A Multi-Source Sensing-Based Method for Monitoring and Early Warning of Slope Stability in Open-Pit Coal Mines

Abstract

This invention discloses a method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, relating to the fields of mine safety monitoring and geological disaster prevention technology. The method includes the following steps: S100, establishing a baseline model of solar radiation geometry and albedo, acquiring reference images of crack-free areas based on predicted solar altitude and azimuth angles, generating a time-varying illuminance gradient field within the monitoring period, and extracting the spatial location and temporal trajectory of the shadow front; S200, constructing a shadow-crack dual-domain decomposition model based on the illuminance gradient field, extracting the shadow feature subspace using brightness ratio, polarization ratio, and chromaticity invariance features, and generating a shadow probability map to suppress crack identification interference. This invention constructs a crack monitoring mechanism based on illumination modeling and multi-source sensing fusion, accurately identifying crack contour and length changes, dynamically optimizing imaging parameters, adapting to illumination fluctuations, achieving high-precision all-weather early warning, and significantly improving slope disaster prevention capabilities and response efficiency.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring and geological disaster prevention technology, specifically to a method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing. Background Technology

[0002] Multi-source sensing-based open-pit coal mine slope stability monitoring and early warning refers to the continuous, dynamic, and collaborative data collection and fusion analysis of slope deformation, crack development, structural characteristics, geological structure, and environmental factors during the open-pit coal mining process. This is achieved by comprehensively utilizing various types of sensing methods (such as ground-penetrating radar, 3D laser scanning, video image acquisition, GNSS displacement monitoring, tilt sensors, seismic wave monitoring, etc.). Through multi-source data cross-validation and trend prediction model calculation, the stability status and potential instability risk of the slope are assessed in real time. When instability signs are predicted or risk thresholds are triggered, graded early warning signals are automatically issued to enable targeted reinforcement, hazard mitigation, or personnel evacuation safety measures, thereby improving the timeliness and accuracy of open-pit coal mine slope disaster prevention and control.

[0003] The existing technology has the following shortcomings: In existing technologies, crack identification on open-pit coal mine slopes based on video monitoring images typically relies on edge detection algorithms with fixed thresholds to extract crack contours from areas of varying brightness in the images. However, during periods of rapid change in the angle of sunlight, such as sunrise, sunset, or instantaneous changes in illumination caused by rapid cloud movement, the light distribution on the slope surface can change significantly in a short time, causing the projection direction, position, and boundary morphology of shadows to dynamically shift. When the shadow boundary happens to fall within the crack area, its grayscale or color distribution will form a highly similar brightness contrast feature to the actual crack edge, resulting in overlap with the crack contour in the image. Because existing crack identification algorithms struggle to effectively distinguish between shadow boundaries caused by illumination changes and the actual crack edge, they are prone to misjudging the shadow contour information as an extension of the crack, leading to an abnormally sudden increase in crack length in the identification results. This misjudgment not only results in an inflated assessment of the crack propagation rate but also causes significant deviations in subsequent slope stability analysis and instability risk prediction, potentially triggering erroneous early warning signals and affecting the timeliness and accuracy of on-site safety decisions.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, comprising the following steps: S100, establish a baseline model of solar radiation geometry and albedo, collect reference images of crack-free areas based on the predicted results of solar altitude angle and azimuth angle, generate a time-varying illuminance gradient field within the monitoring period, and extract the spatial location and temporal change trajectory of the shadow front. S200 constructs a shadow crack dual-domain decomposition model based on the illuminance gradient field, extracts the shadow feature subspace using the brightness ratio, polarization ratio and chromaticity invariance features, and generates a shadow probability map to suppress crack identification interference. S300, guided by the shadow probability map, extracts the crack skeleton, performs local gradient weakening, and applies width smoothing, curvature constraint and endpoint inertial tracking constraint to generate an initial crack skeleton with temporal continuity. S400 introduces multi-source constraints based on crack skeleton, laser point cloud and tilt sensor, constructs deformation criteria based on crack opening amount and angle change, and uses adaptive gain reweighting method to suppress virtual length segments and improve the confidence of crack length measurement. S500 performs inverse temporal correction of the skeleton based on the shadow velocity field in image frames with a sudden increase in crack length, removes skeleton segments that move synchronously with the shadow, and reconstructs the crack length time series to ensure the authenticity of the change trend. The S600 feeds back the crack length residual and confidence changes to the baseline model, adaptively adjusts the albedo parameters and edge detection threshold, and updates the exposure sequence and camera pose to build a closed-loop dynamic monitoring and early warning mechanism.

[0007] Preferably, step S100 includes: Based on the longitude, latitude, and altitude of the slope location, the solar altitude angle and solar azimuth angle are calculated every minute within the monitoring period to form a solar motion path sequence with time resolution. Based on the slope orientation, determine the incident direction and angular distribution of sunlight acting on the slope surface at various times; Images of a benchmark area of ​​a slope without cracks were collected under multiple time periods, angles, and exposure conditions. After brightness normalization, grayscale response was extracted to establish an albedo variation model. Based on the brightness difference between the albedo model and the measured image, the spatial location and temporal trajectory of the shadow front are extracted to generate a time-varying illuminance gradient field covering the entire monitoring period.

[0008] Preferably, step S200 includes: The image to be identified is spatially registered with the illumination gradient image at the corresponding time to establish a consistent spatial reference. Extract the brightness ratio feature, polarization response feature and chromaticity invariance feature of each pixel to construct a three-dimensional feature vector; The three-dimensional feature vectors are standardized and fused to calculate the shadow probability value and generate a shadow probability map of the same size as the image. In the subsequent crack recognition process, gradient suppression and skeleton path avoidance are performed on pixel regions with shadow probability values ​​higher than a set threshold to reduce the interference of shadows on crack recognition.

[0009] Preferably, step S300 includes: Based on the probability values ​​of each pixel in the shadow probability map, a weighted weakening process is performed on the local gray-level gradient in the image. The gradient response intensity is reduced in the region with a high shadow probability value, generating an image gradient response map that suppresses the influence of illumination interference, which is used for subsequent crack path extraction and structure judgment. Based on crack width smoothing constraints, abnormal width path segments are identified and eliminated, while candidate paths with structural stability are retained. Curvature constraints and endpoint inertial tracking are applied to candidate paths to filter out paths with abrupt turning and enhance the consistency of paths in time series. The retained path is compressed into a single-pixel-width skeleton line, and auxiliary branches with low signal strength and abrupt changes in direction are removed to generate an initial crack skeleton with continuous structure and stable time sequence.

[0010] Preferably, step S400 includes: Spatial registration was performed between the initial crack skeleton path and the lidar point cloud data. The crack opening change information at the skeleton line segment was extracted to identify the skeleton region with a continuous expansion trend. The angle change value of the tilt sensor is projected into the image space, and the tilt change response at the skeleton segment is extracted to determine whether there is real displacement in the corresponding rock mass area. The crack skeleton was divided into multiple line segment units, and confidence scores were obtained based on point cloud deformation index, tilt response index and image shadow interference index, respectively. Based on the confidence score results, the skeleton segments with scores below the threshold and located in high shadow probability areas are length-reduced to improve the overall confidence of crack length measurement.

[0011] Preferably, step S500 includes: Identify image frames with sudden increases in crack length, extract the spatial location and confidence score of the newly added crack skeleton segment, and calculate the crack length difference by combining the image frames before and after the crack. Based on the established shadow velocity field, the shadow boundary evolution trajectory is traced back for no less than five frames before and after the sudden increase in image frames, and the advancement path of the shadow front is extracted. Spatial overlap analysis was performed on the crack skeleton segment and the shadow advancement path to identify completely overlapping skeleton segments. Based on their confidence scores, these segments were determined to be pseudo-crack segments and were removed. The remaining skeleton segments are reconnected and their lengths are recalculated to correct the crack length time series, and structured removal records are output.

[0012] Preferably, step S600 includes: The corrected crack length residuals and confidence changes in the crack identification results are collected and categorized according to image frame index, timestamp, solar altitude angle, shooting direction, and slope area location. The accuracy of the albedo parameter setting under the solar angle is judged based on the feedback results. If there is a misjudgment, the brightness distribution curve in the reference image is re-analyzed and the albedo parameter is corrected. Based on the feedback results, the edge detection threshold and the exposure time and aperture value during image acquisition are adjusted; Based on the spatial density distribution of misjudged skeleton segments in the image, the camera's pitch angle, rotation angle, and installation position are fine-tuned to improve image quality and crack recognition accuracy.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention establishes a sustainable and convergent intelligent monitoring mechanism by introducing illuminance modeling driven by solar altitude and azimuth angles, constructing a shadow probability map to suppress interference features, fusing laser point cloud and tilt angle information for multi-source verification, and dynamically optimizing edge detection and imaging parameters through a feedback closed loop. This mechanism can accurately identify the true contour and length variation trend of cracks, significantly improving the reliability of crack development trend curves. It can also proactively adapt to fluctuations in ambient light, achieving high-precision slope deformation identification and early warning output around the clock, greatly enhancing the disaster prevention and control capabilities and on-site response efficiency of open-pit coal mines in actual production. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, as described in this invention. Detailed Implementation

[0016] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0017] This invention provides, for example Figure 1 The method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, as shown, includes the following steps: S100, establish a baseline model of solar radiation geometry and albedo, collect reference surface images of crack-free areas based on the predicted results of solar altitude angle and azimuth angle, generate a time-varying illuminance gradient field covering the entire monitoring period, and extract the spatial location and temporal change trajectory of the shadow front from it. To address the crack identification error caused by rapid changes in natural lighting conditions in open-pit coal mine slope video monitoring, a technical approach combining solar radiation geometry and albedo modeling is proposed. By constructing an illumination gradient field with temporal continuity and spatial distribution characteristics, and extracting the evolution trajectory of the shadow front, a spatial-temporal background model for crack misjudgment under the influence of illumination is established. This provides accurate prior conditions for subsequent shadow-crack separation, crack skeleton extraction, and stability assessment. The implementation mainly includes the following steps: Based on the known longitude, latitude, and altitude of the open-pit coal mine slope, and using a standard solar position estimation method, the solar altitude angle and solar azimuth angle at each moment during the monitoring period are calculated, forming a time-resolution sequence of solar motion paths. This sequence covers the entire observation period from sunrise to sunset and continuously describes the solar incidence direction at minute-level time intervals. Based on this, and combined with the actual orientation information of the slope, the incidence direction and angular distribution of sunlight acting on the slope surface at specific times are determined.

[0018] A complete area without cracks on the slope surface was selected as the reference surface. This area should meet the following conditions: there are no visible structural cracks, no mine roadways or artificial accumulations obstructing the view, and the possibility of consistently acquiring data on changes in illumination at different time periods. After the reference surface area was selected, images of the area were acquired at multiple time periods, angles, and exposure parameters using fixedly installed video acquisition equipment within the selected monitoring time period. The image acquisition interval was no more than 5 minutes, and the continuous cycle was no less than 7 24-hour cycles to cover various natural illumination environmental conditions such as sunny days, cloudy days, partly cloudy days, and intermittent cloud cover.

[0019] The acquired reference surface images are input into the processing flow. The grayscale values ​​in the images are uniformly normalized to eliminate brightness drift caused by camera exposure settings, sensor aging, or dynamic environmental interference. Each image is processed using frame-by-frame averaging and local contrast enhancement to ensure comparability in grayscale contrast between images acquired at different time points. After brightness normalization, the grayscale values ​​of each pixel on the reference surface in each image frame are extracted, and a grayscale variation curve is established in chronological order. Statistical aggregation of all pixels is performed to obtain the average grayscale response under different solar altitude and azimuth angles. This grayscale response curve reflects the reflectivity of the slope surface material under natural light, thus constructing an albedo variation model.

[0020] This albedo variation model characterizes the brightness response trend of a baseline slope area throughout the monitoring period under specific geographical conditions, surface material properties, and the influence of incident light angle. Through model calculations, the grayscale level that different areas on the slope should exhibit under theoretical illumination can be predicted at any given time point, thus providing fundamental data support for identifying shaded and unshaded areas.

[0021] Based on the established albedo model and actual acquired images, the theoretical brightness is compared with the measured images at the pixel level to identify areas in the actual images where the brightness is lower than the predicted value. Regions showing significant brightness decay in consecutive frames, and possessing spatially closed contours or sharp gradient transitions, are identified as potentially occluded by shadows. The spatial movement path of this region at multiple time points is analyzed in conjunction with the previously established solar incidence direction to further verify that the region is a genuine shadow area caused by solar occlusion. Edge tracking processing is performed on the shadow boundaries to extract their positions, which are then recorded as pixel coordinates.

[0022] By repeating the identification and extraction process over time, a continuous trajectory of the shadow front in the image space can be obtained throughout the entire monitoring period. This trajectory is organized as a time series, recording the spatial location, morphological changes, and direction of advancement of the shadow boundary every minute, forming a continuous shadow front evolution dataset. This dataset is used to characterize the spatial projection of dynamic changes in illumination, accurately reflecting the range and rate of change of the slope surface affected by natural illumination.

[0023] The illuminance gradient field at each moment is matched and integrated with the corresponding shadow front trajectory to form a three-dimensional data structure. This structure, with time as the main axis, records the theoretical illuminance distribution and shadow front position of the slope surface at each moment. This joint model not only provides a reference for the illuminance and shadow distribution at the current moment but also has the ability to project and predict the future for several minutes. By analyzing the rate of change of illuminance and the direction of shadow advance, it is possible to predict whether a certain location may be covered by shadow at a specific time in the future.

[0024] Based on this illumination-shadow model, when performing crack identification processing, the bright and dark boundaries located in the predicted shadow area of ​​the current image can be excluded from the crack edge extraction objects, thereby effectively reducing the interference of shadow pseudo edges and improving the accuracy and stability of crack identification. Unlike the existing technology that only relies on image grayscale gradients for edge detection, this invention introduces a predictable illumination model and a real shadow trajectory to achieve physical scene constraints on the crack identification results, enhancing the method's anti-interference ability and practicality.

[0025] The core function of this step is to provide stable and reliable spatiotemporal prior constraints for distinguishing between real cracks and pseudo-boundaries of shadows caused by changes in illumination during the crack identification process of open-pit coal mine slopes. This effectively solves the problem of misidentification caused by shadow interference in traditional crack identification algorithms under scenarios with rapidly changing natural illumination conditions. In open-pit environments, due to the continuous changes in the sun's position and the complex influence of weather conditions, such as sunrise, sunset, and cloud movement, the illumination distribution on the slope surface can change significantly in a short period of time, resulting in large fluctuations in brightness and contrast in the image. Especially when the shadow boundary falls exactly in the crack area, the brightness-dark contrast features it forms in the image are easily confused with the edge of the real crack, becoming an interference factor in the image recognition algorithm. Through the solar radiation geometry model and albedo model constructed in this step, not only can the incident direction and intensity of illumination at different time points be accurately predicted throughout the entire monitoring period, but also, combined with the optical reflection characteristics of the slope material, a time-varying illuminance gradient field can be established, thereby extracting the dynamic trajectory of the shadow front. This prior information on shadows, possessing both temporal continuity and spatial distribution characteristics, can be introduced as an interference suppression condition during the crack identification stage. It accurately eliminates unstructured edge interference caused by shadow projection, significantly improving the accuracy and robustness of crack detection. This step not only lays the foundation for the illumination environment model in crack image processing but also forms the starting point for the closed-loop optimization mechanism in the method of this invention, making it an indispensable and crucial foundational step in the entire slope monitoring and early warning process.

[0026] S200, based on the generated time-varying illuminance gradient field, constructs a shadow crack dual-domain decomposition model, uses the luminance ratio distribution, polarization ratio features and chromaticity invariance features to extract the feature subspace of the shadow region, and generates a shadow probability map as the suppression input for subsequent crack identification; Based on the established baseline model of solar radiation geometry and albedo, and generating the temporal illuminance gradient field and shadow front trajectory of the slope surface throughout the monitoring period, a shadow-crack dual-domain decomposition method is proposed to effectively separate the shadow region from the crack region. This method constructs a feature subspace of the shadow based on image brightness structure, polarization response characteristics, and color stability, thereby generating a quantifiable shadow probability map, providing auxiliary input for interference suppression in subsequent crack identification steps. The process includes the following steps: In the previous phase of implementation, a time-synchronized illuminance gradient field was obtained. This field, in pixels, reflects the theoretical light intensity distribution of various regions on the slope surface at a specific time. To ensure that subsequent shadow recognition of each pixel can be based on the real lighting background, the crack image to be identified needs to be spatially aligned with the corresponding time-varying illuminance gradient image. Specifically, firstly, at least six alignment reference points with stable shapes and known spatial locations are selected in both the crack image and the illuminance gradient image. These can be points such as time-invariant corners in the rock mass, edges of cavities, centers of color spots, or intersections of surface boundary polygons. An affine transformation matrix is ​​established by calculating the corresponding coordinates of the reference points in the two images. This matrix is ​​then applied to the crack image, spatially transforming it to align it pixel-level with the illuminance image in two-dimensional space. After registration, any pixel in the crack image can be accurately mapped to the illuminance gradient map, ensuring the accuracy and consistency of subsequent feature calculations.

[0027] This study aims to clarify the physical representation characteristics of shadows in image space and concretize these characteristics into calculable image information. First, regarding the brightness ratio characteristic, each pixel and its surrounding pixels within a 5×5 neighborhood are selected. The ratio between the brightness of the central pixel and the average brightness of its neighborhood is calculated to assess whether the pixel exhibits a sudden brightness change without accompanying structural boundary changes. Second, regarding the polarization response characteristic, during image acquisition, an imaging device equipped with a rotatable polarizing filter is used to capture at least three images at the same time point from different polarization angles. The degree of grayscale change for each pixel in the images at different polarization angles is calculated, and regions exhibiting significant responses with polarization angle changes are identified based on the differences in grayscale response. These regions are typically shadow areas on non-metallic mineral surfaces where the incident light angle changes significantly. Third, regarding the chromaticity invariance characteristic, the image is converted from the standard red-green-blue color space to a chromaticity-luminance-saturation space, separating the chromaticity channel. The chromaticity value changes of the same pixel are compared at different time periods. If the chromaticity value remains stable but the brightness decreases, it can be inferred that the pixel is in a shadow area caused by changes in illumination. The features of the above three dimensions are recorded as a brightness ratio diagram, a polarization response diagram, and a chromaticity variation diagram, respectively.

[0028] After extracting features for brightness ratio, polarization response, and chromaticity stability, a three-dimensional feature vector is created for each image pixel. This vector corresponds to the pixel's value in each of the three feature maps. All feature vectors are standardized to normalize their value range to [0,1] for easier subsequent fusion. For each pixel's three-dimensional feature vector, its comprehensive shadow probability value is calculated based on empirical weights (e.g., assigned values ​​of 0.4, 0.3, and 0.3 respectively). If the value exceeds a preset threshold (e.g., 0.7), the pixel is marked as a high-probability shadow region; if it is below another threshold (e.g., 0.3), it is marked as a non-shadow region; and values ​​in between are considered medium-probability shadow regions. Finally, a probability image of the same size as the crack image is generated, where each pixel value fluctuates between 0 and 1, representing the probability that the pixel is a shadow region. To improve image spatial consistency, local smoothing is performed on the generated shadow probability map. The specific processing method is as follows: take each pixel as the center, select its surrounding 3×3 neighborhood, calculate the weighted average of the neighborhood probability values ​​as the new value of the center pixel, enhance the spatial continuity of the probability map and suppress the interference of random noise.

[0029] After generating the shadow probability map, its spatial distribution characteristics are used as input guidance information in the crack edge extraction process. During crack identification, for pixel regions with a value higher than 0.7 in the shadow probability map, the response intensity of these regions is suppressed during image gradient calculation. For example, the image gradient value in these regions is multiplied by a coefficient less than 1 to reduce the likelihood of them being identified as edges. Simultaneously, for continuously distributed high-probability shadow regions, their boundaries are not prioritized for skeleton growth paths, thus constructing "obstacle avoidance regions" in the image. In this way, brightness changes caused by shadows are avoided from being misidentified as crack extension lines, ensuring that the extracted crack skeleton has a genuine physical origin and spatial coherence.

[0030] The core function of this step is to construct a shadow probability map that accurately represents the spatial distribution of shadows, thereby eliminating the interference of natural lighting variations on crack identification results and ensuring the accuracy and stability of subsequent crack extraction. In the actual monitoring environment of open-pit coal mine slopes, the slope surface is affected by factors such as changes in solar altitude angle, azimuth drift, and cloud cover. The appearance of shadows in images is highly dynamic and uncertain. These shadow boundaries are often very similar to real cracks in brightness, grayscale, and edge morphology. In existing technologies, most methods use fixed thresholds or grayscale changes to determine crack edges, lacking the identification of the physical characteristics of shadows. This easily leads to misidentification of shadows as cracks, resulting in abnormally large increases in crack length recognition, ultimately affecting the reliability of slope stability analysis and the accuracy of early warning judgments.

[0031] This step extracts the physical optical features of shadows in the image from three dimensions: brightness ratio distribution, polarization response difference, and chromaticity invariance. Combined with the time-varying illuminance gradient field established in the previous stage and the image spatial registration results, it achieves a shadow presence probability assessment for each image pixel. The resulting shadow probability map clearly indicates the shadow distribution intensity and confidence level of each region in the image at the pixel level. This allows for targeted intervention in the subsequent crack identification stage on these high-probability shadow areas, such as reducing their response weight in edge detection and restricting their path generation in skeleton extraction, thereby effectively avoiding misjudgments in crack identification.

[0032] Therefore, this step not only technically achieves visualization, quantification, and prior modeling of the shaded area, but also forms a key bridge connecting illumination modeling and crack identification in the overall methodological logic. It greatly enhances the anti-interference capability and environmental adaptability of the crack identification process and is an important foundational link for realizing intelligent monitoring of open slope stability under multi-source sensing.

[0033] S300 extracts the crack skeleton under the guidance of the shadow probability map, performs shadow weight-based weakening processing on the local gradient information in the image, and combines crack width smoothing constraint, crack curvature constraint and crack endpoint inertial tracking mechanism to generate an initial crack skeleton with time continuity. Based on the shadow probability map constructed in the preceding steps, to avoid misjudgment of cracks caused by changes in natural lighting conditions and to ensure the continuity and authenticity of the extracted crack skeleton in terms of spatial morphology and temporal evolution, a stepwise crack skeleton construction method is further proposed. This method, by introducing shadow weight reduction, spatial structure constraints, morphological coherence control, and time tracking mechanisms, achieves the extraction of an initial crack skeleton that conforms to physical characteristics and has engineering significance. Specifically, it includes the following steps: In the previous stage, a shadow probability map spatially registered with the current crack monitoring image was obtained. In this probability map, each pixel corresponds to a value that reflects the probability that the location is a shadowed area. To suppress the misidentification of unstructured gray-level abrupt changes caused by shadows as crack edges, the gray-level gradient information needs to be weighted and weakened during image processing.

[0034] The specific processing method is as follows: The crack monitoring image is scanned pixel by pixel. For each pixel, the grayscale difference between it and its four adjacent pixels (top, bottom, left, and right) is obtained to form the local gradient response value of that pixel. Simultaneously, the probability value corresponding to that pixel in the shadow probability map is read and recorded as the shadow weight factor. For locations where the shadow weight factor value is higher than a preset threshold (e.g., 0.7), its local gradient response value is weakened proportionally, for example, by multiplying the original gradient by 0.5; for locations where the shadow weight factor value is lower than the threshold (e.g., 0.3), its original gradient value remains unchanged. This processing method systematically reduces the influence of shadow areas on subsequent edge discrimination, ensuring that the construction of the crack skeleton is not misled by false illumination boundaries.

[0035] After grayscale gradient weighting, a gradient response map of the crack image is generated, which is less affected by shadow interference. In this map, the crack typically appears as a linear region with a continuous distribution of high response intensity. To further clarify the crack path, it is necessary to extract high-response line segments with continuous gradient intensity and consistent direction as candidate crack paths.

[0036] During the path extraction process, a crack width smoothing constraint is set: First, along the length of each candidate path, the local width value is calculated every three pixels, defined as the distance between the significant gray-level gradient regions on the left and right sides; then, the standard deviation of the width over the entire path is calculated; if the width variation in a certain path segment exceeds twice the average width of the entire path, or if there is a single segment with a sudden width change (e.g., a width change greater than 5 pixels), then that segment is determined to be an abnormal path segment. For these path segments, boundary shrinkage processing is performed to preserve their gradient core regions; furthermore, if the width fluctuation of the segment is too large and the structure is unstable, it is completely removed. This method ensures that the extracted crack paths are continuous and convergent in spatial scale, avoiding misjudgments at locations with drastic changes in slope surface texture or rock mass joints.

[0037] In actual development, cracks generally extend slowly along geological structural planes and weak surfaces of rock masses, and their representation in images should be naturally curved line segments with gentle curvature changes. Against this backdrop, to avoid sudden changes or path jumps in the skeleton path due to image noise, boundary breaks, or other reasons, curvature control of the path direction is necessary.

[0038] The specific method is as follows: For each crack path, the angle change is calculated in groups of three consecutive pixels. If the continuous angle of a certain path exceeds the set limit (e.g., the angle of every three pixels is greater than 45 degrees), the path is considered to have structural discontinuity and needs to be adjusted or broken and reconstructed. During the path extension process, the system prioritizes pixels with smaller local directional changes as extension targets to ensure that the skeleton moves smoothly and naturally.

[0039] Furthermore, to enhance the stability of the skeleton over time, an endpoint inertial tracking mechanism is introduced: the direction vector of the crack skeleton endpoint in the previous moment's image is read, and this vector is used as the priority direction for the path growth direction in the current frame. In the path extension determination, the pixel region with the smallest angle to this direction is preferentially selected as the extension direction, ensuring that the crack path maintains physical consistency between consecutive frames, which conforms to the real situation of slow crack evolution.

[0040] After the aforementioned three steps, the obtained crack path possesses spatial coherence, structural stability, and temporal consistency. Based on this, the high-response regions within the path need to be further compressed into a single-pixel-width skeleton structure to facilitate subsequent crack length calculation, dynamic change tracking, and multi-source verification comparison.

[0041] The specific skeleton extraction method is as follows: For each crack path, the pixel with the maximum local gradient value is selected as the skeleton node by scanning horizontally or vertically, and all nodes are connected in spatial order to generate the main line of the crack skeleton. For areas with branch paths, the main directional path is retained, and auxiliary branches with weak signal strength and large curvature changes are discarded to ensure that the skeleton structure is simple and accurate.

[0042] The resulting initial crack skeleton structure has clear start and end points, a clear spatial orientation, and single-pixel-level path accuracy, which can serve as the input basis for the next stage of crack authenticity verification, length correction, and dynamic trend judgment.

[0043] The core function of this step is to construct a structurally continuous, temporally stable, and physically reliable crack skeleton structure in the complex lighting environment and significant shadow interference of open-pit coal mines. This provides a high-quality input foundation for subsequent crack authenticity verification, length calculation, and multi-source information fusion. In open-pit coal mine slope monitoring, cracks are an important precursor to slope instability. The continuity of their morphology, the trend of length variation, and the direction of path evolution directly affect the accuracy of slope stability assessment and disaster early warning. However, due to the significant influence of factors such as solar altitude angle, azimuth angle, and cloud movement on the slope environment, the shadow distribution in the image changes frequently, forming brightness abrupt boundaries that are highly similar to the actual crack edges. In this case, traditional image processing methods often misjudge shadow boundaries as crack edges, leading to false extension of crack paths, structural distortion, and even misjudgment of crack length, greatly interfering with the judgment of the early warning system.

[0044] To address this issue, this step, guided by the shadow probability map, reduces the weight of the image grayscale gradient, accurately identifying true structural edges unaffected by shadows and minimizing the involvement of false edges. Building upon this, by setting a smooth constraint on crack width, the abnormal path widening problem caused by background stray textures is effectively eliminated. Furthermore, by combining crack path curvature constraints and endpoint inertial tracking mechanisms, the coherence of the spatial structure and the stability of the temporal series are achieved, allowing the skeleton to maintain a natural extension trend between image frames. This processing not only solves the structural discontinuity problems such as crack morphology jumps and path breaks, but also improves the consistency of crack extraction across multiple frames.

[0045] Therefore, this step is crucial for constructing a realistic, complete, and stable crack representation structure. The obtained crack skeleton not only accurately reflects the location, orientation, and morphological characteristics of cracks on the slope surface, but also provides clear, accurate, and traceable basic data for subsequent cross-validation with laser point cloud and tilt monitoring data, correction of virtual length segments, and trend judgment in early warning models, playing a key role in bridging the gap between previous and subsequent steps.

[0046] S400 introduces multi-source sensing cross constraints through the generated initial crack skeleton, uses the change information of crack opening in lidar point cloud data and the angle change information recorded by tilt sensor as crack deformation criteria, and uses the adaptive gain reweighting method to weaken the virtual length segment of crack caused by shadow interference, thereby improving the confidence of crack length measurement. Based on the initial crack skeleton structure with spatial continuity and temporal consistency extracted in the aforementioned steps, to avoid misleading crack extraction results due to shadow interference caused by drastic changes in illumination, a mechanism for verifying the authenticity of crack skeleton segments is further established by introducing lidar point cloud data and tilt sensor data, and suspected imaginary segments are suppressed. This method constructs a joint criterion between image data and physical sensor data, realizing multi-source verification and weighted correction of crack structures. The method includes the following steps: Three-dimensional point cloud data of the slope surface was obtained using a pre-installed lidar device. This point cloud data has a spatial resolution better than 0.01 meters and covers the area where the crack skeleton was extracted in the image. To establish a spatial mapping relationship between the skeleton path and the point cloud data, the three-dimensional coordinates (X, Y, Z) of the lidar point cloud were first transformed by perspective projection based on known external camera parameters (including camera and lidar installation position, orientation, rotation angle, etc.) and internal parameters (including focal length, principal point, distortion coefficient, etc.) to obtain the two-dimensional coordinate positions (u, v) in the image plane. Subsequently, the coordinates of each pixel point on the crack skeleton path were matched with the projected point cloud coordinates to establish a one-to-one correspondence at the pixel level.

[0047] After the mapping is completed, for each crack skeleton segment, sampling bands of a predetermined width (e.g., 0.1 meters) are extended on both sides of its vertical direction. The changes in point cloud density and distance distribution in the two sides within the same time interval are extracted. By comparing the average spacing between the point clouds on both sides of the location, the opening amount of the crack at that location can be calculated. To enhance the dynamic perception capability over time, the difference in the opening amount at the same location in two consecutive radar scans is compared. If the difference exceeds a preset threshold (e.g., 2 mm) and shows an increasing trend in both time periods, it indicates that there is real deformation at that location, providing a physical basis for crack growth.

[0048] Several tilt sensors are installed at key locations on the slope to monitor minute changes in the vertical and horizontal attitude of the rock mass. To use the tilt information for auxiliary verification of crack paths in images, a unified transformation is needed between the sensor's geographic coordinates (e.g., GPS latitude, longitude, and elevation) and its pixel coordinates in the image space. Based on the camera's shooting direction, installation calibration parameters, and topographic mapping data, the pixel projection position of the sensor in the image is calculated. An influence area (e.g., with a radius of at least 20 pixels) is then constructed centered on this position to match the mapping relationship between the skeleton path segment and the corresponding sensor.

[0049] In terms of time, using the timestamp of image acquisition as a benchmark, an angle change sequence within at least 20 minutes before and after the image is extracted from the data records of the tilt sensor, and the average value and maximum rate of change of the angle are calculated. If the tilt angle of a sensor changes continuously during the image acquisition period, and the magnitude of the change is higher than the background disturbance level (e.g., the change value is greater than 0.05 degrees), it can be preliminarily determined that there is structural displacement in the rock mass near the sensor. Combining this with whether the skeleton path in the image passes through this area can serve as an auxiliary criterion for determining whether this skeleton segment is a real crack.

[0050] Based on the spatial data registration and crack deformation index extraction completed in the first two steps, the crack skeleton line in the image is divided into multiple line segment units of equal length (e.g., every 5 pixels). Confidence calculation and annotation are performed on each line segment. Each line segment is scored based on the following three factors: Whether it matches the opening volume data (point cloud deformation index) with a stable growth trend. Whether it is located in a region of continuous angle change (tilt response index); Is it located in a high-probability interference zone in the shadow probability map (image uncertainty index)?

[0051] Each indicator can be assigned a specific scoring standard. For example, point cloud support earns 1 point, tilt angle support earns 1 point, and shadow interference deducts 0.5 points. The maximum score is 2 points, and the minimum score is 0 points. This results in a confidence score table for the crack skeleton segment.

[0052] Based on the confidence score results, the crack skeleton segments are weighted and corrected: skeleton segments with a score of 1.5 or higher are considered to have high confidence and their original length is retained for subsequent crack statistics; skeleton segments with a score below 1.0 and located in high shadow probability areas are considered to be dummy length segments that may be caused by illumination interference, and their length is multiplied by a reduction factor of 0.3 before being included in the total length calculation; skeleton segments with a score of 0 are marked and can be further evaluated in subsequent steps to determine whether they should be completely removed based on stability.

[0053] After weighted correction of all crack skeleton segments, the lengths of the retained segments are accumulated to obtain the crack length value of the current image frame. Simultaneously, the spatial location, confidence score, correction status, and data support source of each skeleton segment are recorded, and a structured crack identification result table is constructed. This result serves not only as the crack identification output for the current frame but also as a reference input for crack change trend analysis in subsequent frames. If a sudden change in crack length occurs in multiple consecutive image frames, the system can retrospectively examine whether the abrupt change was caused by a low-confidence skeleton and, combined with actual point cloud and tilt angle changes, determine whether it is due to actual crack development or identification errors caused by illumination, thereby deciding whether to trigger an early warning signal.

[0054] The core function of this step is to utilize multi-source sensing techniques to verify the authenticity and correct the data of the crack skeleton structure extracted from the image, thereby eliminating fictitious crack length segments caused by shadow interference and improving the reliability and confidence of crack length measurement results. In the actual monitoring of open-pit coal mine slopes, the lighting environment in video images is strongly affected by factors such as solar altitude angle, azimuth angle, and cloud movement, easily forming shadow boundaries in the image that resemble the actual crack edges. These shadow boundaries exhibit strong brightness abrupt changes in the image gradient, thus being incorrectly identified as part of the crack extension path, leading to a false extension of the crack skeleton. If not identified and processed, these fictitious length segments can cause incorrect judgments about crack growth trends, misleading subsequent slope stability analysis and potentially triggering erroneous early warning decisions.

[0055] To address the aforementioned issues, this step incorporates 3D point cloud data acquired by LiDAR and attitude change data acquired by tilt sensors into the crack analysis process, constructing a cross-validation mechanism that integrates image information and physical perception information. Specifically, LiDAR point cloud data quantifies changes in crack opening size, reflecting whether the crack has undergone actual spatial displacement; while tilt sensors detect minute shifts in the rock mass structure, assisting in determining whether changes have occurred in the surrounding structure. By spatially matching these two types of physical information with the crack skeleton in the image, the authenticity and confidence scores of each segment of the crack path are assessed, and the identified skeleton length is then weighted and corrected or eliminated. Compared to traditional methods that rely solely on image processing for crack identification, this step significantly enhances the physical reliability of the crack identification results and improves the robustness and adaptability of the entire crack monitoring process.

[0056] Therefore, this step plays a crucial role in "image result verification and structural error correction" within the entire slope stability monitoring method. It not only improves the accuracy and reliability of crack identification but also provides more reliable foundational data for subsequent trend modeling, abrupt change identification, and early warning response, serving as an important technical bridge from "perception identification" to "structural verification."

[0057] S500 performs shadow front replay reestimation in monitoring image frames where a sudden increase in crack length is detected, performs inverse time-series correction of crack skeleton based on previously generated shadow velocity field, removes crack skeleton segments that move synchronously with the shadow, and reconstructs crack length time series to reflect the true trend of change. To address the issue of abrupt increases in crack length that may occur during image crack identification, a method based on shadow front playback and inverse temporal correction of the crack skeleton is proposed. This method utilizes a previously generated shadow motion velocity field to perform temporal backtracking and spatial reestimation on image frames in the image sequence where crack length increases significantly. It identifies and removes crack skeleton segments that move synchronously with the shadow boundary, thereby reconstructing a realistic and reliable temporal variation curve of the crack length. The method includes the following steps: Building upon the crack skeleton extraction and crack length statistics completed in the previous stage, time-series data of crack lengths in consecutive image frames within the monitoring period are obtained. Pairwise differences are calculated between the crack lengths of adjacent frames to identify key frames exhibiting sudden increases in length. The criteria for identification are: the crack length in a given image frame increases by more than 30% compared to the length of the previous frame, and the newly added crack skeleton segment within that frame scores in a low confidence range (e.g., less than 1 point). The timestamp and image number of the frame with the sudden increase are recorded, and the location range, grayscale features, and confidence score of the newly added crack skeleton segment within that frame are extracted to provide processing targets for subsequent shadow trajectory tracing.

[0058] The preceding steps have established shadow velocity field information covering the entire monitoring period, including the movement speed, direction, and path of the shadow boundary in the image space at each time point. Image data from at least five frames before and after the sudden increase in frame are selected, and records of shadow boundary position changes within the corresponding time period are extracted. These records originate from pixel regions in the image with a shadow probability value greater than 0.7; their boundary contours are extracted and sorted according to time series to form the shadow boundary evolution path.

[0059] For each frame, the position of the center mass point of the shadow boundary is extracted, and the spatial displacement vector in its adjacent frames is calculated to obtain the magnitude and direction of the shadow's velocity in image space. These vectors are then combined to form complete temporal trajectory data, creating the advancement trajectory of the shadow front before and after the sudden increase frame. Spatial overlap analysis is performed on this trajectory and the position of the crack skeleton segment in the sudden increase frame to determine whether the skeleton segment is on the path traversed by the shadow front and whether it exhibits directional consistency and synchronization.

[0060] The crack skeleton segments in the burst frames are categorized into three types based on their spatial overlap with the shadow leading edge's trajectory: complete overlap, partial overlap, and no overlap. Complete overlap indicates that the crack skeleton segment is spatially continuous within the shadow leading edge's advance zone, and no structurally connected skeleton extension appears in the preceding and following image frames. Partial overlap indicates that the skeleton path overlaps with the shadow boundary only in certain areas. No overlap refers to skeleton segments with no clear spatial connection to the shadow.

[0061] For completely overlapping skeleton segments, considering their confidence scores (e.g., lack of support from changes in laser point cloud opening amount or tilt angle) and the fact that they only appear in sudden frame jumps, they can be identified as pseudo-crack segments caused by shadow perturbations. These segments are removed from the skeleton path, disconnecting their path from preceding and following skeleton segments. For partially overlapping skeleton segments, a retention or correction strategy is further determined based on their local confidence and spatial morphology. Skeleton segments without overlap retain their original state.

[0062] After the removal operation is performed, the remaining crack skeleton segments are reconnected to ensure the continuity and integrity of the overall skeleton structure and to avoid problems such as breakage or reversal that may affect subsequent processing.

[0063] The lengths of the corrected and retained crack skeleton segments are accumulated to obtain the effective total crack length value of the current image frame. This value replaces the crack length record in the original burst frame, and the updated length value is written into the crack length time series as the correction result. Simultaneously, trend analysis is performed on the crack growth curve to determine whether the current crack extension has continuity. If, after removing the synchronous shadow skeleton segment, the crack length continues to gradually increase, and the skeleton structure possesses continuity and physical support information, then it can be determined that the crack is indeed in a stable growth phase; conversely, if the crack length returns to the normal range after the burst segment is removed, it can be determined that this growth is a misidentification phenomenon caused by changes in illumination.

[0064] In addition, to enhance the reference value of subsequent processing, the skeleton segment coordinates, shadow trajectory data, removal criteria (whether they overlap or lack physical support), and length changes before and after correction involved in this removal operation were recorded as structured data for subsequent self-learning and early warning criterion optimization of the crack recognition mechanism.

[0065] This step aims to identify and eliminate false crack identification caused by rapid shadow movement from a temporal perspective, ensuring the authenticity and stability of crack length change trends and providing highly reliable basic data support for slope stability assessment and early warning decisions. In the actual monitoring environment of open-pit coal mines, slope surfaces are frequently affected by changes in solar altitude angle, azimuth shift, and cloud cover, causing rapid and significant spatial movement of shadowed areas in images. This movement manifests as strong changes in the light-dark boundary in the image, which can easily be misidentified as new crack segments, especially in scenarios where grayscale gradients or brightness edges are used as key criteria for crack identification. This can lead to abnormally sudden increases in crack length in a particular frame. If these false skeleton segments are not identified and eliminated in time, the crack growth trend curve will be inflated, ultimately causing deviations in slope instability predictions based on this trend, and even prematurely triggering incorrect early warning signals, misleading on-site emergency response deployment, and affecting operational safety.

[0066] This step establishes a time-backtracking mechanism to re-estimate the motion trajectory of the shadow front on image frames that identify sudden crack increases. Combined with previously generated shadow velocity fields and crack confidence information, it performs spatial and temporal dual verification of the synchronicity between the crack skeleton segment and the shadow boundary, thereby accurately identifying pseudo-crack structures caused by shadow movement. By removing these "synchronized shadow skeleton segments," the crack skeleton is reconstructed, and the crack length time series is updated, ensuring that the final crack growth trend truly reflects the actual geological deformation evolution process.

[0067] Therefore, this step not only enhances the adaptability of the crack identification system to changes in dynamic lighting environments, but also achieves a leap from single-frame crack extraction results to multi-frame trend rationality judgment. Through qualitative and quantitative analysis of abrupt changes, it effectively improves the robustness and anti-interference capability of the overall monitoring system, and is an important supporting link to ensure the reliability of multi-source sensing monitoring results.

[0068] S600 feeds back the corrected crack length residual value and confidence change to the solar geometry and albedo baseline model. Based on the feedback results, it adaptively adjusts the albedo parameters and edge detection threshold, and simultaneously updates the exposure control timing and camera attitude fine-tuning parameters of the camera system, thereby realizing a continuously convergent closed-loop crack monitoring and slope early warning mechanism. To improve the accuracy and stability of crack identification results and avoid identification bias caused by fixed parameter settings, a parameter adaptive update mechanism based on identification result feedback is proposed. This mechanism uses the crack length residual value and crack confidence change after the previous stage of identification as input, and dynamically corrects several key parameters, including the albedo parameter in the baseline model, the edge extraction threshold during crack identification, the exposure time setting in the image acquisition stage, and the camera's installation angle. This forms a continuously converging, closed-loop crack monitoring process that can continuously optimize identification accuracy. The process includes the following steps: After extracting the crack skeleton structure, performing temporal correction, and reconstructing the length trend, the system records the difference between the original crack length and the reconstructed crack length after removing shadow interference in each image frame, and defines this difference as the crack length residual value. Simultaneously, the system extracts the confidence scores corresponding to the removed and retained segments, and calculates their mean, variance, and confidence trends to evaluate the changes in crack recognition accuracy and reliability under different lighting conditions.

[0069] For example, if the average confidence level of the crack skeleton segment drops significantly within a certain time period, and the residual value increases significantly after removal, it indicates that the identification process is severely affected by environmental interference; conversely, if the confidence level increases and the residual value decreases, it indicates that the current parameter settings are relatively reasonable. The above statistical data are categorized and organized according to dimensions such as image frame index, timestamp, solar altitude angle, shooting direction, and slope area coordinates to construct a complete feedback dataset.

[0070] Based on the correspondence between the residual value of each image frame in the feedback dataset and the shooting time and location, it is determined whether the albedo setting is reasonable at specific solar altitude and azimuth angles. If the crack identification residual frequently increases in images under certain illumination angles, it indicates that the current albedo setting cannot accurately reflect the actual reflective characteristics of the slope material, and should be corrected for the illumination angle range covered by that time period.

[0071] The correction method involves selecting pixel regions at corresponding angles and time periods from the original crack-free reference surface image, reanalyzing the brightness distribution curve, and adjusting the albedo coefficient based on the actual reflectivity of the surface material type (such as rock, gravel, and weathered soil). The adjusted albedo parameters are then used to generate the time-varying illuminance gradient field to improve the spatial accuracy of the shadow probability map, making the shadows and actual cracks more clearly distinguishable in terms of brightness changes.

[0072] During crack identification, the edge response intensity of the image is directly related to the exposure control of the acquisition device. In the feedback data, if image frames from multiple time periods simultaneously show a decrease in confidence and an increase in skeleton misjudgments, it indicates that the current edge detection parameters or image acquisition parameters are not suitable for the lighting environment at that time.

[0073] First, adjust the grayscale change threshold used for edge detection. For example, in images from early morning and late afternoon, shadow boundaries are often misjudged, so the detection threshold can be increased to weaken the response in low-contrast areas; while during the midday high-light period, the threshold needs to be decreased to enhance the ability to capture details.

[0074] Secondly, based on the time period in which low-confidence frames were concentrated in the feedback, the shooting conditions of the corresponding image frames were traced back, such as exposure time and shutter speed. If it was found that the overall brightness of the image was low and the shadow area was enlarged during a certain period, the exposure time should be extended and the aperture value should be appropriately increased to compensate for the brightness and improve the distinguishability of image details. Conversely, if overexposed areas appeared in the image, it indicated that the acquisition time was set too long and should be shortened.

[0075] By employing the above methods, the adaptability of image acquisition and edge response capabilities to changes in natural lighting is enhanced, thereby improving the basic quality assurance capabilities in the crack identification stage.

[0076] When processing the feedback data, the spatial location of each misidentified crack skeleton segment in the image is marked and mapped to the geographic coordinates of the actual slope area. If the misidentified skeleton segments are highly concentrated at the imaging edge or field-of-view limit of a certain camera, it indicates that the camera's imaging effect on the crack structure is poor at the current installation angle, and attitude adjustment is required.

[0077] By adjusting the camera's pitch angle (e.g., lowering it by 3 degrees), lens rotation angle (e.g., adjusting it counterclockwise by 5 degrees), or moving the camera to a support closer to the slope surface, the imaging axis can be altered to be as perpendicular as possible to the direction of the main crack development on the slope, thus enhancing the clarity of crack details. This type of physical angle optimization is particularly suitable when the slope surface is irregular and the crack development direction is orthogonal to the camera.

[0078] Fine-tuning of the camera's posture should be performed in conjunction with the cross-analysis results of the sunlight angle, shooting time, and misjudgment probability mentioned in the preceding steps to ensure that the optimization operation has physical support and necessity.

[0079] The purpose of this step is to establish a feedback mechanism for the identification results, enabling dynamic adaptive adjustment and optimization throughout the crack monitoring process. This forms a closed-loop, sustainably converging slope crack identification and early warning system, improving the system's adaptability to complex environments and the accuracy, stability, and reliability of crack identification results. In the process of crack identification on open-pit coal mine slopes, images are greatly affected by factors such as ambient light, solar altitude angle, cloud cover changes, and the reflectivity of the surface material. This leads to significant differences in the image representation of the same crack at different times and angles, easily resulting in identification errors, misjudgments of crack length, or low confidence levels. If crack identification parameters (such as albedo settings, edge detection thresholds, exposure time, and attitude angle) remain fixed for a long period, the system will struggle to cope with identification fluctuations caused by real-time lighting disturbances or changes in slope morphology, ultimately leading to increased misjudgment rates, distorted crack development trends, and in severe cases, even triggering false warnings or missed detections.

[0080] This step involves frame-by-frame statistical analysis of the crack length residuals and confidence level changes from the previous identification stage. It extracts the spatiotemporal distribution patterns of declining identification performance and feeds the error information back into the solar geometry and albedo baseline models. This allows for the correction of the albedo modeling parameters to better reflect the actual reflective characteristics of the slope. Simultaneously, the edge response threshold in the image processing workflow is dynamically adjusted to adapt to the current image quality and detail representation capabilities. Furthermore, intervention is made in the image acquisition stage by changing the exposure sequence, extending or shortening the exposure time, and optimizing shutter control to improve image quality under specific lighting conditions. Moreover, based on the spatial distribution analysis of the identification errors, the camera's installation angle is fine-tuned to better match the shooting angle with the crack direction, enhancing image clarity and structural coherence.

[0081] In summary, this step transforms the crack identification process from "one-way processing" to "closed-loop control," realizing a dynamic self-circulating mechanism of image acquisition, identification analysis, error feedback, parameter optimization, and imaging control. It not only improves the accuracy and robustness of crack identification under complex natural conditions but also lays the technical foundation for an intelligent, self-learning slope monitoring system, providing crucial support for all-weather, high-precision open-pit coal mine slope stability monitoring and disaster early warning.

[0082] The above-mentioned multi-source sensing-based open-pit coal mine slope stability monitoring and early warning method significantly improves the accuracy and robustness of crack identification under complex lighting conditions, fundamentally solving the technical problems of misjudgment of crack outlines, sudden increases in crack length, and false triggering of early warning signals caused by dynamic changes in shadows. This invention establishes a sustainable and convergent intelligent monitoring mechanism by introducing illuminance modeling driven by solar altitude and azimuth angles, constructing a shadow probability map to suppress interference features, fusing laser point cloud and tilt angle information for multi-source verification, and dynamically optimizing edge detection and imaging parameters through a feedback closed loop. This mechanism can not only accurately identify the true outline and length change trend of cracks, significantly improving the reliability of crack development trend curves, but also actively adapt to fluctuations in ambient light, achieving high-precision slope deformation identification and early warning output around the clock, greatly enhancing the disaster prevention and control capabilities and on-site response efficiency of open-pit coal mines in actual production.

[0083] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for monitoring and early warning of slope stability in open-pit coal mines based on multi-source sensing, characterized in that, Includes the following steps: S100, establish a baseline model of solar radiation geometry and albedo, collect reference images of crack-free areas based on the predicted results of solar altitude angle and azimuth angle, generate a time-varying illuminance gradient field within the monitoring period, and extract the spatial location and temporal change trajectory of the shadow front. S200 constructs a shadow crack dual-domain decomposition model based on the illuminance gradient field, extracts the shadow feature subspace using the brightness ratio, polarization ratio and chromaticity invariance features, and generates a shadow probability map to suppress crack identification interference. S300, guided by the shadow probability map, extracts the crack skeleton, performs local gradient weakening, and applies width smoothing, curvature constraint and endpoint inertial tracking constraint to generate an initial crack skeleton with temporal continuity. S400 introduces multi-source constraints based on crack skeleton, laser point cloud and tilt sensor, constructs deformation criteria based on crack opening amount and angle change, and uses adaptive gain reweighting method to suppress virtual length segments and improve the confidence of crack length measurement. S500 performs inverse temporal correction of the skeleton based on the shadow velocity field in image frames with a sudden increase in crack length, removes skeleton segments that move synchronously with the shadow, and reconstructs the crack length time series to ensure the authenticity of the change trend. The S600 feeds back the crack length residual and confidence changes to the baseline model, adaptively adjusts the albedo parameters and edge detection threshold, and updates the exposure sequence and camera pose to build a closed-loop dynamic monitoring and early warning mechanism.

2. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S100 includes: Based on the longitude, latitude, and altitude of the slope location, the solar altitude angle and solar azimuth angle are calculated every minute during the monitoring period to form a solar motion path sequence with time resolution. Based on the slope orientation, determine the incident direction and angular distribution of sunlight acting on the slope surface at various times; Images of a benchmark area of ​​a slope without cracks were collected under multiple time periods, angles, and exposure conditions. After brightness normalization, grayscale response was extracted to establish an albedo variation model. Based on the brightness difference between the albedo model and the measured image, the spatial location and temporal trajectory of the shadow front are extracted to generate a time-varying illuminance gradient field covering the entire monitoring period.

3. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S200 includes: The image to be identified is spatially registered with the illumination gradient image at the corresponding time to establish a consistent spatial reference. Extract the brightness ratio feature, polarization response feature and chromaticity invariance feature of each pixel to construct a three-dimensional feature vector; The three-dimensional feature vectors are standardized and fused to calculate the shadow probability value and generate a shadow probability map of the same size as the image. In the subsequent crack recognition process, gradient suppression and skeleton path avoidance are performed on pixel regions with shadow probability values ​​higher than a set threshold to reduce the interference of shadows on crack recognition.

4. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S300 includes: Based on the probability values ​​of each pixel in the shadow probability map, a weighted weakening process is performed on the local gray-level gradient in the image. The gradient response intensity is reduced in the region with a high shadow probability value, generating an image gradient response map that suppresses the influence of illumination interference, which is used for subsequent crack path extraction and structure judgment. Based on crack width smoothing constraints, abnormal width path segments are identified and eliminated, while candidate paths with structural stability are retained. Curvature constraints and endpoint inertial tracking are applied to candidate paths to filter out paths with abrupt turning and enhance the consistency of paths in time series. The retained path is compressed into a single-pixel-width skeleton line, and auxiliary branches with low signal strength and abrupt changes in direction are removed to generate an initial crack skeleton with continuous structure and stable time sequence.

5. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S400 includes: Spatial registration was performed between the initial crack skeleton path and the lidar point cloud data. The crack opening change information at the skeleton line segment was extracted to identify the skeleton region with a continuous expansion trend. The angle change value of the tilt sensor is projected into the image space, and the tilt change response at the skeleton segment is extracted to determine whether there is real displacement in the corresponding rock mass area. The crack skeleton was divided into multiple line segment units, and confidence scores were obtained based on point cloud deformation index, tilt response index and image shadow interference index, respectively. Based on the confidence score results, the skeleton segments with scores below the threshold and located in high shadow probability areas are length-reduced to improve the overall confidence of crack length measurement.

6. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S500 includes: Identify image frames with sudden increases in crack length, extract the spatial location and confidence score of the newly added crack skeleton segment, and calculate the crack length difference by combining the image frames before and after the crack. Based on the established shadow velocity field, the shadow boundary evolution trajectory is traced back for no less than five frames before and after the sudden increase in image frames, and the advancement path of the shadow front is extracted. Spatial overlap analysis was performed on the crack skeleton segment and the shadow advancement path to identify completely overlapping skeleton segments. Based on their confidence scores, these segments were determined to be pseudo-crack segments and were removed. The remaining skeleton segments are reconnected and their lengths are recalculated to correct the crack length time series, and structured removal records are output.

7. The method for monitoring and early warning of open-pit coal mine slope stability based on multi-source sensing according to claim 1, characterized in that, Step S600 includes: The corrected crack length residuals and confidence changes in the crack identification results are collected and categorized according to image frame index, timestamp, solar altitude angle, shooting direction, and slope area location. The accuracy of the albedo parameter setting under the solar angle is judged based on the feedback results. If there is a misjudgment, the brightness distribution curve in the reference image is re-analyzed and the albedo parameter is corrected. Based on the feedback results, the edge detection threshold and the exposure time and aperture value during image acquisition are adjusted; Based on the spatial density distribution of misjudged skeleton segments in the image, the camera's pitch angle, rotation angle, and installation position are fine-tuned to improve image quality and crack recognition accuracy.

Images