Slope instability sliding real-time early warning method based on improved machine learning algorithm

Abstract

The invention discloses a slope instability sliding real-time early warning method based on an improved machine learning algorithm. The method comprises the following steps: S1, outputting a consistent slope monitoring data set; s2, constructing a slope monitoring map structure based on the consistent slope monitoring data set and the spatial position information of each sensor; s3, outputting a slope state feature vector; s4, performing online optimization on key parameters of the dynamic graph attention residual image convolutional neural network model by using an adaptive particle swarm optimization algorithm, and outputting an optimized dynamic graph attention residual image convolutional neural network model; and S5, re-mapping the consistent side slope monitoring data set to generate a new side slope state feature vector, judging an instability sliding risk in the side slope state according to a comparison result between the side slope state feature vector and an early warning threshold value, and generating side slope instability real-time early warning information. According to the method, the risk that too many invalid connections are established in a state stable region is effectively avoided, and the physical rationality and the anomaly capture capability of the slope map in the actual instability trend are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of slope early warning, and particularly to a real-time early warning method for slope instability sliding based on an improved machine learning algorithm. Background Art

[0002] With the in-depth promotion of geological disaster monitoring and prevention work, real-time early warning of slope stability has become an important part of the prevention and control of landslide and collapse geological disasters. Especially in mountainous highways, railway lines, water conservancy facilities and urban slope projects, slope instability events often pose a serious threat to people's lives and property safety and the operation of projects. Therefore, researching efficient and reliable real-time early warning methods for slope instability has become an important research direction in the current geotechnical engineering field.

[0003] Existing slope monitoring systems mostly use multi-source sensor networks to continuously monitor the stress, displacement, pore water pressure, surface temperature, rainfall physical and environmental parameters in the slope area for a long time, and judge the state through expert rules or simple statistical threshold methods. However, there are generally problems such as response lag, low early warning accuracy and poor adaptability. First of all, the early warning mechanism based on fixed thresholds cannot adapt to the complex and changeable slope environment. The response behaviors of different monitoring points under different climate and geological conditions are significantly different, and a unified static judgment standard cannot effectively identify local abnormal evolution trends. Secondly, most of the existing methods ignore the topological structure between the spatial distribution of sensors and the data, and only judge based on single points or a few points, and cannot capture the global collaborative instability characteristics in the slope system. In addition, in the data preprocessing link, conventional methods handle the problems of data missing, noise and inconsistent time steps roughly, resulting in a great reduction in the effect of subsequent model training.

[0004] In recent years, although some studies have introduced traditional machine learning methods to classify and predict slope monitoring data, they have limited capabilities in dealing with time series dependence and spatial structure information. They can often only process flattened feature inputs and are difficult to comprehensively mine the deep laws of slope evolution. At the same time, although existing graph convolutional neural network models attempt to introduce graph structures to process spatial topological relationships, most models build static graphs and have fixed parameters, and are difficult to dynamically adapt to the rapid evolution process of slope states, resulting in inaccurate prediction results when facing mutations or complex working conditions.

[0005] Therefore, there is an urgent need for a new real-time early warning method for slope instability sliding that integrates multi-dimensional monitoring data, has dynamic optimization capabilities and can deeply model the spatio-temporal evolution process to improve the intelligence and reliability of the overall early warning system. Summary of the Invention

[0006] An object of the present invention is to propose a real-time early warning method for slope instability sliding based on an improved machine learning algorithm. The present invention effectively avoids the risk of establishing too many invalid connections in the stable state area, and enhances the physical rationality and abnormal capture ability of the slope map in the actual instability trend.

[0007] A real-time early warning method for slope instability sliding based on an improved machine learning algorithm according to an embodiment of the present invention includes the following steps:

[0008] S1. Deploy a sensor network in the slope area to collect various physical and environmental parameters of the slope area in real time, form an original slope monitoring data set, and perform data preprocessing on the original slope monitoring data set to output a unified slope monitoring data set;

[0009] S2. Construct a slope monitoring graph structure based on the unified slope monitoring data set and the spatial position information of each sensor;

[0010] S3. Input the slope monitoring graph structure into a dynamic graph attention residual graph convolutional neural network model to extract the global spatial features and deep temporal features of the slope monitoring data, and output a slope state feature vector;

[0011] S4. Use an adaptive particle swarm optimization algorithm to online optimize the key parameters of the dynamic graph attention residual graph convolutional neural network model. The adaptive particle swarm optimization algorithm constructs an early warning objective function based on the error information between the slope state feature vector and the corresponding real slope state information in the unified slope monitoring data, and adjusts the key parameters to output an optimized dynamic graph attention residual graph convolutional neural network model;

[0012] S5. Based on the optimized dynamic graph attention residual graph convolutional neural network model, remap the unified slope monitoring data set to generate a new slope state feature vector, determine the instability sliding risk in the slope state according to the comparison result between the slope state feature vector and the early warning threshold, and generate real-time early warning information for slope instability.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Deploy a sensor network including N sensor measurement points in the slope area. The sensor measurement points are respectively used to collect stress information, displacement information, pore water pressure information, surface temperature information and rainfall information of the slope area in real time, and form an original slope monitoring data set D raw :

[0015]

[0016] where d i represents the i-th original monitoring data, and s iIndicates the sensor number, M represents the total number of collected data, and t i is the acquisition timestamp, (x i , y i , z i ) represents the coordinate position of the sensor s i in three-dimensional space, represents the stress information, represents the displacement information, represents the pore water pressure information, represents the surface temperature information, represents the rainfall information;

[0017] S12. Perform time synchronization processing on the original slope monitoring data set. Align the timestamps in the original slope monitoring data set according to the set unified time step, and synchronize the data of various physical and environmental parameters uploaded at different time nodes to a fixed time series to form a time-synchronized slope monitoring data set;

[0018] S13. Perform noise filtering processing on the time-synchronized slope monitoring data set. Traverse each physical and environmental parameter item by item, remove the mutation points and outliers to form a noise-filtered slope monitoring data set;

[0019] S14. Perform missing value filling processing on the noise-filtered slope monitoring data set. When there is missing stress information, displacement information, pore water pressure information, surface temperature information, and rainfall information at a certain time node or a certain sensor measurement point, use an interpolation method combining time interpolation and spatial interpolation to complete the reconstruction of the missing data, and output the slope monitoring data set after missing value filling;

[0020] S15. Perform standardization processing on the slope monitoring data set after missing value filling. Use the maximum-minimum normalization method to uniformly normalize all physical and environmental parameters to the interval [0,1], and define it as the unified slope monitoring data set D norm .

[0021] Optionally, the sensor network includes a stress sensor, a displacement sensor, a pore water pressure sensor, a surface temperature sensor, and a rainfall sensor.

[0022] Optionally, the S2 includes the following steps:

[0023] S21. According to the unified slope monitoring data set D norm in the sensor number s j , the coordinate position of the sensor in three-dimensional space (x j , y j , z j ) and various standardized physical quantities Construct a set of sensor nodes \(V = \{v_1, v_2, \ldots, v\}\) N}, where each sensor node \(v\) i corresponds to a sensor number \(s\) i , and \(N\) is the total number of sensors deployed in the slope area;

[0024] S22. For any sensor node \(v\) i and sensor node \(v\) j , \(i, j\in\{1, 2, \ldots, N\}\), define the spatial distance \(d\) i,j and the dynamic difference index \(\Delta\) based on the standardized physical quantities of each sensor node i,j ;

[0025] S23. Construct a slope monitoring graph structure \(G=(V, E, A)\), where the edge set \(E\) represents the connection relationship between sensor nodes that satisfy geographical proximity and dynamic difference characteristics. The element \(A_{ij}\) of the weighted adjacency matrix \(A\) of the slope monitoring graph is defined as: i,j Defined as:

[0026]

[0027] where \(\sigma\) is the spatial attenuation coefficient, \(\tau\) is the dynamic difference attenuation coefficient, and \(\delta\) is the spatial distance threshold that limits the connection of sensor nodes. Connections are established only within the local range of the slope and when the state anomaly is significant, reflecting the relevance of local anomaly evolution in real-time early warning of slope instability and sliding;

[0028] S24. Construct a graph structure feature matrix where \(F = 5\) represents five types of standardized physical quantities corresponding to each sensor node;

[0029] S25. Output the constructed slope monitoring graph structure \(G=(V, E, A)\) and the graph structure feature matrix \(X\), and optimize the ability to capture local anomaly evolution of slope instability and sliding risk by combining the spatial geometric information of physical monitoring data with the dynamic physical quantity differences.

[0030] Optionally, the said S3 includes the following steps:

[0031] S31. Take the slope monitoring graph structure \(G=(V, E, A)\) and the graph structure feature matrix \(X\) as inputs and send them into the dynamic graph attention residual graph convolutional neural network model. In the dynamic graph attention residual graph convolutional neural network model, let the output of the initial layer be \(H\) (0) =X. For the \(l\)-th layer, \(l = 1, 2, \ldots, L\), for any \(l\)-th layer, extract the local neighborhood feature \(H_{l}^{local}\) (l,1) and the extended neighborhood feature \(H_{l}^{extended}\) (l,2) from two scales of local neighborhood and extended neighborhood respectively;

[0032] S32. For each sensor node, use the graph attention mechanism to calculate the multi-scale feature fusion weights. By mapping the multi-scale graph convolutional fusion output H of the (l - 1)th layer, obtain the multi-scale feature fusion weight vector: (l-1)

[0033]

[0034] where f att (·) represents the attention mapping function composed of a multi-layer perceptron. represents the reference weight of the local neighborhood space information in the current instability feature extraction. represents the reference weight of the extended neighborhood structure in the overall sliding trend inference. softmax(·) ensures that the sum of the two is 1. The multi-scale feature fusion weight vector reflects the importance of features at different scales in the current slope instability situation.

[0035] S34. Fuse the multi-scale features and calculate the fusion output of the lth layer:

[0036]

[0037] where ⊙ represents the element-wise multiplication operation.

[0038] S35. Further embed the time evolution information contained in the slope monitoring data. Introduce a time encoding-based enhancement module in each layer. Let each sensor node v i append the timestamp t of its collected data i and map to generate the corresponding time series vector T i . Aggregate the time series vectors T of all sensor nodes i to form the time series matrix T. Add the time series matrix T and the multi-scale graph convolutional fusion output H of the current layer (l) row by row according to the corresponding nodes to obtain the updated graph convolutional feature matrix

[0039] S36. Establish an improved dynamic residual fusion mechanism and calculate the gating coefficient g (l) for regulating the fusion of the feature information of the current layer and the previous layer:

[0040]

[0041] where is the trainable weight matrix of the gating layer, is the bias term, vec(·) represents flattening the matrix into a vector, σ(·) is the Sigmoid function, and the value range of g (l) is [0, 1];

[0042] Output the residual graph convolutional feature matrix of the lth layer:​

[0043]

[0044] S37. Repeat steps S32 to S36 until the network depth of the dynamic graph attention residual graph convolutional neural network model reaches the predetermined number of layers L, and output the final layer residual graph convolutional feature matrix Z (L*) , where each row z i of the residual graph convolutional feature matrix represents the final slope state feature vector of the sensor node v i . The final slope state feature vector reflects the stability state of each sensor node in the slope area under the action of multi-scale spatio-temporal information.

[0045] Optionally, the local neighborhood feature H (l,1) is extracted by taking each sensor node as the center and combining the spatial topological relationship and feature information of its directly connected first-order neighbor sensor nodes. The formed output is called the local neighborhood feature representation, and the feature extraction process is based on weighted aggregation of the normalized adjacency matrix;

[0046] The extended neighborhood feature H (l,2) is to consider the indirectly connected second-order neighbor sensor nodes on the basis of the local neighborhood, that is, the sensor nodes with a two-hop connection relationship with the current sensor node, and construct an information propagation path through the square operation of the extended normalized adjacency matrix to extract and form the extended neighborhood feature.

[0047] Optionally, S4 includes the following steps:

[0048] S51. Match the final residual graph convolutional feature matrix Z (L*) with the true slope stability label values corresponding to each sensor node in the unified slope monitoring dataset D norm to construct a set of slope state feature-label pairs where represents the true stability label of the sensor node v at the current moment based on historical data or manual annotation, and the value range is [0, 1]. The closer the value is to 1, the closer it is to the sliding instability state; i

[0049] S52. Set a slope sliding instability risk prediction function in the residual graph convolutional neural network model to map the slope state feature vector z i of each sensor node to its sliding instability risk prediction score y i . The sliding instability risk prediction score is a real number, and the value range is from 0 to 1. The higher the score, the more unstable the slope tends to be;

[0050] S53. Construct an objective function for adaptive particle swarm optimization, where the objective function takes the predicted slope state value y output by the dynamic graph attention residual graph convolutional neural network model i and the corresponding true stability label as the mean square error to be minimized f(Θ);

[0051] S54. Initialize the particle swarm population P = {Θ1, Θ2, …, Θ S}, where each particle Θ s represents a combination of residual graph convolutional neural network parameters, S represents the number of particles, and each particle contains its current velocity vector V s , its current optimal position and its historical optimal solution Θ global ;

[0052] S55. Based on the improved adaptive particle swarm optimization update mechanism, dynamically adjust the particle search speed in each iteration round:

[0053]

[0054] where, represents the velocity vector of the s-th particle in the t-th iteration, represents the velocity vector of the s-th particle in the (t + 1)-th iteration, ω (t) is the inertia weight coefficient, which is adaptively adjusted in combination with the convergence situation of the current objective function, c1 and c2 are the individual and group learning factors, and r1 and r2 are the weighted coefficients randomly generated within [0, 1], represents the position of the s-th particle in the t-th iteration, represents the position of the s-th particle in the (t + 1)-th iteration;

[0055] S56. Introduce an adaptive adjustment mechanism for the inertia weight coefficient driven by the slope state change rate. The adaptive adjustment mechanism for the inertia weight coefficient dynamically adjusts the particle search ability according to the change trend of the slope instability state. Take the descent speed of the objective function in continuous iterations as the environmental fitness index, and define the adaptive inertia weight coefficient adjustment function:

[0056]

[0057] where, ω max and ω min are the maximum and minimum values of the inertia weight respectively, f (t) represents the global objective function value at the t-th iteration, f (t-K) represents the global objective function value at the (t - K)-th iteration, K is the sliding time window, and ∈ is a very small positive number to prevent the denominator from being zero;

[0058] S57. Repeat steps S55 to S56 until a preset termination condition is met, including the maximum number of iterations or the change in the objective function being lower than a threshold, and output the converged optimal parameter combination Θ * 。

[0059] Optionally, S5 includes the following steps:

[0060] S51. Re - input the unified slope monitoring dataset into the dynamically graph - attention residual graph convolutional neural network model optimized by adaptive particle swarm optimization. Through forward - propagation calculation, obtain the final residual graph convolutional feature output of each sensor node on the graph structure at the current moment, and represent the final residual graph convolutional feature output as a set of slope - state feature vectors. Each feature vector in the set of slope - state feature vectors is used to represent the slope stability state reflected by the corresponding sensor node at the current moment;

[0061] S52. Input the set of slope - state feature vectors into the trained slope sliding instability risk prediction function, and use the prediction function to map the feature vector of each sensor node to obtain its sliding instability risk prediction score;

[0062] S53. According to the relationship between the sliding instability risk prediction score and the three - class risk - level boundary thresholds, determine the slope stability level of the sensor node at the current moment. The risk - level boundary thresholds include a first threshold, a second threshold, and a third threshold, which respectively represent the boundaries between the safe state and the warning state, the warning state and the dangerous state, and the dangerous state and the extremely high - risk state. The risk - level boundary thresholds satisfy an ascending numerical relationship. The slope stability level division rules are as follows:

[0063] When the sliding instability risk prediction score is less than or equal to the first threshold, it is determined to be in a safe state;

[0064] When the sliding instability risk prediction score is greater than the first threshold and less than or equal to the second threshold, it is determined to be in a warning state;

[0065] When the sliding instability risk prediction score is greater than the second threshold and less than or equal to the third threshold, it is determined to be in a dangerous state; when the sliding instability risk prediction score is greater than the third threshold, it is determined to be in an extremely high - risk state;

[0066] S54. Record the number, three - dimensional spatial position coordinates, sliding instability risk prediction score, and its corresponding slope stability level label of each sensor node together to form a slope instability risk monitoring result set;

[0067] S55. Generate real-time slope instability early warning information based on the slope instability risk monitoring result set. When the number of nodes with a stability level of warning or dangerous state exceeds the set proportion threshold of the total number of monitored nodes, the system generates a warning recommendation text that the slope risk is increasing, and it is recommended to increase the monitoring frequency and formulate an emergency response strategy.

[0068] Optionally, the real-time slope instability early warning information includes the following three types of output content:

[0069] One is the risk level heat map, which associates the horizontal spatial position of each sensor node with its slope stability level label to generate a two-dimensional spatial heat map for visualizing the stability distribution of each position in the slope area;

[0070] The second is the list of risk nodes. Screen all nodes with a stability level of warning, dangerous state, or extremely high risk state from the slope instability risk monitoring result set, and output their numbers, spatial position coordinates, risk prediction scores, and corresponding levels;

[0071] The third is the output of the warning recommendation text. When the stability level of any sensor node is in the extremely high risk state, the system generates a warning recommendation text that there is a critical area of sliding instability in the slope, and it is recommended to immediately verify and initiate an emergency response.

[0072] The beneficial effects of the present invention are:

[0073] (1) The dynamic graph attention residual graph convolutional neural network model proposed by the present invention introduces a dual-scale graph convolutional path of local neighborhood and extended neighborhood, and dynamically calculates the multi-scale feature fusion weight through the attention mechanism, significantly improving the model's joint modeling ability for local slope instability evolution and overall sliding trend. In addition, a dynamic residual fusion gating mechanism is introduced, and the Sigmoid gating function is used to dynamically adjust the information fusion ratio between the current layer and the previous layer according to the feature expression difference, thereby suppressing the accumulation of redundant features and improving the stable convergence of the deep network in complex geological data.

[0074] (2) The present invention innovatively designs a dual-constraint mechanism of spatial distance factor + dynamic physical quantity difference factor. By introducing dynamic difference indicators and spatial distance thresholds, edges are established only between nodes that simultaneously meet the conditions of significant physical state differences and geographical proximity, ensuring that the graph structure can highly sensitively depict the evolution linkage relationship of local potential instability areas on the slope, effectively avoiding the risk of establishing too many invalid connections in the stable state area, and enhancing the physical rationality and abnormal capture ability of the slope graph in the actual instability trend.

[0075] (3) The present invention proposes an adaptive inertia weight adjustment strategy driven by the evolution speed of the slope instability state. By calculating the descent rate of the objective function in the continuous iteration of the model, the environmental dynamics is quantified, and the inertia weight is dynamically adjusted, enabling the particles to have stronger global search ability when the slope state fluctuates violently, and enhancing the local fine-tuning ability under stable trends, significantly improving the adaptability of the model parameters to the dynamic response of the slope state, and enabling the dynamic graph neural network model to maintain better prediction performance in different slope scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0077] Figure 1 is a flowchart of a real-time early warning method for slope instability sliding based on an improved machine learning algorithm proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0079] Refer to Figure 1 , a real-time early warning method for slope instability sliding based on an improved machine learning algorithm, includes the following steps:

[0080] S1. Deploy a sensor network within the slope area to collect various physical and environmental parameters of the slope area in real time, form an original slope monitoring data set, and perform data preprocessing on the original slope monitoring data set to output a unified slope monitoring data set;

[0081] S2. Construct a slope monitoring graph structure based on the unified slope monitoring data set and the spatial position information of each sensor;

[0082] S3. Input the slope monitoring graph structure into a dynamic graph attention residual graph convolutional neural network model to extract the global spatial features and deep temporal features of the slope monitoring data, and output a slope state feature vector;

[0083] S4. Use the adaptive particle swarm optimization algorithm to online optimize the key parameters of the dynamic graph attention residual graph convolutional neural network model. The adaptive particle swarm optimization algorithm constructs an early warning objective function based on the error information between the slope state feature vector and the corresponding real slope state information in the unified slope monitoring data, and adjusts the key parameters to output an optimized dynamic graph attention residual graph convolutional neural network model;

[0084] S5. Based on the optimized dynamic graph attention residual graph convolutional neural network model, remap the unified slope monitoring dataset to generate a new slope state feature vector, determine the instability sliding risk in the slope state according to the comparison result between the slope state feature vector and the warning threshold, and generate real-time early warning information for slope instability.

[0085] In this embodiment, S1 includes the following steps:

[0086] S11. Deploy a sensor network including N sensor measurement points in the slope area. The sensor measurement points are respectively used to collect stress information, displacement information, pore water pressure information, surface temperature information and rainfall information in the slope area in real time, and form the original slope monitoring dataset D raw :

[0087]

[0088] where d i represents the i-th piece of original monitoring data, s i represents the sensor number, M represents the total number of collected data, t i is the collection timestamp, (x i , y i , z i ) represents the coordinate position of the sensor s i in the three-dimensional space, represents the stress information, represents the displacement information, represents the pore water pressure information, represents the surface temperature information, represents the rainfall information;

[0089] S12. Perform time synchronization processing on the original slope monitoring dataset, align the timestamps in the original slope monitoring dataset according to the set unified time step, and synchronize the physical and environmental parameter data uploaded at different time nodes to a fixed time series to form a time-synchronized slope monitoring dataset;

[0090] S13. Perform noise filtering processing on the time-synchronized slope monitoring dataset, traverse each physical and environmental parameter one by one, and remove the mutation points and outliers to form a noise-filtered slope monitoring dataset;

[0091] S14. Perform missing value filling processing on the noise-filtered slope monitoring dataset. When there are missing stress information, displacement information, pore water pressure information, surface temperature information and rainfall information at a certain time node or a certain sensor measurement point, use an interpolation method combining time interpolation and space interpolation to complete the reconstruction of the missing data, and output the slope monitoring dataset after missing value filling;

[0092] S15. Standardize the slope monitoring data set after filling in the missing values. Use the maximum - minimum normalization method to uniformly normalize all physical and environmental parameters to the interval [0, 1], which is defined as the unified slope monitoring data set D norm .

[0093] In this embodiment, the sensor network includes a stress sensor, a displacement sensor, a pore water pressure sensor, a surface temperature sensor, and a rainfall sensor.

[0094] In this embodiment, S2 includes the following steps:

[0095] S21. According to the sensor number s norm in the unified slope monitoring data set D j , the coordinate position (x j , y j , z j ) of the sensor in three - dimensional space, and various standardized physical quantities construct a sensor node set V = {v1, v2, …, v N}, where each sensor node v i corresponds to the sensor number s i , and N is the total number of sensors deployed in the slope area;

[0096] S22. For any sensor node v i and sensor node v j , i, j ∈ {1, 2, …, N}, define the spatial distance d i,j and the dynamic difference index Δ i,j based on the standardized physical quantities of each sensor node;

[0097] S23. Construct a slope monitoring graph structure G=(V, E, A). The edge set E represents the connection relationship between sensor nodes that satisfy geographical proximity and dynamic difference characteristics. The element A i,j of the weighted adjacency matrix A of the slope monitoring graph is defined as:

[0098]

[0099] where σ is the spatial attenuation coefficient, τ is the dynamic difference attenuation coefficient, and δ is the spatial distance threshold that limits the connection of sensor nodes. Connections are established only within a local range of the slope and when the state anomaly is significantly different, reflecting the relevance of local abnormal evolution in the real - time early warning of slope instability sliding;

[0100] S24. Construct a graph structure feature matrix where F = 5 represents five types of standardized physical quantities corresponding to each sensor node;

[0101] S25. Output the constructed slope monitoring graph structure \(G=(V, E, A)\) and the graph structure feature matrix \(X\), and optimize the ability to capture the local abnormal evolution of the slope instability sliding risk by combining the spatial geometric information and the dynamic physical quantity differences of the physical monitoring data.

[0102] In this embodiment, S3 includes the following steps:

[0103] S31. Take the slope monitoring graph structure \(G=(V, E, A)\) and the graph structure feature matrix \(X\) as inputs and send them into the dynamic graph attention residual graph convolutional neural network model. In the dynamic graph attention residual graph convolutional neural network model, let the output of the initial layer be \(H\) (0) \(=X\). For the \(l\)-th layer, \(l = 1, 2, \ldots, L\), for any \(l\)-th layer, extract the local neighborhood feature \(H\) (l,1) and the extended neighborhood feature \(H\) (l,2) ;

[0104] S32. For each sensor node, use the graph attention mechanism to calculate the multi-scale feature fusion weights, and map the multi-scale graph convolutional fusion output \(H\) (l-1) of the \((l - 1)\)-th layer to obtain the multi-scale feature fusion weight vector:

[0105]

[0106] where \(f\) att (·) represents the attention mapping function composed of a multi-layer perceptron, represents the reference weight of the local neighborhood spatial information in the current instability feature extraction, represents the reference weight of the extended neighborhood structure in the overall sliding trend inference, and softmax(·) ensures that the sum of the two is 1. The multi-scale feature fusion weight vector reflects the importance of features at different scales in the current slope instability situation;

[0107] S34. Fuse the multi-scale features and calculate the fusion output of the \(l\)-th layer:

[0108]

[0109] where \(\odot\) represents the element-wise multiplication operation;

[0110] S35. Further embed the time evolution information contained in the slope monitoring data, introduce a time encoding-based enhancement module in each layer, and let each sensor node \(v\) i append the timestamp \(t\) i of the data it collects to map and generate the corresponding time series vector \(T\) i , and for all sensor nodes' time series vectors \(T\) iSummarize to form the time series matrix T, and fuse the time series matrix T with the multi-scale graph convolution of the current layer to output H (l) Add row by row according to the corresponding nodes to obtain the updated graph convolution feature matrix

[0111] S36. Establish an improved dynamic residual fusion mechanism and calculate the gating coefficient g (l) Used to regulate the fusion of feature information between the current layer and the previous layer:

[0112]

[0113] Among them, is the trainable weight matrix of the gating layer, is the bias term, vec(·) represents flattening the matrix into a vector, σ(·) is the Sigmoid function, and the value range of g (l) is [0, 1];

[0114] Output the residual graph convolution feature matrix of the l-th layer:

[0115]

[0116] S37. Repeat steps S32 to S36 until the network depth of the dynamic graph attention residual graph convolution neural network model reaches the predetermined number of layers L, and output the final layer residual graph convolution feature matrix Z (L*) , where each row z i of the residual graph convolution feature matrix represents the final slope state feature vector of the sensor node v i . The final slope state feature vector reflects the stability state of each sensor node in the slope area under the action of multi-scale spatio-temporal information.

[0117] In this embodiment, the local neighborhood feature H (l,1) is extracted with each sensor node as the center, combining the spatial topological relationship and feature information of its directly connected first-order neighbor sensor nodes. The formed output is called the local neighborhood feature representation, and the feature extraction process is based on weighted aggregation of the normalized adjacency matrix;

[0118] The extended neighborhood feature H (l,2) considers the indirectly connected second-order neighbor sensor nodes on the basis of the local neighborhood, that is, the sensor nodes with a two-hop connection relationship with the current sensor node, and constructs an information propagation path through the square operation of the extended normalized adjacency matrix to extract and form the extended neighborhood feature.

[0119] In this embodiment, S4 includes the following steps:

[0120] S51. Combine the final residual graph convolution feature matrix Z (L*) with the unified slope monitoring dataset Dnorm The true slope stability label values corresponding to each sensor node in are matched to construct a set of slope state feature-label pairs where represents the true stability label of the sensor node v at the current moment based on historical data or manual annotation i , and the value range is [0, 1]. The closer the value is to 1, the closer it is to the sliding instability state;

[0121] S52. Set a slope sliding instability risk prediction function in the residual graph convolutional neural network model to map the slope state feature vector z of each sensor node i to its sliding instability risk prediction score y i . The sliding instability risk prediction score is a real number, and the value range is from 0 to 1. The higher the score, the more unstable the slope tends to be;

[0122] S53. Construct an objective function for adaptive particle swarm optimization. The objective function takes the mean square error between the slope state prediction value y i output by the dynamic graph attention residual graph convolutional neural network model and the corresponding true stability label as the minimization target f(Θ);

[0123] S54. Initialize the particle swarm population P = {Θ1, Θ2, …, Θ S}}, where each particle Θ s represents a combination of residual graph convolutional neural network parameters. S represents the number of particles. Each particle contains its current velocity vector V s , its current optimal position and its historical optimal solution Θ global ;

[0124] S55. Based on the improved adaptive particle swarm optimization update mechanism, dynamically adjust the particle search speed in each iteration round:

[0125]

[0126]

[0127] where represents the velocity vector of the s-th particle in the t-th iteration, represents the velocity vector of the s-th particle in the (t + 1)-th iteration, ω (t) is the inertia weight coefficient, which is adaptively adjusted in combination with the convergence of the current objective function. c1 and c2 are individual and group learning factors, and r1 and r2 are weighted coefficients randomly generated within [0, 1], represents the position of the s-th particle in the t-th iteration, Denote the position of the s-th particle in the (t + 1)-th iteration;

[0128] S56. Introduce an adaptive adjustment mechanism for the inertia weight coefficient driven by the rate of change of the slope state. The adaptive adjustment mechanism for the inertia weight coefficient dynamically adjusts the particle search ability according to the changing trend of the slope instability state. Take the decreasing speed of the objective function in consecutive iterations as the environmental fitness index, and define the adaptive inertia weight coefficient adjustment function:

[0129]

[0130] where ω max and ω min are the maximum and minimum values of the inertia weight respectively, f (t) represents the global objective function value at the t-th iteration, f (t-K) represents the global objective function value at the (t - K)-th iteration, K is the sliding time window, and ∈ is a very small positive number to prevent the denominator from being zero;

[0131] S57. Repeat steps S55 to S56 until the preset termination conditions are met, including the maximum number of iterations or the change in the objective function being lower than the threshold, and output the converged optimal parameter combination Θ * .

[0132] In this embodiment, S5 includes the following steps:

[0133] S51. Re-input the unified slope monitoring data set into the dynamically graph attention residual graph convolutional neural network model optimized by the adaptive particle swarm algorithm. Through forward propagation calculation, obtain the final residual graph convolutional feature output of each sensor node on the graph structure at the current moment, and represent the final residual graph convolutional feature output as a set of slope state feature vectors. Each feature vector in the set of slope state feature vectors is used to represent the slope stability state reflected by the corresponding sensor node at the current moment;

[0134] S52. Input the set of slope state feature vectors into the trained slope sliding instability risk prediction function, and use the prediction function to map the feature vectors of each sensor node to obtain its sliding instability risk prediction score;

[0135] S53. According to the relationship between the sliding instability risk prediction score and the three risk level demarcation thresholds, determine the slope stability level of the sensor node at the current moment. The risk level demarcation thresholds include the first threshold, the second threshold, and the third threshold, which represent the demarcation between the safe state and the warning state, the warning state and the dangerous state, and the dangerous state and the extremely high risk state respectively. The risk level demarcation thresholds satisfy the quantitative relationship from small to large. The slope stability level division rules are as follows:

[0136] When the predicted score of sliding instability risk is less than or equal to the first threshold, it is determined to be in a safe state;

[0137] When the predicted score of sliding instability risk is greater than the first threshold and less than or equal to the second threshold, it is determined to be in a warning state;

[0138] When the predicted score of sliding instability risk is greater than the second threshold and less than or equal to the third threshold, it is determined to be in a dangerous state; when the predicted score of sliding instability risk is greater than the third threshold, it is determined to be in an extremely high risk state;

[0139] S54. Record the number, three-dimensional spatial position coordinates, predicted score of sliding instability risk, and its corresponding slope stability level label of each sensor node together to form a monitoring result set of slope instability risk;

[0140] S55. Generate real-time early warning information for slope instability based on the monitoring result set of slope instability risk. When the number of nodes with a stability level of warning state or dangerous state exceeds the set proportion threshold of the total number of monitored nodes, the system generates a warning advice text that the slope risk is increasing, and it is recommended to increase the monitoring frequency and formulate an emergency response strategy.

[0141] In this embodiment, the real-time early warning information for slope instability includes the following three types of output content:

[0142] One is a risk level heat map, which associates the horizontal spatial position of each sensor node with its slope stability level label to generate a two-dimensional spatial heat map for visualizing the stability distribution of each position in the slope area;

[0143] The second is a list of risk nodes, which filters all nodes with a stability level of warning state, dangerous state, or extremely high risk state from the monitoring result set of slope instability risk, and outputs their numbers, spatial position coordinates, risk prediction scores, and corresponding levels;

[0144] The third is the output of warning advice text. When the stability level of any sensor node is in an extremely high risk state, the system generates a warning advice text that there is a critical area of sliding instability in the slope, and it is recommended to immediately verify and initiate an emergency response.

[0145] Example 1:

[0146] At 01:12 in the early morning of September 17, 2023, the displacement sensor with the sensor number V019 in the slope area of section K87+230 of G351 National Highway within the territory of City A uploaded abnormal data. The normalized displacement values continuously reported by this node in 4 consecutive rounds from 00:50 to 01:10 rapidly jumped from 0.41 to 0.79, far exceeding the slip warning threshold of 0.65 set for the project.

[0147] The system background simultaneously received nodes V017 and V021 adjacent to V019, and the pore water pressure data showed a sudden increase within the same time period, rising from 0.33 and 0.38 to 0.62 and 0.64 respectively. The synchronous rainfall data record shows that since 17:00 on September 16, the rainfall intensity in this area has increased from 2.4 mm per hour to 7.1 mm, and the cumulative rainfall has reached 38.6 mm before 01:00 on September 17, reaching the critical rainfall intensity of the local designed slope.

[0148] At 01:14, the slope graph neural network analysis system deployed by the present invention initiated the early warning reasoning process. The dynamic graph residual graph convolutional network read the subgraph composed of 7 nodes with V019 as the core. The edge weights between the nodes were significantly amplified due to the increase in the dynamic difference value. After 4 rounds of graph convolutional reasoning, the final slope instability risk score of node V019 was 0.887, and the system identified its status level as "extremely high risk".

[0149] At 01:15:23, the system automatically encapsulated 6 types of information, namely node number V019, three-dimensional coordinates (102.8341E, 29.3124N, 972m), risk level "extremely high risk", historical displacement trend, risk score curve, and recent rainfall intensity change curve, into a structured JSON data packet and sent it to the county emergency management bureau and the on-site monitoring and control center of the G351 project.

[0150] At the same time, the operation result of the traditional SVM model showed that the risk score of node V019 was only 0.562, which did not exceed the alarm threshold of 0.65, and the system did not output any early warning; the output score of the static graph GCN model was 0.702, judged as "dangerous state", but since it did not reach the level threshold of "extremely high risk", only a mild warning pop-up window was triggered, and no SMS or email alarm was sent.

[0151] At 01:17, the on-duty engineer at the G351 project control center received the SMS emergency alarm pushed by the system and immediately logged in to the visualization platform to view the regional heat map. It could be clearly seen on the map that among the 9 nodes in the section from K87+230 to K87+250, 4 nodes were in the "dangerous" or "extremely high risk" state, and the regional status color quickly transitioned from green to dark red.

[0152] At 01:19, the on-duty engineer called the person in charge of the on-site inspection team. The person in charge immediately drove the inspection vehicle to the incident area. After arriving at the scene, the person in charge found that: there was a slight bulge in the slope area between nodes V019 and V021. Three longitudinal cracks appeared on the surface of the soft mudstone near V019, with a total length of about 1.2 meters. The widest crack reached 1.7 centimeters, and there was water seepage. The system recalculated the model again. Since the crack development speed and the ground settlement trend did not slow down, the score of V019 rose to 0.926, and the score of V021 also exceeded 0.81. The automatic text alert pushed by the background was updated to: "The local area sliding of the slope has entered the critical stage. It is recommended to immediately evacuate the personnel and block the road."

[0153] At 01:34, the emergency treatment instructions were issued. The highway inspection team set up reflective cones in the section from K87+200 to K87+280 and closed one lane. The construction team allocated shotcrete equipment and sandbags to reinforce the slope stability.

[0154] At 02:20, due to proper control and rapid response, no large-scale sliding occurred in the whole area. Only a local soil mass slowly subsided by 4.3 cm around 02:10 and stabilized after 03:00.

[0155] The system log shows that in this incident, a total of 18 risk level update records were generated, 4 automatic alarm pushes were made, and the number of platform logged-in users reached 22. It is worth noting that the first trigger point of the entire response chain - the identification and response of the risk score of V019 exceeding the "extremely high risk" threshold only took 32 seconds.

[0156] To further verify the effect of this system, the project team conducted a retrospective analysis from September 18th to September 22nd, and sorted out various model evaluation data within 24 hours before the event:

[0157] Table 1 Evaluation data of the present invention and traditional methods

[0158]

[0159] It is pointed out in the expert review report that: "Although the scale of this sliding incident is small, it is representative. The dynamic graph neural network used in the present invention can identify risk points in advance and give rapid feedback under the interaction and evolution of stress and water pressure, which is an important breakthrough in the application of graph neural network in the field of geotechnical engineering early warning."

[0160] The whole process of this slope sliding early warning incident fully demonstrates the data response ability, risk identification speed and accuracy of the method of the present invention in the actual scenario. Its high-frequency re-computation ability performs stably in the face of sudden geological responses, effectively making up for the lag and misjudgment problems of traditional models in critical event identification, and providing an expandable slope early warning solution for mountain road engineering.

[0161] The dynamic graph attention residual graph convolutional neural network model proposed by the present invention introduces a dual-scale graph convolutional path of local neighborhood and extended neighborhood, and dynamically calculates the multi-scale feature fusion weight through the attention mechanism, significantly improving the model's joint modeling ability for the local instability evolution and overall sliding trend of the slope. In addition, a dynamic residual fusion gating mechanism is introduced, and the Sigmoid gating function is used to dynamically adjust the information fusion ratio between the current layer and the previous layer according to the feature expression difference, thereby suppressing the accumulation of redundant features and improving the stable convergence of the deep network in complex geological data.

[0162] The present invention innovatively designs a dual-constraint mechanism of spatial distance factor + dynamic physical quantity difference factor. By introducing a dynamic difference index and a spatial distance threshold, edges are established only between nodes that simultaneously meet the conditions of significant physical state differences and geographical proximity, ensuring that the graph structure can highly sensitively depict the evolution linkage relationship of local potential instability areas of the slope, effectively avoiding the risk of establishing too many invalid connections in stable areas, and enhancing the physical rationality and abnormal capture ability of the slope graph in the actual instability trend.

[0163] The present invention proposes an adaptive inertia weight adjustment strategy driven by the evolution speed of the slope instability state. By calculating the descent rate of the objective function in consecutive iterations of the model, the environmental dynamics are quantified, and the inertia weight is dynamically adjusted, enabling the particles to have stronger global search ability when the slope state fluctuates violently, and enhancing the local fine-tuning ability under stable trends, significantly improving the adaptability of the model parameters to the dynamic response of the slope state, and enabling the dynamic graph neural network model to maintain excellent prediction performance in different slope scenarios.

[0164] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A real-time early warning method for slope instability and sliding based on an improved machine learning algorithm, characterized in that: The steps include: S1. Deploy a sensor network in the slope area to collect various physical and environmental parameters of the slope area in real time, form an original slope monitoring data set, and perform data preprocessing on the original slope monitoring data set to output a consistent slope monitoring data set; S2. Constructing a slope monitoring map structure based on the consistent slope monitoring data set and the spatial location information of each sensor; S3. Input the slope monitoring graph structure into the dynamic graph attention residual graph convolutional neural network model, extract the global spatial features and deep temporal features of the slope monitoring data, and output the slope state feature vector; S4. The key parameters of the dynamic graph attention residual graph convolutional neural network model are optimized online using the adaptive particle swarm optimization algorithm. The adaptive particle swarm optimization algorithm constructs the warning objective function based on the error information between the slope state feature vector and the corresponding real slope state information in the consistent slope monitoring data, adjusts the key parameters, and outputs the optimized dynamic graph attention residual graph convolutional neural network model; S5. Based on the optimized dynamic graph attention residual graph convolutional neural network model, the consistent slope monitoring dataset is remapped to generate a new slope state feature vector. The risk of unstable sliding in the slope state is determined according to the comparison result between the slope state feature vector and the warning threshold, and real-time warning information of slope instability is generated.

2. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 1 is characterized in that: The S1 comprises the following steps: S11. A sensor network consisting of N sensor points is deployed in the slope area. The sensor points are used to collect stress information, displacement information, pore water pressure information, surface temperature information and rainfall information in the slope area in real time to form the original slope monitoring data set D. raw : Among them, d i represents the original monitoring data of the ith item, s i represents the sensor number, M represents the total number of collected data, t i is the acquisition timestamp, (x i ,y i ,z i ) indicates sensor s i The coordinate position in three-dimensional space, Represents stress information, Represents displacement information, Represents pore water pressure information, Represents the surface temperature information, Indicates rainfall information; S12. Perform time synchronization processing on the original monitoring data set of the slope, align the timestamps in the original monitoring data set of the slope according to the set unified time step, synchronize various physical and environmental parameter data uploaded at different time nodes to a fixed time series, and form a time-synchronized slope monitoring data set; S13. Perform noise filtering on the time-synchronized slope monitoring data set, traverse various physical and environmental parameters one by one, remove mutation points and outliers, and form a slope monitoring data set after noise filtering; S14. Perform missing value filling processing on the slope monitoring data set after noise filtering. When there is missing stress information, displacement information, pore water pressure information, surface temperature information and rainfall information at a certain time node or a certain sensor measuring point, an interpolation method based on a combination of time interpolation and space interpolation is used to complete the reconstruction of the missing data, and the slope monitoring data set after missing value filling is output; S15. After missing values ​​are filled, the slope monitoring dataset is standardized and all physical and environmental parameters are normalized to the interval [0,1] using the maximum and minimum normalization method, which is defined as the uniform slope monitoring dataset D norm .

3. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 2 is characterized in that: The sensor network includes a stress sensor, a displacement sensor, a pore water pressure sensor, a ground surface temperature sensor and a rainfall sensor.

4. The method for real-time early warning of slope instability and sliding based on an improved machine learning algorithm according to claim 2 is characterized in that: The S2 comprises the following steps: S21. Based on the consistent slope monitoring data set D norm Sensor number in s j , the coordinate position of the sensor in three-dimensional space (x j ,y j ,z j ) and various standardized physical quantities Construct a sensor node set V = {v1, v2, ..., v N }, where each sensor node v i Corresponds to sensor number s i , N is the total number of sensors deployed in the slope area; S22. For any sensor node v i With sensor node v j , i,j∈{1,2,…,N}, defining the spatial distance d i,j And the dynamic difference index Δ based on the standardized physical quantity of each sensor node i,j ; S23. Construct a slope monitoring graph structure G = (V, E, A), where the edge set E represents the connection relationship between sensor nodes that meet the characteristics of geographic proximity and dynamic difference, and the element A of the weighted adjacency matrix A of the slope monitoring graph i,j Defined as: Among them, σ is the spatial attenuation coefficient, τ is the dynamic difference attenuation coefficient, and δ is the spatial distance threshold that limits the connection of sensor nodes. The connection is established only within the local range of the slope and when the state abnormality difference is significant, reflecting the correlation of the evolution of local abnormalities in the real-time early warning of slope instability and sliding. S24. Constructing graph structure feature matrix Where F = 5 represents the five types of standardized physical quantities corresponding to each sensor node; S25. Output the constructed slope monitoring graph structure G = (V, E, A) and the graph structure feature matrix X, and optimize the ability to capture the local abnormal evolution of the slope instability and sliding risk by combining the spatial geometric information of the physical monitoring data with the dynamic physical quantity differences.

5. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 4 is characterized in that: The S3 comprises the following steps: S31. The slope monitoring graph structure G = (V, E, A) and the graph structure feature matrix X are used as input and sent to the dynamic graph attention residual graph convolutional neural network model. In the dynamic graph attention residual graph convolutional neural network model, the initial layer output is set to H (0) =X, for the lth layer, l = 1, 2, ..., L, for any lth layer, extract the local neighborhood features H from the local neighborhood and the extended neighborhood respectively (l,1) and the extended neighborhood feature H (l,2) ; S32. For each sensor node, the graph attention mechanism is used to calculate the multi-scale feature fusion weight, and the output H is obtained by fusing the l-1 layer multi-scale graph convolution. (l-1) Mapping is performed to obtain the multi-scale feature fusion weight vector: Among them, f att (·) represents the attention mapping function composed of a multi-layer perceptron, represents the reference weight of the local neighborhood spatial information in the current instability feature extraction, represents the reference weight of the extended neighborhood structure in the overall sliding trend reasoning, softmax(·) ensures that the sum of the two is 1, and the multi-scale feature fusion weight vector reflects the importance of different scale features in the current slope instability situation; S34. Fuse multi-scale features and calculate the l-th layer fusion output: Among them, ⊙ represents the element-by-element multiplication operation; S35. Further embed the time evolution information contained in the slope monitoring data, introduce an enhancement module based on time coding in each layer, and make each sensor node v i Add the timestamp t of the collected data i Mapping generates the corresponding time series vector T i , the time series vector T of all sensor nodes i Summarize to form a time matrix T, fuse the time matrix T with the current layer multi-scale graph convolution to output H (l) Add row by row according to the corresponding nodes to get the updated graph convolution feature matrix S36. Establish an improved dynamic residual fusion mechanism and calculate the gating coefficient g (l) Used to control the fusion of feature information of the current layer and the previous layer: in, is the trainable weight matrix of the gating layer, is the bias term, vec(·) indicates that the matrix is ​​flattened into a vector, σ(·) is the Sigmoid function, and g (l) The value range of is [0,1]; Output the l-th layer residual graph convolution feature matrix: S37. Repeat steps S32 to S36 until the network depth of the dynamic graph attention residual graph convolutional neural network model reaches a predetermined number of layers L, and output the final layer residual graph convolution feature matrix Z (L*) , where each row z of the residual graph convolution feature matrix i Represents the sensor node v i The final slope state eigenvector reflects the stability state of each sensor node in the slope area under the action of multi-scale spatiotemporal information.

6. The method for real-time early warning of slope instability and sliding based on an improved machine learning algorithm according to claim 5 is characterized in that: The local neighborhood feature H (l,1) Taking each sensor node as the center, the spatial topological relationship and feature information of its directly connected first-order neighboring sensor nodes are combined to extract the output, which is called local neighborhood feature representation. The feature extraction process is based on weighted aggregation of the normalized adjacency matrix. The extended neighborhood feature H (l,2) In order to consider the indirectly connected second-order neighbor sensor nodes on the basis of local neighborhood, that is, the sensor nodes that have a two-hop connection relationship with the current sensor node, the information propagation path is constructed through the extended normalized adjacency matrix square operation to extract the extended neighborhood features.

7. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 5 is characterized in that: The S4 comprises the following steps: S51. Convolve the final residual graph feature matrix Z (L*) With the consistent slope monitoring dataset D norm The true slope stability label value corresponding to each sensor node in Match and build a set of slope state feature-label pairs in Represents the sensor node v at the current moment based on historical data or manual annotation i The stability true label of is in the range of [0,1]. The closer the value is to 1, the closer it is to the sliding instability state. S52. Set up a slope sliding instability risk prediction function in the residual graph convolutional neural network model to transform the slope state feature vector z of each sensor node i Mapped to its sliding instability risk prediction score y i , the sliding instability risk prediction score is a real number, ranging from 0 to 1. The higher the score, the more likely the slope is to be unstable; S53. Construct an objective function for adaptive particle swarm optimization, which outputs the slope state prediction value y using a dynamic graph attention residual graph convolutional neural network model i The stability of the corresponding true label The mean square error between them is taken as the minimization target f(Θ); S54. Initialize the particle swarm population P = {Θ1, Θ2, ..., Θ S }, where each particle Θ s represents a residual graph convolutional neural network parameter combination, S represents the number of particles, and each particle contains its current velocity vector V s , current optimal position and its historical optimal solution Θ global ; S55. Based on the improved adaptive particle swarm optimization update mechanism, the particle search speed is dynamically adjusted in each iteration round: in, represents the velocity vector of the sth particle in the tth iteration, represents the velocity vector of the sth particle in the t+1th iteration, ω (t) is the inertia weight coefficient, which is adaptively adjusted based on the convergence of the current objective function. c1 and c2 are individual and group learning factors. r1 and r2 are weight coefficients randomly generated in [0,1]. represents the position of the sth particle in the tth iteration, represents the position of the sth particle in the t+1th iteration; S56. The inertia weight coefficient adaptive adjustment mechanism driven by the slope state change rate is introduced. The inertia weight coefficient adaptive adjustment mechanism dynamically adjusts the particle search capability according to the change trend of the slope instability state, takes the target function decrease speed in continuous iteration as the environmental fitness index, and defines the adaptive inertia weight coefficient adjustment function: Among them, ω max and ω min are the maximum and minimum values ​​of the inertia weight, respectively, and f (t) represents the global objective function value at the tth iteration, f (t-K) represents the global objective function value at the tKth iteration, K is the sliding time window, and ∈ is a very small positive number to prevent the denominator from being zero; S57. Repeat steps S55 to S56 until the preset termination conditions are met, including the maximum number of iterations or the objective function change is lower than the threshold, and output the optimal parameter combination Θ after convergence * .

8. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 7 is characterized in that: The S5 comprises the following steps: S51. Re-input the consistent slope monitoring data set into the dynamic graph attention residual graph convolutional neural network model optimized by adaptive particle swarm optimization, obtain the final residual graph convolution feature output of each sensor node on the graph structure at the current moment through forward propagation calculation, and represent the final residual graph convolution feature output as a slope state feature vector set, each feature vector in the slope state feature vector set is used to represent the slope stability state reflected by the corresponding sensor node at the current moment; S52. Inputting the slope state feature vector set into the trained slope sliding instability risk prediction function, and using the prediction function to map the feature vector of each sensor node to obtain its sliding instability risk prediction score; S53. According to the relationship between the sliding instability risk prediction score and the three risk level demarcation thresholds, the slope stability level of the sensor node at the current moment is determined. The risk level demarcation thresholds include the first threshold, the second threshold, and the third threshold, which respectively represent the demarcation between the safe state and the warning state, the warning state and the dangerous state, and the dangerous state and the extremely high risk state. The risk level demarcation thresholds satisfy the quantitative relationship from small to large. The slope stability level classification rules are as follows: When the sliding instability risk prediction score is less than or equal to the first threshold, it is determined to be a safe state; When the sliding instability risk prediction score is greater than the first threshold and less than or equal to the second threshold, it is determined to be in a warning state; When the sliding instability risk prediction score is greater than the second threshold and less than or equal to the third threshold, it is determined to be a dangerous state; when the sliding instability risk prediction score is greater than the third threshold, it is determined to be an extremely high risk state; S54. Record the number, three-dimensional spatial position coordinates, sliding instability risk prediction score and corresponding slope stability grade label of each sensor node to form a slope instability risk monitoring result set; S55. Generate real-time warning information on slope instability based on the slope instability risk monitoring result set. When the number of nodes with a stability level of warning or dangerous status exceeds the set ratio threshold of the total number of monitoring nodes, the system generates a warning suggestion text that the slope risk is increasing, and it is recommended to increase the monitoring frequency and formulate an emergency response strategy.

9. The real-time early warning method for slope instability and sliding based on an improved machine learning algorithm according to claim 8 is characterized in that: The real-time warning information of slope instability includes the following three types of output content: The first one is the risk level heat map, which associates the horizontal spatial position of each sensor node with its slope stability level label to generate a two-dimensional spatial heat map for visualizing the stability distribution of each position in the slope area; The second is the risk node list, which selects all nodes with stability levels of warning state, dangerous state or extremely high risk state from the slope instability risk monitoring results, and outputs their numbers, spatial location coordinates, risk prediction scores and corresponding levels; The third is the output of early warning suggestion text. When the stability level of any sensor node is in an extremely high-risk state, the system generates an early warning suggestion text that there is a critical area of ​​sliding instability on the slope, and it is recommended to immediately check and initiate an emergency response.

Images