Intelligent identification method for dangerous rock disaster based on fracture mechanics

Abstract

The application discloses a dangerous rock disaster intelligent identification method based on fracture mechanics and belongs to the technical field of geological disaster monitoring and intelligent identification. High-resolution image sequences are acquired, and three-dimensional point cloud reconstruction and slope three-dimensional modeling are completed; after image preprocessing, standardized samples are constructed and input into a Swin-Transformer joint model, multi-scale feature extraction and segmentation of dangerous rock blocks, main control fractures and structural surfaces are realized; in combination with three-dimensional point clouds, parameterized extraction and weight assignment of dangerous rock geometric parameters and configuration types are completed, mechanical parameters such as stress intensity factors and energy release rates are calculated, and fracture damage indexes and fracture safety factors are constructed; according to the threshold values of the mechanical indexes, dangerous rock stability grade division and early warning mapping are completed, and in combination with a field monitoring system, the visual early warning and dynamic updating of dangerous rock instability risks are realized. The application deeply couples visual global identification and quantitative determination of fracture mechanics, improves dangerous rock identification precision under complex working conditions, scientific nature of stability evaluation and early warning reliability.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and intelligent identification technology, specifically to an intelligent identification method for dangerous rock disasters based on fracture mechanics. Background Technology

[0002] Rockfall is a typical and highly dangerous geological hazard in mountainous and reservoir areas of my country. It is widely distributed in the Three Gorges Reservoir area, steep slopes in western China, along major transportation routes, and in the drawdown zones of hydropower projects. It is characterized by its wide spatial distribution, sudden onset, high concealment, and large-scale destruction. my country has a vast area of ​​mountains and hills, and steep slopes are generally characterized by densely developed joints and fissures, dramatic topographic relief, and complex vegetation cover. These factors, combined with the dynamic effects of heavy rainfall, reservoir water level fluctuations, and micro-earthquakes, result in highly nonlinear and uncertain evolution of unstable rock masses. Traditional rockfall identification relies mainly on manual reconnaissance, visual interpretation, and simple instrument monitoring. This method is inefficient, has limited coverage, and is particularly vulnerable in steep areas due to poor accessibility and significant safety risks. It fails to achieve refined, routine, and wide-area monitoring of rock masses, controlling fissures, and structural surfaces, and cannot meet the engineering requirements for accurate identification, dynamic assessment, and timely early warning. With the upgrading of geological disaster prevention and control, the development of high-precision, high-efficiency, and intelligent rockfall identification and risk assessment technologies has become an urgent task in the fields of engineering geology and disaster prevention and mitigation.

[0003] Current methods for identifying and monitoring dangerous rock formations are evolving from manual processes to integrated air-space-ground non-contact surveying. Technologies such as UAV photogrammetry, 3D laser scanning, and remote sensing image interpretation are widely used, significantly improving data acquisition efficiency and spatial coverage. Deep learning methods, represented by convolutional neural networks, are used for slope image segmentation, dangerous rock detection, and crack extraction. By learning shallow features such as texture and contour, they achieve dangerous rock area identification, mitigating to some extent the subjectivity and inconsistency issues of traditional methods. However, existing visual methods largely rely on local features and lack robustness to complex lighting, vegetation obstruction, and changes in scale and perspective, making it difficult to accurately characterize dangerous rock masses, controlling cracks, and structural surfaces globally. More importantly, these methods only reach the geometric recognition level, lacking a quantitative mapping relationship between the recognition results and core parameters of fracture mechanics. They cannot explain crack propagation and rock mass failure processes from the perspective of instability mechanisms, and stability evaluation lacks mechanical basis, making it difficult to support high-precision risk assessment and early warning decisions. These technological shortcomings restrict engineering applications.

[0004] Internationally, a large number of high-quality studies have been conducted on rock fracture identification, slope risk assessment, and applications of fracture mechanics, providing theoretical support for technological evolution. Chen et al. published a study in the Journal of Structural Geology, using U-Net and DeeplabV3+ to achieve high-precision segmentation of rock fractures, but only focused on two-dimensional image recognition without combining it with three-dimensional point clouds and mechanical parameters. Lato et al. proposed a discrete fracture network modeling method based on UAV photogrammetry in Remote Sensing, which can extract fracture orientation and density, but did not establish a quantitative conversion relationship between fracture geometric parameters and stress intensity factor and energy release rate. Fan et al. (2025) conducted a risk assessment of dangerous rocks in the Three Gorges Reservoir area in Geomorphology, integrating terrain units and physical models, but still mainly relied on statistical and empirical criteria, and did not couple visual global perception with linear elastic fracture mechanics (LEFM) to quantitatively characterize the critical state. Overall, existing English literature generally suffers from two major shortcomings: first, visual models lack a global attention mechanism, resulting in insufficient accuracy in extracting key structures in complex scenes; second, geometric recognition is disconnected from fracture mechanics parameters, failing to form a closed loop of recognition and mechanical judgment, thus limiting interpretability and engineering applicability.

[0005] Chinese patent CN113010835B discloses a method and system for early warning of rockfall based on fracture mechanics. By introducing fracture parameters such as stress intensity factor and energy release rate to construct a stability evaluation model, it partially compensates for the lack of mechanical basis in traditional methods. This patent calculates crack propagation parameters based on monitoring data to determine instability risk, but it still has significant shortcomings: the data source relies on contact-based single-point monitoring equipment, failing to utilize multi-view close-up photography and 3D reconstruction by drones to achieve wide-area non-contact acquisition, making it difficult to cover large reservoir areas and steep slopes; it does not employ a visual Transformer global attention mechanism for high-precision segmentation of rock masses and controlling cracks, resulting in low accuracy and poor robustness in crack geometric parameter extraction; it lacks a rock mass configuration classification system and weighting mechanism, and the mechanical calculations do not incorporate spatial features such as block morphology and structural surface combinations, leading to significant deviations between fracture parameters and actual instability modes. The recognition accuracy, scene adaptability, and mechanical interpretability still cannot meet the needs of intelligent identification and early warning in complex working conditions.

[0006] In summary, existing technologies and domestic and international research have significant shortcomings in terms of adaptability to complex environments, global accuracy in identification, deep coupling of mechanical mechanisms, and fully automated closed-loop early warning. Traditional methods are unable to balance identification accuracy, efficiency, mechanical interpretability, and engineering practicality. Summary of the Invention

[0007] To address the aforementioned technical problems, this application discloses an intelligent identification method for unstable rockfall hazards based on fracture mechanics, comprising the following steps:

[0008] S1. Use drones to conduct multi-view close-up photography and directional cruise shooting of the target dangerous rock area to obtain high-resolution visible light image sequences. Obtain sparse / dense point clouds through structured beam method or multi-view stereo reconstruction to construct a three-dimensional geometric model of the dangerous rock slope.

[0009] S2. The original image is preprocessed sequentially by background suppression, geometric correction, noise reduction, illumination equalization, color normalization and distortion correction. Combined with manual annotation, the image is cropped into fixed-size blocks, which are then converted into Tensor form after format conversion, normalization / standardization and data augmentation to construct training and inference samples.

[0010] S3. Input the preprocessed image into the Swin-Transformer joint model. After multi-stage feature extraction, multi-head self-attention and translation window mechanism, global-local multi-scale features are obtained. The features are then fused by decoder upsampling and skip connections to output the segmentation mask of the dangerous rock block, the main control fracture and the structural surface.

[0011] S4. Under the joint constraints of 3D point cloud and segmentation mask, extract the geometric parameters of the dangerous rock block and the geometric parameters of the fracture. Determine the configuration type of the dangerous rock based on the combination relationship of structural surfaces and the direction of intersection and assign configuration weights.

[0012] S5. Combining self-weight, earthquake, pore water and additional load, solve for the normal stress and tangential stress of the potential fracture surface. Use the linear elastic fracture mechanics analytical formula to calculate the stress intensity factor and energy release rate of type I and type II stresses. Compare the critical energy release rate of the material to construct the fracture damage index and fracture safety factor.

[0013] S6. Based on the threshold criteria of fracture safety factor and fracture damage index, the stability level of dangerous rock is divided and mapped to a multi-level early warning level. Combined with the on-site monitoring system, the risk of dangerous rock fracture instability is visualized, early warning and dynamically updated.

[0014] Preferably, the process of constructing the three-dimensional geometric model of the dangerous rock slope in S1 is as follows:

[0015] High-resolution visible light image sequences of the target dangerous rock area were obtained by using UAV multi-view close-up photography. The exterior orientation elements of the image sequence were solved based on the structure-reconstruction-motion algorithm to construct a sparse point cloud. The sparse point cloud was then densified using a multi-view stereo matching algorithm to obtain a dense three-dimensional point cloud of the dangerous rock slope. The dense point cloud was then reconstructed into a surface based on the Poisson reconstruction algorithm to generate a three-dimensional geometric model of the dangerous rock slope.

[0016] The three-dimensional coordinates of the sparse point cloud are obtained by bundle adjustment, and the solution model is as follows: In the formula, , The first The rotation matrix and translation vector of the image For the first The coordinates of a three-dimensional point For the first The three-dimensional point at the th t Projected coordinates on the image For the camera intrinsic parameter matrix, The total number of images, This represents the total number of three-dimensional feature points. This is the robust loss function.

[0017] Preferably, the preprocessing of the original image in S2 includes:

[0018] An adaptive thresholding method is used for background suppression to separate the unstable rock slope from the non-target background area; geometric correction is performed using a camera distortion model, which is as follows: In the formula, To correct the pixel coordinates before, For the corrected pixel coordinates, Radial distance, These are the radial distortion coefficients, These are the tangential distortion coefficients;

[0019] Gaussian filtering is used for noise reduction, histogram equalization is used for illumination equalization, and linear transformation based on RGB channels is used for color normalization. Combined with manual annotation, the image is cropped into fixed-size patches, which are then converted into Tensor form after format conversion, normalization / standardization, and data augmentation to construct training and inference samples.

[0020] Preferably, the processing procedure of the Swin-Transformer joint model in S3 is as follows: the input image is divided into 4×4 non-overlapping image blocks, and the initial feature matrix is ​​obtained through linear mapping. C represents the embedding dimension;

[0021] Features are extracted in multiple stages using the SwinTransformer encoder, with a multi-head self-attention mechanism within each block: ,in, They are query, key, and value, respectively. For feature dimension, This is the relative position offset matrix;

[0022] In the adjacent stages, the features of 2×2 image blocks are stitched together and downsampled. After the features are upsampled by the decoder and fused with skip connections, the segmentation mask is output through the Softmax function. ,in, The decoder outputs a feature map. This is the final segmentation result.

[0023] Preferably, the extraction of geometric parameters and configuration weight calculation of unstable rock blocks and fractures in S4 are as follows:

[0024] Under the joint constraints of 3D point cloud and segmentation mask, the contour point set, envelope volume, principal axis length, average thickness, and equivalent shape factor of the unstable rock mass are extracted. The formula for calculating the equivalent shape factor is as follows: In the formula, The visible surface area of ​​the dangerous rock. This represents the volume of the envelope of the dangerous rock.

[0025] Extract the length, opening, dip, dip angle, and spatial distribution density of fractures. Determine the rockfall configuration type based on the structural plane combination relationship and intersection direction. The configuration weight coefficient is calculated using the following formula: In the formula, These are empirical weighting coefficients. The crack length is... The maximum fracture length in the target unstable rock area. The spatial distribution density of the fracture. This is the equivalent shape factor.

[0026] Preferably, the calculation process for the Type I and Type II stress intensity factors in S5 is as follows:

[0027] Combining self-weight, earthquake, pore water, and additional loads, the normal and tangential stresses of the potential fracture surface are solved; based on the linear elastic fracture mechanics theory, the Type I and Type II stress intensity factors are calculated using the following formulas: In the formula, For the normal stress of the fracture surface, For the shear stress on the crack surface, The equivalent crack length, These are dimensionless geometric coefficients.

[0028] Preferably, the calculation process for the energy release rate in S5 is as follows: Under the assumption of plane strain, the stress intensity factor is converted into the energy release rate per unit crack propagation area, and the calculation formula is: In the formula, The equivalent elastic modulus of the rock mass. Poisson's ratio, It is a type I stress intensity factor. It is a type II stress intensity factor.

[0029] Preferably, the calculation process for the critical energy release rate of the material in S5 is as follows: the critical energy release rate of the material is calculated based on the fracture toughness of the rock mass mixed mode, and the calculation formula is: In the formula, Type I fracture toughness, It has a type II fracture toughness. This represents the weighting coefficient for the contribution of type II toughness to the total fracture energy. The equivalent elastic modulus of the rock mass. It is Poisson's ratio.

[0030] Preferably, the calculation process for the LEFM fracture damage index and fracture safety factor in S5 is as follows: Based on the energy release rate and the material's critical energy release rate, the LEFM fracture damage index is calculated using the following formula: In the formula, Energy release rate, The critical energy release rate of the material. It is a type I stress intensity factor. It is a type II stress intensity factor. Type I fracture toughness, It has a type II fracture toughness. Type II toughness weighting coefficient;

[0031] The fracture safety factor is calculated based on fracture damage indices, using the following formula: In the formula, LEFM fracture damage index This is the equivalent safety factor based on fracture mechanics.

[0032] Preferably, the relationship between the stability level and the fracture safety factor in S6 is as follows:

[0033] Based on the threshold criteria of fracture safety factor and fracture damage index, the stability of dangerous rocks is divided into multiple levels, and each stability level corresponds to a different fracture safety factor range. The higher the fracture safety factor, the higher the stability level of the dangerous rock. The stability level is mapped to a multi-level early warning level, and combined with the on-site monitoring system, the visual early warning and dynamic update of the risk of dangerous rock fracture instability can be realized.

[0034] Compared with the prior art, the technical solution of this application has the following technical effects:

[0035] This invention utilizes UAV multi-view close-up photography and 3D reconstruction technology to acquire high-resolution images and dense point clouds of dangerous rock areas, accurately construct 3D geometric models of dangerous rock slopes, and realize the spatial positioning and morphological restoration of the main control cracks and structural surfaces of dangerous rock blocks.

[0036] This invention employs the Swin-Transformer joint model to intelligently segment dangerous rock blocks and fissures. By leveraging global and local multi-scale feature extraction capabilities, it stably obtains segmentation results of key dangerous rock structures. Simultaneously, relying on standardized image preprocessing and tensor sample construction processes, it enhances the stability of model inference and recognition, achieving automated and high-precision extraction of key elements of dangerous rocks.

[0037] This invention fully parameterizes the geometric parameters and configuration characteristics of unstable rocks and deeply couples them with the theory of linear elastic fracture mechanics. It can accurately calculate the stress intensity factor, energy release rate, and critical mechanical indicators, establish a quantitative mapping relationship from geometric identification to mechanical judgment, clearly reflect the crack propagation and instability mechanism of unstable rocks, and improve the scientificity and interpretability of unstable rock stability evaluation.

[0038] This invention completes the stability classification and early warning mapping of dangerous rocks based on fracture mechanics indicators, and realizes the visualization and dynamic updating of risk warnings by combining on-site monitoring systems. It forms a fully automated closed loop for the identification, parameterization, judgment, and early warning process, and can continuously output stable and reliable early warning results, providing continuous and effective technical support for the dynamic monitoring and timely prevention and control of dangerous rock disasters.

[0039] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0040] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0042] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0043] Figure 1 A schematic diagram of the overall process of an intelligent identification method for dangerous rock hazards based on fracture mechanics;

[0044] Figure 2 Functional module architecture diagram of an intelligent rockfall hazard identification system based on fracture mechanics;

[0045] Figure 3 Flowchart of the network architecture of the Swin-Transformer dangerous rock fissure identification system. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0047] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0048] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0049] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0050] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0051] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0052] Example 1 mainly describes a method for intelligent identification of rockfall hazards based on fracture mechanics, such as... Figure 1 As shown, it specifically includes:

[0053] S1. Use drones to conduct multi-view close-up photography and directional cruise shooting of the target dangerous rock area to obtain high-resolution visible light image sequences. Obtain sparse / dense point clouds through structured beam method or multi-view stereo reconstruction to construct a three-dimensional geometric model of the dangerous rock slope.

[0054] S2. The original image is preprocessed sequentially by background suppression, geometric correction, noise reduction, illumination equalization, color normalization and distortion correction. Combined with manual annotation, the image is cropped into fixed-size blocks, which are then converted into Tensor form after format conversion, normalization / standardization and data augmentation to construct training and inference samples.

[0055] S3. Input the preprocessed image into the Swin-Transformer joint model. After multi-stage feature extraction, multi-head self-attention and translation window mechanism, global-local multi-scale features are obtained. The features are then fused by decoder upsampling and skip connections to output the segmentation mask of the dangerous rock block, the main control fracture and the structural surface.

[0056] S4. Under the joint constraints of 3D point cloud and segmentation mask, extract the geometric parameters of the dangerous rock block and the geometric parameters of the fracture. Determine the configuration type of the dangerous rock based on the combination relationship of structural surfaces and the direction of intersection and assign configuration weights.

[0057] S5. Combining self-weight, earthquake, pore water and additional load, solve for the normal stress and tangential stress of the potential fracture surface. Use the linear elastic fracture mechanics analytical formula to calculate the stress intensity factor and energy release rate of type I and type II stresses. Compare the critical energy release rate of the material to construct the fracture damage index and fracture safety factor.

[0058] S6. Based on the threshold criteria of fracture safety factor and fracture damage index, the stability level of dangerous rock is divided and mapped to a multi-level early warning level. Combined with the on-site monitoring system, the risk of dangerous rock fracture instability is visualized, early warning and dynamically updated.

[0059] like Figure 2 As shown in the figure, the intelligent identification method for dangerous rock hazards considering fracture mechanics disclosed in this invention includes three modules: an identification module, a parameterization module, and an assessment and early warning module; the identification module includes three parts: image data acquisition, data preprocessing, and Swin-Transformer dangerous rock fracture identification;

[0060] The identification module is used to preprocess the images collected by the UAV and intelligently identify the unstable rock mass, fissures, and structural surfaces based on the Swin-Transformer joint algorithm, outputting two-dimensional segmentation results and fissure edge information. The parameterization module is used to register and perform ensemble operations on the two-dimensional identification results and three-dimensional point clouds, extract the geometric parameters of the unstable rock mass and the geometric and orientation parameters of the fissures, determine the unstable rock configuration type, and assign configuration weight coefficients. The evaluation and early warning module is used to combine the geometric-configuration parameters with external loads and boundary conditions, calculate the crack tip stress intensity factor and energy release rate based on LEFM, construct fracture damage index and safety factor, classify the stability of the unstable rock mass, and output the early warning level.

[0061] In specific implementation, as a preferred option, the image data acquisition system in the identification module uses a drone to conduct multi-view, directional cruise photogrammetry of the unstable rock, capturing high-resolution visible light image data and point cloud data in real time. After acquisition, the image sequence and exterior orientation elements are imported into 3D reconstruction software. Using the structured bundle method and multi-view stereo reconstruction algorithm, feature matching, joint aerial triangulation adjustment, and dense matching are performed on the images to obtain a dense point cloud and 3D surface model of the unstable rock slope. The 3D point cloud is expressed using a unified coordinate system, providing a spatial basis for subsequent geometric parameterization and configuration recognition.

[0062] The specific workflow of the data preprocessing submodule in the recognition module includes: image annotation, resizing, format conversion, normalization / standardization, data augmentation, and tensorization. By preprocessing the data, the image data is standardized. Finally, the standardized data undergoes image segmentation for Swin-Transformer inference training.

[0063] High-resolution visible light image sequences of the target dangerous rock area were obtained by using UAV multi-view close-up photography. The exterior orientation elements of the image sequence were solved based on the structure-reconstruction-motion algorithm to construct a sparse point cloud. The sparse point cloud was then densified using a multi-view stereo matching algorithm to obtain a dense three-dimensional point cloud of the dangerous rock slope. The dense point cloud was then reconstructed into a surface based on the Poisson reconstruction algorithm to generate a three-dimensional geometric model of the dangerous rock slope.

[0064] The three-dimensional coordinates of the sparse point cloud are obtained by bundle adjustment, and the solution model is as follows: In the formula, , The first The rotation matrix and translation vector of the image For the first The coordinates of a three-dimensional point For the first The three-dimensional point at the th t Projected coordinates on the image For the camera intrinsic parameter matrix, The total number of images, This represents the total number of three-dimensional feature points. This is the robust loss function.

[0065] The preprocessing of the original image includes: background suppression using an adaptive thresholding method to separate the dangerous rock slope from the non-target background area; and geometric correction using a camera distortion model, which is as follows: In the formula, To correct the pixel coordinates before, For the corrected pixel coordinates, Radial distance, These are the radial distortion coefficients, These are the tangential distortion coefficients;

[0066] Gaussian filtering is used for noise reduction, histogram equalization is used for illumination equalization, and linear transformation based on RGB channels is used for color normalization. Combined with manual annotation, the image is cropped into fixed-size patches, which are then converted into Tensor form after format conversion, normalization / standardization, and data augmentation to construct training and inference samples.

[0067] Specifically, the data preprocessing submodule's processing flow is as follows:

[0068] S101. Use annotation software to annotate the original image with pixel-level annotations of targets such as unstable rock blocks, primary control fractures, secondary fractures, and structural surfaces, to obtain the original image and the corresponding mask image.

[0069] S102. Scale and crop the labeled images to a fixed size (512×512 pixels) to ensure consistent input size;

[0070] S103. Convert the image to RGB three-channel format and convert the color space as needed;

[0071] S104. Apply normalization to each pixel channel, converting the original feature value x to: , x is the original feature value; x' is the standardized feature value; mean is the mean of the feature in all samples; std is the standard deviation of the feature in all samples; thus achieving zero mean and unit variance standardization of the data distribution;

[0072] S105. Perform data augmentation operations on the image, such as random flipping, random rotation, brightness and contrast perturbation, Gaussian noise or blurring, to improve the robustness of the model to crack morphology, lighting changes and viewpoint differences.

[0073] S106. Convert the image data into a tensor format of (B, 3, H, W), where B is the batch size, 3 is the RGB channels, and H and W are the input image dimensions (height and width), and use it as the input to the Swin-Transformer network.

[0074] The Swin-Transformer dangerous rock fissure identification system in the identification module uses the Swin-Transformer joint algorithm to infer and train the standardized segmented images in the data preprocessing system. The Swin-Transformer dangerous rock fissure identification system includes a Patch segmentation and linear embedding module, a SwinTransformer encoder, a multi-scale feature decoder, and a segmentation prediction head.

[0075] Specifically, such as Figure 3 As shown, the specific details of the Swin-Transformer dangerous rock fracture identification system are as follows:

[0076] S201, Input the original image After preprocessing, the image is divided into 4×4 non-overlapping patches. Each patch, after being flattened, is... The feature vector is obtained through linear mapping. : In the formula, It is a linear projection matrix. This is a bias term.

[0077] The initial feature matrix is ​​further obtained as follows: In the formula, For the embedded dimension;

[0078] S202. Input the feature matrix into the staged Swing Transformer encoder. Each stage consists of multiple Swing Transformer Blocks, which are stacked alternately using window multi-head self-attention (W-MSA) and translation window multi-head self-attention (SW-MSA) to achieve joint modeling of local and global aspects; the in-window attention uses the formula: In the formula, These are respectively: Query, Key, and Value. For feature dimension, This is the relative position offset matrix;

[0079] S203. Between adjacent stages, the 2×2 neighborhood patch features are concatenated, and downsampling is achieved through linear transformation to obtain a feature representation with a higher number of channels and lower resolution, thereby reducing the computational load while maintaining the receptive field. The new feature representation is as follows: In the formula, The concatenated feature vectors For merging matrices, For bias terms;

[0080] S204. A decoder is used to fuse and upsample features from different stages to restore the original image resolution. Feature fusion combines shallow details with deep semantic information through skip connections, while upsampling uses deconvolution or interpolation to restore spatial resolution. After feature fusion and upsampling, a 1×1 convolution maps the number of channels to the number of classes for channel compression, obtaining a pixel-level class distribution.

[0081] Furthermore, the feature map output by the decoder... Mapped to the final segmentation result The probability of each pixel belonging to class c is calculated using the Softmax function: ;

[0082] Each pixel is classified using a threshold or maximum probability principle to obtain segmentation masks for unstable rock blocks, main control fractures, and structural surfaces. Furthermore, a morphological skeleton extraction method is employed to extract the centerline from the fracture mask, serving as the boundary and orientation information of the fracture's two-dimensional geometry, laying the foundation for subsequent three-dimensional backprojection and parameterization.

[0083] The parameterization module performs a set operation on the two-dimensional segmentation results and the three-dimensional point cloud in a unified coordinate system. This set refers to the spatial registration and joint representation of multiple surveys, multi-view images, and multi-source data (images, point clouds, and monitoring data) within the same three-dimensional coordinate system to obtain a spatiotemporally integrated set of dangerous rock parameters. Specifically:

[0084] S301. Based on the UAV pose and camera intrinsic and extrinsic parameters, image pixels are back-projected onto a 3D point cloud coordinate system using imaging geometry, achieving spatial alignment between fracture edge pixels and point cloud points. Through multi-view data fusion, the external contour point set of the unstable rock block and the discrete point set of the fracture space can be obtained. ;

[0085] For the geometric boundaries of the unstable rock mass, a three-dimensional convex hull or Shape reconstruction method, construct block envelope, calculate block envelope volume Spindle length Average thickness With equivalent shape factor Parameters, etc. The equivalent shape coefficient can be defined as: In the formula, The visible surface area of ​​the dangerous rock. The degree of elongation or sheet-like structure of a block is one of the important criteria for configuration identification;

[0086] S302. For each identified crack, first fit a plane equation from the crack point set: ;

[0087] Normalizing the plane normal vector yields the unit normal vector: ;

[0088] The rock fracture parameters are used to characterize the geometric properties of the main control surface and auxiliary structural surfaces of the rock. The main parameters include: fracture length *a*, and fracture opening. ,tendency ,inclination and the spatial distribution density of fractures ;

[0089] The crack orientation is the actual size obtained by three-dimensional back-projection of the crack edge output by the Transformer recognizer, and is defined as follows:

[0090] tendency The angle between the fracture strike and true north (clockwise); dip angle. : Inclination angle of the fracture surface relative to the horizontal plane; Fracture length a: Scale along the maximum extension direction of the fracture, an important parameter for calculating the stress intensity factor in LEFM; Fracture opening The average distance between fracture surfaces is used to reflect the degree of tension-type fracture.

[0091] By statistically analyzing all sets of fractures and structural surfaces, we can further calculate the spatial distribution density of fractures, the number of joint groups, and the dominant occurrence of each group, thus achieving a three-dimensional geometric quantitative expression of the structural surfaces of unstable rocks.

[0092] S303. The aforementioned rock mass configuration determination is a systematic identification of rock mass instability modes, mainly including single-fracture toppling type, double-structural-plane wedge type, multi-fracture blocky type, and layered rock mass plate-crack type. The configuration determination parameters include the structural plane combination relationship matrix. , direction of fracture intersection Potential slip surface normal and the corresponding configuration weight coefficients The structural plane combination relationship can be expressed as: ;

[0093] The direction of the intersection line between two structural surfaces can be calculated using the spatial vector method: ;

[0094] configuration weight coefficient The general form of the value assignment is based on a comprehensive consideration of structural control, crack size sensitivity, and spatial combination: In the formula, These are empirical weighting coefficients used to express the importance of fracture geometry, the degree of development of structural surfaces, and block morphology in determining the configuration.

[0095] The LEFM-based judgment calculation system in the assessment and early warning module is used to transform external loads and boundary conditions into crack tip stress intensity factors and energy release rates based on the parameterized results of the rock mass geometry and configuration, and to form quantitative instability critical criteria accordingly. This system includes a load and boundary condition calculation submodule, a stress intensity factor solution submodule, and an energy release rate and safety factor calculation submodule.

[0096] Specifically, the load and boundary condition calculation submodule is based on the volume of the unstable rock mass. Equivalent thickness Geometric boundary In addition to slope geometry, automatically apply self-weight loads. Seismic inertial force (horizontal seismic coefficient) Vertical seismic coefficient The normal stress on the potential fracture surface is obtained using a quasi-static limit equilibrium method, taking into account factors such as pore water pressure and additional slope loads. With tangential stress The stress intensity factor solution submodule determines the crack size 'a' and the orientation parameter (tendency). ,inclination After determining the geometric correction factor Y, the Type I and Type II stress intensity factors are obtained using the analytical expression of linear elastic fracture mechanics: , In the formula, The tensile (or compressive) normal stress acting on the fracture surface. denoted as shear stress along the fracture surface, a as the equivalent fracture length, and Y as a dimensionless geometric coefficient considering the influence of block shape, fracture location, and configuration.

[0097] The energy release rate and safety factor calculation submodule, under the assumption of plane strain, converts the above stress intensity factor into the energy release rate F per unit crack propagation area: ;

[0098] In the formula, E is the equivalent elastic modulus of the rock mass. Poisson's ratio. Further, the critical energy release rate of the material. The fracture toughness can be determined by the mixed mode fracture toughness of the rock mass. , Or obtained through indoor experimental fitting, the typical mixed-mode energy criterion can be expressed as: In the formula, The weighting coefficient for the contribution of type II toughness to total fracture energy;

[0099] For ease of engineering application, this invention defines the LEFM fracture damage index. With fracture safety factor for: , In the formula, Characterizes the relative level of crack tip energy driving force with respect to the material's crack resistance. The larger the rock mass, the higher the risk of the rock mass becoming unstable. The equivalent safety factor is based on fracture mechanics. The larger the value, the more stable the unstable rock. This system calculates in real time... , G and This provides a unified mathematical basis for subsequent stability level classification and early warning.

[0100] As a preferred embodiment, the risk classification and early warning system in the assessment and early warning module is based on and The calculation results classify the stability of unstable rock masses into four levels: stable, basically stable, understability, and unstable, and further map these levels into four warning levels: low risk, medium risk, high risk, and ultra-high risk. By setting clear mathematical criteria, the quantitative classification of the risk of fracture and instability of unstable rock masses is achieved.

[0101] The stability level determination based on LEFM can be expressed as:

[0102] Stable (Level I): or ;

[0103] Basically stable (Level II): or ;

[0104] Unstable (Level III): or ;

[0105] Unstable (Level IV): or ;

[0106] In the formula, , , These are the safety factor thresholds for stability grading. , , This represents the corresponding energy damage threshold. In engineering applications, values ​​can be determined based on experience or specifications. , , ;

[0107] The above criterion can then be specified as follows:

[0108] Stablize: Basically stable: Unstable: Unstable: ;

[0109] Correspondingly, mapping the stability level to the warning level W (levels 1-4) can be defined as follows:

[0110] Low risk;

[0111] Medium risk;

[0112] High risk;

[0113] Extremely high risk;

[0114] In the formula, W is the warning level number. The larger the value, the higher the risk of rock instability and the stricter the warning measures.

[0115] In practice, the assessment and early warning module can be deployed on a slope monitoring platform or cloud server to calculate the early warning level. The system renders and displays information such as the location, configuration type, and main control fracture parameters of the dangerous rock mass on a 3D visualization interface. When the warning level of a dangerous rock mass reaches high risk or extremely high risk, the system automatically triggers an audible and visual alarm or pushes an alarm message to the on-duty personnel. It can also link with multiple sensors on site, such as displacement gauges, crack gauges, and inclinometers, to achieve the confirmation of dangerous status and the upgrading of warning based on deformation and fracture information.

[0116] This embodiment details the high-precision extraction of key control elements such as unstable rock masses, main control fractures, and structural surfaces using a visual Transformer model, overcoming the bottleneck of traditional methods that struggle to balance resolution, efficiency, and safety. By introducing a globally perceptive Swin-Transformer network in the identification stage, the algorithm's robustness to changes in illumination, occlusion, complex textures, and scale differences is effectively enhanced. Furthermore, by combining the spatial constraints of 3D point clouds, high-precision block boundary and fracture parameters are obtained through ensemble operations, enabling automatic extraction of geometric quantities and spatial features. These identified quantities are deeply coupled with the linear elastic fracture mechanics (LEFM) criterion, achieving a mechanical closed loop from fracture identification, attitude determination, geometric modeling, and fracture parameter solving. This provides objective evidence for discovering potential fracture surfaces, identifying high-energy driving zones, and quantifying instability mechanisms.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for intelligent identification of rockfall hazards based on fracture mechanics, characterized in that, Includes the following steps: S1. Use drones to conduct multi-view close-up photography and directional cruise shooting of the target dangerous rock area to obtain high-resolution visible light image sequences. Obtain sparse / dense point clouds through structured beam method or multi-view stereo reconstruction to construct a three-dimensional geometric model of the dangerous rock slope. S2. The original image is preprocessed sequentially by background suppression, geometric correction, noise reduction, illumination equalization, color normalization and distortion correction. Combined with manual annotation, the image is cropped into fixed-size blocks, which are then converted into Tensor form after format conversion, normalization / standardization and data augmentation to construct training and inference samples. S3. Input the preprocessed image into the Swin-Transformer joint model. After multi-stage feature extraction, multi-head self-attention and translation window mechanism, global-local multi-scale features are obtained. The features are then fused by decoder upsampling and skip connections to output the segmentation mask of the dangerous rock block, the main control fracture and the structural surface. S4. Under the joint constraints of 3D point cloud and segmentation mask, extract the geometric parameters of the dangerous rock block and the geometric parameters of the fracture. Determine the configuration type of the dangerous rock based on the combination relationship of structural surfaces and the direction of intersection and assign configuration weights. S5. Combining self-weight, earthquake, pore water and additional load, solve for the normal stress and tangential stress of the potential fracture surface. Use the linear elastic fracture mechanics analytical formula to calculate the stress intensity factor and energy release rate of type I and type II stresses. Compare the critical energy release rate of the material to construct the fracture damage index and fracture safety factor. S6. Based on the threshold criteria of fracture safety factor and fracture damage index, the stability level of dangerous rock is divided and mapped to a multi-level early warning level. Combined with the on-site monitoring system, the risk of dangerous rock fracture instability is visualized, early warning and dynamically updated.

2. The method according to claim 1, characterized in that, The process of constructing the three-dimensional geometric model of the dangerous rock slope in S1 is as follows: High-resolution visible light image sequences of the target dangerous rock area were obtained by using UAV multi-view close-up photography. The exterior orientation elements of the image sequence were solved based on the structure-reconstruction-motion algorithm to construct a sparse point cloud. The sparse point cloud was then densified using a multi-view stereo matching algorithm to obtain a dense three-dimensional point cloud of the dangerous rock slope. The dense point cloud was then reconstructed into a surface based on the Poisson reconstruction algorithm to generate a three-dimensional geometric model of the dangerous rock slope. The three-dimensional coordinates of the sparse point cloud are obtained by bundle adjustment, and the solution model is as follows: In the formula, , The first The rotation matrix and translation vector of the image For the first The coordinates of a three-dimensional point For the first The three-dimensional point at the th t Projected coordinates on the image For the camera intrinsic parameter matrix, The total number of images, This represents the total number of three-dimensional feature points. This is the robust loss function.

3. The method according to claim 1, characterized in that, The preprocessing of the original image in S2 includes: An adaptive thresholding method is used for background suppression to separate the unstable rock slope from the non-target background area; geometric correction is performed using a camera distortion model, which is as follows: In the formula, To correct the pixel coordinates before, For the corrected pixel coordinates, Radial distance, These are the radial distortion coefficients, These are the tangential distortion coefficients; Gaussian filtering is used for noise reduction, histogram equalization is used for illumination equalization, and linear transformation based on RGB channels is used for color normalization. Combined with manual annotation, the image is cropped into fixed-size patches, which are then converted into Tensor form after format conversion, normalization / standardization, and data augmentation to construct training and inference samples.

4. The method according to claim 1, characterized in that, The processing procedure of the Swin-Transformer joint model in S3 is as follows: the input image is divided into 4×4 non-overlapping image blocks, and the initial feature matrix is ​​obtained through linear mapping. C represents the embedding dimension; Features are extracted in multiple stages using the SwinTransformer encoder, with a multi-head self-attention mechanism within each block: ,in, They are query, key, and value, respectively. For feature dimension, This is the relative position offset matrix; In the adjacent stages, the features of 2×2 image blocks are stitched together and downsampled. After the features are upsampled by the decoder and fused with skip connections, the segmentation mask is output through the Softmax function. ,in, The decoder outputs a feature map. This is the final segmentation result.

5. The method according to claim 1, characterized in that, The extraction process of geometric parameters and configuration weight calculation of unstable rock blocks and fractures in S4 is as follows: Under the joint constraints of 3D point cloud and segmentation mask, the contour point set, envelope volume, principal axis length, average thickness, and equivalent shape factor of the unstable rock mass are extracted. The formula for calculating the equivalent shape factor is as follows: In the formula, The visible surface area of ​​the dangerous rock. This represents the volume of the envelope of the dangerous rock. Extract the length, opening, dip, dip angle, and spatial distribution density of fractures. Determine the rockfall configuration type based on the structural plane combination relationship and intersection direction. The configuration weight coefficient is calculated using the following formula: In the formula, These are empirical weighting coefficients. The crack length is... The maximum fracture length in the target unstable rock area. The spatial distribution density of the fracture. This is the equivalent shape factor.

6. The method according to claim 1, characterized in that, The calculation process for the Type I and Type II stress intensity factors in S5 is as follows: Combining self-weight, earthquake, pore water, and additional loads, the normal and tangential stresses of the potential fracture surface are solved; based on the linear elastic fracture mechanics theory, the Type I and Type II stress intensity factors are calculated using the following formulas: In the formula, For the normal stress of the crack surface, For the shear stress on the crack surface, The equivalent crack length, These are dimensionless geometric coefficients.

7. The method according to claim 6, characterized in that, The calculation process for the energy release rate in S5 is as follows: Under the assumption of plane strain, the stress intensity factor is converted into the energy release rate per unit crack propagation area, and the calculation formula is: In the formula, The equivalent elastic modulus of the rock mass. Poisson's ratio, It is a type I stress intensity factor. It is a type II stress intensity factor.

8. The method according to claim 7, characterized in that, The calculation process for the critical energy release rate of the material in S5 is as follows: the critical energy release rate of the material is calculated based on the fracture toughness of the rock mass mixed mode, and the calculation formula is: In the formula, Type I fracture toughness, It has a type II fracture toughness. This represents the weighting coefficient for the contribution of type II toughness to the total fracture energy. The equivalent elastic modulus of the rock mass. It is Poisson's ratio.

9. The method according to claim 8, characterized in that, The calculation process for the LEFM fracture damage index and fracture safety factor in S5 is as follows: Based on the energy release rate and the material's critical energy release rate, the LEFM fracture damage index is calculated using the following formula: In the formula, Energy release rate, The critical energy release rate of the material. It is a type I stress intensity factor. It is a type II stress intensity factor. Type I fracture toughness, It has a type II fracture toughness. Type II toughness weighting coefficient; The fracture safety factor is calculated based on fracture damage indices, using the following formula: In the formula, LEFM fracture damage index This is the equivalent safety factor based on fracture mechanics.

10. The method according to claim 1, characterized in that, The relationship between stability level and fracture safety factor in S6 is as follows: Based on the threshold criteria of fracture safety factor and fracture damage index, the stability of dangerous rocks is divided into multiple levels, and each stability level corresponds to a different fracture safety factor range. The higher the fracture safety factor, the higher the stability level of the dangerous rock. The stability level is mapped to a multi-level early warning level, and combined with the on-site monitoring system, the visual early warning and dynamic update of the risk of dangerous rock fracture instability can be realized.

Images