Slope deformation monitoring and dynamic early warning method and system based on multi-sensor data

Abstract

The application discloses a kind of based on multi-sensor data's side slope deformation monitoring and dynamic early warning method and system, it is related to side slope monitoring technical field, including: using pre-deployed multi-class sensor acquisition multi-source sensor data, according to sensor distribution and pre-processed multi-source sensor data constructs graph structure data, using graph convolution network is modeled, constructs side slope deformation monitoring model;Introduce clustering federal learning strategy to carry out joint training to the side slope deformation monitoring model of multiple sites, using the side slope deformation monitoring model trained carries out risk grade division;Extract the key influence factor of landslide disaster and introduce improved firefly algorithm dynamically optimizes early warning threshold, using the early warning threshold optimized and current risk level are distinguished, generate early warning information.The application improves the reliability of monitoring and the timeliness of early warning by multi-source data fusion and intelligent analysis, realizes the leap of side slope deformation monitoring from single point static to networked intelligence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of slope monitoring, more particularly, to a slope deformation monitoring and dynamic early warning method and system based on multi-sensor data. BACKGROUND

[0002] Slope stability monitoring is an important issue in the field of geological disaster prevention, especially in mountainous, mining, and transportation areas. Slope instability can cause landslides, collapses, and other disasters, which seriously threaten people's lives and property safety and the stable operation of major infrastructure. Traditional slope monitoring methods mainly rely on single sensors (such as total station, GNSS, inclinometer, etc.) or manual inspection, which have limited data dimensions, insufficient real-time performance, and low early warning accuracy. Traditional early warning is based on static thresholds or empirical models (such as displacement rate thresholds), which are difficult to adapt to the dynamic evolution process of slopes. Currently, dynamic early warning methods are improved through time series prediction models, multi-modal data fusion, and real-time optimization. However, existing systems still need further breakthroughs in model adaptability in small sample scenarios, uncertainty management of multi-source data, and dynamic adjustment of early warning thresholds.

[0003] To overcome the limitations of single sensors, multi-sensor collaborative monitoring has become a development trend. By integrating multi-dimensional data such as surface displacement, deep displacement, and environmental parameters, the slope state can be more comprehensively described. With the rapid development of sensor technology, Internet of Things (IoT), and Artificial Intelligence (AI), multi-sensor data-based slope deformation monitoring and dynamic early warning have gradually become a research hotspot, aiming to improve the reliability of monitoring and the timeliness of early warning through multi-source data fusion and intelligent analysis. SUMMARY

[0004] To solve the above technical problems, the present application provides a slope deformation monitoring and dynamic early warning method and system based on multi-sensor data, which realizes the leap from single-point static to networked intelligence through the chain of multi-sensor fusion, graph structure modeling, federated learning, and dynamic threshold optimization.

[0005] The present application provides a slope deformation monitoring and dynamic early warning method based on multi-sensor data in the first aspect, comprising the following steps:

[0006] Using pre-deployed multi-class sensors to collect multi-source sensor data, performing spatio-temporal registration and feature extraction on the multi-source sensor data, and obtaining pre-processed multi-source sensor data;

[0007] Constructing graph structure data according to the sensor distribution and the pre-processed multi-source sensor data, modeling the graph structure data using a spatio-temporal graph convolution network, and constructing a slope deformation monitoring model;

[0008] The clustering federation learning strategy is introduced to jointly train the slope deformation monitoring models of multiple sites, and the trained slope deformation monitoring model is used for deformation prediction.

[0009] The key influence factors of landslide disasters are extracted, the improved firefly algorithm is introduced based on the key influence factors to dynamically optimize the early warning threshold, the optimized early warning threshold and the deformation prediction result are used for risk level division, and early warning information is generated.

[0010] In the application, multi-source sensor data is collected by using pre-deployed multi-class sensors, and the multi-source sensor data is spatio-temporally registered and feature extracted to obtain pre-processed multi-source sensor data, specifically:

[0011] Surface displacement monitoring, deep displacement monitoring and environmental parameter monitoring are performed using different sensors to obtain multi-source sensor data, which is transmitted to the edge node cache in real time and subjected to data format standardization processing;

[0012] The sampling frequency of different sensor data is obtained, the highest frequency sensor data is selected as the reference data to generate reference data, the dynamic time warping path is calculated for low frequency data and reference data, the optimal time mapping correlation is obtained, the continuous data is generated by using cubic spline interpolation for the missing period, and the sensor data unified to the same time sequence is output to obtain the time-aligned multi-source sensor data;

[0013] The GNSS parameters and InSAR parameters obtained in the surface displacement monitoring are unified in the coordinate axis, the data fusion is performed through Kalman filter iteration, the GNSS parameters and InSAR parameters are used as observations, the displacement is used as a state variable to construct a state equation, the current displacement is predicted according to the displacement state of the previous period, the predicted value is updated and corrected using the GNSS parameter and InSAR parameter observation value, and the fused displacement field is output to obtain the spatially aligned surface displacement monitoring data;

[0014] The time and space aligned multi-source sensor data is subjected to abnormality detection and missing value processing, the slope deformation monitoring instances are retrieved and obtained, the SHAP value analysis is performed on the slope deformation monitoring instances, the displacement data features and environmental parameter features meeting the preset requirements for slope deformation contribution are screened and stored, and the pre-processed multi-source sensor data is obtained.

[0015] In the application, a graph structure data is constructed according to the sensor distribution and the pre-processed multi-source sensor data, specifically:

[0016] Each sensor is regarded as an independent node, the node attributes of each independent node include displacement parameter features, environmental parameter features and sensor deployment static attributes, and the coordinate information and geological identification are regarded as node metadata;

[0017] Calculate the three-dimensional Euclidean distance between all sensors, determine whether the three-dimensional Euclidean distance is greater than a preset distance threshold, if greater, establish a spatial distance connection between nodes, and set the edge weight according to the distance;

[0018] According to the geological exploration data, mark the rock layer to which the sensor belongs, if two sensors are located in the same rock layer and the rock layer inclination is consistent, a geomechanics connection between nodes is established, if the sensors are located in different rock layers but there is a weak interlayer, a cross-layer geomechanics connection between nodes is established, and the edge weight is set according to the rock layer shear strength;

[0019] According to the sensor nodes and edge weights, an adjacency matrix and a feature matrix are constructed, and the adjacency matrix and the feature matrix are used as graph structure data.

[0020] In the application, a spatio-temporal graph convolution network is used to model the graph structure data to construct a slope deformation monitoring model, specifically:

[0021] The continuous multi-source sensor data is divided into several windows using a time sliding window, each window contains several graph structure data snapshots, the node features are updated according to the latest multi-source sensor data, and the edge weight is updated according to the displacement correlation mutation between sensor nodes, and a dynamic graph sequence is generated through the updated node features and edge weight;

[0022] According to the dynamic graph sequence, the graph structure data snapshot at the current time and the updated graph structure data snapshot at the last time are obtained, and an update gating mechanism is introduced to obtain the fusion ratio of the graph structure data snapshot at the current time and the updated graph structure data snapshot at the last time;

[0023] Based on the fusion ratio, the adjacency matrix is dynamically updated, the current adjacency matrix and the updated adjacency matrix are fused to generate the final dynamic graph structure data, and the spatio-temporal graph convolution network is used to model the final dynamic graph structure data to construct a slope deformation monitoring model;

[0024] The graph convolution network is used to model the spatial dependence of the dynamic graph structure data at each time, the neighborhood node information is aggregated through dynamic graph convolution to obtain the node feature representation at each time, and the spatial features are generated through multi-layer dynamic graph convolution with residual connection;

[0025] The node feature representation sequence at all times output by the graph convolution network is obtained, the time dependence is modeled through multi-head attention, the different time scale dependence of a single attention head is calculated through the correlation between the current time and the historical time of the node, and the time features are obtained through full connection layer fusion after splicing multi-head attention;

[0026] The spatial features and time features are fused and imported into the prediction head for prediction, and the slope deformation monitoring model is initially trained through the preset training data.

[0027] In the present application, the clustering federated learning strategy is introduced to jointly train the slope deformation monitoring models of multiple sites, and the trained slope deformation monitoring model is used for deformation prediction, specifically:

[0028] Each site trains an independent slope deformation monitoring model based on local data, selects geological environmental features related to slope stability to obtain the geological conditions of each site, and obtains the geological similarity between sites according to the geological conditions, and uses the geological similarity to divide the sites into several clusters;

[0029] Each site in the cluster uploads the model parameters of the slope deformation monitoring model to the cloud server after encryption, assigns weights according to the data volume of each site, performs weighted average aggregation, and distributes the aggregated parameters to all sites in the cluster;

[0030] According to the preset period, perform federated aggregation, when there is a site offline, skip its participation in the current round of aggregation, each site loads the global model of the cluster, and obtains the displacement increment of the future preset time period as the deformation prediction result.

[0031] In the present application, the key influencing factors of landslide disasters are extracted, and an improved firefly algorithm is introduced based on the key influencing factors to dynamically optimize the warning threshold, specifically:

[0032] The displacement data features and environmental parameter features corresponding to the multi-source sensor data are combined with the geological environmental features as candidate influencing factors, and the historical landslide event label is used as the target variable, a time delay variable is introduced in the candidate influencing factors, and the maximum mutual information time delay of each candidate influencing factor and displacement is calculated.

[0033] The maximum information coefficient is used to enhance the mRmR algorithm for nonlinear redundancy evaluation, the feature importance of the candidate influencing factors is calculated by the enhanced mRmR algorithm, the feature importance is sorted, and the preset number of key influencing factors is selected according to the sorting result;

[0034] The current parameter value of the key influencing factor is introduced as a weight adjustment factor to minimize the false alarm rate and the false negative rate for multi-objective optimization, the weight adjustment factor is used to weight the objective function, and the fitness function of the firefly algorithm is constructed by the inverse of the weighted objective function;

[0035] The firefly population is initialized according to the warning threshold constraint range, the brightness of each firefly is calculated using the fitness function, the adaptive step size decreasing with the iteration coefficient and the random disturbance vector are introduced to update the firefly position, so that the firefly moves to a brighter neighbor;

[0036] After iteration to convergence, the firefly with the highest brightness is selected as the current dynamic warning threshold.

[0037] In the present application, the optimized early warning threshold is used to divide the risk level with the deformation prediction result, and generate early warning information, specifically:

[0038] The deformation prediction result of a future preset time period is obtained, the deformation prediction result is compared with the current early warning threshold, and early warning information of different risk levels is generated according to the comparison result;

[0039] When the early warning information exceeds the preset risk level, multi-modal verification is triggered, the consistency of the deformation prediction result and the measured displacement trend is detected, and similarity calculation is performed according to the current environmental parameter characteristics, and historical landslide events with similar similarity are retrieved;

[0040] When the consistency verification and the number of historical landslide events meet the preset multi-modal verification standard, early warning information of the corresponding risk level is generated.

[0041] The second aspect of the present application provides a slope deformation monitoring and dynamic early warning system based on multi-sensor data, which comprises: a data acquisition and preprocessing module, a graph structure modeling and intelligent analysis module, a federated learning and model optimization module, a dynamic early warning and decision support module;

[0042] The data acquisition and preprocessing module uses pre-deployed multi-class sensors to collect multi-source sensor data, performs spatio-temporal registration and feature extraction on the multi-source sensor data, and obtains pre-processed multi-source sensor data;

[0043] The graph structure modeling and intelligent analysis module constructs graph structure data according to the sensor distribution and the pre-processed multi-source sensor data, models the graph structure data using a spatio-temporal graph convolution network, constructs a slope deformation monitoring model, and realizes deformation prediction;

[0044] The federated learning and model optimization module introduces a clustering federated learning strategy to jointly train slope deformation monitoring models of multiple sites, improves the model generalization ability, and protects data privacy;

[0045] The dynamic early warning and decision support module extracts key influencing factors of landslide disasters, introduces an improved firefly algorithm to dynamically optimize the early warning threshold based on the key influencing factors, uses the optimized early warning threshold to divide the risk level with the deformation prediction result, and generates early warning information.

[0046] Compared with the prior art, the present application has the following advantages:

[0047] The present application constructs a high-precision, self-adaptive and scalable slope deformation monitoring and early warning system through multi-sensor data fusion, graph convolution network modeling, clustering federated learning and dynamic threshold optimization. The multi-source data deep fusion is used to improve the monitoring accuracy and robustness, the key features such as displacement rate, strain distribution and environmental lag effect are extracted from the original data, and the graph structure data modeling is combined to more comprehensively represent the spatial correlation of slope deformation. The graph convolution network modeling is used to capture the spatial correlation of slope deformation, and to learn the local single sensor data and the global landslide surface sliding trend, thereby avoiding the limitation of traditional point substitution method. The clustering federated learning strategy is introduced to realize multi-site collaborative modeling and privacy protection, and the slope monitoring sites with similar geological conditions are divided into the same cluster, and the model is trained in the cluster to avoid the overfitting problem caused by insufficient data of a single site. The dynamic threshold optimization and hierarchical early warning reduce the false alarm and missed alarm rate, and the hierarchical early warning mechanism considers the real-time and long-term trend prediction. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0049] Figure 1 A flowchart of slope deformation monitoring and dynamic early warning based on multi-sensor data is shown.

[0050] Figure 2 A flowchart of constructing a slope deformation monitoring model in the embodiment is shown.

[0051] Figure 3 A flowchart of dynamically optimizing the early warning threshold in the embodiment is shown.

[0052] Figure 4 A block diagram of a slope deformation monitoring and dynamic early warning system based on multi-sensor data is shown. DETAILED DESCRIPTION

[0053] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or examples will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0054] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0055] Figure 1 A flow chart of slope deformation monitoring and dynamic early warning based on multi-sensor data is shown.

[0056] As Figure 1 shown, the embodiment provides a slope deformation monitoring and dynamic early warning based on multi-sensor data, comprising:

[0057] S102, using pre-deployed multi-class sensors to collect multi-source sensor data, performing spatio-temporal registration and feature extraction on the multi-source sensor data, and obtaining pre-processed multi-source sensor data;

[0058] S104, constructing graph structure data according to the sensor distribution and the pre-processed multi-source sensor data, modeling the graph structure data using a spatio-temporal graph convolution network, and constructing a slope deformation monitoring model;

[0059] S106, introducing a clustering federated learning strategy to jointly train slope deformation monitoring models of multiple sites, and using the trained slope deformation monitoring model to perform deformation prediction;

[0060] S108, extracting key influence factors of landslide disasters, introducing an improved firefly algorithm to dynamically optimize the early warning threshold based on the key influence factors, using the optimized early warning threshold and the deformation prediction result to divide the risk level, and generating early warning information.

[0061] It should be noted that different sensors are used for ground surface displacement monitoring, deep displacement monitoring and environmental parameter monitoring. The sensor types for ground surface displacement monitoring include GNSS, InSAR, crack meter, etc. The sensor types for deep displacement monitoring include inclinometer, optical fiber sensor, etc. The sensor types for environmental parameter monitoring include rain gauge, underground water level gauge, vibration sensor, etc. The multi-source sensor data is obtained, the multi-source sensor data is transmitted to the edge node cache in real time through 4G / 5G / LoRa network, and data format standardization processing is performed.

[0062] The sampling frequencies of different sensors are different, and the time axis needs to be aligned for fusion analysis. The sampling frequencies of different sensor data are obtained, the highest frequency sensor data is selected as the reference data of the reference time axis, the dynamic time warping path is calculated for the low frequency data and the reference data, the optimal time mapping correlation is obtained, the missing period is generated by cubic spline interpolation to generate continuous data, the sensor data unified to the same time sequence is output, the time-aligned multi-source sensor data is obtained, and the time sequence consistency is ensured; the GNSS parameters and InSAR parameters obtained in the ground surface displacement monitoring are unified in the coordinate axis, and are converted to the same local coordinate system. Data fusion is performed through Kalman filter iteration, it is assumed that the slope displacement is subject to a linear dynamic model, the GNSS parameters and InSAR parameters are used as observation values, the displacement is used as a state variable to construct a state equation, the displacement in the current period is predicted according to the displacement state in the last period, the predicted value is updated and corrected by using the GNSS parameters and InSAR parameters, the fused displacement field is output, and the spatially-aligned ground surface displacement monitoring data is obtained; the time and space-aligned multi-source sensor data is subjected to abnormality detection and missing value processing, slope deformation monitoring instances are retrieved and obtained, SHAP value analysis is performed on the slope deformation monitoring instances, displacement data features and environmental parameter features that meet the preset requirements for slope deformation contribution are screened and stored, and the preprocessed multi-source sensor data is obtained. The preferred displacement data features include cumulative displacement, displacement rate, acceleration and spatial gradient, and the environmental data features include rainfall lag effect, underground water level mutation point and vibration energy, the structured features are stored in a time sequence database, and are called by a model in real time, and the unstructured data is stored in an object storage system.

[0063] It should be noted that each sensor is taken as an independent node, the node attribute of each independent node includes displacement parameter characteristics, environmental parameter characteristics and sensor deployment static attribute, the sensor deployment static attribute includes sensor type, installation depth, rock mass type, and the latitude and longitude coordinate information and geological identification of sensor layout are taken as node metadata; the three-dimensional Euclidean distance between all sensors is calculated, whether the three-dimensional Euclidean distance is greater than a preset distance threshold is judged, if greater than, the spatial distance connection between nodes is established, and the edge weight is set according to the distance, the closer the distance, the higher the edge weight; the rock layer (such as the upper loose accumulation body and the lower bedrock) to which the sensor belongs is marked according to the geological exploration data (drilling, geological profile), if two sensors are located in the same rock layer and the rock layer inclination is consistent, the geomechanics connection between nodes is established, reflecting the overall sliding trend of the rock mass; if the sensors are located in different rock layers but there is a weak interlayer, the cross-layer geomechanics connection between nodes is established, reflecting the interlayer sliding risk, and the edge weight is set according to the rock layer shear strength; the adjacency matrix and the feature matrix are constructed according to the sensor nodes and the edge weight, and the adjacency matrix and the feature matrix are used as the graph structure data. In the adjacency matrix, the row and the column represent the sensor nodes, and the matrix element value is the edge weight. In the feature matrix, each row represents a sensor node, and each column represents a feature.

[0064] Figure 2 A flowchart for constructing a slope deformation monitoring model in the embodiment is shown.

[0065] According to the embodiment of the application, the graph structure data is modeled using a spatio-temporal graph convolution network to construct a slope deformation monitoring model, specifically:

[0066] S202, using a time sliding window, the continuous multi-source sensor data is divided into a plurality of windows, each window contains a plurality of graph structure data snapshots, the node features are updated according to the latest multi-source sensor data, and the edge weight is updated according to the displacement correlation mutation between sensor nodes, and a dynamic graph sequence is generated through the updated node features and edge weight;

[0067] S204, according to the dynamic graph sequence, the graph structure data snapshot at the current time and the graph structure data snapshot updated at the last time are obtained, and an update gating mechanism is introduced to obtain the fusion ratio of the graph structure data snapshot at the current time and the graph structure data snapshot updated at the last time;

[0068] S206, based on the fusion ratio, the adjacency matrix is dynamically updated, the current adjacency matrix and the updated adjacency matrix are fused to generate the final dynamic graph structure data, and the spatio-temporal graph convolution network is used to model the final dynamic graph structure data to construct a slope deformation monitoring model;

[0069] S208 uses a graph convolutional network to model the spatial dependency of the dynamic graph structure data at each time step. It aggregates neighborhood node information through dynamic graph convolution to obtain the node feature representation at each time step, and generates spatial features through multi-layer dynamic graph convolution with residual connections.

[0070] S210: Obtain the node feature representation sequence of all time steps output by the graph convolutional network, model the time dependency through multi-head attention, calculate the different time scale dependencies of a single attention head by the correlation between the current time and the historical time of the node, and obtain the time features by splicing the multi-head attention and fusing them through a fully connected layer.

[0071] S212, the spatial and temporal features are fused and imported into the prediction head for prediction, and the slope deformation monitoring model is initially trained using preset training data.

[0072] It should be noted that the preprocessed multi-source sensor data is input into a time sliding window, the time interval of which is set according to the monitoring requirements, and the data within the window includes the current time. and before Graph data at time step (node ​​feature matrix) Adjacency matrix In dynamic graph updates, an update gating mechanism is introduced. Get a snapshot of the graph structure data at the current moment. and the graph structure data snapshot updated at the previous moment The fusion ratio; , This represents the activation function. , The parameters represent learnable parameters, and the adjacency matrix is ​​dynamically updated based on the fusion ratio. , This represents element-wise multiplication, which merges the current adjacency matrix and the updated adjacency matrix to generate the final dynamic graph structure data. , , , Hyperparameters representing the current observations and historical states.

[0073] In the spatio-temporal attention mechanism, the spatial dependency is modeled by dynamic graph convolution operation, the graph convolution is performed on the dynamic graph structure data of each time, the neighborhood node information is aggregated, and the residual connection is introduced to prevent the gradient vanishing of the deep network. The multi-head attention is used to model the time dependency, while the ordinary attention may implicitly use the future time data, which is not suitable for real-time warning scenarios. Preferably, the one-dimensional causal convolution with a kernel size of k is used to generate the attention parameters Q, K and V, so as to force the model to only rely on historical information while aggregating local context. The attention weight is calculated based on the local sequence segment of the convolution output, so as to enhance the perception ability of the adjacent trend. In the local similarity calculation, for each time, the local window (length of ) of Q and K is taken to calculate the similarity , the overall displacement trend (such as “continuous acceleration” and “smooth peristalsis”) of two time periods is compared, rather than the instantaneous value, which is expressed as:

[0074]

[0075] wherein 、 represent the current time and the historical time, represents the relative displacement in the sliding window, which is used to traverse each time step in the local window, represents the query vector of the time , which contains the displacement trend features of the time and the previous k-1 times, represents the key vector transpose of the time , which reflects the deformation mode of the historical local window;

[0076] The lower triangular mask matrix is applied before Softmax , so as to force , and the attention weight is expressed as:

[0077]

[0078] wherein, represents the key vector dimension scaling factor, represents the value matrix.

[0079] Different causal convolution kernels are used in each attention head to capture multi-scale time patterns, and multiple attention heads are run in parallel, and the local attention outputs of the heads are spliced and then uniformly mapped through a fully connected layer. By comparing the displacement trend segments with a length of k, the single-point noise interference is avoided, the multi-head design with different convolution kernel sizes takes into account the sudden events and long-term evolution, and the causal mask is introduced to strictly follow the principle of predicting the future from the past, which meets the real-time requirement of slope monitoring.

[0080] Each site deploys the same slope deformation monitoring model, including dynamic graph updating, spatio-temporal attention, and prediction head module, inputs the graph structure data of the site to train the local model, and each site trains an independent slope deformation monitoring model based on local data. Select the geological conditions related to the slope stability of each site, such as rock type, slope, hydrological condition and historical deformation mode, etc., use K-means clustering to obtain the geological similarity between sites according to the geological conditions, and use the geological similarity to divide the sites into several clusters; when a new site is added, calculate the Euclidean distance between it and each cluster center, and belong to the nearest cluster. The slope deformation monitoring model in each site uploads the encrypted model parameters to the cloud server, assigns weights according to the data volume of each site, performs weighted average aggregation, and distributes the aggregated parameters to all sites in the cluster; according to the preset period, perform federated aggregation, when there is a site offline, skip its participation in the current round of aggregation, each site loads the global model of the cluster, and obtains the displacement increment of the future preset time period as the deformation prediction result. Through clustering federated learning, the generalization ability and early warning accuracy of the slope deformation monitoring model are significantly improved under the premise of protecting data privacy, and different geological clusters have customized models to avoid performance degradation caused by one-size-fits-all.

[0081] Figure 3 A flowchart of dynamically optimizing the warning threshold in the embodiment is shown.

[0082] According to the embodiment of the present application, the key influencing factors of landslide disaster are extracted, and the improved glowworm algorithm is introduced to dynamically optimize the warning threshold based on the key influencing factors, specifically:

[0083] S302, combine the displacement data features and environmental parameter features corresponding to the multi-source sensor data with the geological environment features as candidate influencing factors, and use the historical landslide event label as the target variable, introduce a time lag variable in the candidate influencing factors, and calculate the maximum mutual information time lag of each candidate influencing factor and displacement;

[0084] S304, enhance the mRmR algorithm using the maximum information coefficient, perform nonlinear redundancy evaluation, calculate the feature importance of the candidate influencing factors through the enhanced mRmR algorithm, sort according to the feature importance, and select a preset number of key influencing factors according to the sorting result;

[0085] S306, introduce the current parameter value of the key influencing factor as a weight adjustment factor to minimize the false positive rate and the false negative rate for multi-objective optimization, use the weight adjustment factor to weight the objective function, and construct the fitness function of the glowworm algorithm through the inverse of the weighted objective function;

[0086] S308, initializing the glowworm population according to the early warning threshold constraint range, calculating the brightness of each glowworm using the fitness function, introducing an adaptive step size decreasing with the iteration coefficient and a random disturbance vector to update the glowworm position, making the glowworm move to a brighter neighbor, iterating to convergence, and selecting the glowworm with the highest brightness as the current dynamic early warning threshold.

[0087] It should be noted that the traditional mRMR selects features by maximizing the correlation between features and target and minimizing the redundancy between features. The introduction of lag correlation analysis enhances its applicability. The time lag mutual information between each candidate influencing factor and the target variable is calculated. For environmental factors such as rainfall and water level, the maximum mutual information time lag between the candidate influencing factor and the displacement is calculated, for example, the displacement response is strongest 48 hours after the rainfall. The redundancy between the candidate influencing factor and the selected feature set is calculated, and the feature with low correlation to the selected feature is preferentially selected. The comprehensive score of the candidate influencing factor is obtained as the feature importance with the goal of maximum correlation and minimum redundancy. According to the feature importance ranking result, the Top5 key influencing factors are selected, and the displacement acceleration (lag 0 hours), cumulative rainfall (lag 24 hours), groundwater level mutation (lag 12 hours), and rock-soil type (static) are preferentially selected as key influencing factors.

[0088] The improved glowworm algorithm is used to dynamically optimize the threshold, and the false positive rate and the false negative rate are minimized as optimization objectives. The current parameter value of the key influencing factor is introduced as a weight adjustment factor, and the fitness function is inversely proportional to the weighted objective function , which is expressed as: , represents the false positive rate, represents the false negative rate, represents the weight adjustment factor, represents the glowworm individual. In the iteration process, the glowworm moves to a brighter glowworm, and the threshold is dynamically adjusted according to the environmental parameters. If the current rainfall exceeds the threshold, the search range is reduced. After iteration to convergence, the glowworm with the highest brightness is selected as the current dynamic threshold. For example, during continuous rainfall, the algorithm automatically reduces the early warning threshold to provide early warning for shallow landslides. The glowworm algorithm adjusts the threshold in combination with real-time environmental data, reduces seasonal or sudden disturbances, and avoids the false positive and false negative imbalance problems of traditional threshold methods through multi-objective optimization.

[0089] Obtain the deformation prediction result of the future preset time period, compare the deformation prediction result with the current early warning threshold, and generate early warning information of different risk levels according to the comparison result; when the early warning information exceeds the preset risk level, trigger multi-modal verification to improve the early warning credibility. The consistency of the deformation prediction result and the measured displacement trend is detected, and the similarity is calculated according to the current environmental parameter characteristics, and the historical landslide events with similar similarity are retrieved; when the consistency verification and the number of historical landslide events meet the preset multi-modal verification standard, the early warning information of the corresponding risk level is generated, otherwise, the artificial review is triggered, and the early warning information is visually displayed.

[0090] Figure 4 A slope deformation monitoring and dynamic early warning system based on multi-sensor data is shown.

[0091] The second embodiment of the application provides a slope deformation monitoring and dynamic early warning system based on multi-sensor data 4, which comprises a data acquisition and preprocessing module 401, a graph structure modeling and intelligent analysis module 402, a federated learning and model optimization module 403, and a dynamic early warning and decision support module 404.

[0092] The data acquisition and preprocessing module uses a pre-deployed multi-class sensor to collect multi-source sensor data, performs spatio-temporal registration and feature extraction on the multi-source sensor data, and obtains pre-processed multi-source sensor data.

[0093] The graph structure modeling and intelligent analysis module constructs graph structure data according to the sensor distribution and the pre-processed multi-source sensor data, models the graph structure data using a spatio-temporal graph convolution network, constructs a slope deformation monitoring model, and realizes deformation prediction.

[0094] The federated learning and model optimization module introduces a clustering federated learning strategy to jointly train slope deformation monitoring models of multiple sites, improves the model generalization ability, and protects data privacy.

[0095] The dynamic early warning and decision support module extracts key influencing factors of landslide disasters, introduces an improved firefly algorithm to dynamically optimize the early warning threshold based on the key influencing factors, uses the optimized early warning threshold and the deformation prediction result to divide the risk level, and generates early warning information.

[0096] The third aspect of the application provides a computer readable storage medium, wherein the computer readable storage medium comprises a slope deformation monitoring and dynamic early warning method based on multi-sensor data program, and the slope deformation monitoring and dynamic early warning method based on multi-sensor data program is executed by a processor to realize the steps of the slope deformation monitoring and dynamic early warning method based on multi-sensor data.

[0097] In several embodiments provided in the present application, it should be understood that the disclosed method and system can be implemented in other ways. The above-described system embodiments are only illustrative, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed components can be indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms. In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can be a separate module, or two or more modules can be integrated into one module; the integrated module can be realized in the form of hardware or in the form of hardware plus software functional module.

[0098] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, the above-mentioned program can be stored in a computer readable storage medium, and the program executes the steps including the above-mentioned method embodiments when executed; and the above-mentioned storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various storage medium capable of storing program codes.

[0099] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for slope deformation monitoring and dynamic early warning based on multi-sensor data, characterized in that, The method comprises the following steps: Collecting multi-source sensor data using pre-deployed multi-class sensors, performing space-time registration and feature extraction on the multi-source sensor data, and obtaining pre-processed multi-source sensor data; Constructing graph structure data according to sensor distribution and pre-processed multi-source sensor data, modeling the graph structure data using a space-time graph convolution network, and constructing a slope deformation monitoring model; Introducing a clustering federated learning strategy to jointly train slope deformation monitoring models of multiple sites, and using the trained slope deformation monitoring model to perform deformation prediction; Extracting key influencing factors of landslide disasters, introducing an improved firefly algorithm to dynamically optimize early warning thresholds based on the key influencing factors, dividing risk levels using the optimized early warning thresholds and deformation prediction results, and generating early warning information; Constructing graph structure data according to sensor distribution and pre-processed multi-source sensor data, specifically: Taking each sensor as an independent node, the node attributes of each independent node including displacement parameter features, environmental parameter features and sensor deployment static attributes, and taking coordinate information and geological identification as node metadata; Calculating the three-dimensional Euclidean distance between all sensors, determining whether the three-dimensional Euclidean distance is greater than a preset distance threshold, if greater, establishing a spatial distance connection between nodes, and setting edge weights according to the distance; According to the geological exploration data, the rock layer to which the sensor belongs is marked, if two sensors are located in the same rock layer and the rock layer inclination is consistent, the geomechanics connection between nodes is established, if the sensors are located in different rock layers but there is a soft interlayer, the cross-layer geomechanics connection between nodes is established, and the edge weight is set according to the rock layer shear strength; According to the sensor nodes and edge weights, an adjacency matrix and a feature matrix are constructed, and the adjacency matrix and the feature matrix are used as graph structure data; Modeling the graph structure data using a space-time graph convolution network to construct a slope deformation monitoring model, specifically: Using a time sliding window to divide continuous multi-source sensor data into several windows, each window containing several graph structure data snapshots, updating node features according to the latest multi-source sensor data, and updating edge weights according to the displacement correlation mutation between sensor nodes, and generating a dynamic graph sequence through the updated node features and edge weights; According to the dynamic graph sequence, the graph structure data snapshot at the current time and the updated graph structure data snapshot at the last time are obtained, and an update gate mechanism is introduced to obtain the fusion ratio of the graph structure data snapshot at the current time and the updated graph structure data snapshot at the last time; Based on the fusion ratio, the adjacency matrix is dynamically updated, the current adjacency matrix and the updated adjacency matrix are fused to generate the final dynamic graph structure data, and the space-time graph convolution network is used to model the final dynamic graph structure data to construct a slope deformation monitoring model; Using a graph convolution network to model the spatial dependence of dynamic graph structure data at each time, aggregating neighborhood node information through dynamic graph convolution to obtain node feature representation at each time, and generating spatial features through multi-layer dynamic graph convolution with residual connection; The node feature representation sequence of all time points output by the graph convolution network is obtained, the time dependence is modeled through multi-head attention, the different time scale dependence of a single attention head is obtained through the correlation calculation of the current time and the historical time of the node, and the time features are obtained through the full connection layer fusion after splicing the multi-head attention; The spatial features and the time features are fused and introduced into a prediction head for prediction, and the initial training of the slope deformation monitoring model is performed through preset training data; A clustering federated learning strategy is introduced to jointly train the slope deformation monitoring models of multiple sites, and the trained slope deformation monitoring model is used for deformation prediction, specifically as follows: Each site trains an independent slope deformation monitoring model based on local data, selects geological environmental features related to slope stability to obtain the geological conditions of each site, obtains the geological similarity between sites according to the geological conditions, and divides the sites into several clusters using the geological similarity; The sites in the cluster upload the model parameters of the slope deformation monitoring model to the cloud server after encryption, assign weights according to the data volume of each site, perform weighted average aggregation, and distribute the aggregated parameters to all sites in the cluster; According to a preset period, perform federated aggregation, when there is an offline site, skip its participation in the current round of aggregation, each site loads the global model of the cluster, and obtains the displacement increment of a future preset time period as the deformation prediction result; Key influencing factors of landslide disasters are extracted, and an improved firefly algorithm is introduced to dynamically optimize the warning threshold based on the key influencing factors, specifically as follows: The displacement data features and environmental parameter features corresponding to the multi-source sensor data are combined with the geological environmental features as candidate influencing factors, and the historical landslide event labels are used as target variables, a time lag variable is introduced in the candidate influencing factors, and the maximum mutual information time lag of each candidate influencing factor and displacement is calculated; The maximum information coefficient is used to enhance the mRmR algorithm, perform nonlinear redundancy evaluation, calculate the feature importance of the candidate influencing factors through the enhanced mRmR algorithm, sort according to the feature importance, and select a preset number of key influencing factors according to the sorting result; The current parameter value of the key influencing factor is introduced as a weight adjustment factor to minimize the false alarm rate and the missed alarm rate for multi-objective optimization, the weight adjustment factor is used to weight the objective function, and the fitness function of the firefly algorithm is constructed through the inverse of the weighted objective function; The firefly population is initialized according to the warning threshold constraint range, the brightness of each firefly is calculated using the fitness function, an adaptive step size decreasing with the iteration coefficient and a random disturbance vector are introduced to update the firefly position, and the firefly moves to a brighter neighbor; After iteration to convergence, the brightest firefly is selected as the current dynamic warning threshold.

2. The multi-sensor data based slope deformation monitoring and dynamic early warning method according to claim 1, characterized in that, Pre-deployed multi-class sensors are used to collect multi-source sensor data, and the multi-source sensor data is subjected to spatio-temporal registration and feature extraction to obtain pre-processed multi-source sensor data, specifically as follows: Different sensors are used for surface displacement monitoring, deep displacement monitoring and environmental parameter monitoring to obtain multi-source sensor data, which is transmitted to an edge node cache in real time and subjected to data format standardization processing; The sampling frequencies of different sensor data are acquired, the sensor data with the highest frequency is selected as a reference time axis to generate reference data, a dynamic time warping path is calculated for low-frequency data and the reference data, an optimal time mapping correlation is acquired, continuous data is generated by using cubic spline interpolation for a missing period, sensor data unified to the same time sequence is output, and multi-source sensor data after time alignment is acquired; GNSS parameters and InSAR parameters acquired in surface displacement monitoring are unified in coordinate axes, data fusion is performed through Kalman filter iteration, GNSS parameters and InSAR parameters are used as observation values, a state equation is constructed by taking displacement as a state variable, the current displacement is predicted according to the displacement state of the previous period, the predicted value is updated and corrected by using the GNSS parameter and InSAR parameter observation values, a fused displacement field is output, and spatially aligned surface displacement monitoring data is acquired; After the multi-source sensor data after time and space alignment is subjected to abnormality detection and missing value processing, a slope deformation monitoring instance is retrieved and acquired, SHAP value analysis is performed on the slope deformation monitoring instance, displacement data features and environmental parameter features that meet preset requirements are screened and stored, and preprocessed multi-source sensor data is acquired. 3.The method of claim 1, wherein, The optimized early warning threshold and the deformation prediction result are used for risk level division, and early warning information is generated, specifically as follows: The deformation prediction result of a future preset time period is acquired, the deformation prediction result is compared with the current early warning threshold, and early warning information of different risk levels is generated according to the comparison result; When the early warning information exceeds a preset risk level, multi-modal verification is triggered, the consistency of the deformation prediction result and the measured displacement trend is detected, and similarity calculation is performed according to the current environmental parameter feature, and a historical landslide event with a similar similarity is retrieved; When the consistency verification and the number of historical landslide events meet the preset multi-modal verification standard, early warning information of the corresponding risk level is generated.

4. A slope deformation monitoring and dynamic warning system based on multi-sensor data, realizing the slope deformation monitoring and dynamic warning method based on multi-sensor data according to any one of claims 1-3, the system comprising: The data acquisition and preprocessing module, the graph structure modeling and intelligent analysis module, the federated learning and model optimization module, and the dynamic early warning and decision support module; The data acquisition and preprocessing module uses pre-deployed multi-class sensors to collect multi-source sensor data, performs time and space registration and feature extraction on the multi-source sensor data, and acquires preprocessed multi-source sensor data; The graph structure modeling and intelligent analysis module constructs graph structure data according to sensor distribution and preprocessed multi-source sensor data, models the graph structure data using a spatio-temporal graph convolution network, constructs a slope deformation monitoring model, and realizes deformation prediction; The federated learning and model optimization module introduces a clustering federated learning strategy to jointly train slope deformation monitoring models of multiple sites, improves the model generalization ability, and protects data privacy; The dynamic early warning and decision support module extracts key influencing factors of landslide disasters, introduces an improved firefly algorithm to dynamically optimize the early warning threshold based on the key influencing factors, uses the optimized early warning threshold and the deformation prediction result for risk level division, and generates early warning information.

Images