Dangerous rock falling motion feature prediction method based on multi-source data fusion

Abstract

The invention discloses a dangerous rock falling motion feature prediction method based on multi-source data fusion, and belongs to the technical field of engineering geology. Comprising the steps of collecting dangerous rock multi-source data to construct a basic information base; processing dangerous rock multi-source data based on a CNN algorithm, identifying a dangerous rock contour and calculating geometric features; the dangerous rock multi-source data is processed through multiple algorithms, and the terrain is reconstructed to generate a digital elevation and terrain model; analyzing a vibration signal frequency spectrum, quantifying dangerous rock multi-source data features and determining a fusion weight; fusing the multi-source data of the dangerous rock to construct a three-dimensional real scene model and construct a physical and mechanical model; establishing a historical database, selecting a machine learning algorithm to construct a data driving model, and adopting an optimization algorithm to predict a rockfall movement track and state; and determining a two-dimensional simulation key section, establishing a two-dimensional numerical model, and dynamically simulating and calculating rockfall movement and rock-soil body response. According to the scheme, the defects in the aspects of dangerous rock falling motion feature prediction and protection design in the prior art are overcome, and the prediction accuracy and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of engineering geology, and particularly relates to a method for predicting the movement characteristics of dangerous rocks and falling stones based on multi-source data fusion. Background Art

[0002] The mountainous terrain is complex and the geological conditions are changeable. As a common geological disaster, dangerous rocks and falling stones pose a serious threat to the construction of mountain infrastructure, transportation, and the safety of residents' lives and property. Collapse and falling stones are characterized by strong suddenness, fast movement speed, large impact force, and wide damage range, which may lead to road interruptions, bridge collapses, house damage, and even casualties and significant economic losses. With the continuous advancement of mountain engineering construction, such as the construction of highways, railways, and hydropower projects, the prevention and control of dangerous rock and falling stone disasters have become a key issue to be solved urgently.

[0003] Most traditional analyses of the movement of dangerous rocks and falling stones adopt two-dimensional methods. Based on simplified plane models, these methods cannot accurately describe the three-dimensional movement characteristics of dangerous rocks and falling stones under complex terrain conditions. In the actual mountain environment, factors such as terrain undulation, slope changes, and valley intersections have a significant impact on the movement trajectory and speed of falling stones. Two-dimensional analysis is difficult to consider these factors, resulting in a large deviation between the prediction results and the actual situation and unable to provide an accurate basis for engineering protection design.

[0004] For example, a method and system for predicting the movement characteristics of mountain railway collapse and falling stones disclosed in Chinese Patent CN117807758A indicates in its background art that this method can only target the occurred falling stone events, unable to obtain instant movement characteristic parameters and data, and difficult to reflect the influence of the randomness of falling stone movement. In mountainous sections with complex terrain, two-dimensional analysis cannot accurately predict the movement changes of falling stones at special terrains such as valleys and ridges, which may lead to unreasonable settings of protection structures.

[0005] Although three-dimensional numerical simulation technology can, to a certain extent, simulate the complex movement process of dangerous rocks and falling stones, single three-dimensional numerical simulation has the defects of large calculation amount and long calculation time. When dealing with large-scale mountain scenes and complex geological conditions, it requires a large amount of computing resources and time costs, and is not convenient for rapid application in actual projects. Moreover, it has high requirements for data, requires accurate geological models and rock mechanics parameters, and the data acquisition and processing are difficult, easily introducing errors and affecting the reliability of prediction results.

[0006] With the continuous development of sensor technology, information technology, and computer science, multi-source data fusion technology has gradually attracted attention in the field of geological disaster prediction. Multi-source data fusion can integrate data from different sensors, different measurement methods, and different time and space scales. Through data processing and analysis algorithms, it can extract more valuable information and improve the accuracy and reliability of disaster prediction. In other geological disaster fields, such as earthquake prediction and landslide monitoring, multi-source data fusion technology has achieved certain research results. However, its application in predicting the movement characteristics of dangerous rocks and falling stones is relatively less, and a mature and efficient method system has not yet been formed.

[0007] However, the simple superposition or separate processing of multi-source data lacks an effective fusion mechanism, cannot fully explore the internal relationships of the data, and has an incomplete and in-depth understanding of the movement characteristics. For example, some scholars have collected UAV images, displacement data from ground monitoring stations, and meteorological data. When processing these data, they only perform simple terrain analysis on the images, trend judgment on the displacement data, and statistical analysis on the meteorological data, and finally simply superimpose the results to evaluate the state of dangerous rocks, without achieving true data fusion. Simply superimposing meteorological data and displacement data cannot accurately reflect the complex coupling relationship between rainfall and the deformation of dangerous rocks, and cannot provide an accurate basis for predicting movement characteristics ("Review Report on the Movement Characteristics and Protection Research of Dangerous Rocks and Falling Stones on Rock Slopes", Luo Tian).

[0008] In summary, there are many deficiencies in the existing prediction and protection design technologies for the movement characteristics of dangerous rocks and falling stones. For example, the two-dimensional analysis method model is too simplified to accurately present the three-dimensional movement characteristics under complex terrain, resulting in large deviations in prediction results and unable to provide an accurate basis for protection design. Single three-dimensional numerical simulation has a large amount of calculation, long time, high requirements for data and difficult to obtain, is prone to introducing errors, and is not suitable for rapid decision-making in practical engineering. The simple superposition or separate processing of multi-source data lacks an effective fusion mechanism, cannot fully explore the internal relationships of the data, and has an incomplete and in-depth understanding of the movement characteristics. Summary of the Invention

[0009] The purpose of the present invention is to address some technical problems existing in the prior art. The present invention aims to provide a method for predicting the movement characteristics of dangerous rocks and falling stones based on multi-source data fusion, overcome the deficiencies in the prediction of the movement characteristics of dangerous rocks and falling stones and protection design in the prior art, improve the accuracy and efficiency of prediction, and provide a scientific and reliable basis for the prevention and control of dangerous rock and falling stone disasters.

[0010] Specifically, the present invention provides a method for predicting the movement characteristics of dangerous rocks and falling stones based on multi-source data fusion, including the following steps:

[0011] S1. Collect multi-source data of dangerous rocks to construct a basic information database;

[0012] S2. Process the multi-source data of dangerous rocks based on the CNN algorithm, identify the contours of dangerous rocks and calculate geometric features;

[0013] S3. Process the multi-source data of dangerous rocks through various algorithms, reconstruct the terrain to generate digital elevation and terrain models;

[0014] S4. Analyze the vibration signal spectrum, quantify the characteristics of the multi-source data of dangerous rocks and determine the fusion weights;

[0015] S5. Based on the fusion weights, fuse the multi-source data of dangerous rocks to construct a three-dimensional real-scene model, and combine indoor rock mechanics tests and on-site geological surveys to construct a physical mechanics model;

[0016] S6. Establish a historical database and select a machine learning algorithm to construct a data-driven model, and use an optimization algorithm with an adaptive step size to realize the collaborative calculation of the physical mechanics model and the data-driven model, and predict the movement trajectory and state of the falling rocks;

[0017] S7. Determine the two-dimensional simulation key section based on the simulation results of the three-dimensional real-scene model and the digital elevation and terrain models, establish a two-dimensional numerical model and accurately depict the relevant elements, and dynamically simulate and calculate the movement of the falling rocks and the response of the rock and soil mass.

[0018] Preferably, for collecting the multi-source data of dangerous rocks, according to the terrain, area and distribution of dangerous rocks in the area to be predicted, use a drone to fly and collect data according to a preset flight route. The dual-light camera takes spectral images at a set time interval during flight, the lidar scanner emits laser pulses and receives reflected signals to obtain point cloud data, and the high-definition camera records video data throughout the process to collect the multi-source data of dangerous rocks; according to the geological structure and the distribution of dangerous rocks, arrange seismic geophones and microseismic sensors in a triangular array at key positions around and inside the area to be predicted to obtain vibration signals.

[0019] Preferably, the S2 specifically includes the following steps: training the model with a large number of labeled samples to realize automatic identification and accurate extraction of the contours of dangerous rocks, and converting them into a unified vector format

[0020] S21. Use the VGGNet architecture as the basis, add an attention mechanism to the network structure, and construct a deep learning model based on a convolutional neural network;

[0021] S22. Use the multi-source data of dangerous rocks as training samples and make detailed annotations. The annotation content includes the accurate contours, categories and stability states of dangerous rocks;

[0022] S23. Divide the labeled samples into a training set, a validation set, and a test set according to the ratio of 7:2:1. Use the training set to train the CNN model. During the training process, adopt the stochastic gradient descent optimization algorithm. Adjust the learning rate according to the accuracy of the validation set every set number of training rounds. When the accuracy no longer improves, reduce the learning rate to 0.1 times the current value and further train until the accuracy of the deep learning model on the validation set reaches the specified requirements;

[0023] S24. Input the spectral image and video data into the trained deep learning model and perform feature judgment based on pixel points. Determine the dangerous rock contour through the adaptive threshold segmentation method;

[0024] S25. Convert the determined dangerous rock contour into a unified vector format and calculate its geometric features at the same time.

[0025] Preferably, when calculating its geometric features, first record the coordinate information of each point on the dangerous rock contour, and then connect these points in sequence to form a polygon. Among them, the geometric features include area, perimeter, and centroid coordinates.

[0026] Preferably, the calculation of the area is specifically expressed as:

[0027]

[0028] where n is the number of contour points. When i = n, x n+1 = x 1 , y n+1 = y 1 ; x i is the abscissa in the coordinate information, and y i is the ordinate in the coordinate information;

[0029] The calculation of the perimeter is specifically expressed as:

[0030]

[0031] The calculation of the ordinate of the centroid coordinate is specifically expressed as:

[0032]

[0033] Preferably, the S3 specifically includes: registering the point cloud data by using the ICP algorithm, removing the ground point cloud through the slope-based filtering algorithm, and then performing terrain reconstruction on the processed non-ground point cloud by using the Delaunay triangulation algorithm to generate a digital elevation and terrain model.

[0034] Preferably, the step S4 specifically includes: preprocessing the vibration signal, removing noise interference by using the wavelet transform denoising method, then performing filtering processing, and at the same time, amplifying the effective signal, performing spectrum analysis by using the fast Fourier transform algorithm, identifying the spectrum signals related to the dangerous rock activities, analyzing the vibration energy distribution in different frequency bands, calculating the energy spectral density, and judging whether the energy in this frequency band is abnormal according to the range of the characteristic spectrum signals related to the dangerous rock activities; for the multi-source data of the dangerous rock, extracting the multi-source data features, including geometric shape features, texture features, spatial distribution features, amplitude features, energy features and frequency features, quantifying and dimension-reducing the extracted features by using the principal component analysis method, and then determining the fusion weights based on the method combining information entropy and analytic hierarchy process.

[0035] Preferably, the step S5 specifically includes the following steps: fusing the multi-source data of the dangerous rock based on the fusion weights to obtain the comprehensive information of the dangerous rock, and then constructing a three-dimensional real-scene model, carrying out indoor rock mechanics tests to determine the rock mechanics parameters, combining the on-site geological exploration data to determine the joint fracture characteristics, selecting a suitable constitutive relation model, obtaining the initial in-situ stress state of the area to be predicted through on-site in-situ stress measurement, using the measurement results as boundary conditions, and constructing a three-dimensional numerical model by using finite element analysis software.

[0036] Preferably, the step S6 specifically includes the following steps: collecting the data of historical collapse and rockfall events, meteorological data, seismic activity data and human engineering activity records, and performing data cleaning and preprocessing, and then establishing a historical database, selecting a machine learning algorithm to construct a data-driven model, including a motion trend prediction model, dividing the data set based on the historical database to train the model, optimizing the parameters through the training results, establishing a parameter transfer interface between the physical mechanics model and the motion trend prediction model, and realizing collaborative calculation by using an optimization algorithm with an adaptive step size, considering the influence of various complex factors, and predicting the motion trajectory and state of the rockfall.

[0037] Preferably, the step S7 specifically includes: determining the two-dimensional simulation key section according to the simulation results of the three-dimensional real-scene model and the digital elevation and terrain model, and at the same time, combining with the topographic contour map, selecting the key section and obtaining its topographic parameters and geotechnical parameters, establishing a two-dimensional numerical model and accurately depicting the relevant elements, using the multi-source data feedback of the dangerous rock and the on-site information update mechanism to dynamically adjust the boundary conditions, topographic parameters and geotechnical parameters, and simulating the motion of the rockfall and the response of the geotechnical body by using the finite element-discrete element coupling method.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. Through multi-source data fusion, the advantages of different types of data are fully integrated, various information of dangerous rocks is comprehensively obtained, and the limitations of a single data source are avoided. For example, UAV aerial photography data provides the shape and terrain information of dangerous rocks, ground vibration monitoring data reflects the stability state of geological bodies, and combined with historical disaster data, etc., it can more accurately depict the initial state and potential movement trend of dangerous rocks, thereby improving the accuracy of predicting the movement characteristics of dangerous rock falls.

[0040] 2. The fusion of physical mechanics models and data-driven models and the two-way coupling mechanism enable the prediction model to not only perform accurate calculations based on physical principles but also dynamically adjust and optimize in combination with actual data, effectively improving the coincidence degree between the simulation results and the actual situation. For example, when considering the influence of ground motion, the calculation parameters of seismic forces are continuously corrected according to historical earthquake data and real-time monitoring data, making the prediction of the movement trajectory of rock falls more accurate.

[0041] 3. In the data fusion stage, an adaptive data fusion algorithm is adopted to automatically determine weights according to data characteristics, reducing unnecessary data processing and calculation amounts, and improving data processing efficiency. At the same time, in two-dimensional numerical simulations, key cross-sections are dynamically determined according to three-dimensional simulation results and actual situations, avoiding the huge computational cost of comprehensively performing two-dimensional simulations on the entire area, and performing refined simulations targeted, significantly improving the computational efficiency, making the entire prediction method more suitable for practical engineering applications.

[0042] 4. The AR / VR-based visualization and interaction verification system provides an intuitive simulation result display and interactive operation platform for engineering personnel. By simulating different protection schemes in a virtual environment and observing the movement response of rock falls in real time, the protection effect can be quickly evaluated, so as to select the optimal protection design scheme, improve the scientificity and rationality of protection projects, and effectively reduce the risk of dangerous rock falls disasters.

[0043] 5. The dynamically adaptive two-dimensional numerical simulation can timely adjust model parameters according to real-time monitoring data and on-site exploration information, adapt to the complex and changeable environmental conditions in mountainous areas, such as the influence of working conditions changes such as rainfall, earthquake, and human engineering activities on the movement of dangerous rock falls, ensure that the prediction results are always consistent with the actual situation, enhance the adaptability and reliability of the entire prediction method, and provide strong support for the long-term safety guarantee of mountainous engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the overall process of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] Example 1: As Figure 1 shown, a method for predicting the movement characteristics of dangerous rock falls based on multi-source data fusion includes the following steps:

[0046] S1. The drone collaborates with ground monitoring to collect multi-source data of dangerous rocks and build a basic information database.

[0047] The steps are as follows: Use the drone to carry a multi-spectral camera, a lidar scanner, and a high-definition camera, plan a reasonable flight path, conduct collaborative aerial photography of the area to be predicted, and obtain high-resolution spectral images of dangerous rocks, high-precision three-dimensional point cloud data, and dynamic video data. On the ground, arrange seismic detectors and microseismic sensor networks in a specific array according to the geological structure and the distribution of dangerous rocks, and continuously monitor the weak vibration signals of geological bodies.

[0048] S2. Process the images based on the CNN algorithm to accurately identify the outlines of dangerous rocks and convert them into vector format.

[0049] The steps are as follows: Use the CNN algorithm based on deep learning to preprocess the multi-spectral images and video data, train the model with a large number of labeled samples, achieve automatic identification and accurate extraction of the outlines of dangerous rocks, and convert them into a unified vector format.

[0050] S3. Use algorithms such as ICP to process the point cloud, reconstruct the terrain, and generate a high-precision digital elevation and terrain model.

[0051] The steps are as follows: Use the ICP algorithm to register the lidar point cloud data, remove the ground point cloud through a filtering algorithm, and then use the Delaunay triangulation algorithm to reconstruct the terrain to generate a high-precision digital elevation model and a three-dimensional terrain surface model.

[0052] S4. Analyze the frequency spectrum of the vibration signal, quantify the characteristics of multi-source data, and determine the fusion weights.

[0053] The steps are as follows: After denoising, filtering, and amplifying the vibration signal, perform frequency spectrum analysis using the FFT algorithm, extract characteristic parameters such as frequency, amplitude, and phase, and identify the spectral signals related to the activities of dangerous rocks. Extract the key characteristics of data such as images, point clouds, and vibration signals, use methods such as PCA and ICA to quantify and reduce the dimensions, and then determine the data fusion weights based on the method combining information entropy and AHP (Analytic Hierarchy Process).

[0054] S5. Integrate the data to build a three-dimensional real-scene model, and establish a physical mechanics model in combination with experiments and explorations.

[0055] The steps are as follows: Build a weighted average fusion model, integrate multi-source data to obtain comprehensive information of dangerous rocks, build a three-dimensional real-scene model including spatial form, geological attributes, and mechanical states, and ensure the high resolution and high precision of the model. Conduct indoor rock mechanics experiments and on-site geological explorations, determine the rock mechanics parameters and the characteristics of joints and fissures, select a suitable constitutive relation model, and build a three-dimensional numerical model in combination with the results of in-situ stress measurements.

[0056] S6. Build a machine learning model based on historical data and cooperate with the physical model to predict the trend of rockfall.

[0057] It includes the following steps: collect historical data to establish a database, select a machine learning algorithm (such as SVM) to construct a motion trend prediction model, divide the data set to train the model, and optimize the parameters to improve the prediction accuracy. Establish an interface for parameter transfer between the two models, design a dynamic optimization algorithm to achieve collaborative calculation, consider the influence of various complex factors, and accurately predict the motion trajectory and state of the rockfall.

[0058] S7. Determine the two-dimensional section according to the three-dimensional simulation, and dynamically simulate and calculate the motion of the rockfall and the response of the rock and soil mass.

[0059] It includes the following steps: determine the key two-dimensional simulation section according to the three-dimensional simulation results and terrain features, obtain the terrain and rock and soil mass parameters of the section, establish a two-dimensional numerical model and accurately depict the relevant elements. According to the real-time monitoring and exploration information, use the real-time data feedback and on-site information update mechanism to dynamically adjust the model boundary conditions and parameters, and use the finite element-discrete element coupling method to simulate and calculate.

[0060] S8. Build an AR / VR platform for visualization, interactively verify and optimize the protection plan, and improve the model algorithm.

[0061] It includes the following steps: calculate the interaction force and motion parameters between the rockfall and the slope surface in real time, obtain accurate motion characteristic data, analyze the disturbance effect of the rockfall on the surrounding rock and soil mass, and provide a basis for the protection design. Configure high-performance hardware, develop a software platform based on engines such as Unity3D, import the simulation result data to construct a virtual scene, and develop a scene management and data interaction module. Use interactive devices to enable users to operate the virtual scene, develop an evaluation index system and an optimization algorithm, verify and optimize the protection plan through user interaction, and analyze the feedback data to improve the model algorithm.

[0062] Embodiment 2: A method for predicting the motion characteristics of dangerous rock and rockfall based on multi-source data fusion, including the following steps:

[0063] As Figure 1 shown, it is implemented according to the following detailed steps:

[0064] S1. Cooperate the unmanned aerial vehicle and ground monitoring to collect multi-source data of dangerous rocks and construct a basic information database.

[0065] The DJI Matrice 600Pro multi-rotor drone is selected as the flight platform. It has good stability and wind resistance, and can adapt to the complex meteorological conditions in mountainous areas. Equipped with a dual-light camera from Sentera (spectral resolution can reach 2.5nm), it can obtain spectral images in different bands, which helps to analyze the material composition and surface characteristics of dangerous rocks; the VUX-1UAV lidar scanner from Riegl (point cloud density can reach 150 points / m 2 ), is used to accurately obtain the three-dimensional spatial information of the terrain and dangerous rocks; the HERO9 Black high-definition camera from GoPro (frame rate can reach 120fps) can record the dynamic change process of dangerous rocks.

[0066] According to the terrain, area and distribution of dangerous rocks in the area to be predicted, the Pix4Dcapture software is used to plan the flight route. To ensure the integrity and accuracy of the data, the overlap rate between adjacent flight routes is set to 75%, and the side overlap rate is set to 80%. The flight altitude is determined according to the terrain undulation and the height of dangerous rocks in the mountainous area, generally between 100 - 150 meters, which can not only ensure the coverage range but also obtain data with sufficient resolution. The flight speed is controlled at 8 m / s to reduce data blurring or loss caused by too fast flight speed.

[0067] Under good weather conditions (such as clear and gentle breeze), the drone is started to fly and collect data according to the preset route. The dual-light camera automatically takes spectral images at set time intervals during flight, the lidar scanner continuously emits laser pulses and receives reflected signals to obtain point cloud data, and the high-definition camera records video data throughout the process. At the same time, the flight control system of the drone records information such as flight attitude, position, and time for precise spatial positioning and registration during subsequent data processing.

[0068] The GS-11D seismometer from Geospace and the Trillium CompactPosthole microseismic sensor from Nanometrics are selected. They have high sensitivity and wide-band response characteristics, and can effectively monitor the weak vibration signals of geological bodies. According to the geological structure and the distribution of dangerous rocks, sensors are arranged in a triangular array at key positions around and inside the area to be predicted. For example, a sensor is arranged every 30 meters in areas with concentrated distribution of dangerous rocks, complex geological structures, and areas affected by possible rockfalls to ensure that the vibration information of geological bodies can be comprehensively captured.

[0069] The sensor transmits the collected vibration signals to the data acquisition station in real time through wired or wireless transmission methods. The data acquisition station uses a high-precision data acquisition card (such as NI 9234), and the sampling frequency is set to 2000 Hz to meet the acquisition requirements of high-frequency vibration signals. The collected data is preliminarily processed and stored at the data acquisition station. A large-capacity hard disk (such as 4TB) is selected as the storage device to ensure the long-term continuous storage of monitoring data. At the same time, the data acquisition station has the functions of real-time data display and preliminary analysis, which is convenient for on-site staff to timely understand the vibration situation of the geological body.

[0070] S2. Process the image based on the CNN algorithm to accurately identify the contour of the dangerous rock and convert it into a vector format.

[0071] Build a deep learning model based on the convolutional neural network (CNN), using the VGGNet architecture as the basis and improving it according to the dangerous rock identification task. An attention mechanism is added to the network structure to enable the model to pay more attention to the key feature areas of the dangerous rock. Collect a large number of multi-spectral images and high-definition video screenshots containing dangerous rocks as training samples, with the number of samples not less than 12,000. Make detailed annotations on the samples, and the annotation content includes the accurate contour of the dangerous rock, the category (such as block-shaped dangerous rock, fragmented dangerous rock, isolated dangerous rock), and the stability status (such as stable, potentially unstable, unstable), etc.

[0072] Divide the annotated samples into a training set, a validation set, and a test set according to the ratio of 7:2:1. Use the training set to train the CNN model. During the training process, the stochastic gradient descent (SGD) optimization algorithm is adopted, the initial value of the learning rate is set to 0.001, and the momentum is 0.9. And as the training progresses, every 10 epochs (training rounds), the learning rate is automatically adjusted according to the accuracy of the validation set. When the accuracy no longer improves, the learning rate is reduced to 0.1 times the original for further training. The number of training iterations is determined according to the convergence situation of the model, generally not less than 1500 times, until the accuracy of the model on the validation set reaches more than 96%.

[0073] Input the collected multi-spectral images and video data into the trained CNN model. The model extracts features and classifies and identifies the images through network structures such as convolutional layers, pooling layers, and fully connected layers. For each pixel point in the image, the model determines whether it belongs to the dangerous rock area based on its features, and determines the boundary of the dangerous rock through an adaptive threshold segmentation method. The specific calculation process is as follows: Let I(x, y) represent the gray value of the pixel point with coordinates (x, y) in the image, calculate the gray mean μ and standard deviation σ within the neighborhood of this pixel point (such as a 5×5 pixel window), then the threshold T = μ + kσ (where k is a coefficient determined according to the training results, generally taking 1.5 - 2.5). When I(x, y) > T, it is determined that this pixel point belongs to the dangerous rock area, otherwise it belongs to the background area. In this way, the dangerous rock area is distinguished from the background.

[0074] Convert the identified dangerous rock contour information into a unified vector format, such as the Shapefile format. The specific calculation process is as follows: For each point on the contour, record its coordinate information (x i , y i ), and then connect these points in sequence to form a polygon, which is the vector contour of the dangerous rock. At the same time, calculate geometric features such as the area A, perimeter C, and centroid coordinates (x g , y g ) of the dangerous rock. The calculation formulas are as follows:

[0075]

[0076] where n is the number of contour points. When i = n, x n+1 = x 1 , y n+1 = y 1 .

[0077]

[0078] S3. Use algorithms such as ICP to process the point cloud and reconstruct the terrain to generate a high-precision digital elevation and terrain model.

[0079] From the collected multi-viewpoint cloud data, select a representative, wide-coverage, and high-precision point cloud as the reference point cloud. For the point clouds of other viewpoints, use a feature-based registration method for initial registration.

[0080] First, extract feature points (such as corner points and edge points) in the point cloud, and match the corresponding feature points in different viewpoint point clouds by calculating the descriptors of the feature points (such as the FPFH feature descriptor). According to the matched feature point pairs, use the singular value decomposition (SVD) algorithm to calculate the initial transformation matrix and preliminarily register other point clouds to the coordinate system of the reference point cloud.

[0081] The Iterative Closest Point (ICP) algorithm is used to accurately register the point cloud after initial registration. In the ICP algorithm, for each point P in the point cloud to be registered i , its nearest neighbor point Q is found in the reference point cloud i , and the distance d between the corresponding points is calculated i =||P i -Q i ||. Then, the transformation matrix is optimized by minimizing the sum of the squares of the distances between all corresponding points . Let the transformation matrix T consist of a translation vector and a rotation matrix R. The transformation matrix is updated by solving the following optimization problem:

[0082]

[0083] An iterative solution to the above optimization problem is adopted using a gradient descent-based method until a preset convergence condition is met, such as the difference between the transformation matrices of two adjacent iterations being less than a certain threshold (e.g., 0.0005) or the root mean square error of the distances between corresponding points being less than a certain value (e.g., 0.01 meters).

[0084] A slope-based filtering algorithm is used to remove the ground point cloud. For each point P(x, y, z) in the point cloud, its local slope within a certain neighborhood range (e.g., 5 meters × 5 meters) is calculated. Let the neighborhood point set of point P be {P j =(x j ,y j ,z j )}. First, the covariance matrix C of the neighborhood points is calculated:

[0085]

[0086] where, is the mean of the neighborhood points.

[0087] Then, the eigenvalues λ 1 ,λ 2 ,λ 3 (λ 1 ≥λ 2 ≥λ 3 ) of the covariance matrix C are calculated, and the local slope S of point P is calculated based on the eigenvalues:

[0088]

[0089] According to the terrain characteristics and experience, a slope threshold (e.g., 12°) is set. When S < 12°, point P is determined to be a ground point and removed, and non-ground points with a larger slope, i.e., the point cloud of dangerous rocks and other ground objects, are retained.

[0090] Use the Delaunay triangulation algorithm to perform terrain reconstruction on the processed non-ground point cloud. Take the points in the point cloud data as the vertices of triangles, and construct a triangular network according to the Delaunay rule to ensure that no other points are contained within the circumcircle of each triangle. During the process of constructing the triangular network, calculate the area A of each triangle t , normal vector and other geometric properties, as well as the topological relationship between adjacent triangles. For triangle ΔABC, its area calculation formula is:

[0091]

[0092] Normal vector

[0093] Form a continuous surface by connecting the sides of the triangles to generate a high-precision digital elevation model (DEM) and a three-dimensional terrain surface model. In the generated terrain model, the vertex coordinates and geometric properties of each triangle are accurately recorded, providing accurate terrain data for subsequent numerical simulations and analyses. For example, according to the terrain model, the elevation value, slope, and aspect at any position can be calculated, and this information is crucial for analyzing the movement trajectory and potential influence range of dangerous rockfalls

[0094] S4. Analyze the vibration signal spectrum, quantify the characteristics of multi-source data, and determine the fusion weights

[0095] Preprocess the collected vibration signal, and use the wavelet transform denoising method to remove the noise interference in the signal. Select the Daubechies wavelet basis function (such as 'db4'), and determine the decomposition level (such as 6 layers) according to the frequency characteristics and noise level of the vibration signal. Perform wavelet decomposition on the vibration signal s(t) to obtain the wavelet coefficients c j,k (where j represents the scale and k represents the time shift).

[0096] For the wavelet coefficients at each scale j, calculate its noise standard deviation σ j , and use the soft threshold method to process the wavelet coefficients. The soft threshold formula is:

[0097]

[0098] where (N is the signal length). Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the signal and obtain the denoised vibration signal

[0099] For the denoised signal Perform filtering processing, and use a band-pass filter to extract the effective signal within a specific frequency range. According to the frequency characteristics of the dangerous rock vibration signal, set the passband frequency range of the filter (such as 8 - 400 Hz) to remove low-frequency and high-frequency interference signals. Design a Butterworth band-pass filter, whose transfer function is (where ω c is the cut-off frequency), and convert it into a digital filter through bilinear transformation to filter the signal.

[0100] Meanwhile, perform amplification processing on the signal. According to the amplitude range of the signal and the sensitivity of the acquisition device, select an appropriate amplification factor (such as 50 times). Let the amplitude of the original signal be A, and the amplitude of the amplified signal A' = 50A, which improves the signal resolution for subsequent analysis.

[0101] Use the Fast Fourier Transform (FFT) algorithm to perform spectral analysis on the preprocessed vibration signal Convert the time-domain signal into a frequency-domain signal S(f) to obtain the spectrogram of the signal. In the spectrogram, extract the frequency f, amplitude A f , phase and other characteristic parameters of the vibration signal. Determine the main frequency f of the signal by finding the peak in the spectrogram p , that is Calculate the amplitude magnitude A corresponding to the peak fp = |S(f p )|, as well as the phase information of this frequency component

[0102] Analyze the vibration energy distribution in different frequency bands, and calculate the energy spectral density E(f). Divide the square of the amplitude in the spectrogram by the frequency resolution Δf to obtain the energy spectral density curve, that is Calculate the energy proportion in different frequency bands through integration. For example, calculate the energy proportion in the 10 - 100H z frequency band:

[0103]

[0104] According to the characteristic spectral signal range related to the dangerous rock activity (such as 10 - 100 Hz), judge whether the energy magnitude in this frequency band is abnormal. If P 10-100 is significantly higher than the normal level (such as greater than 0.5, and the normal level can be determined according to historical data or experience), it may indicate that the dangerous rock is in an unstable state or an impending collapse and rockfall event is about to occur.

[0105] For image and point cloud data, geometric shape features (such as length L, width W, height H, volume V), texture features (such as roughness R, contrast C), and spatial distribution features (such as position coordinates (x, y, z), spacing D, azimuth angle θ) of dangerous rocks are extracted. For geometric shape features, they are obtained by calculating the vector data of the dangerous rock contour. For example, the length and width of the minimum bounding rectangle are calculated as the approximate length and width of the dangerous rock, and the volume V of the dangerous rock is calculated by integration V = ∫ V dV (where dV is the volume element).

[0106] Texture features are determined by analyzing the gray-scale changes in the dangerous rock area of the image data. For example, the contrast of the gray-level co-occurrence matrix is calculated (where p(i,j) is the element of the gray-level co-occurrence matrix) and entropy

[0107] Spatial distribution features are calculated based on the coordinate information of the dangerous rocks in three-dimensional space. For example, the average spacing between dangerous rocks is calculated (where n is the number of dangerous rocks), and the azimuth angle (where (x 0 , y 0 ) are the coordinates of the reference point).

[0108] For vibration signals, frequency features (such as main frequency f p , frequency band width B), amplitude features (such as maximum amplitude A max , average amplitude A avg ), and energy features (such as total energy E t , energy density e) are extracted. The main frequency and frequency band width are obtained through spectrum analysis, the maximum amplitude and average amplitude are directly calculated from the time-domain signal, and the total energy is obtained by integrating the energy spectral density curve The energy density is the total energy divided by the signal duration (where is the signal duration).

[0109] The principal component analysis (PCA) method is used to quantify and reduce the dimension of the extracted features. First, a feature matrix X is constructed by combining the feature vectors of different data types into a matrix form, where each row represents the feature vector of a sample. The covariance matrix Cov(X) of the feature matrix is calculated, and then the eigenvalues λ i and eigenvectors of the covariance matrix are solved. The eigenvectors are sorted according to the magnitudes of the eigenvalues, and the first few eigenvectors with larger contribution rates (such as the cumulative contribution rate reaching more than 95%) are selected to form a new feature space. Let the selected eigenvectors be Then the feature data after dimension reduction is Data redundancy is reduced and data processing efficiency is improved.

[0110] Let \(n\) be the number of data types (here \(n = 3\), namely image and point cloud data, vibration signal data, other geological exploration data, etc.), and \(m\) be the number of evaluation indicators (such as data accuracy, integrity, timeliness, etc., assume \(m = 3\)). Let \(x\) ij represent the normalized value of the \(i\)-th data type under the \(j\)-th evaluation indicator.

[0111] First, calculate the information entropy \(e\) under the \(j\)-th evaluation indicator j :

[0112] where (when \(p\) ij \( = 0\), \(p\) ij \(\ln(p\) ij ) = 0).

[0113] For example, for image and point cloud data under the accuracy evaluation indicator, if its normalized value \(x\) 11 \( = 0.75\), for vibration signal data \(x\) 21 \( = 0.65\), and for other geological exploration data \(x\) 31 \( = 0.4\), then

[0114] \(p\) 11 \( = 0.75 / (0.75 + 0.65 + 0.4)=0.375\),

[0115] \(p\) 21 \( = 0.65 / (0.75 + 0.65 + 0.4)=0.325\),

[0116] \(p\) 31 \( = 0.4 / (0.75 + 0.65 + 0.4)=0.2\).

[0117] Calculate \(e\) 1 \( = -1 / \ln(3)\times(0.375\times\ln(0.375)+0.325\times\ln(0.325)+0.2\times\ln(0.2))\approx0.934\).

[0118] Then calculate the information entropy redundancy \(d\) i \( = 1 - e\) j , then \(d\) 1 \( = 1 - 0.934 = 0.066\).

[0119] Construct a judgment matrix \(A\) according to the AHP method, and determine the relative importance between different evaluation indicators through methods such as expert scoring. Assume the judgment matrix \(A=(a\) ij ), where \(a\) ij represents the importance degree of the \(i\)-th evaluation indicator relative to the \(j\)-th evaluation indicator. For example, if it is considered that accuracy is 1.5 times more important than integrity, and integrity is 1.2 times more important than timeliness, then the judgment matrix

[0120]

[0121] Calculate the maximum eigenvalue λ of the judgment matrix A max and its corresponding eigenvector ω. For the above judgment matrix A, through calculation, λ max ≈3.004, and the eigenvector

[0122] Perform consistency check. Consistency index:

[0123] CI = (λ max - n) / (n - 1) = (3.004 - 3) / (3 - 1) = 0.002.

[0124] The random consistency index RI is obtained by looking up the table according to n = 3 as 0.58. Consistency ratio:

[0125] CR = CI / RI = 0.002 / 0.58 ≈ 0.0034 < 0.1, and it is considered that the judgment matrix has satisfactory consistency.

[0126] Finally, calculate the weights ω of each data type i :

[0127] Then the weights of the image and point cloud data:

[0128] ω 1 = 0.493×0.066 + 0.327×d 2 + 0.180×d 3 ;

[0129] where d 2 and d 3 are the information entropy redundancies under the integrity and timeliness evaluation indexes respectively, and the calculation method is the same as above). Similarly, the weights ω 2 and ω 3 of the vibration signal data and other geological exploration data can be calculated.

[0130] S5. Construct a three-dimensional real scene model by fusing data, and establish a physical mechanics model by combining experiments and explorations.

[0131] Construct a weighted average fusion model. According to the calculated fusion weights, fuse data such as images, point clouds, and vibration signals. For the spatial position and geometric shape information of the dangerous rock, mainly use the point cloud data and combine the texture information in the image data for fusion. For example, for each point P(x, y, z) in the point cloud, find the corresponding pixel point in the image data according to its spatial coordinates, and assign the texture information such as the color and brightness of the pixel point to point P according to the weight assignment method. Let the weight of the image data be ω 1, the weight of the point cloud data is ω 2 (ω 1 +ω 2 = 1), the texture attribute T of the fused point P = ω 1 ×T 影像 +ω 2 ×T 点云 (where T 影像 is the texture value of the corresponding pixel point in the image, T 点云 is the initial texture value of point P in the point cloud data. If the point cloud data has no texture information, then T 点云 = 0).

[0132] For the geological attributes and mechanical state information of the dangerous rock, combine the stability characteristics obtained from the vibration signal analysis with the geological exploration data.

[0133] For example, judge the stability state of the dangerous rock according to the spectral characteristics of the vibration signal. If the main frequency is within a specific range and the energy is high, combined with the lithology and joint fissure conditions in this area during geological exploration, assign a higher unstable risk attribute value to the dangerous rock area in the three-dimensional real-scene model. At the same time, integrate other geological exploration data (such as stratigraphic lithology distribution, geotechnical physical and mechanical parameters, etc.) into the model, determine the geotechnical types of different regions according to the stratigraphic lithology data, and then assign corresponding mechanical parameters to the corresponding regions in the model, such as elastic modulus E, Poisson's ratio ν, internal friction angle etc.

[0134] Suppose the elastic modulus determined for a certain area of geotechnical body during geological exploration is E 0 , the elastic modulus E adjusted according to the fusion weight = ω 3 ×E 0 (where ω 3 is the weight of the geological exploration data in the elastic modulus fusion).

[0135] Through the above fusion process, construct a comprehensive three-dimensional real-scene model including the spatial form, geological attributes, mechanical state and dynamic change trend of the dangerous rock, ensuring that the model has high resolution (spatial resolution not less than 0.1 m) and high precision, and can accurately reflect the actual situation of the dangerous rock.

[0136] Conduct indoor rock mechanics tests, select representative rock samples, and conduct uniaxial compressive strength tests, tensile strength tests, shear strength tests, elastic modulus tests, Poisson's ratio tests, etc. according to the test methods recommended by the International Society for Rock Mechanics (ISRM). For example, in the uniaxial compressive strength test, place the rock sample on a press and apply axial pressure at a constant loading rate (such as 0.5 MPa / s) until the rock fails, and record the maximum pressure at failure as the uniaxial compressive strength σ c . Determine the basic mechanical parameters of the rock through these tests, such as the compressive strength σc , tensile strength σ t , internal friction angle Cohesion c, elastic modulus E, Poisson's ratio ν, etc.

[0137] Combined with on-site geological exploration data, the development degree, distribution law, and filling conditions of rock joints and fissures are analyzed in detail. Through geological mapping and equipment such as joint fissure measuring instruments, the geometric parameters of joint fissures (such as length l, width w, spacing d, dip angle α) are measured. For the mechanical parameters of joint fissures (such as normal stiffness k n , tangential stiffness k s , tensile strength σ tj , internal friction angle ), empirical formulas or in-situ tests (such as joint shear tests) can be used to determine them.

[0138] For example, according to the empirical formula Calculate the normal stiffness (where E is the elastic modulus of the rock, ν is Poisson's ratio, d is the joint spacing, and l is the joint length).

[0139] According to the mechanical properties of the rock and the characteristics of joint fissures, select a suitable rock constitutive relation model, such as the Mohr-Coulomb criterion, Drucker-Prager criterion, Hoek-Brown criterion, etc. Taking the Mohr-Coulomb criterion as an example, its yield function is (where τ is the shear stress, c is the cohesion, σ is the normal stress, is the internal friction angle), and according to the c and values determined by experiments, implement this constitutive relation in the numerical model.

[0140] Obtain the initial in-situ stress state of the area to be predicted through on-site in-situ stress measurement (such as using the hydraulic fracturing method and stress relief method), including the vertical stress σ υ , horizontal stress σ h magnitude and direction.

[0141] Take the measurement results as boundary conditions, and use finite element analysis software (such as ABAQUS, ANSYS) to establish a three-dimensional numerical model of the dangerous rock and the surrounding rock mass. Accurately depict the geometric shape, spatial position, and joint fissure distribution of the dangerous rock in the model, assign the rock mechanical parameters and the initial stress state to the corresponding elements and nodes, and construct a complete physical and mechanical model.

[0142] S6. Build a machine learning model based on historical data and cooperate with the physical model to predict the trend of falling rocks.

[0143] Collect a large amount of data on historical rockfall events, meteorological data, seismic activity data, and records of human engineering activities, etc., to establish a comprehensive database. The historical rockfall event data includes detailed information such as the occurrence time, location, scale, and movement trajectory of the rockfall; the meteorological data covers parameters such as rainfall intensity, wind speed, and temperature; the seismic activity data records the magnitude, epicenter location, and seismic wave propagation characteristics of the earthquake; and the records of human engineering activities contain the time, scope, and intensity of activities such as nearby road construction, mining operations, and building construction. Clean and preprocess the collected data to remove outliers and incorrect data to ensure the accuracy and integrity of the data.

[0144] Select the support vector machine (SVM) algorithm to construct a motion trend prediction model. Divide the data in the database into a training set, a validation set, and a test set in a ratio of 7:2:1. For the training set data, extract feature vectors related to the movement of dangerous rocks, such as the cumulative amount of previous rainfall, the peak acceleration of ground motion, the intensity of surrounding human engineering activities, and the geometric characteristics and geological properties of the dangerous rocks themselves. Use the SVM algorithm to train the training set, and optimize the model performance by adjusting hyperparameters such as the kernel function (such as the radial basis function) parameters and the penalty factor. Adopt a cross-validation method, such as 5-fold cross-validation, to evaluate indicators such as the accuracy, recall rate, and F1 value of the model on the validation set, and adjust the hyperparameters according to the evaluation results until the model reaches better performance indicators on the validation set, such as an accuracy rate of over 90%.

[0145] Establish a parameter transfer interface between the physical mechanics model and the data-driven model. The physical mechanics model transfers the stress, strain, displacement and other state information of the dangerous rock during the simulation process to the data-driven model in real time, and the data-driven model feeds back the rockfall movement trend information predicted based on historical data and machine learning to the physical mechanics model. For example, after each calculation time step of the physical mechanics model, the stress value σ ij , strain value ε ij and displacement value u i (where i, j represent the spatial coordinate directions) of the key parts of the dangerous rock are combined to form a feature vector and transferred to the data-driven model. The data-driven model generates correction parameters for the rockfall movement trend, such as the predicted correction amount of the rockfall velocity Δυ and the adjustment angle of the movement direction Δθ, etc., according to the received information and its own prediction model, and transfers these parameters back to the physical mechanics model.

[0146] Design a dynamic optimization algorithm to achieve collaborative computing of two models. An optimization algorithm with an adaptive step size is adopted to dynamically adjust the calculation step size according to the error change during the model calculation process. Let the current calculation step size be h and the calculation error be e. If the error e continuously increases within several consecutive calculation steps, the step size h is reduced, i.e., h = h × α (where α is a contraction factor less than 1, such as 0.8); if the error continuously decreases, the step size h is appropriately increased, i.e., h = h × β (where β is an expansion factor greater than 1, such as 1.2). During the calculation process, the influences of various complex factors are considered simultaneously, such as the spatio-temporal non-uniformity of ground motion, the seepage-stress coupling effect of rainfall infiltration, the dynamic time-varying action of wind force, and the disturbance effect of human engineering activities.

[0147] Regarding the influence of the spatio-temporal non-uniformity of ground motion, according to the seismic wave propagation theory, the seismic acceleration time history a(t, x, y, z) received by unstable rocks at different positions is calculated, where t is time and (x, y, z) are spatial coordinates. It is used as an external load and applied to the physical mechanics model, and is also passed as a characteristic parameter to the data-driven model.

[0148] The seepage-stress coupling effect of rainfall infiltration is calculated by establishing a coupling model of the seepage equation and the stress balance equation. Considering parameters such as rainfall intensity I(t) and the permeability coefficient k of the rock and soil mass, the change in pore water pressure u p (x, y, z, t) caused by rainfall and its influence on the stability of unstable rocks are calculated, and the relevant parameters are passed to the data-driven model for comprehensive analysis.

[0149] The dynamic time-varying action of wind force is calculated based on the wind speed time history υ w (t) and wind direction θ w (t) in meteorological data, and the wind load F w (t) acting on the surface of the unstable rock is calculated and incorporated into the physical mechanics model for calculation, and is also passed as characteristic information to the data-driven model.

[0150] The disturbance effect of human engineering activities is calculated according to engineering activity records. For example, when blasting operations are carried out nearby, the blasting vibration load F b (t) and the degree of damage to the rock and soil mass structure are calculated and reflected in the parameter adjustment of the physical mechanics model and the data-driven model.

[0151] Through this collaborative calculation, the movement trajectory and state of the falling rocks are accurately predicted, such as predicting the position coordinates (x t , y t , z t ) of the falling rocks at different times, the velocity vector acceleration vector and the rotational angular velocity ω t and so on.

[0152] S7. Determine the two-dimensional cross-section based on three-dimensional simulation, and dynamically simulate and calculate the movement of falling rocks and the response of rock and soil masses.

[0153] Determine the key cross-sections for two-dimensional simulation based on the three-dimensional simulation results and topographic features. Analyze the concentrated areas of the falling rock movement trajectories, areas with significant topographic changes (such as steep slopes, cliffs, valleys, etc.) and possible locations for protective structures in the three-dimensional simulation, and select representative cross-sections. For example, according to the probability density distribution function P(x, y, z) of the falling rock movement trajectories, determine the areas with a higher frequency of falling rocks, and select cross-sections perpendicular to the main movement direction within these areas. At the same time, combined with the topographic contour map, select positions with large topographic undulations and obvious slope changes as key cross-sections to ensure that the movement characteristics of dangerous rock falls under different topographic conditions can be comprehensively reflected.

[0154] Obtain the topographic and rock and soil mass parameters of the cross-section. Using the three-dimensional topographic model and geological exploration data, extract the topographic profile line at the key cross-section, and determine topographic parameters such as elevation h(x), slope α(x), and aspect β(x) (where x is the horizontal coordinate on the cross-section). Through on-site rock and soil mass tests and geological data analysis, determine the thickness d of different rock and soil mass layers on the cross-section i and physical and mechanical parameters (such as elastic modulus E i and Poisson's ratio ν i and internal friction angle cohesion c i ).

[0155] In the two-dimensional numerical model, accurately depict these topographic and rock and soil mass elements, use the topographic profile line as the boundary of the model, divide different material regions according to the rock and soil mass stratification situation, and assign corresponding mechanical parameters to each region. For example, use the geometric modeling function in finite element software to draw the geometric shape of the cross-section, and then set the corresponding material properties for different elements to construct a two-dimensional numerical model.

[0156] According to real-time monitoring and exploration information, use the real-time data feedback and on-site information update mechanism to dynamically adjust the model boundary conditions and parameters. Real-time monitoring data includes rainfall intensity I(t), ground motion acceleration a(t), and monitored displacement u of dangerous rocks m (t), etc.

[0157] When the rainfall intensity changes, calculate the change in rainfall infiltration depth h p (t) and pore water pressure u p (t) according to the infiltration theory, and adjust parameters such as the permeability coefficient k(t) and saturation S(t) of the rock and soil mass in the model. For example, use the empirical formula k(t) = k 0 ×(1 - S(t)) n (where k 0(where \(k_0\) is the initial permeability coefficient and \(n\) is the empirical exponent) to calculate the change in permeability coefficient.

[0158] During an earthquake, based on the monitored ground motion acceleration \(a(t)\), it is applied as a dynamic load on the model boundary, and the damping coefficient \(c(t)\) of the rock and soil mass in the model is adjusted d to reflect the dissipation characteristics of seismic energy. According to the monitored displacement value \(u(t)\) of the dangerous rock m , it is compared with the displacement value calculated by the model. If the error exceeds a certain threshold, the rock mechanics parameters in the model, such as the elastic modulus \(E\) and Poisson's ratio \(\nu\), are adjusted, and the back-analysis method is used to determine more realistic parameter values.

[0159] The finite element-discrete element coupling method is used for simulation calculation. In the two-dimensional numerical model, the rock and soil mass area is discretized using the finite element method, and the falling rocks are regarded as discrete element particles. Finite element elements are used to simulate the deformation and stress distribution of the rock and soil mass, and discrete element particles are used to simulate the movement and collision behavior of the falling rocks. During the calculation process, the contact relationship between the falling rocks and the rock and soil mass is considered, and contact parameters such as contact stiffness \(k\) c , friction coefficient \(\mu\), damping coefficient \(\xi\), etc. are set. For example, when a falling rock collides with the rock and soil mass, the contact force \(F\) is calculated according to the contact mechanics theory c , and the penalty function method or Lagrange multiplier method is used to handle the contact constraints to calculate the change in the motion state of the falling rock. By solving the equilibrium equation of the finite element and the motion equation of the discrete element, the contact force \(F(t)\) cp , frictional force \(F(t)\) f , collision impact force \(F(t)\) i between the falling rock and the slope are calculated in real time, and the bounce height \(h(t)\) b , change angle \(\theta(t)\) of the motion direction d , kinetic energy loss \(\Delta E(t)\) k and other motion characteristic data such as the final stopping position \((x\) f , \(y\) f ) are accurately obtained. At the same time, the disturbance influence range \(A(t)\) d and degree \(D(t)\) d of the falling rock movement on the surrounding rock and soil mass are analyzed, such as the change conditions of the displacement field \(u(x,y,t)\) and stress field \(\sigma(x,y,t)\) of the rock and soil mass.

[0160] S8, construct an AR / VR platform for visualization and interactively verify and optimize the protection plan to improve the model algorithm.

[0161] First, configure high-performance hardware, including a graphics workstation (such as one equipped with a multi-core processor, high-end graphics card, and large-capacity memory) to meet the requirements of large-scale data processing and complex scene rendering, a head-mounted display device (such as HTC Vive, Oculus Rift) to provide an immersive visual experience, and interaction devices (such as gamepads, data gloves) to enable users to interact with the virtual scene.

[0162] Develop a software platform based on the Unity 3D engine, import simulation result data (including data such as the shape, movement trajectory, and mechanical parameters of unstable rocks obtained from three-dimensional and two-dimensional numerical simulations) to construct a virtual scene. In Unity 3D, utilize its powerful three-dimensional modeling and rendering capabilities to create virtual objects such as unstable rocks, terrain, and protective structures, and assign corresponding materials and textures to them to make them more realistic.

[0163] Develop a scene management module to implement operations such as loading, switching, and scaling of the virtual scene. For example, control the position and perspective of the camera by writing C# scripts to enable users to freely roam and observe in the virtual scene. Develop a data interaction module to ensure that the simulation result data can be accurately displayed and updated in the virtual scene. For example, convert the movement trajectory data of falling rocks into displacement information of virtual objects in real time to present the dynamic movement process of falling rocks in the scene.

[0164] Utilize interaction devices to enable users to operate on the virtual scene. Through gamepads or data gloves, users can simulate different protective measure settings in the virtual scene, such as adjusting the position, height, and strength of the protective net, and setting the position and shape of the rockfall retaining wall. At the same time, users can simulate changes in external load conditions, such as changing the rainfall intensity, ground motion parameters, and wind force, and observe the change responses of the movement characteristics of falling rocks in real time.

[0165] For example, when the user adjusts the height of the protective net, the system calculates the collision result between the falling rock and the protective net in real time according to the mechanical model of the protective structure and the movement model of the falling rock, and displays whether the falling rock is successfully intercepted and the force condition of the protective net. Develop an evaluation index system, including protection effect indexes (such as the rockfall interception rate R i , the reduction rate of rockfall kinetic energy R e ), economic indexes (such as the cost C of the protective structure), and environmental impact indexes (such as the degree of damage D to the surrounding landscape). Calculate the comprehensive evaluation index value according to these indexes and the corresponding weights ω i , ω e , ω c , ω d (the weights can be determined by the analytic hierarchy process or expert scoring), and compare the advantages and disadvantages between different schemes. (where f i is the normalized function value of each index), and compare the advantages and disadvantages between different schemes.

[0166] Analyze the user feedback data and operation records during the visualization and interactive verification process, extract valuable information, and use it to further improve the numerical simulation model and algorithm. For example, if the user finds that a certain protection scheme is ineffective under specific terrain and rockfall conditions during the interaction process, analyze the reasons. It may be that the collision model between the rockfall and the protection structure in the model is not accurate enough, or there are deviations in the simulation of the mechanical behavior of rock and soil masses under complex terrain.

[0167] According to the analysis results, improve the collision algorithm in the physical mechanics model, such as adopting a more accurate contact mechanics model; optimize the constitutive relationship of rock and soil masses, such as considering more non-linear factors. At the same time, adjust the feature extraction and prediction algorithms in the data-driven model, such as adding feature parameters related to the protection structure to improve the accuracy of the model's prediction of the protection scheme effect. Through continuous interactive verification and model improvement, improve the accuracy and reliability of motion feature prediction, and provide more intuitive and operable decision-making support for actual engineering protection design.

Claims

1. A method for predicting the motion characteristics of dangerous rockfalls based on multi-source data fusion, characterized in that: The steps include: S1. Collect multi-source data of dangerous rocks to build a basic information database; S2, based on CNN algorithm, process multi-source data of dangerous rocks, identify the contours of dangerous rocks and calculate geometric features; S3, using a variety of algorithms to process dangerous rock multi-source data, reconstruct terrain to generate digital elevation and terrain models; S4, analyze the vibration signal spectrum, quantify the characteristics of dangerous rock multi-source data and determine the fusion weight; S5, based on the fusion weight, the dangerous rock multi-source data is integrated to construct a three-dimensional real scene model, and a physical mechanics model is constructed by combining indoor rock mechanics tests and on-site geological surveys; S6. Establish a historical database and select a machine learning algorithm to build a data-driven model. Use an adaptive step-size optimization algorithm to achieve collaborative calculations between the physical mechanics model and the data-driven model to predict the trajectory and state of rockfall. S7. Determine the key sections of the two-dimensional simulation based on the simulation results of the three-dimensional real-scene model and the digital elevation and terrain model, establish a two-dimensional numerical model and accurately characterize the relevant elements, and dynamically simulate and calculate the rockfall movement and rock and soil response.

2. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1 is characterized in that: The multi-source data of dangerous rocks in the basic information database are collected according to the topography, area and distribution of dangerous rocks in the area to be predicted. Unmanned aerial vehicles are used to fly along preset routes to collect data. The dual-light camera takes spectral images at set time intervals during the flight. The lidar scanner emits laser pulses and receives reflected signals to obtain point cloud data. The high-definition camera records video data throughout the process to collect multi-source data of dangerous rocks. According to the geological structure and the distribution of dangerous rocks, seismic detectors and microseismic sensors are arranged in a triangular array at key positions around and inside the area to be predicted to obtain vibration signals.

3. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1 is characterized in that: The S2 specifically includes the following steps: S21. Using the VGGNet architecture as the basis, adding an attention mechanism to the network structure, and building a deep learning model based on convolutional neural networks; S22, using the dangerous rock multi-source data as training samples and annotating them in detail, wherein the annotation contents include the precise outline, category and stability status of the dangerous rock; S23, divide the labeled samples into training set, validation set and test set in a ratio of 7:2:1, use the training set to train the CNN model, use the stochastic gradient descent optimization algorithm during the training process, adjust the learning rate according to the accuracy of the validation set after each set training round, and when the accuracy no longer improves, reduce the learning rate to 0.1 times the current one for further training until the accuracy of the deep learning model on the validation set meets the specified requirements; S24, inputting the spectral image and video data into the trained deep learning model and performing feature judgment based on the pixel points, and determining the contour of the dangerous rock through an adaptive threshold segmentation method; S25. Convert the determined dangerous rock contour into a uniform vector format and calculate its geometric features.

4. The method for predicting the motion characteristics of dangerous rocks and falling rocks based on multi-source data fusion according to claim 3 is characterized in that: When calculating the geometric features, the coordinate information of each point on the dangerous rock contour is first recorded, and then these points are connected in sequence to form a polygon, wherein the geometric features include area, perimeter and center of gravity coordinates.

5. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 4 is characterized in that: The calculation area is specifically expressed as: Where n is the number of contour points. When i=n, ​​x n+1 =x1,y n+1 =y1;x i is the horizontal coordinate in the coordinate information, y i is the ordinate in the coordinate information; The calculation of the perimeter is specifically expressed as: The specific expression of the ordinate of the center of gravity coordinate is:

6. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1 is characterized in that: The S3 specifically includes: using the ICP algorithm to register the point cloud data, removing the ground point cloud by using a slope-based filtering algorithm, and then using the Delaunay triangulation algorithm to reconstruct the terrain of the processed non-ground point cloud to generate a digital elevation and terrain model.

7. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1 is characterized in that: The S4 specifically includes: preprocessing based on the vibration signal, using the wavelet transform denoising method to remove noise interference, and then filtering. At the same time, amplifying the effective signal, using the fast Fourier transform algorithm to perform spectrum analysis, identifying spectrum signals related to dangerous rock activities, analyzing the vibration energy distribution in different frequency bands, calculating the energy spectrum density and judging whether the energy size in the frequency band is abnormal based on the range of characteristic spectrum signals related to dangerous rock activities; for multi-source data of dangerous rocks, extracting multi-source data features, including geometric shape features, texture features, spatial distribution features, amplitude features, energy features and frequency features, using the principal component analysis method to quantify and reduce the dimensionality of the extracted features, and then determining the fusion weight based on the method combining information entropy and hierarchical analysis.

8. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1 is characterized in that: The S5 specifically includes the following steps: fusing multi-source dangerous rock data based on the fusion weights to obtain comprehensive dangerous rock information, and then constructing a three-dimensional real-scene model, conducting indoor rock mechanics tests to determine rock mechanics parameters, determining joint and fissure characteristics in combination with on-site geological survey data, selecting a suitable constitutive relationship model, obtaining the initial ground stress state of the area to be predicted through on-site ground stress measurement, using the measurement results as boundary conditions, and constructing a three-dimensional numerical model using finite element analysis software.

9. The method for predicting the motion characteristics of dangerous rocks and falling rocks based on multi-source data fusion according to claim 1, characterized in that: The S6 specifically includes the following steps: collecting historical collapse and rockfall event data, meteorological data, seismic activity data, and records of human engineering activities, and performing data cleaning and preprocessing, thereby establishing a historical database, selecting a machine learning algorithm to construct a data-driven model, including a motion trend prediction model, dividing the data set training model based on the historical database, optimizing parameters through training results, establishing a parameter transfer interface between the physical and mechanical model and the motion trend prediction model, using an adaptive step size optimization algorithm to achieve collaborative computing, considering the influence of multiple complex factors, and predicting the trajectory and state of rockfall.

10. The method for predicting the motion characteristics of dangerous rocks and rockfalls based on multi-source data fusion according to claim 1, characterized in that: The S7 specifically includes: determining the two-dimensional simulation key section based on the simulation results of the three-dimensional real-scene model and the digital elevation and terrain model; at the same time, combining the terrain contour map, selecting the key section and obtaining its terrain parameters and rock and soil parameters; establishing a two-dimensional numerical model and characterizing related elements; using the dangerous rock multi-source data feedback and on-site information update mechanism to adjust the boundary conditions, terrain parameters and rock and soil parameters; and using the finite element-discrete element coupling method to simulate and calculate the rockfall movement and rock and soil response.

Images