A multi-source data fusion slope safety monitoring method

Abstract

The present application relates to the technical field of data processing, and more particularly to a multi-source data fusion slope safety monitoring method, comprising obtaining a three-dimensional model of a slope surface generated by three-dimensional laser scanning and a two-dimensional radar deformation cloud map collected by slope radar; extracting a cloud map spatial feature vector through an encoder of a self-supervised deep learning model, and outputting three-dimensional heat map coordinate data by a decoder; screening slope shoulder, scarp and crack area terrain feature points from triangular mesh vertices based on a terrain curvature threshold; matching three-dimensional heat map feature points with terrain feature points, and mapping to a three-dimensional model coordinate system through affine transformation to generate a time-stamped dynamic three-dimensional deformation heat map; comparing each time-stamped displacement value to calculate displacement change; and marking displacement overrun areas on the three-dimensional model with a color gradient and outputting a three-dimensional visualization result. The present application realizes two-dimensional deformation data mapping to three-dimensional space, eliminates existing two-dimensional cloud map positioning ambiguity, and improves slope deformation position recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for slope safety monitoring by multi-source data fusion. Background Technology

[0002] In existing slope safety monitoring systems, slope stability directly affects structural safety and service life in water conservancy projects, railways, and bridges. Therefore, monitoring systems use displacement sensors, strain gauges, and automated data acquisition equipment to track slope deformation trends in real time. By analyzing parameters such as displacement rate and crack development, potential instability risks can be identified, allowing engineers to develop maintenance strategies to prevent landslides and improve infrastructure reliability.

[0003] Existing slope safety monitoring suffers from the following technical challenges: Because monitoring results are presented as two-dimensional cloud maps, lacking three-dimensional spatial coordinate information, direct correlation and mapping with the actual slope topography cannot be achieved during image data processing. Logically, two-dimensional images rely solely on pixel positions and deformation values, excluding elevation or topographic relief features. This makes it difficult for data processing algorithms to accurately map deformation points in the cloud map to the actual three-dimensional terrain, preventing operators from intuitively identifying the location of deformation. For example, in slope radar monitoring applications, when a two-dimensional cloud map displays deformation hotspots, it's impossible to distinguish whether the hotspot is located at the top of a slope crack or at the toe of a slope. For instance, in areas of fractured rock, vegetation disturbance might be misjudged as actual displacement, increasing the risk of positioning errors and delaying early warning decisions and engineering intervention. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a slope safety monitoring method based on multi-source data fusion. This invention solves the technical problem that the actual deformation location of the slope is difficult to intuitively identify and accurately locate because the monitoring results are presented in the form of a two-dimensional cloud map.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows:

[0006] This invention provides a slope safety monitoring method based on multi-source data fusion, comprising:

[0007] Step 1: Obtain a 3D model of the slope surface and a 2D radar deformation cloud map. The 3D model of the slope surface includes point cloud data and triangular mesh topology generated by 3D laser scanning. The 2D radar deformation cloud map includes pixel position coordinates, corresponding displacement and acquisition time information collected by the slope radar.

[0008] Step 2: Input the pixel position coordinates and corresponding displacements in the two-dimensional radar deformation cloud map into the pre-trained self-supervised deep learning model. The self-supervised deep learning model is trained using historical slope deformation data. The encoder of the self-supervised deep learning model extracts spatial feature vectors, and the decoder of the self-supervised deep learning model maps the spatial feature vectors into three-dimensional heat map coordinate data including horizontal position, vertical displacement and elevation.

[0009] Step 3: Based on the preset terrain curvature threshold according to the geological characteristics of the slope, key feature points are selected from the vertices of the triangular mesh, the curvature value of each vertex is calculated, and vertices with curvature greater than the threshold are selected as terrain feature points. The terrain feature points include the vertex coordinates of the slope shoulder, steep slope and crack area.

[0010] Step 4: Match the feature points in the 3D heat map coordinate data generated in Step 2 with the terrain feature points selected in Step 3, and map the 3D heat map coordinates to the coordinate system of the 3D model of the slope surface through an affine transformation algorithm to generate a dynamic 3D deformable heat map with radar acquisition timestamps.

[0011] Step 5: Compare the displacement values ​​of the same coordinate point at different timestamps in the dynamic three-dimensional deformation heat map, calculate the displacement change, and store the displacement change data in the preset displacement database. Extract the displacement change data from the displacement database and compare it with the preset threshold according to the engineering safety standard to identify the coordinate area where the displacement change exceeds the limit.

[0012] Step 6: Mark the coordinate areas of displacement changes exceeding the limit on the 3D model of the slope surface using color gradient, and output the 3D visualization monitoring results.

[0013] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention includes an encoder and a decoder in its self-supervised deep learning model;

[0014] The encoder receives the pixel position coordinates and corresponding displacements from the two-dimensional radar deformable cloud map, extracts and outputs spatial feature vectors;

[0015] The decoder receives the spatial feature vector output by the encoder and outputs three-dimensional heat map coordinate data.

[0016] Furthermore, in the slope safety monitoring method based on multi-source data fusion described in this invention, step 4 includes:

[0017] Obtain the set of terrain feature points from the output of step 3, and extract key feature points from the 3D heat map coordinate data output in step 2 using a feature extraction algorithm;

[0018] The extracted key feature points are paired with the acquired terrain feature points using a feature matching algorithm, and the pairing results are mapped to the coordinate system of the three-dimensional model of the slope surface using an affine transformation algorithm.

[0019] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0020] The newly added two-dimensional radar deformation cloud map is collected in real time by slope radar, and the newly added two-dimensional radar deformation cloud map is input into the encoder of the self-supervised deep learning model.

[0021] The weight parameters of the self-supervised deep learning model are updated using an online learning algorithm based on the newly added two-dimensional radar deformable cloud map, and the three-dimensional heat map coordinate data is regenerated based on the updated weight parameters of the model.

[0022] The regenerated 3D heatmap coordinate data is aligned and fused with the historical 3D heatmap coordinate data stored in the displacement database using a coordinate transformation algorithm.

[0023] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0024] Before outputting the three-dimensional visualization monitoring results in step 6, noise reduction processing is performed. Color features are extracted from the texture data of the three-dimensional model of the slope surface using a color extraction algorithm. The extracted color features are compared with the preset vegetation spectral feature library using a similarity algorithm. When the similarity value exceeds the preset matching threshold, the displacement change data of the corresponding coordinate point is filtered out.

[0025] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0026] A color feature library was constructed based on multi-source remote sensing imagery and ground spectral measurement data, including the spectral characteristics of slope rock mass, vegetation, and artificial structures.

[0027] Color feature values ​​are obtained from the three-dimensional model texture of the slope surface corresponding to each coordinate point included in the displacement change data;

[0028] The similarity values ​​of the obtained color feature values ​​are calculated using a similarity calculation algorithm and compared with the similarity values ​​of each type of feature in the color feature library;

[0029] If the similarity value exceeds the preset similarity threshold and the maximum similarity corresponds to the vegetation category, then the coordinate point is determined to be a noise point and the data of the coordinate point is removed from the displacement change data.

[0030] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0031] On the 3D model of the slope surface from the 3D visualization monitoring results output in step 6:

[0032] Displacement change data is read from the displacement database, and contour lines of displacement change are superimposed using a contour line generation algorithm.

[0033] Based on the numerical range of displacement change, a color mapping algorithm is used to label the displacement change level with color gradients;

[0034] Identify the coordinate region of the maximum value in the displacement change data and highlight this region using a highlighting algorithm.

[0035] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0036] Point cloud data of the slope surface is collected by a drone equipped with a laser scanner. The point cloud data is used to generate the three-dimensional model of the slope surface in step 1.

[0037] The collected point cloud data is denoised using a filtering algorithm, and a surface reconstruction algorithm is used to generate a triangular mesh model.

[0038] The generated triangular mesh model is used as the topological basis for the three-dimensional model of the slope surface in step 1.

[0039] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0040] Based on dynamic 3D deformation heatmaps with timestamps:

[0041] Select the displacement change at the same coordinate point in two consecutive timestamps;

[0042] Calculate the rate of change of the displacement difference over time;

[0043] When the rate of change exceeds the preset safety value, an early warning signal is triggered and sent to the terminal device.

[0044] Furthermore, the slope safety monitoring method based on multi-source data fusion described in this invention also includes:

[0045] In response to the warning signal, the online learning rate of the self-supervised deep learning model is dynamically adjusted according to the displacement change rate exceeding the limit value carried in the warning signal. The sampling frequency of the newly added two-dimensional radar deformed cloud map is increased according to the displacement change rate exceeding the limit value. The dynamic update mechanism of the vegetation spectral feature library for noise reduction is activated. The matching threshold is corrected based on the spectral features of the current season.

[0046] In the 3D visualization monitoring interface, a thermal layer of the early warning area is overlaid, and the coordinate points of the displacement change rate exceeding the limit are highlighted with pulsating red light. The structural parameters of the triangular mesh model corresponding to the coordinate point are automatically associated, and the simulated data of rock mass stress distribution is output to the terminal in real time.

[0047] The drone laser scanner is launched to perform an automatic cruise mission. It generates a circling flight path based on the warning coordinates and re-acquires point cloud data of the warning area at millimeter-level resolution. This triggers a surface reconstruction algorithm to update the 3D model of the slope surface.

[0048] The re-collected data is input into the online learning mechanism to iteratively optimize the encoder and decoder weight parameters of the self-supervised deep learning model, and regenerate the 3D heat map coordinate data to replace the previously generated 3D heat map coordinate data.

[0049] Based on the updated dynamic three-dimensional deformation heat map, the displacement change and displacement change rate are recalculated. If the change rate still exceeds the limit after three consecutive iterations, the warning level is upgraded.

[0050] Beneficial effects of this invention;

[0051] The beneficial effects of this invention are reflected in solving the positioning ambiguity problem caused by the lack of three-dimensional coordinate information in existing two-dimensional radar deformation cloud maps in slope monitoring through multi-source data fusion and three-dimensional spatial mapping technology. The two-dimensional radar deformation cloud map is input into a self-supervised deep learning model. The encoder extracts spatial feature vectors, and the decoder outputs three-dimensional heat map coordinate data including horizontal position, vertical displacement, and elevation, achieving a dimensionality upgrade from two-dimensional image space to three-dimensional geographic coordinate space. Based on the terrain curvature threshold, terrain feature points in slope shoulders, steep slopes, and crack areas are selected from the vertices of the triangular mesh. An affine transformation algorithm maps the three-dimensional heat map coordinates to the edge... A 3D coordinate system for the slope surface model generates a dynamic 3D deformation heatmap with timestamps, ensuring that the deformation data accurately matches the actual terrain. The displacement change is calculated by comparing displacement values ​​at different timestamps, and areas exceeding the displacement change limit are marked on the 3D model with color gradients. The results are output as 3D visualization monitoring, allowing operators to intuitively identify key deformation locations such as cracks at the top of the slope or sliding zones at the toe. A vegetation spectral feature database comparison mechanism filters out vegetation disturbance noise. After an early warning is triggered, millimeter-level re-acquisition of point cloud data updates the model and iteratively verifies the displacement change rate, forming a closed-loop decision chain from monitoring to location, improving the accuracy and reliability of slope deformation location identification. Attached Figure Description

[0052] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0053] Figure 1A flowchart of a slope safety monitoring method based on multi-source data fusion provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0055] Please see Figure 1 The present invention provides a slope safety monitoring method based on multi-source data fusion, comprising:

[0056] Step 1: Obtain a 3D model of the slope surface and a 2D radar deformation cloud map. The 3D model of the slope surface includes point cloud data and triangular mesh topology generated by 3D laser scanning. The 2D radar deformation cloud map includes pixel position coordinates, corresponding displacement and acquisition time information collected by the slope radar.

[0057] Step 2: Input the pixel position coordinates and corresponding displacements in the two-dimensional radar deformation cloud map into the pre-trained self-supervised deep learning model. The self-supervised deep learning model is trained using historical slope deformation data. The encoder of the self-supervised deep learning model extracts spatial feature vectors, and the decoder of the self-supervised deep learning model maps the spatial feature vectors into three-dimensional heat map coordinate data including horizontal position, vertical displacement and elevation.

[0058] Step 3: Based on the preset terrain curvature threshold according to the geological characteristics of the slope, key feature points are selected from the vertices of the triangular mesh, the curvature value of each vertex is calculated, and vertices with curvature greater than the threshold are selected as terrain feature points. The terrain feature points include the vertex coordinates of the slope shoulder, steep slope and crack area.

[0059] Step 4: Match the feature points in the 3D heat map coordinate data generated in Step 2 with the terrain feature points selected in Step 3, and map the 3D heat map coordinates to the coordinate system of the 3D model of the slope surface through an affine transformation algorithm to generate a dynamic 3D deformable heat map with radar acquisition timestamps.

[0060] Step 5: Compare the displacement values ​​of the same coordinate point at different timestamps in the dynamic three-dimensional deformation heat map, calculate the displacement change, and store the displacement change data in the preset displacement database. Extract the displacement change data from the displacement database and compare it with the preset threshold according to the engineering safety standard to identify the coordinate area where the displacement change exceeds the limit.

[0061] Step 6: Mark the coordinate areas of displacement changes exceeding the limit on the 3D model of the slope surface using color gradient, and output the 3D visualization monitoring results.

[0062] This invention provides a multi-source data fusion method for slope safety monitoring, achieving precise monitoring of slope deformation through multi-source data acquisition and fusion processing. Step 1 acquires two basic data sources: a 3D model of the slope surface and a 2D radar deformation cloud image. The 3D model is acquired using 3D laser scanning technology, including point cloud data and a triangular mesh topology. The point cloud data provides discrete coordinate information of the slope surface, and the triangular mesh topology connects the point cloud data to form a continuous surface model using a surface reconstruction algorithm. The 2D radar deformation cloud image is acquired using slope radar, including pixel position coordinates, corresponding displacement, and acquisition time information. Pixel position coordinates represent the planar position of the image, displacement reflects the degree of slope surface deformation, and acquisition time information provides a time reference for subsequent time series analysis. This step integrates multi-source heterogeneous data, providing fundamental data support for subsequent data fusion processing.

[0063] Step 2 involves data dimensionality enhancement. The pixel coordinates and corresponding displacements from the 2D radar deformation cloud image are input into a pre-trained self-supervised deep learning model. This model is trained using historical slope deformation data. The encoder extracts spatial feature vectors from the 2D data, encoding the spatial distribution and deformation-related features of the image. The decoder maps these spatial feature vectors into 3D heatmap coordinate data, including horizontal position, vertical displacement, and elevation information, thus achieving the transformation from 2D image space to 3D geographic coordinate space. The self-supervised deep learning model is based on an encoder-decoder architecture. The encoder processes the input data using a convolutional neural network, and the decoder reconstructs the output data through deconvolution operations, completing the data dimensionality enhancement.

[0064] Step 3 involves selecting key areas based on topographic geometric features. A topographic curvature threshold is preset according to the slope's geological characteristics. The curvature value of each vertex in the triangular mesh is calculated. A curvature estimation algorithm is used to analyze the geometric changes in the vertex's neighborhood. Vertices with curvature greater than the threshold are selected as topographic feature points. These feature points cover the vertex coordinates of slope shoulders, steep slopes, and crack areas, where higher curvature represents areas of significant topographic change. The topographic curvature threshold is set based on the slope's rock mass mechanical properties and geological survey data, ensuring the engineering rationality of feature point selection.

[0065] Step 4 achieves spatial registration of multi-source data. The feature points in the 3D heatmap coordinate data generated in Step 2 are matched with the terrain feature points selected in Step 3. A feature descriptor matching algorithm is used to associate the points. The matching algorithm includes scale-invariant feature transformation or fast retinal keypoint descriptor. The successfully matched feature point pairs are mapped to coordinates through an affine transformation algorithm. The affine transformation maps the 3D heatmap coordinates to the 3D model coordinate system of the slope surface based on rotation, translation and scaling matrices, generating a dynamic 3D deformable heatmap with radar acquisition timestamps. The timestamps are derived from the radar acquisition time information to achieve spatiotemporal data consistency.

[0066] Step 5 involves dynamic deformation analysis, comparing the displacement values ​​of the same coordinate point at different timestamps in the dynamic 3D deformation heatmap, extracting continuous time-series displacement data, and calculating the difference between adjacent timestamp displacement values ​​to obtain the displacement change. The displacement change data is stored in a displacement database pre-set according to the monitoring system architecture. The displacement database uses a time-series database structure to store time-stamped displacement data. The displacement change data is extracted from the displacement database and compared with a threshold pre-set according to engineering safety standards. The engineering safety standards are determined based on industry specifications and slope safety levels to identify coordinate areas where the displacement change exceeds the limit. These areas represent potential deformation risk areas.

[0067] Step 6 outputs the 3D visualization monitoring results. On the 3D model of the slope surface, the coordinates of areas where displacement changes exceed limits are marked using color gradients. The color gradient, based on color temperature changes, represents the level of displacement change, with a red-to-blue gradient indicating high-risk to low-risk levels. The output of the 3D visualization monitoring results is achieved through a rendering engine, and the output format includes interactive 3D models or static images, allowing operators to intuitively observe the deformation locations of key areas such as cracks at the top of the slope or slip zones at the toe. The visualization results are overlaid with terrain feature information and deformation data, providing a comprehensive basis for slope condition assessment.

[0068] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention includes a self-supervised deep learning model comprising an encoder and a decoder.

[0069] The encoder receives the pixel position coordinates and corresponding displacements from the two-dimensional radar deformable cloud map, extracts and outputs spatial feature vectors;

[0070] The decoder receives the spatial feature vector output by the encoder and outputs three-dimensional heat map coordinate data.

[0071] The self-supervised deep learning model employs an encoder-decoder architecture to process 2D radar deformable cloud image data. The encoder receives the pixel position coordinates and corresponding displacements from the 2D radar deformable cloud image as input data, and extracts spatial feature vectors from the input data through a multi-layer convolutional neural network structure. These spatial feature vectors encode the spatial distribution characteristics of the pixel position coordinates in the 2D radar deformable cloud image and their correlation with the displacements, forming a high-dimensional feature representation.

[0072] The decoder receives the spatial feature vector output from the encoder as input and reconstructs it using a deconvolutional neural network structure. The decoder maps the spatial feature vector into three-dimensional heatmap coordinate data, including horizontal position coordinates, vertical displacement, and elevation information, thus achieving the conversion from two-dimensional image space to three-dimensional geographic coordinate space. The three-dimensional heatmap coordinate data represents slope surface deformation information in three-dimensional coordinate form, with each coordinate point including data in three dimensions: horizontal position, vertical displacement, and elevation.

[0073] The convolutional neural network in the encoder employs a combination of multiple convolutional and pooling layers. Convolutional layers extract local features through convolutional kernels, while pooling layers compress feature dimensions through downsampling operations. The convolutional layers use the ReLU activation function to enhance non-linear feature extraction, while the pooling layers use max pooling to preserve salient features. The encoder's final output spatial feature vector includes a representation of the essential features of the input data.

[0074] The decoder employs a deconvolutional neural network with multiple layers of deconvolutional and upsampling layers. The deconvolutional layers restore the spatial resolution of the feature map through transpose convolution, while the upsampling layers increase the feature map size through interpolation. The deconvolutional layers use linear activation functions to maintain numerical stability, and the upsampling layers use bilinear interpolation to ensure spatial continuity. The decoder ultimately outputs 3D heatmap coordinate data, completing the data upscaling transformation.

[0075] The encoder and decoder are connected end-to-end to form a complete self-supervised deep learning model. The feature vectors extracted by the encoder serve as the input to the decoder, maintaining the coherence of data processing. The self-supervised deep learning model is trained using historical slope deformation data, and the model parameters are optimized by minimizing the reconstruction loss function, enabling the model to accurately learn the mapping relationship from two-dimensional data to three-dimensional data.

[0076] In this process, the pixel position coordinates and corresponding displacements in the 2D radar deformation cloud map are used as model input. The encoder extracts spatial feature vectors, and the decoder generates 3D heat map coordinate data output. The spatial feature vectors serve as intermediate feature representations, carrying the essential feature information of the input data. The 3D heat map coordinate data serves as the final output, providing 3D coordinate data of slope deformation, including elevation information.

[0077] Specifically, in the slope safety monitoring method based on multi-source data fusion described in this invention, step 4 includes:

[0078] Obtain the set of terrain feature points from the output of step 3, and extract key feature points from the 3D heat map coordinate data output in step 2 using a feature extraction algorithm;

[0079] The extracted key feature points are paired with the acquired terrain feature points using a feature matching algorithm, and the pairing results are mapped to the coordinate system of the three-dimensional model of the slope surface using an affine transformation algorithm.

[0080] Step 4 performs spatial registration of multi-source data, extracting topographic feature points for slope shoulders, steep slopes, and crack areas from the topographic feature point set selected in Step 3. The topographic feature point set is selected from the vertices of a triangular mesh based on a topographic curvature threshold, which is preset according to the geological characteristics of the slope. Feature points represent areas of significant topographic change. Key feature points are extracted from the 3D heatmap coordinate data output in Step 2 using feature extraction algorithms, including scale-invariant feature transformation or fast retinal keypoint descriptors. Key feature points correspond to locations in the 3D heatmap coordinate data that exhibit significant displacement or prominent geometric features.

[0081] The extracted key feature points are paired with the acquired terrain feature points using a feature matching algorithm, including nearest neighbor matching or fast approximate nearest neighbor search. The matching process is based on the similarity calculation of feature descriptors. Feature matching establishes the correspondence between the feature points of the 3D heatmap coordinate data and the terrain feature points of the 3D model of the slope surface, forming a set of matched point pairs. The set of matched point pairs is then mapped using an affine transformation algorithm, which calculates the rotation matrix, translation vector, and scaling factor based on the least squares method.

[0082] The affine transformation algorithm maps the coordinates of the 3D heatmap to the coordinate system of the 3D model of the slope surface, achieving spatial alignment between coordinate systems from different data sources. The mapping process maintains spatial consistency between the geometric features of the 3D heatmap coordinate data and the topographic features of the 3D model of the slope surface. The coordinate mapping result generates a dynamic 3D deformable heatmap with radar acquisition timestamps derived from the acquisition time information of the 2D radar deformable cloud image, achieving spatiotemporal data correlation.

[0083] Dynamic 3D deformation heatmaps integrate deformation information from 3D heatmap coordinate data with topographic information from 3D slope surface models, forming a time-stamped spatiotemporally correlated data layer. The spatial registration process establishes a precise mapping between 2D radar deformation data and the 3D terrain model, resolving the positioning ambiguity caused by the lack of elevation information in 2D data. The registered deformation data accurately matches the actual landform features, providing an accurate spatial benchmark for subsequent deformation analysis.

[0084] This spatial registration process is achieved through three sub-steps: feature extraction, feature matching, and coordinate transformation. The output of each sub-step serves as the input for the next, forming a complete processing chain. Feature extraction identifies salient features from the original data, feature matching establishes feature correspondences, and coordinate transformation completes spatial mapping, ultimately generating a dynamic 3D deformation heatmap that integrates multi-source data.

[0085] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0086] The newly added two-dimensional radar deformation cloud map is collected in real time by slope radar, and the newly added two-dimensional radar deformation cloud map is input into the encoder of the self-supervised deep learning model.

[0087] The weight parameters of the self-supervised deep learning model are updated using an online learning algorithm based on the newly added two-dimensional radar deformable cloud map, and the three-dimensional heat map coordinate data is regenerated based on the updated weight parameters of the model.

[0088] The regenerated 3D heatmap coordinate data is aligned and fused with the historical 3D heatmap coordinate data stored in the displacement database using a coordinate transformation algorithm.

[0089] Newly added two-dimensional radar deformation cloud map data is continuously collected by slope radar equipment. The slope radar uses synthetic aperture radar technology to obtain slope surface deformation information. The newly added two-dimensional radar deformation cloud map includes the latest pixel position coordinates, corresponding displacement, and acquisition time information, reflecting the real-time deformation status of the slope. The newly added two-dimensional radar deformation cloud map is input into the encoder of a self-supervised deep learning model. The encoder processes the input data based on a convolutional neural network architecture.

[0090] Using newly added 2D radar deformable cloud image data, the weight parameters of a self-supervised deep learning model are updated through an online learning algorithm. The online learning algorithm employs stochastic gradient descent or the Adam optimizer to incrementally update the model parameters. During the online learning process, the gradient of the loss function is calculated based on the new data, and the model weight parameters are adjusted in the opposite direction of the gradient to adapt the model to the latest data distribution characteristics. Based on the self-supervised deep learning model with updated weight parameters, 3D heatmap coordinate data is regenerated, and the updated model retains its ability to represent the features of the new data.

[0091] The regenerated 3D heatmap coordinate data includes the latest displacement information and is aligned and fused with the historical 3D heatmap coordinate data stored in the displacement database using a coordinate transformation algorithm. The coordinate transformation algorithm includes rigid body transformation or affine transformation methods, and the coordinate transformation parameters are calculated using the least squares method. The alignment and fusion process matches displacement data with the same coordinate points based on timestamps, and uses a weighted average or Kalman filter algorithm to fuse historical and latest data, forming a continuous and complete temporal deformation dataset.

[0092] This process enables real-time updating and fusion of monitoring data, and utilizes an online learning mechanism to allow the self-supervised deep learning model to continuously adapt to changes in slope deformation characteristics. Coordinate alignment and fusion ensure the temporal continuity and spatial consistency of historical and real-time data, providing an accurate and reliable data foundation for slope safety monitoring. The data processing forms a closed-loop feedback loop, continuously optimizing model performance and improving monitoring accuracy.

[0093] The real-time data acquisition and processing module adopts a distributed computing architecture, supporting parallel acquisition and processing of multi-source data. Data fusion results are stored in a displacement database, providing data support for subsequent deformation analysis and early warning decisions. The system is scalable, capable of connecting to various sensor data sources to achieve more comprehensive slope safety status monitoring.

[0094] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0095] Before outputting the three-dimensional visualization monitoring results in step 6, noise reduction processing is performed. Color features are extracted from the texture data of the three-dimensional model of the slope surface using a color extraction algorithm. The extracted color features are compared with the preset vegetation spectral feature library using a similarity algorithm. When the similarity value exceeds the preset matching threshold, the displacement change data of the corresponding coordinate point is filtered out.

[0096] Before outputting the 3D visualization monitoring results in step 6, noise reduction processing is performed. Color features of each coordinate point are extracted from the texture data of the 3D model of the slope surface using a color extraction algorithm. The color extraction algorithm is based on the RGB color space or HSV color model and obtains the color feature value corresponding to each coordinate point from the texture of the 3D model surface. The color feature value includes information in three dimensions: hue, saturation, and lightness.

[0097] The extracted color features are compared with a pre-defined vegetation spectral feature library using similarity algorithms, including Euclidean distance calculation or cosine similarity calculation. The vegetation spectral feature library is constructed based on multi-source remote sensing imagery and ground spectral measurement data, and includes spectral reflectance feature data of typical vegetation types. Each vegetation category in the feature library corresponds to a set of standard spectral feature values.

[0098] When the similarity value exceeds a preset matching threshold, the displacement change data of the corresponding coordinate point is determined to be caused by vegetation disturbance. The matching threshold is determined based on historical data analysis, and the distinguishing boundary value between vegetation features and rock mass features is calculated using statistical methods. The similarity calculation process employs feature normalization to eliminate the influence of lighting conditions and acquisition angle on color features.

[0099] The process filters displacement change data at corresponding coordinate points, removing data points identified as being caused by vegetation disturbance from the displacement database. The filtering operation retains displacement change data reflecting the true deformation of the slope rock mass, eliminating interference from non-rock mass deformation. Noise reduction improves the accuracy of the displacement change data, providing a reliable data foundation for subsequent visualization monitoring.

[0100] This noise reduction process is completed before the output of the 3D visualization monitoring results, ensuring that the final displayed displacement change data all originate from the actual deformation of the rock mass. Color feature extraction and spectral comparison form a complete vegetation identification chain, similarity threshold determination provides an objective filtering standard, and data filtering operations maintain the data quality of the displacement database.

[0101] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0102] A color feature library was constructed based on multi-source remote sensing imagery and ground spectral measurement data, including the spectral characteristics of slope rock mass, vegetation, and artificial structures.

[0103] Color feature values ​​are obtained from the three-dimensional model texture of the slope surface corresponding to each coordinate point included in the displacement change data;

[0104] The similarity values ​​of the obtained color feature values ​​are calculated using a similarity calculation algorithm and compared with the similarity values ​​of each type of feature in the color feature library;

[0105] If the similarity value exceeds the preset similarity threshold and the maximum similarity corresponds to the vegetation category, then the coordinate point is determined to be a noise point and the data of the coordinate point is removed from the displacement change data.

[0106] A color feature library was constructed based on multi-source remote sensing imagery and ground spectral measurement data, encompassing the spectral characteristics of slope rock masses, vegetation, and man-made structures. The multi-source remote sensing imagery includes high-resolution satellite imagery and aerial photogrammetry data, while ground spectral measurements were performed using field spectrometers to collect reflectance spectral data for typical land cover types. The color feature library stores spectral feature vectors for different land cover types in a database format. Each feature vector includes reflectance values ​​for multiple bands, and the library structure supports rapid querying and similarity comparison.

[0107] Color feature values ​​were obtained from the 3D model texture of the slope surface corresponding to each coordinate point in the displacement change data. The 3D model texture of the slope surface was generated using RGB color information acquired by a 3D laser scanner, with each coordinate point corresponding to a pixel color value in the texture image. Color feature value extraction employed a region sampling method, selecting a certain range of texture regions centered on the coordinate points and calculating the color statistical characteristics of that region, including average hue, saturation, and brightness values.

[0108] The similarity calculation algorithm is used to calculate the similarity between the obtained color feature values ​​and the feature values ​​of various types in the color feature database. The similarity calculation algorithm chooses either Euclidean distance or cosine similarity method. Euclidean distance calculates the geometric distance between color feature vectors, while cosine similarity measures the cosine value of the angle between feature vectors. During the calculation process, the feature vectors are normalized to eliminate the influence of dimensions and ensure the comparability of similarity values.

[0109] If the similarity value exceeds a preset similarity threshold and the vegetation category corresponding to the highest similarity, the coordinate point is determined as noise and its data is removed from the displacement change data. The preset similarity threshold is determined based on historical data analysis, and the classification boundary value between vegetation features and rock mass features is calculated using statistical methods. The determination process adopts the maximum similarity principle, selecting the land cover category with the highest similarity as the matching result. When the matching result is a vegetation category and the similarity exceeds the threshold, the data removal operation is performed.

[0110] This noise identification and removal process forms a complete data quality control chain. The color feature library provides a comparison benchmark, color feature value extraction obtains field data, similarity calculation quantifies the degree of matching, and threshold determination makes filtering decisions. The data processing results retain displacement change data reflecting the true deformation of the rock mass, exclude noise data caused by vegetation disturbance, and improve the accuracy and reliability of slope safety monitoring.

[0111] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0112] On the 3D model of the slope surface from the 3D visualization monitoring results output in step 6:

[0113] Displacement change data is read from the displacement database, and contour lines of displacement change are superimposed using a contour line generation algorithm.

[0114] Based on the numerical range of displacement change, a color mapping algorithm is used to label the displacement change level with color gradients;

[0115] Identify the coordinate region of the maximum value in the displacement change data and highlight this region using a highlighting algorithm.

[0116] On the 3D model of the slope surface from the 3D visualization monitoring results output in step 6, displacement change data is read from the displacement database. The displacement database stores the displacement change data calculated in step 5, including coordinate point location information, displacement change values, and timestamp information. Contour lines of displacement change are superimposed using a contour line generation algorithm. This algorithm, based on the Marching Cubes method or Contour Tree algorithm, extracts points with the same displacement change value on the 3D model surface and connects them to form contour lines. The contour lines represent the distribution of areas with equal displacement changes.

[0117] Based on the range of displacement change values, a color mapping algorithm is used to label the displacement change levels with color gradients. The color mapping algorithm uses linear interpolation or piecewise linear mapping methods to map the displacement change values ​​to a predefined color gradient table. The color gradient table, from cool to warm colors, represents the displacement change levels from low to high. For example, blue represents low-risk areas and red represents high-risk areas. Color labeling enhances the visualization and intuitiveness.

[0118] The system identifies the region with the maximum value in the displacement change data and highlights it using a highlighting algorithm. This algorithm includes increasing the region's brightness, adding a blinking effect, or drawing a special border. The region with the maximum value corresponds to the point or area with the largest displacement change, and highlighting it allows operators to quickly locate high-risk positions. This visualization process is implemented based on a 3D rendering engine, updating the displacement change data in real time and dynamically adjusting the display effects.

[0119] The visualization process is logically continuous, forming a complete chain from data reading to graphic rendering. Contour lines provide the outline of displacement distribution, color gradients represent risk levels, and highlighting emphasizes extreme value areas, all of which together enhance the interpretability of monitoring results and decision support capabilities.

[0120] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0121] Point cloud data of the slope surface is collected by a drone equipped with a laser scanner. The point cloud data is used to generate the three-dimensional model of the slope surface in step 1.

[0122] The collected point cloud data is denoised using a filtering algorithm, and a surface reconstruction algorithm is used to generate a triangular mesh model.

[0123] The generated triangular mesh model is used as the topological basis for the three-dimensional model of the slope surface in step 1.

[0124] Point cloud data of the slope surface is collected by using a drone equipped with a laser scanner. The drone uses a multi-rotor or fixed-wing platform, and the laser scanner uses a 3D LiDAR device. The drone scans the slope surface along a preset flight path to collect high-density point cloud data. The point cloud data includes a large number of 3D coordinate points, each with XYZ coordinate information and reflection intensity value. The point cloud data is used to generate the 3D model of the slope surface in step 1, serving as the basic data source for 3D modeling.

[0125] The acquired point cloud data is denoised using filtering algorithms, including statistical filtering and radius filtering. Statistical filtering removes outliers based on the neighborhood distance distribution of the point cloud, while radius filtering filters noise points based on a point density threshold. This denoising process eliminates error points and environmental interference points introduced during the acquisition process, improving the quality of the point cloud data and providing clean data input for subsequent modeling.

[0126] A surface reconstruction algorithm is used to generate a triangular mesh model. The algorithm chosen is either Poisson reconstruction or Delaunay triangulation. Poisson reconstruction generates a smooth surface from the point cloud by solving the Poisson equation, while Delaunay triangulation generates a triangular mesh based on the topological connections of the point set. The surface reconstruction process converts discrete point cloud data into a continuous surface model. The triangular mesh model includes vertex, edge, and facet information, forming a geometric surface representation.

[0127] The generated triangular mesh model serves as the topological foundation for the 3D slope surface model in Step 1. The triangular mesh model defines the geometry and spatial relationships of the slope surface, and its topological structure includes vertex connectivity and facet adjacency information. The 3D slope surface model is constructed based on the triangular mesh model, with texture maps and material properties added to form a complete 3D visualization model for subsequent deformation analysis and monitoring result display.

[0128] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0129] Based on dynamic 3D deformation heatmaps with timestamps:

[0130] Select the displacement change at the same coordinate point in two consecutive timestamps;

[0131] Calculate the rate of change of the displacement difference over time;

[0132] When the rate of change exceeds the preset safety value, an early warning signal is triggered and sent to the terminal device.

[0133] Based on a timestamped dynamic 3D deformation heatmap, displacement change data corresponding to two consecutive timestamps are read from the displacement database. The timestamps are derived from the time information acquired by the slope radar, and the displacement change data is calculated in step 5 and stored in the displacement database. Displacement change values ​​at the same coordinate point at two consecutive timestamps are selected to achieve time series continuity.

[0134] The rate of change of displacement variation over time is calculated as the ratio of the displacement variation difference to the time interval. The time interval is calculated as the difference between two timestamps, typically in days or hours. The rate of change of time quantifies the rate of slope deformation, reflecting the swiftness of displacement change.

[0135] When the rate of change over time exceeds a preset safety value, an early warning signal is triggered and sent to the terminal device. The preset safety value is set according to engineering safety standards, taking into account slope type, rock mass properties, and historical deformation data. The early warning signal includes the coordinate point location, the rate of change over time value, and the extent of the exceedance, and is transmitted to the monitoring center or mobile terminal via a wireless communication network.

[0136] The early warning triggering mechanism is based on real-time data monitoring, and the system automatically compares the rate of change over time with preset safety values. Once an early warning signal is triggered, the early warning process begins, including data recording, notification sending, and interface prompts. Upon receiving the early warning signal, the terminal device displays the warning information, prompting operators to take further action.

[0137] This early warning process is integrated into the slope safety monitoring system, enabling automated monitoring and real-time response. The calculation of the rate of change over time provides dynamic risk assessment, and the early warning mechanism enhances the timeliness and effectiveness of slope safety management.

[0138] Specifically, the slope safety monitoring method based on multi-source data fusion described in this invention further includes:

[0139] In response to the warning signal, the online learning rate of the self-supervised deep learning model is dynamically adjusted according to the displacement change rate exceeding the limit value carried in the warning signal. The sampling frequency of the newly added two-dimensional radar deformed cloud map is increased according to the displacement change rate exceeding the limit value. The dynamic update mechanism of the vegetation spectral feature library for noise reduction is activated. The matching threshold is corrected based on the spectral features of the current season.

[0140] In the 3D visualization monitoring interface, a thermal layer of the early warning area is overlaid, and the coordinate points of the displacement change rate exceeding the limit are highlighted with pulsating red light. The structural parameters of the triangular mesh model corresponding to the coordinate point are automatically associated, and the simulated data of rock mass stress distribution is output to the terminal in real time.

[0141] The drone laser scanner is launched to perform an automatic cruise mission. It generates a circling flight path based on the warning coordinates and re-acquires point cloud data of the warning area at millimeter-level resolution. This triggers a surface reconstruction algorithm to update the 3D model of the slope surface.

[0142] The re-collected data is input into the online learning mechanism to iteratively optimize the encoder and decoder weight parameters of the self-supervised deep learning model, and regenerate the 3D heat map coordinate data to replace the previously generated 3D heat map coordinate data.

[0143] Based on the updated dynamic three-dimensional deformation heat map, the displacement change and displacement change rate are recalculated. If the change rate still exceeds the limit after three consecutive iterations, the warning level is upgraded.

[0144] In response to an early warning signal, the system analyzes the displacement change rate exceeding the limit carried in the warning signal and dynamically adjusts the online learning rate of the self-supervised deep learning model based on this value. The online learning rate adjustment employs a proportional control algorithm; the larger the displacement change rate exceeding the limit, the greater the increase in the online learning rate, thus accelerating the model's adaptation to new data. Simultaneously, the sampling frequency of newly added two-dimensional radar deformation cloud images is increased proportionally according to the displacement change rate exceeding the limit. The sampling frequency adjustment is based on a preset mapping table to match the data acquisition rate with the deformation rate. A dynamic update mechanism for the vegetation spectral feature library, which has undergone noise reduction processing, is activated. This mechanism automatically downloads the latest seasonal spectral data based on timed tasks or events. The matching threshold is corrected based on the spectral features of the current season. These spectral features are derived from meteorological satellite and ground observation station data, and the matching threshold is calculated using statistical analysis methods to determine the distinction boundary between the vegetation spectrum and the rock mass spectrum for the current season.

[0145] A thermal layer of the warning area is overlaid on the 3D visualization monitoring interface. This thermal layer is generated using a heatmap rendering algorithm, with colors grading from blue to red to indicate risk levels. Points exceeding the displacement rate of change are highlighted with pulsating red light; the pulsation effect is achieved by periodically changing the transparency, with an adjustable frequency. The system automatically associates the corresponding triangular mesh model structural parameters with these points, including vertex coordinates, facet normals, and material properties, obtained through database queries. Real-time simulation data of rock mass stress distribution is output to the terminal. The stress distribution simulation uses the finite element method, calculated based on the displacement rate of change and material mechanical parameters. The output data includes numerical tables and graphs.

[0146] The drone's laser scanner is automatically launched for navigation, generating a circling flight path based on the warning coordinates. The path planning algorithm is based on A* or Dijkstra's algorithm to achieve full coverage of the warning area. Point cloud data of the warning area is reacquired at millimeter-level resolution, achieved through high-precision lidar and close-range flight. A surface reconstruction algorithm is triggered to update the 3D model of the slope surface. Surface reconstruction uses Poisson reconstruction or Marching Cubes algorithms, integrating the new point cloud data into the existing model.

[0147] The re-acquired data is input into the online learning mechanism, which employs incremental learning or mini-batch update strategies. The encoder and decoder weight parameters of the self-supervised deep learning model are iteratively optimized using stochastic gradient descent or the Adam optimizer. The previously generated 3D heatmap coordinate data is regenerated and stored in a displacement database with a version number.

[0148] Based on the updated dynamic 3D deformation heatmap, the displacement change and displacement rate are recalculated, following the same calculation process as in step 5. If the rate of change still exceeds the limit after three consecutive iterations, the warning level is upgraded. The warning level gradually increases from low to medium to high, with each level corresponding to a different response protocol and notification method.

[0149] This invention addresses the technical problem of difficulty in intuitively identifying and accurately locating the actual deformation position of slopes due to the presentation of monitoring results in two-dimensional cloud maps. It primarily achieves this through multi-source data fusion and three-dimensional spatial mapping techniques. First, it acquires a three-dimensional model of the slope surface generated by three-dimensional laser scanning and a two-dimensional radar deformation cloud map collected by slope radar. The three-dimensional slope surface model includes point cloud data and a triangular mesh topology, providing a complete three-dimensional terrain spatial reference. The two-dimensional radar deformation cloud map includes pixel position coordinates and displacement, but lacks elevation information. The encoder of a self-supervised deep learning model extracts spatial feature vectors from the two-dimensional cloud map, and the decoder maps these spatial feature vectors into three-dimensional heatmap coordinate data including horizontal position, vertical displacement, and elevation. This achieves a dimensionality upgrade from two-dimensional image space to three-dimensional geographic coordinate space, overcoming the limitations of two-dimensional data.

[0150] Based on a terrain curvature threshold, topographic feature points in slope shoulders, steep slopes, and crack areas are selected from the vertices of a triangular mesh. These feature points represent areas of significant terrain change. Feature points in the 3D heatmap coordinate data are matched with these topographic feature points. An affine transformation algorithm is then used to map the matching results onto the coordinate system of the 3D model of the slope surface, generating a timestamped dynamic 3D deformation heatmap. The spatial registration process establishes a precise mapping relationship between 2D radar data and 3D terrain, ensuring that the deformation data accurately matches the actual landform and eliminating the ambiguity of 2D cloud maps failing to correlate with the actual terrain.

[0151] Dynamic 3D deformation heatmaps are integrated with timestamps to compare displacement values ​​at the same coordinate point at different timestamps, calculating displacement changes and quantifying slope deformation trends. Areas exceeding displacement limits are marked with color gradients on the 3D model of the slope surface, outputting 3D visualized monitoring results. Operators can directly observe the deformation locations of key areas such as cracks at the top of the slope or slip zones at the toe, improving identification accuracy. A vegetation spectral feature database comparison mechanism filters vegetation disturbance noise, further reducing the risk of misjudgment. After an early warning is triggered, millimeter-level re-acquisition of point cloud data updates the model and iteratively verifies it, forming a closed-loop decision chain from monitoring to location, completely resolving the ambiguity problem of 2D cloud map positioning.

[0152] The specific implementation of this invention is as follows: In slope safety monitoring applications, a three-dimensional model of the slope surface is acquired using three-dimensional laser scanning technology. The three-dimensional laser scanning equipment employs pulsed or phase-detection lidar, with the scanning angle covering the entire slope surface. The point cloud density is set according to the monitoring accuracy requirements. The three-dimensional model of the slope surface includes high-precision point cloud data and a triangular mesh topology. The point cloud data is generated through data registration from multiple scanning stations, and the triangular mesh is constructed using the Delaunay triangulation algorithm to form a continuous surface model. Simultaneously, a two-dimensional radar deformable cloud image is acquired using a slope radar operating in the Ku-band or Ka-band. The transmission frequency is adjusted according to the monitoring distance, and the received signal is processed by synthetic aperture radar to generate a deformable cloud image including pixel position coordinates, displacement, and time information.

[0153] Two-dimensional radar deformable cloud images are input into a self-supervised deep learning model for processing. The self-supervised deep learning model employs either a Transformer or U-Net architecture. The encoder uses a multi-layer convolutional neural network to extract spatial features, with the kernel size adjusted according to the input image resolution. The decoder uses transposed convolutional layers for upsampling, and the output channel count is set to 3 to generate 3D coordinate data. Model training uses the Adam optimizer, with the learning rate dynamically decaying according to the training progress. The loss function combines reconstruction loss and geometric consistency loss.

[0154] Key points are extracted based on terrain curvature features. Moving least squares is used to calculate vertex curvature, with curvature thresholds set according to rock type: higher thresholds for granite slopes and lower thresholds for soil slopes. In the feature matching stage, the RANSAC algorithm is used to remove mismatches. The affine transformation matrix is ​​calculated through least squares fitting to achieve spatial registration from two-dimensional data to a three-dimensional model.

[0155] The displacement analysis phase employs time-series analysis to calculate the mean and standard deviation of displacement changes, and sets dynamic thresholds for anomaly detection. Visualization rendering uses the Phong color model, mapping displacement levels to the HSV color space, displaying high-risk areas in red tones and low-risk areas in blue tones.

[0156] Vegetation filtering employs multispectral data analysis, combining near-infrared and red-edge band features to establish a vegetation index threshold model. The early warning system sets multi-level thresholds based on the rate of displacement change, uses a sliding window algorithm to detect trend changes, and transmits early warning signals to the monitoring center via 4G / 5G networks.

[0157] The entire system is deployed on a cloud platform architecture, with data processing employing a distributed computing framework. Real-time data streams are processed via Kafka message queues, while historical data is stored in a time-series database. The user interface provides WebGL 3D rendering and mobile app access, supporting multi-user collaborative monitoring and historical data backtracking analysis.

[0158] Specific embodiments of the present invention are as follows:

[0159] In the slope monitoring scenario of a major engineering project in the mountainous region of Southwest China, a drone equipped with a laser scanner first scans the target slope to collect raw point cloud data of the slope surface. The drone uses a multi-rotor platform, and the laser scanner uses a 3D LiDAR device. Following a preset flight path, it comprehensively scans the slope surface, collecting high-density point cloud data. The point cloud data includes a large number of 3D coordinate points, each with XYZ coordinate information and a reflection intensity value. Statistical filtering algorithms are used to remove noise and outliers from the collected point cloud data, and a Poisson reconstruction algorithm is used to generate a triangular mesh model with a continuous surface. The generated triangular mesh model serves as the topological basis for the 3D model of the slope surface, defining the geometry and spatial relationships of the slope surface.

[0160] Simultaneously, slope radar equipment is deployed at a safe location on the opposite bank of the slope. This slope radar employs synthetic aperture radar technology to acquire two-dimensional radar deformation cloud images, including pixel location coordinates, corresponding displacement amounts, and acquisition time information. The two-dimensional radar deformation cloud images reflect the real-time deformation state of the slope surface; pixel location coordinates represent the planar position of the image; displacement amounts quantify the degree of deformation; and acquisition time information provides a time reference for time series analysis.

[0161] The pixel coordinates and corresponding displacements from a 2D radar deformation cloud image are input into a pre-trained self-supervised deep learning model. The self-supervised deep learning model is trained using historical slope deformation data. The encoder extracts spatial feature vectors of pixel coordinates and displacements through a convolutional neural network, and the decoder maps these spatial feature vectors into 3D heatmap coordinate data including horizontal position, vertical displacement, and elevation through deconvolution. The model training process requires no manual annotation; model parameters are optimized by minimizing the reconstruction loss function.

[0162] The triangular mesh vertices are processed based on a pre-defined topographic curvature threshold determined by the slope's geological characteristics. The curvature value of each vertex is calculated, and a curvature estimation algorithm is used to analyze the geometric changes in the vertex's neighborhood. Vertices with curvature greater than the threshold are selected as topographic feature points. These feature points encompass the vertex coordinates of slope shoulders, steep slopes, and crack areas, which exhibit higher curvature and represent areas of significant topographic change.

[0163] Feature points in the 3D heatmap coordinate data are matched with terrain feature points using feature descriptors. A scale-invariant feature transformation algorithm is used to associate point pairs, and the affine transformation parameter matrix is ​​calculated based on the matched point pairs. Rotation, translation, and scaling matrices are applied to map the 3D heatmap coordinates to the 3D model coordinate system of the slope surface, generating a dynamic 3D deformable heatmap with radar acquisition timestamps. The timestamps are derived from the radar acquisition time information, ensuring spatiotemporal data consistency.

[0164] The displacement values ​​of the same coordinate point at different time stamps are compared in a dynamic 3D deformation heatmap. Continuous time-series displacement values ​​are extracted, and the difference between adjacent time stamp displacement values ​​is calculated to obtain the displacement change, quantifying the slope deformation rate. The displacement change data is stored in a pre-set displacement database, which uses a time-series database structure to store time-stamped displacement data. Displacement change data is extracted from the displacement database and compared with thresholds preset according to engineering safety standards to identify coordinate regions where displacement changes exceed limits.

[0165] Noise reduction was performed on the 3D model of the slope surface. RGB color features of each coordinate point were extracted from the model texture data using a color extraction algorithm. Color feature values ​​included hue, saturation, and brightness information. The extracted color features were compared with a pre-defined vegetation spectral feature library constructed based on multi-source remote sensing imagery and ground spectral measurement data using a cosine similarity algorithm. When the similarity value exceeded a pre-defined matching threshold, the displacement change data of the corresponding coordinate point was filtered out to eliminate noise caused by vegetation disturbance.

[0166] On the 3D model of the slope surface, coordinate areas where displacement changes exceed limits are marked using color gradients. A color mapping algorithm is used to label the displacement change levels with color gradients, with blue indicating low-risk areas and red indicating high-risk areas. A highlighting algorithm is used to highlight the coordinate areas of the maximum displacement change data. The 3D visualization monitoring results are output, and an interactive 3D model is generated through a rendering engine, allowing operators to intuitively observe the deformation locations of key areas such as cracks at the slope crest or slip zones at the slope toe.

[0167] A warning signal is triggered when the rate of displacement change exceeds a preset safety value. In response to the warning signal, the online learning rate of the self-supervised deep learning model is dynamically adjusted based on the magnitude of the displacement change exceeding the limit carried in the warning signal, and the model weight parameters are updated using a stochastic gradient descent algorithm. The sampling frequency of newly added two-dimensional radar deformable cloud images is increased based on the magnitude of the displacement change exceeding the limit, and a dynamic update mechanism for the vegetation spectral feature library after noise reduction is activated. The matching threshold is then corrected based on the spectral characteristics of the current season.

[0168] The drone laser scanner is initiated with an automated cruise mission. Based on the warning coordinates, a circling flight path is generated to reacquire point cloud data of the warning area at millimeter-level resolution. A surface reconstruction algorithm is triggered to update the 3D model of the slope surface, and the reacquired data is input into an online learning mechanism. The encoder and decoder weight parameters of the self-supervised deep learning model are iteratively optimized to regenerate 3D heatmap coordinate data, replacing the previously generated 3D heatmap coordinate data.

[0169] Based on the updated dynamic 3D deformation heatmap, the displacement change and displacement rate are recalculated. If the rate of change still exceeds the limit after three consecutive iterations, the warning level is upgraded, gradually increasing from low to high, with each level corresponding to a different response protocol and notification method. This implementation process forms a closed-loop decision chain from data acquisition to warning response, improving the accuracy and reliability of slope safety monitoring.

[0170] The technical features of this invention are explained below:

[0171] The self-supervised deep learning model is trained and constructed based on historical slope deformation data, employing an encoder-decoder architecture to process two-dimensional radar deformation cloud image data. The encoder extracts spatial feature vectors of pixel position coordinates and displacement through a convolutional neural network, while the decoder maps these spatial feature vectors into three-dimensional heatmap coordinate data, including horizontal position, vertical displacement, and elevation, through deconvolution operations. In this invention, this model is used to upscale two-dimensional radar deformation cloud images to three-dimensional geographic coordinate space, achieving a mapping from image data to spatial data and providing a foundation for subsequent terrain registration and deformation analysis.

[0172] A three-dimensional model of the slope surface is constructed using 3D laser scanning technology, generating point cloud data and a triangular mesh topology. The point cloud data includes a large number of 3D coordinate points, each with XYZ coordinate information and a reflection intensity value. The triangular mesh is connected to form a continuous surface model through a surface reconstruction algorithm. In this invention, this model serves as a 3D spatial reference, providing terrain geometry and spatial relationships for fusion with radar deformation data and visualization.

[0173] The two-dimensional radar deformation cloud map is constructed using slope radar data, including pixel location coordinates, corresponding displacement, and acquisition time information. Pixel location coordinates represent the planar position of the image, displacement quantifies the degree of slope surface deformation, and acquisition time information provides a time reference for time series analysis. In this invention, this model serves as the input data source, and through a self-supervised deep learning model, deformation features are extracted and three-dimensional heat map coordinate data is generated.

[0174] Topographic feature points are constructed by selecting vertices from a triangular mesh based on a topographic curvature threshold. The curvature value of each vertex is calculated, and vertices with curvature greater than the threshold are selected to represent slope shoulders, steep slopes, and crack areas. The curvature threshold is preset according to the geological characteristics of the slope, so that the feature points represent areas with significant topographic changes. In this invention, topographic feature points are used to match with 3D heatmap coordinate data, and spatial registration is achieved through affine transformation, so that the deformation data accurately matches the actual landform.

[0175] The dynamic 3D deformation heatmap is constructed using an affine transformation algorithm. After matching the 3D heatmap coordinate data with terrain feature points, it is mapped onto the 3D model coordinate system of the slope surface, generating a deformation heatmap with radar acquisition timestamps. The heatmap integrates the timestamp dimension and displacement information to form a spatiotemporally correlated data layer. In this invention, this model is used to compare displacement values ​​at each timestamp to calculate displacement changes, quantify slope deformation trends, and support 3D visualization output.

[0176] The vegetation spectral feature library is constructed based on multi-source remote sensing imagery and ground spectral measurement data, including spectral reflectance characteristics of typical vegetation types, such as near-infrared and red-edge band features. The library stores spectral feature vectors in a database format, supporting rapid querying and similarity comparison. In this invention, the library is used for noise reduction, filtering displacement change data caused by vegetation disturbance through color feature comparison, thereby improving monitoring accuracy.

[0177] The displacement database is constructed according to the pre-defined architecture of the monitoring system and uses a time-series database structure to store displacement change data with timestamps. The database supports efficient read, write, and query operations, storing historical and real-time displacement data. In this invention, this database is used to store and extract displacement change data, compare it with safety thresholds to identify areas exceeding limits, and provide data support for visualization.

[0178] The early warning system is built upon a displacement change rate threshold. An early warning signal is triggered when the displacement change rate exceeds a preset safety value. This safety value is set according to engineering standards, taking into account slope type and historical data. In this invention, the system is used to monitor the deformation rate in real time. Upon triggering, it initiates a response mechanism, such as adjusting the sampling frequency, updating the model, and sending alarms, forming a closed-loop decision chain.

[0179] The feature extraction algorithm employs scale-invariant feature transformation or fast retinal keypoint descriptor methods to identify key feature points with significant displacement or prominent geometric features from 3D heatmap coordinate data. In this invention, the algorithm is used to extract feature points from 3D heatmap data for matching with terrain feature points, supporting the spatial registration process.

[0180] The feature matching algorithm is based on the principle of nearest neighbor matching or fast approximate nearest neighbor search. It calculates the descriptor similarity between key feature points and terrain feature points to realize point pair association. In this invention, the algorithm is used to establish the correspondence between 3D heat map feature points and terrain feature points to form a set of matching point pairs, providing input for affine transformation.

[0181] The affine transformation algorithm calculates the rotation matrix, translation vector, and scaling factor using the least squares method to map the matching point pairs from the 3D heat map coordinate system to the 3D model coordinate system of the slope surface. In this invention, the algorithm is used to achieve spatial alignment of coordinate systems from different data sources, generate a dynamic 3D deformation heat map with timestamps, and achieve accurate matching of deformation data with actual landforms.

[0182] The online learning algorithm employs stochastic gradient descent or Adam optimizer strategies to incrementally update the weight parameters of the self-supervised deep learning model based on newly added two-dimensional radar deformation cloud map data. In this invention, the algorithm is used to enable the model to continuously adapt to changes in slope deformation characteristics, improve the model's ability to process new data, and maintain monitoring accuracy.

[0183] The coordinate transformation algorithm is based on rigid body transformation or affine transformation methods. It aligns the newly generated 3D heat map coordinate data with historical data by calculating coordinate transformation parameters. In this invention, the algorithm is used to fuse real-time data and historical data to form a continuous time series dataset, which supports deformation trend analysis.

[0184] The color extraction algorithm obtains RGB or HSV color feature values, including hue, saturation, and brightness information, from the texture of the three-dimensional model of the slope surface. In this invention, the algorithm is used to extract the color features of each coordinate point, providing input data for vegetation disturbance noise reduction processing.

[0185] The similarity algorithm uses Euclidean distance or cosine similarity to calculate and compare the color feature to be tested with the standard feature value in the vegetation spectral feature library. In this invention, the algorithm is used to quantify the matching degree between the color feature and the vegetation spectrum. When the similarity exceeds the threshold, noise data is filtered out to improve the reliability of displacement change data.

[0186] The contour line generation algorithm is based on the Marching Cubes or Contour Tree method, which extracts points with the same displacement change value on the surface of the three-dimensional model and connects them to form contour lines. In this invention, the algorithm is used to visualize the displacement distribution contour, intuitively display the deformation area, and enhance the interpretability of the monitoring results.

[0187] The color mapping algorithm converts the numerical value of displacement change into a predefined color gradient through linear interpolation or piecewise linear mapping, with the risk level represented by a gradient from cool to warm colors. In this invention, the algorithm is used to label the displacement change level with color gradients, enabling operators to quickly identify high-risk areas.

[0188] The highlighting algorithm highlights the coordinate area of ​​the maximum displacement change by increasing the brightness of the area, adding a flashing effect, or drawing a special border; in this invention, the algorithm is used to emphasize extreme points, guide attention to potential deformation risk locations, and support the decision-making process.

[0189] The filtering algorithm uses statistical filtering or radius filtering methods to remove noise points and outliers from point cloud data. In this invention, the algorithm is used to preprocess point cloud data, improve data quality, and provide clean input for surface reconstruction.

[0190] The surface reconstruction algorithm is based on the Poisson reconstruction or Delaunay triangulation principle to generate a continuous triangular mesh model from point cloud data. In this invention, the algorithm is used to construct the topological structure of a three-dimensional model of a slope surface and define its geometric shape and spatial relationships.

[0191] The curvature estimation algorithm calculates the curvature value of each triangular mesh vertex by analyzing the geometric changes in the vertex neighborhood; in this invention, the algorithm is used to screen high curvature vertices in slope shoulders, steep slopes and crack areas as terrain feature points, supporting feature matching.

[0192] The time-varying rate calculation quantifies the slope deformation rate by taking the ratio of the difference in displacement change to the time interval. In this invention, the algorithm is used to analyze deformation trends and trigger an early warning when the rate of change exceeds a safe value, thereby achieving dynamic risk assessment.

[0193] The early warning triggering algorithm compares the rate of change over time based on a preset safety value. When the limit is exceeded, an early warning signal is generated and sent to the terminal device. In this invention, the algorithm is used for automated monitoring and real-time response, initiating the early warning process, including data recording and notification sending.

Claims

1. A slope safety monitoring method of multi-source data fusion, characterized in that, The method comprises the following steps: Step 1: Obtain the three-dimensional model of the slope surface and the two-dimensional radar deformation cloud image. The three-dimensional model of the slope surface comprises point cloud data of the slope surface generated by three-dimensional laser scanning and triangular mesh topology. The two-dimensional radar deformation cloud image comprises pixel position coordinates, corresponding displacement and acquisition time information collected by the slope radar. Step 2: Input the pixel position coordinates and corresponding displacement in the two-dimensional radar deformation cloud image into the pre-trained self-supervised deep learning model. The self-supervised deep learning model is trained using historical slope deformation data. The spatial feature vector is extracted by the encoder of the self-supervised deep learning model, and the spatial feature vector is mapped to three-dimensional heat map coordinate data including horizontal position, vertical displacement and elevation by the decoder of the self-supervised deep learning model. Step 3: Based on the terrain curvature threshold preset according to the geological characteristics of the slope, the key feature points are selected from the triangular mesh vertices. The curvature value of each vertex is calculated, and the vertex with a curvature greater than the threshold is selected as the terrain feature point. The terrain feature point comprises the vertex coordinates of the slope shoulder, steep ridge and crack area. Step 4: Match the feature points in the three-dimensional heat map coordinate data generated in step 2 with the terrain feature points selected in step 3. Map the three-dimensional heat map coordinates to the coordinate system of the three-dimensional model of the slope surface by the affine transformation algorithm, and generate a dynamic three-dimensional deformation heat map with a radar acquisition timestamp. Step 5: Compare the displacement values of the same coordinate point in the dynamic three-dimensional deformation heat map at different time stamps, calculate the displacement change value, and store the displacement change value data in the preset displacement database. Extract the displacement change value data from the displacement database, compare it with the threshold value preset according to the engineering safety standard, and identify the coordinate area with displacement change value exceeding the limit. Step 6: Mark the identified coordinate area with displacement change value exceeding the limit on the three-dimensional model of the slope surface with a color gradient, and output the three-dimensional visual monitoring result. Further comprising: Based on the dynamic three-dimensional deformation heat map with timestamp: Select the displacement change value of the same coordinate point in the two consecutive time stamps; Calculate the time change rate of the displacement change value difference; When the change rate exceeds the preset safety value, trigger the warning signal and send it to the terminal device; Further comprising: In response to the triggering of the warning signal, dynamically adjust the online learning rate of the self-supervised deep learning model according to the displacement change rate exceeding the limit amplitude value carried in the warning signal, increase the sampling frequency of the two-dimensional radar deformation cloud image according to the displacement change rate exceeding the limit amplitude value, activate the dynamic update mechanism of the vegetation spectrum feature library of the noise reduction processing, and correct the matching threshold based on the spectrum features of the current season; Overlay the heat map layer of the warning area in the three-dimensional visual monitoring interface, highlight the displacement change rate exceeding the limit coordinate point with pulsating red light, automatically associate the corresponding triangular mesh model structure parameters of the coordinate point, and output the real-time rock mass stress distribution simulation data to the terminal; Start the automatic cruising task of the unmanned aerial vehicle laser scanner, generate a surrounding flight path according to the warning coordinate position, and reacquire the point cloud data of the warning area with millimeter-level resolution, and trigger the surface reconstruction algorithm to update the three-dimensional model of the slope surface. The recollected data is input into the online learning mechanism, and the encoder and decoder weight parameters of the self-supervised deep learning model are iteratively optimized to regenerate the three-dimensional heat map coordinate data to replace the previously generated three-dimensional heat map coordinate data; Based on the updated dynamic three-dimensional deformation heat map, the displacement change and displacement change rate are recalculated, and if the change rate still exceeds the limit after three consecutive iterations, the warning level is upgraded.

2. The multi-source data fusion method for slope safety monitoring according to claim 1, characterized in that, The self-supervised deep learning model includes an encoder and a decoder; The encoder receives the pixel position coordinates and corresponding displacement in the two-dimensional radar deformation cloud image, extracts and outputs a spatial feature vector; The decoder receives the spatial feature vector output by the encoder and outputs three-dimensional heat map coordinate data.

3. The multi-source data fusion method for slope safety monitoring according to claim 2, characterized in that, Step 4 includes: Obtain a set of terrain feature points from the output of step 3, and extract key feature points from the three-dimensional heat map coordinate data output in step 2 using a feature extraction algorithm; The extracted key feature points and the obtained terrain feature points are paired using a feature matching algorithm, and the pairing results are mapped to the slope surface three-dimensional model coordinate system through an affine transformation algorithm.

4. The multi-source data fusion method for slope safety monitoring according to claim 3, characterized in that, Further comprising: Input the newly added two-dimensional radar deformation cloud image into the encoder of the self-supervised deep learning model through real-time collection by the slope radar; Update the weight parameters of the self-supervised deep learning model using the newly added two-dimensional radar deformation cloud image through an online learning algorithm, and regenerate three-dimensional heat map coordinate data based on the model with updated weight parameters; Align and fuse the regenerated three-dimensional heat map coordinate data with the historical three-dimensional heat map coordinate data stored in the displacement database using a coordinate transformation algorithm.

5. The multi-source data fusion method for slope safety monitoring according to claim 4, characterized in that, Further comprising: Before outputting the three-dimensional visualization monitoring result in step 6, perform noise reduction processing, extract color features from the texture data of the slope surface three-dimensional model using a color extraction algorithm, and compare the extracted color features with the preset vegetation spectrum feature library using a similarity algorithm. When the similarity value exceeds the preset matching threshold, filter the displacement change data of the corresponding coordinate points.

6. The multi-source data fusion's side slope safety monitoring method according to claim 5, characterized in that, Further comprising: Based on multi-source remote sensing images and ground spectrum measurement data, a color feature library is constructed, including slope rock mass spectrum features, vegetation spectrum features, and artificial structure spectrum features; Obtain color feature values from the texture of the slope surface three-dimensional model corresponding to each coordinate point in the displacement change data; Calculate the similarity values of the obtained color feature values and each type of feature in the color feature library using a similarity calculation algorithm; If the similarity value exceeds the preset similarity threshold and the maximum similarity corresponds to a vegetation category, the coordinate point is determined to be a noise point and the data of the coordinate point is removed from the displacement change data.

7. The multi-source data fusion method for slope safety monitoring according to claim 6, characterized in that, Further comprising: On the slope surface three-dimensional model of the three-dimensional visualization monitoring result output in step 6: Read the displacement change data from the displacement database and superimpose the displacement change contour lines using a contour generation algorithm; Label the displacement change levels with a color gradient using a color mapping algorithm according to the displacement change value range; Identify the maximum value coordinate area in the displacement change data and highlight the area using a highlight display algorithm.

8. The multi-source data fusion's side slope safety monitoring method according to claim 7, characterized in that, Further comprising: Collect slope surface point cloud data by mounting a laser scanner on a drone, and use the point cloud data to generate the slope surface three-dimensional model in step 1. The collected point cloud data is denoised by using a filtering algorithm, and a triangular mesh model is generated by using a surface reconstruction algorithm. The generated triangular mesh model is used as the topological structure basis of the slope surface three-dimensional model in step 1.

Images