Road slope stability image monitoring and risk early warning system

Abstract

The invention relates to the field of highway engineering safety monitoring, in particular to a highway slope stability image monitoring and risk early warning system which comprises an image acquisition module, a slope parameterization representation module, a multi-scale curvature flow analysis module, a Riemannian manifold learning module and a risk early warning module. Differential features are extracted by calculating a measurement tensor and a curvature tensor of the curved surface; constructing multi-scale representation of the slope curved surface by applying a curvature flow theory, and identifying deformation characteristics under different scales; according to the method, riemannian manifold learning and a geodesic convolutional network are utilized to extract advanced features, geological disaster precursors such as landslide, collapse and debris flow are accurately identified, the system has a self-adaptive monitoring frequency adjustment function, monitoring modes are automatically switched according to risk levels, the slope disaster risk is greatly reduced, and a powerful guarantee is provided for safe operation of roads.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of highway engineering safety monitoring, in particular to a highway slope stability image monitoring and risk early warning system for real-time monitoring of highway slopes and early warning of potential geological disasters. BACKGROUND

[0002] Highway slope stability is an important guarantee for safe operation of highways. Landslides, collapses and mudslides caused by slope instability not only threaten people's life and property safety, but also cause traffic disruption and economic losses. Traditional slope monitoring methods mainly rely on manual patrol and single-point measurement techniques such as displacement meters, inclinometers and stress meters, which have limited coverage, poor real-time performance and low automation level. In recent years, with the development of image processing and artificial intelligence technology, image-based slope monitoring methods have been gradually applied, but there are still the following technical problems:

[0003] 1. Most existing technologies are based on Euclidean geometry and linear analysis methods, which are difficult to effectively represent the complex nonlinear deformation process of the slope;

[0004] 2. Traditional methods do not adequately consider the multi-scale characteristics of slope disaster precursors, making it difficult to distinguish between environmental noise and real deformation;

[0005] 3. Existing pattern recognition methods do not perform well in handling the nonlinear manifold structure of slope features, affecting the accuracy of early warning;

[0006] 4. The existing system faces engineering challenges such as power supply, communication and device stability in long-term monitoring in complex outdoor environments.

[0007] Therefore, it is urgent to develop a slope monitoring system that can fully capture the deformation characteristics of the slope, accurately identify disaster precursors and achieve reliable early warning. SUMMARY

[0008] The purpose of the present application is to provide a highway slope stability image monitoring and risk early warning system that integrates differential geometry theory and slope monitoring technology to accurately identify slope disaster precursors and provide risk early warning.

[0009] The present application provides a highway slope stability image monitoring and risk early warning system, comprising:

[0010] An image acquisition module for obtaining high-resolution image data and three-dimensional point cloud data of the highway slope;

[0011] A slope parameterization representation module in communication with the image acquisition module for representing the highway slope as a two-dimensional Riemannian manifold and calculating the metric tensor and curvature tensor of the slope surface;

[0012] A multi-scale curvature flow analysis module, in communication with the slope parameterization representation module, configured to receive the metric tensor and the curvature tensor of the slope surface, construct a multi-scale representation of the slope surface, and extract a multi-scale feature map;

[0013] A Riemannian manifold learning module, in communication with the multi-scale curvature flow analysis module, configured to represent slope features in a Riemannian manifold space, extract high-level features through a geodesic convolution network, and identify disaster precursors such as landslides, collapses, and debris flows;

[0014] A risk warning module, in communication with the Riemannian manifold learning module, configured to generate risk level assessment and warning information according to the identified disaster precursors.

[0015] As a preferred, the image acquisition module comprises:

[0016] A high-resolution camera array configured to acquire image data of the highway slope;

[0017] A three-dimensional laser radar configured to acquire point cloud data of the highway slope;

[0018] A GPS positioning system configured to acquire spatial position information;

[0019] A data preprocessing unit configured to perform noise reduction, registration, and fusion processing on the image data and the point cloud data to generate a textured three-dimensional slope model.

[0020] As a preferred, the slope parameterization representation module comprises:

[0021] A parameterization mapping unit configured to construct a mapping relationship from a parameter domain to a three-dimensional coordinate of the slope;

[0022] A differential property calculation unit configured to calculate the metric tensor, the curvature tensor, the principal curvature, and the principal direction of the slope surface;

[0023] A differential invariant generation unit configured to calculate the Gaussian curvature, the mean curvature, and the shape index of the slope surface;

[0024] A time variation rate analysis unit configured to calculate the variation rate of the differential properties in the time dimension and mark abnormal regions with a variation rate exceeding a preset threshold.

[0025] As a preferred, the multi-scale curvature flow analysis module comprises:

[0026] A curvature flow evolution unit configured to construct a multi-scale representation of the slope surface;

[0027] A feature operator unit configured to apply differential feature operators at different scales to extract local features;

[0028] A feature map generation unit, used to organize features at different scales into feature maps;

[0029] The scale space analysis unit is used to analyze the evolution of features in the scale space and identify characteristic patterns related to disaster precursors.

[0030] Preferably, the Riemannian manifold learning module includes:

[0031] A manifold construction unit for defining appropriate Riemannian manifolds based on the intrinsic structure of the slope characteristics;

[0032] Geodesic convolutional network unit, used to perform convolution operations on Riemannian manifolds to extract high-level features;

[0033] Disaster precursor classification unit, used to identify different types of disaster precursors;

[0034] The risk probability estimation unit is used to calculate the probability distribution and spatiotemporal risk map of different disaster types.

[0035] Preferably, the risk warning module includes:

[0036] Risk assessment unit, used to assess risk levels based on the results of disaster precursor identification;

[0037] An early warning information generation unit, used to generate early warning information including risk type, location, level and disposal suggestions;

[0038] Multi-channel notification unit for sending warning information via SMS, email and monitoring platform;

[0039] The response strategy recommendation unit is used to recommend corresponding disposal measures based on the risk level.

[0040] Preferably, the system further comprises a monitoring mode control module for automatically adjusting the monitoring frequency and analysis depth according to the slope risk status, including a conventional monitoring mode, a warning monitoring mode and an emergency monitoring mode.

[0041] Preferably, the geodesic convolutional network unit comprises:

[0042] Geodesic convolution layer, used to perform convolution operations on Riemannian manifolds;

[0043] The tangent space pooling layer is used to perform feature aggregation and dimensionality reduction in the tangent space;

[0044] Parallel transmission layer, used to align and fuse features of different layers through parallel transmission;

[0045] Manifold Batch Normalization layer, used to perform batch normalization processing adapted to Riemannian geometry.

[0046] As preferred, the curvature flow evolution unit comprises:

[0047] a flow evolution equation solver for solving curvature flow evolution equations that control the smoothing process of the surface at different scales;

[0048] a feature preserving smoother for preserving key features in the smoothing process;

[0049] a scale space constructor for generating a sequence of surface representations covering different feature scales;

[0050] a critical point tracker for tracking the evolution trajectory of feature points in the curvature flow evolution process.

[0051] As preferred, the system further comprises a data management module for storing historical monitoring data, model parameters and early warning records, and providing data query, analysis and visualization functions; the data management module is communicatively connected with the image acquisition module, the slope parameterization representation module, the multi-scale curvature flow analysis module, the Riemannian manifold learning module and the risk early warning module for data sharing and collaborative analysis.

[0052] The present application adopts an innovative slope surface feature representation method based on differential geometry, a multi-scale curvature flow analysis technology and a disaster precursor identification model based on Riemannian manifold learning, combined with a hardware design with strong field adaptability, and has the following beneficial effects:

[0053] 1. By regarding the slope as a Riemannian manifold, the present application can comprehensively capture the complex nonlinear deformation characteristics of the slope, improving the accuracy and comprehensiveness of the monitoring;

[0054] 2. Based on the multi-scale curvature flow analysis technology, the present application can effectively distinguish between environmental noise and real disaster precursors, greatly improving the accuracy and reliability of the early warning;

[0055] 3. By using the Riemannian manifold learning method, the present application can adapt to the nonlinear distribution of slope features, improving the accuracy of disaster precursor identification;

[0056] 4. The system design fully considers the complexity of the field environment, and realizes long-term stable monitoring through solar power supply, wireless data transmission and modular design.

[0057] The present application realizes 3-7 day early warning of landslide precursors, 1-3 day early warning of collapse precursors and 12-24 hour early warning of debris flow precursors, and has important value in ensuring the safe operation of highways, reducing personnel casualties and property losses. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a schematic diagram of the overall architecture of the system of the present application;

[0059] Figure 2 is a structural schematic diagram of the image acquisition module of the present application;

[0060] Figure 3 is a workflow diagram of the slope parameterization characterization module of the present application;

[0061] Figure 4 is a processing flow diagram of the multi-scale curvature flow analysis module of the present application;

[0062] Figure 5 is a network structure diagram of the Riemannian manifold learning module of the present application;

[0063] Figure 6 is a workflow diagram of the risk early warning module of the present application;

[0064] Figure 7 is a structural schematic diagram of the geodesic convolution network unit of the present application;

[0065] Figure 8 is a structural schematic diagram of the curvature flow evolution unit of the present application. DETAILED DESCRIPTION

[0066] Reference Figures 1-8 will be made to the drawings and specific embodiments.

[0067] Reference Figure 1 The highway slope stability image monitoring and risk early warning system provided by the present application comprises an image acquisition module 1, a slope parameterization characterization module 2, a multi-scale curvature flow analysis module 3, a Riemannian manifold learning module 4, a risk early warning module 5, a monitoring mode control module 6 and a data management module 7. The modules are connected through communication interfaces to form a complete data processing link.

[0068] In a preferred embodiment of the present application, the system adopts a hierarchical architecture design and is divided into a data acquisition layer, a data processing layer and an application service layer from bottom to top. The data acquisition layer is mainly responsible by the image acquisition module 1, the data processing layer comprises the slope parameterization characterization module 2, the multi-scale curvature flow analysis module 3 and the Riemannian manifold learning module 4, and the application service layer is composed of the risk early warning module 5, the monitoring mode control module 6 and the data management module 7. This hierarchical design makes the system have good scalability and maintainability.

[0069] Reference Figure 2 The image acquisition module 1 of the present application comprises a high-resolution camera array 11, a three-dimensional laser radar 12, a GPS positioning system 13 and a data preprocessing unit 14.

[0070] The high-resolution camera array 11 is used to acquire image data of the highway slope. In one embodiment of the present application, the high-resolution camera array employs multiple industrial cameras with more than 20 million pixels, arranged in different directions along the slope, forming an observation network covering the overall slope. The sampling frequency of the cameras is automatically adjusted according to the monitoring mode, with a regular monitoring mode of once every 12 hours, a warning monitoring mode of once every 1 hour, and an emergency monitoring mode of up to once every 10 minutes.

[0071] The three-dimensional laser radar 12 is used to acquire point cloud data of the highway slope. Preferably, the three-dimensional laser radar employs a device with a point density of not less than 100 points per square meter and an accuracy of better than 5 millimeters to ensure the acquisition of fine structures of the slope surface. The scanning frequency of the laser radar is synchronized with the sampling frequency of the cameras to ensure the time consistency of the point cloud data and the image data.

[0072] The GPS positioning system 13 is used to acquire spatial position information. The present application employs a differential GPS system with centimeter-level accuracy to provide accurate geographic reference for image and point cloud data. The GPS system is not only used for spatial registration of data, but also for monitoring the position stability of the monitoring equipment itself to ensure the spatial consistency of long-term monitoring.

[0073] The data preprocessing unit 14 is used to perform noise reduction, registration and fusion processing on the image data and point cloud data to generate a textured three-dimensional slope model. Specifically, data preprocessing includes the following steps:

[0074] Data cleaning: removing abnormal points in point cloud through statistical filtering and removing noise in image through morphological processing;

[0075] Coordinate conversion: converting data from different sources to a unified coordinate system;

[0076] Time alignment: time aligning data collected at different times;

[0077] Data fusion: fusing point cloud data and image data to generate a textured three-dimensional model;

[0078] Region segmentation: extracting the slope region of interest and removing the background and irrelevant regions.

[0079] During data preprocessing, special attention needs to be paid to the normalization of different data sources. For example, for point cloud data, the spatial coordinates need to be normalized to the [-1, 1] interval according to the actual measurement range; for image data, brightness correction and color balance need to be performed to ensure that images collected under different time and weather conditions are comparable. In addition, during data fusion, accurate registration of point cloud and image needs to be ensured through control point matching, with an error controlled at the pixel level.

[0080] Reference Figure 3The slope parameterization characterization module 2 of the present application comprises a parameterization mapping unit 21, a differential characteristic calculation unit 22, a differential invariant generation unit 23 and a time variation rate analysis unit 24.

[0081] The parameterization mapping unit 21 is used to construct a mapping relationship from the parameter domain to the three-dimensional coordinates of the slope. In the present application, the slope is regarded as a two-dimensional Riemannian manifold embedded in a three-dimensional Euclidean space, and a parameterization representation system is established to lay a foundation for subsequent differential geometric analysis.

[0082] Specifically, the parameterization mapping unit 21 performs the following operations:

[0083] A local coordinate system is constructed with the center of the slope as the origin;

[0084] The slope area is divided into grid units, and the size of each unit is preferably 0.5 meters x 0.5 meters;

[0085] A parameter mapping is constructed to define the mapping relationship from the parameter domain (u, v) to the actual three-dimensional coordinates (x, y, z) of the slope;

[0086] A time dimension is added to form a spatio-temporal parameterization representation to capture the dynamic characteristics of the slope over time.

[0087] The parameter domain selection adopts a regular parameter domain to ensure good correspondence between the parameter space and the physical space. The grid resolution can be adjusted according to the monitoring accuracy requirements, generally in the range of 0.1-1 meters, and the default value is 0.5 meters, which achieves a good balance between accuracy and computational efficiency.

[0088] The differential characteristic calculation unit 22 is used to calculate the metric tensor, curvature tensor, principal curvature and principal direction of the slope surface. These differential characteristics are basic quantities for describing the local geometric characteristics of the surface and are crucial for identifying slope deformation.

[0089] In the calculation of differential characteristics, first, the first fundamental form (metric tensor) of the surface is calculated :

[0090] .

[0091] where: represents the vector function of the parameter surface, which maps the parameters (u, v) to the point (x, y, z) in the three-dimensional space; u and v are parameter coordinates; E, F and G are components of the metric tensor, representing the metric relationship between the parameter curves; "." represents vector dot product operation. The metric tensor is a basic tool for measuring distance and area on the surface.

[0092] Then, the second fundamental form (curvature tensor) b of the surface is calculated:

[0093] .

[0094] in: is the unit normal vector of the surface, calculated as , where × represents the vector cross product operation; L, M, and N are the components of the curvature tensor, which describe the degree of curvature of the surface.

[0095] Then, by solving the characteristic equation, we can calculate the principal curvatures and And the corresponding main directions:

[0096] .

[0097] in: Denotes the principal curvature; det denotes the determinant operation of the matrix. and are the two roots of the above characteristic equation, representing the maximum and minimum curvature of the surface at a certain point along a specific direction, and are important indicators for characterizing the local shape of the surface.

[0098] During the calculation process, a high-order central difference format is used to ensure the accuracy of numerical differentiation, regularization is used to deal with singular points that may appear in the curvature calculation, and natural boundary conditions are used to avoid boundary effects affecting the calculation results.

[0099] The differential invariant generation unit 23 is used to calculate the Gaussian curvature, mean curvature and shape index of the slope surface. These differential invariants are intrinsic characteristics of the surface and are not affected by the parameterization method. They are stable indicators that characterize the surface characteristics.

[0100] Gaussian curvature is the product of the two principal curvatures:

[0101] .

[0102] in: and are the principal curvatures; L, M, and N are the components of the curvature tensor; and E, F, and G are the components of the metric tensor. Gaussian curvature is a built-in property of a surface that indicates its overall curvature.

[0103] The mean curvature H is the average of the two principal curvatures:

[0104] .

[0105] Where: The symbols have the same meanings as above. The mean curvature indicates the average curvature of the surface.

[0106] Shape Index It is a dimensionless parameter based on the principal curvatures, used to characterize the local shape type of the surface:

[0107] .

[0108] where π is the constant pi; arctan is the inverse tangent function; ≥ denotes not less than . The shape index ranges from -1 to 1, and different values correspond to different surface types: S = -1 represents a spherical concave surface, S = -0.5 represents a groove shape, S = 0 represents a saddle surface, S = 0.5 represents a ridge shape, and S = 1 represents a spherical convex surface.

[0109] The time variation rate analysis unit 24 is used to calculate the variation rate of the differential characteristics in the time dimension, and mark the abnormal regions whose variation rate exceeds a preset threshold. This step is a key link for identifying abnormal deformation of the slope.

[0110] The time variation rate calculation needs to first accurately align the curvature data collected at different times, and then calculate the time derivative of each differential characteristic. For the Gaussian curvature, the variation rate ΔK is calculated as follows:

[0111] .

[0112] where K(u, v, t) represents the Gaussian curvature value at the parameter coordinates (u, v) at time t; Δt is the sampling time interval, and the unit is day. Similarly, the average curvature variation rate ΔH and the shape index variation rate ΔS can also be calculated in the same way.

[0113] According to empirical data and experimental results, the application sets variation rate warning thresholds for different types of slopes: for sandy slopes, the Gaussian curvature variation rate threshold is 0.05-0.1 / day; for rock slopes, the Gaussian curvature variation rate threshold is 0.02-0.05 / day. When a region whose variation rate exceeds the threshold is detected, the system will mark it as a potential abnormal region and enter the next step of analysis.

[0114] The time window length is automatically adjusted according to the monitoring mode, and is 7 days in the standard mode and 3 days in the accelerated deformation mode. The baseline selection can adopt a fixed baseline or a sliding baseline mode, and the sliding baseline mode has better adaptability to seasonal changes.

[0115] Referring to Figure 4 , the multi-scale curvature flow analysis module 3 of the application includes a curvature flow evolution unit 31, a feature operator unit 32, a feature atlas generation unit 33, and a scale space analysis unit 34.

[0116] The curvature flow evolution unit 31 is used to construct a multi-scale representation of the slope surface. In the present application, the curvature flow theory in differential geometry is applied to generate a series of surface representations at different scales by solving the curvature flow evolution equation, thereby capturing the slope features at different scales.

[0117] Curvature flow refers to the process of evolving a surface along its normal at a speed proportional to the curvature. In the simplest case, the mean curvature flow can be used, whose evolution equation is:

[0118] .

[0119] wherein: is the position vector of the surface, representing a point on the parameterized surface; is the scale parameter, controlling the smoothing degree; is the mean curvature; is the unit normal vector. The equation describes the process of gradually smoothing the surface as the scale parameter increases.

[0120] In practical applications, in order to preserve key features, the present application adopts a feature-preserving curvature flow, whose evolution equation is:

[0121] .

[0122] wherein: is a modulation function based on the Gaussian curvature , used to preserve key features during the smoothing process. A typical form of the modulation function is:

[0123] .

[0124] wherein: is a control parameter, used to adjust the degree of preservation of high-curvature areas; represents the absolute value of the Gaussian curvature . Preferably, takes a value in the range of 1-10, and a larger value will preserve high-curvature features more strongly.

[0125] During the curvature flow evolution process, the present application sets 5-8 scale levels, covering feature scales of 0.5-10 meters. The evolution step size is set to 0.01-0.05 according to the numerical stability requirement. The evolution is terminated when the curvature change rate is less than a preset threshold (usually 0.001) or the maximum number of iterations (usually 100) is reached.

[0126] The feature operator unit 32 is used to apply differential feature operators at different scales to extract local features. At each scale, a series of feature operators are applied to extract local features of the surface, including curvature operators, normal variation operators, and shape variation operators.

[0127] The curvature operator is based on the curvature tensor and is used to extract the bending features of the surface:

[0128] .

[0129] wherein: represents the curvature operator; , and represent the Gaussian curvature, mean curvature, and shape index at scale at the parameter coordinate ; the square brackets represent vector combination.

[0130] The normal variation operator is based on the rate of change of the normal vector and is used to detect abnormal changes in the surface normal:

[0131] .

[0132] wherein: represents the normal variation operator; represents the derivative of the normal vector with respect to the scale ; represents the Euclidean norm of the vector.

[0133] The shape variation operator is based on the variation of the shape index and is used to identify changes in the local shape of the surface:

[0134] .

[0135] wherein: represents the shape variation operator; represents the derivative of the shape index with respect to the scale ; represents the absolute value.

[0136] The feature response threshold is automatically adjusted according to the noise level, usually set to 2-3 times the standard deviation. This means that the region where the feature response value exceeds the background noise by 2-3 standard deviations is considered as an effective feature.

[0137] The feature map generation unit 33 is used to organize the features at different scales into a feature map. The feature map is a multi-dimensional data structure that records the distribution of features in the spatial and scale dimensions. Formally, the feature map can be represented as:

[0138] .

[0139] where: represents the feature map; is the parametric coordinate; is the scale parameter; represents the feature type (e.g., curvature, normal variation, or shape variation); is the response value of the corresponding feature operator.

[0140] When constructing the feature map, the feature weights of different scales need to be reasonably allocated. Generally, large-scale features have higher weights because they represent more significant structural changes. A typical weight allocation method is:

[0141] .

[0142] where: represents the weight of scale ; represents the th scale value; is the number of scale levels; represents the summation operation from to .

[0143] The scale space analysis unit 34 is used to analyze the evolution of features in the scale space and identify feature patterns related to disaster precursors. This step is crucial for extracting disaster precursor features from multi-scale features.

[0144] First, perform feature persistence analysis to evaluate the persistence of features in the scale space. Persistence refers to the ability of a feature to remain present in consecutive scales, used to filter transient noise. The persistence measure can be defined as:

[0145] .

[0146] where: represents the persistence measure of feature ; is an indicator function that takes 1 when the condition is true, and 0 otherwise; represents the response value of the feature operator at the parametric coordinate and scale ; is the feature response threshold; is the number of scale levels; represents the summation operation over all scales . The persistence threshold is usually set to 0.6-0.8, i.e., a feature needs to remain significant in 60% to 80% of scales to be considered persistent.

[0147] The evolution path of the feature in the scale space is then extracted to form a feature spectrum. The feature spectrum describes the behavior of the feature as a function of scale and is an important basis for identifying specific patterns. The feature spectrum can be represented as:

[0148] .

[0149] where: denotes the feature spectrum of feature ; denotes the response value of feature operator at parameter coordinates and scale ; the curly braces denote a set.

[0150] Finally, the extracted feature spectrum is matched with known disaster precursor patterns to identify potential disaster precursors. The pattern matching employs appropriate distance metrics in the feature space, such as geodesic distance or Mahalanobis distance. At the same time, anomalies that deviate significantly from normal evolution patterns are detected, which may be an indication of disaster precursors.

[0151] In the pattern matching process, the feature spectrum templates for different types of disasters are established based on historical cases and expert knowledge. For example, landslide precursors are usually characterized by an increase in the average curvature of a local region; collapse precursors are characterized by a sudden change in local Gaussian curvature; and debris flow precursors may be characterized by a gradual change in the shape index of a channel region.

[0152] Referring to Figure 5 , the Riemann manifold learning module 4 of the present application includes a manifold construction unit 41, a geodesic convolution network unit 42, a disaster precursor classification unit 43, and a risk probability estimation unit 44.

[0153] The manifold construction unit 41 is used to define an appropriate Riemann manifold according to the internal structure of the slope features. In the present application, the slope features are regarded as points distributed on the Riemann manifold, and by defining an appropriate Riemann metric, a mathematical foundation is provided for nonlinear feature learning.

[0154] Specifically, the manifold construction unit 41 performs the following operations:

[0155] Manifold structure definition: define an appropriate Riemann manifold according to the internal structure of the slope features;

[0156] Metric tensor design: design a Riemann metric tensor that reflects the similarity of features;

[0157] Geodesic line calculation: calculate the geodesic line between points on the manifold as a distance measure;

[0158] Tangent space mapping: establish a mapping from the manifold to the local tangent space for local analysis.

[0159] In the present application, the dimension of the Riemannian manifold is set according to the feature complexity, usually 3-8 dimensions. To prevent ill-conditioned cases, upper and lower bounds are set for the curvature of the manifold, usually [-10, 10]. In the selection of local coordinates, the intrinsic coordinate system that can reflect the intrinsic nature of the feature is preferred.

[0160] Riemannian metric tensor is a positive definite symmetric matrix at each point on the manifold, which defines the distance measure on the manifold:

[0161] .

[0162] where: represents the square of the infinitesimal distance; represents the components of the Riemannian metric tensor at point ; and represent the coordinate differential; represents the summation operation for all indices and .

[0163] In the present application, an adaptive Riemannian metric is designed according to the feature distribution characteristics:

[0164] .

[0165] where: is the Riemannian metric tensor at point ; is the Jacobian matrix of the feature mapping, reflecting the local variation of the feature; represents the transpose matrix of ; is a regularization parameter to prevent metric degeneration, usually taking values of 0.01-0.1; is the identity matrix, with the same dimension as .

[0166] The geodesic convolution network unit 42 is used to perform convolution operations on the Riemannian manifold and extract high-level features. Traditional convolutional neural network is designed for Euclidean space, and it is difficult to be directly applied to Riemannian manifold. The present application designs a geodesic convolution network based on Riemannian geometry, which meets the feature extraction requirements of non-Euclidean geometry.

[0167] The geodesic convolution network unit 42 includes a geodesic convolution layer, a tangent space pooling layer, a parallel transmission layer, and a manifold batch normalization layer.

[0168] The geodesic convolution layer performs convolution operations on the Riemannian manifold. Unlike convolution in Euclidean space, convolution on Riemannian manifold needs to consider curvature and geodesic distance. The mathematical expression of geodesic convolution is:

[0169] .

[0170] where: is the input feature function defined on the manifold ; is the convolution kernel function defined on the tangent space; is the inverse exponential map from point to point , mapping points on the manifold to the tangent space; is the volume element, where denotes the determinant of the metric tensor ; denotes the integral on the manifold ; denotes the convolution operation on the manifold. In practical computation, the geodesic convolution is approximated by defining the convolution kernel on the tangent space and using geodesic neighborhood sampling.

[0171] The tangent space pooling layer performs feature aggregation and dimension reduction in the tangent space. For a point on the manifold, its neighborhood points are first mapped to the tangent space , and then a standard pooling operation is performed in the tangent space. This approach avoids the difficulty of performing pooling directly on the manifold.

[0172] The parallel transport layer aligns and fuses features from different layers through parallel transport. In Riemannian geometry, parallel transport is the operation of transporting a vector along a geodesic while keeping its parallelism. Parallel transport can be represented as:

[0173] .

[0174] where: denotes the parallel transport operation from point to point ; is the parallel transport matrix describing how to transport a vector in the tangent space to the tangent space ; is the vector to be transported, located in the tangent space .

[0175] The manifold batch normalization layer performs batch normalization adapted to Riemannian geometry. Unlike batch normalization in Euclidean space, batch normalization on a Riemannian manifold needs to consider the local structure of the manifold. The form of manifold batch normalization is:

[0176] .

[0177] where: is the normalized feature; is the input feature; is the mean computed on the manifold; is the variance computed on the manifold; is a small constant to prevent division by zero, usually set to ; and are learnable parameters; denotes the multiplication operation adapted to the manifold structure.

[0178] In network design, the depth of the geodesic convolution network is set according to the complexity of the task, usually 6-12 layers. The size of the convolution kernel is defined in the geodesic distance space, usually covering the local neighborhood with a radius of 0.1-0.5 units of geodesic distance. The learning rate adopts an adaptive strategy, with an initial value of 0.001, which gradually decreases according to the training process.

[0179] The disaster precursor classification unit 43 is used to identify different types of disaster precursors. In this invention, we mainly focus on the precursor characteristics of three common geological disasters: landslides, collapses, and debris flows.

[0180] The classification decision adopts a softmax function to output the probability of different categories:

[0181] .

[0182] where: denotes the probability of the output category when given input ; is the network output corresponding to the category ; denotes the exponential function of ; denotes the sum of the exponential functions of all categories. The classification decision threshold is set according to risk preference, and in this invention, we prefer to reduce the false negative rate, so we set the recognition threshold to 0.3-0.4, that is, when the probability of a certain type of disaster exceeds this threshold, the system considers that there is a precursor of that type of disaster.

[0183] The risk probability estimation unit 44 is used to calculate the probability distribution of different disaster types and the spatiotemporal risk map. The risk probability distribution not only includes the probability of the disaster type, but also includes the estimation of the disaster occurrence time and intensity.

[0184] The time prediction is based on the feature evolution rate to construct a time model of disaster development:

[0185] .

[0186] where: is the predicted disaster occurrence time; is the current time; is the characteristic change threshold that triggers a disaster, determined from historical data; is the currently observed characteristic change; is the characteristic change rate, i.e., the ratio of the characteristic change amount to time.

[0187] The time prediction window is set to 24 hours, 72 hours, and 7 days, corresponding to short-term, medium-term, and long-term predictions, respectively. The credibility of the prediction decreases as the time window increases, and the system provides confidence intervals for predictions at different time windows.

[0188] The spatial risk distribution is represented by generating a risk heat map. The risk heat map is based on the spatial distribution and intensity of feature anomalies, and intuitively displays the location and range of risk areas. The risk level is divided into four levels: low risk (green), general risk (yellow), high risk (orange), and extremely high risk (red), corresponding to different warning and response measures.

[0189] Reference Figure 6 The risk warning module 5 of the present application includes a risk assessment unit 51, a warning information generation unit 52, a multi-channel notification unit 53, and a response strategy recommendation unit 54.

[0190] The risk assessment unit 51 is used to assess the risk level according to the disaster precursor identification result. The risk assessment considers factors such as disaster type, occurrence probability, potential impact range, and severity, and gives a comprehensive risk assessment result.

[0191] The risk level calculation uses the risk matrix method, which maps the disaster occurrence probability and potential impact severity into a risk matrix:

[0192] .

[0193] where: Risk represents the risk value; Probability represents the probability of disaster occurrence, divided into 5 levels (1 - extremely low, 2 - low, 3 - medium, 4 - high, 5 - extremely high); Severity represents the severity of the disaster, also divided into 5 levels (1 - slight, 2 - general, 3 - serious, 4 - very serious, 5 - catastrophic). The final risk level is divided into four levels according to the risk value: low risk (1-4), general risk (5-9), high risk (10-16), and extremely high risk (17-25).

[0194] The warning information generation unit 52 is used to generate warning information containing risk type, location, level, and disposal suggestions. The warning information adopts a structured format, including the following elements:

[0195] Warning number: a unique number identifying the warning event;

[0196] Warning time: the time of the warning release;

[0197] Risk type: the type of the early warning disaster (landslide, collapse or debris flow);

[0198] Risk location: the geographical location description and coordinates of the risk area;

[0199] Risk level: low risk, general risk, high risk or extremely high risk;

[0200] Estimated occurrence time: the predicted time range when the disaster may occur;

[0201] Impact range: the predicted range of the affected area;

[0202] Treatment suggestion: recommended treatment measures for the risk.

[0203] The multi-channel notification unit 53 is used to send early warning information through short messages, emails and monitoring platforms. To ensure the timely delivery of early warning information, the system uses a multi-channel redundant sending mechanism, including short messages, emails, dedicated App push and monitoring platform display. Different risk levels correspond to different notification strategies: low risk level only displays on the monitoring platform, general risk level adds email notification, high risk and extremely high risk level starts full-channel notification, including short message and telephone reminder.

[0204] The response strategy recommendation unit 54 is used to recommend corresponding treatment measures according to the risk level. For different risk levels and disaster types, the system provides customized response strategy suggestions:

[0205] Low risk: strengthen monitoring, shorten sampling interval, no special measures needed;

[0206] General risk: dispatch engineering personnel for on-site inspection, prepare emergency supplies and develop preliminary emergency plans;

[0207] High risk: implement traffic control, evacuate surrounding personnel, start emergency plan and deploy emergency treatment measures;

[0208] Extremely high risk: close roads, evacuate comprehensively, start the highest level of emergency response and mobilize rescue forces.

[0209] The development of response strategies takes into account many factors such as disaster characteristics, terrain conditions, weather factors and available resources, providing scientific reference for decision makers.

[0210] The present application also includes a monitoring mode control module 6 for automatically adjusting the monitoring frequency and analysis depth according to the slope risk state, including regular monitoring mode, alert monitoring mode and emergency monitoring mode.

[0211] Conventional Monitoring Mode is the system's default operating mode and is suitable for normal slope monitoring. In this mode, the system samples data every 12 hours, performs comprehensive analysis, calculates resource allocation balance, and completes data processing and analysis within 24 hours.

[0212] Alert Monitoring Mode activates when potential risks are detected and is suitable for situations with a moderate risk level. In this mode, the system sampling interval is shortened to one hour, focusing on abnormal areas and prioritizing computing resources to ensure that critical data analysis is completed within one hour.

[0213] Emergency Monitoring Mode, the system's highest level of monitoring, activates during high-risk or extremely high-risk situations. In this mode, the system's sampling interval is further shortened to 10 minutes, enabling real-time analysis of high-risk areas. Edge computing and cloud-based collaboration ensures real-time analysis and early warning within 10 minutes.

[0214] Transitions between modes are automated based on risk assessment results, with manual intervention also supported. Transitions from low-level to high-level can be triggered automatically, while transitions from high-level to low-level require manual confirmation to ensure safety.

[0215] Reference Figure 7 The geodesic convolutional network unit 42 of the present invention includes a geodesic convolution layer 421, a tangent space pooling layer 422, a parallel transmission layer 423 and a manifold batch normalization layer 424.

[0216] The geodesic convolution layer 421 is used to perform convolution operations on Riemannian manifolds. Traditional convolutional neural networks are designed for Euclidean space and are difficult to directly apply to Riemannian manifolds with non-zero curvature. The present invention designs a convolution operation based on geodesics and tangent space to achieve feature extraction on Riemannian manifolds.

[0217] The core idea of ​​geodesic convolution is to generalize the convolution operation in Euclidean space to Riemannian manifold. The specific implementation includes the following steps:

[0218] Define geodesic neighborhood: For a point x on the manifold, define its geodesic neighborhood For The geodesic distance is less than the preset radius The set of all points of ;

[0219] Construct a local coordinate system: Construct the tangent space at point x , and define the local orthogonal coordinate system;

[0220] Mapping to tangent space: via inverse exponential mapping Map neighborhood points to the tangent space;

[0221] Perform convolution: Perform standard convolution operation in tangent space;

[0222] Feature Aggregation: Map the convolution result back to the manifold.

[0223] The mathematical expression of geodesic convolution is:

[0224] .

[0225] Where: f*gW represents the result of geodesic convolution; f is the input feature function; W is the convolution kernel function; is the geodesic neighborhood of point x; is the inverse exponential map, which maps a point y on the manifold to the tangent space of point x; * represents the dot product operation.

[0226] The tangent space pooling layer 422 is used for feature aggregation and dimension reduction in the tangent space. The pooling operation is performed in the tangent space, avoiding the complexity of directly performing pooling on the manifold. The specific steps include:

[0227] Define the pooling region: define the geodesic neighborhood on the manifold as the pooling region;

[0228] Map to the tangent space: map the points within the pooling region to the tangent space of the center point;

[0229] Perform pooling: perform standard pooling operations (such as max pooling or average pooling) in the tangent space;

[0230] Feature aggregation: take the pooling result as the feature representation of the center point.

[0231] The parallel transport layer 423 is used to realize the alignment and fusion of different layer features through parallel transport. On the Riemannian manifold, feature vectors at different points are located in different tangent spaces and cannot be directly compared or combined. Parallel transport provides a method to transport vectors from one tangent space to another while maintaining vector parallelism.

[0232] The mathematical expression of parallel transport has been given in the previous text. In actual calculation, parallel transport can be realized by solving the parallel transport equation or using discrete approximation methods.

[0233] The manifold batch normalization layer 424 is used to perform batch normalization processing adapted to Riemannian geometry. Traditional batch normalization stabilizes network training by normalizing feature distribution in Euclidean space. On the Riemannian manifold, normalization needs to consider the local structure of the manifold, especially the Riemannian metric.

[0234] The steps of manifold batch normalization include:

[0235] Calculate the mean on the manifold: use the Fréchet mean under the Riemannian metric;

[0236] Calculate the variance on the manifold: based on the geodesic distance to the mean;

[0237] Standardization: map the feature to tangent space, perform standardization, and map back to manifold;

[0238] Scale and shift: apply learnable scale and shift parameters.

[0239] The mathematical expression of manifold batch normalization has been given in the foregoing.

[0240] With reference to Figure 8 , the curvature flow evolution unit 31 of the present application comprises a flow evolution equation solver 311, a feature preserving smoother 312, a scale space constructor 313, and a critical point tracker 314.

[0241] The flow evolution equation solver 311 is used to solve the curvature flow evolution equation that controls the smoothing process of the surface at different scales. Curvature flow is an important tool for studying the geometric evolution of surfaces. By solving the curvature flow equation, a multi-scale representation of the surface can be constructed.

[0242] In the present application, the mean curvature flow is used as the basic evolution equation, and its mathematical expression has been given in the foregoing. The mean curvature flow has the property of area minimization, which gradually smooths the surface and reduces the high-frequency changes of the surface during the evolution process. This feature makes it suitable for constructing a multi-scale representation from details to roughness.

[0243] In numerical solving, the evolution equation is discretized using explicit or implicit finite difference methods. To ensure numerical stability, the time step needs to satisfy the CFL condition, usually set to 0.01-0.05. The solution of the evolution equation can be represented as an iterative process:

[0244] .

[0245] where, represents the surface position vector when the scale parameter is ; represents the surface position vector when the scale parameter is ; is the scale step; is the mean curvature when the scale is ; is the unit normal vector when the scale is .

[0246] The feature preserving smoother 312 is used to preserve key features during the smoothing process. The standard mean curvature flow uniformly reduces all curvatures during the smoothing process, and cannot distinguish between noise and important features. In order to preserve the key features of the slope, the present application designs a feature-preserving curvature flow.

[0247] The key to feature preserving smoothing is to introduce a modulation function that adjusts the smoothing rate according to the local geometric properties. The modulation function can be based on Gaussian curvature, shape index or other geometric features. In this invention, a modulation function based on Gaussian curvature is adopted, whose mathematical expression is given in the previous section.

[0248] The scale space constructor 313 is used to generate a sequence of surface representations that cover different feature scales. A scale space is a sequence of parameterized surfaces that represent the original surface at different levels of smoothness. By analyzing the behavior of features in the scale space, one can distinguish noise from real deformations.

[0249] When constructing a scale space, one needs to determine the number and distribution of scale levels. This invention adopts 5-8 scale levels that cover feature scales from 0.5 to 10 meters. The distribution of scale parameters can be linear or logarithmic, with the logarithmic distribution being more effective in capturing multi-scale properties.

[0250] The scale space can be represented as:

[0251] .

[0252] where: is the scale space; is the parameter coordinate is the surface point at scale parameter is the number of scale levels.

[0253] The critical point tracker 314 is used to track the evolution trajectories of feature points during curvature flow evolution. Critical points are points where the curvature or its derivatives reach extrema, and are important markers of surface features. By tracking the evolution trajectories of critical points in the scale space, one can extract the scale space behavior of features.

[0254] The steps of critical point tracking include:

[0255] Identify critical points at each scale, such as Gaussian curvature extrema, mean curvature extrema or shape index feature points;

[0256] Establish the correspondence of critical points between adjacent scales to form tracking trajectories;

[0257] Analyze the behavior of trajectories, such as duration, displacement and intensity changes;

[0258] Classify features according to trajectory characteristics.

[0259] ​A key challenge in critical point tracking is handling bifurcation and coalescence events in the scale space. In these events, a feature can split into multiple branches or multiple features can merge into one. The present invention adopts a nearest neighbor and similarity metric based approach to solve this problem, ensuring continuity and accuracy of the tracking.

[0260] The present invention also includes a data management module 7 for storing historical monitoring data, model parameters and early warning records, and providing data query, analysis and visualization functions. The data management module 7 is in communication connection with the image acquisition module 1, the slope parameterization representation module 2, the multi-scale curvature flow analysis module 3, the Riemannian manifold learning module 4 and the risk early warning module 5 for data sharing and collaborative analysis.

[0261] The data management module 7 includes a data storage unit, a data query unit, a data analysis unit and a data visualization unit.

[0262] The data storage unit is responsible for the storage and management of raw data, processing results and early warning records. The storage adopts a hierarchical structure, including:

[0263] Raw data layer: stores raw data such as images and point clouds, adopts incremental backup strategy;

[0264] Feature data layer: stores extracted features and analysis results, supports fast retrieval;

[0265] Early warning record layer: stores historical early warning information and disposal records for traceability analysis;

[0266] Model parameter layer: stores trained model parameters and configuration information, supports model updating.

[0267] The data query unit provides multi-condition query and data export functions, supports flexible query by time, location, feature type and risk level, etc. The query interface adopts standardized design, supports integration and interaction with other systems.

[0268] The data analysis unit is responsible for in-depth analysis of historical data, including trend analysis, correlation analysis and prediction analysis. Trend analysis automatically generates time series charts of key indicators, visually displaying the trend of change; correlation analysis mines potential relationships between different factors, such as the correlation between weather conditions and slope deformation; prediction analysis predicts future risk changes based on historical data, providing a basis for long-term planning.

[0269] The data visualization unit converts complex monitoring data and analysis results into intuitive visual expressions, including three-dimensional models, heat maps, trend charts and risk maps, etc. The visualization interface supports multi-level display, from macro overview to micro detail, meeting the needs of users at different levels.

[0270] Through the data management module 7, the application realizes the whole life cycle management of the monitoring data, and provides a solid data foundation for the slope risk assessment and decision support.

[0271] The above-described embodiments only express specific implementation manners of the application, which are described in a more specific and detailed manner, but cannot be understood as a limitation on the patent scope of the application. It should be noted that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the application, and these all belong to the protection scope of the application.

Claims

1. Highway slope stability image monitoring and risk warning system, characterized by: include: Image acquisition module, used to obtain high-resolution image data and three-dimensional point cloud data of highway slopes; a slope parameterization characterization module, in communication with the image acquisition module, for characterizing the highway slope as a two-dimensional Riemannian manifold and calculating a metric tensor and a curvature tensor of the slope surface; a multi-scale curvature flow analysis module, communicatively connected to the slope parameterization characterization module, configured to receive the metric tensor and curvature tensor of the slope surface, construct a multi-scale representation of the slope surface, and extract a multi-scale feature map; a Riemannian manifold learning module, communicating with the multi-scale curvature flow analysis module, for representing slope features in a Riemannian manifold space, extracting high-level features through a geodesic convolutional network, and identifying precursors of landslides, collapses, and debris flows; The risk warning module is in communication with the Riemannian manifold learning module and is used to generate risk level assessment and warning information based on the identified disaster precursors.

2. The system according to claim 1, wherein: The image acquisition module includes: A high-resolution camera array to acquire image data of highway slopes; 3D LiDAR, used to obtain point cloud data of highway slopes; GPS positioning system, used to obtain spatial location information; The data preprocessing unit is used to perform noise reduction, registration and fusion processing on the image data and point cloud data to generate a textured three-dimensional slope model.

3. The system according to claim 1, wherein: The slope parameterization characterization module includes: The parameterized mapping unit is used to construct the mapping relationship from the parameter domain to the three-dimensional coordinates of the slope; A differential characteristic calculation unit is used to calculate the metric tensor, curvature tensor, principal curvature and principal direction of the slope surface; Differential invariant generation unit for calculating Gaussian curvature, mean curvature and shape index of slope surface; The time change rate analysis unit is used to calculate the change rate of the differential characteristic in the time dimension and mark the abnormal area where the change rate exceeds the preset threshold.

4. The system according to claim 1, wherein: The multi-scale curvature flow analysis module includes: Curvature flow evolution unit, used to construct multi-scale representation of slope surfaces; Feature operator unit, used to apply differential feature operators at different scales to extract local features; A feature map generation unit, used to organize features at different scales into feature maps; The scale space analysis unit is used to analyze the evolution of features in the scale space and identify characteristic patterns related to disaster precursors.

5. The system according to claim 1, wherein: The Riemannian manifold learning module includes: A manifold construction unit for defining appropriate Riemannian manifolds based on the intrinsic structure of the slope characteristics; Geodesic convolutional network unit, used to perform convolution operations on Riemannian manifolds to extract high-level features; Disaster precursor classification unit, used to identify different types of disaster precursors; The risk probability estimation unit is used to calculate the probability distribution and spatiotemporal risk map of different disaster types.

6. The system according to claim 1, wherein: The risk warning module includes: Risk assessment unit, used to assess risk levels based on the results of disaster precursor identification; An early warning information generation unit, used to generate early warning information including risk type, location, level and disposal suggestions; Multi-channel notification unit for sending warning information via SMS, email and monitoring platform; The response strategy recommendation unit is used to recommend corresponding disposal measures based on the risk level.

7. The system according to any one of claims 1 to 6, characterized in that The system also includes a monitoring mode control module for automatically adjusting the monitoring frequency and analysis depth according to the slope risk status, including conventional monitoring mode, warning monitoring mode and emergency monitoring mode.

8. The system according to claim 5, wherein: The geodesic convolutional network unit includes: Geodesic convolution layer, used to perform convolution operations on Riemannian manifolds; The tangent space pooling layer is used to perform feature aggregation and dimensionality reduction in the tangent space; Parallel transmission layer, used to align and fuse features of different layers through parallel transmission; Manifold Batch Normalization layer, used to perform batch normalization processing adapted to Riemannian geometry.

9. The system according to claim 4, wherein: The curvature flow evolution unit includes: Flow evolution equation solver, used to solve the curvature flow evolution equation that controls the smoothing process of the surface at different scales; Feature-preserving smoother, used to preserve key features during the smoothing process; A scale-space builder for generating sequences of surface representations covering different feature scales; Critical point tracker is used to track the evolution trajectory of characteristic points during the evolution of curvature flow.

10. The system according to claim 1, wherein: The system also includes a data management module for storing historical monitoring data, model parameters and warning records, and providing data query, analysis and visualization functions; the data management module establishes communication connections with the image acquisition module, the slope parameterization characterization module, the multi-scale curvature flow analysis module, the Riemannian manifold learning module and the risk warning module for data sharing and collaborative analysis.

Images