Glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion

Abstract

The present invention discloses a glacial lake burst monitoring and early warning method based on artificial intelligence and multi-source data fusion, which belongs to the field of natural disaster monitoring technology. The method specifically includes: extracting glacial lake state feature vectors from multi-temporal and spatial data through cross-modal self-supervised learning and constructing a three-dimensional virtual glacial lake model; based on the feature vectors, using the finite element method to construct a three-dimensional stress field model of the glacier and a graph neural network to construct a seepage channel graph model, and fusing the characteristics of the two to form a mechanical-seepage coupling state vector; constructing a pulse neural network to simulate glacial crack expansion and seepage mutation pulse events; constructing a heterogeneous graph federation architecture and combining it with a dynamic weighted aggregation strategy to output a federation weight matrix; fusing the federation learning weights and the three-dimensional model output results, calculating the posterior probability of burst through a Bayesian neural network, and generating a four-level early warning signal; the present invention realizes multi-dimensional characterization of glacial lake state and coupling process simulation, thereby improving the accuracy and timeliness of early warning.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural disaster monitoring, and in particular to a glacial lake feedback monitoring and early warning method based on the fusion of artificial intelligence and multi-source data. Background Art

[0002] As global warming accelerates glacier melting, the number and area of ​​glacial lakes continue to increase, and the risk of glacial lake outbursts (GLBs) is also increasing. These outbursts often trigger secondary disasters such as floods and debris flows, posing a serious threat to the safety of life, property, and the ecological environment in downstream areas. Traditional GLB monitoring methods rely primarily on in-situ sensors and remote sensing image analysis, which suffer from long data acquisition cycles, delayed warnings, and poor model adaptability. Furthermore, existing technologies struggle to effectively integrate heterogeneous multi-source data, resulting in inaccurate assessments of GLB risk. Furthermore, traditional early warning models lack the ability to self-update, making it impossible to adjust warning strategies based on real-time data changes. Consequently, they are unable to accurately simulate GLB dynamics and provide early warnings. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes a glacier burst monitoring and early warning method based on artificial intelligence and multi-source data fusion. Through cross-modal self-supervised learning, the glacier state characteristic vectors of multi-temporal and spatial data are extracted and a three-dimensional virtual glacier lake model is constructed; based on the characteristic vectors, the finite element method is used to construct a three-dimensional stress field model of the glacier, and the graph neural network is used to construct a seepage channel graph model, and the characteristics of the two are integrated to form a mechanical-seepage coupling state vector; a pulse neural network is constructed to simulate glacier crack expansion and seepage mutation pulse events; a heterogeneous graph federation architecture is constructed and combined with a dynamic weighted aggregation strategy to output a federation weight matrix; the federation learning weights and the three-dimensional model output results are integrated, and the posterior probability of burst is calculated by a Bayesian neural network to generate a four-level early warning signal; the present invention realizes multi-dimensional characterization of the glacier state and coupling process simulation, thereby improving the accuracy and timeliness of the early warning.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion includes:

[0006] S1: Joint feature extraction is performed on the acquired multi-temporal and spatiotemporal data through a cross-modal self-supervised learning method to obtain the glacial lake state feature vector and construct a three-dimensional virtual glacial lake model;

[0007] S2: Based on the glacial lake state characteristic vector, a 3D stress field model of the glacier is constructed using the finite element method to obtain stress field data. Simultaneously, a graph neural network is used to construct a seepage channel graph model, and the seepage network characteristics of the graph neural network are obtained. The stress field data and seepage network characteristics are fused through a feature cascade mechanism to form a mechanical-seepage coupling state vector.

[0008] S3: Construct a spiking neural network to simulate pulse events of glacier crack expansion and seepage mutation based on the mechanical-seepage coupling state vector, and dynamically trigger pulse emission using an event-driven learning algorithm;

[0009] S4: Build a heterogeneous graph federation architecture, combine it with a dynamic weighted aggregation strategy, and output a federation weight matrix;

[0010] S5: By integrating the federated learning weights and the output results of the three-dimensional virtual ice lake model, the posterior probability of collapse is calculated through the Bayesian neural network to generate a level 4 warning signal.

[0011] Specifically, the specific steps of S1 include:

[0012] S1.1: Acquire multi-temporal data and perform spatiotemporal registration; the multi-temporal data is quad-modal data that integrates remote sensing imagery, terrain data, sensor network data, and regional meteorological data; the multi-temporal data is collected once an hour;

[0013] S1.2: Use corresponding encoders for different modal data to construct a cross-modal feature space mapping mechanism to obtain spatiotemporal dimension feature vectors, time series feature vectors, and spatial feature vectors; the encoders include an image encoder, a time series encoder, and a terrain encoder;

[0014] S1.3: Jointly optimize the loss function of the spatiotemporal dimension feature vector, the time series feature vector, and the spatial feature vector to obtain the glacial lake state feature vector;

[0015] S1.4: Based on the terrain data and glacier boundary data, construct the 3D geometry of the glacial lake using a voxelization method; the 3D geometry of the glacial lake includes the lake basin, glacier ice, and surrounding terrain;

[0016] S1.5: Mapping the ice lake state feature vector obtained by cross-modal learning into the physical state parameters of the ice lake, and simulating the dynamic change process of the ice lake through a physical engine to obtain a three-dimensional virtual ice lake model; the physical state parameters of the ice lake include temperature field and flow velocity field.

[0017] Specifically, the specific steps of S1.2 include:

[0018] S1.2.1: Use a 3D convolutional neural network as an image encoder to extract spatiotemporal feature vectors from remote sensing images, capturing changes in glacial lake area and glacier texture dynamics, and obtaining spatiotemporal feature vectors.

[0019] S1.2.2: Use the Time Series Transformer as a time series encoder to parse sensor network data and regional meteorological data to obtain time series feature vectors.

[0020] S1.2.3: Construct a terrain encoder using a spatial transformer to convert terrain data into spatial feature vectors to characterize the topographic features of the glacial lake.

[0021] Specifically, the specific steps of constructing the glacier three-dimensional stress field model using the finite element method in S2 include:

[0022] S2.1: Based on the geometric information in the glacial lake state eigenvector, discretize the glacier continuum into a three-dimensional mesh composed of tetrahedral units, and use an adaptive meshing method to refine the mesh near the ice bed or in areas with crack development to obtain a refined three-dimensional mesh; the geometric information includes ice surface elevation, thickness, and bedrock topography;

[0023] S2.2: Assign a corresponding material property tensor to each mesh element in the encrypted three-dimensional mesh according to the physical parameters in the glacial lake state characteristic vector to obtain a set of unit material properties. Simultaneously, for the case of uneven temperature distribution, construct a thermo-mechanical coupling model to convert the temperature field into a thermal expansion coefficient, and combine it with the strain-displacement matrix to generate a temperature load vector. The physical parameters in the glacial lake state characteristic vector include density and elastic modulus.

[0024] S2.3: Based on the variational principle of elasticity, the glacier stress balance equation is transformed into a weak form. The displacement field is interpolated using isoparametric shape functions for each element to obtain the shape function matrix.

[0025] S2.4: Based on the shape function matrix and the strain-displacement matrix, combined with the elastic matrix determined by the material properties, the element stiffness matrix set and the nodal force vector set are obtained;

[0026] S2.5: Through the node number mapping relationship and the preset assembly matrix, the element stiffness matrix set is assembled into a global stiffness matrix, and the node force vector set is assembled into a global load vector to obtain the global equation set;

[0027] S2.6: Use the Newton-Raphson method to solve the modified global equations. If the relative error and residual norm of the displacement increment in each iteration are less than the preset error threshold, the iteration is terminated.

[0028] S2.7: Calculate the stress tensor of each element based on the constitutive relationship to obtain the three-dimensional stress field distribution of the entire glacier.

[0029] Specifically, the feature cascade mechanism is implemented as follows:

[0030] Standardizing the stress field data and the seepage network characteristics;

[0031] The standardized stress field data and percolation network features are mapped to the same dimension through a fully connected layer;

[0032] The channels are spliced ​​together to form a mechanical-percolation coupling state vector.

[0033] Specifically, the construction process of the pulse neural network includes:

[0034] Load the LIF neuron model;

[0035] Dynamically adjusting the neuron membrane potential in the LIF neuron model according to the mechanical-percolation coupling state vector;

[0036] When the neuron membrane potential exceeds the preset threshold, a pulse event is triggered, simulating the expansion of glacier cracks and seepage mutations.

[0037] Specifically, the specific steps of the event-driven learning algorithm include:

[0038] Record the time and spatial location of pulse emission;

[0039] The neuron weights are updated according to the pulse events to optimize the prediction ability of the LIF neuron model for crack expansion.

[0040] Specifically, the construction of a heterogeneous graph federation architecture, combined with a dynamic weighted aggregation strategy, outputs a federation weight matrix, including:

[0041] Divide the glacier area into N sub-areas, each of which serves as a federation node;

[0042] Each federated node locally trains the spiking neural network and Bayesian neural network;

[0043] The model parameters of the spiking neural network and Bayesian neural network in each federated node are integrated through a dynamic weighted aggregation strategy to output a federated weight matrix.

[0044] The specific steps of the dynamic weighted aggregation strategy include:

[0045] Assign weights based on the data quality and historical prediction accuracy of each federation node;

[0046] Gradient descent method is used to optimize the weights of aggregate assignments.

[0047] Specifically, the calculation process of the Bayesian neural network includes:

[0048] Inputting the federated weight matrix and the output of the three-dimensional virtual ice lake model into the input layer of a Bayesian neural network; the federated weight matrix serves as an initialization parameter for a hidden layer of the Bayesian neural network; and concatenating the output of the three-dimensional virtual ice lake model to form a feature vector, which is input into the input layer of the Bayesian neural network.

[0049] The posterior probability distribution of the collapse is estimated by sampling using the Monte Carlo method;

[0050] Based on the posterior probability distribution of the collapse and combined with the preset probability threshold, four-level warning signals of blue, yellow, orange and red are generated; the update frequency of the warning signals is once every 6 hours.

[0051] The glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion also includes:

[0052] Compare and analyze the generated warning signals with the actual monitored glacial lake conditions and collect feedback information;

[0053] Based on the feedback data, the three-dimensional virtual ice lake model, pulse neural network, glacier three-dimensional stress field model, and graph neural network are optimized and adjusted, and the federated learning weights are updated.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention proposes a glacial lake burst monitoring and early warning method based on artificial intelligence and multi-source data fusion. It realizes the deep fusion of multi-temporal and spatial data through cross-modal self-supervised learning, constructs a three-dimensional virtual glacial lake model and extracts the glacial lake state characteristic vector. It combines the finite element method with the graph neural network to construct a mechanical-seepage coupling model, breaking through the limitations of single physical field modeling. It can accurately depict the complex evolution process of glacial lakes under the action of multiple physical fields, improve the ability to capture key disaster-causing factors such as glacial crack expansion and seepage mutation, and solve the problems of low data utilization and insufficient physical mechanism description of traditional methods, thereby improving the authenticity of glacial lake burst simulation and the accuracy of early warning.

[0056] 2. This paper proposes a glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion. It simulates nonlinear mutation processes through a pulse neural network and implements distributed data collaborative modeling in combination with a heterogeneous graph federation architecture. This improves the model's generalization capability while protecting data privacy and is suitable for high-altitude communication-restricted environments. A four-level early warning system based on a Bayesian neural network provides a quantitative basis for disaster classification response, and a dynamic weighted aggregation and event-driven learning mechanism further enhances the timeliness and robustness of the early warning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion of the present invention;

[0058] Figure 2 This is a principle flow chart of the glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion of the present invention;

[0059] Figure 3 The present invention is a flowchart for constructing a three-dimensional virtual glacial lake model for the glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion. DETAILED DESCRIPTION

[0060] Example 1

[0061] See also Figure 1 and Figure 2 The present invention provides an embodiment of a glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion, comprising the following steps:

[0062] S1: Joint feature extraction is performed on the acquired multi-temporal and spatiotemporal data through a cross-modal self-supervised learning method to obtain the glacial lake state feature vector and construct a three-dimensional virtual glacial lake model;

[0063] S2: Based on the glacial lake state characteristic vector, a 3D stress field model of the glacier is constructed using the finite element method to obtain stress field data. Simultaneously, a graph neural network is used to construct a seepage channel graph model, and the seepage network characteristics of the graph neural network are obtained. The stress field data and seepage network characteristics are fused through a feature cascade mechanism to form a mechanical-seepage coupling state vector.

[0064] Furthermore, the construction process of the seepage channel graph model in S2 includes:

[0065] A1: Model the glacier crevasse network as a graph structure, where nodes represent crevasse intersections and edges represent crevasse channels;

[0066] A2: Learn the topological relationship and seepage characteristics of the fractures through graph neural networks, and output the seepage network characteristics. Graph neural networks are the existing technology in this field and are not the inventive solution of this application, so they will not be described here.

[0067] S3: Construct a spiking neural network to simulate pulse events of glacier crack expansion and seepage mutation based on the mechanical-seepage coupling state vector, and dynamically trigger pulse emission using an event-driven learning algorithm;

[0068] S4: Build a heterogeneous graph federation architecture, combine it with a dynamic weighted aggregation strategy, and output a federation weight matrix;

[0069] S5: By integrating the federated learning weights and the output results of the three-dimensional virtual ice lake model, the posterior probability of collapse is calculated through the Bayesian neural network to generate a level 4 warning signal.

[0070] Furthermore, the ice lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion is deployed on the cloud server, supporting real-time access from multiple terminals and push of early warning information.

[0071] Furthermore, the glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion also includes:

[0072] S6: Compare and analyze the generated warning signal with the actual monitored glacial lake conditions and collect feedback information;

[0073] Furthermore, the specific steps of S6 include:

[0074] (1) Align the timestamp of the warning signal generation with the timestamp of the actual monitoring data. At the same time, align the gridded prediction results output by the three-dimensional virtual ice lake model with the spatial coordinates of the actual monitoring sensors (such as GNSS and satellite images);

[0075] (2) Calculate the classification differences between the warning signals and the actual breach events, and compare the mechanical-seepage coupling state vector output by the model with the actual monitored stress field and seepage channel data to obtain quantitative results of the differences;

[0076] (3) Record the confusion matrix between the warning signal and the actual collapse event; the confusion matrix consists of TP, FP, TN, and FN, where TP represents the positive sample predicted by the model as the positive class, TN represents the negative sample predicted by the model as the negative class, FP represents the negative sample predicted by the model as the positive class, and FN represents the positive sample predicted by the model as the negative class;

[0077] (4) Collect physical parameters such as actual seepage rate and crack expansion rate, compare them with the model prediction values, and obtain the model prediction deviation;

[0078] (5) Locate the source of model prediction deviation through residual analysis, such as stress field model error or seepage channel modeling deviation, and store the comparison results as a time series database, annotating time, space, difference type and weight priority.

[0079] S7: Based on feedback data, the three-dimensional virtual ice lake model, pulse neural network, glacier three-dimensional stress field model, and graph neural network are optimized and adjusted. At the same time, the weight strategy of federated learning is updated.

[0080] Furthermore, the multi-model joint optimization process includes:

[0081] 3D Virtual Ice Lake Model: Updating the cross-modal encoder based on self-supervised reconstruction loss and physical residuals;

[0082] Glacier three-dimensional stress field model: combining finite element equation constraints with measured stress data to optimize stress field solution parameters;

[0083] Graph Neural Networks: Update graph node embeddings and edge weights via percolation path residuals;

[0084] Spiking neural networks: Adjusting spike firing thresholds and synaptic weights to adapt to new patterns of cleft expansion events.

[0085] Example 2

[0086] See also Figure 3 , the specific steps of S1 in this embodiment include:

[0087] S1.1: Acquire multi-temporal data and perform spatiotemporal registration; the multi-temporal data is quad-modal data that integrates remote sensing imagery, terrain data, sensor network data, and regional meteorological data; the multi-temporal data is collected once an hour;

[0088] Furthermore, the specific steps of S1.1 include:

[0089] (1) Collect multi-temporal and spatial data, including:

[0090] Remote sensing images: Multispectral images acquired via satellite or drones, with a resolution sufficient for glacial lake monitoring.

[0091] Terrain data: Use digital elevation model data with an accuracy requirement of less than or equal to 1 meter elevation error;

[0092] Sensor network data: deploy ground displacement sensors and temperature sensors;

[0093] Regional meteorological data: integrating temperature, precipitation, wind speed and other data from weather stations;

[0094] (2) Based on the collection of multi-temporal and spatial data, the timestamps are unified to the UTC standard, and linear interpolation or sliding window mean filling is used for asynchronous data to achieve time alignment;

[0095] (3) Resample the data of different resolutions to a unified grid and use affine transformation to align them to the same coordinate system to achieve spatial alignment.

[0096] S1.2: Use corresponding encoders for different modal data to construct a cross-modal feature space mapping mechanism to obtain spatiotemporal dimension feature vectors, time series feature vectors, and spatial feature vectors; the encoders include an image encoder, a time series encoder, and a terrain encoder;

[0097] S1.3: Jointly optimize the loss function of the spatiotemporal dimension feature vector, the time series feature vector, and the spatial feature vector to obtain the glacial lake state feature vector;

[0098] Furthermore, the specific steps of S1.3 include:

[0099] (1) Obtaining the spatiotemporal dimension feature vector , time series feature vector and spatial eigenvectors ;

[0100] It should be noted that the spatiotemporal dimension feature vector is the 256-dimensional output from 3D-ResNet, which extracts the spatiotemporal dynamics of remote sensing images; the time series feature vector is the 128-dimensional output from the temporal transformer, which is the parsed sensor network data and regional meteorological data; the spatial feature vector comes from terrain data analysis and has a dimension of 64;

[0101] (2) The spatiotemporal dimension feature vector, time series feature vector and spatial feature vector are mapped to the same dimension through the fully connected layer. The mapping formula is: ,in, 、 、 Respectively represent the mapped spatiotemporal dimension feature vector, time series feature vector and spatial feature vector, and represents the learnable weight matrix, and , , and Represents the bias term, and its dimension is 128;

[0102] (3) The mapped feature vectors are concatenated into joint features. The concatenation formula is: ,in, Represents vector concatenation;

[0103] (4) Based on joint features , use the mean square error method to predict the probability of collapse and obtain the main loss value. The calculation formula of the main loss is: ,in, represents the main loss value, represents the predicted probability of collapse, Indicates the true label, the value is 0 or 1, representing whether it is collapsed, and is the regression layer parameter, represents the learnable weight matrix, represents the bias term, represents the Euclidean distance;

[0104] (5) Based on joint features , calculate the contrast loss value and KL divergence between similar samples, where the contrast loss value calculation formula is: , the KL divergence calculation formula is: , represents the joint eigenvector of sample i, represents the joint feature vector of sample j, represents the maximum function, express norm, represents the distribution of feature vectors in the spatiotemporal dimension, represents the time series feature vector distribution, represents the spatial eigenvector distribution, represents the KL divergence function;

[0105] It should be noted that the sample is the feature vector after mapping.

[0106] (6) Main loss value , contrast loss value and KL divergence Perform weighted summation to get the total loss value ;

[0107] (7) According to the total loss value , synchronously update the 3D-ResNet, temporal Transformer and fully connected layer parameters through the back-propagation algorithm, and output the optimized joint features , which is the characteristic vector of the ice lake state, wherein the back propagation algorithm is the existing technical content in this field and is not the inventive solution of this application, and is not described here in detail;

[0108] S1.4: Based on the terrain data and glacier boundary data, construct the 3D geometry of the glacial lake using a voxelization method; the 3D geometry of the glacial lake includes the lake basin, glacier ice, and surrounding terrain;

[0109] Furthermore, the specific steps of S1.4 include:

[0110] (1) Obtaining terrain data and glacier boundary data and converting them to the same coordinate system; the terrain data is a digital elevation model;

[0111] (2) Convert the glacier boundary vector into a binary grid with the same resolution as the terrain data to obtain the gridded glacier mask matrix At the same time, the glacier mask is spatially buffered with a preset buffer radius to generate an expanded area, which is then multiplied element-by-element with the elevation value in the digital elevation model to obtain local terrain data. ,in, Represents the plane coordinates of raster pixels;

[0112] The obtained rasterized glacier mask The process includes:

[0113] If the plane coordinates of the grid pixel are within the glacier boundary, the pixel in the rasterized glacier mask matrix is ​​1; otherwise, the pixel in the rasterized glacier mask matrix is ​​0;

[0114] It should be noted that the scanline filling method is used to convert the glacier boundary vector into a binary raster with the same resolution as the terrain data.

[0115] (3) Define a voxel grid; the voxel grid includes a spatial range and a voxel size; the spatial range needs to cover the local terrain data The bounding box of

[0116] (4) Set attribute filling rules, including:

[0117] Lake basin voxels: If the elevation is below the lake boundary and is not a glacier area, the attribute is marked as ;

[0118] Glacier Ice Voxel: If located If the elevation is greater than or equal to the end of the glacier, the attribute is marked as ;

[0119] Surrounding terrain voxels: If the elevation is above the lake boundary and is not in a glacier area, or is not located in If the elevation is less than the end of the glacier, the attribute is marked as ;

[0120] (5) Export a 3D voxel grid containing attribute labels.

[0121] S1.5: Mapping the ice lake state feature vector obtained by cross-modal learning into the physical state parameters of the ice lake, and simulating the dynamic change process of the ice lake through a physical engine to obtain a three-dimensional virtual ice lake model; the physical state parameters of the ice lake include temperature field and flow velocity field.

[0122] Furthermore, the specific steps of S1.5 include:

[0123] (1) Obtaining the ice lake state characteristic vector , using the multi-layer perceptron regression model Mapping into a physical parameter vector; the physical parameter vector includes lake basin geometric parameters, glacier-basin contact angle, temperature field initial distribution parameters, velocity field initial parameters, and seepage coefficient. The multi-layer perceptron regression model is a prior art in this field and is not an inventive solution of the present application, and is not described in detail here.

[0124] (2) using the Marching Cubes algorithm to generate a parameterized lake basin model based on the lake basin geometric parameters. The Marching Cubes algorithm is a prior art in this field and is not an inventive solution of this application, and will not be described in detail here.

[0125] (3) Based on the glacier boundary data and the glacier contact angle, a three-dimensional model of the ice tongue is generated using Boolean operations. Boolean operations are prior art in this field and are not an inventive solution of this application, so they will not be described in detail here.

[0126] (4) Define the physical engine and configure the temperature field and flow velocity field;

[0127] (5) Based on the physics engine, the state is updated by combining the physical parameter vector and the vertex state variables; the state refers to the set of all dynamically changing physical properties of the Ice Lake system;

[0128] (6) Based on the three-dimensional model of the ice tongue, the ice tongue stress is monitored in real time. When the ice tongue stress is greater than the preset critical tensile strength, the crack expansion is triggered and the boundary conditions of the seepage field and velocity field are updated;

[0129] (7) Real-time rendering of three-dimensional grids and physical fields, exporting time-series grid sequences and physical field data, and forming a dynamically updated three-dimensional virtual ice lake model; the physical fields include temperature field cloud maps and velocity field vector maps.

[0130] The specific steps of S1.2 include:

[0131] S1.2.1: Use a 3D convolutional neural network as an image encoder to extract spatiotemporal feature vectors from remote sensing images, capturing changes in glacial lake area and glacier texture dynamics, and obtaining spatiotemporal feature vectors.

[0132] Furthermore, the specific steps of S1.2.1 include:

[0133] (1) Obtain remote sensing images and stack them into a four-dimensional tensor in chronological order , where R represents the real number domain, T represents the number of time frames, H represents the height, W represents the width, and C represents the number of channels; the number of channels is the number of multispectral bands;

[0134] (2) Normalizing the pixel values ​​of the remote sensing image in the four-dimensional tensor to [0, 1], and using a temporal linear interpolation method to linearly interpolate and complete the missing frames. The temporal linear interpolation method is a prior art in this field and is not an inventive solution of the present application, and will not be described in detail here.

[0135] The pixel value normalization process is as follows: subtract the minimum value of the global pixel value range of the cth band from the pixel value of each frame of the remote sensing image in the four-dimensional tensor to obtain a first difference variable; at the same time, subtract the minimum value of the global pixel value range of the cth band from the maximum value of the global pixel value range of the cth band to obtain a second difference variable; finally, the ratio of the first difference variable to the second difference variable is calculated to obtain a normalized four-dimensional tensor.

[0136] (3) Loading a pre-trained 3D-ResNet model; the 3D-ResNet includes a 3D convolutional layer, a 3D pooling layer, and a residual connection. 3D-ResNet is a prior art in this field and is not an inventive solution of this application, and is not described in detail here;

[0137] (4) The four-dimensional tensor after linear interpolation completion is input into the pre-trained 3D-ResNet model, and the changes in the water body boundaries in the multi-frame remote sensing images are captured through the 3D convolution layer, and the temporal area change rate corresponding to the number of channels of the temporal feature map is output; the temporal area change rate is: the water body binary mask of the tth frame Subtract the water binary mask of frame t-1 of Norm, the product of the height H and width W of the four-dimensional tensor after linear interpolation completion, is used to obtain the temporal area change rate of the t-th frame ;

[0138] It should be noted that the water body binary mask It is obtained by segmenting the spectral characteristics of remote sensing images. The specific calculation process includes:

[0139] a. Extracting candidate water areas by normalizing the water index, wherein the normalized water index calculation formula is the prior art content in this field and is not an inventive solution of the present application and will not be described in detail here;

[0140] b. Set the segmentation threshold. If the normalized water index is greater than or equal to the preset segmentation threshold, the initial binary mask is 1; if the normalized water index is less than the preset segmentation threshold, the initial binary mask is 0;

[0141] c. Optimize the initial binary mask through morphological opening and closing operations to eliminate noise and holes and obtain the optimized binary mask;

[0142] d. Retain the largest connected area in the optimized binary mask as the final water body mask.

[0143] (5) At the same time, a 3D convolution kernel is used to scan the temporal texture to locate the crack expansion area, and the melting trend is detected by the mean shift of the temporal feature map channel;

[0144] Furthermore, the mean calculation formula of the time series feature map channel is: ,in, represents the average characteristic response value of the glacier area in the t-th frame, which is used to quantify the degree of surface melting. Represents the glacier area, Represents the total number of pixels in the glacier area, used for normalization calculation, Represents the temporal feature map output by 3D-ResNet at the tth frame and spatial position The channel activation value at reflects the local texture or temperature anomaly, Indicates spatial location.

[0145] (6) For the last layer of temporal feature map Perform spatiotemporal average pooling to generate a fixed-length feature vector, that is, the feature vector after pooling , and compress the pooled feature vector into a 256-dimensional low-dimensional vector to generate a spatiotemporal feature vector, where Represents the t-th frame and spatial position of the last layer of the 3D-ResNet temporal feature map The channel activation value at Indicates the time dimension length of the last layer feature map of 3D-ResNet, Indicates the height of the last feature map of 3D-ResNet. Indicates the width of the last layer feature map of 3D-ResNet, Indicates the number of channels of the last layer feature map of 3D-ResNet.

[0146] S1.2.2: Use the Time Series Transformer as a time series encoder to parse sensor network data and regional meteorological data to obtain time series feature vectors. The Time Series Transformer is a prior art in this field and does not constitute an inventive solution of this application, so it is not described in detail here.

[0147] S1.2.3: A terrain encoder is constructed through a spatial transformer to convert terrain data into a spatial feature vector to characterize the topographic features of the glacial lake. The spatial transformer is a prior art in this field and is not an inventive solution of this application, so it will not be elaborated here.

[0148] The specific steps of constructing the 3D stress field model of the glacier using the finite element method in S2 include:

[0149] S2.1: Based on the ice lake state characteristic vector The glacier continuum is discretized into a three-dimensional grid composed of tetrahedral units based on the geometric information in , and the adaptive meshing method is used to encrypt the mesh near the ice bed or in the crack development area to obtain the encrypted three-dimensional mesh ,in, represents the nth grid cell in the three-dimensional grid, where n represents the number of divided grid cells; the geometric information includes ice surface elevation, thickness, and bedrock topography;

[0150] It should be noted that the adaptive grid division method is the existing technical content in this field and is not the inventive solution of this application, and will not be described in detail here.

[0151] S2.2: According to the physical parameters in the glacial lake state characteristic vector, for each grid cell in the encrypted three-dimensional grid The corresponding material property tensor is assigned to obtain the unit material property set. At the same time, for the case of uneven temperature distribution, a thermal-mechanical coupling model is constructed to convert the temperature field into a thermal expansion coefficient. Combined with the strain-displacement matrix, the temperature load Q is generated. The physical parameters in the ice lake state characteristic vector include density, elastic modulus, and Poisson's ratio.

[0152] Furthermore, the specific steps of S2.2 include:

[0153] (1) Extracting the glacial lake state feature vector Physical parameters in;

[0154] (2) For each grid cell, define the stiffness matrix of the isotropic elastic material according to its position and the eigenvector of the glacial lake state: , and satisfy: , , , , ,in, and Represents a grid cell The local parameters of 、 、 represents an array in matrix form, represents Poisson's ratio, represents the elastic modulus function;

[0155] (3) According to the stiffness matrix, the element material property set is obtained ;

[0156] (4) The temperature field in the ice lake state characteristic vector Interpolate to each grid node to get the average temperature of the grid unit node , where r represents the node index;

[0157] In the present invention, the interpolation method adopts the linear interpolation method, which is the existing technical content in this field and is not the inventive solution of this application, and will not be described in detail here;

[0158] (5) The average temperature of the grid unit node and isoparametric shape functions After multiplying, accumulate and sum to get the average temperature of the grid unit ,in, 、 、 represents natural coordinates;

[0159] (6) Calculate the average temperature of the grid cells The difference between the preset reference temperature and the result is then compared with the unit tensor and thermal expansion coefficient Multiply to get the grid cells Thermal strain within ;

[0160] (7) Thermal strain , strain-displacement matrix , the elastic matrix determined by the material properties Integrate the product of to get the grid unit temperature load vector ;

[0161] It should be explained that the strain-displacement matrix is ​​obtained by isoparametric shape functions The derivatives with respect to the natural coordinates, the Jacobian matrix and its inverse are calculated and used to map the nodal displacements to strains within the element.

[0162] (8) The grid cell temperature load vector The temperature load is obtained by summing the products with the preset assembly matrix .

[0163] S2.3: Based on the variational principle of elasticity, the glacier stress balance equation is transformed into a weak form and an isoparametric shape function is used for each element. Interpolate the displacement field to obtain the shape function matrix ;

[0164] Furthermore, the weak form equilibrium equation is:

[0165] ;

[0166] in, represents the double product, u represents the displacement field, represents the strain tensor, represents the stress tensor, represents the imaginary part, represents the imaginary displacement vector, represents the virtual strain tensor, It represents the mapping relationship between virtual strain and virtual displacement, that is, virtual strain is derived from virtual displacement. represents the solution domain of the glacier, represents the body force, represents the surface force, Represents the boundary of the solution domain.

[0167] Furthermore, the isoparametric shape function The formula for interpolating the displacement field is: ,in, Indicates the number of mesh element nodes. For tetrahedral elements, , Represents a grid cell The displacement vector of the rth node in , Represents a grid cell Any point in the natural coordinates The displacement vector below.

[0168] Furthermore, regarding the shape function matrix and the strain-displacement matrix , it needs to be explained that the shape function matrix is ​​composed of isoparametric shape functions Arranged in node order, they are used to map the node displacement to the displacement of any point in the element.

[0169] S2.4: Elastic matrix determined based on shape function matrix and strain-displacement matrix, combined with material properties , and obtain the element stiffness matrix set and the set of nodal force vectors ,in, represents transpose, represents the shape function matrix, Indicates the r The strain-displacement matrix of each node, represents the surface area of ​​the kth grid cell, Represents the surface integral of the grid cell boundary, represents the boundary accumulation of surface forces, and reflects the boundary interaction. It represents the volume integral in three-dimensional space, represents the spatial accumulation of volume forces and material properties, and reflects the three-dimensional spatial distribution.

[0170] It should be noted that the surface integral is integrated on the unit boundary; the volume integral is integrated in three-dimensional space.

[0171] S2.5: Through the node number mapping relationship, combined with the preset assembly matrix, the unit stiffness matrix is ​​assembled Assembled into a global stiffness matrix, node force vector set Assemble into global load vectors and obtain the global equations;

[0172] S2.6: Solve the corrected global system of equations using the Newton-Raphson method. If the relative error and residual norm of the displacement increment in each iteration are both less than a preset error threshold, terminate the iteration. The Newton-Raphson method is known in the art and does not constitute an inventive solution of this application, so its detailed description is omitted here.

[0173] S2.7: According to the constitutive relations Calculate each grid cell The stress tensor , thus obtaining the three-dimensional stress field distribution of the entire glacier ,in, represents strain, Represented by displacement field u is the strain tensor of the variable;

[0174] The specific implementation of the feature cascade mechanism in S2 is:

[0175] B1: standardizing the stress field data and the seepage network characteristics;

[0176] B2: Map the normalized stress field data and percolation network features to the same dimension through a fully connected layer;

[0177] B3: Splice according to channel dimension to form the mechanical-percolation coupling state vector.

[0178] The construction process of the spiking neural network in S3 includes:

[0179] C1: Load the LIF (Leaky Integrate and Fire, LIF) neuron model;

[0180] C2: Dynamically adjust the neuron membrane potential in the LIF neuron model according to the mechanical-percolation coupling state vector;

[0181] Furthermore, the specific steps of C2 include:

[0182] (1) Obtaining the LIF neuron model;

[0183] (2) Decomposing the mechanical-percolation coupling state vector into mechanical components and percolation components;

[0184] (3) Establish a mapping function from mechanical components and percolation components to neuron parameters, and integrate the mapping function from mechanical components and percolation components to neuron parameters into the equations of the LIF neuron model to obtain the parameter changes by solving them;

[0185] (4) Iterate so that the model parameters are updated at each time step according to the current mechanical-percolation coupling state vector.

[0186] It should be noted that the LIF neuron model and its corresponding equations are prior art in this field and are not the inventive solution of this application, and will not be described in detail here.

[0187] C3: When the neuron membrane potential exceeds the preset threshold, a pulse event is triggered, simulating the expansion of glacier cracks and seepage mutations.

[0188] The specific steps of the event-driven learning algorithm in S3 include:

[0189] E1: records the time and spatial position of pulse emission;

[0190] E2: Update the neuron weights according to the pulse events to optimize the LIF neuron model's ability to predict crack expansion.

[0191] The specific steps of S4 include:

[0192] S4.1: Divide the glacier area into N sub-areas, each of which serves as a federation node;

[0193] S4.2: Each federated node locally trains a spiking neural network and a Bayesian neural network. The Bayesian neural network is prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0194] S4.3: Integrate the model parameters of the spiking neural network and Bayesian neural network in each federated node through a dynamic weighted aggregation strategy and output the federated weight matrix;

[0195] The specific steps of the dynamic weighted aggregation strategy include:

[0196] Assign weights based on the data quality and historical prediction accuracy of each federation node;

[0197] The gradient descent method is used to optimize the weight of the aggregation distribution, wherein the gradient descent method is the existing technical content in this field and is not an inventive solution of this application, and will not be described here.

[0198] The calculation process of the Bayesian neural network in S5 includes:

[0199] S5.1: Input the federated weight matrix and the output of the three-dimensional virtual ice lake model into the input layer of a Bayesian neural network; the federated weight matrix serves as the initialization parameter of the hidden layer of the Bayesian neural network; the output of the three-dimensional virtual ice lake model is concatenated to form a feature vector, which is input into the input layer of the Bayesian neural network;

[0200] S5.2: Estimate the posterior probability distribution of the collapse by sampling using the Monte Carlo method. The Monte Carlo method is a prior art in this field and is not an inventive solution of the present application, and is not described in detail here.

[0201] S5.3: Based on the posterior probability distribution of the collapse and the preset probability threshold, four-level warning signals of blue, yellow, orange and red are generated; the update frequency of the warning signal is once every 6 hours.

[0202] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

Claims

1. A glacial lake feedback monitoring and early warning method based on artificial intelligence and multi-source data fusion is characterized by: include: S1: Joint feature extraction is performed on the acquired multi-temporal and spatiotemporal data through a cross-modal self-supervised learning method to obtain the glacial lake state feature vector and construct a three-dimensional virtual glacial lake model; S2: Based on the glacial lake state characteristic vector, a 3D stress field model of the glacier is constructed using the finite element method to obtain stress field data. Simultaneously, a graph neural network is used to construct a seepage channel graph model, and the seepage network characteristics of the graph neural network are obtained. The stress field data and seepage network characteristics are fused through a feature cascade mechanism to form a mechanical-seepage coupling state vector. S3: Construct a spiking neural network to simulate pulse events of glacier crack expansion and seepage mutation based on the mechanical-seepage coupling state vector, and dynamically trigger pulse emission using an event-driven learning algorithm; S4: Build a heterogeneous graph federation architecture, combine it with a dynamic weighted aggregation strategy, and output a federation weight matrix; S5: Integrating the federated learning weights and the output of the 3D virtual ice lake model, the Bayesian neural network is used to calculate the posterior probability of collapse and generate a level 4 warning signal. The specific steps of S4 include: S4.1: Divide the glacier area into N sub-areas, each of which serves as a federation node; S4.2: Each federated node locally trains the spiking neural network and the Bayesian neural network; S4.3: Integrate the model parameters of the spiking neural network and Bayesian neural network in each federated node through a dynamic weighted aggregation strategy and output the federated weight matrix; The specific steps of the dynamic weighted aggregation strategy include: Assign weights based on the data quality and historical prediction accuracy of each federation node; Gradient descent method is used to optimize the weight of aggregation allocation; The calculation process of the Bayesian neural network in S5 includes: S5.1: Input the federated weight matrix and the output of the three-dimensional virtual ice lake model into the input layer of a Bayesian neural network; the federated weight matrix serves as the initialization parameter of the hidden layer of the Bayesian neural network; the output of the three-dimensional virtual ice lake model is concatenated to form a feature vector, which is input into the input layer of the Bayesian neural network; S5.2: Estimate the posterior probability distribution of the failure by Monte Carlo sampling; S5.3: Based on the posterior probability distribution of the collapse and the preset probability threshold, four-level warning signals of blue, yellow, orange and red are generated; the update frequency of the warning signal is once every 6 hours.

2. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 1 is characterized in that: The specific steps of S1 include: S1.1: Acquire multi-temporal data and perform spatiotemporal registration; the multi-temporal data is quad-modal data that integrates remote sensing imagery, terrain data, sensor network data, and regional meteorological data; the multi-temporal data is collected once an hour; S1.2: Use corresponding encoders for different modal data to construct a cross-modal feature space mapping mechanism to obtain spatiotemporal dimension feature vectors, time series feature vectors, and spatial feature vectors; the encoders include an image encoder, a time series encoder, and a terrain encoder; S1.3: Jointly optimize the loss function of the spatiotemporal dimension feature vector, the time series feature vector, and the spatial feature vector to obtain the glacial lake state feature vector; S1.4: Based on the terrain data and glacier boundary data, construct the 3D geometry of the glacial lake using a voxelization method; the 3D geometry of the glacial lake includes the lake basin, glacier ice, and surrounding terrain; S1.5: Mapping the ice lake state feature vector obtained by cross-modal learning into the physical state parameters of the ice lake, and simulating the dynamic change process of the ice lake through a physical engine to obtain a three-dimensional virtual ice lake model; the physical state parameters of the ice lake include temperature field and flow velocity field.

3. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 2 is characterized in that: The specific steps of S1.2 include: S1.2.1: Use a 3D convolutional neural network as an image encoder to extract spatiotemporal feature vectors from remote sensing images, capturing changes in glacial lake area and glacier texture dynamics, and obtaining spatiotemporal feature vectors. S1.2.2: Use the Time Series Transformer as a time series encoder to parse sensor network data and regional meteorological data to obtain time series feature vectors. S1.2.3: Construct a terrain encoder using a spatial transformer to convert terrain data into spatial feature vectors to characterize the topographic features of the glacial lake.

4. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 3 is characterized in that: The specific steps of constructing the 3D stress field model of the glacier using the finite element method in S2 include: S2.1: Based on the geometric information in the glacial lake state eigenvector, discretize the glacier continuum into a three-dimensional mesh composed of tetrahedral units, and use an adaptive meshing method to refine the mesh near the ice bed or in areas with crack development to obtain a refined three-dimensional mesh; the geometric information includes ice surface elevation, thickness, and bedrock topography; S2.2: Assign a corresponding material property tensor to each mesh element in the encrypted three-dimensional mesh according to the physical parameters in the glacial lake state characteristic vector to obtain a set of unit material properties. Simultaneously, for the case of uneven temperature distribution, construct a thermo-mechanical coupling model to convert the temperature field into a thermal expansion coefficient, and combine it with the strain-displacement matrix to generate a temperature load vector. The physical parameters in the glacial lake state characteristic vector include density and elastic modulus. S2.3: Based on the variational principle of elasticity, the glacier stress balance equation is transformed into a weak form. The displacement field is interpolated using isoparametric shape functions for each element to obtain the shape function matrix. S2.4: Based on the shape function matrix and the strain-displacement matrix, combined with the elastic matrix determined by the material properties, the element stiffness matrix set and the nodal force vector set are obtained; S2.5: Through the node number mapping relationship and the preset assembly matrix, the element stiffness matrix set is assembled into a global stiffness matrix, and the node force vector set is assembled into a global load vector to obtain the global equation set; S2.6: Use the Newton-Raphson method to solve the modified global equations. If the relative error and residual norm of the displacement increment in each iteration are less than the preset error threshold, the iteration is terminated. S2.7: Calculate the stress tensor of each element based on the constitutive relationship to obtain the three-dimensional stress field distribution of the entire glacier.

5. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 4 is characterized in that: The specific implementation of the feature cascade mechanism is as follows: Standardizing the stress field data and the seepage network characteristics; The standardized stress field data and percolation network features are mapped to the same dimension through a fully connected layer; The channels are spliced ​​together to form a mechanical-percolation coupling state vector.

6. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 5 is characterized in that: The construction process of the pulse neural network includes: Load the LIF neuron model; Dynamically adjusting the neuron membrane potential in the LIF neuron model according to the mechanical-percolation coupling state vector; When the neuron membrane potential exceeds the preset threshold, a pulse event is triggered, simulating the expansion of glacier cracks and seepage mutations.

7. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 6 is characterized in that: The specific steps of the event-driven learning algorithm include: Record the time and spatial location of pulse emission; The neuron weights are updated according to the pulse events to optimize the prediction ability of the LIF neuron model for crack expansion.

8. The glacial lake feed monitoring and early warning method based on artificial intelligence and multi-source data fusion according to claim 7 is characterized in that: Also includes: Compare and analyze the generated warning signals with the actual monitored glacial lake conditions and collect feedback information; Based on the feedback data, the three-dimensional virtual ice lake model, pulse neural network, glacier three-dimensional stress field model, and graph neural network are optimized and adjusted, and the federated learning weights are updated.

Images