An intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines

Abstract

An intelligent method for measuring the surface trace spacing of hard rock pillars in underground mines collects massive amounts of raw trace feature images. After preprocessing, the image data is divided into training and test sets according to a set ratio. A convolutional neural network based on a deep learning framework is constructed, trained, and tested to obtain a prediction model. The target image is input into the prediction model to predict a two-dimensional trace distribution map. First, based on the principle of angle difference, an angle threshold is input to perform fracture segmentation on the traces. A secondary detection is then performed using the fracture threshold to obtain a number of trace segments consistent with the actual survey results. Based on the principle of angle similarity, classification angles are set, all trace segments are grouped and evaluated, and the optimal number of groups is determined. Manual scan line calibration is performed on all traces and traces within each individual group to determine the optimal values ​​for intra-group and total spacing. This method can accurately and efficiently measure the surface trace spacing of hard rock pillars.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring technology, and in particular relates to an intelligent measurement method for surface trace spacing of hard rock pillars in an underground mine. Background Art

[0002] Trace spacing is a key characteristic for evaluating the quality and performance of surface and underground rock masses. Macroscopically, a rock mass is a collection of individual rock blocks, and trace spacing directly describes the size of these blocks. For hard rock pillars, in underground environments characterized by high ground stress, high ground temperature, high well depth, and significant mining disturbance, the initial degree of failure often depends on the size of the blocks containing developed joints and fissures. Therefore, accurate measurement of trace spacing on the surface of hard rock pillars can further improve the accuracy of pillar and stope stability assessments.

[0003] Traditionally, measuring the spacing between traces on the surface of pillars is often done manually, with geologists or specialized technicians using contact equipment such as measuring tapes and curve boards to perform complex and lengthy measurements on the pillar surface. However, in the dimly lit environment of underground mines, and subject to significant interference from unknown hazards such as roof falls and rockfalls, the instability of contact equipment and the uncertainty of manual operation further reduce the accuracy of trace measurements. This is particularly true when measuring the spacing between traces on the surfaces of extremely long pillars. Therefore, acquiring high-definition, two-dimensional digital images of the pillar surface based on non-contact measurement technology, and relying on image processing and intelligent automatic / semi-automatic analysis of image information to extract traces and complete spacing measurements, has become a better, more reliable, and safer technical approach.

[0004] However, extracting accurate trace information from two-dimensional images of complex rock masses remains a challenging and challenging issue in geotechnical engineering. Currently, edge detection-based extraction methods are suitable for extracting trace information from exposed rock mass images with a small number of uniformly distributed targets. When images contain a large amount of information and target features are difficult to discern, extraction methods based on target prediction within deep learning frameworks can more deeply capture trace information and reduce unnecessary noise interference. However, the performance of deep learning models is undeniably dependent on the number and quality of samples, a requirement difficult to achieve in the complex and hazardous environment of underground mines. Therefore, establishing high-performance prediction models based on limited, average-quality images is essential for ensuring trace spacing measurements. Furthermore, while the predicted traces are continuous and complete, the one-to-one correspondence between traces and discontinuities is neglected. Therefore, segmenting traces at key locations to accurately describe plane boundaries is a prerequisite for calculating spacing. Furthermore, due to the chaotic distribution of discontinuities, measuring the spacing between traces distributed within similar planes requires accurate classification of all traces in advance. Most existing classification methods are developed based on the concept of clustering, which usually only focus on the similarity between traces and ignore the discontinuity of traces. Therefore, measurement errors are inevitable in practical applications.

[0005] Based on the above problems, it is urgent to propose a new intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines, so as to realize the intelligent calculation of the surface trace spacing of pillars and provide reliable technical support for the stability evaluation of hard rock pillars in underground mines. Summary of the Invention

[0006] In response to the problems existing in the above-mentioned prior art, the present invention provides an intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines. The method has a simple implementation process, low implementation cost, high measurement efficiency and high measurement accuracy. It can realize the intelligent and automated measurement process of the surface trace spacing of hard rock pillars in underground mines, and can accurately and efficiently measure the trace spacing data on the surface of hard rock pillars, which can be beneficial to further improve the accuracy of pillar and stope stability evaluation, and can provide effective technical support for the stability evaluation of hard rock pillars in underground mines.

[0007] In order to achieve the above object, the present invention provides an intelligent measurement method for the surface trace spacing of hard rock pillars in an underground mine, which specifically includes the following steps:

[0008] Step 1: Collect original trace feature images of the surfaces of pillars in different states in a large number of underground stopes;

[0009] Step 2: Preprocess the collected original trace feature images, reassemble the preprocessed images to complete the construction of the image database, and divide the image data in the image database into training set and test set according to the set ratio;

[0010] Step 3: Build a convolutional neural network based on a deep learning framework, use the training set to train the convolutional neural network to obtain a high-precision convolutional neural network, and use the test set to test the performance of the high-precision convolutional neural network to obtain a prediction model for predicting the distribution of traces within the image;

[0011] Step 4: Input the target image to be analyzed as input data into the prediction model, use the prediction model to predict the trace features of the target image, and obtain the predicted distribution map of the trace;

[0012] Step 5: First, based on the angle difference principle, input the angle threshold and perform fracture segmentation on all predicted traces to initially obtain a number of trace segments with a certain pixel length. Then, introduce the fracture threshold to perform secondary detection on the initially obtained trace segments to obtain the number of trace segments consistent with the actual survey results.

[0013] Step 6: According to the principle of angle similarity, set the classification angle, group and evaluate all trace segments, and determine the optimal number of groups and the number of trace lines they contain based on the principle of minimizing classification error;

[0014] Step 7: Perform manual scan line calibration on all traces and the traces in each individual group, and determine the optimal intra-group spacing and total spacing measurement values ​​based on the calibration results.

[0015] As a preference, the different states in step 1 include a stable state, a temporarily stable state and an unstable state.

[0016] Furthermore, in order to obtain rich original trace feature images and thus ensure a prediction model with higher prediction accuracy, a high-performance prediction model is established based on limited general quality images. The specific process of step 2 is as follows:

[0017] S21: Screening the original trace feature images on the pillar surface on site to remove images with excessive or insufficient exposure;

[0018] S22: Manually mark the trace features in each image, and have geological experts or other professional technicians repeatedly verify whether the markings are reasonable;

[0019] S23: Using data augmentation technology, each image is flipped, moved, rotated at a set angle, partially enlarged, and mirrored to expand the number of original trace images, thereby completing the construction of the image database;

[0020] S24: Allocate the image data in the database, and divide 80% of the image data into a training set, and divide 20% of the image data into a test set; at the same time, calculate the trace and non-trace pixel distribution of the images in the training set and the test set. If the pixel distribution difference between the two sets meets the pixel distribution requirements, the data allocation result is considered reasonable. Otherwise, the image data in the database is reallocated until the pixel distribution between the two sets meets the pixel distribution requirements.

[0021] Furthermore, in order to more deeply capture the characteristics of trace information and reduce unnecessary noise interference to ensure prediction accuracy, in step three, the total loss function is first constructed, and then the convolutional neural network is trained. During the training process, the preset accuracy index is used to evaluate whether the performance of the trained convolutional neural network meets the requirements.

[0022] Furthermore, segmenting the traces at key locations so that they can truly and accurately describe the plane boundaries is a prerequisite for accurately calculating the spacing. In order to correspond the traces to the discontinuous planes one-to-one, a segmentation algorithm is used to segment the traces. At the same time, in order to further enhance the accuracy of trace fractures, a fracture threshold is incorporated to evaluate the segmentation effect, so as to achieve accurate measurement of the trace spacing on the pillar surface. Therefore, the specific process of step five is as follows:

[0023] S51: confirm the number of pixels contained in each trace, and calculate the angle between three consecutive pixels starting from the starting point;

[0024] S52: Based on the angle breaking algorithm, different angle thresholds are set and compared with all the line segment angles obtained in S51 one by one to determine whether the trace needs to be broken. If the angle threshold is greater than the current line segment angle, the trace is broken at the middle point of the line segment angle and a trace breaking diagram is obtained. At the same time, the middle point is used as the starting point for calculating the next line segment angle value. Otherwise, the comparison of the next line segment angle is continued; until all traces are traversed and the corresponding trace breaking diagram is obtained;

[0025] S53: setting a fracture threshold, importing all trace segments obtained in S52, deleting trace segments whose length is less than the specified fracture threshold, and obtaining a new trace fracture map;

[0026] S54: A geological expert or other professional technician compares the distribution of the new trace segments obtained in S53 with the actual survey. If non-negligible trace segments are deleted, it is determined that the fracture threshold is over-selected, and S53 is re-executed to obtain a new trace fracture map by re-adjusting the fracture threshold. At the same time, by combining the adjustment of different angle thresholds, it is ensured that tiny traces that do not contribute significantly to the spacing measurement are not extracted too much, until the optimal angle threshold and fracture threshold are determined to complete the trace fracture evaluation and obtain the final trace fracture map.

[0027] Furthermore, since the distribution of discontinuous surfaces is chaotic, the measurement of the distance between traces distributed in similar planes requires accurate classification of all trace segments in advance. However, most existing classification methods are developed based on the clustering concept, which basically only focuses on the similarity between traces and ignores the discontinuity of traces. Therefore, errors in actual field applications are inevitable. In order to accurately and efficiently classify trace segments distributed in similar planes, the following grouping algorithm is used for processing. Therefore, the specific process of step six is ​​as follows:

[0028] S61: Determine the optimal classified angle value;

[0029] S62: importing the trace fracture map obtained after segmentation based on the optimal angle threshold and cropping based on the fracture threshold, and calculating the difference between the angles between any two trace segments in the map and the horizontal axis;

[0030] S63: grouping two trace segments whose calculated difference is less than the optimal classification angle value classified angle into the same group, and so on, traversing all trace segments, and determining the final number of trace groupings after the classification is completed.

[0031] Furthermore, in order to determine the optimal number of groups, a strategy based on classification error is adopted to determine the optimal number of groups. In step 6 S61, the classification error Error is first calculated according to formula (1), and then the optimal classification angle value classified angle is determined according to the principle of minimizing the classification error Error.

[0032]

[0033] Where θ i m is the angle of the mth trace segment in the i-th trace group; averθ i is the average angle of all trace segments in the i-th trace group, and I is the maximum number of trace groups.

[0034] Furthermore, in order to accurately and efficiently calculate the spacing information, in step seven, during the intra-group spacing measurement process, the scan line is made perpendicular to the majority of trace segments within the same group. At the same time, during the total spacing measurement process, the scan line is made perpendicular to the majority of trace segments in all traces.

[0035] As a preference, in step 2 S23, the angle is set to 45 degrees, 90 degrees, 180 degrees, or 270 degrees.

[0036] The present invention provides a semi-automated and interactive method for measuring the spacing of complex traces on the surface of hard rock pillars in underground mines. First, a large number of pillar surface feature images taken in underground stopes are collected and trained using a convolutional neural network based on a deep learning framework to obtain a prediction model for predicting the distribution of traces within the image. The prediction model can automatically capture trace information features from the target image to be analyzed and help reduce unnecessary noise interference. Second, the prediction model is used to predict the target image to be analyzed to obtain a predicted distribution map of the traces. This is then segmented using an angle-based fracture algorithm. Simultaneously, a fracture threshold is incorporated to delete trace segments that do not conform to the actual survey. The final new trace fracture map is obtained by adjusting the angle threshold and fracture threshold, and combining actual survey results and expert evaluation. Thus, the present invention effectively integrates the fracture threshold concept into the traditional angle-based fracture algorithm, thereby achieving accurate segmentation of trace segments. Accurately segmenting traces at key locations can help accurately reflect the plane boundary, thereby enabling one-to-one correspondence between traces and discontinuous planes, ensuring the subsequent accurate measurement of the spacing between traces on the pillar surface. Then, based on the principle of minimizing classification error, the optimal classification angle is determined, and a grouping algorithm based on the classification angle is used to divide the trace segments into different sets to obtain the optimal number of groups. Thus, by developing a new grouping algorithm based on the classification angle to rationally divide the trace segments and proposing an error minimization criterion to enhance the accuracy of trace grouping, this method can accurately and efficiently classify trace segments distributed within similar planes, while ensuring that trace discontinuities are not ignored, thereby ensuring that the spacing between traces distributed within similar planes can be accurately measured. Finally, manual scan line calibration is performed on all traces and the traces within each group to obtain the intra-group and total spacing values ​​of the traces.

[0037] This method has a simple implementation process, low implementation cost, high measurement efficiency and high measurement accuracy. It can realize the intelligent and automated measurement process of the trace spacing on the surface of hard rock pillars in underground mines, and can accurately and efficiently measure the trace spacing data on the surface of hard rock pillars. It can be beneficial to further improve the accuracy of pillar and stope stability evaluation, and can provide effective technical support for the stability evaluation of hard rock pillars in underground mines. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a workflow diagram of the present invention;

[0039] Figure 2 This is a schematic diagram of the state of enhancing the trace image data according to the present invention;

[0040] Figure 3 Schematic diagram of trace breakage in the present invention;

[0041] Figure 4 Schematic diagram of trace segment grouping in the present invention;

[0042] Figure 5 Schematic diagram of scan line calibration for measuring trace spacing in the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below with reference to the embodiments.

[0044] In order to facilitate those skilled in the art to understand and implement the present invention, Figures 1 to 5 The present invention will be further described;

[0045] The present invention provides an intelligent measurement method for the surface trace spacing of hard rock pillars in an underground mine, which specifically includes the following steps:

[0046] Step 1: Collecting original trace feature images of the surfaces of pillars in different states in a large number of underground stopes; preferably, using image acquisition equipment to collect original trace feature images in the underground stope under sufficient light conditions, and ensuring that each image contains as much trace feature information as possible;

[0047] Step 2: Preprocess the collected original trace feature images, reassemble the preprocessed images to complete the construction of the image database, and divide the image data in the image database into training set and test set according to the set ratio;

[0048] Step 3: Build a classic convolutional neural network based on the deep learning framework, use the training set to train the convolutional neural network to obtain a high-precision convolutional neural network, and use the test set to test the performance of the high-precision convolutional neural network to obtain a prediction model (high-precision convolutional neural network) for predicting the distribution of traces in the image;

[0049] Step 4: Input the target image to be analyzed as input data into the prediction model, use the prediction model to predict the trace features of the target image, and obtain the predicted distribution map of the trace;

[0050] Step 5: First, based on the angle difference principle, input the angle threshold and perform fracture segmentation on all predicted traces to initially obtain a number of trace segments with a certain pixel length. Then, introduce the fracture threshold to perform secondary detection on the initially obtained trace segments to obtain the number of trace segments consistent with the actual survey results.

[0051] Step 6: According to the principle of angle similarity, set the classification angle, group and evaluate all trace segments, and determine the optimal number of groups and the number of trace lines they contain based on the principle of minimizing classification error;

[0052] Step 7: Perform manual scan line calibration on all traces and the traces in each individual group, and determine the optimal intra-group spacing and total spacing measurement values ​​based on the calibration results.

[0053] As a preference, the different states in step 1 include a stable state, a temporarily stable state and an unstable state.

[0054] The performance of the prediction model depends on the number and quality of samples, which is difficult to achieve in the complex and dangerous underground mining environment. Therefore, establishing a high-performance prediction model based on limited, general-quality images is the basis for ensuring trace spacing measurement. In order to obtain rich original trace feature images based on limited image data, and thus ensure a more accurate prediction model, the specific process of step 2 is as follows:

[0055] S21: Screening the original trace feature images on the pillar surface on site to remove images with excessive or insufficient exposure;

[0056] S22: Manually mark the trace features in each image, and have geological experts or other professional technicians repeatedly verify whether the markings are reasonable;

[0057] S23: If Figure 2 As shown in the figure, data enhancement technology is used to flip, move, rotate at a set angle, partially magnify and mirror each image to expand the number of original trace images and complete the construction of the image database;

[0058] S24: In order to avoid overfitting during the model training process, the image data in the database are distributed, and 80% of the image data are divided into a training set, and 20% of the image data are divided into a test set; at the same time, the trace and non-trace pixels of the images in the above two sets are distributed statistically. Specifically, the trace and non-trace pixel distributions of the images in the training set and the test set are calculated. If the difference in pixel distribution between the two sets meets the pixel distribution requirements, the data distribution result is deemed reasonable. Otherwise, the image data in the database are redistributed until the pixel distribution between the two sets meets the pixel distribution requirements.

[0059] In order to ensure the prediction accuracy, in step three, the total loss function is first constructed, and then the convolutional neural network is trained. During the training process, the preset accuracy index is used to evaluate whether the performance of the convolutional neural network after training meets the requirements.

[0060] Segmenting the traces at key locations so that they can truly and accurately describe the plane boundaries is a prerequisite for accurately calculating the spacing. In order to make the traces correspond one-to-one with the discontinuous planes, a segmentation algorithm is used to segment the traces. At the same time, in order to further enhance the accuracy of trace fractures, a fracture threshold is incorporated to evaluate the segmentation effect, so as to achieve accurate measurement of the trace spacing on the pillar surface. Therefore, the specific process of step five is as follows:

[0061] S51: confirm the number of pixels contained in each trace, and calculate the angle between three consecutive pixels starting from the starting point;

[0062] S52: Based on the angle breaking algorithm, different angle thresholds are set and compared with all the line segment angles obtained in S51 one by one to determine whether the trace needs to be broken. If the angle threshold is greater than the current line segment angle, the trace is broken at the middle point of the line segment angle and a trace breaking diagram is obtained. At the same time, the middle point is used as the starting point for calculating the next line segment angle value. Otherwise, the comparison of the next line segment angle is continued; until all traces are traversed and the corresponding trace breaking diagram is obtained;

[0063] S53: setting a fracture threshold, importing all trace segments obtained in S52, deleting trace segments whose length is less than the specified fracture threshold, and obtaining a new trace fracture map;

[0064] S54: A geological expert or other professional technician compares the distribution of the new trace segments obtained in S53 with the actual survey. If non-negligible trace segments are deleted, it is determined that the fracture threshold is over-selected, and S53 is re-executed to obtain a new trace fracture map by re-adjusting the fracture threshold. At the same time, by combining the adjustment of different angle thresholds, it is ensured that tiny traces that do not contribute significantly to the spacing measurement are not extracted too much, until the optimal angle threshold and fracture threshold are determined to complete the trace fracture evaluation and obtain the final trace fracture map.

[0065] Since the distribution of discontinuous surfaces is chaotic, the distance measurement of traces distributed in similar planes requires accurate classification of all trace segments in advance. However, most existing classification methods are developed based on the concept of clustering, which basically only focuses on the similarity between traces and ignores the discontinuity of traces. Therefore, errors in actual field applications are inevitable. In order to accurately and efficiently classify trace segments distributed in similar planes, Figure 4 As shown, the following grouping algorithm is used for processing. Therefore, the specific process of step six is ​​as follows:

[0066] S61: Determine the optimal classified angle value;

[0067] S62: importing the trace fracture map obtained after segmentation based on the optimal angle threshold and cropping based on the fracture threshold, and calculating the difference between the angles between any two trace segments in the map and the horizontal axis;

[0068] S63: grouping two trace segments whose calculated difference is less than the optimal classification angle value classified angle into the same group, and so on, traversing all trace segments, and determining the final number of trace groupings after the classification is completed.

[0069] In order to determine the optimal number of groups, a strategy based on classification error is adopted to determine the optimal number of groups. In step 6 S61, the classification error Error is first calculated according to formula (1), and then the optimal classification angle value classified angle is determined according to the principle of minimizing the classification error Error.

[0070]

[0071] Where θ i m is the angle of the mth trace segment in the i-th trace group; averθ i is the average angle of all trace segments in the i-th trace group, and I is the maximum number of trace groups.

[0072] In order to accurately and efficiently calculate the spacing information, such as Figure 5 As shown, in step seven, during the intra-group spacing measurement process, the scan line is made perpendicular to the majority of trace segments in the same group, that is, the scan line should be perpendicular to as many trace segments in the same group as possible. At the same time, during the total spacing measurement process, the scan line is made perpendicular to the majority of trace segments in all traces, that is, the number of trace segments that maintain a perpendicular position relationship with the scan line should be as large as possible.

[0073] As a preference, in step 2 S23, the angle is set to 45 degrees, 90 degrees, 180 degrees, or 270 degrees.

[0074] Example:

[0075] Figure 3 (a) is an example of a practical application. The image to be analyzed is a 1000*1000 pixel two-dimensional trace image taken from the surface of a stable pillar in an underground mine. Since the pillar is located in a large space where mining is being carried out, and the surface traces are densely distributed and there are too many short traces, manual measurement is more dangerous and the measurement accuracy does not meet the standard. Therefore, the intelligent measurement method of the present invention is used to measure the trace spacing.

[0076] First, the image to be analyzed is imported into the trained prediction model to obtain the distribution map of the predicted traces. It can be seen that the distribution of the predicted traces is continuous between different planes. Therefore, a break segmentation operation is performed on all the above predicted traces, and the angle threshold is set to 140 degrees. If the angle of the line segment formed by three pixels in any trace is less than the angle threshold, the trace is broken at the position of the middle pixel point, and then the pixel point is used as the starting point of the next line segment angle to be detected, and so on until the trace is broken. Figure 3 (b) To avoid the deletion of important trace segments and the excessive extraction of small traces without significant contribution, the break threshold is set to 2 pixels. If the length of the broken trace segment is less than the break threshold, the trace segment is deleted. By evaluating all trace segments in this way, the final trace break can be obtained. Figure 3 (c) Then, set the classification angle value to 12.5 degrees and calculate the angle difference between any two trace segments. If the angle difference is less than the classification angle, the two trace segments are grouped together. Repeat this operation until all trace segments are grouped. Calculate the grouping error and obtain the optimal trace grouping according to the minimum classification error principle. Figure 4 ,Finally, the scan lines are manually calibrated to calculate the ,intergroup spacing and the total spacing values ​​of the traces.

[0077] The present invention provides a semi-automated and interactive method for measuring the spacing of complex traces on the surface of hard rock pillars in underground mines. First, a large number of pillar surface feature images taken in underground stopes are collected and trained using a convolutional neural network based on a deep learning framework to obtain a prediction model for predicting the distribution of traces within the image. The prediction model can automatically capture trace information features from the target image to be analyzed and help reduce unnecessary noise interference. Second, the prediction model is used to predict the target image to be analyzed to obtain a predicted distribution map of the traces. This is then segmented using an angle-based fracture algorithm. Simultaneously, a fracture threshold is incorporated to delete trace segments that do not conform to the actual survey. The final new trace fracture map is obtained by adjusting the angle threshold and fracture threshold, and combining actual survey results and expert evaluation. Thus, the present invention effectively integrates the fracture threshold concept into the traditional angle-based fracture algorithm, thereby achieving accurate segmentation of trace segments. Accurately segmenting traces at key locations can help accurately reflect the plane boundary, thereby enabling one-to-one correspondence between traces and discontinuous planes, ensuring the subsequent accurate measurement of the spacing between traces on the pillar surface. Then, based on the principle of minimizing classification error, the optimal classification angle is determined, and a grouping algorithm based on the classification angle is used to divide the trace segments into different sets to obtain the optimal number of groups. Thus, by developing a new grouping algorithm based on the classification angle to rationally divide the trace segments and proposing an error minimization criterion to enhance the accuracy of trace grouping, this method can accurately and efficiently classify trace segments distributed within similar planes, while ensuring that trace discontinuities are not ignored, thereby ensuring that the spacing between traces distributed within similar planes can be accurately measured. Finally, manual scan line calibration is performed on all traces and the traces within each group to obtain the intra-group and total spacing values ​​of the traces.

[0078] This method has a simple implementation process, low implementation cost, high measurement efficiency and high measurement accuracy. It can realize the intelligent and automated measurement process of the trace spacing on the surface of hard rock pillars in underground mines, and can accurately and efficiently measure the trace spacing data on the surface of hard rock pillars. It can be beneficial to further improve the accuracy of pillar and stope stability evaluation, and can provide effective technical support for the stability evaluation of hard rock pillars in underground mines.

Claims

1. An intelligent method for measuring the distance between traces on the surface of hard rock pillars in an underground mine, which is characterized by comprising the following steps: Step 1: Collect original trace feature images of the surfaces of pillars in different states in a large number of underground stopes; Step 2: Preprocess the collected original trace feature images, reassemble the preprocessed images to complete the construction of the image database, and divide the image data in the image database into training set and test set according to the set ratio; Step 3: Build a convolutional neural network based on a deep learning framework, use the training set to train the convolutional neural network to obtain a high-precision convolutional neural network, and use the test set to test the performance of the high-precision convolutional neural network to obtain a prediction model for predicting the distribution of traces within the image; Step 4: Input the target image to be analyzed as input data into the prediction model, use the prediction model to predict the trace features of the target image, and obtain the predicted distribution map of the trace; Step 5: First, based on the angle difference principle, input the angle threshold and perform segmentation on all predicted traces to initially obtain several trace segments with a certain pixel length; Then, a fracture threshold is introduced to conduct a secondary test on the initially obtained trace segments to obtain the number of trace segments that is consistent with the actual survey results; Step 6: According to the principle of angle similarity, set the classification angle, group and evaluate all trace segments, and determine the optimal number of groups and the number of trace lines they contain based on the principle of minimizing classification error; Step 7: Perform manual scan line calibration on all traces and the traces in each individual group, and determine the optimal intra-group spacing and total spacing measurement values ​​based on the calibration results.

2. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 1, characterized in that: The different states in step 1 include a stable state, a temporarily stable state, and an unstable state.

3. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 1, characterized in that: The specific process of step 2 is as follows: S21: Screening the original trace feature images on the pillar surface on site to remove images with excessive or insufficient exposure; S22: Manually mark the trace features in each image, and have geological experts or other professional technicians repeatedly verify whether the markings are reasonable; S23: Using data augmentation technology, each image is flipped, moved, rotated at a set angle, partially enlarged, and mirrored to expand the number of original trace images, thereby completing the construction of the image database; S24: Allocate the image data in the database, and divide 80% of the image data into a training set, and divide 20% of the image data into a test set; at the same time, calculate the trace and non-trace pixel distribution of the images in the training set and the test set. If the pixel distribution difference between the two sets meets the pixel distribution requirements, the data allocation result is considered reasonable. Otherwise, the image data in the database is reallocated until the pixel distribution between the two sets meets the pixel distribution requirements.

4. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 1 or 2, characterized in that: In step three, the total loss function is first constructed, and then the convolutional neural network is trained. During the training process, the preset accuracy index is used to evaluate whether the performance of the convolutional neural network after training meets the requirements.

5. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 1, characterized in that: The specific process of step five is as follows: S51: confirm the number of pixels contained in each trace, and calculate the angle between three consecutive pixels starting from the starting point; S52: Based on the angle breaking algorithm, different angle thresholds are set and compared with all the line segment angles obtained in S51 one by one to determine whether the trace needs to be broken. If the angle threshold is greater than the current line segment angle, the trace is broken at the middle point of the line segment angle and a trace breaking diagram is obtained. At the same time, the middle point is used as the starting point for calculating the next line segment angle value. Otherwise, the comparison of the next line segment angle is continued; until all traces are traversed and the corresponding trace breaking diagram is obtained; S53: setting a fracture threshold, importing all trace segments obtained in S52, deleting trace segments whose length is less than the specified fracture threshold, and obtaining a new trace fracture map; S54: A geological expert or other professional technician compares the distribution of the new trace segments obtained in S53 with the actual survey. If non-negligible trace segments are deleted, it is determined that the fracture threshold is over-selected, and S53 is re-executed to obtain a new trace fracture map by re-adjusting the fracture threshold. At the same time, by combining the adjustment of different angle thresholds, it is ensured that tiny traces that do not contribute significantly to the spacing measurement are not extracted too much, until the optimal angle threshold and fracture threshold are determined to complete the trace fracture evaluation and obtain the final trace fracture map.

6. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 5, characterized in that: The specific process of step six is ​​as follows: S61: Determine the optimal classified angle value; S62: importing the trace fracture map obtained after segmentation based on the optimal angle threshold and cropping based on the fracture threshold, and calculating the difference between the angles between any two trace segments in the map and the horizontal axis; S63: grouping two trace segments whose calculated difference is less than the optimal classification angle value classified angle into the same group, and so on, traversing all trace segments, and determining the final number of trace groupings after the classification is completed.

7. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 6, characterized in that: In step 6 S61, the classification error Error is first calculated according to formula (1), and then the optimal classification angle value classified angle is determined according to the principle of minimizing the classification error Error; （1）； Where, For the i The first trace in the group m Trace segment angle; For the i The average angle of all trace segments in a trace group, I is the maximum number of trace groups.

8. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 6, characterized in that: In step seven, during the intra-group spacing measurement, the scan line is made perpendicular to most trace segments within the same group. Meanwhile, during the total spacing measurement, the scan line is made perpendicular to most trace segments in all traces.

9. The intelligent measurement method for the surface trace spacing of hard rock pillars in underground mines according to claim 3, characterized in that: In step 2 S23 , the angle is set to 45 degrees, 90 degrees, 180 degrees, or 270 degrees.

Images