotation, and adjusts the model parameters through back propagation to minimize the loss value. The post-processing module uses NMS (Non-Maximum Suppression) removes redundant overlapping boxes and filters out overlapping detection boxes to obtain the precise target location. Finally, the detection and segmentation results of the algorithm are output, including the target category, confidence, bounding box coordinates, and segmentation mask.

7. The method for intelligently picking up drilling images based on deep learning according to claim 1 is characterized in that: The process of structural surface curve fitting is: The structural surface mask obtained in the previous step of image segmentation is fitted by polynomial fitting to obtain the corresponding curve parameters. Before fitting the data with the polynomial function, it is necessary to first define the polynomial function f(x)=a according to the fitting requirements. n x n +a n-1 x n-1 +...+a1x+a0, and then use the least squares method and other optimization algorithms to minimize the residual sum of squares formula (i.e., the sum of the squares of the differences between the actual observed values ​​and the predicted values ​​of the fitted curve) to solve the polynomial coefficients. Finally, when the polynomial coefficients are determined, the fitted polynomial function can be used to perform further data analysis on the mask coordinates. Among them, a in the formula n , a n-1 ,…, a1, a0 are the coefficients to be determined, n is the degree of the polynomial, y i is the i-th observed value of the dependent variable, that is, the value actually recorded in the experiment or in the data, f(x i ) is based on the model and the independent variable x i The predicted value is calculated from the value of .

8. The method for intelligently picking up drilling images based on deep learning according to claim 1 is characterized in that: The calculation process of the inclination angle and other occurrence of the structural surface is as follows: The extreme points A(MinX, MinY) and B(MaxX, MaxY) of the structural surface calculated in the fitting formula are connected by the formula The coordinates on the y-axis are interpolated to calculate the absolute value of the difference between the extreme point coordinates X val and Y val , and convert the difference into Calculate the inclination angle of each structural surface; Among them, Y in the formula cal is the actual diameter of the drilled hole, C degrees is the calculated inclination angle.

9. The method for intelligently picking up drilling images based on deep learning according to claim 1 is characterized in that: The calculation process of the dip and isoecdence of the structural surface is as follows: The lowest point of the structural surface curve on the y-axis corresponds to the mapping between 0° and 360°. If the coordinate x i Between 0 and x mid (the minimum value of the drilling image on the x-axis), its inclination can be averagely mapped to between 180° and 360°; if the coordinate x of the maximum point is i In x mid to x max (the maximum value of the drilling image on the x-axis), its inclination can be averagely mapped to between 0° and 180°, and the mapping rule formula is as follows: Among them, α in the formula Angle It is the structural surface inclination obtained by piecewise function calculation.

10. A picking system used in the drilling image intelligent picking method based on deep learning according to any one of claims 1 to 9, characterized in that: include: A data acquisition module, wherein the data acquisition module is used to acquire drilling data to obtain a data set of drilling images, and divide the acquired data set into a training set, a validation set, and a test set; An instance segmentation module, which extracts features of different types of structural surfaces such as vein fillings and fractures in the borehole image through a multi-classification picking task constructed by the instance segmentation model YOLOv8_seg, and obtains the corresponding mask coordinates of each structural surface as the basis for the next step of calculation of the algorithm model; An occurrence calculation module is used to input the initially obtained structural surface mask coordinates into a polynomial function to fit a structural surface curve that is similar to a sine function, and calculate the occurrence of the structural surface such as the inclination and dip angle through the extreme value of the fitted structural surface curve; The borehole image automatic picking module uses a trained model to automatically and quickly pick up structural surfaces such as cracks and discontinuities in the interference of complex information on the borehole wall image and calculate their occurrence, and finally outputs the borehole image automatically picked up by the model.

Images