Design method of slope protection monitoring and early warning system

Abstract

The present application relates to the technical field of slope protection, and particularly relates to a design method of a slope protection monitoring and early warning system. The technical scheme comprises the following steps: constructing a multi-dimensional perception network, analyzing space-time characteristics and coupling modeling of multiple physical fields, dynamically evaluating risks and intelligently adopting early warning strategies, adopting a hierarchical response mechanism, and verifying the system. The present application improves data complementarity and analysis reliability through the multi-dimensional perception network, solves the problem of missing sampling by combining dynamic sampling measures, reduces the misjudgment rate by improving the Moran's I index analysis, and further improves the monitoring and early warning accuracy of the system by quantifying the water-separation-heat interaction effect through the seepage-pressure-temperature coupling model. In addition, the reliability of the decision basis is dynamically improved according to the Bayesian network, and finally the early warning threshold is automatically adjusted through the dynamic threshold mechanism, solving the problem that the fixed threshold adopted by the traditional system cannot be dynamically adjusted with the change of the environment, and reducing the false positive rate and the false negative rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of slope protection, in particular to a design method of a slope protection monitoring and early warning system. BACKGROUND

[0002] Slope protection is a technical system that controls the stability of slopes that may have landslides, collapses and other geological disasters through engineering techniques to protect life and property safety and engineering facility safety. The main protection measures include engineering protection, ecological protection and monitoring and early warning. The engineering protection includes installing anchor rods, anti-slide piles, retaining walls or vegetation concrete on the slope, and the ecological protection includes laying vegetation protection slope and vine plant slope fixation. The monitoring and early warning monitors the deformation, stress, hydrology and other parameters of the slope in real time through sensors, and predicts the disaster risk in combination with the model. The existing design method of the slope protection monitoring and early warning system often sets the parameter threshold values such as displacement rate and stress value, triggers the early warning when the threshold values are exceeded, lacks spatial correlation analysis, and does not consider the coupling effect of environmental dynamic changes on the slope stability, so the false positive rate is high. Therefore, we propose a design method of a slope protection monitoring and early warning system. SUMMARY

[0003] The purpose of the present application is to solve the problem of the fixed threshold early warning set by the existing design method of the slope protection monitoring and early warning system in the background art, lack of spatial correlation analysis, and not considering the coupling effect of environmental dynamic changes on the slope stability, and high false positive rate. A design method of a slope protection monitoring and early warning system is proposed.

[0004] The technical scheme of the present application: a design method of a slope protection monitoring and early warning system, comprising the following steps:

[0005] Constructing a multi-dimensional perception network to collect slope data: a multi-dimensional perception network covering the slope is constructed through a sensor network and a space-time synchronization technology;

[0006] Analysis of space-time characteristics and multi-physical field coupling modeling: based on spatial autocorrelation analysis and graph convolution network, a full-coupling model of seepage-stress-temperature field is established;

[0007] Dynamic risk assessment and intelligent early warning strategy: the probability reasoning of multi-source evidence is carried out through a Bayesian network, and the early warning threshold is dynamically optimized based on a reinforcement learning algorithm;

[0008] Establishing a hierarchical response mechanism and optimizing the system: different disposal measures are triggered according to the risk level through the hierarchical response mechanism, and the model parameters are updated online;

[0009] System verification: the reliability of the system is verified through multi-dimensional performance test, and the system is verified through benchmark value calibration and joint debugging test.

[0010] Optionally, the step of collecting slope data by the constructed multi-dimensional perception network comprises the following steps:

[0011] Designing a sensor network: using an unmanned aerial vehicle LiDAR to pre-scan the surface morphology of the slope to obtain three-dimensional terrain data of the slope, calculating the optimal spacing of the sensor nodes based on a Delaunay triangulation algorithm, the sensor nodes being distributed in a honeycomb three-dimensional grid, and each sensor node recording three-dimensional coordinates and time information;

[0012] The sensor comprises an FBG optical fiber grating sensor, a MEMS accelerometer, a piezoelectric pressure gauge and a multi-parameter water quality sensor, wherein the FBG optical fiber grating sensor is arranged along the potential sliding surface of the slope, the MEMS accelerometer is installed at the center position of the grid node, the piezoelectric pressure gauge is arranged in the key mechanical response area, and the multi-parameter water quality sensor is arranged in the underground water level monitoring hole;

[0013] A time-space synchronous acquisition mechanism: calculating the time offset and transmission delay of the sensor based on a two-way time synchronization algorithm based on IEEE 1588v2 protocol, the sampling frequency of the sensor comprising a low-frequency sampling, a high-frequency sampling and a 20Hz sampling frequency, the low-frequency sampling being used when the slope is in a stable period, the sampling frequency being 5Hz, the high-frequency sampling being used when the slope is in a rainfall period, the sampling frequency being 50Hz, and the sampling frequency being 20Hz when the slope is in a pre-warning state, the two-way time synchronization algorithm formula being as follows:

[0014]

[0015] Wherein, is the master clock time, , , and are the slave clock sending synchronization request time, the master clock receiving request time, the master clock returning response time and the slave clock receiving response time, respectively, is the transmission delay error.

[0016] Optionally, the step of analyzing the space-time characteristics and the multi-physical field coupling modeling comprises the following steps:

[0017] Spatial feature analysis and anomaly detection: using an improved Moran's I index to analyze the spatial correlation of the monitoring parameters, identifying the deformation concentration area, and introducing a distance attenuation factor, the improved Moran's I index formula being as follows:

[0018]

[0019] Wherein, This represents the number of sensor nodes. Let be the spatial weight matrix, and , For distance attenuation factor, For nodes and European distance, For nodes The monitoring parameter values, The Moran's I value is used as the parameter mean. Spatial regions are judged by the magnitude of the Moran's I value, including weakly correlated regions and strongly correlated regions. When a region is judged to be strongly correlated, local intensified monitoring is performed.

[0020] The spatiotemporal dependencies between nodes are modeled using a spatiotemporal graph convolutional network, and the formula is expressed as follows:

[0021]

[0022] in, For the first Layer feature matrix, For the first Chebyshev polynomial basis functions of order 1. For the normalized Laplace matrix, For the first Layer The trainable parameter matrix of an order polynomial, For activation functions;

[0023] Multiphysics dynamic coupling modeling: The coupling equations are solved using a finite element-discrete element coupling algorithm. The three-level coupling control equations are as follows:

[0024]

[0025] in, The permeability coefficient varies with temperature. For the water head, For water storage rate, For sources and sinks such as rainfall infiltration, For stress tensor, For soil density, It is the acceleration due to gravity. For displacement vectors, For Biot coefficient, For specific heat capacity, Thermal conductivity, is the dissipation coefficient.

[0026] Optionally, the finite element-discrete element coupled algorithm solution includes stress field solution and seepage field iteration, and the stress field solution formula is:

[0027]

[0028] wherein, is the global stiffness matrix, is the nodal displacement vector, is the external load vector, is the seepage force vector;

[0029] The seepage field iteration formula is:

[0030]

[0031] wherein, is the water head at the moment, is the time step, represents the permeability coefficient matrix, is the flow vector at the moment, is the water storage rate matrix.

[0032] Optionally, the dynamic risk assessment and intelligent early warning strategy comprises the following steps:

[0033] Based on the Bayesian network for dynamic risk assessment: a Bayesian network containing environmental factors, geometric parameters and real-time monitoring data is constructed, and the slope instability risk is quantified through conditional probability inference, and the Bayesian network inference formula is:

[0034]

[0035] wherein, represents a slope instability event, and the slope instability event state includes occurrence and non-occurrence, is the evidence set, and the evidence set includes displacement rate, pressure change rate, pore water pressure and rainfall intensity, represents the instability prior probability, is the probability of the occurrence of the evidence set under the slope instability event;

[0036] A dynamic threshold mechanism is constructed: the dynamic threshold mechanism is constructed based on a sliding window statistics and a reinforcement learning algorithm, and the threshold includes a reference threshold and a real-time data statistics, and the dynamic threshold formula is:

[0037]

[0038] wherein, is the moment of early warning threshold, is the reference threshold, To balance prior knowledge with historical threshold weights for real-time data, This is the 24-hour sliding window average. The standard deviation of the sliding window. The adjustment coefficient is mentioned above;

[0039] The reinforcement learning algorithm optimizes the correction coefficients and minimizes the false alarm rate and average warning delay time. The optimization formula for the reinforcement learning algorithm is as follows:

[0040]

[0041] in, For learning rate, The reward function is... The formula is expressed as:

[0042]

[0043] in, The false alarm rate is the proportion of false alarms to the total number of alarms. The average warning delay time, The non-warning rate is the proportion of slope instability events that were not warned in advance.

[0044] A hybrid time-series prediction model for predicting slope deformation time series is established: The hybrid time-series prediction model adopts an LSTM-Transformer concatenated architecture, which includes LSTM layers and Transformer layers. The LSTM layers are used to extract local features of the deformation sequence, and the Transformer layers predict long-term deformation trends through a self-attention mechanism. The input of the hybrid time-series prediction model includes normalized features of environmental parameters, and the output is the predicted value of the environmental parameters. The network architecture formula of the hybrid time-series prediction model is as follows:

[0045]

[0046]

[0047]

[0048] in, for Input feature vector at any time, Let be the hidden state vector of the LSTM layer, with dimension . , These represent the query, key, and value matrices, respectively. For the first The weight matrix of each attention head. Projecting a matrix of multi-head attention outputs.

[0049] Optionally, the establishing a hierarchical response mechanism and system optimization comprises the following steps:

[0050] Establishing a hierarchical response mechanism: the hierarchical response mechanism comprises risk grading and disposal strategies, the risk grading comprises a low risk level, a medium risk level and a high risk level, wherein the low risk level comprises a displacement rate < 1.2-1.5 mm / d, a stress rate of change < 2-3 kPa / d, a pore water pressure growth rate < 0.1-0.2 m / d, and daily data review 1-2 times, the medium risk level comprises a displacement rate 15-5 mm / d, a stress rate of change 3-8 kPa / d, a pore water pressure growth rate 0.2-0.5 m / d, when in the medium risk level, automatically starting a slope drainage system, a monitoring frequency of 1-2 hours once, and the high risk level comprises a displacement rate > 5 mm / d, a stress rate of change > 8 kPa / d, a pore water pressure growth rate > 0.5 m / d, when in the high risk level, automatically starting real-time video monitoring, and synchronously implementing traffic control strategies and personnel evacuation preparation;

[0051] Model parameter optimization based on reinforcement learning algorithm: taking model parameters as decision variables of the reinforcement learning algorithm, the model parameters include early warning threshold coefficients and prediction model weights, taking the false alarm rate, early warning delay and un-early warning rate as evaluation indexes, and optimizing the parameters through a reward function, the model parameter update formula is:

[0052]

[0053] wherein, is a model parameter vector, is a learning rate, is a discount factor, is a state at time t, is an action at time t, is an immediate reward after executing the action .

[0054] Optionally, the system verification comprises the following steps:

[0055] Multi-dimensional performance verification: verifying the system performance from the aspects of spatial resolution, timing prediction accuracy and early warning reliability;

[0056] Response delay test: introducing a simulated instability signal at the edge computing node, measuring the time from signal generation to early warning information emission, and the response delay is less than 5 seconds;

[0057] Engineering deployment: the engineering deployment includes sensitive area identification, sensor optimal arrangement, benchmark value calibration, and system commissioning and false alarm suppression;

[0058] Engineering maintenance: including inspection, data calibration and model iteration, the inspection includes checking the sensor power supply system, communication link state and sensor shielding every week, the data calibration includes zero point calibration of the FBG fiber Bragg grating sensor and piezoelectric pressure gauge every quarter, updating the benchmark value database, the model iteration includes retraining the Bayesian network and time series prediction model according to the newly collected monitoring data.

[0059] Optionally, the spatial resolution is verified by calculating the average value of the minimum distance between sensor nodes to evaluate the spatial coverage capability, and the spatial resolution verification formula is as follows:

[0060]

[0061] Among them, is the average monitoring distance, is the Euclidean distance between the node and the nearest neighbor node , and is the total number of sensor nodes;

[0062] The time series prediction accuracy is verified by the root mean square error of the time series prediction model through historical data back testing, and the time series prediction accuracy verification formula is:

[0063]

[0064] Among them, is the measured value of the environmental parameter, is the model prediction value, is the number of verification samples;

[0065] The early warning reliability is verified by the false alarm rate and response delay through simulation of instability events, and the early warning reliability verification formula is:

[0066]

[0067] Among them, is the number of false alarms, is the total number of early warnings, the verification period is ≥6 months, and <5%.

[0068] Optionally, the sensitive area identification includes shallow structure scanning of the slope by using ground penetrating radar to identify potential instability areas, and obtaining slope surface deformation data by unmanned aerial vehicle aerial survey to delineate high-risk areas as the key deployment area of the sensor;

[0069] The sensor optimization arrangement comprises minimizing deployment cost by a tabu search algorithm, the tabu search algorithm is formulaed as:

[0070]

[0071] Wherein, is the deployment cost of the node , represents the slope monitoring area, is the sensor node set covering the position , and is the effective monitoring radius of a single sensor. The reference value calibration comprises environmental background monitoring for 6-10 days continuously after sensor installation, collecting environmental data under no significant deformation condition as reference value data, establishing a multi-dimensional parameter baseline, and the reference value statistical quantity calculation formula is:

[0072]

[0073]

[0074] Wherein, is the monitoring data matrix in the reference period, is the reference mean value, is the reference standard deviation.

[0075] The system joint debugging and false alarm suppression comprises simulation test, false alarm suppression and communication test.

[0076] Optionally, the simulation test comprises injecting different levels of simulation instability signals by a signal generator, verifying the accuracy of the early warning level and response measures, the false alarm suppression comprises sliding window filtering and multi-parameter cross verification, when a single sensor triggers an early warning, verifying whether there is an anomaly in the data of the adjacent nodes, and the communication test comprises verifying the data transmission reliability between the sensor nodes and the edge computing settlement, the cloud server.

[0077] In summary, the present application comprises at least one of the following beneficial technical effects:

[0078] The present application realizes global coverage of slope surface deformation and internal mechanical response through a multi-dimensional perception network, and simultaneously collects multi-dimensional data through various sensors, improves data complementarity and analysis reliability, and automatically adjusts the sampling density in different environments by combining dynamic sampling measures, solving the missing sampling problem of traditional fixed frequency sampling on sudden changes.

[0079] ​​The application improves the Moran's I index by introducing a distance attenuation factor to analyze and identify the concentrated area of slope deformation, improves the spatial correlation analysis accuracy by weakening the far-field interference, reduces the misjudgment rate, and captures the spatial propagation path and time evolution trend of slope deformation through the combination of graph convolution network and attention mechanism, improves the monitoring accuracy, at the same time, a seepage-pressure-temperature three-field coupling model is established to quantify the water-ion-heat interaction effect, which can adapt to complex geological environment, and further improve the system monitoring and early warning accuracy;

[0080] The application updates the slope instability risk probability dynamically through the Bayesian network, improves the reliability of the decision basis, and adjusts the early warning threshold automatically according to the real-time data distribution through the dynamic threshold mechanism based on the reinforcement learning algorithm, solves the problem that the fixed threshold adopted by the traditional system cannot be dynamically adjusted with the change of environment, and reduces the false positive rate and the false negative rate. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 A flowchart of the design method of the slope protection monitoring and early warning system is given. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0083] EMBODIMENT

[0084] The design method of the slope protection monitoring and early warning system proposed in the application, as shown in Figure 1 includes the following steps:

[0085] I. Constructing a multidimensional perception network to collect slope data: a multidimensional perception network covering the slope is constructed through a sensor network and a space-time synchronization technology, including the following steps:

[0086] Designing a sensor network: the surface morphology of the slope is pre-scanned by an unmanned aerial vehicle LiDAR to obtain three-dimensional terrain data of the slope, the optimal spacing of the sensor nodes is calculated based on a Delaunay triangulation algorithm, the sensor nodes are distributed in a honeycomb three-dimensional grid, and each sensor node records three-dimensional coordinates and time information;

[0087] The sensors include FBG fiber grating sensors, MEMS accelerometers, piezoelectric pressure gauges and multi-parameter water quality sensors. One monitoring line is arranged every 10 m along the slope trend, 80 surface nodes are arranged every 8 m in elevation, 3 monitoring holes are constructed in the potential sliding surface area (K34+250 section), the depth of each monitoring hole is 20 m, FBG fiber grating sensors and piezoelectric stress gauges are arranged in the monitoring hole every 5 m, the FBG fiber grating sensors are arranged along the trend of the potential sliding surface of the slope, and the layered strain is monitored in the monitoring hole, the strain accuracy of the FBG fiber grating sensor is preferably ±1.5με, the range is ±1500με, the sampling frequency is 5-50 Hz, the MEMS accelerometer is installed at the center of the grid node to monitor the vibration frequency response, the resolution of the MEMS accelerometer is 0.001g, the bandwidth is 0-100 Hz, the piezoelectric pressure gauge is arranged in the key mechanical response area, and the multi-parameter water quality sensor is arranged in the underground water level monitoring hole. The multi-parameter water quality sensor includes a pore water pressure sensor and a rain gauge, the pore water pressure sensor measures a range of 0-20 m water column, an error of <0.3% FS, and the rain gauge is installed on the slope top to monitor rainfall data, with a resolution of 0.1 mm.

[0088] The space-time synchronous acquisition mechanism: the time offset and transmission delay of the sensor are calculated by the bidirectional time synchronization algorithm based on the IEEE 1588v2 protocol, and the sampling frequency of the sensor includes low frequency sampling, high frequency sampling and 20 Hz sampling frequency. When the slope is in the stable period, the low frequency sampling is adopted, and the sampling frequency is 5 Hz. When the slope is in the rainfall period, the high frequency sampling is adopted, and the sampling frequency is 50 Hz. When the slope is in the early warning state, the sampling frequency is 20 Hz. The stable period is when there is no rainfall and the displacement rate is less than 1 mm / d. When the rainfall intensity is greater than 10 mm, it is the rainfall period. When any sensor parameter exceeds 70% of the threshold value, it is in the early warning state.

[0089] The formula of the bidirectional time synchronization algorithm is as follows:

[0090]

[0091] Wherein, is the master clock time, , , and are the slave clock sending synchronization request time, the master clock receiving request time, the master clock returning response time and the slave clock receiving response time, is the transmission delay error.

[0092] The application realizes global coverage of slope surface deformation and internal mechanical response through a multi-dimensional perception network, and improves data complementarity and analysis reliability by simultaneously collecting multi-dimensional data through multiple sensors, and solves the missing sampling problem of traditional fixed frequency sampling on sudden changes by automatically adjusting the sampling density in different environments combined with dynamic sampling measures.

[0093] II. Analysis of spatio-temporal characteristics and multi-physical field coupling modeling: based on spatial autocorrelation analysis and graph convolution network to establish full coupling model of seepage-stress-temperature field, including the following steps:

[0094] Spatial feature analysis and anomaly detection: using improved Moran's I index to analyze the spatial correlation of monitoring parameters, identifying deformation concentrated areas, and introducing a distance attenuation factor, the improved Moran's I index formula is as follows:

[0095]

[0096] Wherein, is the number of sensor nodes, is the spatial weight matrix, and , is the distance attenuation factor, is the Euclidean distance between nodes and , is the monitoring parameter value of node , is the parameter mean value, the spatial region is judged by the size of Moran's I value, including spatial weak correlation region and spatial strong correlation region, when > 0.4, it is determined as a spatial strong correlation region, when it is determined as a spatial strong correlation region, local encryption monitoring is carried out, and local encryption monitoring is carried out by unmanned aerial vehicle LiDAR encryption scanning.

[0097] The spatio-temporal dependence relationship between nodes is modeled through the spatio-temporal graph convolution network, taking the 20 nodes of the potential sliding surface area as the center to generate an adjacency matrix, each node is connected to 3-5 adjacent nodes, the input feature is the standardized sequence of the monitored data, the window length is 24 hours, and the output is the displacement prediction value in the next 12 hours, and its formula is expressed as:

[0098]

[0099] Wherein, is the feature matrix of the layer, is the Chebyshev polynomial basis function of the order, is the normalized Laplacian matrix, For the first Layer the Order polynomial trainable parameter matrix, For activation function.

[0100] Multi-physical field dynamic coupling modeling: the coupled equations are solved by finite element-discrete element coupling algorithm, and the three-level coupling control equation is as follows:

[0101]

[0102] Wherein, The permeability coefficient changes with temperature, The water head, The water storage rate, The source and sink term such as rainfall infiltration, The stress tensor, The soil density, The gravity acceleration, The displacement vector, The Biot coefficient, The specific heat capacity, The thermal conductivity, The dissipation coefficient, the stress field adopts 8-node hexahedral element, the seepage field adopts 4-node tetrahedral element, the field variable transmission is realized through grid mapping, and the time step is 0.5s;Coupling results are output once every 10 steps;

[0103] The finite element-discrete element coupling algorithm solves the stress field and the seepage field iteration, and the stress field solving formula is: Wherein, The overall stiffness matrix, The node displacement vector, The external load vector, The seepage force vector, the seepage field iteration formula is: Wherein, The first Time water head, The time step, Indicate the permeability coefficient matrix, The first Time flow vector, The water storage rate matrix.

[0104] The improved Moran's I index of the present application is introduced to analyze and identify the deformation concentration area of the slope, the spatial correlation analysis precision is improved by weakening the far field interference, the misjudgment rate is reduced, and the spatial propagation path and time evolution trend of the slope deformation are captured by combining the graph convolution network and the attention mechanism, and the monitoring precision is improved.

[0105] III. Dynamic risk assessment and intelligent early warning strategy: probabilistic inference of multi-source evidence through Bayesian network, and dynamic optimization of early warning threshold based on reinforcement learning algorithm, including the following steps:

[0106] Dynamic risk assessment based on Bayesian network: construct a Bayesian network containing environmental factors, geometric parameters, and real-time monitoring data, and quantify the risk of slope instability through conditional probability inference. The Bayesian network structure includes parent nodes and child nodes. The parent nodes include rainfall intensity and temperature change. The child nodes include abnormal displacement, abnormal stress, and instability events. The Bayesian network inference formula is:

[0107]

[0108] where, represents the slope instability event, and the slope instability event state includes occurrence and non-occurrence, is the evidence set, and the evidence set includes displacement rate, pressure change rate, pore water pressure, and rainfall intensity, represents the instability prior probability, is the probability of the evidence set appearing under the slope instability event.

[0109] Dynamic threshold mechanism: dynamic threshold mechanism is constructed based on sliding window statistics and reinforcement learning algorithm. The threshold includes the baseline threshold and the real-time data statistics. The dynamic threshold formula is:

[0110]

[0111] where, is the early warning threshold at time t, is the baseline threshold, is the historical threshold weight used to balance prior knowledge and real-time data, is the 24-hour sliding window mean, is the sliding window standard deviation, is the correction coefficient, which is optimized by reinforcement learning algorithm to minimize false alarm rate and average early warning delay time. The reinforcement learning algorithm optimization formula is:

[0112]

[0113] where, is the learning rate, is the reward function, and the reward function The formula is:

[0114]

[0115] where, The false alarm rate is the percentage of false alarms out of the total number of alarms. The average warning delay time, The non-warning rate represents the proportion of slope instability events that were not warned in advance.

[0116] This invention uses Bayesian networks to dynamically update the probability of slope instability risk, thereby improving the reliability of decision-making.

[0117] A hybrid time-series prediction model for predicting slope deformation time series is established: The hybrid time-series prediction model adopts an LSTM-Transformer concatenated architecture, which includes LSTM layers and Transformer layers. The LSTM layers are used to extract local features of the deformation sequence. The preferred LSTM layers are two layers, each with 128 hidden units, and the output sequence dimension is 72. The Transformer layers predict long-term deformation trends through a self-attention mechanism. The Transformer layers adopt a 4-head attention mechanism, and the sequence length is 72. The input of the hybrid time-series prediction model includes normalized features of environmental parameters, and the output is the predicted value of the environmental parameters. The network architecture formula of the hybrid time-series prediction model is as follows:

[0118]

[0119]

[0120]

[0121] in, for Input feature vector at any time, Let be the hidden state vector of the LSTM layer, with dimension . , These represent the query, key, and value matrices, respectively. For the first The weight matrix of each attention head. Output projection matrix for multi-head attention.

[0122] This invention solves the problem that the fixed threshold used in traditional systems cannot be dynamically adjusted with changes in the environment by automatically adjusting the warning threshold according to the real-time data distribution through a dynamic threshold mechanism based on reinforcement learning algorithm, thereby reducing the false alarm rate and the missed alarm rate.

[0123] IV. Establishing a Tiered Response Mechanism and System Optimization: A tiered response mechanism triggers different response measures based on risk levels, and model parameters are updated online, including the following steps:

[0124] A hierarchical response mechanism is established: the hierarchical response mechanism includes risk grading and disposal strategy, the risk grading includes low risk level, medium risk level and high risk level, wherein the low risk level includes displacement rate < 1.2-1.5mm / d, stress change rate < 2-3kPa / d, pore water pressure growth rate < 0.1-0.2m / d, and daily data review is carried out, the medium risk level includes displacement rate 15-5mm / d, stress change rate 3-8kPa / d, pore water pressure growth rate 0.2-0.5m / d, when in the medium risk level, the slope drainage system is automatically started, the monitoring frequency is once every 1 hour, the high risk level includes displacement rate > 5mm / d, stress change rate > 8kPa / d, pore water pressure growth rate > 0.5m / d, when in the high risk level, real-time video monitoring is automatically started, and traffic control strategy and personnel evacuation preparation are simultaneously implemented.

[0125] Model parameter optimization based on reinforcement learning algorithm: the model parameters are taken as the decision variables of the reinforcement learning algorithm, the model parameters include the early warning threshold coefficient and the prediction model weight, the false alarm rate, the early warning delay and the un-early warning rate are taken as the evaluation indexes, and the parameter optimization is carried out through the reward function, the model parameter update formula is:

[0126]

[0127] Among them, is the model parameter vector, is the learning rate, is the discount factor, is the state at time t, is the action at time t, is the immediate reward after executing the action .

[0128] Five, system verification: the reliability of the system is verified by multi-dimensional performance test, and the system is verified by benchmark value calibration and joint debugging test, including the following steps:

[0129] Multi-dimensional performance verification: the system performance is verified from the aspects of spatial resolution, timing prediction accuracy and early warning reliability, the spatial resolution is verified by calculating the average of the minimum distance of the sensor nodes to evaluate the spatial coverage capability, 10 nodes are randomly selected, the measured minimum distance is 0.9m, the average monitoring distance is calculated through the spatial resolution verification formula, the average monitoring distance should be less than 0.8m, and the spatial resolution verification formula is as follows:

[0130]

[0131] Among them, is the average monitoring distance, is the node Nearest neighbor node European distance, This represents the total number of sensor nodes.

[0132] The accuracy of time series forecasts is verified by backtesting with historical data and checking the root mean square error of the time series forecast model. The formula for verifying the accuracy of time series forecasts is as follows:

[0133]

[0134] in, These are measured values ​​of environmental parameters. These are the model's predicted values. To verify the sample size.

[0135] The reliability of the early warning system is verified by testing the false alarm rate and response delay through simulated instability events. The verification formula for early warning reliability is as follows:

[0136]

[0137] in, Number of false alarms The total number of warnings, with a verification period of ≥6 months, and <5%.

[0138] Response delay test: Introduce a simulated instability signal into the edge computing node and measure the time from signal generation to warning information issuance. The response delay should be less than 5 seconds.

[0139] Engineering deployment: Engineering deployment includes sensitive area identification, sensor optimization and layout, benchmark calibration, system integration and false alarm suppression. Sensitive area identification includes using ground-penetrating radar to scan the shallow structure of the slope, identifying potential instability areas, and using UAV aerial surveys to obtain slope surface deformation data, delineating high-risk areas as key sensor deployment areas.

[0140] Sensor optimization deployment includes minimizing deployment costs using a tabu search algorithm, the formula of which is:

[0141]

[0142] in, For nodes Deployment costs Indicates the slope monitoring area. For coverage location The set of sensor nodes, The effective monitoring radius of a single sensor.

[0143] The benchmark value calibration includes environmental background monitoring for 7 days after the sensor installation, collecting environmental data under the condition of no significant deformation as the benchmark period data, establishing a multi-dimensional parameter baseline, and the benchmark value statistical formula is:

[0144]

[0145] wherein, is the benchmark period monitoring data matrix, is the benchmark mean value, is the benchmark standard deviation.

[0146] System debugging and false alarm suppression includes simulation test, false alarm suppression and communication test. The simulation test includes injecting different levels of simulation instability signals through the signal generator to verify the accuracy of the early warning level and response measures. The false alarm suppression includes sliding window filtering and multi-parameter cross verification. When a single sensor triggers an early warning, it is verified whether there is an anomaly in the adjacent node data. The communication test includes verifying the data transmission reliability between the sensor node and the edge computing settlement, and the cloud server.

[0147] Engineering maintenance: including inspection, data calibration and model iteration. The inspection includes checking the sensor power supply system, communication link state and sensor shielding condition every week. The data calibration includes zero point calibration of the fiber Bragg grating sensor and the piezoelectric pressure gauge every quarter, updating the benchmark value database, and the model iteration includes retraining the Bayesian network and time series prediction model according to the newly collected monitoring data.

[0148] The above specific embodiments are only several optional embodiments of the present application, and based on the technical solutions of the present application and the related inspiration of the above embodiments, those skilled in the art can make various alternative improvements and combinations on the above specific embodiments.

Claims

1. A design method of a slope protection monitoring and early warning system, characterized in that, The method comprises the following steps: Construction of multi-dimensional perception network for collecting slope data: a multi-dimensional perception network covering the slope is constructed through a sensor network and a space-time synchronization technology; Analysis of space-time characteristics and modeling of multi-physical field coupling: a full-coupling model of seepage-stress-temperature field is established based on spatial autocorrelation analysis and graph convolution network; Dynamic risk assessment and intelligent early warning strategy: probability reasoning of multi-source evidence is performed through a Bayesian network, and the early warning threshold is dynamically optimized based on a reinforcement learning algorithm; Establishment of hierarchical response mechanism and system optimization: different treatment measures are triggered according to the risk level through the hierarchical response mechanism, and the model parameters are updated online; System verification: the reliability of the system is verified through multi-dimensional performance testing, and the system is verified through benchmark value calibration and joint debugging test; The analysis of space-time characteristics and the modeling of multi-physical field coupling Comprise the following steps: Spatial feature analysis and anomaly detection: the spatial correlation of the monitoring parameters is analyzed by using an improved Moran's I index, the deformation concentration area is identified, and a distance attenuation factor is introduced, and the improved Moran's I index formula is as follows: ； wherein, is the number of sensor nodes, is a spatial weight matrix, and , is a distance decay factor, is a node Euclidean distance, of is a monitoring parameter value of a node , is a parameter mean value, and the spatial region is judged by the size of the Moran's I value, including a spatial weak correlation region and a spatial strong correlation region, and local encryption monitoring is performed when the spatial region is determined as the spatial strong correlation region. The space-time dependence relationship between nodes is modeled by a space-time graph convolution network, and the formula is expressed as: ； wherein, is the first layer feature matrix, is the first Chebyshev polynomial basis function of order is the normalized Laplacian matrix, is the first layer trainable parameter matrix of the polynomial of order is an activation function; Multi-physical field dynamic coupling modeling: the coupled equations are solved by a finite element-discrete element coupling algorithm, and the three-level coupled control equations are as follows: ； wherein, is the permeability coefficient as a function of temperature, is the water head, is the water storage rate, is the source / sink term such as rainfall infiltration, is the stress tensor, is the soil density, is the gravitational acceleration, is the displacement vector, is the Biot coefficient, is the specific heat capacity, is the thermal conductivity, is the dissipation coefficient.

2. The method of claim 1, wherein the method further comprises: The construction of multi-dimensional perception network for collecting slope data comprises the following steps: Design of sensor network: the three-dimensional terrain data of the slope is obtained by pre-scanning the surface morphology of the slope by using an unmanned aerial vehicle LiDAR, the optimal spacing of the sensor nodes is calculated based on a Delaunay triangulation algorithm, the sensor nodes are distributed in a honeycomb three-dimensional grid, and each sensor node records three-dimensional coordinates and time information; The sensor comprises an FBG optical fiber grating sensor, a MEMS accelerometer, a piezoelectric pressure gauge and a multi-parameter water quality sensor, wherein the FBG optical fiber grating sensor is arranged along the potential sliding surface of the slope, the MEMS accelerometer is installed at the center position of the honeycomb three-dimensional grid node, the piezoelectric pressure gauge is arranged in the key mechanical response area, and the multi-parameter water quality sensor is arranged in the underground water level monitoring hole; Space-time synchronous acquisition mechanism: the time offset and transmission delay of the sensor are calculated by a bidirectional time synchronization algorithm based on IEEE 1588v2 protocol, the sampling frequency of the sensor comprises low-frequency sampling, high-frequency sampling and 20Hz sampling frequency, the low-frequency sampling is adopted when the slope is in a stable period, the sampling frequency is 5Hz, the high-frequency sampling is adopted when the slope is in a rainfall period, the sampling frequency is 50Hz, and the sampling frequency is 20Hz when the slope is in an early warning state, and the bidirectional time synchronization algorithm formula is as follows: ； wherein, is the master clock time, , , and are the slave clock sending request time, master clock receiving request time, master clock returning response time and slave clock receiving response time, respectively, is the transmission delay error.

3. The method of claim 1, wherein the method further comprises: The finite element-discrete element coupling algorithm solving comprises stress field solving and seepage field iteration, and the stress field solving formula is as follows: ； wherein, is the global stiffness matrix, is the nodal displacement vector, is the external load vector, is the seepage force vector; The seepage field iteration formula is as follows: ； wherein, is the first moment head, is the time step, denotes the permeability matrix, is the first moment flow vector, is the storage rate matrix.

4. The method of claim 1, wherein the method further comprises: The dynamic risk assessment and intelligent early warning strategy comprises the following steps: Dynamic risk assessment based on Bayesian network: a Bayesian network containing environmental factors, geometric parameters, and real-time monitoring data is constructed, and the risk of slope instability is quantified through conditional probability reasoning, and the Bayesian network inference formula is: ； wherein, represents a slope instability event, said slope instability event the states include occurrence and non-occurrence, is a set of evidences, said set of evidences include displacement rate, pressure rate of change, pore water pressure and rainfall intensity, represents a prior probability of instability, is a probability of occurrence of the set of evidences under said slope instability event; Dynamic threshold mechanism is constructed: the dynamic threshold mechanism is constructed based on sliding window statistics and reinforcement learning algorithm, the threshold includes baseline threshold and real-time data statistics, the dynamic threshold formula of early warning threshold in the dynamic threshold mechanism is: ； wherein, is a time alert threshold, is the baseline threshold, is a history threshold weight for balancing prior knowledge and real-time data, is a 24-hour sliding window mean, is a sliding window standard deviation, is a trimming coefficient; The reinforcement learning algorithm is used to optimize the correction coefficient and minimize the false alarm rate and the average early warning delay time, and the reinforcement learning algorithm optimization formula is: ； wherein, is a learning rate, is a reward function, the reward function is expressed by the formula: ； wherein, is a false alarm rate, the false alarm rate being a proportion of the number of false alarms to the total number of alarms, is an average alarm delay time, is an unalarmed rate, the unalarmed rate being a proportion of the slope instability events that were not alerted in advance; A hybrid time series prediction model is established for predicting the time series of slope deformation: the hybrid time series prediction model adopts a LSTM-Transformer series architecture, which includes LSTM layers and Transformer layers, the LSTM layers are used to extract local features of the deformation sequence, and the Transformer layers predict long-term deformation trends through self-attention mechanism, the input of the hybrid time series prediction model includes normalized features of environmental parameters, and the output is the predicted value of environmental parameters, and the network architecture formula of the hybrid time series prediction model is: ； ； ； wherein, is the moment input feature vector, is the LSTM layer hidden state vector, with dimension , respectively represent the query, key, value matrices, is the weight matrix of the th attention head, is the multi-head attention output projection matrix.

5. The method of claim 4, wherein the method further comprises: The establishment of hierarchical response mechanism and system optimization includes the following steps: Establishment of hierarchical response mechanism: the hierarchical response mechanism includes risk classification and disposal strategy, the risk classification includes low risk level, medium risk level and high risk level, wherein the low risk level includes displacement rate <1.2-1.5mm / d, stress change rate <2-3kPa / d, and pore water pressure growth rate <0.1-0.2m / d, and daily 1-2 times data review is carried out, the medium risk level includes displacement rate 15-5mm / d, stress change rate 3-8kPa / d, and pore water pressure growth rate 0.2-0.5m / d, when in the medium risk level, the slope drainage system is automatically started, and the monitoring frequency is once every 1-2 hours, the high risk level includes displacement rate >5mm / d, stress change rate >8kPa / d, and pore water pressure growth rate >0.5m / d, when in high risk level, real-time video monitoring is automatically started, and traffic control strategy and personnel evacuation preparation are implemented synchronously; Model parameter optimization based on reinforcement learning algorithm: the model parameters are taken as the decision variables of the reinforcement learning algorithm, the model parameters include early warning threshold coefficient and prediction model weight, the false alarm rate, early warning delay and un-warning rate are taken as evaluation indexes, and the model parameters are optimized through reward function, and the model parameter update formula is: ； in, For model parameter vectors, For learning rate, As a discount factor, for Current state for Momentary action To perform the action Instant rewards afterwards.

6. The method of claim 2, wherein the method further comprises: The system verification includes the following steps: Multi-dimensional performance verification: the system performance is verified from the aspects of spatial resolution, time series prediction accuracy and early warning reliability; Response delay test: introduce simulated instability signal in edge computing node, measure the time from signal generation to early warning information sending, and the response delay is less than 5 seconds; Engineering deployment: the engineering deployment includes sensitive area identification, sensor optimization arrangement, baseline value calibration and system commissioning and false alarm suppression; Engineering maintenance: including inspection, data calibration and model iteration, the inspection includes checking the sensor power supply system, communication link state and sensor shielding every week, the data calibration includes zero point calibration of the FBG fiber Bragg grating sensor and piezoelectric pressure gauge every quarter, updating the reference value database, the model iteration includes retraining the Bayesian network and time series prediction model according to the newly collected monitoring data.

7. The method of claim 6, wherein the method further comprises: The spatial resolution is verified by calculating the mean value of the minimum spacing of the sensor nodes to evaluate the spatial coverage capability, and the spatial resolution verification formula is as follows: ； wherein, is the average monitoring distance, is the node Euclidean distance to the nearest neighbor node , and is the total number of sensor nodes; The time series prediction accuracy is verified by the root mean square error of the time series prediction model through historical data backtesting, and the time series prediction accuracy verification formula is: ； wherein, is the measured value of the environmental parameter, is the predicted value of the model, is the number of validation samples; The early warning reliability is verified by the false alarm rate and response delay through simulation of instability events, and the early warning reliability verification formula is: ； wherein, is the number of false positives, is the total number of alerts, the verification period is ≥ 6 months, and < 5%.

8. The method of claim 6, wherein the method further comprises: The sensitive area identification includes shallow structure scanning of the slope by using the geological radar to identify potential instability areas, and obtaining the slope surface deformation data by unmanned aerial vehicle aerial survey to delineate high-risk areas as the key deployment area of the sensor; The sensor optimization arrangement includes minimizing the deployment cost by using the tabu search algorithm, and the tabu search algorithm formula is: ； wherein, is the deployment cost of a node is the deployment cost of a node represents a slope monitoring area, is a set of sensor nodes covering a location is a set of sensor nodes covering a location is the effective monitoring radius of a single sensor; The reference value calibration includes continuous 6-10 days of environmental background monitoring after the installation of the sensor, collecting environmental data under the condition of no significant deformation as the reference value data, establishing a multi-dimensional parameter baseline, and the reference value statistical quantity calculation formula is: ； wherein, is the baseline period monitoring data matrix, is the baseline mean, is the baseline standard deviation; The system joint debugging and false alarm suppression includes simulation test, false alarm suppression and communication test.

9. The method of claim 8, wherein the method further comprises: The simulation test includes verifying the accuracy of the early warning level and response measures by injecting different levels of simulated instability signals through the signal generator, the false alarm suppression includes verifying whether there is an anomaly in the data of the adjacent nodes when a single sensor triggers an early warning through sliding window filtering and multi-parameter cross verification, and the communication test includes verifying the data transmission reliability between the sensor nodes and the edge computing settlement, cloud server.

Images