RQD Calculation Method Based on Deep Learning Model and Core Image

Abstract

The present invention discloses an RQD calculation method based on a deep learning model and borehole core images, including: establishing a borehole core image data set; extracting image features of the borehole core image data set, inputting the image features into a UNet deep network for training to obtain an EUNet model; using the EUNet segmentation model to perform semantic segmentation on the borehole core image to obtain a core area and a background area, and finding a single-row borehole core image; extracting the contours of all cores in the single-row borehole core image, counting the number of core contour pixels, making a pixel waveform diagram, and judging the core type through the pixel waveform diagram; determining the number of intact cores and their locations from the pixel waveform diagram to determine the research area; further determining the boundaries of the cores; and calculating the core lengths according to the boundaries of the cores to obtain the RQD. The intelligent quantitative analysis of RQD is realized, and the efficiency of geological exploration is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing methods, and relates to a method for calculating RQD based on a deep learning model and core images. Background Art

[0002] Large and medium-sized water conservancy and hydropower projects are mostly located in alpine valleys with relatively complex geological conditions, where fault structures are developed and there are obvious differences in the integrity of rock masses in different regions. Accurately and objectively judging geological conditions and rock mass integrity is an important prerequisite for the design and construction of water conservancy and hydropower projects.

[0003] As an important index for evaluating rock mass quality, RQD is widely used in water conservancy projects. RQD is also a basic parameter in multi-factor evaluation systems for rock masses such as RMR and Q-system. The traditional method for obtaining RQD is calculated by geological staff after measuring the length of drilled cores, but manual measurement of RQD is time-consuming and laborious, reducing the efficiency of geological exploration. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for calculating RQD based on a deep learning model and core images, which solves the problem of low efficiency of geological exploration caused by manual measurement of RQD in the prior art.

[0005] The technical solution adopted by the present invention is a method for calculating RQD based on a deep learning model and borehole core images, including the following steps:

[0006] Step 1: Collect borehole core images, correct the borehole core images, and establish a borehole core image dataset;

[0007] Step 2: Extract the image features of the borehole core image dataset, input the image features into the UNet deep network for training, and obtain the EUNet model;

[0008] Step 3: Use the EUNet segmentation model to perform semantic segmentation on the borehole core image to obtain the core area and the background area, and find the single-row borehole core image;

[0009] Step 4: Apply the Canny edge detection algorithm to extract the contours of all cores in the single-row borehole core image, count the number of core contour pixels, make a pixel waveform diagram, and judge the core type through the pixel waveform diagram;

[0010] Step 5: Determine the number and location of intact cores from the pixel waveform diagram. When the pixels in the core area at both ends of the intact core section increase, take the wave peak group adjacent to the head and tail of the intact core section as the research area; if there are continuous wave peaks and wave valleys at one end of the intact core section, the length of the research area should be taken as 30 mm;

[0011] Step 6: Locate the research area in the image after semantic segmentation, and use the midline of the background area as the boundary of the core; when the cores are closely adjacent to each other, the curve fitting method is used to fit the boundary of the core from partial boundaries.

[0012] Step 7: Calculate the core length according to the boundary of the core, and the RQD is the ratio of the sum of the core lengths greater than 10 cm in the footage to the total footage.

[0013] The features of the present invention also lie in:

[0014] After step 2, the performance of the EUNet model is judged, and the next step is carried out when the performance meets the requirements.

[0015] The method for judging the performance of the EUNet model is:

[0016] Use the EUNet model to segment the borehole core image, and then calculate the F1_score value and IoU value, and judge the performance of the EUNet model through the F1_score value and IoU value.

[0017] The judgment method for the single-row borehole core image in step 3 is: traverse the pixels of each row in the segmentation result, count the number of core area pixels in each row, and use the row with zero core area pixels as the segmentation line to cut out the single-row core to obtain the single-row borehole core image.

[0018] The calculation method of the core length is: the actual length represented by a single pixel in the borehole core image is the physical length of the core box divided by the number of pixels in the length direction of the corrected borehole core image; take the horizontal midline of a single row of the core as the reference line for determining the core length. When calculating the core in the middle position, the core length is the total length of the core area pixels between the left and right boundaries on the reference line; when calculating the core at both ends, the core length is the total length of the core area pixels from the first core area pixel at the head or tail on the reference line to the boundary.

[0019] The beneficial effects of the present invention are:

[0020] The RQD calculation method based on a deep learning model and core images in the present invention uses a deep model transfer algorithm to accurately identify the core part in borehole core images, and uses multiple indicators to evaluate the core image segmentation results, avoiding the influence of subjective factors and improving the core recognition efficiency; the core segmentation model is used to segment borehole core images in actual projects, single-row cores are segmented by scanning horizontal pixels, and a core contour map is obtained through contour detection. The vertical pixels of the contour map are scanned to obtain a pixel waveform map. Through morphological and logical analysis of the semantic segmentation image and the waveform map, the internal distribution law between pixels is summarized, the calculation method of core length is explored, the intelligent quantitative analysis of RQD is realized, and the efficiency of geological exploration is improved; it provides a theoretical basis and method support for the rapid assessment of the rock mass quality index in the rock foundation of hydraulic engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0022] Figure 2 is a schematic diagram of the image correction process of the RQD calculation method based on a deep learning model and borehole core images in the present invention:

[0023] Figure 3 is a schematic diagram of the UNet model structure in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0024] Figure 4 is a schematic diagram of the change of different evaluation indicators during the training process of the EUNet model in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0025] Figure 5 is a comparison diagram of the borehole core image and the semantic segmentation result in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0026] Figure 6a is a comparison diagram of the single-row core image and the semantic segmentation result in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0027] Figure 6b is a diagram for determining the segmentation position of a single-row core in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0028] Figure 7 is a comparison diagram of the semantic segmentation result of a single-row core and the Canny edge detection result in the RQD calculation method based on a deep learning model and borehole core images in the present invention;

[0029] Figure 8It is the pixel waveform diagram in the RQD calculation method based on the deep learning model and borehole core image of the present invention;

[0030] Figure 9 It is the schematic diagram for determining the research area in the RQD calculation method based on the deep learning model and borehole core image of the present invention;

[0031] Figure 10 It is the schematic diagram of fitting the core boundary curve in the research area in the RQD calculation method based on the deep learning model and borehole core image of the present invention;

[0032] Figure 11 It is the method diagram for determining the core length in the RQD calculation method based on the deep learning model and borehole core image of the present invention. Specific implementation manners

[0033] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0034] The RQD calculation method based on the deep learning model and borehole core image, as Figure 1 shown, includes the following steps:

[0035] Step 1: Collect borehole core images, correct the borehole core images, and establish a borehole core image data set, which includes a training set and a test set;

[0036] Step 1.1: Collect borehole core images, which should include both relatively complete core images and cores with fractures, flakes, and fragmentation;

[0037] Step 1.2: Perform image correction and unify the size of the borehole core images; the image correction method is: determine the position of the core box through the four corner points of the core box, take the four sides of the rectangular frame determined by the four points as the objects, stretch and rotate the image within the frame to make the field of view horizontal with the image, see Figure 2 ; unify the image size to 640×480;

[0038] Step 1.3: Mark the core part in the borehole core images processed in Step 1.2 to make image labels; 80% of the core images are used as the training set, and 20% are used as the validation set, and the core data set is completed. In this embodiment, a total of 280 borehole core photos are collected, 240 are selected as the training set, and 40 are selected as the test set.

[0039] Step 2: Use EfficientNetb5 as the pre-trained model to extract the image features in the training set, and input the image features into the UNet deep network for training, as Figure 3As shown, a total of 200 epochs were set in the training process, the batchsize (the number of samples selected for one training) was set to 4, and the model training took a total of 3.2 hours to obtain the EUNet model; the EUNet model was used to segment the borehole core images in the validation set, and then the F1_score value and the IoU value were calculated. The changes of different evaluation indicators in the model training process are as Figure 4 shown; it is set that when the F1_score value is above 95% and the IoU value is above 85%, the trained model is considered applicable to the segmentation of borehole core images. In this embodiment, the F1_score converges to 0.973 and the IoU converges to 0.947, meeting the recognition accuracy requirements. The calculation formulas for the F1_score value and the IoU value are as follows:

[0040]

[0041]

[0042]

[0043]

[0044] Among them, TP represents that the label is positive and the prediction is also positive; FP represents that the label is negative and the prediction is positive; TN represents that the label is negative and the prediction is negative; FN represents that the label is positive and the prediction is negative; DT represents the range of the prediction result, and GT represents the true result range of the object; ∩ represents the intersection, and ∪ represents the union; Precision is the precision rate, and Recall is the recall rate.

[0045] Step 3: Use the EUNet segmentation model to perform semantic segmentation on the borehole core images to obtain the core area and the background area, as Figure 5 shown. In the figure, a is the borehole core image, and b is the semantic segmentation result image; in this embodiment, the core area in the segmentation result is white and the background area is black; traverse the pixels in each row of the segmentation result, count the number of white pixels in each row, and use the row with zero white pixels as the segmentation line to cut out a single row of core to obtain a single-row borehole core image, as Figure 6a and Figure 6b shown. Figure 6a In the figure, a is the borehole core image, and b is the semantic segmentation result;

[0046] Step 4: Apply the Canny edge detection algorithm to extract the contours of all cores in the single-row borehole core image, and draw the core contours with single-width white pixels, as Figure 7 shown. Figure 7In the figure, a is the image of the borehole core. In the figure, b is the result of Canny edge detection. The number of pixels of the core contour is counted to produce a pixel waveform diagram. The type of core is judged through the pixel waveform diagram. Specifically, the part where the number of white pixels in the waveform diagram remains 2 continuously is regarded as a complete core, and the rest is regarded as a broken section. As Figure 8 shown, in the figure, a is the semantic segmentation result of the single-row core, b is the result of Canny edge detection, and c is the waveform diagram;

[0047] Step 5: Determine the number and location of the complete cores in the pixel waveform diagram, and at the same time determine the research area. Specifically, in the pixel waveform diagram, when the number of pixels in the white areas at both ends of the complete core section increases, the peak group adjacent to the head and tail of the complete core section is used as the research area. As Figure 9 shown, the cores marked with numbers in the figure are the complete core blocks determined to calculate the length, and the rest is the broken section; if there are continuous peaks and valleys at one end of the complete core section and it is difficult to determine the boundary of the research area here, when the research area is too long, it is not conducive to subsequent processing. After processing a large amount of core data and summarizing, the length of the research area should be taken as 30 mm;

[0048] Step 6: Locate the research area in the image after semantic segmentation. The background area (black area) is used as the separation area between the cores, and the midline of the black area is used as the boundary of the core; when the cores are close together, the EUNet model cannot completely separate them and the complete boundaries of the two cores cannot be obtained. Therefore, a curve fitting method is used to fit the boundary of the core from part of the boundary. As Figure 10 shown; in this embodiment, a quartic polynomial curve can accurately fit the core boundary line;

[0049] Step 7: The actual length represented by a single pixel in the borehole core image is the ratio of the physical length of the core box to the number of pixels in the length direction of the corrected borehole core image; the horizontal midline of the single-row core is taken as the reference line for determining the core length. When calculating the core in the middle position, the core length is the total length of the pixels in the core area between the left and right boundaries on the reference line; when calculating the core at both ends, the core length is the total length of the pixels in the core area from the first pixel in the core area at the head or tail on the reference line to the boundary. As Figure 11 shown; then RQD is the ratio of the sum of the core lengths greater than 10 cm in the footage to the total footage.

[0050] Ascertaining the quality of slopes and underground rock masses is of great significance to water conservancy, highway, mining and other projects. The method of coring by drilling can directly expose the underground rock mass and evaluate the quality of the rock mass, which is widely used in engineering. The drilled cores are loaded into standard core boxes according to the drilling order and photographed and recorded, which is convenient for geological engineering personnel to subsequently compile RQD and analyze other rock mass indicators such as lithology. The core photos taken can reflect the position and size information of the core, and the information therein can be interpreted through digital image processing to obtain the core RQD. However, when taking core photos, different core boxes have different shooting orientations and angles, and dust, soil, etc. at the engineering site are attached to the surface of the core or core box, which greatly increases the difficulty of image recognition. The present invention uses an image correction algorithm to overcome the angle and position deviation problems when shooting cores; establishes an EUNet deep learning model to achieve accurate segmentation of cores; and proposes a refined analysis process to achieve intelligent calculation of RQD.

[0051] Through the above methods, the present invention is based on the deep learning model and the RQD calculation method of the core image, and adopts the deep model migration algorithm to realize the accurate identification of the core part in the borehole core image, and adopts multiple indicators to evaluate the core image segmentation result, so as to avoid the influence of subjective factors and improve the recognition efficiency of the core; the core segmentation model is used to segment the borehole core image in the actual project, and the single row of cores is segmented by scanning the horizontal pixels, and the core contour map is obtained by contour detection, and the vertical pixels of the contour map are scanned to obtain the pixel waveform map, and the intrinsic distribution law between pixels is summarized by performing morphological analysis and logical analysis on the semantic segmentation image and the waveform map, and the calculation method of the core length is explored, and the intelligent quantitative analysis of RQD is realized, so as to improve the efficiency of geological exploration; it provides a theoretical basis and method support for the rapid evaluation of rock quality indicators in the rock foundation of water conservancy projects.

[0052] Example

[0053] The method of the present invention is used to calculate the RQD of 40 borehole core images in the test set, and the lengths of 1302 core blocks are calculated cumulatively. The results are shown in Table 1. Taking a single box core as the research object, the table lists the manual measurement results (M) and automatic quantization results (A) of RQD, the absolute error (AE) and relative error (RE) of RQD prediction, where AE = |MA|, RE = AE / M. Compared with manual measurement, the average absolute error of the RQD prediction value is 1.48%, and the maximum value is 3.74%, which can meet the RQD quantification accuracy requirements.

[0054] Table 1 Comparison of manual and automatic RQD quantification results

[0055]

[0056]

Claims

1. A method for calculating RQD based on a deep learning model and borehole core images, characterized in that, The steps include the following: Step 1: Collect borehole core images. The borehole core images should include both relatively complete core images and cores with fractures, flakes, and fragmentation. Perform image correction and unify the image size for the borehole core images. The image correction method is as follows: Determine the position of the core box through the four corner points of the core box. Take the four sides of the rectangular box determined by the four points as the objects, stretch and rotate the image within the box so that the field of view is horizontal with the image. Unify the image size to 640×480. Mark the core part in the processed borehole core images and make them into image labels. Use 80% of the core images as the training set and 20% as the validation set to establish a borehole core image dataset. Step 2: Extract the image features of the borehole core image dataset, input the image features into the UNet deep network for training, and obtain the EUNet model. A total of 200 epochs are set during the training process, and the batchsize is set to 4 to obtain the EUNet model. Use the EUNet model to segment the borehole core images in the validation set, and then calculate the F1_score value and the IoU value. It is set that when the F1_score value is above 95% and the IoU value is above 85%, the trained model is applied to the segmentation of borehole core images. Step 3: Use the EUNet segmentation model to perform semantic segmentation on the borehole core images to obtain the core area and the background area, and find the single-row borehole core images. The judgment method for single-row borehole core images is as follows: Traverse the pixels in each row of the segmentation result, count the number of core area pixels in each row, and use the row with zero core area pixels as the segmentation line to cut out the single-row core to obtain the single-row borehole core images. Step 4: Apply the Canny edge detection algorithm to extract the contours of all cores in the single-row borehole core images, count the number of core contour pixels, make a pixel waveform diagram, and judge the core type through the pixel waveform diagram. Consider the core with the white pixel number continuously being 2 in the waveform diagram as a complete core, and the rest as the broken section. Step 5: Determine the number and positions of the complete cores from the pixel waveform diagram. When the core area pixels at both ends of the complete core section increase, take the peak group adjacent to the head and tail of the complete core section as the research area. If continuous peaks and valleys appear at one end of the complete core section, the length of the research area is taken as 30mm. Step 6: Locate the research area in the image after semantic segmentation, and use the midline of the background area as the boundary of the core. When the cores are closely adjacent to each other, use the curve fitting method to fit the boundary of the core from part of the boundary. Step 7: Calculate the core length according to the core boundary. The calculation method of the core length is as follows: The true length represented by a single pixel in the borehole core image is the physical length of the core box divided by the number of pixels in the length direction of the corrected borehole core image. Take the horizontal midline of a single row of the core as the reference line for determining the core length. When calculating the core in the middle position, the core length is the total length of the core area pixels between the left and right boundaries on the reference line. When calculating the core at both ends, the core length is the total length of the core area pixels from the first core area pixel at the head or tail on the reference line to the boundary. Then RQD is the ratio of the sum of the core lengths greater than 10 cm in the footage to the total footage.

2. The RQD calculation method based on the deep learning model and borehole core images according to claim 1, wherein After step 2, judge the performance of the EUNet model. When the performance meets the requirements, proceed to the next step.

3. The RQD calculation method based on a deep learning model and borehole core images according to claim 2, characterized in that The method for judging the performance of the EUNet model is as follows: Use the EUNet model to segment the borehole core image, and then calculate the F1_score value and the IoU value. Judge the performance of the EUNet model through the F1_score value and the IoU value.

Images