Underground surrounding rock joint fissure intelligent identification method based on direction perception deep learning

Abstract

The invention discloses an underground surrounding rock joint fissure intelligent identification method based on direction perception deep learning, and the method comprises the steps: collecting and preprocessing underground mine roadway roof and side images, and obtaining a standardized surrounding rock image; labeling joint fissures in the standardized surrounding rock image, and constructing a training sample set; performing direction disturbance enhancement processing on the training sample set to generate an enhanced sample set; constructing a deep learning model with a direction perception capability, and setting a feature extraction structure extending along different directions in a multi-scale feature extraction stage; the constructed enhanced sample set is utilized to train the constructed direction perception deep learning model, model parameter updating is completed based on direction perception feature expression, and a high-performance joint fissure recognition model is obtained; and inputting a to-be-identified underground mine roadway surrounding rock image into the trained joint fissure identification model, and outputting an accurate surrounding rock joint fissure identification result. According to the method, a more continuous, stable and accurate surrounding rock joint fissure prediction result is output.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering information perception and intelligent analysis technology, specifically involving an intelligent identification method for joints and fissures in underground surrounding rock based on direction-aware deep learning. Background Technology

[0002] Joints and fissures are widely developed on the surface of the roof and sidewalls of underground mine roadways. These fissures often exhibit a long, continuous linear or banded distribution and maintain a relatively stable orientation within a specific spatial range. The spatial distribution pattern, extension direction, and continuity of joints and fissures are core characteristics that form the basis for analyzing the structural features of the surrounding rock, evaluating its stability, and designing roadway support schemes. They directly affect the safety and economy of underground mining operations.

[0003] In practical engineering applications, the acquisition of information on joints and fissures in surrounding rock still relies primarily on traditional manual methods, including visual inspection, on-site hand-drawn sketches, and manual interpretation and annotation of surrounding rock images. However, underground mine tunnels are characterized by narrow working spaces, unstable lighting conditions, and significant interference from complex environmental factors such as dust cover and reflective wetness on the surrounding rock surface. Traditional manual acquisition methods suffer from inherent defects such as low work efficiency, poor consistency of identification results, and strong subjectivity, making it difficult to meet the practical engineering needs for stable, continuous, and accurate identification of joints and fissures in long-distance, multi-section scenarios in underground mine tunnels.

[0004] To improve the automation level of rock joint and fracture identification, image processing-based fracture identification methods have been gradually promoted and applied in engineering practice. However, these methods still have obvious technical limitations: in scenarios with complex rock surface texture, significant fracture width fluctuations, low contrast between fractures and rock background, or strong noise interference, fracture identification is prone to breakage, false edge misjudgment, and the same fracture being segmented into multiple segments. It is difficult to form a stable and reliable linear feature representation of fractures, and it cannot provide accurate data support for subsequent engineering applications.

[0005] With the rapid development of deep learning technology, semantic segmentation methods based on convolutional neural networks have been gradually introduced into the field of rock joint and fracture identification. This has reduced the reliance on manually designed features and threshold parameters to a certain extent and improved the automation of identification. However, underground rock joints and fractures have significant directional and continuous characteristics, and most of them are typical slender structures. Existing deep learning methods mostly adopt isotropic convolution and multi-scale feature fusion modes in feature extraction and decoding reconstruction, lacking the ability to explicitly model the fracture orientation features.

[0006] Specifically, in the multi-scale feature extraction stage, network activation responses often exhibit local point-like or block-like distributions, making it difficult to form continuous feature responses along the fracture extension direction. During decoding and upsampling, fracture predictions often start from local high-confidence regions and gradually expand and recover, resulting in a discrete point-segment structure of fractures in the early prediction stages. This leads to problems such as increased fracture identification breaks, insufficient continuity, and poor stability in identifying fractures in the same direction. Based on these technical bottlenecks, existing technologies for identifying joint fractures in underground surrounding rock generally suffer from insufficient characterization of fracture direction continuity, making it difficult to simultaneously address the requirements of multi-scale fracture feature expression and directional consistency. Currently, there is a lack of an intelligent identification method that can introduce a direction-aware mechanism in the feature extraction stage and effectively maintain the linear continuous expression of fractures in the decoding stage, thus failing to fully meet the practical engineering needs for accurate joint fracture identification.

[0007] Therefore, there is an urgent need to provide an intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides an intelligent identification method for joints and fissures in underground surrounding rock based on direction-aware deep learning. This method can output more continuous, stable, and accurate prediction results of joints and fissures in surrounding rock, and can effectively solve the problems of poor continuity of fissure identification, insufficient characterization of directional features, and weak anti-interference ability in the prior art.

[0009] To achieve the above objectives, this invention provides an intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning, comprising the following steps: Step 1: Collect raw surrounding rock image data of the roof and sidewalls of the underground mine roadway, and perform preprocessing operations on the raw surrounding rock images to obtain standardized surrounding rock images that meet the requirements for joint and fracture identification; Step 2: Label the joints and fractures in the standardized surrounding rock images to construct a training sample set containing the surrounding rock images and corresponding joint and fracture labels; Step 3: Perform directional perturbation enhancement processing on the training sample set to generate an enhanced sample set containing various joint and fracture orientation distributions; Step 4: Construct a deep learning model with orientation awareness, and set up feature extraction structures that extend along different directions in the multi-scale feature extraction stage to achieve effective expression of orientation awareness features of joints and fractures; Step 5: Using the constructed enhanced sample set, train the constructed direction-aware deep learning model, update the model parameters based on the direction-aware feature representation, and finally obtain a high-performance joint and fracture recognition model; Step 6: Input the image of the surrounding rock of the underground mine roadway to be identified into the trained joint and fracture identification model, and output accurate identification results of the surrounding rock joints and fractures through model inference.

[0010] Furthermore, in order to provide reliable data support for the model to learn the continuous features of joint and fracture orientation, the process of constructing a training sample set containing surrounding rock images and corresponding joint and fracture labels in step 2 is as follows: S21: Pixel-by-pixel annotation; In the standardized surrounding rock image, pixel-by-pixel annotation is performed on the joint and fracture area to generate a joint and fracture label image with the same size as the original surrounding rock image, and to ensure the spatial correspondence between the label and the image. S22: Connectivity Part Sorting: Connectivity part analysis and sorting are performed on the generated joint and fissure label images to remove discrete noise pixels, making the joints and fissures form a continuous linear structure and improving the effectiveness of the labels. S23: Sample pairing: The sorted joint and fracture annotation results are paired one-to-one with the corresponding standardized surrounding rock images to construct a training sample set for learning the continuous features of joint and fracture direction.

[0011] Furthermore, in order to construct a high-quality augmented sample set that meets the training requirements of the model's orientation awareness, the process of generating an augmented sample set containing various joint and fracture orientation distributions in step 3 is as follows: S31: Rotation Enhancement: Rotate the surrounding rock images in the training sample set to enhance them by rotating the images around their geometric center by 90°, 180° and 270° respectively, to generate rotation-enhanced samples containing different joint and fracture orientations. S32: Flip Enhancement: Perform horizontal and vertical flipping processes sequentially on the rotation-enhanced sample to further expand the directional distribution range of joints and fractures, generating flipped enhanced samples; S33: Label Synchronization Transformation: Perform geometric transformations on the joint and fracture labels corresponding to the rotation-enhanced samples and the flip-enhanced samples, ensuring that the spatial correspondence between the labels and the images is not destroyed, and finally construct an enhanced training sample set after directional perturbation enhancement.

[0012] Furthermore, in order to form a high-quality orientation-aware feature representation that combines directional continuity and multi-scale adaptability, and to provide core support for accurate identification by subsequent models, the effective representation process of joint fracture orientation-aware features in step 4 is as follows: S41: Multi-scale feature mapping; Based on the convolutional neural network framework, high-level features are extracted from the input standardized surrounding rock image, and multi-scale feature mapping operation is performed on the high-level features to obtain feature representations at different receptive field scales; S42: Horizontal feature extraction; During the multi-scale feature mapping process, a directional extension feature extraction operation is performed on the high-level features along the horizontal direction to generate a directional feature response that is continuously distributed along the horizontal direction, capturing the continuous features of the horizontally oriented crack. S43: Vertical feature extraction; Simultaneously, during the multi-scale feature mapping process, a directional extension feature extraction operation is performed on the high-level features along the vertical direction to generate a directional feature response that is continuously distributed along the vertical direction, capturing the continuous features of vertically oriented cracks; S44: Directional feature fusion; The horizontal and vertical directional feature responses are spatially aligned and fused with the isotropic features at the corresponding scales to form a directional perception feature expression that combines directional continuity and multi-scale adaptability.

[0013] Furthermore, to ensure the coordinated operation of each layer of the model and effectively improve the direction perception capability and recognition accuracy, the process of obtaining the high-performance joint and fracture recognition model in step 5 is as follows: S51: Feature forward propagation; Input the surrounding rock images in the orientation-aware deep learning model from the orientation perturbation enhancement training sample set, and generate orientation-aware feature representations through the model's forward propagation process; S52: Supervision information matching; The obtained joint and fracture label images are used as supervision information and matched with the direction-aware feature expressions generated by the model to provide supervision for model parameter updates; S53: Joint parameter update; through the backpropagation mechanism, the parameters of the orientation-aware feature extraction layer, multi-scale feature extraction layer and decoding layer in the model are jointly updated and optimized to ensure that the various layers of the model work together and improve orientation awareness and recognition accuracy. S54: Model solidification; After the model completes the preset training rounds and the training loss tends to stabilize and reach the preset threshold, all parameters of the model are fixed to obtain the joint and fracture recognition model trained based on direction-aware feature expression.

[0014] Furthermore, in order to accurately output the quantified identification results of surrounding rock joints and fractures, the process of outputting accurate identification results of surrounding rock joints and fractures through model inference in step 6 is as follows: S61: Image reasoning to be identified; Input the image of the surrounding rock of the underground mine roadway to be identified into the obtained direction-aware deep learning model, perform the forward propagation operation of the model, and generate direction-aware feature representation; S62: Pixel-level response generation; using orientation-aware feature representation as input features in the decoding stage, the image spatial resolution is restored step by step during the decoding process to generate pixel-level predicted responses for joints and fissures; S63: Predicted image generation; perform thresholding on pixel-level predicted responses to distinguish joint and fracture regions from background regions, and generate joint and fracture predicted images that correspond one-to-one with the spatial location of the input surrounding rock image. S64: Quantification of recognition results; Quantify and analyze the main features of joints and fractures in the joint and fracture prediction image, and finally output the quantified recognition results of joints and fractures in the surrounding rock of underground mine roadways.

[0015] As a preferred option, in step 1, the preprocessing operations on the original image include sequential size unification, normalization, and format conversion.

[0016] As a preferred option, in step 6, S64, the main characteristics of joints and fissures include orientation, continuity, and width.

[0017] This invention proposes an intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning. Compared with existing technologies, this invention has the following advantages: First, this invention enhances the orientation of rock images and fracture labels through directional perturbation and introduces a directional extension feature extraction and fusion mechanism in the multi-scale feature extraction stage. This transforms high-level feature activation from point-like responses to continuous band-like responses along the fracture direction, improving the expressive power and robustness of continuous directional features in slender joint fractures. Specifically, by using a directional perturbation enhancement strategy combining rotation and flipping, the training sample set is expanded, and consistent geometric transformations are simultaneously performed on the rock images and corresponding fracture labels. This effectively enriches the distribution samples of joint fracture orientations, improves the model's adaptability to joint fractures with different orientations, solves the problem of poor generalization ability caused by a single sample orientation, and adapts to the actual engineering scenarios of diverse fracture orientations in underground mine roadways. In the multi-scale feature extraction stage, the innovative concept of direction awareness is incorporated instead of simply stacking directional convolutions. Instead, a feature extraction structure that extends along different horizontal and vertical directions is specially set up. By fusing directional features, a direction-aware feature expression for crack recognition is formed, which effectively reduces the phenomena of crack breakage, false edges, and the segmentation of the same crack into multiple segments. This improves the continuity and integrity of crack recognition and breaks the limitations of existing deep learning methods that use isotropic convolutions to activate features in a point-like or block-like manner. This allows high-level features to form a continuous band-like response along the crack direction, accurately capturing the directional continuous features of slender joint cracks. This significantly improves the robustness of the model to the recognition of cracks with different directions and widths, and effectively avoids crack recognition deviations under dust and lighting interference.

[0018] Secondly, this invention integrates high-level semantic features with consistent orientation and low-level spatial detail features during the decoding and reconstruction stage. This enables the crack prediction results to form continuous line segments in the early stages and maintain orientation consistency during the progressive resolution recovery process, significantly improving the stability and continuity of the segmentation results for joint cracks in the same direction. In terms of the technical solution, the constructed orientation-aware deep learning model, through the orientation-aware features formed during the multi-scale feature extraction stage, carries clear crack orientation information. During the decoding and reconstruction process, this orientation-aware high-level semantic feature and low-level spatial detail feature are deeply integrated, enabling the crack prediction to form a continuous line segment structure from the early stages, rather than discrete point segments. Simultaneously, during the progressive recovery of image spatial resolution, the orientation-aware mechanism effectively constrains the consistency of crack orientation, avoiding problems such as the same crack being segmented into multiple segments, crack breakage, or orientation shift. This significantly improves the stability and continuity of the segmentation results for joint cracks in the same direction, solving the core pain point of existing technologies that struggle to simultaneously address multi-scale crack representation and orientation consistency.

[0019] This method constructs a direction-aware deep learning model and combines it with a direction perturbation-enhanced training strategy. This enables the model to form direction-aware feature representations during the multi-scale feature extraction stage, thereby outputting more continuous, stable, and accurate prediction results for rock joints and fractures. It effectively solves the problems of poor continuity in fracture identification, insufficient characterization of direction features, and weak anti-interference ability in existing technologies. This invention can significantly improve the identification stability and continuity of slender fractures in the same direction, adapting to the actual engineering needs of complex operating environments in underground mine roadways, and providing accurate and reliable data support for rock structure analysis, stability evaluation, and support design. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the surrounding rock image and joint and fracture annotation samples of the underground mine roadway in this invention; Figure 3 This is a schematic diagram of the directional perturbation enhancement processing of the surrounding rock image in an underground mine tunnel in this invention; Figure 4 This is a schematic diagram of the orientation-aware deep learning model structure in this invention; Figure 5 This is a diagram showing the results of joint and fracture identification in the surrounding rock of an underground mine tunnel in this invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings; like Figures 1 to 4 As shown, this invention provides an intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning, comprising the following steps: Step 1: Collect raw images of the surrounding rock of the roof and sidewalls of the underground mine roadway. Preprocessing of the raw images is preferred, and this preprocessing includes sequential size unification, normalization, and format conversion. Specifically, digital images of the surface of the surrounding rock of the roof and sidewalls of the underground mine roadway are collected. All input surrounding rock images are uniformly adjusted to 512×512 pixels using bilinear interpolation. Then, the image pixel values ​​are normalized by dividing the original pixel values ​​by 255, mapping the pixel values ​​to the [0,1] interval, resulting in a standardized surrounding rock image. This standardized image serves as the basic input data for subsequent sample construction, model training, and recognition inference. This preprocessing effectively eliminates interference from image size differences, grayscale fluctuations, and format incompatibility, ultimately yielding a standardized surrounding rock image that meets the requirements for joint and fracture recognition, laying a data foundation for subsequent annotation, training, and recognition work. Step 2: As Figure 2 As shown, joints and fissures in standardized surrounding rock images are accurately labeled, and a training sample set containing surrounding rock images and corresponding joint and fissure labels is constructed. To provide reliable data support for the model to learn the continuous features of joint and fracture orientation, the process of constructing a training sample set containing surrounding rock images and corresponding joint and fracture labels is as follows: S21: Pixel-by-pixel annotation; In the standardized surrounding rock image, perform precise pixel-by-pixel annotation on the joint and fracture area to generate a binary joint and fracture label image with the same size as the original surrounding rock image, and ensure the spatial correspondence between the label and the image; In the label image, pixels belonging to joints and fractures are marked as 1, and the rest are marked as 0; S22: Connectivity Part Sorting: Connectivity part analysis and sorting are performed on the generated binary joint and crack label images to remove discrete noise pixels, merge spatially connected crack pixels into the same connected region, and maintain their original connectivity structure, so that the joint and crack form a continuous linear structure, thereby improving the effectiveness of the labels. S23: Sample pairing: The sorted joint and fracture annotation results are paired one-to-one with the corresponding standardized surrounding rock images to construct a training sample set for learning the continuous features of joint and fracture direction, ensuring the relevance of the sample set.

[0022] Step 3: To improve the model's adaptability to joints and fractures with different orientations, directional perturbation enhancement processing is performed on the training sample set to generate an enhanced sample set containing multiple joint and fracture orientation distributions, thereby expanding the distribution range of joint and fracture orientations in the samples. To construct a high-quality augmented sample set that meets the training requirements of the model's orientation awareness, the process of generating an augmented sample set containing various joint and fracture orientation distributions is as follows: S31: Rotational reinforcement: such as Figure 3As shown, the surrounding rock images in the training sample set are rotated for enhancement. The surrounding rock images are rotated 90°, 180° and 270° around their geometric center to generate rotated enhanced samples containing different joint and fracture orientations. S32: Flip Enhancement: Based on rotation enhancement, horizontal flipping and vertical flipping are performed sequentially on the rotation enhancement sample to further expand the directional distribution range of joints and fractures, and generate flip enhancement samples with richer directional distribution. S33: Label Synchronization Transformation: Perform a geometric transformation on the joint and fracture labels corresponding to the rotation-enhanced samples and the flip-enhanced samples that is completely consistent with the surrounding rock image to ensure that the spatial correspondence between the labels and the images is not destroyed and that they still have spatial consistency. Finally, construct an enhanced training sample set after orientation perturbation enhancement for subsequent orientation-aware model training.

[0023] Step 4: As Figure 4 As shown, a deep learning model with direction awareness is constructed, focusing on setting up feature extraction structures that extend along different directions in the multi-scale feature extraction stage to achieve effective expression of the direction awareness features of joints and fractures. To generate high-quality orientation-aware feature representations that combine directional continuity and multi-scale adaptability, and to provide core support for accurate identification by subsequent models, the effective representation process of joint fracture orientation-aware features is as follows: S41: Multi-scale feature mapping; Based on the convolutional neural network (CNN) framework, high-level features are extracted from the input standardized surrounding rock image during the feature extraction stage, and multi-scale feature mapping operation is performed on the high-level features to obtain feature representations at different receptive field scales, taking into account both the detailed features and global features of the fracture. In the embodiments provided by this invention, three different receptive field scales, corresponding to dilated convolutional structures with dilation rates of 6, 12, and 18 respectively, are set in parallel on the high-level features to extract joint and fracture features at different spatial scales. Specifically, the smaller receptive field scale is used to extract fine-scale fracture features, the medium receptive field scale is used to extract medium-scale fracture features, and the larger receptive field scale is used to extract the spatial distribution features of larger-scale fractures.

[0024] S42: Horizontal feature extraction; In the process of multi-scale feature mapping, a horizontally extended convolutional structure is set on the high-level features, and the directional extension feature extraction operation is performed on the high-level features along the horizontal direction. This horizontal extension branch can form a continuous response in the horizontal direction, thereby generating a directional feature response that is continuously distributed along the horizontal direction, effectively capturing the continuous features of the horizontally oriented crack and enhancing the feature expression of the crack along its main direction. S43: Vertical feature extraction; During the multi-scale feature mapping process, a vertically extended convolutional structure is set synchronously on the high-level features. The directional extension feature extraction operation is performed on the high-level features along the vertical direction. This vertical extension branch can form a continuous response in the vertical direction, thereby generating a directional feature response that is continuously distributed along the vertical direction, effectively capturing the continuous features of vertically oriented cracks and characterizing the crack structure features in another main direction.

[0025] S44: Directional feature fusion; The horizontal and vertical directional feature responses are spatially aligned and fused with the isotropic features at the corresponding scales to form a directional-aware feature representation that combines directional continuity and multi-scale adaptability. This directional-aware feature representation serves as the input feature in the model decoding stage to effectively address the problem of insufficient directional modeling in existing models.

[0026] Step 5: Using the constructed enhanced sample set, train the constructed direction-aware deep learning model, update the model parameters based on the direction-aware feature representation, and finally obtain a high-performance joint and fracture recognition model; To ensure the coordinated operation of each layer of the model and effectively improve the direction perception capability and recognition accuracy, the process of obtaining a high-performance joint and fracture recognition model is as follows: S51: Feature forward propagation; Input the surrounding rock images in the orientation-aware deep learning model from the orientation perturbation enhancement training sample set, and generate orientation-aware feature representations through the model's forward propagation process; S52: Supervision information matching; The obtained joint and fracture label images are used as supervision information and matched with the direction-aware feature expressions generated by the model to provide supervision for model parameter updates; S53: Joint parameter update; through the backpropagation mechanism, the parameters of the orientation-aware feature extraction layer, multi-scale feature extraction layer and decoding layer in the model are jointly updated and optimized to ensure that the various layers of the model work together and improve orientation awareness and recognition accuracy. S54: Model solidification; After the model completes the preset training rounds and the training loss tends to stabilize and reach the preset threshold, all parameters of the model are fixed to obtain a joint and fracture identification model based on direction-aware feature expression training for identifying fractures in the surrounding rock of underground mine roadways.

[0027] Step 6: Input the image of the surrounding rock of the underground mine roadway to be identified into the trained joint and fracture identification model, and output accurate identification results of the surrounding rock joints and fractures through model inference.

[0028] To accurately output the quantified identification results of surrounding rock joints and fractures, the process of outputting accurate identification results through model inference is as follows: S61: Image reasoning to be identified; Input the image of the surrounding rock of the underground mine roadway to be identified into the obtained direction-aware deep learning model, perform the forward propagation operation of the model, and generate direction-aware feature representation; S62: Pixel-level response generation; using orientation-aware feature representation as input features in the decoding stage, the image spatial resolution is restored step by step during the decoding process to generate pixel-level predicted responses for joints and fissures; S63: Predicted image generation; threshold determination processing is performed on the pixel-level predicted response to distinguish the joint and fracture area from the background area. In the embodiment provided by the present invention, the threshold is set to 0.5. When the predicted response value is greater than or equal to 0.5, it is determined to be a fracture pixel and output as 1. Other pixels are output as 0, generating a joint and fracture predicted image that corresponds one-to-one with the spatial position of the input surrounding rock image. S64: Quantification of Recognition Results; Quantitative analysis of the main features of joints and fractures in the joint and fracture prediction image allows for the extraction of spatial distribution data of fractures, including their length, angle, density, and intensity. The final output is the quantified identification result of joints and fractures in the surrounding rock of underground mine roadways, providing data support for engineering applications. As a preferred approach, the main features of joints and fractures include their orientation, continuity, and width.

[0029] Figure 5 (a) is an embodiment of a practical application of the present invention, wherein the image to be analyzed is a high-resolution digital image of the surrounding rock of an underground mine roadway with a size of 512×512 pixels. In this image, joints and fractures mainly appear as slender linear structures, continuously distributed along a specific direction, with different fractures exhibiting relatively consistent orientation characteristics in local areas. Simultaneously, the fracture width varies, and some fractures have low contrast in areas with complex surrounding rock textures or varying brightness, resulting in unclear boundaries at local locations and easy overlap between fractures and background textures. In this type of surrounding rock image, the main information of the fractures lies in the continuous structural features extending along the direction, rather than isolated point features. If the identification process lacks an effective expression of the continuous directional features of the fractures, fracture identification phenomena can easily occur at locations where the fracture is locally weakened or its orientation changes, making it difficult to maintain the overall directional consistency and continuity of the fractures. Therefore, when performing joint and fracture analysis on the shown surrounding rock image, the identification method provided by the present invention needs to be used.

[0030] First, the input surrounding rock image is fed into the orientation-aware deep learning model constructed in this invention. During the model's feature extraction stage, multi-scale feature mapping is performed on the surrounding rock image to obtain high-level feature responses at different receptive field scales. Then, as... Figure 5As shown in (b), during the multi-scale feature mapping process, directional extension feature extraction operations are performed on the high-level features along the horizontal and vertical directions respectively, generating directional feature responses continuously distributed along different directions. Subsequently, the horizontal and vertical feature responses are fused with isotropic features at the corresponding scales to form a direction-aware feature representation (such as...). Figure 5 (c) shown), and input the direction-aware feature representation into the model decoding stage. Then, as shown... Figure 5 As shown in (d), the spatial resolution of the feature map is restored step by step during the decoding stage to obtain the pixel-level prediction response of the joints and fissures. The prediction results exhibit a continuous linear distribution in space. Finally, as... Figure 5 As shown in (e), a threshold determination process is performed on the pixel-level prediction response to generate a joint and fracture prediction image corresponding to the input surrounding rock image space, and the result of the identification of joints and fractures in the surrounding rock of the underground mine roadway is output, thereby completing the application process of the method of the present invention on the actual surrounding rock image.

[0031] This invention proposes an intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning. Compared with existing technologies, this invention has the following advantages: First, this invention enhances the orientation of rock images and fracture labels through directional perturbation and introduces a directional extension feature extraction and fusion mechanism in the multi-scale feature extraction stage. This transforms high-level feature activation from point-like responses to continuous band-like responses along the fracture direction, improving the expressive power and robustness of continuous directional features in slender joint fractures. Specifically, by using a directional perturbation enhancement strategy combining rotation and flipping, the training sample set is expanded, and consistent geometric transformations are simultaneously performed on the rock images and corresponding fracture labels. This effectively enriches the distribution samples of joint fracture orientations, improves the model's adaptability to joint fractures with different orientations, solves the problem of poor generalization ability caused by a single sample orientation, and adapts to the actual engineering scenarios of diverse fracture orientations in underground mine roadways. In the multi-scale feature extraction stage, the innovative concept of direction awareness is incorporated instead of simply stacking directional convolutions. Instead, a feature extraction structure that extends along different horizontal and vertical directions is specially set up. By fusing directional features, a direction-aware feature expression for crack recognition is formed, which effectively reduces the phenomena of crack breakage, false edges, and the segmentation of the same crack into multiple segments. This improves the continuity and integrity of crack recognition and breaks the limitations of existing deep learning methods that use isotropic convolutions to activate features in a point-like or block-like manner. This allows high-level features to form a continuous band-like response along the crack direction, accurately capturing the directional continuous features of slender joint cracks. This significantly improves the robustness of the model to the recognition of cracks with different directions and widths, and effectively avoids crack recognition deviations under dust and lighting interference.

[0032] Secondly, this invention integrates high-level semantic features with consistent orientation and low-level spatial detail features during the decoding and reconstruction stage. This enables the crack prediction results to form continuous line segments in the early stages and maintain orientation consistency during the progressive resolution recovery process, significantly improving the stability and continuity of the segmentation results for joint cracks in the same direction. In terms of the technical solution, the constructed orientation-aware deep learning model, through the orientation-aware features formed during the multi-scale feature extraction stage, carries clear crack orientation information. During the decoding and reconstruction process, this orientation-aware high-level semantic feature and low-level spatial detail feature are deeply integrated, enabling the crack prediction to form a continuous line segment structure from the early stages, rather than discrete point segments. Simultaneously, during the progressive recovery of image spatial resolution, the orientation-aware mechanism effectively constrains the consistency of crack orientation, avoiding problems such as the same crack being segmented into multiple segments, crack breakage, or orientation shift. This significantly improves the stability and continuity of the segmentation results for joint cracks in the same direction, solving the core pain point of existing technologies that struggle to simultaneously address multi-scale crack representation and orientation consistency.

[0033] This method constructs a direction-aware deep learning model and combines it with a direction perturbation-enhanced training strategy. This enables the model to form direction-aware feature representations during the multi-scale feature extraction stage, thereby outputting more continuous, stable, and accurate prediction results for rock joints and fractures. It effectively solves the problems of poor continuity in fracture identification, insufficient characterization of direction features, and weak anti-interference ability in existing technologies. This invention can significantly improve the identification stability and continuity of slender fractures in the same direction, adapting to the actual engineering needs of complex operating environments in underground mine roadways, and providing accurate and reliable data support for rock structure analysis, stability evaluation, and support design.

Claims

1. A method for intelligent identification of joints and fractures in underground surrounding rock based on direction-aware deep learning, characterized in that, Includes the following steps: Step 1: Collect raw surrounding rock image data of the roof and sidewalls of the underground mine roadway, and perform preprocessing operations on the raw surrounding rock images to obtain standardized surrounding rock images that meet the requirements for joint and fracture identification; Step 2: Label the joints and fractures in the standardized surrounding rock images to construct a training sample set containing the surrounding rock images and corresponding joint and fracture labels; Step 3: Perform directional perturbation enhancement processing on the training sample set to generate an enhanced sample set containing various joint and fracture orientation distributions; Step 4: Construct a deep learning model with orientation awareness, and set up feature extraction structures that extend along different directions in the multi-scale feature extraction stage to achieve effective expression of orientation awareness features of joints and fractures; Step 5: Using the constructed enhanced sample set, train the constructed direction-aware deep learning model, update the model parameters based on the direction-aware feature representation, and finally obtain a high-performance joint and fracture recognition model; Step 6: Input the image of the surrounding rock of the underground mine roadway to be identified into the trained joint and fracture identification model, and output accurate identification results of the surrounding rock joints and fractures through model inference.

2. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 1, characterized in that, In step 2, the process of constructing a training sample set containing images of the surrounding rock and corresponding joint and fracture labels is as follows: S21: Pixel-by-pixel annotation; In the standardized surrounding rock image, pixel-by-pixel annotation is performed on the joint and fracture area to generate a joint and fracture label image with the same size as the original surrounding rock image, and to ensure the spatial correspondence between the label and the image. S22: Connectivity Part Sorting: Connectivity part analysis and sorting are performed on the generated joint and fissure label images to remove discrete noise pixels, making the joints and fissures form a continuous linear structure and improving the effectiveness of the labels. S23: Sample pairing: The sorted joint and fracture annotation results are paired one-to-one with the corresponding standardized surrounding rock images to construct a training sample set for learning the continuous features of joint and fracture direction.

3. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 1 or 2, characterized in that, In step 3, the process of generating an enhanced sample set containing multiple joint and fracture orientation distributions is as follows: S31: Rotation Enhancement: Rotate the surrounding rock images in the training sample set to enhance them by rotating the images around their geometric center by 90°, 180° and 270° respectively, to generate rotation-enhanced samples containing different joint and fracture orientations. S32: Flip Enhancement: Perform horizontal and vertical flipping processes sequentially on the rotation-enhanced sample to further expand the directional distribution range of joints and fractures, generating flipped enhanced samples; S33: Label Synchronization Transformation: Perform geometric transformations on the joint and fracture labels corresponding to the rotation-enhanced samples and the flip-enhanced samples, ensuring that the spatial correspondence between the labels and the images is not destroyed, and finally construct an enhanced training sample set after directional perturbation enhancement.

4. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 3, characterized in that, In step 4, the effective expression process of joint fracture direction sensing features is as follows: S41: Multi-scale feature mapping; Based on the convolutional neural network framework, high-level features are extracted from the input standardized surrounding rock image, and multi-scale feature mapping operation is performed on the high-level features to obtain feature representations at different receptive field scales; S42: Horizontal feature extraction; During the multi-scale feature mapping process, a directional extension feature extraction operation is performed on the high-level features along the horizontal direction to generate a directional feature response that is continuously distributed along the horizontal direction, capturing the continuous features of the horizontally oriented crack. S43: Vertical feature extraction; Simultaneously, during the multi-scale feature mapping process, a directional extension feature extraction operation is performed on the high-level features along the vertical direction to generate a directional feature response that is continuously distributed along the vertical direction, capturing the continuous features of vertically oriented cracks; S44: Directional feature fusion; The horizontal and vertical directional feature responses are spatially aligned and fused with the isotropic features at the corresponding scales to form a directional perception feature expression that combines directional continuity and multi-scale adaptability.

5. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 4, characterized in that, In step 5, the process of obtaining the high-performance joint and fracture identification model is as follows: S51: Feature forward propagation; Input the surrounding rock images in the orientation-aware deep learning model from the orientation perturbation enhancement training sample set, and generate orientation-aware feature representations through the model's forward propagation process; S52: Supervision information matching; The obtained joint and fracture label images are used as supervision information and matched with the direction-aware feature expressions generated by the model to provide supervision for model parameter updates; S53: Joint parameter update; through the backpropagation mechanism, the parameters of the orientation-aware feature extraction layer, multi-scale feature extraction layer and decoding layer in the model are jointly updated and optimized to ensure that the various layers of the model work together and improve orientation awareness and recognition accuracy. S54: Model solidification; After the model completes the preset training rounds and the training loss tends to stabilize and reach the preset threshold, all parameters of the model are fixed to obtain the joint and fracture recognition model trained based on direction-aware feature expression.

6. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 5, characterized in that, In step 6, the process of outputting accurate identification results of surrounding rock joints and fractures through model inference is as follows: S61: Image reasoning to be identified; Input the image of the surrounding rock of the underground mine roadway to be identified into the obtained direction-aware deep learning model, perform the forward propagation operation of the model, and generate direction-aware feature representation; S62: Pixel-level response generation; Using orientation-aware feature representation as input features in the decoding stage, the spatial resolution of the image is restored step by step during the decoding process to generate pixel-level predictive responses for joints and fissures; S63: Predicted image generation; perform thresholding on pixel-level predicted responses to distinguish joint and fracture regions from background regions, and generate joint and fracture predicted images that correspond one-to-one with the spatial location of the input surrounding rock image. S64: Quantification of recognition results; Quantify and analyze the main features of joints and fractures in the joint and fracture prediction image, and finally output the quantified recognition results of joints and fractures in the surrounding rock of underground mine roadways.

7. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 1, characterized in that, In step 1, the preprocessing operations on the original image include sequential size unification, normalization, and format conversion.

8. The intelligent identification method for joints and fractures in underground surrounding rock based on direction-aware deep learning according to claim 6, characterized in that, In step 6, S64, the main characteristics of joints and fissures include orientation, continuity, and width.

Images