Water level distributed monitoring method and system based on visual sensing network

Abstract

The invention provides a water level distributed monitoring method and system based on a visual sensing network, and relates to the technical field of water level monitoring, and the method comprises the steps: constructing a water level propagation prediction model through multi-modal comparative learning and a graph neural network, processing a real-time water level image through the combination of adaptive wavelet transform, decoupling representation and a dense connection network, and obtaining a real-time water level prediction model; the water level measurement value and the reliability score are obtained, anomaly analysis and emergency monitoring are carried out based on state estimation, causal discovery and hierarchical reinforcement learning, high-precision and high-reliability distributed water level monitoring can be achieved, abnormal water level changes can be effectively recognized, and timely emergency monitoring parameters are provided.

Description

Technical Field

[0001] The present invention relates to the technical field of water level monitoring, and in particular, to a distributed water level monitoring method and system based on a visual sensor network. Background Art

[0002] With the continuous development of water conservancy projects and flood control and disaster reduction work, water level monitoring, as an important basic task, has received extensive attention. Traditional water level monitoring mainly relies on manual observation or automatic measurement equipment, but there are many limitations in terms of monitoring efficiency, accuracy, and coverage; In recent years, the water level monitoring technology based on a visual sensor network has gradually emerged. By deploying camera devices to collect water level images and combining computer vision and deep learning algorithms for intelligent analysis, automatic monitoring of water levels has been achieved. Existing visual sensor network water level monitoring systems usually adopt image processing and machine learning methods, estimate water level values by extracting water level scale features and water surface texture features, and establish a prediction model to analyze the water level change trend; However, the existing technology still has problems such as insufficient monitoring accuracy in complex environments, failure to fully consider the spatial dependence relationship between visual sensor nodes, and lack of research on the causal relationship analysis and abnormal propagation mechanism of water level changes; Therefore, there is an urgent need for a solution to solve the problems existing in the existing technology. Summary of the Invention

[0003] Embodiments of the present invention provide a distributed water level monitoring method and system based on a visual sensor network, which can at least solve some problems existing in the existing technology.

[0004] In the first aspect of the embodiments of the present invention, a distributed water level monitoring method based on a visual sensor network is provided, including: Collecting water level image samples of a monitoring area and constructing a spectral decomposition network to extract the frequency domain features of the water level image samples, constructing a multi-scale spectral feature matrix, constructing an energy matrix based on the multi-scale spectral feature matrix and calculating the energy distribution density, generating an energy mask map corresponding to the water level image samples, guiding an attention mechanism through the energy mask map to perform self-supervised decomposition on the image, obtaining a water level scale feature tensor, a water surface texture feature tensor, and an environmental background feature tensor, inputting the decomposed feature tensors into a multi-modal contrast learning network, calculating a discriminative feature vector group through maximizing mutual information, constructing a random field model based on the spatial distribution characteristics of visual sensor nodes, calculating the spatial dependence relationship between nodes to obtain a node association probability matrix, and inputting the node association probability matrix and the discriminative feature vector group into a graph neural network, and obtaining a parameter matrix of a water level propagation prediction model through message passing iteration; Receive real-time water level images, input the parameter matrix of the water level propagation prediction model into the adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate multi-resolution feature representations, obtain the compressed feature code stream by performing sparse coding on the multi-resolution feature representations, input the feature code stream into the decoupled representation network, separate the water level feature map and calculate the complementary metric value based on the water level feature map, perform adaptive fusion based on the complementary metric value to obtain the enhanced feature map, input the enhanced feature map into the densely connected network, obtain the water level region segmentation map and water level measurement value through cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement according to the Bayesian inference framework, combining prior knowledge and observed data, generate a reliability score table by calculating the measurement uncertainty value based on the posterior probability distribution, screen out the high-confidence water level data set and perform spectral analysis, and extract the periodic feature vector and trend feature vector; Based on the periodic feature vector and the trend feature vector, construct a nonlinear state space model, calculate the state estimation sequence by combining particle filtering and construct a causal discovery network, calculate the causal relationship matrix of the water level change based on the causal discovery network and decompose it into multiple sub-matrices, construct a hierarchical attention network at different spatial scales and execute to obtain the water level feature tensor, obtain the encrypted model parameter update amount through the federated optimization objective function combined with the homomorphic encryption algorithm, calculate the optimal sampling strategy vector through the pre-set multi-agent collaborative decision-making system and perform anomaly analysis on the water level to obtain the anomaly feature vector, add the optimal sampling strategy vector and the anomaly feature vector to the hierarchical reinforcement learning network to obtain the detection resource allocation plan and the execution parameter matrix, and analyze the anomaly propagation path map and the emergency monitoring parameter sequence by combining the causal relationship matrix.

[0005] In an alternative embodiment, collect water level image samples of the monitoring area and construct a spectral decomposition network to extract the frequency domain features of the water level image samples, construct a multi-scale spectral feature matrix, construct an energy matrix based on the multi-scale spectral feature matrix and calculate the energy distribution density, generate the energy mask map corresponding to the water level image samples, perform self-supervised decomposition on the image through the energy mask map-guided attention mechanism to obtain the water level scale feature tensor, the water surface texture feature tensor and the environmental background feature tensor, input the decomposed feature tensors into the multi-modal contrast learning network, calculate the discriminative feature vector group through mutual information maximization, construct a random field model based on the spatial distribution characteristics of the visual sensing nodes, calculate the spatial dependence relationship between nodes to obtain the node association probability matrix, and input the node association probability matrix and the discriminative feature vector group into the graph neural network, and obtain the parameter matrix of the water level propagation prediction model through message passing iteration, including: Collect water level image samples of the monitoring area, and input the water level image samples into a multi-level spectral decomposition network. In the multi-level spectral decomposition network, the shallow network extracts the high-frequency components of the water level image samples through a high-pass filter bank to obtain water level scale line features and water surface ripple features. The middle network extracts the intermediate-frequency components of the water level image samples through a band-pass filter bank to obtain water level line contour features and shore structure features. The deep network extracts the low-frequency components of the water level image samples through a low-pass filter bank to obtain water area distribution features. Normalize the feature maps corresponding to the high-frequency components, the intermediate-frequency components, and the low-frequency components, and arrange and combine them in the frequency band order to form a multi-scale spectral feature matrix; Calculate the energy values of the components in each frequency band of the multi-scale spectral feature matrix to obtain a two-dimensional energy matrix. Perform density estimation on the two-dimensional energy matrix through a kernel function to obtain an energy density distribution map. Set an adaptive threshold based on the energy density distribution map and generate a binary energy mask map. Determine the high-energy region and the low-energy region. Mark the high-energy region as the foreground region and the low-energy region as the background region; Divide the water level image samples into multiple overlapping image blocks. Calculate the attention weights of the multiple overlapping image blocks based on the binary energy mask map. Input the overlapping image blocks with the attention weights into a self-supervised deconstruction module for feature decomposition. Extract the water level scale feature branch through a fine-grained feature extractor to obtain a water level scale feature tensor. Extract the water surface texture feature branch through a texture analyzer to obtain a water surface texture feature tensor. Extract the environmental background feature branch through a context encoder to obtain an environmental background feature tensor. Input the water level scale feature tensor, the water surface texture feature tensor, and the environmental background feature tensor into a feature encoder respectively for dimensionality reduction to obtain feature vectors of a unified dimension, and construct a feature comparison pool; Randomly sample positive sample pairs and negative sample pairs from the feature vectors of the unified dimension. Calculate the mutual information metric values between the positive sample pairs and the negative sample pairs. Obtain a discriminative feature vector group by maximizing the mutual information metric values of the positive sample pairs and minimizing the mutual information metric values of the negative sample pairs. Calculate the distance matrix between node pairs based on the geographical coordinates of the visual sensing nodes. Convert the distance matrix into an initial association strength. Calculate the interaction between node pairs through the potential function of the random field to obtain a node association probability matrix; Input the node association probability matrix and the discriminative feature vector group into a graph neural network. In each round of message passing iteration, determine the importance weights of neighbor nodes based on the node association probability matrix. Perform a non-linear transformation on the feature information of the neighbor nodes to obtain a fused feature. Combine the fused feature with the feature information of the node itself to update the node state. Obtain the parameter matrix of the water level propagation prediction model through multiple rounds of iteration.

[0006] In an alternative embodiment, positive sample pairs and negative sample pairs are randomly sampled from the feature vectors of the unified dimension, the mutual information metric value between the positive sample pairs and the negative sample pairs is calculated, and a discriminative feature vector group is obtained by maximizing the mutual information metric value of the positive sample pairs and minimizing the mutual information metric value of the negative sample pairs. Based on the geographical coordinates of the visual sensing nodes, a distance matrix between node pairs is calculated, the distance matrix is converted into an initial association strength, and the interaction between node pairs is calculated through the potential function of the random field to obtain a node association probability matrix, including: A feature comparison pool is constructed from the feature vectors of the unified dimension and the feature comparison pool is divided into multiple time windows. Scene feature information is extracted within each time window and positive sample pairs are randomly sampled. Negative sample pairs are randomly sampled from the feature vectors by a multi-layer hierarchical sampling strategy according to weather conditions, monitoring periods, and water level change trends in different monitoring scenarios; The feature vectors in the positive sample pairs and the negative sample pairs are added to a multi-layer perceptron including an input layer, multiple hidden layers, and an output layer for feature transformation. A batch normalization layer and a non-linear activation function are set between adjacent layers of the multi-layer perceptron. After mapping the feature vectors from the feature space to the probability distribution space, a Gaussian kernel function group with an adaptive bandwidth is constructed in the probability distribution space, and a continuous probability distribution is obtained by accumulating the local and global contributions of the probability densities of all feature point pairs; The mutual information metric value between the positive sample pairs and the negative sample pairs is calculated. Among them, the mutual information metric value is obtained by calculating the multiple differences of the joint distribution, conditional distribution, and marginal distribution of the transformed features. A mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints is constructed. By maximizing the mutual information metric value of the positive sample pairs and minimizing the mutual information metric value of the negative sample pairs, a multi-round iterative optimization is performed in combination with the gradient descent method with an adaptive learning rate to obtain a feature vector group with local and global discriminability; Based on the geographical coordinates of the visual sensing nodes, a distance matrix between node pairs is calculated by fusing the geodetic distance between nodes, the terrain undulation coefficient, the river channel connectivity coefficient, and the hydrological propagation feature coefficient. The multi-scale elevation difference between nodes is calculated in combination with digital elevation model data to dynamically correct the distance matrix. At the same time, the actual flow distance considering the water flow direction and velocity is calculated based on the river channel connectivity data, and the actual flow distance and the corrected distance matrix are fused with multiple features to obtain a corrected distance matrix; Based on the corrected distance matrix, map the distance values to a dynamic interval through an adaptive normalization method, calculate the local and global densities based on the multi-dimensional spatial distribution density characteristics of the nodes, set an adaptive attenuation coefficient based on the density gradient for the densely distributed areas, set a dynamic attenuation coefficient considering spatial heterogeneity for the sparsely distributed areas, and calculate the initial association strength at multiple scales by combining the spatio-temporal correlation characteristics of the nodes; Based on the initial association strength, construct a random field potential function with an adaptive piecewise structure, dynamically divide the node distances into a short-distance interval with a transition zone, a medium-distance interval, and a long-distance interval based on the multi-dimensional spatial relationships between the nodes, design a linear attenuation function considering local spatial dependence within the short-distance interval, design an exponential attenuation function considering regional influence within the medium-distance interval, design a residual function considering global association within the long-distance interval, and determine the adaptive piecewise points of the potential function through spatial autocorrelation analysis; Based on the adaptive piecewise points, calculate the interaction between node pairs through the potential function of the random field by means of a multi-layer message passing mechanism. In each iteration, collect the state information of multi-scale neighborhood nodes based on the spatial dependence relationship, weight the messages at multiple levels using the adaptive potential function values, identify and suppress spatio-temporal redundant information through a structured attention mechanism, perform non-linear feature combination on the weighted messages and add random noise following a dynamic distribution, and obtain the node association probability matrix characterizing the association relationship between nodes through state update.

[0007] In an alternative embodiment, receive real-time water level images, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream by performing sparse coding on the multi-resolution feature representation, input the feature code stream into a decoupled representation network, separate to obtain a water level feature map and calculate a complementarity metric value based on the water level feature map, perform adaptive fusion based on the complementarity metric value to obtain an enhanced feature map, input the enhanced feature map into a densely connected network, obtain a water level region segmentation map and a water level measurement value through cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement according to the Bayesian inference framework by combining prior knowledge and observed data, generate a reliability score table based on the measurement uncertainty value calculated from the posterior probability distribution, screen to obtain a high-confidence water level data set and perform spectral analysis, and extract the periodic feature vector and the trend feature vector including: Receive real-time water level images, input the parameter matrix of the water level propagation prediction model into the adaptive wavelet transform network, calculate the gradients of the image in multiple directions to construct a histogram of oriented gradients, calculate the gray-level co-occurrence matrix at different combinations of directions and distances to extract multi-dimensional texture features including energy, entropy, contrast, and correlation, select the optimal wavelet basis function through adaptive weight fusion according to the multi-dimensional texture features, analyze the local extreme value characteristics and energy distribution of wavelet coefficients to calculate the optimal decomposition scale sequence, perform multi-layer decomposition based on the optimal decomposition scale sequence through row-column transformation and downsampling operations to obtain low-frequency approximation components and high-frequency detail components, and perform tree decomposition on the high-frequency detail components to generate a multi-resolution feature representation; Perform sparse coding on the multi-resolution feature representation by constructing an over-complete dictionary containing redundant atoms, divide the features into blocks of a fixed size, perform sparse decomposition based on orthogonal matching pursuit on each feature block, iteratively select the dictionary atoms most relevant to the residual and update the reconstruction coefficients until the sparsity constraint is satisfied to obtain a compressed feature code stream, input the compressed feature code stream into the decoupled representation network, extract multi-layer features through cascaded residual calculation units, where the residual calculation unit includes a convolutional layer, a normalization layer, and an activation function, extract context information at different scales through a multi-scale pooling module, gradually restore the feature resolution in combination with transposed convolution operations, and calculate channel weights based on feature statistics and non-linear transformations for separation to obtain a water level feature map; Calculate the local region structural similarity using a Gaussian weighted window based on the water level feature map, calculate the horizontal and vertical gradients to construct a gradient direction field and analyze the direction consistency, perform Fourier transform on the features to calculate the normalized cross-power spectrum, determine the best matching position through local maximum detection to obtain a complementary measure value, construct a feature fusion relationship graph based on the complementary measure value, where the nodes in the feature fusion relationship graph represent features and the weights of the edges are determined by the complementary measure value, calculate the feature correlation scores through a multi-layer graph attention calculation unit and normalize them, and adaptively fuse the normalized attention scores with the feature weighted combination to obtain an enhanced feature map; Input the enhanced feature map into a densely connected network, decompose the standard convolution into depth convolution and pointwise convolution through depthwise separable convolution to reduce the computational amount, adopt dense connections to enable the nodes in the subsequent layer to receive the features of all previous layers simultaneously for layer feature reuse, and obtain a water level region segmentation map and a water level measurement value through a squeeze-and-excitation mechanism to adaptively learn the importance weights of feature channels; According to the Bayesian inference framework, in a three-layer probability model including an observation layer, a latent variable layer, and a prior layer, the observed data and physical constraints are input as prior knowledge. Through the variational inference method, the parameters of the approximate posterior distribution are iteratively optimized to calculate the posterior probability distribution of water level measurement. Based on the posterior probability distribution, during the forward calculation of the network, by randomly deactivating some neurons and performing multiple samplings, the predicted mean and variance are calculated to estimate the measurement uncertainty value and generate a reliability score table; Perform spectral analysis on the selected high-confidence water level data set, achieve time-frequency analysis through continuous wavelet transform, use the spectral decomposition method to decompose and reconstruct the time series, and adopt an adaptive signal decomposition algorithm based on the envelope of extreme points to calculate multiple intrinsic characteristic functions. Extract the periodic feature vector and trend feature vector by controlling the decomposition process with the amplitude threshold.

[0008] In an alternative implementation, input the enhanced feature map into a densely connected network. Decompose the standard convolution into depth convolution and pointwise convolution through depthwise separable convolution to reduce the computational amount. Adopt dense connection to enable the nodes in the subsequent layer to receive the features of all previous layers simultaneously for layer feature reuse. Obtain the water level region segmentation map and water level measurement value through the squeeze-and-excitation mechanism to adaptively learn the importance weights of feature channels, including: Extract features from the enhanced feature map using depthwise separable convolution. Decompose the standard convolution into depth convolution and pointwise convolution. The depth convolution extracts spatial correlation features by performing convolution operations independently on each input channel, and the pointwise convolution combines the channel information of the spatial correlation features to obtain the recombined features; Input the recombined features into the first densely connected block, the second densely connected block, and the third densely connected block in sequence. Set multiple layers of depthwise separable convolutional layers in each densely connected block. The first layer of the first densely connected block extracts features from the recombined features to obtain the first-layer output features. The second layer concatenates the recombined features and the first-layer output features and then extracts features to obtain the second-layer output features. The third layer concatenates the recombined features, the first-layer output features, and the second-layer output features and then extracts features to obtain the third-layer output features; Reduce the spatial size of the feature map and increase the number of channels for the features between adjacent densely connected blocks through convolution operations and feature sampling operations to obtain multi-scale features. Perform global feature statistics on the feature map output by the densely connected block to obtain a channel description vector. Input the channel description vector into the first fully connected layer to obtain a compressed feature vector. Input the compressed feature vector into the second fully connected layer to obtain an excitation feature vector. Normalize the excitation feature vector to obtain the channel weight coefficient. Multiply the channel weight coefficient by the multi-scale features to obtain the weighted features; Adjust the weighted features output by the first dense connection block, the second dense connection block, and the third dense connection block to the same spatial resolution. After upsampling the weighted features of the third dense connection block and fusing them with the weighted features of the second dense connection block, a first fused feature is obtained. After upsampling the first fused feature and fusing it with the weighted features of the first dense connection block, a fused feature map is obtained. Input the fused feature map into the segmentation branch and the measurement branch respectively. Through the multi-layer convolution operation of the segmentation branch, a water level area segmentation map with the same size as the enhanced feature map is output. Through the feature pooling operation and the fully connected operation of the measurement branch, a water level measurement value is output.

[0009] In an alternative embodiment, a non-linear state space model is constructed based on the periodic feature vector and the trend feature vector. The state estimation sequence is calculated by combining particle filtering and a causal discovery network is constructed. Based on the causal discovery network, a causal relationship matrix of water level changes is calculated and decomposed into multiple sub-matrices. A hierarchical attention network is constructed at different spatial scales and executed to obtain a water level feature tensor. The encrypted model parameter update amount is obtained by combining the federated optimization objective function and the homomorphic encryption algorithm. The optimal sampling strategy vector is calculated by a pre-set multi-agent collaborative decision-making system and the water level is analyzed for anomalies to obtain an anomaly feature vector. The optimal sampling strategy vector and the anomaly feature vector are added to the hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix. By analyzing the causal relationship matrix, an anomaly propagation path map and an emergency monitoring parameter sequence are obtained, including: Continuously collect water level data for each monitoring site according to a preset sampling period. Perform wavelet decomposition on the water level data to obtain a periodic component and a trend component. Construct the periodic component into a periodic feature vector and the trend component into a trend feature vector. Take the periodic feature vector as the observation variable and the trend feature vector as the state variable to construct a non-linear state space model. Generate a preset number of particle samples for each state variable. Combine particle filtering and perform iterative calculation on the particle samples through importance sampling. Calculate the likelihood probability value of each particle sample based on the observation data as the particle weight, and resample the particle samples to obtain a state estimation sequence. Construct a causal discovery network based on the state estimation sequence. Calculate the transfer entropy value between nodes as the causal relationship strength to obtain a causal relationship matrix. Divide the causal relationship matrix into multiple regions according to the spatial positions of the monitoring points and decompose it into multiple sub-matrices. Construct a hierarchical attention network at different spatial scales. The first-layer attention network inputs a sub-matrix representing the internal causal relationship of the region, the second-layer attention network inputs a sub-matrix representing the causal relationship between adjacent regions, and the third-layer attention network inputs a sub-matrix representing the causal relationship between remote regions. The water level feature tensor is obtained by weighted fusion of features at different spatial scales through attention weights; Design a federated optimization objective function. Each monitoring site is used as a participant in federated learning. The homomorphic encryption algorithm is used to encrypt the model parameters. Each of the participants calculates the model parameter gradients using local sampled data and encrypts them. The server aggregates the encrypted model parameter gradients to obtain the encrypted model parameter update amount, and the global model parameters are obtained through iterative optimization for a preset number of rounds; Calculate the optimal sampling strategy through a pre-set multi-agent collaborative decision-making system. Among them, the multi-agent collaborative decision-making system includes multiple agents, and the agents respectively correspond to multiple monitoring regions. The agents interact the water level feature tensor information through a communication network, calculate the state value function based on the water level feature tensor and obtain the optimal sampling strategy vector using a collaborative decision-making algorithm, and perform anomaly analysis on the collected water level data to extract time series features and spatial distribution features to obtain an anomaly feature vector; Add the optimal sampling strategy vector and the anomaly feature vector to the hierarchical reinforcement learning network. The state evaluation sub-network in the hierarchical reinforcement learning network outputs a state vector, the action generation sub-network outputs a detection resource allocation scheme matrix, the value estimation sub-network predicts the long-term benefit of the detection resource allocation scheme to obtain a benefit vector, and the execution parameter matrix is obtained through policy optimization; Analyze the anomaly propagation path based on the execution parameter matrix and the causal relationship matrix, locate the anomaly event location based on the execution parameter matrix, trace back the anomaly propagation link through the causal relationship matrix, analyze the anomaly influence range and draw an anomaly propagation path diagram, determine the key monitoring regions and increase the sampling frequency, and generate an emergency monitoring parameter sequence including monitoring regions and monitoring frequencies.

[0010] In an optional implementation manner, analyzing the anomaly propagation path based on the execution parameter matrix and the causal relationship matrix, locating the anomaly event location based on the execution parameter matrix, tracing back the anomaly propagation link through the causal relationship matrix, analyzing the anomaly influence range and drawing an anomaly propagation path diagram, determining the key monitoring regions and increasing the sampling frequency, and generating an emergency monitoring parameter sequence including monitoring regions and monitoring frequencies includes: Based on the execution parameter matrix and the causality matrix, segment the water level data through a sliding time window, and calculate the transfer entropy value of the water level data sequence between each pair of monitoring points as the causality strength. Among them, the transfer entropy value is obtained by calculating the conditional probability and accumulating it, and fill the transfer entropy value into the corresponding position of the causality matrix; Construct an execution parameter matrix, where the row vectors of the execution parameter matrix correspond to monitoring points, and the column vectors correspond to execution parameters. The execution parameters include the mean deviation, variance ratio, trend slope, and fluctuation period obtained through statistical analysis of historical data. Based on the execution parameters, establish a multi-level anomaly discrimination rule, and combine the fuzzy inference mechanism to evaluate the anomaly degree of the real-time data of the monitoring points; Adopt a depth-first search strategy to perform backtracking analysis in the causality matrix. Starting from the anomaly point, determine the search direction based on the causality strength threshold, and preferentially visit adjacent nodes with the strongest causality. Maintain an access marker table to record the visited nodes, and obtain the complete propagation link from the anomaly source to the impact end through iterative search; Based on the complete propagation link, analyze the anomaly propagation path through a hierarchical processing strategy. Analyze the correlation between the anomaly point and the surrounding monitoring points in the local area to obtain local propagation characteristics, expand the analysis scope to adjacent areas to study the propagation mode at a larger spatial scale, and comprehensively obtain the anomaly propagation network structure at the global scale by integrating the analysis results of each level; According to the anomaly propagation network structure, spatially arrange the monitoring points and propagation paths through an adaptive layout algorithm. Construct a base map based on geographic information data and project the monitoring points onto the base map. Use Bezier curves to draw the propagation paths to achieve smooth transitions. Dynamically adjust the arrow size and color according to the propagation intensity. Combine the clustering analysis method to determine the key monitoring areas. Cluster the nodes in the anomaly propagation network according to their positions and anomaly degrees to identify the anomaly concentration areas, calculate the importance scores of each clustering area. The importance scores comprehensively consider the anomaly degree, influence range, and propagation speed. Divide the monitoring areas into levels according to the importance scores and determine the monitoring frequencies; Adopt the dynamic programming method to generate an emergency monitoring parameter sequence, establish an optimization model for monitoring resource allocation with the maximum anomaly monitoring effect as the objective function, use the number of monitoring devices and communication bandwidth as resource constraint conditions, solve the optimal monitoring parameter configuration scheme, and online adjust the monitoring parameters based on the anomaly development trend to obtain an emergency monitoring parameter sequence including monitoring areas and monitoring frequencies.

[0011] In the second aspect of the embodiments of the present invention, a distributed water level monitoring system based on a visual sensor network is provided, including: The first unit is used to collect water level image samples of the monitoring area, construct a spectral decomposition network to extract the frequency-domain features of the water level image samples, construct a multi-scale spectral feature matrix, construct an energy matrix based on the multi-scale spectral feature matrix and calculate the energy distribution density, generate an energy mask map corresponding to the water level image samples, guide the attention mechanism through the energy mask map to perform self-supervised decomposition on the image, obtain a water level scale feature tensor, a water surface texture feature tensor and an environmental background feature tensor, input the decomposed feature tensors into a multi-modal contrastive learning network, calculate a discriminative feature vector group through maximizing mutual information, construct a random field model based on the spatial distribution characteristics of visual sensing nodes, calculate the spatial dependence relationship between nodes to obtain a node association probability matrix, input the node association probability matrix and the discriminative feature vector group into a graph neural network, and obtain the parameter matrix of the water level propagation prediction model through message passing iteration; The second unit is used to receive real-time water level images, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate an optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream through sparse coding of the multi-resolution feature representation, input the feature code stream into a decoupled representation network, separate to obtain a water level feature map and calculate a complementarity metric value based on the water level feature map, perform adaptive fusion based on the complementarity metric value to obtain an enhanced feature map, input the enhanced feature map into a densely connected network, obtain a water level area segmentation map and a water level measurement value through cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement according to the Bayesian inference framework, combining prior knowledge and observation data, calculate a measurement uncertainty value based on the posterior probability distribution to generate a reliability scoring table, screen to obtain a high-confidence water level data set and perform spectral analysis, and extract a periodic feature vector and a trend feature vector; The third unit is used to construct a non-linear state space model based on the periodic feature vector and the trend feature vector, calculate a state estimation sequence through particle filtering and construct a causal discovery network, calculate a causal relationship matrix of water level changes based on the causal discovery network and decompose it into multiple sub-matrices, construct a hierarchical attention network at different spatial scales and execute to obtain a water level feature tensor, obtain an encrypted model parameter update amount through a federated optimization objective function combined with a homomorphic encryption algorithm, calculate an optimal sampling strategy vector through a pre-set multi-agent collaborative decision-making system and perform anomaly analysis on the water level to obtain an anomaly feature vector, add the optimal sampling strategy vector and the anomaly feature vector to a hierarchical reinforcement learning network to obtain a detection resource allocation plan and an execution parameter matrix, and analyze the anomaly propagation path map and the emergency monitoring parameter sequence in combination with the causal relationship matrix.

[0012] In the present invention, through technical means such as multi-modal contrast learning, adaptive wavelet transform, decoupled representation, and dense connection, water level features can be effectively extracted, the influence of noise and environmental interference can be reduced, and high-precision water level measurement can be achieved. The Bayesian inference framework and reliability scoring mechanism further improve the confidence of the measurement results, effectively avoid the influence of abnormal data, construct a non-linear state space model, combine particle filtering and causal discovery network, and can effectively analyze the periodicity, trend, and causal relationship of water level changes, and achieve the prediction of future water level changes. Using hierarchical reinforcement learning and multi-agent collaborative decision-making system, combined with the abnormal analysis results and causal relationship matrix, the allocation scheme and execution parameters of monitoring resources can be dynamically adjusted to achieve rapid response and effective disposal of abnormal water level events. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a schematic flowchart of the water level distributed monitoring method based on the vision sensor network according to an embodiment of the present invention; Figure 2 is a schematic structural diagram of the water level distributed monitoring system based on the vision sensor network according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0015] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0016] Figure 1 is a schematic flowchart of the water level distributed monitoring method based on the vision sensor network according to an embodiment of the present invention, as Figure 1 shown, the method includes: S1. Collect water level image samples in the monitoring area and construct a spectral decomposition network to extract the frequency domain features of the water level image samples, construct a multi-scale spectral feature matrix, construct an energy matrix based on the multi-scale spectral feature matrix and calculate the energy distribution density, generate an energy mask map corresponding to the water level image samples, and use the energy mask map to guide the attention mechanism to perform self-supervised decomposition on the image to obtain a water level scale feature tensor, a water surface texture feature tensor, and an environmental background feature tensor. Input the feature tensors obtained by decomposition into a multi-modal contrastive learning network, calculate a discriminative feature vector group through maximizing mutual information, construct a random field model based on the spatial distribution characteristics of visual sensing nodes, calculate the spatial dependence relationship between nodes to obtain a node association probability matrix, and input the node association probability matrix and the discriminative feature vector group into a graph neural network to obtain the parameter matrix of the water level propagation prediction model through message passing iteration; The spectral decomposition network is a deep learning model dedicated to extracting features of different frequency components from signals. It can decompose complex signals into multiple frequency sub-components to capture more hierarchical information. The energy mask map is a mask map generated by analyzing the energy distribution in different regions of a signal or image, used to emphasize or filter specific regions or frequency bands. The self-supervised decomposition is a learning method that relies on the intrinsic structure of the data itself for training without the need for manually labeled supervision signals and is commonly used in tasks such as feature decomposition and representation learning. The mutual information is a measure method for quantifying the mutual dependence relationship between two variables, describing the amount of information carried by one variable about another variable. The spatial distribution characteristic describes the distribution law or pattern of data in space, which can reflect how the data changes with spatial coordinates. The random field model is a class of mathematical models used to describe the random changes of spatial or temporal data, simulating the uncertainty in complex systems through random variables. The message passing iteration is a calculation method used to transmit information in a graph structure, updating information iteratively through the interaction between nodes, and is commonly used in graph neural networks and optimization problems.

[0017] In an alternative embodiment, water level image samples of a monitoring area are collected, and a spectral decomposition network is constructed to extract the frequency-domain features of the water level image samples, a multi-scale spectral feature matrix is constructed, an energy matrix is constructed based on the multi-scale spectral feature matrix, and the energy distribution density is calculated to generate an energy mask map corresponding to the water level image samples. The image is self-supervised decomposed by guiding the attention mechanism through the energy mask map to obtain a water level scale feature tensor, a water surface texture feature tensor, and an environmental background feature tensor. The feature tensors obtained by decomposition are input into a multi-modal contrastive learning network, and a discriminative feature vector group is obtained by maximizing the mutual information. A random field model is constructed based on the spatial distribution characteristics of the visual sensing nodes, the spatial dependence relationship between nodes is calculated to obtain a node association probability matrix, and the node association probability matrix and the discriminative feature vector group are input into a graph neural network. The parameter matrix of the water level propagation prediction model is obtained through message passing iteration, including: Collect water level image samples of the monitoring area, and input the water level image samples into a multi-level spectral decomposition network. The shallow network in the multi-level spectral decomposition network extracts the high-frequency components of the water level image samples through a high-pass filter bank to obtain water level scale tick features and water surface ripple features. The middle network extracts the intermediate-frequency components of the water level image samples through a band-pass filter bank to obtain water level line contour features and shore structure features. The deep network extracts the low-frequency components of the water level image samples through a low-pass filter bank to obtain water area distribution features. The feature maps corresponding to the high-frequency components, the intermediate-frequency components, and the low-frequency components are normalized and arranged and combined in the order of frequency bands to form a multi-scale spectral feature matrix; Calculate the energy values of the components in each frequency band of the multi-scale spectral feature matrix to obtain a two-dimensional energy matrix. The density estimation of the two-dimensional energy matrix is performed through a kernel function to obtain an energy density distribution map. An adaptive threshold is set based on the energy density distribution map, and a binary energy mask map is generated. The high-energy region and the low-energy region are determined. The high-energy region is marked as the foreground region, and the low-energy region is marked as the background region; Divide the water level image samples into multiple overlapping image patches, calculate the attention weights of the multiple overlapping image patches based on the binary energy mask map, input the overlapping image patches with the attention weights into a self-supervised deconstruction module for feature decomposition, extract the water level scale feature branch through a fine-grained feature extractor to obtain a water level scale feature tensor, extract the water surface texture feature branch through a texture analyzer to obtain a water surface texture feature tensor, extract the environmental background feature branch through a context encoder to obtain an environmental background feature tensor, input the water level scale feature tensor, the water surface texture feature tensor, and the environmental background feature tensor into a feature encoder respectively for dimensionality reduction to obtain feature vectors of a unified dimension, and construct a feature contrast pool; Randomly sample positive and negative sample pairs from the feature vectors of the unified dimension, calculate the mutual information metric values between the positive and negative sample pairs, obtain a discriminative feature vector group by maximizing the mutual information metric values of the positive sample pairs and minimizing the mutual information metric values of the negative sample pairs, calculate the distance matrix between node pairs based on the geographical coordinates of the visual sensing nodes, convert the distance matrix into an initial association strength, and calculate the interaction between node pairs through the potential function of the random field to obtain a node association probability matrix; Input the node association probability matrix and the discriminative feature vector group into the graph neural network. In each round of message passing iteration, determine the importance weights of neighbor nodes based on the node association probability matrix, perform a non-linear transformation on the feature information of the neighbor nodes to obtain fused features, combine the fused features with the feature information of the nodes themselves to update the node states, and obtain the parameter matrix of the water level propagation prediction model through multiple rounds of iteration.

[0018] The kernel function is a mathematical function used to measure the similarity between data points. The two-dimensional energy matrix is a matrix representing the energy distribution of a signal or image in a two-dimensional space. By analyzing the energy intensities at different positions, it helps in the modeling and recognition of spatial features. The fine-grained feature extractor refers to a network or algorithm that can capture subtle feature differences, usually used for high-precision analysis of images or signals, and can extract detailed information that is easily overlooked in large-scale analysis. The feature contrast pool is a mechanism for comparing and fusing different features, usually used to enhance the expression ability of features. By comparing the features in the contrast pool, the model's ability to identify different samples is improved.

[0019] Collect water level image samples of the monitoring area. The multi-level spectral decomposition network first performs multi-scale decomposition on the input water level image. The shallow network uses a high-pass filter bank, where multiple groups of filter kernels with different sizes are set, and the filter kernel sizes increase gradually from small to large to extract the features of the water level scale lines and the water surface ripples. The features of the scale lines are obtained through edge detection in the vertical direction, and the features of the water surface ripples are extracted through a direction-sensitive texture operator. The middle network processes the image using a band-pass filter bank, and the center frequency and bandwidth of the filters are determined according to the scale features of the water level line contour and the shore structure to extract the features of the water level line contour and the shore structure. The deep network uses a low-pass filter bank to extract low-frequency features such as water area distribution by gradually reducing the cut-off frequency. All the feature maps are normalized, mapping the amplitudes to a unified interval, and recombined into a multi-scale spectral feature matrix in the order from high frequency to low frequency.

[0020] Calculate the energy values for each frequency band component in the multi-scale spectral feature matrix to establish a two-dimensional energy matrix. Use a kernel function to perform density estimation on the two-dimensional energy matrix, and the bandwidth of the kernel function is optimized and determined by the cross-validation method. Calculate the global statistical characteristics based on the energy density distribution diagram, including mean, variance, skewness, and kurtosis, and adaptively set the threshold. Generate a binary energy mask image through threshold segmentation, mark the area above the threshold as the foreground area, and the area below the threshold as the background area. Perform morphological processing on the mask image to remove noise and fill holes, and divide the water level image sample into multiple overlapping image patches, and the overlapping degree is adaptively adjusted according to the complexity of the image content. Calculate the attention weight of each image patch based on the binary energy mask image, and the weight calculation comprehensively considers the proportion of foreground pixels in the patch, energy mean, and gradient information. The image patches with attention weights are input into the self-supervised deconstruction module, and features are extracted through three parallel branches: the fine-grained feature extractor extracts the water level scale feature branch to obtain the water level scale feature tensor, the texture analyzer extracts the water surface texture feature branch to obtain the water surface texture feature tensor, and the context encoder extracts the environmental background feature branch to obtain the environmental background feature tensor. The three feature tensors are respectively input into the feature encoder for dimensionality reduction to obtain feature vectors with a unified dimension, and a feature contrast pool is constructed.

[0021] Randomly sample from the feature contrast pool to construct positive sample pairs and negative sample pairs. The selection of positive sample pairs is based on temporal correlation and spatial proximity, and the selection of negative sample pairs ensures feature difference. Calculate the mutual information metric values between the positive sample pairs and negative sample pairs, and optimize to maximize the mutual information of the positive sample pairs and minimize the mutual information of the negative sample pairs to obtain a discriminative feature vector group. Calculate the distance matrix between node pairs based on the geographical coordinates of the visual sensing nodes, and convert the distance matrix into an initial association strength through an attenuation function. Calculate the interaction between node pairs through the potential function of the random field, and the design of the potential function considers the feature similarity and spatial relationship of node pairs, and iteratively optimize to obtain the node association probability matrix.

[0022] Input the node association probability matrix and the discriminative feature vector group into the graph neural network for training. In each round of message passing iteration, calculate the attention score based on the node association probability matrix to determine the importance weight of neighbor nodes. Perform a non-linear transformation on the feature information of neighbor nodes to obtain a fused feature. Combine the fused feature with the feature information of the node itself through a gating mechanism to update the node state. The training process adopts a batch processing method, uses an adaptive optimizer to update parameters, and the learning rate adopts a dynamic adjustment strategy. After multiple rounds of iteration, obtain a parameter matrix that can accurately predict the water level propagation law.

[0023] In this embodiment, through the self-supervised feature decomposition method guided by spectral decomposition and energy masking, the effective separation of scale features, water surface texture features, and environmental background features in the water level image is realized, improving the discriminability and robustness of feature expression. Based on the method of multi-modal contrast learning and random field modeling, the spatial correlation between visual sensing nodes is fully exploited, effectively improving the accuracy and generalization ability of water level propagation prediction. By adopting a message passing mechanism, the dynamic fusion and update of node features are realized, overcoming the problem of difficultly modeling complex water level propagation patterns and improving the adaptability of the prediction model.

[0024] In an alternative embodiment, positive sample pairs and negative sample pairs are randomly sampled from the feature vectors of the unified dimension, the mutual information metric value between the positive sample pairs and the negative sample pairs is calculated, and a discriminative feature vector group is obtained by maximizing the mutual information metric value of the positive sample pairs and minimizing the mutual information metric value of the negative sample pairs. The distance matrix between node pairs is calculated based on the geographical coordinates of the visual sensing nodes, the distance matrix is converted into an initial association strength, and the interaction between node pairs is calculated through the potential function of the random field to obtain the node association probability matrix, including: A feature contrast pool is constructed from the feature vectors of the unified dimension and the feature contrast pool is divided into multiple time windows. Scene feature information is extracted within each time window and positive sample pairs are randomly sampled. Negative sample pairs are randomly sampled from the feature vectors by adopting a multi-layer hierarchical sampling strategy according to weather conditions, monitoring time periods, and water level change trends in different monitoring scenarios; The feature vectors in the positive sample pairs and the negative sample pairs are added to a multi-layer perceptron including an input layer, multiple hidden layers, and an output layer for feature transformation. A batch normalization layer and a non-linear activation function are set between adjacent layers of the multi-layer perceptron. After mapping the feature vectors from the feature space to the probability distribution space, a Gaussian kernel function group with an adaptive bandwidth is constructed in the probability distribution space, and a continuous probability distribution is obtained by accumulating the local and global contributions of the probability density of all feature point pairs; Calculate the mutual information metric value between the positive sample pairs and the negative sample pairs. Among them, the mutual information metric value is obtained by calculating the multiple differences of the joint distribution, conditional distribution, and marginal distribution of the transformed features. An mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints is constructed. By maximizing the mutual information metric value of the positive sample pairs and minimizing the mutual information metric value of the negative sample pairs, combined with the gradient descent method with an adaptive learning rate, multi-round iterative optimization is carried out to obtain a feature vector group with local and global discriminability; Based on the geographical coordinates of visual sensing nodes, a distance matrix between nodes is calculated by fusing the geodetic distance, terrain undulation coefficient, river channel connectivity coefficient, and hydrological propagation characteristic coefficient between nodes. The multi-scale elevation difference between nodes is calculated in combination with digital elevation model data to dynamically correct the distance matrix. At the same time, the actual flow distance considering the water flow direction and velocity is calculated based on the river channel connectivity data, and the actual flow distance is multi-feature fused with the corrected distance matrix to obtain a corrected distance matrix; Based on the corrected distance matrix, the distance values are mapped to a dynamic interval through an adaptive normalization method. The local and global densities are calculated based on the multi-dimensional spatial distribution density characteristics of the nodes. An adaptive attenuation coefficient based on density gradient is set for the densely distributed area, and a dynamic attenuation coefficient considering spatial heterogeneity is set for the sparsely distributed area. The initial association strength at multiple scales is calculated in combination with the spatio-temporal correlation characteristics of the nodes; Based on the initial association strength, a random field potential function with an adaptive segmented structure is constructed. Based on the multi-dimensional spatial relationship between nodes, the node distance is dynamically divided into a short-distance interval, a medium-distance interval, and a long-distance interval with a transition zone. A linear attenuation function considering local spatial dependence is designed in the short-distance interval, an exponential attenuation function considering regional influence is designed in the medium-distance interval, and a residual function considering global association is designed in the long-distance interval. The adaptive segmentation points of the potential function are determined through spatial autocorrelation analysis; Based on the adaptive segmentation points, the interaction between node pairs is iteratively calculated through the potential function of the random field based on a multi-layer message passing mechanism. In each iteration, the state information of multi-scale neighborhood nodes is collected based on the spatial dependence relationship, the messages are multi-level weighted using the adaptive potential function values, the spatio-temporal redundant information is identified and suppressed through a structured attention mechanism, the weighted messages are non-linearly feature combined and random noise obeying a dynamic distribution is added, and the node association probability matrix characterizing the association relationship between nodes is obtained through state update.

[0025] The mutual information optimization objective function is an objective function that optimizes model parameters by maximizing mutual information, which is used to improve the model's learning ability of the dependence relationship between data and is widely applied in deep learning and optimization problems. The geodetic distance refers to the shortest distance between two points on the Earth's surface, usually considering the Earth's curvature and geographical coordinates. The terrain undulation coefficient is an index describing the degree of surface undulation, reflecting the size of terrain undulation. The river channel connectivity coefficient is an index measuring the connectivity degree of each river reach in the river network and is used in hydrological and ecological research to evaluate the connectivity of water flow and biological habitats. The hydrological propagation characteristic coefficient is a coefficient describing the water flow propagation speed and mode, which helps to simulate the relationship between precipitation and runoff. The digital elevation model data is a terrain elevation data set represented digitally. The dynamic attenuation coefficient considering spatial heterogeneity is a coefficient that describes how information or influence decays in space after considering spatial heterogeneity factors. The adaptive potential function value refers to a function value that adaptively adjusts according to data changes and dynamically adjusts parameters according to environmental changes. The spatio-temporal redundant information refers to the repetitive information generated due to time or space similarity in spatio-temporal data, which is usually processed through feature compression or redundancy removal techniques to help improve the data analysis efficiency.

[0026] Construct a feature comparison pool and divide it according to time windows, with each time window width set to 30 minutes. Positive sample pairs are randomly sampled according to scene information within the window, including samples with similar water levels under the same monitoring scene as positive samples. For negative sample sampling, three-layer stratification is carried out according to weather, monitoring period, and water level trend, and feature vectors are randomly sampled under different stratification combinations to construct negative sample pairs.

[0027] The sampled positive and negative sample pairs are input into a multi-layer perceptron for feature transformation. The multi-layer perceptron includes an input layer (dimension 256), 3 hidden layers (dimensions 128, 64, 32 respectively), and an output layer (dimension 16). A batch normalization layer and a ReLU activation function are added between adjacent layers. The transformed features construct a Gaussian kernel function group in the probability space, and the kernel bandwidth is adaptively determined by the average distance of neighboring sample points. The local (k = 5 nearest neighbors) and global contributions are calculated for each feature point to obtain a continuous probability distribution, and the mutual information measure between positive and negative sample pairs is calculated. The mutual information value is obtained by calculating the KL divergence of the joint distribution, conditional distribution, and marginal distribution of the transformed features. An optimization objective function is constructed, including feature discriminative constraints and distribution consistency constraints. The Adam optimizer is used, with the learning rate starting from 0.001 and decaying by 0.1 every 50 rounds, and iterating 500 rounds to obtain a discriminative feature vector group. The distance matrix between nodes is calculated based on the geographical coordinates of visual sensing nodes. The geodetic distance (weight 0.4), terrain undulation coefficient (weight 0.2), river channel connectivity coefficient (weight 0.2), and hydrological propagation feature coefficient (weight 0.2) are fused. The elevation difference at multiple scales (1km, 5km, 10km) is combined with DEM data to correct the distance matrix. Based on the river channel connectivity data, considering the water flow direction and velocity (0.5 - 2m / s), the actual flow distance is calculated and fused with the corrected distance matrix to obtain the final corrected distance matrix. The initial association strength is calculated based on the corrected distance matrix. The MinMax method is used to normalize the distance values to the [0, 1] interval. The local density and global density within 500m, 1km, and 2km of the nodes are calculated.

[0028] The node distances are divided into a short-distance (2km) interval, and the transition zone width is set to 100m. A linear decay function (slope -0.2) is used in the short-distance interval, an exponential decay function (decay rate 0.5) is used in the medium-distance interval, and a residual function is used in the long-distance interval. The segmentation points of the potential function are determined through spatial autocorrelation analysis, and the node interaction is iteratively calculated based on a three-layer message passing mechanism. The status information of neighboring nodes within 200m, 500m, and 1km is collected in each round of iteration, and the potential function values are used to weight the messages. The spatio-temporal redundant information is identified through structured attention, the weighted messages are non-linearly combined and Gaussian random noise (mean 0, variance 0.1) is added to update the node status to obtain the association probability matrix.

[0029] In this embodiment, through the hierarchical sampling strategy and multi-layer perceptron feature transformation, combined with the mutual information optimization objective, the discriminability and representational ability of the feature vector are improved, the recognition accuracy of node association relationships is enhanced, the multi-dimensional spatial information is fused to calculate the corrected distance matrix, the initial association strength is calculated using an adaptive normalization and density-aware attenuation mechanism, the rationality of the node spatial association representation is improved, an adaptive segmented random field potential function is designed, and the node interaction is calculated based on the multi-layer message passing and structured attention mechanism, enhancing the robustness and generalization ability of the node association probability.

[0030] S2. Receive real-time water level images, input the parameter matrix of the water level propagation prediction model into the adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream by performing sparse coding on the multi-resolution feature representation, input the feature code stream into the decoupled representation network, obtain a water level feature map by separation and calculate a complementarity metric value based on the water level feature map, perform adaptive fusion based on the complementarity metric value to obtain an enhanced feature map, input the enhanced feature map into the densely connected network, obtain a water level region segmentation map and a water level measurement value through cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement according to the Bayesian inference framework, combining prior knowledge and observation data, generate a reliability score table based on the measurement uncertainty value calculated from the posterior probability distribution, screen to obtain a high-confidence water level data set and perform spectral analysis, and extract a periodic feature vector and a trend feature vector; The sparse coding is a feature learning method that represents the input signal as a linear combination of a few basis elements to achieve the purpose of data compression and efficient representation. The compressed feature code stream refers to representing the features in a compact code stream format through a specific compression algorithm to reduce storage and transmission overheads. The decoupled representation network is a neural network architecture that aims to decompose the complex representation of the input data into simpler and independent components to improve the interpretability and training efficiency of the model. The complementarity metric value is an index used to evaluate the degree of complementarity between two sets of features or information. By calculating the complementarity metric, the parts of the data that have not been fully utilized can be discovered. The densely connected network is a network architecture that promotes the full propagation of information flow by connecting each layer to all previous layers, thereby improving the training efficiency and performance of the network. The Bayesian inference framework is an inference method based on Bayes' theorem for inferring unknown variables or parameters given prior information.

[0031] In an alternative embodiment, a real-time water level image is received, the parameter matrix of the water level propagation prediction model is input into an adaptive wavelet transform network, an optimal decomposition scale sequence is calculated and a multi-resolution feature representation is generated, a compressed feature code stream is obtained by performing sparse coding on the multi-resolution feature representation, the feature code stream is input into a decoupled representation network, a water level feature map is separated and a complementary metric value is calculated based on the water level feature map, an enhanced feature map is adaptively fused based on the complementary metric value, the enhanced feature map is input into a densely connected network, a water level region segmentation map and a water level measurement value are obtained through cross-layer feature reuse, according to the Bayesian inference framework, the posterior probability distribution of the water level measurement is calculated by combining prior knowledge and observed data, a measurement uncertainty value is calculated based on the posterior probability distribution to generate a reliability score table, a high-confidence water level data set is screened and spectral analysis is performed, and a periodic feature vector and a trend feature vector are extracted, including: A real-time water level image is received, the parameter matrix of the water level propagation prediction model is input into an adaptive wavelet transform network, gradients of the image in multiple directions are calculated to construct a histogram of oriented gradients, a gray-level co-occurrence matrix is calculated at different combinations of directions and distances to extract multi-dimensional texture features including energy, entropy, contrast, and correlation, an optimal wavelet basis function is selected through adaptive weight fusion according to the multi-dimensional texture features, the optimal decomposition scale sequence is calculated by analyzing the local extreme value characteristics and energy distribution of wavelet coefficients, and multi-layer decomposition is performed based on the optimal decomposition scale sequence through row-column transformation and downsampling operations to obtain a low-frequency approximation component and high-frequency detail components, and the high-frequency detail components are tree-structured decomposed to generate a multi-resolution feature representation; The multi-resolution feature representation is sparsely coded by constructing an over-complete dictionary containing redundant atoms, the features are blocked according to a fixed size, and sparse decomposition based on orthogonal matching pursuit is performed on each feature block, iteratively selecting the dictionary atoms most relevant to the residual and updating the reconstruction coefficients until the sparsity constraint is satisfied to obtain a compressed feature code stream, the compressed feature code stream is input into a decoupled representation network, multi-layer features are extracted through cascaded residual calculation units, where the residual calculation unit includes a convolutional layer, a normalization layer, and an activation function, different-scale context information is extracted through a multi-scale pooling module, the feature resolution is gradually restored by combining deconvolution operations, and channel weights are calculated based on feature statistics and non-linear transformation to separate and obtain a water level feature map; Calculate the local region structural similarity using a Gaussian weighted window based on the water level feature map, calculate the horizontal and vertical gradients to construct a gradient direction field and analyze the direction consistency, perform a Fourier transform on the features to calculate the normalized cross-power spectrum, determine the optimal matching position through local maximum detection to obtain a complementary measure value, construct a feature fusion relationship graph based on the complementary measure value, where the nodes in the feature fusion relationship graph represent features and the weights of the edges are determined by the complementary measure value, calculate the feature correlation score through a multi-layer graph attention calculation unit and normalize it, and adaptively fuse the normalized attention score with the feature weighted combination to obtain an enhanced feature map; Input the enhanced feature map into a densely connected network, decompose the standard convolution into depth convolution and pointwise convolution through depthwise separable convolution to reduce the computational amount, adopt dense connection to enable the nodes in the subsequent layer to receive the features of all previous layers simultaneously for layer feature reuse, and obtain a water level region segmentation map and a water level measurement value through a squeeze-and-excitation mechanism to adaptively learn the importance weights of feature channels; According to the Bayesian inference framework, in a three-layer probability model including an observation layer, a latent variable layer, and a prior layer, input the observation data and physical constraints as prior knowledge, calculate the posterior probability distribution of the water level measurement by iteratively optimizing the approximate posterior distribution parameters through a variational inference method, and based on the posterior probability distribution, randomly deactivate some neurons during the forward calculation of the network and perform multiple samplings to calculate the prediction mean and variance to estimate the measurement uncertainty value and generate a reliability score table; Perform spectral analysis on the selected high-confidence water level data set, achieve time-frequency analysis through continuous wavelet transform, use a spectral decomposition method to decompose and reconstruct the time series, calculate multiple intrinsic feature functions using an adaptive signal decomposition algorithm based on the envelope of extreme points, and extract the periodic feature vector and trend feature vector by controlling the decomposition process through an amplitude threshold.

[0032] The gray-level co-occurrence matrix is a texture analysis method that describes the texture features of an image by calculating the spatial relationship of pixel gray values in the image. The row-column transformation refers to the operation of exchanging the rows and columns of a matrix. The downsampling operation refers to the operation of reducing the computational complexity by reducing the sampling frequency or the number of pixels in signal processing or image processing. The tree decomposition is the process of decomposing data or problems into a tree structure. The redundant atom refers to the basis atom that does not have independence in dictionary learning. The overcomplete dictionary is a dictionary composed of multiple basis elements, where the number of basis elements is more than the dimension of the signal, allowing for a better signal reconstruction through sparse representation. The adaptive signal decomposition algorithm based on the envelope of extreme points is an algorithm that adaptively decomposes a signal by extracting the extreme points in the signal.

[0033] Obtain real-time water level images. And transmit the images to the server for processing, and preprocess the water level images. This includes operations such as image denoising, correction, and enhancement to improve the image quality and the accuracy of subsequent processing. For example, median filtering is used to remove image noise, geometric transformation is used to correct image distortion, and histogram equalization is used to enhance image contrast. Input the preprocessed water level image into the adaptive wavelet transform network for feature extraction. First, input the parameter matrix of the water level propagation prediction model into the network. The network analyzes the gray-scale changes of the image in different directions and distances, extracts multi-dimensional texture features, adaptively selects the optimal wavelet basis function according to the texture features, and calculates the optimal decomposition scale sequence. Based on this sequence, perform multi-layer wavelet decomposition on the image to obtain the low-frequency approximation component and the high-frequency detail component. Perform tree decomposition on the high-frequency detail component to generate a multi-resolution feature representation.

[0034] Perform sparse coding on the multi-resolution feature representation. Construct an over-complete dictionary containing redundant atoms, divide the features into blocks of a fixed size, and perform sparse decomposition based on orthogonal matching pursuit on each feature block. Iteratively select the dictionary atoms most relevant to the residual and update the reconstruction coefficients until the sparsity constraint is satisfied to obtain the compressed feature code stream. For example, use a dictionary containing 1024 atoms and decompose each feature block into at most 10 atoms. Input the compressed feature code stream into the decoupled representation network. The decoupled representation network extracts multi-layer features through cascaded residual calculation units. Each residual calculation unit contains a convolutional layer, a normalization layer, and an activation function. Extract context information at different scales through a multi-scale pooling module and gradually restore the feature resolution by combining deconvolution operations. Calculate the channel weights based on the feature statistics and non-linear transformation, and perform separation to obtain the water level feature map.

[0035] Calculate the complementary measure value based on the water level feature map. Use a Gaussian weighted window to calculate the local region structural similarity, calculate the horizontal and vertical gradients to construct the gradient direction field and analyze the direction consistency, perform Fourier transform on the features to calculate the normalized cross-power spectrum, and determine the best matching position through local maximum detection to obtain the complementary measure value. For example, use a Gaussian window of size 5x5 to calculate the structural similarity.

[0036] Perform adaptive fusion based on the complementary measure value. Construct a feature fusion relationship graph, where the nodes in the graph represent features and the weights of the edges are determined by the complementary measure value. Calculate the feature correlation scores through a multi-layer graph attention calculation unit and normalize them. Combine the normalized attention scores with the features in a weighted manner to obtain the enhanced feature map. Input the enhanced feature map into the densely connected network. The network uses depthwise separable convolution and dense connections to achieve cross-layer feature reuse. Adaptively learn the importance weights of the feature channels through a squeeze-and-excitation mechanism to obtain the water level region segmentation map and the water level measurement value.

[0037] According to the Bayesian inference framework, calculate the posterior probability distribution and uncertainty of water level measurement. Input the observed data and physical constraints as prior knowledge into a three-layer probability model, and iteratively optimize the parameters of the approximate posterior distribution through variational inference method to calculate the posterior probability distribution of water level measurement. Based on the posterior probability distribution, in the forward calculation of the network, randomly inactivate some neurons and perform multiple samplings to calculate the prediction mean and variance, estimate the measurement uncertainty value, and generate a reliability score table.

[0038] Perform spectral analysis on the high-confidence water level dataset. Achieve time-frequency analysis through continuous wavelet transform, use spectral decomposition method to decompose and reconstruct the time series, adopt an adaptive signal decomposition algorithm based on the envelope of extreme points to calculate multiple intrinsic characteristic functions, and control the decomposition process through amplitude threshold to extract the periodic feature vector and trend feature vector. For example, extract the daily and annual cycle features of water level changes.

[0039] In this embodiment, through technologies such as multi-scale feature extraction, sparse coding, decoupled representation, and adaptive fusion, the expression ability of water level features is effectively improved, thereby improving the accuracy of water level measurement. Based on the Bayesian inference framework, combining prior knowledge and observed data, calculate the posterior probability distribution and uncertainty of water level measurement, generate a reliability score table, improve the reliability of water level measurement, and extract the periodic feature vector and trend feature vector through spectral analysis, which can predict the water level change trend and provide a scientific basis for water resource management and flood prevention warning.

[0040] In an alternative embodiment, input the enhanced feature map into a densely connected network. Decompose the standard convolution into depth convolution and pointwise convolution through depthwise separable convolution to reduce the computational amount. Adopt dense connection so that the nodes in the later layer receive the features of all previous layers simultaneously for layer feature reuse. Obtain the water level region segmentation map and water level measurement value through the squeeze-and-excitation mechanism to adaptively learn the importance weights of feature channels, including: Perform feature extraction on the enhanced feature map using depthwise separable convolution, decompose the standard convolution into depth convolution and pointwise convolution. The depth convolution extracts spatial correlation features by performing convolution operations independently on each input channel, and the pointwise convolution combines the channel information of the spatial correlation features to obtain the recombined features; Input the recombined features into the first densely connected block, the second densely connected block, and the third densely connected block in sequence. Set multiple layers of depthwise separable convolutional layers in each densely connected block. The first layer of the first densely connected block extracts features from the recombined features to obtain the first layer output features. The second layer concatenates the recombined features and the first layer output features and then extracts features to obtain the second layer output features. The third layer concatenates the recombined features, the first layer output features, and the second layer output features and then extracts features to obtain the third layer output features; For the features between adjacent dense connection blocks, the spatial dimension of the feature map is reduced and the number of channels is increased through convolution operations and feature sampling operations to obtain multi-scale features. Global feature statistics are performed on the feature map output by the dense connection blocks to obtain a channel description vector. The channel description vector is input into the first fully connected layer to obtain a compressed feature vector. The compressed feature vector is input into the second fully connected layer to obtain an excitation feature vector. The excitation feature vector is normalized to obtain a channel weight coefficient. The multi-scale features are multiplied by the channel weight coefficient to obtain weighted features; The weighted features output by the first dense connection block, the second dense connection block, and the third dense connection block are adjusted to the same spatial resolution. The weighted feature of the third dense connection block is upsampled and fused with the weighted feature of the second dense connection block to obtain a first fusion feature. The first fusion feature is upsampled and fused with the weighted feature of the first dense connection block to obtain a fusion feature map. The fusion feature map is respectively input into the segmentation branch and the measurement branch. Through the multi-layer convolution operations of the segmentation branch, a water level area segmentation map with the same size as the enhanced feature map is output. Through the feature pooling operation and the fully connected operation of the measurement branch, a water level measurement value is output.

[0041] The depthwise separable convolutional layer is a convolutional neural network layer that decomposes the standard convolution operation into depthwise convolution and pointwise convolution, thereby significantly reducing the computational amount and improving the efficiency. It is commonly found in lightweight network structures. The channel description vector is a vector used to describe the features of each channel in a convolutional neural network. By encoding the features of each channel, the features of the network can be compressed and characterized.

[0042] Perform depthwise separable convolution processing on the enhanced feature map. Decompose the standard convolution operation into two consecutive steps: The first step is depthwise convolution, which independently performs convolution operations on each channel of the input feature map to extract the feature correlation in the spatial dimension; the second step is pointwise convolution, which performs a linear combination on the feature map output by the depthwise convolution in the channel dimension to achieve the recombination and fusion of information between channels. The densely connected structure uses three densely connected blocks in series to process the recombined features. Each densely connected block contains three layers of depthwise separable convolutional layers, and the layers are densely connected to each other. Taking the first densely connected block as an example: The first layer directly performs convolution processing on the recombined features to obtain the output of the first layer; the second layer performs convolution processing after concatenating the recombined features and the output of the first layer in the channel dimension; the third layer performs convolution processing again after concatenating all the recombined features, the output of the first layer, and the output of the second layer. The second and third densely connected blocks use the same connection method. A feature transformation module is set between adjacent densely connected blocks. The spatial resolution of the feature map is reduced through convolution operations, and at the same time, the number of feature channels is increased to obtain feature representations at multiple scales. Perform global average pooling on the feature map output by each densely connected block to obtain a description vector reflecting the importance of each channel. Input the channel description vector into a two-layer fully connected network. The first layer plays a role in feature compression, and the second layer performs feature excitation. Finally, the channel weight coefficients are obtained through normalization. Multiply the weight coefficients by the multi-scale features to achieve adaptive weighting of the features. The feature fusion stage adopts a bottom-up progressive fusion strategy. The features of the third densely connected block are fused with the features of the second densely connected block through upsampling to obtain the first-level fusion features; then the first-level fusion features are continuously upsampled and fused with the features of the first densely connected block to obtain the fused feature map. The task output stage contains two parallel branches. The segmentation branch processes the fused feature map through multiple layers of convolution operations and finally outputs a water level region segmentation map with the same size as the input image to achieve pixel-level region recognition. The measurement branch performs feature pooling and dimensionality reduction on the fused feature map and regresses the specific water level measurement value through a fully connected layer.

[0043] Exemplarily, an enhanced feature map with a resolution of 1920×1080 is input, which contains 64 feature channels. Each channel is independently processed using a 3×3 convolutional kernel, and a 1×1 convolutional kernel is used to reorganize the 64 channels into 128 channels. The three-layer structure of the first dense connection block: the first layer inputs 128-channel features and outputs 32-channel features; the second layer inputs 160-channel features (128 + 32) and outputs 32-channel features; the third layer inputs 192-channel features (128 + 32 + 32) and outputs 32-channel features. The second and third dense connection blocks adopt the same channel configuration. The feature transformation module uses a convolutional layer with a stride of 2 to reduce the feature map size to 960×540, 480×270, and 240×135 in sequence. The number of channels is correspondingly increased to 256, 512, and 1024. A 1024-dimensional channel description vector is obtained through global average pooling. The first fully connected layer compresses it to 64 dimensions, and the second fully connected layer remaps it to 1024 dimensions. After normalization, the channel weights are obtained. During feature fusion, the features with a size of 240×135 are upsampled to 480×270 for fusion with the features of the second dense connection block, and then upsampled to 960×540 for fusion with the features of the first dense connection block. The segmentation branch restores the feature map to a resolution of 1920×1080 through 4 convolutional layers and outputs the probability map of the water level area. The measurement branch outputs the water level value within the range of 0 - 100 through global average pooling and two fully connected layers.

[0044] In this embodiment, depthwise separable convolution is used to decompose standard convolution into depthwise convolution and pointwise convolution, effectively reducing the number of model parameters and computational complexity, improving computational efficiency. The dense connection mechanism enables the subsequent nodes of the network to simultaneously receive the features of all previous layers, realizing layer feature reuse and enhancing the feature expression ability of the network. The squeeze-and-excitation mechanism adaptively learns the importance weights of feature channels, further improving the network's ability to extract key information. Through the design of multi-scale feature fusion and the segmentation and measurement branches, the water level area can be accurately segmented and the water level value can be accurately measured, providing reliable data support for water resource management and flood control and disaster reduction.

[0045] S3. Construct a non - linear state - space model based on the periodic feature vector and the trend feature vector, calculate the state - estimation sequence through particle filtering and construct a causal discovery network. Calculate the causal relationship matrix of water - level change based on the causal discovery network and decompose it into multiple sub - matrices. Construct a hierarchical attention network at different spatial scales and execute to obtain the water - level feature tensor. Obtain the encrypted model - parameter update amount through the federated optimization objective function combined with the homomorphic encryption algorithm. Calculate the optimal sampling - strategy vector through a pre - set multi - agent collaborative decision - making system and perform anomaly analysis on the water level to obtain the anomaly feature vector. Add the optimal sampling - strategy vector and the anomaly feature vector to the hierarchical reinforcement - learning network to obtain the detection - resource allocation plan and the execution - parameter matrix. Analyze in combination with the causal relationship matrix to obtain the anomaly - propagation path diagram and the emergency - monitoring parameter sequence.

[0046] The non - linear state - space model is a mathematical model used to describe the dynamic behavior of a system, applicable to cases where there are non - linear relationships between system states. The particle filter is a recursive estimation algorithm based on the Monte Carlo method, using a set of particles (i.e., samples) to represent the probability distribution of the state space, capable of handling state - estimation problems for non - linear and non - Gaussian systems. The causal relationship matrix is a matrix that describes the causal relationships between different variables, usually obtained through statistical methods or model learning, helping to identify the causal - dependence structure in the system. The federated optimization objective function is the objective function used when optimizing in an environment with multi - party data and privacy constraints, usually ensuring that each party's calculation is independent and completing global optimization by aggregating local information from all parties. The homomorphic encryption algorithm is an encryption method that allows computational operations to be performed on encrypted data without decrypting the data, thus ensuring data privacy. The encrypted model - parameter update amount refers to the measure of how the parameters change and are updated during the calculation process in a homomorphic encryption model, used to guide the process optimization of encryption operations. The multi - agent collaborative decision - making system refers to a system in which multiple agents interact and cooperate based on each other's states and goals to jointly make decisions.

[0047] In an alternative embodiment, a non-linear state space model is constructed based on the periodic feature vector and the trend feature vector. The state estimation sequence is calculated by combining particle filtering and a causal discovery network is constructed. Based on the causal discovery network, a causal relationship matrix of water level changes is calculated and decomposed into multiple sub-matrices. A hierarchical attention network is constructed at different spatial scales and the water level feature tensor is obtained through execution. The encrypted model parameter update amount is obtained by combining the federated optimization objective function with the homomorphic encryption algorithm. The optimal sampling strategy vector is calculated through a pre-set multi-agent collaborative decision-making system and the water level is analyzed for anomalies to obtain the anomaly feature vector. The optimal sampling strategy vector and the anomaly feature vector are added to the hierarchical reinforcement learning network to obtain the detection resource allocation scheme and the execution parameter matrix. Based on the causal relationship matrix, the anomaly propagation path map and the emergency monitoring parameter sequence are analyzed, including: Continuously collect water level data for each monitoring site according to a preset sampling period. Perform wavelet decomposition on the water level data to obtain the periodic component and the trend component. Construct the periodic component as the periodic feature vector and the trend component as the trend feature vector; Use the periodic feature vector as the observation variable and the trend feature vector as the state variable to construct a non-linear state space model. Generate a preset number of particle samples for each state variable. Iteratively calculate the particle samples through importance sampling in combination with particle filtering. Calculate the likelihood probability value of each particle sample based on the observation data as the particle weight, and resample the particle samples to obtain the state estimation sequence; Construct a causal discovery network based on the state estimation sequence. Calculate the transfer entropy value between nodes as the causal relationship strength to obtain the causal relationship matrix. Divide the causal relationship matrix into multiple regions according to the spatial positions of the monitoring points and decompose it into multiple sub-matrices; Construct a hierarchical attention network at different spatial scales. The input of the first-layer attention network is the sub-matrix representing the internal causal relationship of the region. The input of the second-layer attention network is the sub-matrix representing the causal relationship between adjacent regions. The input of the third-layer attention network is the sub-matrix representing the causal relationship between remote regions. The features at different spatial scales are weighted and fused through the attention weights to obtain the water level feature tensor; Design a federated optimization objective function. Take each monitoring site as a participant in federated learning. Encrypt the model parameters using the homomorphic encryption algorithm. Each participant calculates the model parameter gradient using local sampling data and encrypts it. The server aggregates the encrypted model parameter gradients to obtain the encrypted model parameter update amount. After a preset number of rounds of iterative optimization, the global model parameters are obtained; Calculate the optimal sampling strategy through a pre-set multi-agent collaborative decision-making system. Among them, the multi-agent collaborative decision-making system includes multiple agents, and the agents respectively correspond to multiple monitoring areas. The agents interact the water level feature tensor information through a communication network, calculate the state value function based on the water level feature tensor, and use a collaborative decision-making algorithm to obtain the optimal sampling strategy vector, and perform anomaly analysis on the collected water level data to extract time series features and spatial distribution features to obtain an anomaly feature vector; Add the optimal sampling strategy vector and the anomaly feature vector to the hierarchical reinforcement learning network. The state evaluation sub-network in the hierarchical reinforcement learning network outputs a state vector, the action generation sub-network outputs a detection resource allocation scheme matrix, and the value estimation sub-network predicts the long-term benefit of the detection resource allocation scheme to obtain a benefit vector, and obtains an execution parameter matrix through policy optimization; Analyze the anomaly propagation path based on the execution parameter matrix and the causal relationship matrix, locate the anomaly event location based on the execution parameter matrix, trace back the anomaly propagation link through the causal relationship matrix, analyze the anomaly influence range and draw an anomaly propagation path diagram, determine the key monitoring area and increase the sampling frequency, and generate an emergency monitoring parameter sequence including the monitoring area and the monitoring frequency.

[0048] The transfer entropy value is a statistic used to quantify the information flow between systems, reflecting the strength and direction of the causal relationship between systems, and is widely used in time series analysis and system modeling. The collaborative decision-making algorithm is a decision-making algorithm in a multi-agent system, aiming to achieve global optimal decision-making through information exchange and sharing among agents, and is commonly used in fields such as distributed control and robot collaboration. The benefit vector is a vector used to describe the benefits obtained by each agent or participant under a specific decision, and is usually used in game theory and decision analysis, reflecting the result and effect of the decision. The anomaly propagation path refers to the path through which an anomaly event propagates from one node to other nodes in a network or system, and is widely used in fault diagnosis and anomaly detection systems.

[0049] Water level data is collected from each monitoring station, and the collection period can be preset, for example, once per hour. After obtaining the water level data, wavelet decomposition technology is used to decompose the water level data into periodic components and trend components, and periodic feature vectors and trend feature vectors are constructed respectively. For example, the water level fluctuation situation within a day is constructed into a periodic feature vector, and the overall rising or falling trend of the water level over a period of time is constructed into a trend feature vector, and a nonlinear state space model is constructed. The periodic feature vector is used as the observation variable of the model, and the trend feature vector is used as the state variable of the model. The model assumes that the trend change of the water level will affect the periodic fluctuation of the water level. The particle filter algorithm is used to estimate the state of the model. For each state variable, a certain number of particle samples are generated. Each particle sample represents a possible value of the state variable. According to the observed data, the likelihood probability value of each particle sample is calculated as the weight of the particle. The higher the weight of the particle, the more consistent the corresponding state variable value is with the observed data. The particles are resampled, and the particles with high weights are used to replace the particles with low weights to obtain the state estimation sequence.

[0050] Based on the obtained state estimation sequence, a causal discovery network is constructed. The nodes in the network represent different monitoring stations, and the connections between the nodes represent the possible causal relationships between the stations. The strength of the causal relationship is measured by calculating the transfer entropy value between the nodes, and a causal relationship matrix is constructed. The higher the transfer entropy value, the stronger the causal relationship. The causal relationship matrix is divided into multiple regions according to the spatial positions of the monitoring points and decomposed into multiple sub-matrices. For example, all monitoring stations are divided into three regions: upstream, middle, and downstream, and the causal relationship sub-matrices within the regions and between the regions are obtained respectively.

[0051] A hierarchical attention network is constructed. The network is divided into three layers, corresponding to different spatial scales. The input of the first-layer attention network is the sub-matrix representing the causal relationship within the region, such as the causal relationship sub-matrix within the upstream region. The input of the second-layer attention network is the sub-matrix representing the causal relationship between adjacent regions, such as the causal relationship sub-matrix between the upstream and middle regions. The input of the third-layer attention network is the sub-matrix representing the causal relationship between remote regions, such as the causal relationship sub-matrix between the upstream and downstream regions. The network uses the attention mechanism to perform weighted fusion on the features at different spatial scales to obtain the water level feature tensor. For example, if the abnormal water level change in the upstream region has a greater impact on the middle region, the network will assign a higher weight to the causal relationship sub-matrix between the upstream and middle regions.

[0052] Each monitoring site is regarded as a participant in federated learning. Each participant calculates the gradient of the model parameters using the locally collected water level data, encrypts the gradient using the homomorphic encryption algorithm, and then uploads the encrypted gradient to the server. The server aggregates the encrypted gradients to obtain the encrypted model parameter update amount, and then distributes the update amount to each participant to update the local model parameters. After multiple rounds of iterative optimization, the global model parameters are obtained. Each agent in the multi-agent collaborative decision-making system corresponds to a monitoring area, and the agents exchange water level feature tensor information through a communication network. Each agent calculates the state value function based on the received water level feature tensor and adopts a collaborative decision-making algorithm to obtain the optimal sampling strategy vector. For example, if the water level feature tensor of a certain area indicates a high probability of abnormality in that area, the agent corresponding to that area will increase the sampling frequency. At the same time, anomaly analysis is performed on the collected water level data, time series features and spatial distribution features are extracted to obtain the anomaly feature vector. The optimal sampling strategy vector and the anomaly feature vector are input into the hierarchical reinforcement learning network. This network includes a state evaluation sub-network, an action generation sub-network, and a value estimation sub-network. The state evaluation sub-network outputs the state vector, the action generation sub-network outputs the detection resource allocation scheme matrix, and the value estimation sub-network predicts the long-term benefits of the detection resource allocation scheme. Through policy optimization, the execution parameter matrix is obtained.

[0053] Analyze the abnormal propagation path based on the execution parameter matrix and the causality matrix. Locate the position of the abnormal event according to the execution parameter matrix, trace back the abnormal propagation link through the causality matrix, analyze the abnormal influence range, and draw the abnormal propagation path diagram. Determine the key monitoring areas and increase the sampling frequency to generate an emergency monitoring parameter sequence including the monitoring areas and monitoring frequencies.

[0054] In this embodiment, through multi-agent collaborative decision-making and hierarchical reinforcement learning, the sampling strategy and detection resource allocation can be optimized, the anomaly detection efficiency can be improved, and the human and material costs can be reduced. Combining the non-linear state space model and the particle filter algorithm can more accurately estimate the water level state, thereby improving the accuracy of anomaly detection. By constructing a causal discovery network and analyzing the abnormal propagation path, the origin and propagation process of the abnormal event can be traced, providing a scientific basis for emergency response.

[0055] In an alternative implementation, analyzing the abnormal propagation path based on the execution parameter matrix and the causality matrix, locating the position of the abnormal event based on the execution parameter matrix, tracing back the abnormal propagation link through the causality matrix, analyzing the abnormal influence range, and drawing the abnormal propagation path diagram, determining the key monitoring areas and increasing the sampling frequency, and generating an emergency monitoring parameter sequence including the monitoring areas and monitoring frequencies includes: Based on the execution parameter matrix and the causality matrix, the water level data is segmented by a sliding time window, and the transfer entropy value of the water level data sequence between each pair of monitoring points is calculated as the causality strength. Among them, the transfer entropy value is obtained by calculating the conditional probability and accumulating it, and the transfer entropy value is filled into the corresponding position of the causality matrix; Construct an execution parameter matrix, where the row vectors of the execution parameter matrix correspond to the monitoring points and the column vectors correspond to the execution parameters. The execution parameters include the mean deviation, variance ratio, trend slope, and fluctuation period obtained through statistical analysis of historical data. Based on the execution parameters, a multi-level anomaly discrimination rule is established, and the fuzzy inference mechanism is combined to evaluate the anomaly degree of the real-time data of the monitoring points; Adopt a depth-first search strategy to perform backtracking analysis in the causality matrix. Starting from the anomaly point, determine the search direction based on the causality strength threshold, preferentially visit the adjacent nodes with the strongest causality, maintain an access mark table to record the visited nodes, and obtain the complete propagation link from the anomaly source to the influence end through iterative search; Based on the complete propagation link, analyze the anomaly propagation path through a hierarchical processing strategy. Analyze the correlation between the anomaly point and the surrounding monitoring points in the local area to obtain the local propagation characteristics, expand the analysis scope to the adjacent area to study the propagation mode at a larger spatial scale, and comprehensively obtain the anomaly propagation network structure at the global scale by integrating the analysis results of each level; According to the anomaly propagation network structure, spatially arrange the monitoring points and propagation paths through an adaptive layout algorithm. Construct a base map based on the geographic information data and project the monitoring points onto the base map. Use Bezier curves to draw the propagation paths to achieve smooth transitions. Dynamically adjust the arrow size and color according to the propagation intensity. Combine the clustering analysis method to determine the key monitoring areas. Cluster the nodes in the anomaly propagation network according to the location and anomaly degree to identify the anomaly concentration areas, calculate the importance scores of each clustering area, and the importance scores comprehensively consider the anomaly degree, influence range, and propagation speed. Divide the monitoring areas into levels according to the importance scores and determine the monitoring frequencies; Use the dynamic programming method to generate an emergency monitoring parameter sequence, establish an optimization model for monitoring resource allocation with the maximum anomaly monitoring effect as the objective function, use the number of monitoring devices and communication bandwidth as resource constraint conditions, solve the optimal monitoring parameter configuration plan, and perform online adjustment of the monitoring parameters based on the anomaly development trend to obtain an emergency monitoring parameter sequence including the monitoring areas and monitoring frequencies.

[0056] The multi-level anomaly discrimination rule is a rule used to determine anomalies in a multi-level structure. It usually analyzes features at different levels and judges anomalies layer by layer. It is often used in anomaly detection of complex systems and multi-dimensional data analysis. The retrospective analysis is a reverse analysis method that deduces the process and reasons from the results and is widely used in fault diagnosis, problem-solving, and data analysis. The base map refers to the map or image used as the base in a geographic information system or other graphical analysis to assist in displaying and analyzing other data layers. The Bezier curve is a commonly used mathematical curve that determines the shape of the curve through control points and is widely used in computer graphics, animation production, and path planning, etc.

[0057] Construct a causal relationship matrix. Select several monitoring points, such as four monitoring points A, B, C, and D, and collect historical water level data. Set a sliding time window. For each pair of monitoring points, calculate the transfer entropy value of the water level data sequence within each time period. The transfer entropy value is used to measure the influence degree of the water level change at point A on the water level change at point B. The method of calculating the transfer entropy value is to count the probabilities of various combinations of the water level changes of A and B, as well as the conditional probability of the water level change of B when the water level change of A is known, and calculate the transfer entropy value based on the probability values. Fill the calculated transfer entropy value into the corresponding cell in the B column of row A in the causal relationship matrix, indicating the causal relationship strength of A to B.

[0058] Construct an execution parameter matrix. The rows of this matrix correspond to the monitoring points, and the columns correspond to the execution parameters. The execution parameters include the mean deviation, variance ratio, trend slope, and fluctuation period obtained through statistical analysis of historical data. Similarly, calculate parameters such as the variance ratio, trend slope, and fluctuation period. Fill these parameters into the corresponding cells of the execution parameter matrix.

[0059] Establish multi-level anomaly discrimination rules based on the execution parameter matrix, and combine the fuzzy inference mechanism to evaluate the anomaly degree of the real-time data of the monitoring points. For example, if the mean deviation exceeds a certain threshold, it is judged as a mild anomaly; if it exceeds a higher threshold, it is judged as a moderate anomaly, and so on. Combine the anomaly degrees of multiple parameters and use the fuzzy inference mechanism to comprehensively evaluate the overall anomaly degree of the monitoring points. For example, if both the mean deviation and the variance ratio of point A exceed the threshold of mild anomaly, then point A is finally determined to be a mild anomaly. If an anomaly point is detected, such as point A is determined to be abnormal, then use the depth-first search strategy to perform backtracking analysis in the causality matrix to determine the anomaly propagation path. Starting from the anomaly point A, determine the search direction according to the causality strength threshold. Preferentially visit the adjacent nodes that have a strong causal relationship with point A. Maintain an access mark table to record the nodes that have been visited to avoid repeated visits. Through iterative search, obtain the complete propagation link from the anomaly source to the impact end. According to the obtained complete propagation link, analyze the anomaly propagation path, analyze the correlation between the anomaly point and the surrounding monitoring points in the local area, and finally synthesize the analysis results of each level on the global scale to obtain the anomaly propagation network structure.

[0060] Based on the anomaly propagation network structure, draw the anomaly propagation path diagram. Use the adaptive layout algorithm to arrange the monitoring points in space and project them onto the geographic information base map. Use Bezier curves to draw the propagation path, and dynamically adjust the arrow size and color according to the propagation intensity. Cluster the nodes in the anomaly propagation network according to their positions and anomaly degrees to identify the anomaly concentration areas. Calculate the importance score of each clustering area, considering the anomaly degree, influence range, and propagation speed. For example, if the anomaly degree is high, the influence range is wide, and the propagation speed is fast in a certain clustering area, then its importance score is also high. Divide the monitoring area into levels according to the importance score and determine the monitoring frequency. For example, the area with a high importance score is divided into the first-level monitoring area, and the monitoring frequency is set to once every 10 minutes; the area with a low importance score is divided into the second-level monitoring area, and the monitoring frequency is set to once every hour. Establish an optimization model for monitoring resource allocation with the goal of maximizing the anomaly monitoring effect, and use the number of monitoring devices and communication bandwidth as resource constraint conditions. Solve the optimal monitoring parameter configuration scheme and perform online adjustment of the monitoring parameters according to the anomaly development trend, and finally obtain the emergency monitoring parameter sequence including the monitoring area and monitoring frequency. For example, if the anomaly degree of a certain area intensifies, then dynamically increase the monitoring frequency of that area.

[0061] In this embodiment, through a multi-level anomaly discrimination rule and a fuzzy inference mechanism, abnormal events can be identified more sensitively, avoiding missed reports and false alarms, thereby improving the discovery efficiency of abnormal events. Through a causality matrix and a depth-first search strategy, the abnormal propagation link can be accurately traced back, the source of the anomaly can be determined, and the scope of influence of the anomaly can be evaluated, providing a scientific basis for emergency response. By generating an emergency monitoring parameter sequence through a dynamic programming method, the monitoring area and monitoring frequency can be dynamically adjusted according to the development trend of the anomaly, optimizing the allocation of monitoring resources, improving the monitoring efficiency, and saving the monitoring cost.

[0062] Figure 2 FIG. is a schematic structural diagram of a water level distributed monitoring system based on a visual sensor network according to an embodiment of the present invention, as Figure 2 shown, the system includes: A first unit for collecting water level image samples of a monitoring area, constructing a spectral decomposition network to extract the frequency domain features of the water level image samples, constructing a multi-scale spectral feature matrix, constructing an energy matrix based on the multi-scale spectral feature matrix and calculating the energy distribution density, generating an energy mask map corresponding to the water level image samples, guiding an attention mechanism through the energy mask map to perform self-supervised decomposition on the image, obtaining a water level scale feature tensor, a water surface texture feature tensor, and an environmental background feature tensor, inputting the decomposed feature tensors into a multi-modal contrastive learning network, calculating a discriminative feature vector group through maximizing mutual information, constructing a random field model based on the spatial distribution characteristics of visual sensor nodes, calculating the spatial dependence relationship between nodes to obtain a node correlation probability matrix, and inputting the node correlation probability matrix and the discriminative feature vector group into a graph neural network to obtain a parameter matrix of a water level propagation prediction model through message passing iteration; A second unit for receiving real-time water level images, inputting the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculating an optimal decomposition scale sequence and generating a multi-resolution feature representation, obtaining a compressed feature code stream through sparse coding of the multi-resolution feature representation, inputting the feature code stream into a decoupled representation network, separating to obtain a water level feature map and calculating a complementary metric value based on the water level feature map, performing adaptive fusion based on the complementary metric value to obtain an enhanced feature map, inputting the enhanced feature map into a densely connected network, obtaining a water level area segmentation map and a water level measurement value through cross-layer feature reuse, calculating the posterior probability distribution of the water level measurement according to a Bayesian inference framework by combining prior knowledge and observed data, calculating a measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screening to obtain a high-confidence water level data set and performing spectral analysis, and extracting a periodic feature vector and a trend feature vector; A third unit is used to construct a non-linear state space model based on the periodic feature vector and the trend feature vector, calculate a state estimation sequence through particle filtering and construct a causal discovery network, calculate a causal relationship matrix of water level changes based on the causal discovery network and decompose it into multiple sub-matrices, construct a hierarchical attention network at different spatial scales and execute to obtain a water level feature tensor, obtain an encrypted model parameter update amount through a federated optimization objective function combined with a homomorphic encryption algorithm, calculate an optimal sampling strategy vector through a pre-set multi-agent collaborative decision-making system and perform anomaly analysis on the water level to obtain an anomaly feature vector, add the optimal sampling strategy vector and the anomaly feature vector to a hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix, and analyze the anomaly propagation path map and the emergency monitoring parameter sequence in combination with the causal relationship matrix.

[0063] The present invention can be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0064] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A water level distributed monitoring method based on a visual sensor network, characterized in that: include: Collect water level image samples in the monitoring area and construct a spectral decomposition network to extract frequency domain features of the water level image samples, construct a multi-scale spectral feature matrix, construct an energy matrix based on the multi-scale spectral feature matrix and calculate the energy distribution density, generate an energy mask map corresponding to the water level image sample, guide the attention mechanism through the energy mask map to perform self-supervised decomposition of the image, obtain the water level scale feature tensor, the water surface texture feature tensor and the environmental background feature tensor, input the decomposed feature tensor into the multimodal contrast learning network, obtain the discriminative feature vector group by maximizing the mutual information, construct a random field model based on the spatial distribution characteristics of the visual sensor nodes, calculate the spatial dependency between the nodes to obtain the node association probability matrix, input the node association probability matrix and the discriminative feature vector group into the graph neural network, and obtain the parameter matrix of the water level propagation prediction model through message passing iteration; Receive a real-time water level image, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream by sparse coding the multi-resolution feature representation, input the feature code stream into a decoupled representation network, obtain a water level feature map by separation, calculate a complementary metric value based on the water level feature map, perform adaptive fusion based on the complementary metric value to obtain an enhanced feature map, input the enhanced feature map into a densely connected network, obtain a water level region segmentation map and a water level measurement value by cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement based on a Bayesian reasoning framework, calculate the measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screen a high-confidence water level data set and perform spectral analysis to extract periodic feature vectors and trend feature vectors; A nonlinear state space model is constructed based on the periodic eigenvector and the trend eigenvector, and a state estimation sequence is obtained by combining with particle filtering to construct a causal discovery network. A causal relationship matrix of water level changes is calculated based on the causal discovery network and decomposed to obtain multiple sub-matrices. A hierarchical attention network is constructed at different spatial scales and executed to obtain a water level feature tensor. The encrypted model parameter update amount is obtained by combining the federated optimization objective function with the homomorphic encryption algorithm. The optimal sampling strategy vector is calculated by a pre-set multi-agent collaborative decision-making system, and an abnormal analysis is performed on the water level to obtain an abnormal feature vector. The optimal sampling strategy vector and the abnormal feature vector are added to the hierarchical reinforcement learning network to obtain a detection resource allocation plan and an execution parameter matrix. The abnormal propagation path diagram and an emergency monitoring parameter sequence are obtained by combining with the causal relationship matrix analysis.

2. The method according to claim 1, characterized in that Collect water level image samples in the monitoring area and construct a spectral decomposition network to extract the frequency domain features of the water level image samples, construct a multi-scale spectrum feature matrix, construct an energy matrix based on the multi-scale spectrum feature matrix and calculate the energy distribution density, generate an energy mask map corresponding to the water level image sample, guide the attention mechanism through the energy mask map to perform self-supervised decomposition of the image, obtain the water level scale feature tensor, the water surface texture feature tensor and the environmental background feature tensor, input the decomposed feature tensor into the multimodal contrast learning network, obtain the discriminative feature vector group by maximizing the mutual information, construct a random field model based on the spatial distribution characteristics of the visual sensor nodes, calculate the spatial dependency between the nodes to obtain the node association probability matrix, input the node association probability matrix and the discriminative feature vector group into the graph neural network, and obtain the parameter matrix of the water level propagation prediction model through message passing iteration, including: Collect water level image samples in the monitoring area, and input the water level image samples into a multi-level spectral decomposition network, wherein the shallow network in the multi-level spectral decomposition network extracts the high-frequency components of the water level image samples through a high-pass filter group to obtain water level scale line features and water surface ripple features, the middle-level network extracts the intermediate-frequency components of the water level image samples through a band-pass filter group to obtain water level line contour features and shore structure features, and the deep network extracts the low-frequency components of the water level image samples through a low-pass filter group to obtain water area distribution features, and normalizes the feature maps corresponding to the high-frequency components, the intermediate-frequency components, and the low-frequency components, and arranges and combines them in frequency band order into a multi-scale spectrum feature matrix; Calculating the energy value of each frequency band component in the multi-scale spectral feature matrix to obtain a two-dimensional energy matrix, performing density estimation on the two-dimensional energy matrix through a kernel function to obtain an energy density distribution map, setting an adaptive threshold based on the energy density distribution map and generating a binary energy mask map, determining high-energy areas and low-energy areas, marking the high-energy areas as foreground areas, and marking the low-energy areas as background areas; The water level image sample is divided into a plurality of overlapping image blocks, and the attention weights of the plurality of overlapping image blocks are calculated based on the binary energy mask map. The overlapping image blocks with the attention weights are input into the self-supervised deconstruction module for feature decomposition. The water level scale feature branch is extracted by a fine-grained feature extractor to obtain a water level scale feature tensor. The water surface texture feature branch is extracted by a texture analyzer to obtain a water surface texture feature tensor. The environment background feature branch is extracted by a context encoder to obtain an environment background feature tensor. The water level scale feature tensor, the water surface texture feature tensor, and the environment background feature tensor are respectively input into a feature encoder for dimensionality reduction to obtain a feature vector of a unified dimension, and a feature comparison pool is constructed. Randomly sampling from the feature vector of the unified dimension to obtain positive sample pairs and negative sample pairs, calculating the mutual information metric between the positive sample pairs and the negative sample pairs, obtaining a discriminative feature vector group by maximizing the mutual information metric of the positive sample pairs and minimizing the mutual information metric of the negative sample pairs, calculating the distance matrix between the node pairs based on the geographic coordinates of the visual sensor nodes, converting the distance matrix into initial association strength, and calculating the interaction between the node pairs through the potential function of the random field to obtain a node association probability matrix; The node association probability matrix and the discriminative feature vector group are input into the graph neural network. In each round of message passing iteration, the importance weights of the neighbor nodes are determined based on the node association probability matrix. The feature information of the neighbor nodes is nonlinearly transformed to obtain fused features. The fused features are combined with the feature information of the node itself to update the node state. The parameter matrix of the water level propagation prediction model is obtained through multiple rounds of iterations.

3. The method according to claim 2, characterized in that Randomly sampling from the feature vector of the unified dimension to obtain positive sample pairs and negative sample pairs, calculating the mutual information metric between the positive sample pairs and the negative sample pairs, obtaining a discriminative feature vector group by maximizing the mutual information metric of the positive sample pairs and minimizing the mutual information metric of the negative sample pairs, calculating the distance matrix between the node pairs based on the geographic coordinates of the visual sensor nodes, converting the distance matrix into an initial association strength, and calculating the interaction between the node pairs through the potential function of the random field to obtain a node association probability matrix, including: A feature comparison pool is constructed from feature vectors of uniform dimensions and the feature comparison pool is divided into multiple time windows. Scene feature information is extracted in each time window and randomly sampled to obtain positive sample pairs. In different monitoring scenarios, a multi-layered sampling strategy is used according to weather conditions, monitoring time periods, and water level change trends to randomly sample feature vectors to construct negative sample pairs. Adding feature vectors in the positive sample pair and the negative sample pair to a multilayer perceptron including an input layer, a plurality of hidden layers and an output layer for feature transformation, setting a batch normalization layer and a nonlinear activation function between adjacent layers of the multilayer perceptron, mapping the feature vectors from a feature space to a probability distribution space, constructing a Gaussian kernel function group with an adaptive bandwidth in the probability distribution space, and obtaining a continuous probability distribution by accumulating local and global contributions of all feature points to the probability density; Calculating the mutual information metric between the positive sample pair and the negative sample pair, wherein the mutual information metric is obtained by calculating the joint distribution of the transformed features, the multiple differences of the conditional distribution and the marginal distribution, constructing a mutual information optimization objective function including feature discriminability constraints and distribution consistency constraints, and maximizing the mutual information metric of the positive sample pair and minimizing the mutual information metric of the negative sample pair, combining the gradient descent method with an adaptive learning rate, performing multiple rounds of iterative optimization to obtain a feature vector group with local and global discriminability; Based on the geographic coordinates of the visual sensor nodes, the distance matrix between node pairs is calculated by fusing the geodetic distance between nodes, the terrain undulation coefficient, the river connectivity coefficient, and the hydrological propagation characteristic coefficient. The multi-scale elevation difference between nodes is calculated in combination with the digital elevation model data to dynamically correct the distance matrix. At the same time, the actual flow distance considering the flow direction and flow velocity is calculated based on the river connectivity data. The actual flow distance is fused with the corrected distance matrix for multiple features to obtain a corrected distance matrix. Based on the modified distance matrix, the distance value is mapped to the dynamic interval through an adaptive normalization method, the local and global densities are calculated based on the multi-dimensional spatial distribution density characteristics of the nodes, an adaptive attenuation coefficient based on the density gradient is set for the densely distributed area, and a dynamic attenuation coefficient considering spatial heterogeneity is set for the sparsely distributed area, and the initial association strength under multi-scales is calculated in combination with the spatiotemporal association characteristics of the nodes; Based on the initial association strength, a random field potential function with an adaptive segmentation structure is constructed, and the node distance is dynamically divided into a short-distance interval, a medium-distance interval and a long-distance interval with a transition zone based on the multi-dimensional spatial relationship between the nodes. A linear attenuation function considering local spatial dependence is designed in the short-distance interval, an exponential attenuation function considering regional influence is designed in the medium-distance interval, and a residual function considering global association is designed in the long-distance interval, and the adaptive segmentation points of the potential function are determined through spatial autocorrelation analysis; Based on the adaptive segmentation points and the multi-layer message passing mechanism, the interaction between node pairs is iteratively calculated through the potential function of the random field. In each round of iteration, the state information of multi-scale neighborhood nodes is collected based on the spatial dependency relationship. The messages are weighted at multiple levels using the adaptive potential function value. The spatiotemporal redundant information is identified and suppressed through the structured attention mechanism. The weighted messages are subjected to nonlinear feature combination and random noise that obeys the dynamic distribution is added. The node association probability matrix that characterizes the association relationship between nodes is obtained through state updating.

4. The method according to claim 1, characterized in that: Receive a real-time water level image, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream by sparse coding the multi-resolution feature representation, input the feature code stream into a decoupled representation network, obtain a water level feature map by separation, calculate a complementary metric value based on the water level feature map, perform adaptive fusion based on the complementary metric value to obtain an enhanced feature map, input the enhanced feature map into a densely connected network, obtain a water level region segmentation map and a water level measurement value by cross-layer feature reuse, calculate the posterior probability distribution of water level measurement based on a Bayesian reasoning framework, calculate the measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screen a high-confidence water level data set and perform spectral analysis, extract periodic feature vectors and trend feature vectors including: Receive a real-time water level image, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate the gradient of the image in multiple directions to construct a directional gradient histogram, calculate the grayscale co-occurrence matrix at different directions and distance combinations to extract multidimensional texture features including energy, entropy, contrast and correlation, select the optimal wavelet basis function through adaptive weight fusion according to the multidimensional texture features, analyze the local extreme value characteristics and energy distribution of the wavelet coefficients to calculate and obtain the optimal decomposition scale sequence, perform multi-layer decomposition based on the optimal decomposition scale sequence through row-column transformation and downsampling operations to obtain low-frequency approximate components and high-frequency detail components, and perform tree decomposition on the high-frequency detail components to generate multi-resolution feature representation; The multi-resolution feature representation is sparsely encoded by constructing an overcomplete dictionary containing redundant atoms, the features are divided into blocks according to a fixed size, sparse decomposition based on orthogonal matching pursuit is performed on each feature block, the dictionary atoms most relevant to the residual are iteratively selected and the reconstruction coefficients are updated until the sparsity constraint is satisfied to obtain a compressed feature code stream, the compressed feature code stream is input into a decoupled representation network, and multi-layer features are extracted through a series of residual calculation units, wherein the residual calculation unit includes a convolution layer, a normalization layer and an activation function, and context information of different scales is extracted through a multi-scale pooling module, and the feature resolution is gradually restored in combination with a deconvolution operation, and the channel weights are calculated based on feature statistics and nonlinear transformations for separation to obtain a water level feature map; Based on the water level feature map, a Gaussian weighted window is used to calculate the local area structural similarity, the horizontal and vertical gradients are calculated to construct the gradient direction field and analyze the direction consistency, the features are Fourier transformed to calculate the normalized cross-power spectrum, the best matching position is determined by local maximum detection to obtain the complementary measurement value, a feature fusion relationship graph is constructed based on the complementary measurement value, the nodes in the feature fusion relationship graph represent the features and the weights of the edges are determined by the complementary measurement value, the feature correlation score is calculated and normalized by the multi-layer graph attention calculation unit, and the normalized attention score is adaptively fused with the feature weighted combination to obtain an enhanced feature map; The enhanced feature map is input into a densely connected network, and the standard convolution is decomposed into depth convolution and point-by-point convolution through deep separation convolution to reduce the amount of calculation. Dense connection is used to make the nodes in the latter layer receive the features of all previous layers at the same time to implement layer feature reuse, and the water level area segmentation map and water level measurement value are obtained by adaptively learning the feature channel importance weights through a compression excitation mechanism; According to the Bayesian reasoning framework, in a three-layer probability model including an observation layer, a latent variable layer, and a priori layer, observation data and physical constraints are input as prior knowledge, and the posterior probability distribution of water level measurement is calculated by iteratively optimizing the approximate posterior distribution parameters through a variational reasoning method. Based on the posterior probability distribution, in the network forward calculation, some neurons are randomly inactivated and multiple samplings are performed to calculate the predicted mean and variance to estimate the measurement uncertainty value and generate a reliability score table; The screened high-confidence water level data set is subjected to spectral analysis. Time-frequency analysis is achieved through continuous wavelet transform. The spectral decomposition method is used to decompose and reconstruct the time series. An adaptive signal decomposition algorithm based on extreme point envelope is adopted to calculate multiple inherent characteristic functions. The periodic characteristic vector and trend characteristic vector are extracted by controlling the decomposition process through amplitude threshold.

5. The method according to claim 4, characterized in that The enhanced feature map is input into a densely connected network, and the standard convolution is decomposed into deep convolution and point-by-point convolution through deep separation convolution to reduce the amount of calculation. Dense connection is used to make the back-layer nodes receive the features of all previous layers at the same time to implement layer feature reuse. The water level area segmentation map and water level measurement values ​​are obtained by adaptively learning the feature channel importance weights through the compression excitation mechanism, including: The enhanced feature map is subjected to feature extraction by using deep separation convolution, and the standard convolution is decomposed into deep convolution and point-by-point convolution. The deep convolution extracts spatial correlation features by performing convolution operation on each input channel independently, and the point-by-point convolution obtains recombined features by combining channel information on the spatial correlation features. The reorganized features are sequentially input into a first dense connection block, a second dense connection block, and a third dense connection block, wherein a plurality of depth separation convolution layers are arranged in each of the dense connection blocks, wherein the first layer of the first dense connection block performs feature extraction on the reorganized features to obtain first layer output features, the second layer performs feature extraction after cascading the reorganized features and the first layer output features to obtain second layer output features, and the third layer performs feature extraction after cascading the reorganized features, the first layer output features, and the second layer output features to obtain third layer output features; The features between adjacent densely connected blocks are subjected to convolution operations and feature sampling operations to reduce the spatial size of the feature map and increase the number of channels to obtain multi-scale features, the feature map output by the densely connected block is subjected to global feature statistics to obtain a channel description vector, the channel description vector is input into a first fully connected layer to obtain a compressed feature vector, the compressed feature vector is input into a second fully connected layer to obtain an excitation feature vector, the excitation feature vector is normalized to obtain a channel weight coefficient, and the channel weight coefficient is multiplied by the multi-scale feature to obtain a weighted feature; The weighted features output by the first densely connected block, the second densely connected block, and the third densely connected block are adjusted to the same spatial resolution, the weighted features of the third densely connected block are upsampled and fused with the weighted features of the second densely connected block to obtain a first fused feature, the first fused feature is upsampled and fused with the weighted features of the first densely connected block to obtain a fused feature map, the fused feature map is respectively input into the segmentation branch and the measurement branch, a water level area segmentation map with the same size as the enhanced feature map is output through the multi-layer convolution operation of the segmentation branch, and the water level measurement value is output through the feature pooling operation and full connection operation of the measurement branch.

6. The method according to claim 1, characterized in that A nonlinear state space model is constructed based on the periodic feature vector and the trend feature vector, a state estimation sequence is obtained by combining with particle filtering and a causal discovery network is constructed, a causal relationship matrix of water level change is calculated based on the causal discovery network and decomposed to obtain multiple sub-matrices, a hierarchical attention network is constructed at different spatial scales and executed to obtain a water level feature tensor, an encrypted model parameter update is obtained by combining a federated optimization objective function with a homomorphic encryption algorithm, an optimal sampling strategy vector is calculated by a pre-set multi-agent collaborative decision-making system, and an abnormal analysis is performed on the water level to obtain an abnormal feature vector, the optimal sampling strategy vector and the abnormal feature vector are added to a hierarchical reinforcement learning network to obtain a detection resource allocation plan and an execution parameter matrix, and an abnormal propagation path diagram and an emergency monitoring parameter sequence are obtained by combining with the causal relationship matrix analysis, including: Continuously collecting water level data at each monitoring site according to a preset sampling period, performing wavelet decomposition on the water level data to obtain a period component and a trend component, constructing the period component into a period feature vector, and constructing the trend component into a trend feature vector; The periodic eigenvector is used as an observation variable, and the trend eigenvector is used as a state variable to construct a nonlinear state space model, a preset number of particle samples are generated for each state variable, the particle samples are iteratively calculated through importance sampling in combination with particle filtering, the likelihood probability value of each particle sample is calculated based on the observation data as a particle weight, and the particle samples are resampled to obtain a state estimation sequence; Building a causal discovery network based on the state estimation sequence, calculating the transfer entropy value between nodes as the causal relationship strength to obtain a causal relationship matrix, dividing the causal relationship matrix into multiple regions according to the spatial positions of the monitoring points and decomposing it into multiple sub-matrices; A hierarchical attention network is constructed at different spatial scales. The first layer of attention network inputs the submatrix representing the causal relationship within the region, the second layer of attention network inputs the submatrix representing the causal relationship between adjacent regions, and the third layer of attention network inputs the submatrix representing the causal relationship between remote regions. The features at different spatial scales are weightedly fused through attention weights to obtain the water level feature tensor. Design a federated optimization objective function, take each monitoring site as a participant in federated learning, use a homomorphic encryption algorithm to encrypt model parameters, each participant uses local sampled data to calculate the model parameter gradient and encrypts it, the server aggregates the encrypted model parameter gradient to obtain the encrypted model parameter update, and obtains the global model parameters after a preset round of iterative optimization; The optimal sampling strategy is calculated by a pre-set multi-agent collaborative decision-making system, wherein the multi-agent collaborative decision-making system includes multiple agents, each of which corresponds to a plurality of monitoring areas. The agents exchange the water level feature tensor information through a communication network, calculate the state value function based on the water level feature tensor, and use a collaborative decision-making algorithm to obtain an optimal sampling strategy vector, perform an abnormal analysis on the collected water level data, and extract time series features and spatial distribution features to obtain an abnormal feature vector; The optimal sampling strategy vector and the abnormal feature vector are added to a hierarchical reinforcement learning network, wherein the state evaluation subnetwork in the hierarchical reinforcement learning network outputs a state vector, the action generation subnetwork outputs a detection resource allocation scheme matrix, the value estimation subnetwork predicts the long-term benefits of the detection resource allocation scheme to obtain a benefit vector, and the execution parameter matrix is ​​obtained through strategy optimization; The abnormal propagation path is analyzed based on the execution parameter matrix and the causal relationship matrix, the abnormal event position is located based on the execution parameter matrix, the abnormal propagation link is traced back through the causal relationship matrix, the abnormal impact range is analyzed and the abnormal propagation path map is drawn, the key monitoring areas are determined and the sampling frequency is increased, and an emergency monitoring parameter sequence including the monitoring areas and monitoring frequencies is generated.

7. The method according to claim 6, characterized in that Analyzing the abnormal propagation path based on the execution parameter matrix and the causal relationship matrix, locating the abnormal event position based on the execution parameter matrix, tracing back the abnormal propagation link through the causal relationship matrix, analyzing the abnormal impact range and drawing the abnormal propagation path map, determining the key monitoring area and increasing the sampling frequency, and generating the emergency monitoring parameter sequence including the monitoring area and the monitoring frequency include: Based on the execution parameter matrix and the causal relationship matrix, the water level data is processed in segments through a sliding time window, and the transfer entropy value of the water level data sequence between each pair of monitoring points is calculated as the causal relationship strength, wherein the transfer entropy value is obtained by calculating and accumulating the conditional probability, and the transfer entropy value is filled into the corresponding position of the causal relationship matrix; Construct an execution parameter matrix, wherein the row vectors of the execution parameter matrix correspond to monitoring points, and the column vectors correspond to execution parameters. The execution parameters include mean deviation, variance ratio, trend slope, and fluctuation period obtained based on statistical analysis of historical data. Multi-level abnormality discrimination rules are established based on the execution parameters, and the abnormality degree of the real-time data of the monitoring points is evaluated in combination with the fuzzy reasoning mechanism. A depth-first search strategy is used to perform backtracking analysis in the causal relationship matrix. Starting from the abnormal point, the search direction is determined based on the causal relationship strength threshold. The adjacent nodes with the strongest causal relationship are visited first, and the access mark table is maintained to record the visited nodes. The complete propagation link from the abnormal source to the affected end is obtained through iterative search. Based on the complete propagation link, the abnormal propagation path is analyzed through a hierarchical processing strategy. The correlation between the abnormal point and the surrounding monitoring points is analyzed in the local area to obtain the local propagation characteristics. The analysis scope is expanded to the adjacent area to study the propagation mode of a larger spatial scale. The abnormal propagation network structure is obtained by integrating the analysis results of each level on a global scale. According to the abnormal propagation network structure, the monitoring points and propagation paths are spatially arranged through an adaptive layout algorithm, a base map is constructed based on geographic information data and the monitoring points are projected onto the base map, a Bezier curve is used to draw the propagation path to achieve a smooth transition, the arrow size and color are dynamically adjusted according to the propagation intensity, and the key monitoring area is determined in combination with the cluster analysis method. The nodes in the abnormal propagation network are clustered according to the position and abnormality degree to identify the abnormal concentration area, and the importance score of each cluster area is calculated. The importance score comprehensively considers the abnormality degree, the scope of influence, and the propagation speed. According to the importance score, the monitoring area is graded and the monitoring frequency is determined; The dynamic programming method is used to generate the emergency monitoring parameter sequence, and a monitoring resource allocation optimization model with the maximization of abnormal monitoring effect as the objective function is established. The number of monitoring equipment and communication bandwidth are used as resource constraints to solve the optimal monitoring parameter configuration scheme. The monitoring parameters are adjusted online based on the abnormal development trend to obtain the emergency monitoring parameter sequence including the monitoring area and monitoring frequency.

8. A water level distributed monitoring system based on a visual sensor network, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect water level image samples in the monitoring area and construct a spectrum decomposition network to extract frequency domain features of the water level image samples, construct a multi-scale spectrum feature matrix, construct an energy matrix based on the multi-scale spectrum feature matrix and calculate the energy distribution density, generate an energy mask map corresponding to the water level image sample, guide the attention mechanism to perform self-supervisory decomposition on the image through the energy mask map, obtain the water level scale feature tensor, the water surface texture feature tensor and the environmental background feature tensor, input the decomposed feature tensor into the multimodal contrast learning network, obtain the discriminative feature vector group by maximizing the mutual information, construct a random field model based on the spatial distribution characteristics of the visual sensor nodes, calculate the spatial dependency between the nodes to obtain the node association probability matrix, input the node association probability matrix and the discriminative feature vector group into the graph neural network, and obtain the parameter matrix of the water level propagation prediction model through message passing iteration; The second unit is used to receive a real-time water level image, input the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculate the optimal decomposition scale sequence and generate a multi-resolution feature representation, obtain a compressed feature code stream by sparsely encoding the multi-resolution feature representation, input the feature code stream into a decoupled representation network, obtain a water level feature map by separation, calculate a complementary metric value based on the water level feature map, perform adaptive fusion based on the complementary metric value to obtain an enhanced feature map, input the enhanced feature map into a densely connected network, obtain a water level region segmentation map and a water level measurement value by cross-layer feature reuse, calculate the posterior probability distribution of the water level measurement based on the Bayesian reasoning framework, calculate the measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screen a high-confidence water level data set, perform spectral analysis, and extract periodic feature vectors and trend feature vectors; The third unit is used to construct a nonlinear state space model based on the periodic eigenvector and the trend eigenvector, calculate a state estimation sequence in combination with a particle filter and construct a causal discovery network, calculate a causal relationship matrix of water level changes based on the causal discovery network and decompose it into multiple sub-matrices, construct a hierarchical attention network at different spatial scales and execute it to obtain a water level feature tensor, obtain an encryption model parameter update amount through a federated optimization objective function combined with a homomorphic encryption algorithm, calculate an optimal sampling strategy vector through a pre-set multi-agent collaborative decision-making system and perform an abnormal analysis on the water level to obtain an abnormal feature vector, add the optimal sampling strategy vector and the abnormal feature vector to a hierarchical reinforcement learning network, obtain a detection resource allocation plan and an execution parameter matrix, and obtain an abnormal propagation path diagram and an emergency monitoring parameter sequence in combination with the causal relationship matrix analysis.

Images