Mine step crack identification method and system based on dynamic feature extraction network

Abstract

The application discloses a kind of mine step crack identification method and system based on dynamic feature extraction network, belong to mine safety monitoring technical field, its method includes the following steps: S1, obtains mine step RGB image and rainfall prior information;S2, calculate the expected smoothness coefficient of crack surface;S3, construct deformation constraint vector, and calculate smooth loss;S4, input mine step RGB image to encoder, obtain multi-scale feature map;S5, input multi-scale feature map to decoder, generate crack probability graph;S6, crack probability graph is carried out adaptive threshold binaryzation and morphological post-processing, obtain crack identification result.The application adopts morphological closing operation and connected domain rejection, effectively filters out the interference generated after rain, outputs relatively accurate crack identification result.Can be seamlessly embedded in existing intelligent mine monitoring platform, has very high industrialization popularization value.

Description

Technical Field

[0001] This invention belongs to the field of mine safety monitoring technology, specifically relating to a method and system for identifying mine bench cracks based on a dynamic feature extraction network. Background Technology

[0002] Early identification of cracks on mine bench slopes is crucial for preventing geological disasters such as landslides and collapses. Traditional manual inspection methods are inefficient, subjective, and require inspectors to climb high-risk areas, with a high risk of falls on wet, slippery surfaces after rain. Missing crack detection can trigger large-scale landslides and collapses, causing significant losses such as equipment damage and casualties. In recent years, deep learning-based crack identification methods (such as U-Net, DeepLabV3+, and YOLOv5) have made significant progress, but they still face two major challenges in real-world mining environments. Firstly, cracks are often long, thin, and curved, making it difficult for conventional square convolution kernels to cover their discontinuous, high-curvature morphology, leading to multiple breaks in the identification results. These breaks are easily misjudged as background noise, thus missing early landslide warnings. Secondly, after rain, the crack surface becomes smooth due to moisture filling, reducing the contrast with the background rock. Residual rainwater can also obscure parts of the crack area, significantly reducing the model's accuracy. However, the period after rain is precisely when the risk of landslides is highest, and inaccurate identification results can cause serious safety hazards. Therefore, the industry urgently needs a recognition method that can integrate prior information about rainwater with deep learning to solve the above problems and avoid major safety accidents and economic losses. Summary of the Invention

[0003] To address the above problems, this invention proposes a method and system for identifying cracks in mine benches based on a dynamic feature extraction network.

[0004] The technical solution of this invention is: a method for identifying cracks in mine benches based on a dynamic feature extraction network, comprising the following steps:

[0005] S1. Obtain RGB images of the mine steps and prior rainfall information;

[0006] S2. Calculate the expected smoothness coefficient of the crack surface based on prior rainfall information;

[0007] S3. Based on the expected smoothness coefficient of the crack surface, construct the deformation constraint vector and calculate the smoothing loss;

[0008] S4. Based on smoothing loss, input the RGB image of the mine steps into the encoder to obtain multi-scale feature maps;

[0009] S5. Input the multi-scale feature map into the decoder to generate a crack probability map;

[0010] S6. Adaptive threshold binarization and morphological post-processing are performed on the crack probability map to obtain the crack identification result.

[0011] Furthermore, S1 includes the following sub-steps:

[0012] S11. Acquire RGB images of the mine steps, read the rainfall from the rain sensor, and use the RGB images of the mine steps to calculate the global average brightness and local texture entropy.

[0013] S12. Based on the global average brightness and local texture entropy, determine whether the rainfall meets the rainfall stop judgment condition. If yes, proceed to S13; otherwise, return to S11.

[0014] S13. Set the rainfall stop flag to 1 and record the cumulative duration after the rainfall stop state begins;

[0015] S14. Use the rainfall stop sign and the cumulative duration after the start of the rainfall stop state as prior information about rainfall;

[0016] In S12, the specific conditions for determining whether rainfall stops are: the duration of the rain stoppage after the rainfall changes from greater than zero to zero exceeds the threshold duration, the global average brightness is higher than the preset brightness threshold, and the local texture entropy is lower than the preset texture threshold.

[0017] Furthermore, the expected smoothness coefficient of the crack surface The expression is:

[0018] ;

[0019] in, The initial smoothness coefficient after rain. Here is the drying rate constant. This is the cumulative duration since the start of the rainfall cessation state. It is an index.

[0020] Furthermore, S3 includes the following sub-steps:

[0021] S31. Initialize the number of sampling points to 1, and proceed to S32;

[0022] S32. Based on the expected smoothness coefficient of the crack surface, calculate the deformation constraint vector elements corresponding to the sampling points in the dynamic serpentine convolution.

[0023] S33. Increase the number of sampling points by 1;

[0024] S34. Determine whether the current number of sampling points is less than or equal to the total number of sampling points of the dynamic snake convolution. If yes, return to S33; otherwise, proceed to S35.

[0025] S35. Calculate the smoothing loss.

[0026] Furthermore, in S32, the first Deformation constraint vector elements corresponding to each sampling point The expression is:

[0027] ;

[0028] in, This is the index of the starting sampling point of the convolution kernel. The expected smoothness coefficient of the crack surface. This represents the total number of sampling points for the dynamic serpentine convolution.

[0029] In S35, smoothing loss The expression is:

[0030] ;

[0031] in, Denotes the operator 2-norm. This is the offset of the k-th sampling point. This is the offset of the (k+1)th sampling point.

[0032] Furthermore, the total loss of the encoder The expression is:

[0033] ;

[0034] ;

[0035] in, For balance coefficient, Indicates the loss from partitioning. Indicates smoothing loss. The total number of pixels in the image. For the network prediction of the first The probability that a pixel belongs to a crack. For the first The real label of each pixel It is a very small constant.

[0036] Furthermore, S6 includes the following sub-steps:

[0037] S61. Perform grayscale histogram statistics on the crack probability map, calculate the bimodal distribution of probability values, and use the probability value corresponding to the valley between the two peaks as the segmentation threshold to divide the crack probability. Figure 2 Value-enhanced, resulting in a binary image;

[0038] S62. Perform morphological closing operation on the binary image to obtain the binary image after fracture repair.

[0039] S63. Remove connected components with an area smaller than a preset threshold in the binary image after fracture repair, and output the crack identification result.

[0040] Furthermore, in S62, the morphological closing operation includes: performing a dilation operation on the binary image after fracture repair based on the size of the structural element, and performing an erosion operation on the dilation result using the same structural element.

[0041] Furthermore, S63 includes the following sub-steps:

[0042] S631. Identify and mark all independent pixel sets in the binary image after the fracture repair according to the eight-connected domain criterion, traverse all pixel positions of the independent pixel sets and accumulate the count to obtain the total number of pixels in the independent pixel sets.

[0043] S632. Compare the total number of pixels in the independent pixel set with a preset threshold. If the preset threshold is not reached, clear the pixels in the area where the independent pixel set is located. Otherwise, retain the pixels in the independent pixel set to obtain the crack identification result.

[0044] Based on the above methods, this invention also proposes a mine bench crack identification system based on a dynamic feature extraction network, comprising:

[0045] The image acquisition module is used to acquire RGB images of the mine steps;

[0046] The rainfall monitoring module is used to read the rainfall data from the rain gauge sensor.

[0047] The rainfall prior determination module is used to determine rainfall prior information;

[0048] The smoothness calculation module is used to calculate the expected smoothness coefficient of the crack surface based on prior rainfall information;

[0049] The dynamic feature extraction module is used to construct a deformation constraint vector based on the expected smoothness coefficient of the crack surface and calculate the smoothing loss; based on the smoothing loss, the RGB image of the mine steps is input into the encoder to obtain a multi-scale feature map; the multi-scale feature map is input into the decoder to generate a crack probability map.

[0050] The post-processing module is used to perform adaptive threshold binarization and morphological post-processing on the crack probability map to obtain the crack identification result.

[0051] The beneficial effects of this invention are:

[0052] 1. This invention combines rainfall sensor data with image brightness and texture entropy, and quantifies the smoothness of crack surfaces to provide prior guidance for dynamic feature extraction, thereby effectively extracting crack features caused by rain.

[0053] 2. During the window period before a landslide occurs, this invention can continuously and reliably monitor crack evolution, buying valuable time for personnel evacuation and equipment relocation in the mine. The deformation constraint vector, constructed based on a dynamic convolution kernel using a smoothness coefficient, results in a more compact and smooth arrangement of sampling points, significantly improving the ability to continuously identify curved and slender cracks after rain.

[0054] 3. This invention introduces a smoothing loss based on the offset of adjacent sampling points to prevent deformable convolution divergence and enhance the stability of feature extraction. Morphological closing operations and connected component removal are employed to effectively filter out interference generated after rain, outputting more accurate crack identification results. It can be seamlessly embedded into existing smart mine monitoring platforms and has extremely high industrialization and promotion value. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method for identifying cracks in mine benches based on a dynamic feature extraction network.

[0056] Figure 2 This is a schematic diagram of the dynamic serpentine convolution sampling point offset in this invention. Detailed Implementation

[0057] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, this invention provides a method for identifying cracks in mine benches based on a dynamic feature extraction network, comprising the following steps:

[0059] S1. Obtain RGB images of the mine steps and prior rainfall information;

[0060] S2. Calculate the expected smoothness coefficient of the crack surface based on prior rainfall information;

[0061] S3. Based on the expected smoothness coefficient of the crack surface, construct the deformation constraint vector and calculate the smoothing loss;

[0062] S4. Based on smoothing loss, input the RGB image of the mine steps into the encoder to obtain multi-scale feature maps;

[0063] S5. Input the multi-scale feature map into the decoder to generate a crack probability map;

[0064] S6. Adaptive threshold binarization and morphological post-processing are performed on the crack probability map to obtain the crack identification result.

[0065] In this embodiment of the invention, S1 includes the following sub-steps:

[0066] S11. Acquire RGB images of the mine steps, read the rainfall from the rain sensor, and use the RGB images of the mine steps to calculate the global average brightness and local texture entropy.

[0067] S12. Based on the global average brightness and local texture entropy, determine whether the rainfall meets the rainfall stop judgment condition. If yes, proceed to S13; otherwise, return to S11.

[0068] S13. Set the rainfall stop flag to 1 and record the cumulative duration after the rainfall stop state begins;

[0069] S14. Use the rainfall stop sign and the cumulative duration after the start of the rainfall stop state as prior information about rainfall;

[0070] In S12, the specific conditions for determining whether rainfall stops are: the duration of the rain stoppage after the rainfall changes from greater than zero to zero exceeds the threshold duration, the global average brightness is higher than the preset brightness threshold, and the local texture entropy is lower than the preset texture threshold.

[0071] Cumulative duration The unit is hours. The threshold duration, brightness threshold, and texture threshold are set based on engineering experience.

[0072] In this embodiment of the invention, the expected smoothness coefficient of the crack surface The expression is:

[0073] ;

[0074] in, The initial smoothness coefficient after rain. Here is the drying rate constant. This is the cumulative duration since the start of the rainfall cessation state. It is an index.

[0075] and The correlation between water content and smoothness was pre-calibrated using rock samples from the surface of mine steps. The cumulative duration after the start of the rainfall cessation state was also determined. Read from the timer.

[0076] In this embodiment of the invention, S3 includes the following sub-steps:

[0077] S31. Initialize the number of sampling points to 1, and proceed to S32;

[0078] S32. Based on the expected smoothness coefficient of the crack surface, calculate the deformation constraint vector elements corresponding to the sampling points in the dynamic serpentine convolution.

[0079] S33. Increase the number of sampling points by 1;

[0080] S34. Determine whether the current number of sampling points is less than or equal to the total number of sampling points of the dynamic snake convolution. If yes, return to S33; otherwise, proceed to S35.

[0081] S35. Calculate the smoothing loss.

[0082] In this embodiment of the invention, in S32, the first... Deformation constraint vector elements corresponding to each sampling point The expression is:

[0083] ;

[0084] in, This is the index of the starting sampling point of the convolution kernel. The expected smoothness coefficient of the crack surface. This represents the total number of sampling points for the dynamic serpentine convolution.

[0085] The dimension of the deformation constraint vector is determined by the size of the convolution kernel. Obtain, that is ; and These are parameters set based on engineering experience.

[0086] In S35, smoothing loss The expression is:

[0087] ;

[0088] in, Denotes the operator 2-norm. This is the offset of the k-th sampling point. This is the offset of the (k+1)th sampling point.

[0089] The sampling point offset is generated based on the multi-scale feature map and the encoder's convolutional layer.

[0090] In this embodiment of the invention, the total loss of the encoder The expression is:

[0091] ;

[0092] ;

[0093] in, For balance coefficient, Indicates the loss from partitioning. Indicates smoothing loss. The total number of pixels in the image. For the network prediction of the first The probability that a pixel belongs to a crack. For the first The real label of each pixel It is a very small constant.

[0094] The encoder has Each convolutional layer employs dynamic serpentine convolution, where... Based on engineering experience.

[0095] Balance coefficient To enable crack identification on the first verification set The highest score value;

[0096] The F1 score, the harmonic mean of precision and recall, is used to measure crack identification performance. Its calculation formula is as follows: ;

[0097] Precision rate refers to the proportion of pixels predicted as cracks that are actually crack pixels.

[0098] Recall rate refers to the proportion of real crack pixels that are correctly predicted;

[0099] The validation set is obtained by dividing the dataset according to a predetermined proportion based on engineering experience;

[0100] The segmentation loss is used to measure the difference between the crack probability map predicted by the network and the actual crack labels annotated manually.

[0101] In this embodiment of the invention, S6 includes the following sub-steps:

[0102] S61. Perform grayscale histogram statistics on the crack probability map, calculate the bimodal distribution of probability values, and use the probability value corresponding to the valley between the two peaks as the segmentation threshold to divide the crack probability. Figure 2 Value-enhanced, resulting in a binary image;

[0103] S62. Perform morphological closing operation on the binary image to obtain the binary image after fracture repair.

[0104] S63. Remove connected components with an area smaller than a preset threshold in the binary image after fracture repair, and output the crack identification result.

[0105] The preset threshold is set based on engineering experience.

[0106] In this embodiment of the invention, in S62, the morphological closing operation includes: performing a dilation operation on the binary image after fracture repair based on the size of the structural element, and performing an erosion operation on the dilation result using the same structural element.

[0107] In this embodiment of the invention, S63 includes the following sub-steps:

[0108] S631. Identify and mark all independent pixel sets in the binary image after the fracture repair according to the eight-connected domain criterion, traverse all pixel positions of the independent pixel sets and accumulate the count to obtain the total number of pixels in the independent pixel sets.

[0109] S632. Compare the total number of pixels in the independent pixel set with a preset threshold. If the preset threshold is not reached, clear the pixels in the area where the independent pixel set is located. Otherwise, retain the pixels in the independent pixel set to obtain the crack identification result.

[0110] Based on the above methods, this invention also proposes a mine bench crack identification system based on a dynamic feature extraction network, comprising:

[0111] The image acquisition module is used to acquire RGB images of the mine steps;

[0112] The rainfall monitoring module is used to read the rainfall data from the rain gauge sensor.

[0113] The rainfall prior determination module is used to determine rainfall prior information;

[0114] The smoothness calculation module is used to calculate the expected smoothness coefficient of the crack surface based on prior rainfall information;

[0115] The dynamic feature extraction module is used to construct a deformation constraint vector based on the expected smoothness coefficient of the crack surface and calculate the smoothing loss; based on the smoothing loss, the RGB image of the mine steps is input into the encoder to obtain a multi-scale feature map; the multi-scale feature map is input into the decoder to generate a crack probability map.

[0116] The post-processing module is used to perform adaptive threshold binarization and morphological post-processing on the crack probability map to obtain the crack identification result.

[0117] The following description is based on specific embodiments.

[0118] Scenario: A platform in an open-pit iron mine. The rain stopped at 4:00 AM. A drone inspection was conducted at 7:00 AM, three hours after the rain stopped (i.e., (hours). The lithology of the steps is granite, and the parameters were pre-calibrated. Drying rate constant .

[0119] S1. Obtain prior information about rainfall.

[0120] S11. The rain sensor shows that the current rainfall is 0, and the duration of the rain stop has exceeded the threshold duration;

[0121] S12. Acquire RGB image, calculate global average brightness as 88 (threshold 80), local texture entropy as 3.2 (threshold 3.5), which meets the condition for stopping rainfall;

[0122] S13, the timer reading is Hourly, the rain stop sign is set to 1.

[0123] S2. Calculate the smoothness coefficient.

[0124] Expected smoothness coefficient of crack surface The level is above average, indicating that the surface of the crack is still relatively smooth but has begun to dry partially.

[0125] S3. Construct the deformation constraint vector and smoothing loss.

[0126] S31. Assume the model uses 3×3 dynamic serpentine convolution (i.e. , ), calculate the first Deformation constraint vector elements corresponding to each sampling point ;

[0127] S32. Calculate the corresponding smoothing loss: ,in, This is the offset of the k-th sampling point. The offset of the (k+1)th sampling point is generated based on the first multi-scale feature map and the convolutional layer of the encoder.

[0128] S4-S5: Encoder-Decoder Processing

[0129] S41. Replace all 3×3 convolutional kernels in the network with dynamic serpentine convolutional kernels;

[0130] S42. Input the RGB image into the network, and process it through the encoder and decoder to obtain the crack probability map.

[0131] S6, Post-processing.

[0132] S61. Perform grayscale histogram statistics on the crack probability map, and take the probability value corresponding to the valley between the two peaks (0.45) as the segmentation threshold to obtain the binary map. .

[0133] S62, to Perform morphological closing operations to obtain a binary image after fracture repair. .

[0134] S63. Identify and label according to the eight-connected-domain criterion. The system calculates the total number of pixels in each connected region from all independent pixel sets, removes isolated noise points with an area smaller than a second preset threshold, retains the remaining connected regions, and outputs the final crack recognition result.

[0135] like Figure 2As shown in the figure, a 3×3 convolution kernel is used. The initial sampling points (gray dots) are arranged along the regular grid, while the sampling points (red dots) after dynamic offset are arranged in a "serpentine" pattern along the crack direction. The blue arrows indicate the offset of each sampling point, and the red dashed lines indicate the "serpentine" path formed by sequentially connecting the offset sampling points.

[0136] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying cracks in mine benches based on a dynamic feature extraction network, characterized in that, Includes the following steps: S1. Obtain RGB images of the mine steps and prior rainfall information; S2. Calculate the expected smoothness coefficient of the crack surface based on prior rainfall information; S3. Based on the expected smoothness coefficient of the crack surface, construct the deformation constraint vector and calculate the smoothing loss; S4. Based on smoothing loss, input the RGB image of the mine steps into the encoder to obtain multi-scale feature maps; S5. Input the multi-scale feature map into the decoder to generate a crack probability map; S6. Perform adaptive threshold binarization and morphological post-processing on the crack probability map to obtain the crack identification result; The rainfall cessation sign and the cumulative duration after the start of the rainfall cessation state are used as prior information about rainfall. S3 includes the following sub-steps: S31. Initialize the number of sampling points to 1, and proceed to S32; S32. Based on the expected smoothness coefficient of the crack surface, calculate the deformation constraint vector elements corresponding to the sampling points in the dynamic serpentine convolution. S33. Increase the number of sampling points by 1; S34. Determine whether the current number of sampling points is less than or equal to the total number of sampling points of the dynamic snake convolution. If yes, return to S32; otherwise, proceed to S35. S35. Calculate the smoothing loss; In S32, the first Deformation constraint vector elements corresponding to each sampling point The expression is: ； in, This is the index of the starting sampling point of the convolution kernel. The expected smoothness coefficient of the crack surface. This represents the total number of sampling points for the dynamic serpentine convolution. In S35, the smoothing loss The expression is: ； in, Denotes the operator 2-norm. This is the offset of the k-th sampling point. This is the offset of the (k+1)th sampling point.

2. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Acquire RGB images of the mine steps, read the rainfall from the rain sensor, and use the RGB images of the mine steps to calculate the global average brightness and local texture entropy. S12. Based on the global average brightness and local texture entropy, determine whether the rainfall meets the rainfall stop judgment condition. If yes, proceed to S13; otherwise, return to S11. S13. Set the rainfall stop flag to 1 and record the cumulative duration after the rainfall stop state begins; S14. Use the rainfall stop sign and the cumulative duration after the start of the rainfall stop state as prior information about rainfall; In S12, the specific conditions for determining whether rainfall stops are: the duration of rain stoppage after the rainfall changes from greater than zero to zero exceeds a threshold duration, the global average brightness is higher than a preset brightness threshold, and the local texture entropy is lower than a preset texture threshold.

3. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 1, characterized in that, The expected smoothness coefficient of the crack surface The expression is: ； in, The initial smoothness coefficient after rain. Here is the drying rate constant. This is the cumulative duration since the start of the rainfall cessation state. It is an index.

4. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 1, characterized in that, The total loss of the encoder The expression is: ； ； in, For balance coefficient, Indicates the loss from partitioning. Indicates smoothing loss. The total number of pixels in the image. For the network prediction of the first The probability that a pixel belongs to a crack. For the first The real label of each pixel It is a very small constant.

5. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 1, characterized in that, S6 includes the following sub-steps: S61. Perform grayscale histogram statistics on the crack probability map, calculate the bimodal distribution of probability values, and use the probability value corresponding to the valley between the two peaks as the segmentation threshold to binarize the crack probability map and obtain a binary map. S62. Perform morphological closing operation on the binary image to obtain the binary image after fracture repair. S63. Remove connected components with an area smaller than a preset threshold in the binary image after fracture repair, and output the crack identification result.

6. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 5, characterized in that, In S62, the morphological closing operation includes: performing a dilation operation on the binary image after fracture repair based on the size of the structural element, and performing an erosion operation on the dilation result using the same structural element.

7. The method for identifying mine bench cracks based on a dynamic feature extraction network according to claim 5, characterized in that, S63 includes the following sub-steps: S631. Identify and mark all independent pixel sets in the binary image after the fracture repair according to the eight-connected domain criterion, traverse all pixel positions of the independent pixel sets and accumulate the count to obtain the total number of pixels in the independent pixel sets. S632. Compare the total number of pixels in the independent pixel set with a preset threshold. If the preset threshold is not reached, clear the pixels in the area where the independent pixel set is located. Otherwise, retain the pixels in the independent pixel set to obtain the crack identification result.

8. A mine bench crack identification system based on a dynamic feature extraction network, characterized in that, The mine bench crack identification system based on a dynamic feature extraction network is used to implement the mine bench crack identification method based on a dynamic feature extraction network as described in any one of claims 1-7; the system includes: The image acquisition module is used to acquire RGB images of the mine steps; The rainfall monitoring module is used to read the rainfall data from the rain gauge sensor. The rainfall prior determination module is used to determine rainfall prior information; The smoothness calculation module is used to calculate the expected smoothness coefficient of the crack surface based on prior rainfall information; The dynamic feature extraction module is used to construct a deformation constraint vector based on the expected smoothness coefficient of the crack surface and calculate the smoothing loss; based on the smoothing loss, the RGB image of the mine steps is input into the encoder to obtain a multi-scale feature map; the multi-scale feature map is input into the decoder to generate a crack probability map. The post-processing module is used to perform adaptive threshold binarization and morphological post-processing on the crack probability map to obtain the crack identification result.

Images