A geotechnical engineering slope deformation monitoring method and system

Abstract

The application discloses a geotechnical engineering slope deformation monitoring method and system, through acquiring multi-modal image data, utilizing a CLAHE algorithm to enhance the contrast of weak deformation areas of the multi-modal image data, extracting gradient edge information through a Sobel operator, and filtering laser radar point cloud noise based on VMD variational mode decomposition; based on a BiLSTM bidirectional long short-term memory network, bidirectional dependence of a time sequence is captured, and weights of different time steps and spatial positions are dynamically allocated through an attention mechanism; the hyperparameters of the mixed prediction model are optimized by using a GJO golden jackal optimization algorithm to obtain a target mixed prediction model; feature image data is input into the target mixed prediction model for prediction to generate a risk probability heat map, and a potential sliding surface position of a slope is located according to the risk probability heat map. The reliability and accuracy of geotechnical engineering slope image data analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geotechnical engineering safety monitoring, in particular to a geotechnical engineering slope deformation monitoring method and system. BACKGROUND

[0002] Traditional slope deformation monitoring methods have many limitations. Manual inspection is inefficient, labor-intensive, and difficult to achieve real-time and comprehensive monitoring of the slope, often only remedying after disasters occur; single-point monitoring instruments have limited monitoring range and are difficult to reflect the overall deformation trend of the slope, easily missing potential dangerous areas; existing automated monitoring technologies have improved monitoring efficiency to some extent, but have low data processing and analysis efficiency and accuracy, weak multi-source data fusion capability, and are difficult to accurately capture the complex characteristics of slope deformation, resulting in low prediction accuracy and inability to timely and effectively warn of slope disasters. SUMMARY

[0003] The present application relates to the technical field of geotechnical engineering safety monitoring, in particular to a geotechnical engineering slope deformation monitoring method and system.

[0004] To achieve the above-mentioned purpose, the technical solution of the present application is as follows: further, in the above-mentioned geotechnical engineering slope deformation monitoring method, the geotechnical engineering slope deformation monitoring method comprises the following steps:

[0005] Obtain multi-modal image data, enhance the contrast of weak deformation areas of the multi-modal image data using the CLAHE algorithm, extract gradient edge information through the Sobel operator, and filter out laser radar point cloud noise based on VMD variational mode decomposition to obtain feature image data;

[0006] Capture the bidirectional dependency of time series based on the BiLSTM bidirectional long short-term memory network, dynamically allocate weights for different time steps and spatial positions through the Attention attention mechanism, and establish a hybrid prediction model;

[0007] Optimize the hyperparameters of the hybrid prediction model using the GJO golden jackal optimization algorithm to obtain a target hybrid prediction model;

[0008] Input the feature image data into the target hybrid prediction model for prediction to generate a risk probability heat map, and locate the potential sliding surface position of the slope according to the risk probability heat map.

[0009] Further, in the above-mentioned geotechnical engineering slope deformation monitoring method, the multi-modal image data is obtained, the contrast of weak deformation areas of the multi-modal image data is enhanced using the CLAHE algorithm, gradient edge information is extracted through the Sobel operator, and laser radar point cloud noise is filtered out based on VMD variational mode decomposition to obtain feature image data, comprising:

[0010] acquire unmanned aerial vehicle aerial image, ground camera video, laser radar point cloud and GNSS displacement data of geotechnical engineering slope, and construct multi-modal image data;

[0011] After splicing and correcting the multi-modal image data, the format is unified to obtain initial multi-modal image data;

[0012] The initial multi-modal image data is divided into a plurality of sub-blocks according to the size of the parameters set by using the CLAHE algorithm, and the divided block image data is obtained;

[0013] The divided block image data is subjected to histogram equalization processing respectively, the histogram is calculated by counting the gray value distribution of the pixels in the sub-block, and the equalization image data is obtained;

[0014] The equalization image data is spliced into a complete image again, and the boundaries between the sub-blocks are processed in a smooth transition manner to obtain enhanced image data.

[0015] Further, in the above-mentioned geotechnical engineering slope deformation monitoring method, the multi-modal image data is acquired, the contrast of the weak deformation area of the multi-modal image data is enhanced by using the CLAHE algorithm, the gradient edge information is extracted by using the Sobel operator, and the VMD variational modal decomposition is used to filter out the noise of the laser radar point cloud to obtain feature image data, and further comprising:

[0016] The enhanced image data is subjected to gray scale processing, and the Sobel convolution kernel in the horizontal direction and the vertical direction is applied to the gray scale image respectively;

[0017] The gradient values in the horizontal direction and the vertical direction are synthesized to obtain the gradient edge information of the image, and the threshold value processing is performed on the synthesized gradient values;

[0018] The VMD variational modal decomposition is used to filter out the noise of the laser radar point cloud, the converted one-dimensional signal is input into the VMD algorithm, the signal is decomposed into K modal components through the iterative optimization process, and the feature image data is obtained.

[0019] Further, in the above-mentioned geotechnical engineering slope deformation monitoring method, the BiLSTM bidirectional long short-term memory network is used to capture the bidirectional dependence of the time sequence, the Attention attention mechanism is used to dynamically allocate the weight of different time steps and spatial positions, and a hybrid prediction model is established, comprising:

[0020] The BiLSTM network is constructed, including an input layer, a hidden layer and an output layer, the input layer receives the processed feature image data, and the data dimension is determined according to the size and dimension of the feature image;

[0021] The hidden layer is provided with a plurality of neurons, and the input data is processed by the forward LSTM and the backward LSTM respectively;

[0022] The forward LSTM processes the data of each time step in turn from the starting point of the time sequence, and the backward LSTM processes the data of each time step in reverse from the end point of the time sequence; the outputs of the forward and backward LSTMs are spliced to serve as the output of the hidden layer.

[0023] Further, in the above-mentioned rock-soil engineering slope deformation monitoring method, the BiLSTM bidirectional long short-term memory network captures the bidirectional dependency of the time sequence, dynamically allocates the weights of different time steps and spatial positions through the Attention attention mechanism, and establishes a hybrid prediction model, and further comprises:

[0024] The output of the BiLSTM network is taken as input and mapped to a low-dimensional space through a fully connected layer; the score value of each position is calculated and converted to a weight value through a softmax function;

[0025] The output of each position is multiplied by the corresponding weight value, and weighted summation is performed to obtain the output of the Attention mechanism.

[0026] Further, in the above-mentioned rock-soil engineering slope deformation monitoring method, the GJO cheetah optimization algorithm is used to optimize the hyperparameters of the hybrid prediction model to obtain a target hybrid prediction model, comprising:

[0027] The prediction error of the model on the validation set is taken as the objective function, and the GJO algorithm is used to search and optimize the hyperparameters;

[0028] In each iteration, the cheetah individual moves according to its own position and group information, updates its position, calculates the objective function value, and finds the optimal hyperparameter combination after multiple iterations.

[0029] Further, in the above-mentioned rock-soil engineering slope deformation monitoring method, the feature image data is input into the target hybrid prediction model for prediction to generate a risk probability heat map, and the potential sliding surface position of the slope is located according to the risk probability heat map, comprising:

[0030] The risk probability value of each position predicted by the model is mapped to a color space to generate a risk probability heat map, and the potential sliding surface is determined in combination with the geological structure and topographic and geomorphic information of the slope.

[0031] Further, in a rock-soil engineering slope deformation monitoring system, the rock-soil engineering slope deformation monitoring system comprises the following modules:

[0032] The slope image acquisition module is configured to acquire multi-modal image data, enhance the contrast of a weak deformation area of the multi-modal image data by using a CLAHE algorithm, extract gradient edge information by using a Sobel operator, and remove noise of a laser radar point cloud based on VMD variational modal decomposition to obtain feature image data.

[0033] The prediction model establishment module is configured to capture bidirectional dependency of a time sequence based on a BiLSTM bidirectional long short-term memory network, dynamically allocate weights of different time steps and spatial positions by using an Attention attention mechanism, and establish a hybrid prediction model.

[0034] The model parameter optimization module is configured to optimize hyperparameters of the hybrid prediction model by using a GJO golden jackal optimization algorithm to obtain a target hybrid prediction model.

[0035] The slope risk identification module is configured to input the feature image data into the target hybrid prediction model for prediction, generate a risk probability heat map, and locate a potential sliding surface position of a slope according to the risk probability heat map.

[0036] Further, in the system for implementing the above-described geotechnical engineering slope deformation monitoring method, the model parameter optimization module comprises the following sub-modules:

[0037] The optimization sub-module is configured to take a prediction error of a model on a verification set as a target function, and search and optimize hyperparameters by using a GJO algorithm.

[0038] The obtaining sub-module is configured to, in each iteration, move a golden jackal individual according to a position of the golden jackal individual and group information, update the position, calculate a target function value, and find an optimal hyperparameter combination through multiple rounds of iteration.

[0039] Further, in the system for implementing the above-described geotechnical engineering slope deformation monitoring method, the slope risk identification module comprises the following sub-modules:

[0040] The generating sub-module is configured to map a risk probability value of each position obtained by model prediction to a color space, generate a risk probability heat map, and determine a potential sliding surface in combination with geological structure and topographic and geomorphic information of a slope.

[0041] The beneficial effects are that 1. the characteristics information of the slope can be comprehensively captured from different angles and different dimensions. Compared with a single data source, multi-modal data fusion can complement and verify each other, reduce the uncertainty of the data, and improve the reliability and accuracy of the image data analysis of the geotechnical engineering slope. 2. The quality of the point cloud data can be improved, the interference of noise on subsequent analysis can be reduced, and the quality of the geotechnical engineering slope characteristic image data is significantly improved. 3. The key information in the image data which has a greater impact on deformation prediction can be focused on, and the sensitivity to important features in the image recognition process is improved. 4. The risk probability of different positions of the slope can be intuitively displayed, and the position of the potential sliding surface of the slope can be accurately located according to the heat map. BRIEF DESCRIPTION OF DRAWINGS

[0042] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the present application.

[0043] Figure 1 FIG. 1 is a first embodiment schematic diagram of a geotechnical engineering slope deformation monitoring method according to the present application;

[0044] Figure 2 FIG. 2 is a second embodiment schematic diagram of a geotechnical engineering slope deformation monitoring method according to the present application;

[0045] Figure 3 FIG. 3 is a first embodiment schematic diagram of a geotechnical engineering slope deformation monitoring system according to the present application. DETAILED DESCRIPTION

[0046] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application.

[0047] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the use of the term "comprise" in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0048] The present application will be specifically described below with reference to the accompanying drawings, such as Figure 1 A geotechnical engineering slope deformation monitoring method, the geotechnical engineering slope deformation monitoring method comprising the following steps:

[0049] Step 101, acquire multi-modal image data, enhance the contrast of the weak deformation area of the multi-modal image data using the CLAHE algorithm, extract gradient edge information through the Sobel operator, and filter out laser radar point cloud noise based on VMD variational mode decomposition to obtain feature image data;

[0050] Specifically, in the present embodiment, the unmanned aerial vehicle aerial image, ground camera video, laser radar point cloud and GNSS displacement data of the geotechnical engineering slope are acquired to construct multi-modal image data.

[0051] After the multi-modal image data is spliced and corrected, the format is unified to obtain initial multi-modal image data.

[0052] The initial multi-modal image data is divided into multiple sub-blocks according to the parameter setting size using the CLAHE algorithm to obtain block image data.

[0053] The block image data is subjected to histogram equalization processing respectively, the histogram is calculated by counting the gray value distribution of the pixels in the sub-block, and the equalization image data is obtained.

[0054] The equalization image data is spliced into a complete image again, and the boundaries between the sub-blocks are processed in a smooth transition manner to obtain enhanced image data.

[0055] The enhanced image data is subjected to grayscale processing, and the Sobel convolution kernel in the horizontal direction and the vertical direction is applied to the grayscale image respectively.

[0056] The horizontal and vertical gradient values are synthesized to obtain the gradient edge information of the image, and the synthesized gradient values are subjected to threshold processing.

[0057] The laser radar point cloud noise is filtered out based on VMD variational mode decomposition, the converted one-dimensional signal is input into the VMD algorithm, the signal is decomposed into K modal components through an iterative optimization process, and feature image data is obtained.

[0058] Specifically;

[0059] 1. Multi-modal image data acquisition;

[0060] 1.2.1 Unmanned aerial vehicle aerial image acquisition;

[0061] A unmanned aerial vehicle with a high-resolution camera is selected, and the pixel is not less than 20,000 to ensure image clarity.

[0062] According to the range, terrain and monitoring accuracy requirements of the slope, the flight route is planned. For complex terrain slopes, a multi-route intersection shooting method is used to ensure that there is no monitoring dead angle. The flight route can be generated through professional unmanned aerial vehicle route planning software.

[0063] The flight height is set according to the size of the slope and the monitoring accuracy, generally between 2-4 meters. For larger slopes, the flight height can be appropriately increased, but the image resolution must meet the requirements; for areas with high accuracy requirements, the flight height should be reduced.

[0064] The shooting frequency is set to 3 frames per second to ensure that continuous image sequences are obtained. At the same time, the GPS positioning function of the UAV is turned on during shooting to record the position information of each image.

[0065] Choose a sunny and well-lit period for shooting to avoid working in rainy, foggy or excessively bright / dim conditions to reduce the impact of external factors on image quality. If sudden weather changes occur, shooting should be paused and restarted when the weather improves.

[0066] Possible problems and solutions: If the image is blurred during shooting, it may be due to the shaking of the UAV or inaccurate focusing. The solution is to check if the anti-shake function of the UAV is turned on and to regularly calibrate the camera focusing system.

[0067] 1.2.2 Ground camera video acquisition;

[0068] Choose a ground camera with high-definition shooting capabilities, with a resolution of no less than 2x2 to ensure the clarity of the video image.

[0069] According to the shape of the slope and the monitoring focus, choose the appropriate installation position. Generally, it is installed on both sides or the top of the slope where the view is open to ensure that the main deformation area of the slope can be clearly shot. The installation angle is about 2 degrees between the lens axis and the slope surface to avoid dead angles.

[0070] The video shooting frame rate is set to 3 frames per second to ensure that the dynamic deformation process of the slope can be captured. At the same time, the night vision function of the camera is turned on to meet the night monitoring needs.

[0071] Regularly maintain and calibrate the camera to check if the lens is clean and the angle is offset to ensure normal operation of the equipment.

[0072] 1.2.3 Laser radar point cloud acquisition;

[0073] Choose a high-precision laser radar device with a ranging accuracy of no less than 1 centimeter to ensure that the acquired point cloud data has high accuracy.

[0074] According to the device operation specification, set up measurement control points around the slope, and use total station and other equipment to accurately measure the control points to obtain their three-dimensional coordinates.

[0075] Laser radar equipment is set up in a suitable position to scan the slope. During the scanning process, ensure the stability of the equipment to avoid vibration affecting data quality. The scanning range should cover the entire slope area, and the scanning density should be set according to the monitoring accuracy requirements, generally not less than 7 points per square meter.

[0076] 4 GNSS displacement data acquisition:

[0077] Select high-precision GNSS receivers with positioning accuracy not less than 3 mm to ensure that the obtained displacement data is accurate and reliable.

[0078] GNSS monitoring points are laid out on the slope, and the monitoring points should be selected at key positions of the slope, such as the top of the slope, the foot of the slope, and the vicinity of the potential sliding surface. The number of monitoring points is determined according to the size and complexity of the slope, generally not less than 2.

[0079] According to the GNSS measurement specification, the observation time is set according to the monitoring accuracy requirements, generally not less than 1 minute each time. During the observation process, ensure the stability of the receiver to avoid external interference.

[0080] Periodically calibrate the GNSS monitoring points to check the working state of the receiver and ensure the continuity and accuracy of the data.

[0081] I. Data preprocessing;

[0082] Before performing CLAHE algorithm enhancement, Sobel operator edge information extraction, and VMD variational mode decomposition noise removal, the multi-modal image data needs to be preprocessed to provide high-quality initial data for subsequent processing.

[0083] Multi-modal image stitching and correction;

[0084] For unmanned aerial vehicle aerial images, as they are a sequence of multiple images, they need to be stitched to form a complete slope area image. A stitching algorithm based on feature point matching, such as the SIFT algorithm, is used to extract feature points from each image, then find the overlapping area between images through feature point matching, and finally perform image fusion to eliminate stitching marks.

[0085] Video frame images taken by ground cameras may have geometric distortion due to camera installation position deviation or slight shaking during shooting, which need to be corrected. By selecting fixed reference points (landmark objects on the slope) in the image, a correction matrix is calculated to perform geometric correction on the image, so that the image can accurately reflect the actual shape of the slope.

[0086] The laser radar point cloud data needs to be registered with the image data to ensure that the point cloud data and the image data are in the same coordinate system. Feature points in the aerial image of the unmanned aerial vehicle or the image of the ground camera are matched with corresponding points in the laser radar point cloud data, coordinate conversion parameters are calculated, and accurate registration of the point cloud data and the image data is realized.

[0087] Data format unification;

[0088] The aerial image of the unmanned aerial vehicle, the video frame image of the ground camera and the laser radar point cloud data are converted into a unified data format.

[0089] II. The CLAHE algorithm enhances the contrast of the weak deformation region;

[0090] Operation flow;

[0091] Image blocking: the preprocessed image is divided into multiple sub-blocks according to the 2x2 size set by the tileGridSize parameter. The size of the sub-block needs to be considered in terms of the richness of the image details. For areas with more details, the sub-block can be appropriately reduced; for relatively flat areas, the sub-block can be appropriately increased.

[0092] Sub-block histogram equalization: each sub-block is processed by histogram equalization. The distribution of the pixel gray value in the sub-block is counted, the histogram is calculated, and then the mapping relationship of the gray value is adjusted to make the distribution of the histogram more uniform and enhance the contrast of the sub-block.

[0093] Contrast limitation: during the histogram equalization process, when the number of pixels of a certain gray level exceeds the value set by the clipLimit parameter, the excess number of pixels is evenly distributed to other gray levels to avoid the contrast of the gray level being excessively enhanced, thereby suppressing the amplification of noise.

[0094] Sub-block fusion: the processed sub-blocks are spliced into a complete image. In the splicing process, the boundaries between the sub-blocks are processed in a smooth transition manner to avoid obvious block effects.

[0095] Parameter setting basis;

[0096] clipLimit parameter: the setting of this parameter is related to the noise level and contrast of the image. For images with more noise and lower contrast, the clipLimit parameter can be set to a smaller value.

[0097] tileGridSize parameter: the size of the sub-block needs to balance the processing efficiency and the detail enhancement effect. The smaller the sub-block, the more obvious the detail enhancement, but the processing time will increase; the larger the sub-block, the higher the processing efficiency, but some details may be lost.

[0098] III. Sobel operator to extract gradient edge information

[0099] Operation flow;

[0100] Image graying: Since the Sobel operator is mainly used for processing grayscale images, it is necessary to convert color images into grayscale images. By calculating the weighted average of the RGB values of each pixel in the image, the grayscale value is obtained, and the formula is: grayscale value = 0.299 x R + 0.587 x G + 0.114 x B.

[0101] Convolution operation: Apply the Sobel convolution kernel in the horizontal and vertical directions to the grayscale image. Taking a 3x3 convolution kernel as an example, the horizontal direction convolution kernel is used to detect the vertical edge, and the vertical direction convolution kernel is used to detect the horizontal edge. The convolution kernel is slid on the image, and the convolution operation is performed with each pixel and its neighborhood pixels in the image to obtain the gradient values in the horizontal and vertical directions.

[0102] Gradient synthesis: Synthesize the gradient values in the horizontal and vertical directions to obtain the gradient edge information of the image. Common synthesis methods include square sum and square root, i.e. gradient value = √(Gx² + Gy²), where Gx is the horizontal direction gradient value and Gy is the vertical direction gradient value; or absolute value addition, i.e. gradient value = |Gx| + |Gy|.

[0103] Edge threshold processing: In order to highlight the obvious edge information and remove some weak noise edges, threshold processing is needed for the synthesized gradient values. Set a threshold value, when the gradient value is greater than the threshold value, consider that the position is an edge point; otherwise, consider that it is not an edge point. The threshold value can be determined by experiment according to the characteristics of the image, generally between 0.22-0.6.

[0104] Convolution kernel selection basis;

[0105] 3x3 convolution kernel: It is suitable for cases where the edge extraction accuracy requirement is high and the image noise is less. It can better preserve the detailed edges of the image, but it is sensitive to noise.

[0106] IV. VMD variational mode decomposition to filter out lidar point cloud noise

[0107] Algorithm preparation;

[0108] Processing object: The pre-processed and registered lidar point cloud data.

[0109] Data conversion: Convert the lidar point cloud data into one-dimensional signal form. The coordinate values or intensity values of the point cloud can be arranged into a one-dimensional array according to some order in space (along the height direction of the slope, the horizontal direction), forming the signal to be processed.

[0110] Operation flow;

[0111] Initialization parameters: According to the characteristics of laser radar point cloud data, the number of modal decomposition K, the penalty factor α and the convergence tolerance τ are set initially. The determination of K value can refer to the complexity of point cloud data. For point cloud data with simple terrain and less noise,

[0112] Modal decomposition: The converted one-dimensional signal is input into the VMD algorithm, which decomposes the signal into K modal components through iterative optimization process. Each modal component has a specific center frequency and bandwidth, which can reflect different characteristics in the signal.

[0113] Modal component screening: Analyze the frequency characteristics and energy distribution of each modal component. Noise usually appears as high-frequency, low-energy components, while useful point cloud signal components have relatively stable frequency and higher energy. By observing the frequency spectrum and energy curve of the modal component, those suspected noise components with high frequency and low energy are removed.

[0114] Signal reconstruction: The useful modal components after screening are superimposed to reconstruct the laser radar point cloud signal after filtering out noise, and then converted back to point cloud data form.

[0115] Parameter adjustment method;

[0116] When the decomposed modal components cannot effectively separate the noise, first consider adjusting the number of modal decomposition K. If K is too small, it may lead to the inability to completely separate useful signals and noise; if K is too large, it will increase the calculation amount and may introduce redundant components. K value can be gradually increased or decreased, and the decomposition effect is observed.

[0117] Adjustment of penalty factor α: If α is too small, the bandwidth of modal component will be wide, and modal aliasing may occur; if α is too large, the modal component will be too narrow, and useful information may be lost. According to the decomposition result, adjust α value appropriately to make the modal component clearly separated.

[0118] Adjustment of convergence tolerance τ: If τ is too small, it will increase the number of iterations and prolong the calculation time; if τ is too large, it may lead to inaccurate decomposition results. Under the premise of ensuring the accuracy of decomposition, choose appropriate τ value to balance the calculation efficiency and result accuracy.

[0119] Step 102, based on BiLSTM bidirectional long short-term memory network to capture the bidirectional dependency of time series, through Attention attention mechanism to dynamically allocate the weight of different time steps and spatial positions, and establish a hybrid prediction model;

[0120] Specifically, in the present embodiment, a BiLSTM network is constructed, which includes an input layer, a hidden layer, and an output layer. The input layer receives processed feature image data, and the data dimension is determined according to the size and dimension of the feature image.

[0121] The hidden layer is provided with multiple neurons, and the input data is processed by a forward LSTM and a backward LSTM respectively.

[0122] The forward LSTM starts from the starting point of the time series and processes the data of each time step in turn, while the backward LSTM starts from the end point of the time series and processes the data of each time step in reverse. The outputs of the forward and backward LSTMs are spliced to serve as the output of the hidden layer.

[0123] The output of the BiLSTM network is taken as input and mapped to a low-dimensional space through a fully connected layer. The score value of each position is calculated and converted to a weight value through a softmax function.

[0124] The output of each position is multiplied by the corresponding weight value, and a weighted sum is performed to obtain the output of the Attention mechanism.

[0125] Specifically;

[0126] 1. A hybrid prediction model based on BiLSTM and Attention mechanism is constructed, which uses BiLSTM to capture the bidirectional dependency relationship of time series, and dynamically allocates the weights of different time steps and spatial positions through Attention mechanism, to improve the accuracy of slope deformation prediction.

[0127] 2. Specific implementation method;

[0128] 2.1. BiLSTM network construction;

[0129] Principle: BiLSTM (Bidirectional Long Short-Term Memory Network) is composed of forward LSTM and backward LSTM. The forward LSTM processes the forward information of the time series, and the backward LSTM processes the reverse information of the time series. By splicing the outputs of the two, the bidirectional dependency relationship of the time series can be captured, and the historical information and future information can be better utilized for prediction.

[0130] Implementation process: Build a BiLSTM network, including input layer, hidden layer and output layer. The input layer receives the processed feature image data, and the data dimension is determined according to the size and dimension of the feature image. The hidden layer is set to 2 neurons, which are processed by forward LSTM and backward LSTM respectively. Forward LSTM starts from the beginning of the time series and processes each time step data in turn; backward LSTM starts from the end of the time series and processes each time step data in reverse. The outputs of forward and backward LSTM are spliced as the output of the hidden layer. The output layer is set according to the prediction target (slope deformation) The number of neurons is set according to the prediction target (slope deformation) The appropriate activation function (linear activation function) is used to get the prediction result.

[0131] Training process: Use stochastic gradient descent (SGD) or Adam optimizer to train the network. Set the learning rate to 2, batch_size to 7, and training rounds to 200. Update the network parameters through the back propagation algorithm to minimize the loss function (mean square error) between the predicted value and the actual value.

[0132] Parameter initialization: Use the Xavier initialization method to initialize the weight parameters of the network, so that the network can quickly converge at the beginning of training.

[0133] 2.2, Attention mechanism is introduced;

[0134] Principle: Attention mechanism calculates the weight value of each time step and spatial position, and the greater the weight value, the greater the influence of the information at that position on the prediction result. By dynamically adjusting the weight, the model can pay more attention to key information and improve the prediction accuracy.

[0135] Implementation process: Introduce Attention mechanism in the output layer of BiLSTM network. Calculate the attention weight of each time step and spatial position, the specific calculation method is as follows: first, take the output of BiLSTM network as input, and map it to a low-dimensional space through a fully connected layer; then, calculate the score value of each position, and the score value is converted to weight value through softmax function, and the sum of weight values is 1; finally, multiply the output of each position with the corresponding weight value, and perform weighted sum to get the output of Attention mechanism.

[0136] Step 103, using GJO cheetah optimization algorithm to optimize the hyperparameters of the mixed prediction model, and obtaining a target mixed prediction model;

[0137] Specifically, in this embodiment, the prediction error of the model on the validation set is taken as the objective function, and the GJO algorithm is used to search and optimize the hyperparameters;

[0138] In each iteration, the individual cheetah moves according to its own position and group information, updates its position, calculates the objective function value, and finds the optimal hyperparameter combination after multiple iterations.

[0139] Specifically;

[0140] 4.1, GJO algorithm principle;

[0141] GJO (Cheetah Optimization Algorithm) is a heuristic optimization algorithm based on the hunting behavior of cheetahs. Cheetahs search for the optimal solution by searching, surrounding, attacking and other behaviors during the hunting process. The algorithm simulates these behaviors of cheetahs to optimize the objective function.

[0142] 4.2, Hyperparameter optimization process;

[0143] Determine the hyperparameters to be optimized: including the number of hidden layer neurons of BiLSTM network, learning rate, batch_size, training rounds, and related parameters in Attention mechanism, etc.

[0144] Set the objective function: the prediction error (root mean square error) of the model on the validation set as the objective function, the goal is to minimize the objective function value.

[0145] Initialize parameters: set the population size of GJO algorithm to 10, the maximum number of iterations to 120, and the exploration factor, development factor and other parameters according to experience.

[0146] Algorithm execution: search and optimize hyperparameters through GJO algorithm. In each iteration, the individual cheetah moves according to its own position and group information, updates its position, and calculates the objective function value. Keep the individuals with smaller objective function values and eliminate the individuals with larger objective function values. After multiple iterations, the optimal hyperparameter combination is found.

[0147] Step 104, input the feature image data into the target hybrid prediction model for prediction, generate a risk probability heat map, and locate the potential slip surface position of the slope according to the risk probability heat map.

[0148] Specifically, in this embodiment, the risk probability value of each position predicted by the model is mapped to the color space to generate a risk probability heat map, and the potential slip surface is determined in combination with the geological structure and topographic information of the slope.

[0149] Specifically;

[0150] 1, model prediction;

[0151] Input the processed feature image data into the optimized target hybrid prediction model, and the model predicts the deformation of the slope according to the learned rules to obtain the risk probability value of each position.

[0152] 2. Risk probability heat map generation;

[0153] Principle: The risk probability heat map represents the risk probability of different positions of the slope by different colors. The deeper the color, the higher the risk probability.

[0154] Implementation process: Map the risk probability value of each position predicted by the model to the color space according to certain rules to generate the risk probability heat map. Interpolation algorithms (Kriging interpolation) can be used to process discrete risk probability values to make the heat map more smooth and continuous.

[0155] 3. Potential sliding surface positioning;

[0156] Judgment basis: In the risk probability heat map, the area with high and continuous risk probability may be the location of the potential sliding surface. Combined with the geological structure, topography and other information of the slope, the potential sliding surface can be further confirmed.

[0157] The beneficial effects are that 1. It can comprehensively capture the characteristic information of the slope from different angles and dimensions. Compared with a single data source, multi-modal data fusion can complement and verify each other, reduce the uncertainty of the data, and improve the reliability and accuracy of the image data analysis of the geotechnical engineering slope. 2. It can improve the quality of point cloud data and reduce the interference of noise on subsequent analysis, significantly improving the quality of geotechnical engineering slope feature image data. 3. It can focus on key information in image data that has a greater impact on deformation prediction, improving the sensitivity to important features in the image recognition process. 4. It can intuitively show the risk probability of different positions of the slope, and accurately locate the position of the potential sliding surface of the slope according to the heat map.

[0158] Please refer to Figure 2 In a geotechnical engineering slope deformation monitoring method, multi-modal image data is obtained, the CLAHE algorithm is used to enhance the contrast of the weak deformation area of the multi-modal image data, the Sobel operator is used to extract gradient edge information, and the VMD variational mode decomposition is used to filter out laser radar point cloud noise to obtain feature image data including the following steps:

[0159] Step 201, perform grayscale processing on the enhanced image data, and apply Sobel convolution kernels in horizontal and vertical directions to the grayscale image respectively;

[0160] Step 202, synthesize the gradient values in the horizontal and vertical directions to obtain the gradient edge information of the image, and perform threshold processing on the synthesized gradient values;

[0161] Step 203, filtering out the noise of the laser radar point cloud based on VMD variational mode decomposition, inputting the converted one-dimensional signal into the VMD algorithm, decomposing the signal into K modal components through an iterative optimization process, and obtaining feature image data.

[0162] Please refer to Figure 3 In a geotechnical engineering slope deformation monitoring system, the geotechnical engineering slope deformation monitoring system comprises the following modules:

[0163] The slope image acquisition module is configured to acquire multi-modal image data, enhance the contrast of the weak deformation area of the multi-modal image data by using the CLAHE algorithm, extract gradient edge information by using the Sobel operator, filter out the noise of the laser radar point cloud based on VMD variational mode decomposition, and obtain feature image data.

[0164] The prediction model establishment module is configured to capture the bidirectional dependency relationship of the time sequence based on the BiLSTM bidirectional long short-term memory network, dynamically allocate the weights of different time steps and spatial positions by using the Attention attention mechanism, and establish a hybrid prediction model.

[0165] The model parameter optimization module is configured to optimize the hyperparameters of the hybrid prediction model by using the GJO golden jackal optimization algorithm, and obtain a target hybrid prediction model.

[0166] The slope risk identification module is configured to input the feature image data into the target hybrid prediction model for prediction, generate a risk probability heat map, and locate the position of a potential sliding surface of the slope according to the risk probability heat map.

[0167] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring slope deformation in geotechnical engineering, characterized in that, The method for monitoring slope deformation in geotechnical engineering includes the following steps: Multimodal image data is acquired, the contrast of weakly deformed regions of the multimodal image data is enhanced using the CLAHE algorithm, gradient edge information is extracted using the Sobel operator, and noise from the lidar point cloud is filtered out based on VMD variational mode decomposition to obtain feature image data. Based on the BiLSTM bidirectional long short-term memory network to capture the bidirectional dependencies of time series, a hybrid prediction model is established by dynamically allocating weights for different time steps and spatial locations through the Attention mechanism. The hyperparameters of the hybrid prediction model are optimized using the GJO Golden Jackal optimization algorithm to obtain the target hybrid prediction model; The feature image data is input into the target hybrid prediction model for prediction, generating a risk probability heatmap, and the location of potential slip surfaces on the slope is located based on the risk probability heatmap; The BiLSTM bidirectional long short-term memory network captures the bidirectional dependencies of time series, and dynamically allocates weights for different time steps and spatial locations through an attention mechanism to establish a hybrid prediction model, including: Construct a BiLSTM network, which includes an input layer, a hidden layer, and an output layer. The input layer receives processed feature image data and determines the data dimension based on the size and dimension of the feature image. The hidden layer contains multiple neurons, and the input data is processed by forward LSTM and backward LSTM respectively; The forward LSTM starts from the beginning of the time series and processes the data at each time step sequentially; the backward LSTM starts from the end of the time series and processes the data at each time step in reverse; the outputs of the forward and backward LSTMs are concatenated as the output of the hidden layer. The BiLSTM bidirectional long short-term memory network is used to capture the bidirectional dependencies of time series. A hybrid prediction model is established by dynamically allocating weights for different time steps and spatial locations through an attention mechanism. This model also includes: The output of the BiLSTM network is used as input and mapped to a low-dimensional space through a fully connected layer; the score value at each position is calculated and converted into a weight value through a softmax function; The output at each position is multiplied by its corresponding weight value, and then a weighted sum is obtained to get the output of the Attention mechanism. The process of optimizing the hyperparameters of the hybrid prediction model using the GJO (Golden Jackal) optimization algorithm to obtain the target hybrid prediction model includes: Using the model's prediction error on the validation set as the objective function, the hyperparameters are searched and optimized using the GJO algorithm; In each iteration, the individual golden jackal moves and updates its position based on its own location and group information, calculates the objective function value, and finds the optimal combination of hyperparameters after multiple iterations; The step of inputting the feature image data into the target hybrid prediction model for prediction, generating a risk probability heatmap, and locating the potential slip surface of the slope based on the risk probability heatmap includes: The risk probability value of each location predicted by the model is mapped to a color space to generate a risk probability heatmap. Potential slip surfaces are then determined by combining the geological structure and topographic information of the slope.

2. The method for monitoring slope deformation in geotechnical engineering as described in claim 1, characterized in that, The process involves acquiring multimodal image data, enhancing the contrast of weakly deformed regions using the CLAHE algorithm, extracting gradient edge information using the Sobel operator, and filtering out lidar point cloud noise based on VMD variational mode decomposition to obtain feature image data, including: Acquire drone aerial images, ground camera videos, lidar point clouds, and GNSS displacement data of geotechnical engineering slopes to construct multimodal image data; After stitching and correcting the multimodal image data, the format is unified to obtain the initial multimodal image data; The CLAHE algorithm is used to divide the initial multimodal image data into multiple sub-blocks according to the size set by the parameters, thus obtaining block image data; Histogram equalization is performed on the segmented image data, and histograms are calculated by statistically analyzing the gray value distribution of pixels within each sub-block to obtain equalized image data. The equalized image data is re-stitched into a complete image, and the boundaries between sub-blocks are processed using a smooth transition method to obtain enhanced image data.

3. The method for monitoring slope deformation in geotechnical engineering as described in claim 2, characterized in that, The process of acquiring multimodal image data, enhancing the contrast of weakly deformed regions of the multimodal image data using the CLAHE algorithm, extracting gradient edge information using the Sobel operator, and filtering out lidar point cloud noise based on VMD variational mode decomposition to obtain feature image data further includes: The enhanced image data is converted to grayscale, and Sobel convolution kernels are applied to the grayscale image in both the horizontal and vertical directions. The gradient values ​​in the horizontal and vertical directions are synthesized to obtain the gradient edge information of the image, and the synthesized gradient values ​​are then thresholded. Based on Variational Mode Decomposition (VMD) to filter out noise in lidar point clouds, the transformed one-dimensional signal is input into the VMD algorithm. Through an iterative optimization process, the signal is decomposed into K modal components to obtain feature image data.

4. A geotechnical engineering slope deformation monitoring system, used to implement the geotechnical engineering slope deformation monitoring method as described in claim 3, characterized in that, The geotechnical engineering slope deformation monitoring system includes the following modules: The slope image acquisition module is used to acquire multimodal image data, enhance the contrast of weak deformation regions of the multimodal image data using the CLAHE algorithm, extract gradient edge information using the Sobel operator, and filter out lidar point cloud noise based on VMD variational mode decomposition to obtain feature image data. The prediction model building module is used to capture the bidirectional dependencies of time series based on the BiLSTM bidirectional long short-term memory network. It dynamically allocates weights for different time steps and spatial locations through the Attention mechanism to build a hybrid prediction model. The model parameter optimization module is used to optimize the hyperparameters of the hybrid prediction model using the GJO Golden Jackal optimization algorithm to obtain the target hybrid prediction model; The slope risk identification module is used to input the feature image data into the target hybrid prediction model for prediction, generate a risk probability heat map, and locate the potential slip surface of the slope based on the risk probability heat map.

5. The geotechnical engineering slope deformation monitoring system as described in claim 4, characterized in that, The model parameter optimization module includes the following sub-modules: The optimization submodule is used to search and optimize hyperparameters using the GJO algorithm, with the model's prediction error on the validation set as the objective function. The resulting submodule is used in each iteration to allow individual jackals to move based on their own position and group information, update their position, calculate the objective function value, and find the optimal combination of hyperparameters after multiple iterations.

6. The geotechnical engineering slope deformation monitoring system as described in claim 4, characterized in that, The slope risk identification module includes the following sub-modules: The generation submodule is used to map the risk probability value of each location predicted by the model to a color space, generate a risk probability heatmap, and determine potential slip surfaces by combining the geological structure and topographic information of the slope.

Images