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

Abstract

The application provides a water level distributed monitoring method and system based on a visual sensing network, relates to the technical field of water level monitoring, and comprises the following steps: a water level propagation prediction model is constructed through multimodal contrast learning and a graph neural network; real-time water level images are processed by combining adaptive wavelet transform, decoupling representation and a dense connection network to obtain water level measurement values and reliability scores; and abnormal analysis and emergency monitoring are performed based on state estimation, causal discovery and hierarchical reinforcement learning, so that high-precision and high-reliability distributed water level monitoring can be realized, abnormal water level changes can be effectively identified, and timely emergency monitoring parameters can be provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water level monitoring, and in particular to a water level distributed monitoring method and system based on a visual sensing network. BACKGROUND

[0002] With the continuous development of water conservancy projects and flood control and disaster reduction work, water level monitoring, as an important basic work, has attracted widespread attention. Traditional water level monitoring mainly relies on manual observation or automatic measurement equipment, but there are many limitations in monitoring efficiency, accuracy and coverage range;

[0003] In recent years, water level monitoring technology based on visual sensing networks has gradually emerged. By deploying camera equipment to collect water level images, combining computer vision and deep learning algorithms for intelligent analysis, automatic monitoring of water levels has been achieved. Existing visual sensing network water level monitoring systems usually use image processing and machine learning methods to estimate water level values by extracting water level scale features and water surface texture features, and establish prediction models to analyze water level trends;

[0004] However, the existing technology still has problems such as insufficient monitoring accuracy in complex environments, not fully considering the spatial dependence between visual sensing nodes, and lacking research on causal relationship analysis and abnormal propagation mechanism of water level changes;

[0005] Therefore, there is an urgent need for a solution to solve the problems in the prior art. SUMMARY

[0006] The embodiments of the present application provide a water level distributed monitoring method and system based on a visual sensing network, which can at least solve some of the problems in the prior art.

[0007] In a first aspect, the present application provides a water level distributed monitoring method based on a visual sensing network, comprising:

[0008] Collecting water level image samples of a monitoring area and constructing a spectral decomposition network to extract 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 an energy distribution density, generating an energy mask image corresponding to the water level image samples, guiding the attention mechanism to perform self-supervised decomposition on the image through the energy mask image, obtaining water level scale feature tensors, water surface texture feature tensors and environmental background feature tensors, inputting the decomposed feature tensors into a multi-modal contrast learning network, calculating a discriminative feature vector group through mutual information maximization, constructing a random field model based on the spatial distribution characteristics of the visual sensing nodes, calculating the spatial dependence between the nodes to obtain a node correlation probability matrix, inputting the node correlation 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.

[0009] receiving real-time water level images, inputting a 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 by sparse coding the multi-resolution feature representation, inputting the feature code stream into a decoupling representation network, obtaining a water level feature map by separation and calculating a complementarity measure value based on the water level feature map, adaptively fusing based on the complementarity measure value to obtain an enhanced feature map, inputting the enhanced feature map into a dense connection network, obtaining a water level region segmentation map and a water level measurement value through cross-layer feature reuse, calculating a posterior probability distribution of the water level measurement according to a Bayesian inference framework combined with prior knowledge and observation data, calculating a measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screening a high-confidence water level dataset and performing spectral analysis to extract a periodic feature vector and a trend feature vector;

[0010] constructing a nonlinear state space model based on the periodic feature vector and the trend feature vector, combining particle filtering to calculate a state estimation sequence and construct a causal discovery network, calculating a causal relationship matrix of water level changes based on the causal discovery network and decomposing to obtain multiple sub-matrices, constructing a hierarchical attention network on different spatial scales and executing to obtain a water level feature tensor, combining homomorphic encryption algorithm through a federal optimization objective function to obtain an encrypted model parameter update, calculating an optimal sampling strategy vector through a pre-set multi-agent collaborative decision system and performing abnormal analysis on the water level to obtain an abnormal feature vector, adding the optimal sampling strategy vector and the abnormal feature vector to a hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix, analyzing to obtain an abnormal propagation path graph and an emergency monitoring parameter sequence combined with the causal relationship matrix.

[0011] In an optional implementation, water level image samples of a monitoring area are collected and a wave spectrum decomposition network is constructed to extract 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 an energy distribution density is calculated, an energy mask graph corresponding to the water level image samples is generated, and a self-supervised decomposition of the image is performed through the energy mask graph to guide an attention mechanism 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 contrast learning network, and a discriminative feature vector group is calculated through mutual information maximization. A random field model is constructed based on the spatial distribution characteristics of the visual sensing nodes, the spatial dependence relationship between the 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. Through message passing iteration, a parameter matrix of the water level propagation prediction model is obtained, including:

[0012] Collecting a water level image sample of a monitoring area, inputting the water level image sample into a multi-level spectral decomposition network, wherein a shallow network in the multi-level spectral decomposition network extracts high-frequency components of the water level image sample through a high-pass filter set to obtain water level scale line features and water surface ripple features, a middle network extracts middle-frequency components of the water level image sample through a band-pass filter set to obtain water level line contour features and shore structure features, and a deep network extracts low-frequency components of the water level image sample through a low-pass filter set to obtain water area distribution features, performing normalization processing on feature maps corresponding to the high-frequency components, the middle-frequency components and the low-frequency components, and arranging and combining the feature maps in a frequency band order to form a multi-scale spectral feature matrix;

[0013] Calculating energy values 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 a high-energy region and a low-energy region, marking the high-energy region as a foreground region, and marking the low-energy region as a background region;

[0014] Dividing the water level image sample into a plurality of overlapping image blocks, calculating attention weights of the plurality of overlapping image blocks based on the binary energy mask map, inputting the overlapping image blocks with the attention weights into a self-supervised deconstruction module for feature decomposition, extracting a water level scale feature branch through a fine-grained feature extractor to obtain a water level scale feature tensor, extracting a water surface texture feature branch through a texture analyzer to obtain a water surface texture feature tensor, and extracting an environmental background feature branch through a context encoder to obtain an environmental background feature tensor, inputting the water level scale feature tensor, the water surface texture feature tensor and the environmental background feature tensor into a feature encoder respectively for dimension reduction to obtain feature vectors of a unified dimension, and constructing a feature contrast pool;

[0015] Randomly sampling positive sample pairs and negative sample pairs from the feature vectors of the unified dimension, calculating mutual information measurement values between the positive sample pairs and the negative sample pairs, obtaining a discriminative feature vector group by maximizing the mutual information measurement values of the positive sample pairs and minimizing the mutual information measurement values of the negative sample pairs, calculating a distance matrix between node pairs based on geographical coordinates of visual sensing nodes, converting the distance matrix into an initial correlation strength, and obtaining a node correlation probability matrix by calculating the interaction between node pairs through a potential function of a random field;

[0016] input the node association probability matrix and the discriminative feature vector group into a graph neural network, determine the importance weight of a neighbor node based on the node association probability matrix in each round of message passing iteration, perform nonlinear transformation on feature information of the neighbor node to obtain fused features, combine the fused features with feature information of the node itself to update the node state, and obtain a parameter matrix of the water level propagation prediction model through multiple iterations.

[0017] In an optional implementation, positive sample pairs and negative sample pairs are randomly sampled from the uniform dimension feature vectors, mutual information metric values between the positive sample pairs and the negative sample pairs are calculated, a discriminative feature vector group is obtained by maximizing the mutual information metric values of the positive sample pairs and minimizing the mutual information metric values of the negative sample pairs, a distance matrix between node pairs is calculated based on geographic coordinates of the visual sensing nodes, the distance matrix is converted into initial association strength, and a node association probability matrix is obtained by calculating the interaction between node pairs through a potential function of a random field, including:

[0018] A feature contrast pool is constructed from the uniform dimension feature vectors, and the feature contrast pool is divided into multiple time windows, scene feature information is extracted in each time window, and positive sample pairs are randomly sampled, and negative sample pairs are constructed by randomly sampling feature vectors according to a multi-layer hierarchical sampling strategy according to weather conditions, monitoring time periods, and water level change trends in different monitoring scenes.

[0019] 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, batch normalization layers and nonlinear activation functions are arranged between adjacent layers of the multi-layer perceptron, a group of Gaussian kernel functions with adaptive bandwidths are constructed in a probability distribution space after the feature vectors are mapped from a feature space to the probability distribution space, and a continuous probability distribution is obtained by accumulating local and global contributions of probability densities of all feature points.

[0020] Mutual information metric values between the positive sample pairs and the negative sample pairs are calculated, wherein the mutual information metric values are obtained by calculating multiple differences of joint distribution, conditional distribution, and marginal distribution of transformed features, a mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints is constructed, a feature vector group with local and global discriminativeness is obtained by maximizing the mutual information metric values of the positive sample pairs and minimizing the mutual information metric values of the negative sample pairs through multiple iteration optimization combined with a gradient descent method with an adaptive learning rate.

[0021] Based on the geographic coordinates of the visual sensing nodes, a distance matrix between node pairs is calculated by fusing geodetic survey distances between nodes, terrain relief coefficients, river connectivity coefficients, and hydrological propagation characteristic coefficients, and the distance matrix is dynamically corrected in combination with digital elevation model data to calculate multi-scale elevation differences between nodes, and actual flow distances considering water flow direction and flow rate are calculated based on river connectivity data, and the actual flow distances are combined with the corrected distance matrix in multiple features to obtain a corrected distance matrix;

[0022] Based on the corrected distance matrix, distance values are mapped to a dynamic interval by an adaptive normalization method, local and global densities are calculated based on multi-dimensional spatial distribution density characteristics of nodes, an adaptive attenuation coefficient based on density gradient is set for densely distributed areas, and a dynamic attenuation coefficient considering spatial heterogeneity is set for sparsely distributed areas, and initial correlation strength under multi-scale is calculated in combination with spatio-temporal correlation characteristics of nodes;

[0023] Based on the initial correlation strength, a random field potential function with an adaptive segmentation structure is constructed, node distances are dynamically divided into near distance intervals, medium distance intervals and far distance intervals with transition zones based on multi-dimensional spatial relationships between nodes, a linear attenuation function considering local spatial dependence is designed in the near distance interval, an exponential attenuation function considering regional influence is designed in the medium distance interval, and a residual function considering global correlation is designed in the far distance interval, and adaptive segmentation points of the potential function are determined through spatial autocorrelation analysis;

[0024] 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 spatial dependence relationship, the messages are weighted in multiple levels using adaptive potential function values, spatio-temporal redundant information is identified and suppressed through structured attention mechanism, the weighted messages are combined in non-linear features and added with random noise subject to dynamic distribution, and the node correlation probability matrix representing the correlation relationship between nodes is obtained through state update.

[0025] In an alternative embodiment, a real-time water level image is received, a parameter matrix of a 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 sparse coding of the multi-resolution feature representation, the feature code stream is input into a decoupled representation network, a water level feature map is obtained by separation and a complementarity measure value is calculated based on the water level feature map, an enhanced feature map is obtained by adaptive fusion based on the complementarity measure 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 by cross-layer feature reuse, a posterior probability distribution of the water level measurement is calculated based on a Bayesian inference framework combined with prior knowledge and observation data, a measurement uncertainty value is calculated based on the posterior probability distribution to generate a reliability score table, a high-confidence water level dataset is screened and spectral analysis is performed to extract a periodic feature vector and a trend feature vector, including:

[0026] A real-time water level image is received, a parameter matrix of a water level propagation prediction model is input into an adaptive wavelet transform network, image gradients in multiple directions are calculated to construct a direction gradient histogram, a gray level co-occurrence matrix is calculated on different direction and distance combinations to extract multi-dimensional texture features including energy, entropy, contrast and correlation, an optimal wavelet basis function is selected by adaptive weight fusion based on the multi-dimensional texture features, an optimal decomposition scale sequence is calculated by analyzing the local extreme value characteristics and energy distribution of wavelet coefficients, low-frequency approximation components and high-frequency detail components are obtained by multi-layer decomposition based on the optimal decomposition scale sequence through row-column transformation and downsampling operations, the multi-resolution feature representation is generated by tree decomposition of the high-frequency detail components;

[0027] The multi-resolution feature representation is sparse coded by constructing an overcomplete dictionary containing redundant atoms, the features are blocked according to a fixed size, sparse decomposition based on orthogonal matching pursuit is performed on each feature block, the dictionary atom most relevant to the residual is iteratively selected and the reconstruction coefficient is updated until the sparsity constraint is met to obtain a compressed feature code stream, the compressed feature code stream is input into a decoupled representation network, multi-layer features are extracted by a series of residual calculation units, wherein the residual calculation unit includes a convolution layer, a normalization layer and an activation function, different scale context information is extracted by a multi-scale pooling module, the feature resolution is gradually restored by inverse convolution operation, channel weights are calculated based on feature statistics and nonlinear transformation to separate a water level feature map;

[0028] The local region structure similarity is calculated using a Gaussian weighted window based on the water level feature map, horizontal and vertical gradients are calculated to construct a gradient direction field and analyze the direction consistency, Fourier transform is performed on the features to calculate a normalized cross-power spectrum, a best matching position is determined by local maximum value detection to obtain a complementarity measure value, a feature fusion relationship graph is constructed based on the complementarity measure value, nodes in the feature fusion relationship graph represent features and weights of edges are determined by the complementarity measure value, feature correlation scores are calculated by a multi-layer graph attention calculation unit and normalized, and enhanced feature maps are obtained by adaptively fusing the normalized attention scores and feature weighting combinations;

[0029] The enhanced feature maps are input into a dense connection network, a standard convolution is decomposed into a depth convolution and a point-wise convolution to reduce the calculation amount by a depthwise separable convolution, a dense connection is used to make nodes in the later layers simultaneously receive features of all previous layers to perform layer feature reuse, and a water level region segmentation map and a water level measurement value are obtained by adaptively learning feature channel importance weights through a squeeze-and-excitation mechanism;

[0030] According to a Bayesian inference framework, in a three-layer probability model including an observation layer, a hidden variable layer and a priori layer, observation data and physical constraints are input as priori knowledge, a posterior probability distribution of water level measurement is calculated by iteratively optimizing parameters of an approximate posterior distribution through a variational inference method, based on the posterior probability distribution, a reliability score table is generated by calculating a prediction mean value and a variance estimation measurement uncertainty value through random deactivation of part of neurons and multiple sampling in network forward calculation;

[0031] A spectrum analysis is performed on the high-confidence water level data set obtained through screening, time-frequency analysis is realized through continuous wavelet transform, a time series is decomposed and reconstructed using a spectral decomposition method, a plurality of intrinsic characteristic functions are calculated using an adaptive signal decomposition algorithm based on an extreme point envelope, and a periodic characteristic vector and a trend characteristic vector are extracted through amplitude threshold control of the decomposition process.

[0032] In an alternative embodiment, inputting the enhanced feature maps into a dense connection network, decomposing a standard convolution into a depth convolution and a point-wise convolution to reduce the calculation amount by a depthwise separable convolution, using a dense connection to make nodes in the later layers simultaneously receive features of all previous layers to perform layer feature reuse, and adaptively learning feature channel importance weights through a squeeze-and-excitation mechanism to obtain a water level region segmentation map and a water level measurement value include:

[0033] A depthwise separable convolution is used to extract features from the enhanced feature maps, a standard convolution is decomposed into a depth convolution and a point-wise convolution, the depth convolution extracts spatial correlation features by independently performing convolution operations on each input channel, and the point-wise convolution obtains reorganized features by combining channel information of the spatial correlation features;

[0034] The reorganized features are sequentially input into a first dense connection block, a second dense connection block and a third dense connection block, a plurality of deep separation convolution layers are arranged in each of the dense connection blocks, a first layer of the first dense connection block performs feature extraction on the reorganized features to obtain first layer output features, a second layer performs feature extraction on the reorganized features and the first layer output features after concatenation to obtain second layer output features, and a third layer performs feature extraction on the reorganized features, the first layer output features and the second layer output features after concatenation to obtain third layer output features;

[0035] The spatial size of the features between adjacent dense connection blocks is reduced and the number of channels is increased through convolution operation and feature sampling operation to obtain multi-scale features, global feature statistics are performed on the feature maps output by the dense connection blocks to obtain a channel description vector, the channel description vector is input into a first full connection layer to obtain a compressed feature vector, the compressed feature vector is input into a second full connection 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 features to obtain weighted features;

[0036] 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 features of the third dense connection block are up-sampled and fused with the weighted features of the second dense connection block to obtain first fused features, the first fused features are up-sampled and fused with the weighted features of the first dense connection block to obtain a fused feature map, and the fused feature map is input into a segmentation branch and a measurement branch respectively, a water level area segmentation map with the same size as the enhanced feature map is output through multi-layer convolution operation of the segmentation branch, and a water level measurement value is output through feature pooling operation and full connection operation of the measurement branch.

[0037] In an optional implementation, a nonlinear state space model is constructed based on the periodic feature vector and the trend feature vector, a state estimation sequence is calculated by combining particle filtering, a causal discovery network is constructed, a causal relationship matrix of water level change is calculated based on the causal discovery network, a plurality of sub-matrices are decomposed, a hierarchical attention network is constructed on different spatial scales, a water level feature tensor is obtained by execution, an encryption model parameter update amount is obtained by combining a homomorphic encryption algorithm through a federated optimization objective function, an optimal sampling strategy vector is calculated by a pre-set multi-agent collaborative decision 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 scheme and an execution parameter matrix, and an abnormal propagation path graph and an emergency monitoring parameter sequence are analyzed based on the causal relationship matrix, including:

[0038] Collect water level data of each monitoring site according to a preset sampling period, perform wavelet decomposition on the water level data to obtain periodic components and trend components, construct the periodic components as a periodic feature vector, and construct the trend components as a trend feature vector;

[0039] Construct a nonlinear state space model by taking the periodic feature vector as an observation variable and the trend feature vector as a state variable, generate a preset number of particle samples for each state variable, perform iterative calculation on the particle samples by combining particle filtering and importance sampling, calculate a likelihood probability value of each particle sample as a particle weight based on observation data, and perform resampling on the particle samples to obtain a state estimation sequence;

[0040] Construct a causal discovery network based on the state estimation sequence, calculate a transfer entropy value between nodes as a causal relationship strength to obtain a causal relationship matrix, divide the causal relationship matrix into multiple regions and decompose it into multiple sub-matrices according to spatial positions of monitoring points;

[0041] Construct a hierarchical attention network on different spatial scales, a first layer of the attention network inputs a sub-matrix representing an internal causal relationship of a region, a second layer of the attention network inputs a sub-matrix representing a causal relationship between adjacent regions, and a third layer of the attention network inputs a sub-matrix representing a causal relationship between remote regions, and a water level feature tensor is obtained by weighting and fusing features at different spatial scales through attention weights;

[0042] Design a federated optimization objective function, take each monitoring site as a participant of federated learning, use a homomorphic encryption algorithm to encrypt model parameters, use local sampling data to calculate model parameter gradients and encrypt them by each participant, aggregate encrypted model parameter gradients to obtain an encrypted model parameter update amount by a server, and obtain global model parameters through a preset number of iterations of optimization;

[0043] An optimal sampling strategy is calculated through a pre-set multi-agent collaborative decision-making system, wherein the multi-agent collaborative decision-making system includes multiple agents, the agents correspond to multiple monitoring regions respectively, the agents interact water level feature tensor information through a communication network, a state value function is calculated based on the water level feature tensor, and an optimal sampling strategy vector is obtained by using a collaborative decision-making algorithm, abnormal analysis is performed on collected water level data, time series features and spatial distribution features are extracted to obtain an abnormal feature vector;

[0044] add the optimal sampling strategy vector and the abnormal feature vector to a hierarchical reinforcement learning network, a state evaluation subnetwork in the hierarchical reinforcement learning network outputs a state vector, an action generation subnetwork outputs a detection resource allocation scheme matrix, a value estimation subnetwork predicts long-term returns of the detection resource allocation scheme to obtain a return vector, and an execution parameter matrix is obtained through policy optimization;

[0045] analyze abnormal propagation paths based on the execution parameter matrix and the causal relationship matrix, locate abnormal event positions based on the execution parameter matrix, backtrack abnormal propagation links through the causal relationship matrix, analyze abnormal influence ranges and draw abnormal propagation path diagrams, determine key monitoring areas and increase sampling frequencies, and generate emergency monitoring parameter sequences containing monitoring areas and monitoring frequencies.

[0046] In an optional implementation, analyzing abnormal propagation paths based on the execution parameter matrix and the causal relationship matrix, locating abnormal event positions based on the execution parameter matrix, backtracking abnormal propagation links through the causal relationship matrix, analyzing abnormal influence ranges and drawing abnormal propagation path diagrams, determining key monitoring areas and increasing sampling frequencies, and generating emergency monitoring parameter sequences containing monitoring areas and monitoring frequencies include:

[0047] Based on the execution parameter matrix and the causal relationship matrix, the water level data is segmented and processed 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 the conditional probability and accumulating, and the transfer entropy value is filled into the corresponding position of the causal relationship matrix.

[0048] An execution parameter matrix is constructed, 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 historical data statistical analysis, multi-level abnormality discrimination rules are established based on the execution parameters, and real-time data of monitoring points are evaluated for abnormality degree in combination with a fuzzy inference mechanism;

[0049] A depth-first search strategy is used for backtrack analysis in the causal relationship matrix, a search direction is determined based on a causal relationship strength threshold from an abnormal point, neighboring nodes with the strongest causal relationship are preferentially accessed, an access mark table is maintained to record accessed nodes, and a complete propagation link from an abnormal source to an influence end is obtained through iterative search;

[0050] Based on the complete propagation link, the abnormal propagation path is analyzed by a hierarchical processing strategy, the correlation between the abnormal point and the surrounding monitoring point in the local area is analyzed to obtain the local propagation characteristics, the analysis range is expanded to the adjacent area to study the propagation mode of a larger spatial scale, and the abnormal propagation network structure is obtained by comprehensively analyzing the results of each level on a global scale;

[0051] According to the abnormal propagation network structure, the spatial arrangement of the monitoring points and the propagation path is performed by an adaptive layout algorithm, a base map is constructed based on geographic information data, and the monitoring points are projected onto the base map, the propagation path is drawn by using a Bezier curve to realize smooth transition, the size and color of the arrow are dynamically adjusted according to the propagation intensity, the key monitoring area is determined by combining a clustering analysis method, the nodes in the abnormal propagation network are clustered according to the position and abnormal degree to identify the abnormal concentrated area, the importance score of each clustering area is calculated, the abnormal degree, influence range and propagation speed are comprehensively considered, the monitoring area is divided into levels according to the importance score, and the monitoring frequency is determined;

[0052] A dynamic programming method is used to generate an emergency monitoring parameter sequence, a monitoring resource allocation optimization model is established with the maximum abnormal monitoring effect as the objective function, the number of monitoring devices and the communication bandwidth are used as resource constraints, the optimal monitoring parameter configuration scheme is solved, the monitoring parameters are adjusted online based on the abnormal development trend, and the emergency monitoring parameter sequence including the monitoring area and the monitoring frequency is obtained.

[0053] In a second aspect of the embodiment of the present application, a water level distributed monitoring system based on a visual sensing network is provided, comprising:

[0054] The first unit is configured to collect water level image samples of a monitoring area, 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 an energy distribution density, generate an energy mask graph corresponding to the water level image samples, guide an attention mechanism to perform self-supervised decomposition on the image through the energy mask graph, 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 contrast learning network, calculate a 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 the nodes to obtain a node correlation probability matrix, and input the node correlation probability matrix and the discriminative feature vector group into a graph neural network to obtain a parameter matrix of the water level propagation prediction model through message passing iteration.

[0055] The second unit is used for receiving a real-time water level image, inputting a 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 decoupling representation network, obtaining a water level feature map through separation and calculating a complementarity measure value based on the water level feature map, adaptively fusing to obtain an enhanced feature map based on the complementarity measure value, inputting the enhanced feature map into a dense connection network, obtaining a water level region segmentation map and a water level measurement value through cross-layer feature reuse, calculating a posterior probability distribution of the water level measurement according to a Bayesian inference framework combined with prior knowledge and observation data, calculating a measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screening a high-confidence water level data set and performing spectral analysis to extract a periodic feature vector and a trend feature vector;

[0056] The third unit is used for constructing a nonlinear state space model based on the periodic feature vector and the trend feature vector, calculating a state estimation sequence combined with a particle filter and constructing a causal discovery network, calculating a causal relationship matrix of the water level change based on the causal discovery network and decomposing to obtain a plurality of sub-matrices, constructing a hierarchical attention network on different spatial scales and performing to obtain a water level feature tensor, combining a homomorphic encryption algorithm to obtain an encrypted model parameter update amount through a federated optimization objective function, calculating an optimal sampling strategy vector through a pre-set multi-agent collaborative decision system and performing abnormal analysis on the water level to obtain an abnormal feature vector, adding the optimal sampling strategy vector and the abnormal feature vector to a hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix, and analyzing to obtain an abnormal propagation path graph and an emergency monitoring parameter sequence combined with the causal relationship matrix.

[0057] In the present application, through the technical means of multi-modal contrast learning, adaptive wavelet transform, decoupling representation and dense connection, water level features can be effectively extracted, the influence of noise and environmental interference is reduced, and high-precision water level measurement is realized. The Bayesian inference framework and the reliability scoring mechanism further improve the confidence of the measurement result, effectively avoiding the influence of abnormal data, constructing a nonlinear state space model, combining particle filtering and causal discovery network, which can effectively analyze the periodicity, trend and causality of water level change, realize the prediction of future water level change, and use hierarchical reinforcement learning and multi-agent collaborative decision system, combined with abnormal analysis results and causal relationship matrix, can dynamically adjust the allocation scheme and execution parameters of monitoring resources, realize the rapid response and effective disposal of abnormal water level events. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 It is a flowchart of the water level distributed monitoring method based on a visual sensing network according to an embodiment of the present application.

[0059] Figure 2 Structure diagram of a water level distributed monitoring system based on a visual sensing network according to an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0061] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0062] Figure 1 Flowchart of a water level distributed monitoring method based on a visual sensing network according to an embodiment of the present application, as shown in FIG. 4, the method comprises the following steps. Figure 1

[0063] S1. Collecting water level image samples of a monitoring area and constructing a spectrum decomposition network to extract frequency domain features of the water level image samples, constructing a multi-scale spectrum feature matrix, constructing an energy matrix based on the multi-scale spectrum feature matrix and calculating an energy distribution density, generating an energy mask graph corresponding to the water level image samples, guiding an attention mechanism to perform self-supervised decomposition on the image through the energy mask graph, obtaining a water level scale feature tensor, a water surface texture feature tensor and an environmental background feature tensor, inputting the feature tensors obtained by decomposition into a multi-modal contrast learning network, calculating a discriminative feature vector group through mutual information maximization, constructing a random field model based on the spatial distribution characteristics of visual sensing nodes, calculating the spatial dependence relationship between nodes to obtain a node association probability matrix, 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.

[0064] ​The spectrum decomposition network is a deep learning model specially used for extracting features of different frequency components from signals, which can decompose complex signals into multiple frequency sub-components to capture more levels of information. The energy mask map is a mask map generated by analyzing the energy distribution of different regions in a signal or image, which is used to emphasize or filter out specific regions or frequency bands. The self-supervised decomposition is a learning method that relies on the internal structure of data itself for training without human-labeled supervision signals. It is commonly used in feature decomposition and representation learning. The mutual information is a measure of the dependence between two variables, which describes the amount of information one variable carries about the other. The spatial distribution characteristics describe the distribution pattern or mode of data in space, which can reflect how data changes with spatial coordinates. The random field model is a mathematical model used to describe the random variation of spatial or temporal data. It simulates the uncertainty in complex systems through random variables. Message passing iteration is a computational method used to pass information in a graph structure. It iteratively updates information through node interactions. It is commonly used in graph neural networks and optimization problems.

[0065] In an alternative embodiment, water level image samples of the monitoring area are collected and a spectrum decomposition network is constructed to extract 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 an energy distribution density is calculated. An energy mask map corresponding to the water level image samples is generated. The energy mask map guides the attention mechanism to perform self-supervised decomposition on the image, obtaining water level scale feature tensors, water surface texture feature tensors, and environmental background feature tensors. The decomposed feature tensors are input into a multi-modal contrast learning network, and discriminative feature vector groups are calculated by maximizing mutual information. A random field model is constructed based on the spatial distribution characteristics of the visual sensing nodes, and a node correlation probability matrix is calculated based on the spatial dependence between nodes. The node correlation probability matrix and the discriminative feature vector groups are input into a graph neural network, and the parameter matrix of the water level propagation prediction model is obtained by message passing iteration, including:

[0066] Water level image samples of the monitoring area are collected and input into a multi-level spectrum decomposition network. The shallow network in the multi-level spectrum decomposition network extracts high-frequency components of the water level image samples through a high-pass filter set to obtain water level scale line features and water surface ripple features. The middle network extracts medium-frequency components of the water level image samples through a band-pass filter set to obtain water level line contour features and shore structure features. The deep network extracts low-frequency components of the water level image samples through a low-pass filter set to obtain water area distribution features. The feature maps corresponding to the high-frequency components, medium-frequency components, and low-frequency components are normalized and arranged in frequency band order to form a multi-scale spectral feature matrix.

[0067] An energy value is calculated for each frequency band component in the multi-scale spectrum feature matrix to obtain a two-dimensional energy matrix, a kernel function is used to perform density estimation on the two-dimensional energy matrix 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, a high-energy region and a low-energy region are determined, the high-energy region is marked as a foreground region, and the low-energy region is marked as a background region;

[0068] The water level image sample is divided into a plurality of overlapping image blocks, the attention weight of the plurality of overlapping image blocks is calculated based on the binary energy mask map, the overlapping image blocks with the attention weight are input into a self-supervised deconstruction module for feature decomposition, a water level scale feature branch is extracted through a fine-grained feature extractor to obtain a water level scale feature tensor, a water surface texture feature branch is extracted through a texture analyzer to obtain a water surface texture feature tensor, and an environment background feature branch is extracted through 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 input into a feature encoder for dimension reduction to obtain feature vectors of a unified dimension, and a feature comparison pool is constructed;

[0069] Positive sample pairs and negative sample pairs are randomly sampled from the feature vectors of the unified dimension, mutual information measurement values between the positive sample pairs and the negative sample pairs are calculated, discriminative feature vector groups are obtained by maximizing the mutual information measurement values of the positive sample pairs and minimizing the mutual information measurement values of the negative sample pairs, a distance matrix between node pairs is calculated based on geographical coordinates of visual sensing nodes, the distance matrix is converted into an initial correlation strength, and node correlation probability matrices between node pairs are obtained by calculating interactions between nodes through a potential function of a random field;

[0070] The node correlation probability matrices and the discriminative feature vector groups are input into a graph neural network, the importance weight of a neighbor node is determined based on the node correlation probability matrix in each round of message passing iteration, a fusion feature is obtained by performing nonlinear transformation on feature information of the neighbor node, the fusion feature and feature information of the node itself are combined to update the node state, and a parameter matrix of a water level propagation prediction model is obtained through multiple iterations.

[0071] 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 signals or images in two-dimensional space, which helps to model and identify spatial features by analyzing the energy intensity at different positions, 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, and the feature contrast pool is a mechanism that compares and integrates different features, usually used to enhance the expression ability of features, and through the comparison of features in the contrast pool, the recognition ability of the model for different samples is improved.

[0072] The water level image samples of the monitoring area are collected. 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, in which multiple sets of filter kernels of different sizes are set, and the filter kernel size increases step by step from small to large, to extract water level scale line features and water surface ripple features. The scale line features are obtained by vertical edge detection, and the water surface ripple features are extracted by a direction-sensitive texture operator. The middle network uses a band-pass filter bank to process the image, and the center frequency and bandwidth of the filter are determined according to the scale features of the water level contour and the shore structure to extract the water level contour features and the shore structure features. The deep network uses a low-pass filter bank to extract low-frequency features such as water distribution by gradually reducing the cutoff frequency. All feature maps are normalized to map the amplitude to a unified interval, and are reorganized into a multi-scale spectral feature matrix in order from high frequency to low frequency.

[0073] The energy values of each frequency band component in the multi-scale spectral feature matrix are calculated to establish a two-dimensional energy matrix. A kernel function is used to estimate the density of the two-dimensional energy matrix, and the bandwidth of the kernel function is optimized and determined by a cross-validation method. Global statistical properties, including mean, variance, skewness, and kurtosis, are calculated based on the energy density distribution map, and the threshold is adaptively set. A binary energy mask map is generated by threshold segmentation, with regions above the threshold marked as foreground regions and regions below the threshold marked as background regions. Morphological processing is performed on the mask map to remove noise and fill in holes, and the water level image sample is divided into multiple overlapping image blocks, with the degree of overlap adaptively adjusted according to the complexity of the image content. The attention weight of each image block is calculated based on the binary energy mask map, considering the proportion of foreground pixels, energy mean, and gradient information within the block. The image blocks with attention weights are input into a self-supervised deconstruction module, which extracts features through three parallel branches: a fine-grained feature extractor extracts water level scale feature tensors from the water level scale feature branch, a texture analyzer extracts water surface texture feature tensors from the water surface texture feature branch, and a context encoder extracts environmental background feature tensors from the environmental background feature branch. The three feature tensors are input into a feature encoder for dimension reduction to obtain feature vectors of a unified dimension, and a feature contrast pool is constructed.

[0074] The positive sample pair and the negative sample pair are constructed by random sampling from the feature contrast pool. The selection of the positive sample pair is based on time correlation and spatial proximity, and the selection of the negative sample pair ensures feature difference. The mutual information metric value between the positive sample pair and the negative sample pair is calculated, and the mutual information of the positive sample pair is maximized and the mutual information of the negative sample pair is minimized through optimization to obtain a discriminative feature vector group. The distance matrix between node pairs is calculated based on the geographic coordinates of the visual sensing nodes, and the distance matrix is converted into an initial association strength through a decay function. The interaction between node pairs is calculated through the potential function of the random field, and the node association probability matrix is obtained by iterative optimization considering the feature similarity and spatial relationship of node pairs.

[0075] The node association probability matrix and the discriminative feature vector group are input into the graph neural network for training. In each round of message passing iteration, the attention score is calculated based on the node association probability matrix to determine the importance weight of the neighbor nodes. The feature information of the neighbor nodes is nonlinearly transformed to obtain the fusion feature. The fusion feature is combined with the feature information of the node itself through a gating mechanism to update the node state. The training process adopts a batch processing mode, and the adaptive optimizer is used to update the parameters, and the learning rate adopts a dynamic adjustment strategy. After multiple iterations, a parameter matrix is obtained which can accurately predict the water level propagation rule.

[0076] In this embodiment, through the spectrum decomposition and energy mask guided self-supervised feature decomposition method, the effective separation of the scale feature, water surface texture feature and environment background feature in the water level image is realized, the discriminative and robustness of the feature expression is improved, based on the multi-modal contrast learning and random field modeling method, the spatial correlation between the visual sensing nodes is fully mined, the accuracy and generalization ability of the water level propagation prediction are effectively improved, the message passing mechanism is adopted, the dynamic fusion and update of the node feature are realized, the problem of difficult modeling of complex water level propagation mode is overcome, and the adaptability of the prediction model is improved.

[0077] In an alternative embodiment, a positive sample pair and a negative sample pair are randomly sampled from the uniform dimension feature vector, a mutual information metric value between the positive sample pair and the negative sample pair is calculated, a discriminative feature vector group is obtained by maximizing the mutual information metric value of the positive sample pair and minimizing the mutual information metric value of the negative sample pair, a distance matrix between node pairs is calculated based on the geographic coordinates of the visual sensing nodes, the distance matrix is converted into an initial association strength, and a node association probability matrix is obtained by calculating the interaction between node pairs through the potential function of the random field, including:

[0078] construct a feature contrast pool from the unified dimension feature vectors and divide the feature contrast pool into multiple time windows, extract scene feature information within each time window and randomly sample to obtain positive sample pairs, and randomly sample feature vectors according to weather conditions, monitoring time periods, and water level change trends in different monitoring scenes to construct negative sample pairs;

[0079] add the feature vectors in the positive sample pairs and the negative sample pairs to a multi-layer perceptron including an input layer, multiple hidden layers, and an output layer for feature transformation, set a batch normalization layer and a nonlinear activation function between adjacent layers of the multi-layer perceptron, map the feature vectors from a feature space to a probability distribution space, construct a Gaussian kernel function group with an adaptive bandwidth in the probability distribution space, and obtain a continuous probability distribution by accumulating local and global contributions of probability densities of all feature points;

[0080] calculate mutual information metric values between the positive sample pairs and the negative sample pairs, wherein the mutual information metric values are obtained by calculating multiple differences of joint distributions, conditional distributions, and marginal distributions of transformed features, construct a mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints, maximize the mutual information metric values of the positive sample pairs and minimize the mutual information metric values of the negative sample pairs, and obtain a feature vector group with local and global discriminativeness by using a gradient descent method with an adaptive learning rate for multiple rounds of iterative optimization;

[0081] based on geographic coordinates of the visual sensing nodes, calculate a distance matrix between node pairs by fusing geodetic distances between nodes, terrain relief coefficients, river connectivity coefficients, and hydrological propagation characteristic coefficients, dynamically correct the distance matrix by combining digital elevation model data to calculate multi-scale elevation differences between nodes, and based on river connectivity data, calculate actual flow distances considering flow directions and flow rates, and perform multi-feature fusion on the actual flow distances and the corrected distance matrix to obtain a corrected distance matrix;

[0082] based on the corrected distance matrix, map distance values to a dynamic interval by an adaptive normalization method, calculate local and global densities based on multi-dimensional spatial distribution density features of the nodes, set an adaptive attenuation coefficient based on a density gradient for densely distributed areas, set a dynamic attenuation coefficient considering spatial heterogeneity for sparsely distributed areas, and calculate initial correlation strengths in multiple scales in combination with spatiotemporal correlation features of the nodes;

[0083] Based on the initial correlation strength, a random field potential function with adaptive segmentation structure is constructed, the node distance is dynamically divided into near distance interval, middle distance interval and far distance interval with transition zone based on the multi-dimensional spatial relationship between nodes, a linear decay function considering local spatial dependence is designed in the near distance interval, an exponential decay function considering regional influence is designed in the middle distance interval, and a residual function considering global correlation is designed in the far distance interval, and the adaptive segmentation point of the potential function is determined through spatial autocorrelation analysis.

[0084] Based on the adaptive segmentation point, the interaction between node pairs is iteratively calculated through the potential function of the random field based on the multi-layer message passing mechanism, the state information of the multi-scale neighborhood nodes is collected based on the spatial dependence relationship in each iteration, the adaptive potential function value is used to weight the message in multiple levels, the spatio-temporal redundant information is identified and suppressed through the structured attention mechanism, the weighted message is combined with nonlinear features and added with random noise subject to dynamic distribution, and the node correlation probability matrix representing the correlation between nodes is obtained through state update.

[0085] The mutual information optimization objective function is an objective function for optimizing model parameters by maximizing mutual information, which is used to improve the learning ability of the model for the dependence relationship between data, and is widely used 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 geographic coordinates. The terrain relief coefficient is an index that describes the degree of surface relief, reflecting the size of the terrain relief. The river connectivity coefficient is an index that measures the connectivity of each river segment in the river network, used in hydrology and ecology research to assess the connectivity of water flow and biological habitats. The hydrological propagation characteristic coefficient is a coefficient that describes the speed and mode of water flow propagation, helping to simulate the relationship between precipitation and runoff. The digital elevation model data is a digital representation of the terrain elevation data set. The dynamic decay coefficient considering spatial heterogeneity is a coefficient that describes how information or influence decays with distance in space after considering spatial heterogeneity factors. The adaptive potential function value refers to the function value that is adaptively adjusted according to the change of data, which dynamically adjusts the parameters according to the change of environment. The spatio-temporal redundant information refers to the repeated information in spatio-temporal data due to the similarity of time or space, which is usually processed through feature compression or de-redundancy techniques to improve data analysis efficiency.

[0086] A feature contrast pool is constructed and divided into time windows, each with a width of 30 minutes. Within the window, positive sample pairs are randomly sampled according to scene information, including samples with similar water levels under the same monitoring scene as positive samples. For negative sample sampling, three layers of stratification are performed according to weather, monitoring period and water level trend, and feature vectors are randomly sampled under different stratification combinations to construct negative sample pairs.

[0087] 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, and 32, respectively), and an output layer (dimension 16). Batch normalization layers and ReLU activation functions are added between adjacent layers. The transformed features construct a set of Gaussian kernel functions in the probability space, and the kernel bandwidth is adaptively determined by the average distance of the nearest neighbor 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 discriminability constraints and distribution consistency constraints. The Adam optimizer is used, with a learning rate starting at 0.001 and decaying by 0.1 every 50 rounds, and iterating for 500 rounds to obtain a set of discriminative feature vectors. Based on the geographic coordinates of the visual sensing nodes, a distance matrix between nodes is calculated. The geodetic distance (weight 0.4), terrain relief coefficient (weight 0.2), river connectivity coefficient (weight 0.2), and hydrological propagation feature coefficient (weight 0.2) are fused. Based on DEM data, a multi-scale (1 km, 5 km, 10 km) elevation difference correction distance matrix is calculated. Based on the river connectivity data, considering the flow direction and flow rate (0.5-2 m / s), the actual flow distance is calculated, and the corrected distance matrix is obtained by fusing the corrected distance matrix. Based on the corrected distance matrix, the initial association strength is calculated. 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 node are calculated.

[0088] The node distance is divided into a near distance (2km) interval, and the transition bandwidth is set to 100m. The near distance interval uses a linear decay function (slope -0.2), the middle distance interval uses an exponential decay function (decay rate 0.5), and the far distance interval uses a residual function. The piecewise points of the potential function are determined by spatial autocorrelation analysis, and the node interaction is iteratively calculated based on a three-layer message passing mechanism. In each iteration, the state information of the neighborhood nodes within 200m, 500m, and 1km is collected, and the potential function value is weighted to the message. The spatiotemporal redundant information is identified through structured attention, and the weighted messages are nonlinearly combined and added with Gaussian random noise (mean 0, variance 0.1) to update the node state to obtain the association probability matrix.

[0089] In this embodiment, through the layered sampling strategy and multi-layer perceptron feature transformation, combined with the mutual information optimization target, the discriminability and expression ability of the feature vector are improved, the recognition accuracy of the node association relationship is enhanced, the multi-dimensional spatial information is fused to calculate and correct the distance matrix, the adaptive normalization and density perception decay mechanism are used to calculate the initial association strength, the rationality of the node spatial association representation is improved, the adaptive segmented random field potential function is designed, the node interaction is calculated based on the multi-layer message passing and structured attention mechanism, and the robustness and generalization ability of the node association probability are enhanced.

[0090] S2. Receiving real-time water level images, inputting the parameter matrix of the water level propagation prediction model into the adaptive wavelet transform network, calculating the optimal decomposition scale sequence and generating the multi-resolution feature representation, obtaining the compressed feature code stream through sparse coding of the multi-resolution feature representation, inputting the feature code stream into the decoupling representation network, obtaining the water level feature map through separation and calculating the complementarity measure value based on the water level feature map, performing adaptive fusion based on the complementarity measure value to obtain an enhanced feature map, inputting the enhanced feature map into the dense connection network, obtaining the water level region segmentation map and water level measurement value through cross-layer feature reuse, calculating the posterior probability distribution of water level measurement according to the Bayesian inference framework combined with prior knowledge and observation data, calculating the measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screening a high-confidence water level dataset and performing spectral analysis to extract periodic feature vectors and trend feature vectors;

[0091] The sparse coding is a feature learning method that represents input signals as a linear combination of a few basis elements to achieve data compression and efficient representation. The compressed feature code stream refers to a compact code stream format obtained by a specific compression algorithm to reduce storage and transmission overhead. The decoupling representation network is a neural network architecture designed to decompose the complex representation of input data into simpler, independent components to improve the interpretability and training efficiency of the model. The complementarity measure value is an indicator used to evaluate the degree of complementarity between two sets of features or information. By calculating the complementarity measure, we can discover parts of the data that have not been fully utilized. The dense connection network is a network architecture that connects each layer with all previous layers to facilitate the full propagation of information, thereby improving the training efficiency and performance of the network. The Bayesian inference framework is a reasoning method based on Bayes' theorem, used to infer unknown variables or parameters given prior information.

[0092] In an alternative embodiment, a real-time water level image is received, a parameter matrix of a 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 sparse coding of the multi-resolution feature representation, the feature code stream is input into a decoupled representation network, a water level feature map is obtained by separation and a complementarity measure value is calculated based on the water level feature map, an enhanced feature map is obtained by adaptive fusion based on the complementarity measure 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 by cross-layer feature reuse, a posterior probability distribution of the water level measurement is calculated according to a Bayesian inference framework combined with prior knowledge and observation data, a measurement uncertainty value is calculated based on the posterior probability distribution to generate a reliability score table, a high-confidence water level dataset is screened and spectral analysis is performed to extract a periodic feature vector and a trend feature vector, including:

[0093] A real-time water level image is received, a parameter matrix of a water level propagation prediction model is input into an adaptive wavelet transform network, image gradients in multiple directions are calculated to construct a direction gradient histogram, a gray level co-occurrence matrix is calculated on different direction and distance combinations to extract multi-dimensional texture features including energy, entropy, contrast and correlation, an optimal wavelet basis function is selected by adaptive weight fusion according to the multi-dimensional texture features, an optimal decomposition scale sequence is calculated by analyzing the local extreme value characteristics and energy distribution of wavelet coefficients, low-frequency approximation components and high-frequency detail components are obtained by multi-layer decomposition based on the optimal decomposition scale sequence through row-column transformation and downsampling operations, the multi-resolution feature representation is generated by tree decomposition of the high-frequency detail components;

[0094] The multi-resolution feature representation is sparse coded by constructing an overcomplete dictionary containing redundant atoms, the features are blocked according to a fixed size, sparse decomposition based on orthogonal matching pursuit is performed on each feature block, the dictionary atom most relevant to the residual is iteratively selected and the reconstruction coefficient is updated until the sparsity constraint is met to obtain a compressed feature code stream, the compressed feature code stream is input into a decoupled representation network, multi-layer features are extracted by a series of residual calculation units, wherein the residual calculation unit includes a convolution layer, a normalization layer and an activation function, different scale context information is extracted by a multi-scale pooling module, the feature resolution is gradually restored by inverse convolution operation, channel weights are calculated based on feature statistics and nonlinear transformation to separate a water level feature map;

[0095] The local region structure similarity is calculated using a Gaussian weighted window based on the water level feature map, the horizontal and vertical gradients are calculated to construct a gradient direction field and analyze the direction consistency, the features are subjected to Fourier transform to calculate a normalized cross-power spectrum, the best matching position is determined by local maximum value detection to obtain a complementarity measure value, a feature fusion relationship graph is constructed based on the complementarity measure value, the nodes in the feature fusion relationship graph represent features and the weights of the edges are determined by the complementarity measure value, the feature correlation score is calculated by a multi-layer graph attention calculation unit and normalized, and the normalized attention score is combined with the feature weighting to obtain an enhanced feature map through adaptive fusion;

[0096] The enhanced feature map is input into a dense connection network, the standard convolution is decomposed into a depth convolution and a point-wise convolution to reduce the calculation amount through a depth separation convolution, the dense connection is adopted to enable the nodes in the later layer to simultaneously receive the features of all the previous layers to perform layer feature reuse, and the water level region segmentation map and the water level measurement value are obtained through adaptive learning of the feature channel importance weight by a compression excitation mechanism;

[0097] According to a Bayesian inference framework, in a three-layer probability model including an observation layer, a hidden variable layer and a priori layer, the observation data and the physical constraint are input as priori knowledge, the approximate posterior distribution parameters are calculated to obtain the posterior probability distribution of the water level measurement through iterative optimization of a variational inference method, and based on the posterior probability distribution, the prediction mean value and the variance estimation measurement uncertainty value are calculated through random inactivation of part of the neurons and multiple sampling in the network forward calculation to generate a reliability score table;

[0098] The high-confidence water level data set obtained through screening is subjected to spectral analysis, time-frequency analysis is realized through continuous wavelet transform, the time series is decomposed and reconstructed through a spectral decomposition method, a plurality of intrinsic characteristic functions are calculated through an adaptive signal decomposition algorithm based on an extreme point envelope, and the period characteristic vector and the trend characteristic vector are extracted through amplitude threshold control decomposition.

[0099] The gray level co-occurrence matrix is a texture analysis method that describes the texture characteristics of an image by calculating the spatial relationship of pixel gray values in the image, the row and column transformation refers to the exchange operation of rows and columns of a matrix, the down-sampling operation refers to an operation in signal processing or image processing that reduces the calculation complexity by reducing the sampling frequency or the number of pixels of data, the tree decomposition is a process of decomposing data or problems into a tree structure, the redundant atom refers to a base atom in dictionary learning that does not have independence, the over-complete dictionary is a dictionary composed of a plurality of base elements, the base elements of which are more than the dimension of the signal, allowing a more optimal signal reconstruction to be found through sparse representation, and the adaptive signal decomposition algorithm based on the envelope of extreme points is an algorithm that adaptively decomposes a signal by extracting extreme points in the signal.

[0100] Real-time water level images are acquired and transmitted to a server for processing. Preprocessing of the water level images is performed, which includes image denoising, correction, and enhancement operations to improve 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. The preprocessed water level images are input into an adaptive wavelet transform network for feature extraction. The network analyzes the gray level changes of the image in different directions and distances to extract multi-dimensional texture features. Based on the texture features, the best wavelet basis function is adaptively selected, and the optimal decomposition scale sequence is calculated. Based on this sequence, the image is decomposed into multiple layers of wavelets to obtain low-frequency approximation components and high-frequency detail components. The high-frequency detail components are tree-structured decomposed to generate multi-resolution feature representations.

[0101] Sparse coding is performed on the multi-resolution feature representations. A redundant atom-containing overcomplete dictionary is constructed, and the features are blocked according to a fixed size. Sparse decomposition based on orthogonal matching pursuit is performed on each feature block. Iterative selection of the most relevant dictionary atom to the residual error is performed, and the reconstruction coefficients are updated until the sparsity constraint is met, resulting in a compressed feature code stream. For example, a dictionary containing 1024 atoms is used, and each feature block is decomposed into a maximum of 10 atoms. The compressed feature code stream is input into a decoupled representation network. The decoupled representation network extracts multi-layer features through a series of residual calculation units. Each residual calculation unit includes a convolution layer, a normalization layer, and an activation function. Different scales of context information are extracted through a multi-scale pooling module, and the feature resolution is gradually restored through a deconvolution operation. Channel weights are calculated based on feature statistics and nonlinear transformation, and separation is performed to obtain water level feature maps.

[0102] Based on the water level feature maps, a complementarity measure value is calculated. A Gaussian weighted window is used to calculate the local region structure similarity, and horizontal and vertical gradients are calculated to construct a gradient direction field and analyze the direction consistency. The features are Fourier transformed to calculate the normalized cross-power spectrum, and the best matching position is determined through local maximum value detection to obtain the complementarity measure value. For example, a 5x5 Gaussian window is used to calculate the structure similarity.

[0103] Based on the complementarity measure value, adaptive fusion is performed. A feature fusion relationship graph is constructed, with nodes representing features and edge weights determined by the complementarity measure value. Feature correlation scores are calculated and normalized through multiple layers of graph attention calculation units, and the normalized attention scores are combined with the features to obtain enhanced feature maps. The enhanced feature maps are input into a densely connected network. The network uses depthwise separable convolution and dense connection to achieve cross-layer feature reuse. Feature channel importance weights are adaptively learned through a compression excitation mechanism to obtain water level region segmentation maps and water level measurement values.

[0104] According to the Bayesian inference framework, the posterior probability distribution and uncertainty of the water level measurement are calculated. The observed data and physical constraints are input as prior knowledge into a three-layer probability model, and the approximate posterior distribution parameters are iteratively optimized by the variational inference method to calculate the posterior probability distribution of the water level measurement. Based on the posterior probability distribution, the prediction mean and variance are calculated by randomly inactivating part of the neurons and sampling multiple times in the network forward calculation, the measurement uncertainty value is estimated, and a reliability score table is generated.

[0105] Spectral analysis is performed on the high-confidence water level dataset. Time-frequency analysis is achieved through continuous wavelet transform, time series are decomposed and reconstructed using spectral decomposition method, multiple intrinsic characteristic functions are calculated using adaptive signal decomposition algorithm based on extreme point envelope, and periodic characteristic vectors and trend characteristic vectors are extracted by amplitude threshold control decomposition process. For example, daily and annual period characteristics of water level changes are extracted.

[0106] In this embodiment, the expression ability of water level features is effectively improved through multi-scale feature extraction, sparse coding, decoupling representation and adaptive fusion, thereby improving the accuracy of water level measurement. Based on the Bayesian inference framework, the posterior probability distribution and uncertainty of the water level measurement are calculated based on prior knowledge and observed data, a reliability score table is generated, and the reliability of the water level measurement is improved. Periodic characteristic vectors and trend characteristic vectors are extracted through spectral analysis, which can be used to predict water level changes and provide scientific basis for water resource management and flood warning.

[0107] In an alternative embodiment, the enhanced feature map is input into a dense connection network, the standard convolution is decomposed into depth convolution and pointwise convolution to reduce the calculation amount through depth separation convolution, the dense connection is adopted to make the posterior layer nodes receive all the features of the previous layers at the same time to perform layer feature reuse, and the water level region segmentation map and water level measurement value are obtained through the compression excitation mechanism to adaptively learn the importance weight of the feature channel.

[0108] The enhanced feature map is extracted using depth separation convolution, which decomposes the standard convolution into depth convolution and pointwise convolution. The depth convolution extracts spatial correlation features by independently performing convolution operations on each input channel, and the pointwise convolution obtains reorganized features by combining channel information of the spatial correlation features.

[0109] The reorganized features are sequentially input into a first dense connection block, a second dense connection block and a third dense connection block, a plurality of deep separation convolution layers are arranged in each of the dense connection blocks, a first layer of the first dense connection block performs feature extraction on the reorganized features to obtain first layer output features, a second layer performs feature extraction on the reorganized features and the first layer output features after concatenation to obtain second layer output features, and a third layer performs feature extraction on the reorganized features, the first layer output features and the second layer output features after concatenation to obtain third layer output features;

[0110] The spatial size of the features between adjacent dense connection blocks is reduced and the number of channels is increased through convolution operation and feature sampling operation to obtain multi-scale features, global feature statistics are performed on the feature maps output by the dense connection blocks to obtain a channel description vector, the channel description vector is input into a first full connection layer to obtain a compressed feature vector, the compressed feature vector is input into a second full connection 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 features to obtain weighted features.

[0111] 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 features of the third dense connection block are upsampled and fused with the weighted features of the second dense connection block to obtain first fused features, the first fused features are upsampled and fused with the weighted features of the first dense connection block to obtain a fused feature map, and the fused feature map is input into a segmentation branch and a measurement branch respectively, a water level area segmentation map with the same size as the enhanced feature map is output through multi-layer convolution operation of the segmentation branch, and a water level measurement value is output through feature pooling operation and full connection operation of the measurement branch.

[0112] The deep separation convolution layer is a convolutional neural network layer, which significantly reduces the calculation amount and improves the efficiency by decomposing the standard convolution operation into deep convolution and point-by-point convolution, and is commonly used in lightweight network structures, and the channel description vector is a vector used to describe the features of each channel in a convolutional neural network, which can compress and represent the features of the network by encoding the features of each channel.

[0113] The enhanced feature map is subjected to deep separation convolution processing. The standard convolution operation is decomposed into two consecutive steps: the first step is deep convolution, which independently performs convolution operation on each channel of the input feature map to extract the feature correlation in the spatial dimension; the second step is point-by-point convolution, which linearly combines the feature map output by the deep convolution in the channel dimension to realize the recombination and fusion of information between channels. The dense connection structure adopts three dense connection blocks in series to process the recombined features. Each dense connection block internally includes three layers of deep separation convolution layers, and the layers are connected in a dense connection manner. Taking the first dense connection block as an example: the first layer directly performs convolution processing on the recombined features to obtain the first layer output; the second layer performs convolution processing after connecting the recombined features and the first layer output in the channel dimension; the third layer performs convolution processing after connecting the recombined features, the first layer output and the second layer output. The second dense connection block and the third dense connection block adopt the same connection manner. A feature transformation module is arranged between adjacent dense connection blocks. The spatial resolution of the feature map is reduced through convolution operation, and the number of feature channels is increased, obtaining feature representations of multiple scales. The feature map output by each dense connection block is subjected to global average pooling to obtain a description vector reflecting the importance of each channel. The channel description vector is input into a two-layer fully connected network, the first layer plays a feature compression role, and the second layer performs feature excitation, and finally the channel weight coefficient is obtained through normalization. Multiply the weight coefficient by the multi-scale features to realize adaptive weighting of the features, and the feature fusion stage adopts a bottom-up progressive fusion strategy. The features of the third dense connection block are fused with the features of the second dense connection block through upsampling to obtain the first level fusion features; then the first level fusion features are further upsampled and fused with the features of the first dense connection block to obtain the fusion feature map, and the task output stage includes two parallel branches. The segmentation branch processes the fusion feature map through multiple convolution operations, finally outputs a water level area segmentation map with the same size as the input image, realizing pixel-level region recognition. The measurement branch performs feature pooling dimension reduction on the fusion feature map, and obtains specific water level measurement values through a fully connected layer regression.

[0114] Exemplarily, an enhanced feature map with a resolution of 1920x1080 and containing 64 feature channels is input. Each channel is independently processed using a 3x3 convolution kernel, and 64 channels are reorganized into 128 channels using a 1x1 convolution kernel. The first dense connection block has a three-layer structure: 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; and the third layer inputs 192-channel features (128+32+32) and outputs 32-channel features. The second and third dense connection blocks have the same channel configuration, and the feature transformation module uses a convolution with a step of 2 to sequentially reduce the feature map size to 960x540, 480x270, and 240x135. The number of channels is correspondingly increased to 256, 512, and 1024. A 1024-dimensional channel description vector is obtained through global average pooling, which is compressed to 64 dimensions by the first fully connected layer and is remapped to 1024 dimensions by the second fully connected layer. After normalization, the channel weight is obtained. When the features are fused, the 240x135 size feature is upsampled to 480x270 and fused with the second dense connection block feature, and then upsampled to 960x540 and fused with the first dense connection block feature. The segmentation branch restores the feature map to a resolution of 1920x1080 through four convolution layers, and outputs a water level area probability map. The measurement branch outputs a water level value in the range of 0-100 through global average pooling and two fully connected layers.

[0115] In this embodiment, the standard convolution is decomposed into a depth convolution and a pointwise convolution using a depthwise separable convolution, which effectively reduces the number of model parameters and computational complexity, improves the computational efficiency, and enables the nodes in the later layers of the network to simultaneously receive features from all previous layers through the dense connection mechanism, thereby realizing layer feature reuse and enhancing the feature expression capability of the network. The compression excitation mechanism adaptively learns the importance weight of the feature channel, further improving the network's ability to extract key information. Through the design of multi-scale feature fusion and segmentation and measurement branches, the water level area can be accurately segmented, and the water level value can be accurately measured, thereby providing reliable data support for water resource management and flood control and disaster reduction.

[0116] S3. Constructing a nonlinear state space model based on the periodic feature vector and the trend feature vector, combining particle filtering to calculate the state estimation sequence and construct a causal discovery network, calculating the causal relationship matrix of water level change based on the causal discovery network and decomposing to obtain multiple sub-matrices, constructing a hierarchical attention network on different spatial scales and executing to obtain a water level feature tensor, obtaining an encrypted model parameter update amount through federated optimization objective function combined with homomorphic encryption algorithm, calculating the optimal sampling strategy vector through the pre-set multi-agent collaborative decision system and performing abnormal analysis on the water level to obtain an abnormal feature vector, adding the optimal sampling strategy vector and the abnormal feature vector to the hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix, and analyzing the abnormal propagation path diagram and the emergency monitoring parameter sequence based on the causal relationship matrix.

[0117] The nonlinear state space model is a mathematical model used to describe the dynamic behavior of a system, suitable for cases where there is a nonlinear relationship between system states. The particle filtering is a recursive estimation algorithm based on the Monte Carlo method, which uses a set of particles (i.e. samples) to represent the probability distribution of the state space, and can handle state estimation problems of nonlinear and non-Gaussian systems. The causal relationship matrix is a matrix that describes the causal relationship between different variables, usually obtained through statistical methods or model learning, which helps to identify the causal dependence structure in the system. The federated optimization objective function is a target function used in the optimization of multi-party data and privacy-limited environments, which usually ensures that each party's calculation is independent, and the global optimization is completed by aggregating the local information of each party. The homomorphic encryption algorithm is an encryption method that allows calculations to be performed on encrypted data without decrypting the data, thereby ensuring data privacy. The encrypted model parameter update amount refers to the measurement of how the parameters change and update in the homomorphic encryption model during the calculation process, which is used to guide the process optimization of encryption operations. The multi-agent collaborative decision system refers to multiple agents in a system interacting and cooperating with each other based on their states and goals to make decisions together.

[0118] In an optional embodiment, a nonlinear state space model is constructed based on the periodic feature vector and the trend feature vector, a state estimation sequence is calculated in combination with particle filtering, and a causal discovery network is constructed, a causal relationship matrix of water level changes is calculated based on the causal discovery network, and a plurality of sub-matrices are decomposed, a hierarchical attention network is constructed on different spatial scales, and a water level feature tensor is obtained, an encrypted model parameter update is obtained by combining a homomorphic encryption algorithm through a federated optimization objective function, an optimal sampling strategy vector is calculated through a pre-set multi-agent collaborative decision 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 scheme and an execution parameter matrix, and an abnormal propagation path graph and an emergency monitoring parameter sequence are analyzed in combination with the causal relationship matrix, including:

[0119] Water level data is continuously collected at each monitoring site according to a pre-set sampling period, and wavelet decomposition is performed on the water level data to obtain periodic components and trend components, the periodic components are constructed into a periodic feature vector, and the trend components are constructed into a trend feature vector;

[0120] A nonlinear state space model is constructed with the periodic feature vector as an observation variable and the trend feature vector as a state variable, a pre-set number of particle samples are generated for each state variable, the particle samples are iteratively calculated through importance sampling in combination with particle filtering, and a likelihood probability value of each particle sample is calculated based on observation data as a particle weight, and the particle samples are resampled to obtain a state estimation sequence;

[0121] A causal discovery network is constructed based on the state estimation sequence, a transfer entropy value between nodes is calculated as a causal relationship strength to obtain a causal relationship matrix, the causal relationship matrix is divided into a plurality of regions according to the spatial positions of the monitoring points and decomposed into a plurality of sub-matrices;

[0122] A hierarchical attention network is constructed on different spatial scales, a first layer attention network inputs a sub-matrix representing the causal relationship within a region, a second layer attention network inputs a sub-matrix representing the causal relationship between adjacent regions, and a third layer attention network inputs a sub-matrix representing the causal relationship between remote regions, and a water level feature tensor is obtained by weighting and fusing features at different spatial scales through attention weights;

[0123] A federated optimization objective function is designed, each monitoring site is taken as a participant in federated learning, a homomorphic encryption algorithm is used to encrypt model parameters, each participant calculates model parameter gradients using local sampling data and encrypts them, a server aggregates encrypted model parameter gradients to obtain encrypted model parameter updates, and global model parameters are obtained through a pre-set number of iteration optimizations;

[0124] An optimal sampling strategy is calculated by a pre-set multi-agent collaborative decision-making system, wherein the multi-agent collaborative decision-making system comprises a plurality of agents, the agents correspond to a plurality of monitoring areas respectively, the agents interact the water level feature tensor information through a communication network, a state value function is calculated based on the water level feature tensor, and an optimal sampling strategy vector is obtained by using a collaborative decision-making algorithm, abnormal analysis is performed on the collected water level data, time series features and spatial distribution features are extracted to obtain an abnormal feature vector;

[0125] The optimal sampling strategy vector and the abnormal feature vector are added to a hierarchical reinforcement learning network, a state evaluation subnetwork in the hierarchical reinforcement learning network outputs a state vector, an action generation subnetwork outputs a detection resource allocation scheme matrix, a value estimation subnetwork predicts long-term returns of the detection resource allocation scheme to obtain a return vector, and an execution parameter matrix is obtained by policy optimization;

[0126] Based on the execution parameter matrix and the causal relationship matrix, an abnormal propagation path is analyzed, based on the execution parameter matrix, an abnormal event location is located, through the causal relationship matrix, an abnormal propagation link is traced back, an abnormal influence range is analyzed and an abnormal propagation path diagram is drawn, a key monitoring area is determined and a sampling frequency is increased, and an emergency monitoring parameter sequence comprising a monitoring area and a monitoring frequency is generated.

[0127] The transfer entropy value is a statistical quantity for quantifying the flow of information between systems, reflecting the strength and direction of causal relationships 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 between agents, commonly used in distributed control, robot collaboration and other fields. The return vector is a vector used to describe the returns obtained by each agent or participant under a specific decision, commonly used in game theory and decision analysis, reflecting the results and effects of decision-making. The abnormal propagation path refers to the path through which an abnormal event propagates from one node to other nodes in a network or system, widely used in fault diagnosis and anomaly detection systems.

[0128] Water level data is collected from each monitoring site, and the collection period can be pre-set, for example, once an hour. After obtaining the water level data, the wavelet decomposition technique is used to decompose the water level data into periodic components and trend components, and the periodic feature vector and the trend feature vector are constructed respectively. For example, the water level fluctuation within a day is constructed into a periodic feature vector, and the overall rising or falling trend of the water level in a period 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 filtering 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 observation data, the likelihood probability value of each particle sample is calculated as the weight of the particle. The particle with higher weight indicates that the corresponding state variable value is more consistent with the observation data. The particles are resampled, and the particles with high weight are replaced by the particles with low weight to obtain the state estimation sequence.

[0129] Based on the obtained state estimation sequence, a causal discovery network is constructed. The nodes in the network represent different monitoring sites, and the connections between the nodes represent the possible causal relationships between the sites. The causal relationship matrix is constructed by calculating the transfer entropy value between the nodes to measure the strength of the causal relationship. The higher the transfer entropy value, the stronger the causal relationship. The causal relationship matrix is divided into multiple regions according to the spatial position of the monitoring points, and is decomposed into multiple sub-matrices, for example, all monitoring sites are divided into three regions of upstream, midstream and downstream, and the causal relationship sub-matrices within the region and between the regions are obtained respectively.

[0130] A hierarchical attention network is constructed. The network is divided into three layers, corresponding to different spatial scales. The first layer of attention network inputs represents the sub-matrix of the causal relationship within the region, for example, the causal relationship sub-matrix within the upstream region. The second layer of attention network inputs represents the sub-matrix of the causal relationship between adjacent regions, for example, the causal relationship sub-matrix between the upstream and midstream regions. The third layer of attention network inputs represents the sub-matrix of the causal relationship between remote regions, for example, the causal relationship sub-matrix between the upstream and downstream regions. The network weights and fuses the features at different spatial scales through the attention mechanism 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 midstream region, the network will give a higher weight to the causal relationship sub-matrix between the upstream and midstream regions.

[0131] Each monitoring site is taken as a participant of federated learning. Each participant calculates the gradient of the model parameters using the locally collected water level data, and encrypts the gradient using a 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, and then distributes the update 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 obtains the optimal sampling strategy vector using a collaborative decision-making algorithm. For example, if the water level feature tensor of a certain area indicates that the area has a high probability of abnormality, the agent corresponding to the area will increase the sampling frequency. At the same time, the collected water level data is analyzed for abnormalities, and time series features and spatial distribution features are extracted to obtain an abnormal feature vector. The optimal sampling strategy vector and the abnormal feature vector are input into a hierarchical reinforcement learning network. The network includes a state evaluation subnetwork, an action generation subnetwork, and a value estimation subnetwork. The state evaluation subnetwork outputs a state vector, the action generation subnetwork outputs a detection resource allocation scheme matrix, and the value estimation subnetwork predicts the long-term benefits of the detection resource allocation scheme. Through policy optimization, an execution parameter matrix is obtained.

[0132] Based on the execution parameter matrix and the causal relationship matrix, the abnormal propagation path is analyzed. The location of the abnormal event is located based on the execution parameter matrix, the abnormal propagation link is traced back through the causal relationship matrix, the abnormal influence range is analyzed, and an abnormal propagation path diagram is drawn. The key monitoring area is determined and the sampling frequency is increased, and an emergency monitoring parameter sequence containing the monitoring area and the monitoring frequency is generated.

[0133] In this embodiment, through multi-agent collaborative decision-making and hierarchical reinforcement learning, the sampling strategy and detection resource allocation can be optimized, the abnormality detection efficiency can be improved, the manpower and material resources cost can be reduced, the water level state can be more accurately estimated by combining the nonlinear state space model and the particle filter algorithm, thereby improving the accuracy of abnormality detection, and the origin and propagation process of the abnormal event can be traced back by constructing a causal discovery network and analyzing the abnormal propagation path, thereby providing a scientific basis for emergency disposal.

[0134] In an alternative embodiment, 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 influence range and drawing an abnormal propagation path diagram, determining the key monitoring area and increasing the sampling frequency, and generating an emergency monitoring parameter sequence containing the monitoring area and the monitoring frequency include:

[0135] Based on the execution parameter matrix and the causal relationship matrix, the water level data is segmented and processed 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 the conditional probability and accumulating, and the transfer entropy value is filled into the corresponding position of the causal relationship matrix;

[0136] An execution parameter matrix is constructed, 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 mean deviation, variance ratio, trend slope, and fluctuation period obtained based on historical data statistical analysis, multi-level abnormality discrimination rules are established based on the execution parameters, and real-time data of the monitoring points are evaluated for abnormality degree in combination with a fuzzy inference mechanism;

[0137] A backtracking analysis is performed in the causal relationship matrix using a depth-first search strategy, a search direction is determined based on a causal relationship strength threshold value starting from an abnormal point, neighboring nodes with the strongest causal relationship are preferentially accessed, an access mark table is maintained to record the accessed nodes, and a complete propagation link from an abnormal source to an impact end is obtained through iterative search;

[0138] Based on the complete propagation link, an abnormal propagation path is analyzed through a hierarchical processing strategy, a local propagation feature is obtained by analyzing the correlation between the abnormal point and the surrounding monitoring points in a local area, the analysis range is expanded to adjacent areas to study the propagation mode at a larger spatial scale, and an abnormal propagation network structure is obtained by comprehensively analyzing the results of each level on a global scale;

[0139] According to the abnormal propagation network structure, a self-adaptive layout algorithm is used to arrange the monitoring points and the propagation path in space, 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 smooth transition, the size and color of the arrow are dynamically adjusted according to the propagation strength, a clustering analysis method is used to determine the key monitoring area, the nodes in the abnormal propagation network are clustered and identified according to the position and abnormal degree to identify the abnormal concentration area, the importance score of each clustering area is calculated, the importance score considers the abnormal degree, the influence range, and the propagation speed, the monitoring area is divided into levels according to the importance score, and the monitoring frequency is determined;

[0140] A dynamic programming method is used to generate an emergency monitoring parameter sequence, a monitoring resource allocation optimization model is established with the maximum abnormal monitoring effect as the objective function, the number of monitoring devices and the communication bandwidth are used as resource constraints, an optimal monitoring parameter configuration scheme is solved, the monitoring parameters are adjusted online based on the abnormal development trend, and an emergency monitoring parameter sequence including the monitoring area and the monitoring frequency is obtained.

[0141] The multi-level anomaly discrimination rule is a rule for determining anomalies in a multi-level structure, usually by analyzing features at different levels and determining anomalies layer by layer, commonly used in anomaly detection of complex systems and multi-dimensional data analysis. Backtracking analysis is a reverse analysis method that traces back from the result to the process and reason, widely used in fault diagnosis, problem solving and data analysis. The base map refers to a map or image used as a base in geographic information systems or other graphical analysis, used 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, widely used in computer graphics, animation production and path planning.

[0142] Construct a causal relationship matrix. Select several monitoring points, such as A, B, C, and D, and collect historical water level data. Set a sliding time window, and for each pair of monitoring points, calculate the transfer entropy value of the water level data sequence in each time period. The transfer entropy value measures the degree of influence of A point water level change on B point water level change. The method of calculating the transfer entropy value is to calculate the probability of various combinations of A and B water level changes, and the conditional probability of B water level change given A water level change, and to calculate the transfer entropy value based on the probability value. Fill the calculated transfer entropy value into the causal relationship matrix cell corresponding to the A row and B column to represent the causal relationship strength of A to B.

[0143] Construct an execution parameter matrix. The rows of the matrix correspond to the monitoring points, and the columns correspond to the execution parameters. The execution parameters include mean deviation, variance ratio, trend slope and fluctuation period based on historical data statistical analysis. Similarly, calculate the variance ratio, trend slope and fluctuation period parameters. Fill these parameters into the corresponding cells of the execution parameter matrix.

[0144] Based on the execution parameter matrix, multi-level abnormality discrimination rules are established, and the real-time data of the monitoring points are evaluated for abnormality degree in combination with the fuzzy reasoning mechanism. For example, if the mean deviation exceeds a certain threshold, it is judged to be a mild abnormality, and if it exceeds a higher threshold, it is judged to be a moderate abnormality, and so on. The overall abnormality degree of the monitoring points is comprehensively evaluated by using the fuzzy reasoning mechanism in combination with the abnormality degrees of multiple parameters. For example, if the mean deviation and the variance ratio of point A both exceed the threshold of mild abnormality, point A is finally judged to be a mild abnormality. If an abnormal point, such as point A, is detected, a backtracking analysis is performed in the causal relationship matrix using a depth-first search strategy to determine the abnormality propagation path. Starting from the abnormal point A, the search direction is determined according to the causal relationship strength threshold. The neighboring nodes with strong causal relationship with point A are preferentially accessed. An access mark table is maintained to record the nodes that have been accessed to avoid repeated access. Through iterative search, the complete propagation link from the abnormal source to the impact end is obtained. According to the complete propagation link obtained, the abnormality propagation path is analyzed, the correlation between the abnormal point and the surrounding monitoring points in the local area is analyzed, and finally the abnormality propagation network structure is obtained by comprehensively analyzing the results of each level on a global scale.

[0145] Based on the abnormality propagation network structure, an abnormality propagation path diagram is drawn. An adaptive layout algorithm is used to arrange the monitoring points in space and project them onto a geographic information base map. A Bézier curve is used to draw the propagation path, and the arrow size and color are dynamically adjusted according to the propagation strength. The nodes in the abnormality propagation network are clustered according to their location and abnormality degree, and the abnormality concentration areas are identified. The importance score of each cluster area is calculated, considering the abnormality degree, the influence range and the propagation speed. For example, if the abnormality degree in a cluster area is high, the influence range is wide, and the propagation speed is fast, the importance score of the cluster area is also high. The monitoring area is divided into levels according to the importance score and the monitoring frequency is determined. For example, the area with high importance score is divided into a first-level monitoring area, and the monitoring frequency is set to every 10 minutes; the area with low importance score is divided into a second-level monitoring area, and the monitoring frequency is set to every hour. A monitoring resource allocation optimization model is established with the objective of maximizing the abnormality monitoring effect, and the number of monitoring devices and the communication bandwidth are used as resource constraints. The optimal monitoring parameter configuration scheme is solved, and the monitoring parameters are adjusted online according to the abnormality development trend. Finally, the emergency monitoring parameter sequence containing the monitoring area and the monitoring frequency is obtained. For example, if the abnormality degree in a certain area intensifies, the monitoring frequency of the area is dynamically increased.

[0146] In this embodiment, through the multi-level abnormality discrimination rule and the fuzzy reasoning mechanism, the abnormal event can be more sensitively identified, the false alarm and the false alarm are avoided, the discovery efficiency of the abnormal event is improved, through the causal relationship matrix and the depth first search strategy, the abnormal propagation link can be accurately traced back, the abnormal source is determined, and the influence range of the abnormality is evaluated, a scientific basis is provided for emergency disposal, through the dynamic programming method, the emergency monitoring parameter sequence is generated, the monitoring area and the monitoring frequency can be dynamically adjusted according to the abnormal development trend, the monitoring resource configuration is optimized, the monitoring efficiency is improved, and the monitoring cost is saved.

[0147] Figure 2 The structure diagram of the water level distributed monitoring system based on the visual sensing network in the embodiment of the application is shown as Figure 2 The system comprises:

[0148] The first unit is used for collecting water level image samples of a monitoring area, constructing a spectrum decomposition network to extract frequency domain features of the water level image samples, constructing a multi-scale spectrum feature matrix, constructing an energy matrix based on the multi-scale spectrum feature matrix and calculating an energy distribution density, generating an energy mask graph corresponding to the water level image samples, guiding the attention mechanism to perform self-supervised decomposition on the image through the energy mask graph, obtaining a water level scale feature tensor, a water surface texture feature tensor and an environmental background feature tensor, inputting the feature tensors obtained by decomposition into a multi-modal contrast learning network, calculating a discriminative feature vector group through mutual information maximization, constructing a random field model based on the spatial distribution characteristics of the visual sensing nodes, calculating the spatial dependence relationship between the nodes to obtain a node correlation probability matrix, inputting the node correlation 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;

[0149] The second unit is used 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 decoupling representation network, obtaining a water level feature graph through separation and calculating a complementarity measure value based on the water level feature graph, obtaining an enhanced feature graph through adaptive fusion based on the complementarity measure value, inputting the enhanced feature graph into a densely connected network, obtaining a water level region segmentation graph and a water level measurement value through cross-layer feature reuse, calculating a posterior probability distribution of the water level measurement according to a Bayesian inference framework, combining prior knowledge and observation data to obtain a reliability score table based on the posterior probability distribution, calculating a measurement uncertainty value, screening a high-confidence water level dataset and performing spectral analysis to extract a periodic feature vector and a trend feature vector;

[0150] The third unit is configured to construct a nonlinear state space model based on the periodic feature vector and the trend feature vector, calculate a state estimation sequence by combining a particle filter, construct a causal discovery network, calculate a causal relationship matrix of water level change based on the causal discovery network, decompose the causal relationship matrix to obtain a plurality of sub-matrices, construct a hierarchical attention network on different spatial scales, and execute to obtain a water level feature tensor, obtain an encrypted model parameter update amount by combining a homomorphic encryption algorithm through a federated optimization objective function, calculate an optimal sampling strategy vector by a pre-set multi-agent collaborative decision system, and perform 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 to obtain a detection resource allocation scheme and an execution parameter matrix, and analyze the causal relationship matrix to obtain an abnormal propagation path graph and an emergency monitoring parameter sequence.

[0151] The present application can be a method, apparatus, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein for executing various aspects of the present application.

[0152] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part 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 application.

Claims

1. A water level distributed monitoring method based on a visual sensing network, characterized in that, The method comprises the following steps: Collecting water level image samples of a monitoring area and constructing a spectral decomposition network to extract 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 an energy distribution density, generating an energy mask graph corresponding to the water level image samples, guiding an attention mechanism to perform self-supervised decomposition on the image through the energy mask graph, obtaining a water level scale feature tensor, a water surface texture feature tensor and an environmental background feature tensor, inputting the feature tensors obtained by decomposition into a multi-modal contrast learning network, calculating a discriminative feature vector group through mutual information maximization, constructing a random field model based on the spatial distribution characteristics of the visual sensing nodes, calculating the spatial dependence between nodes to obtain a node correlation probability matrix, inputting the node correlation 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; 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 decoupling representation network, obtaining a water level feature graph through separation and calculating a complementarity measure value based on the water level feature graph, performing adaptive fusion based on the complementarity measure value to obtain an enhanced feature graph, inputting the enhanced feature graph into a dense connection network, obtaining a water level region segmentation graph and a water level measurement value through cross-layer feature reuse, calculating a posterior probability distribution of the water level measurement according to a Bayesian inference framework, combining prior knowledge and observation data to obtain a reliability score table based on the posterior probability distribution, filtering to obtain a high-confidence water level dataset and performing spectral analysis to extract a periodic feature vector and a trend feature vector; Based on the periodic feature vector and the trend feature vector, a nonlinear state space model is constructed, a state estimation sequence is calculated by combining particle filtering, and a causal discovery network is constructed, a causal relationship matrix of water level changes is calculated based on the causal discovery network, and a plurality of sub-matrices are decomposed, a hierarchical attention network is constructed on different spatial scales, and a water level feature tensor is obtained, an encryption model parameter update amount is obtained by combining a homomorphic encryption algorithm through a federated optimization objective function, an optimal sampling strategy vector is calculated by a pre-set multi-agent collaborative decision 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 scheme and an execution parameter matrix, and an abnormal propagation path graph and an emergency monitoring parameter sequence are analyzed based on the causal relationship matrix.

2. The method of claim 1, wherein, The water level image samples of the monitoring area are collected, and a spectrum decomposition network is constructed to extract frequency domain features of the water level image samples, a multi-scale spectrum feature matrix is constructed, an energy matrix is constructed based on the multi-scale spectrum feature matrix, and an energy distribution density is calculated, an energy mask graph corresponding to the water level image samples is generated, and a self-supervised decomposition is performed on the image by using the energy mask graph to guide the attention mechanism, 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 the decomposition are input into a multi-modal contrast learning network, a discriminative feature vector group is calculated by maximizing mutual information, a random field model is constructed based on the spatial distribution characteristics of the visual sensing nodes, a spatial dependence relationship between the nodes is calculated to obtain a node correlation probability matrix, and the node correlation probability matrix and the discriminative feature vector group are input into a graph neural network to obtain a parameter matrix of a water level propagation prediction model by message passing iteration, including: The water level image samples of the monitoring area are collected, and the water level image samples are input into a multi-level spectrum decomposition network, wherein a shallow network in the multi-level spectrum decomposition network extracts high-frequency components of the water level image samples by using a high-pass filter set to obtain water level scale line features and water surface ripple features, a middle network extracts medium-frequency components of the water level image samples by using a band-pass filter set to obtain water level line contour features and shore structure features, and a deep network extracts low-frequency components of the water level image samples by using a low-pass filter set to obtain water area distribution features; the feature maps corresponding to the high-frequency components, the medium-frequency components, and the low-frequency components are normalized, and are arranged and combined into a multi-scale spectrum feature matrix in the order of frequency bands; Energy values of each frequency band component in the multi-scale spectrum feature matrix are calculated to obtain a two-dimensional energy matrix, the two-dimensional energy matrix is subjected to density estimation by using a kernel function to obtain an energy density distribution graph, an adaptive threshold is set based on the energy density distribution graph, and a binary energy mask graph is generated, the high-energy region is marked as a foreground region, and the low-energy region is marked as a background region; The water level image samples are divided into a plurality of overlapping image blocks, attention weights of the plurality of overlapping image blocks are calculated based on the binary energy mask graph, the overlapping image blocks with the attention weights are input into a self-supervised deconstruction module for feature decomposition, water level scale feature tensors are obtained by using a fine-grained feature extractor to extract a water level scale feature branch, water surface texture feature tensors are obtained by using a texture analyzer to extract a water surface texture feature branch, and environmental background feature tensors are obtained by using a context encoder to extract an environmental background feature branch, the water level scale feature tensors, the water surface texture feature tensors, and the environmental background feature tensors are respectively input into feature encoders for dimension reduction to obtain feature vectors of a unified dimension, and a feature contrast pool is constructed; randomly sampling positive sample pairs and negative sample pairs from the uniform-dimension feature vectors, calculating mutual information metric values between the positive sample pairs and the negative sample pairs, obtaining 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, calculating a distance matrix between node pairs based on geographical coordinates of visual sensing nodes, converting the distance matrix into initial correlation strengths, and obtaining a node correlation probability matrix by calculating interactions between node pairs through a potential function of a random field; inputting the node correlation probability matrix and the discriminative feature vector group into a graph neural network, determining importance weights of neighbor nodes based on the node correlation probability matrix in each round of message passing iteration, performing nonlinear transformation on feature information of the neighbor nodes to obtain fused features, combining the fused features with feature information of the nodes themselves to update node states, and obtaining a parameter matrix of the water level propagation prediction model through multiple rounds of iteration.

3. The method of claim 2, wherein, randomly sampling positive sample pairs and negative sample pairs from the uniform-dimension feature vectors, calculating mutual information metric values between the positive sample pairs and the negative sample pairs, obtaining 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, calculating a distance matrix between node pairs based on geographical coordinates of visual sensing nodes, converting the distance matrix into initial correlation strengths, and obtaining a node correlation probability matrix by calculating interactions between node pairs through a potential function of a random field, including: constructing a feature contrast pool from uniform-dimension feature vectors and dividing the feature contrast pool into multiple time windows, extracting scene feature information in each time window and randomly sampling positive sample pairs, and randomly sampling feature vectors to construct negative sample pairs according to weather conditions, monitoring time periods, and water level change trends in different monitoring scenes through a multi-layer hierarchical sampling strategy; adding feature vectors in the positive sample pairs and the negative sample pairs to a multilayer perceptron including an input layer, multiple hidden layers, and an output layer to perform feature transformation, setting a batch normalization layer and a nonlinear activation function between adjacent layers of the multilayer perceptron, constructing a Gaussian kernel function group with adaptive bandwidth in a probability distribution space after mapping feature vectors from a feature space to the probability distribution space, and obtaining a continuous probability distribution by accumulating local and global contributions of probability densities of all feature point pairs; calculating mutual information metric values between the positive sample pairs and the negative sample pairs, wherein the mutual information metric values are obtained by calculating multiple differences of joint distributions, conditional distributions, and marginal distributions of transformed features, constructing a mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints, obtaining a feature vector group with local and global discriminativeness by maximizing mutual information metric values of the positive sample pairs and minimizing mutual information metric values of the negative sample pairs through a gradient descent method with an adaptive learning rate, and performing multiple rounds of iterative optimization; and calculating mutual information metric values between the positive sample pairs and the negative sample pairs, wherein the mutual information metric values are obtained by calculating multiple differences of joint distributions, conditional distributions, and marginal distributions of transformed features, constructing a mutual information optimization objective function including feature discriminative constraints and distribution consistency constraints, obtaining a feature vector group with local and global discriminativeness by maximizing mutual information metric values of the positive sample pairs and minimizing mutual information metric values of the negative sample pairs through a gradient descent method with an adaptive learning rate, and performing multiple rounds of iterative optimization. Based on the geographic coordinates of the visual sensing nodes, a distance matrix between node pairs is calculated by fusing geodetic survey distances between nodes, terrain relief coefficients, river connectivity coefficients, and hydrological propagation characteristic coefficients, the distance matrix is dynamically corrected by calculating multi-scale elevation differences between nodes in combination with digital elevation model data, and actual flow distances considering water flow direction and flow rate are calculated based on river connectivity data, the actual flow distances are multi-featured fused with the corrected distance matrix to obtain a corrected distance matrix; Based on the corrected distance matrix, distance values are mapped to a dynamic interval by an adaptive normalization method, local and global densities are calculated based on multi-dimensional spatial distribution density characteristics of nodes, an adaptive attenuation coefficient based on density gradient is set for densely distributed areas, and a dynamic attenuation coefficient considering spatial heterogeneity is set for sparsely distributed areas, and initial correlation strengths under multi-scale are calculated in combination with spatio-temporal correlation characteristics of nodes; Based on the initial correlation strengths, a random field potential function with an adaptive segmented structure is constructed, node distances are dynamically divided into near-distance intervals, middle-distance intervals, and far-distance intervals with transition zones based on multi-dimensional spatial relationships between nodes, a linear attenuation function considering local spatial dependence is designed in the near-distance intervals, an exponential attenuation function considering regional influence is designed in the middle-distance intervals, and a residual function considering global correlation is designed in the far-distance intervals, and 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, state information of multi-scale neighborhood nodes is collected based on spatial dependence relationships, messages are multi-level weighted using adaptive potential function values, spatio-temporal redundant information is identified and suppressed through a structured attention mechanism, weighted messages are nonlinearly combined and added with random noise subject to dynamic distribution, and a node correlation probability matrix representing the correlation relationship between nodes is obtained through state updating.

4. The method of claim 1, wherein, Real-time water level images are received, parameter matrices of the water level propagation prediction model are input into an adaptive wavelet transform network, optimal decomposition scale sequences are calculated and multi-resolution feature representations are generated, compressed feature code streams are obtained through sparse coding of the multi-resolution feature representations, the feature code streams are input into a decoupling representation network, water level feature maps are separated and obtained, and complementarity measure values are calculated based on the water level feature maps, enhanced feature maps are obtained through adaptive fusion based on the complementarity measure values, the enhanced feature maps are input into a dense connection network, water level region segmentation maps and water level measurement values are obtained through cross-layer feature reuse, posterior probability distributions of water level measurements are calculated based on Bayesian inference framework in combination with prior knowledge and observation data, measurement uncertainty values are calculated based on the posterior probability distributions to generate a reliability score table, a high-confidence water level dataset is screened, and periodic feature vectors and trend feature vectors are extracted, including: Receiving real-time water level images, inputting the parameter matrix of the water level propagation prediction model into an adaptive wavelet transform network, calculating the gradient of the image in multiple directions to construct a direction gradient histogram, calculating a gray level co-occurrence matrix on different direction and distance combinations to extract multi-dimensional texture features including energy, entropy, contrast and correlation, selecting the optimal wavelet basis function through adaptive weight fusion according to the multi-dimensional texture features, calculating the optimal decomposition scale sequence by analyzing the local extreme value characteristics and energy distribution of the wavelet coefficients, and performing multi-layer decomposition based on the optimal decomposition scale sequence through row and column transformation and downsampling operation to obtain low-frequency approximation components and high-frequency detail components, and performing tree-shaped decomposition on the high-frequency detail components to generate multi-resolution feature representations; Sparse coding is performed on the multi-resolution feature representations by constructing an over-complete dictionary containing redundant atoms, the features are blocked according to a fixed size, sparse decomposition based on orthogonal matching pursuit is performed on each feature block, the dictionary atom most relevant to the residual is iteratively selected and the reconstruction coefficient is updated until the sparsity constraint is met to obtain a compressed feature code stream, and the compressed feature code stream is input into a decoupling representation network, multi-layer features are extracted through a series of residual calculation units, wherein the residual calculation units include convolution layers, normalization layers and activation functions, different scale context information is extracted through a multi-scale pooling module, the feature resolution is gradually restored through deconvolution operation, and channel weights are calculated based on feature statistics and nonlinear transformation to separate and obtain a water level feature map; Based on the water level feature map, a local region structure similarity is calculated using a Gaussian weighted window, horizontal and vertical gradients are calculated to construct a gradient direction field and analyze the direction consistency, a normalized cross power spectrum is calculated by Fourier transform of the features, a complementarity measure value is obtained by local maximum value detection to determine the best matching position, a feature fusion relationship graph is constructed based on the complementarity measure value, the nodes in the feature fusion relationship graph represent features and the weights of the edges are determined by the complementarity measure value, feature correlation scores are calculated and normalized through a multi-layer graph attention calculation unit, and an enhanced feature map is obtained by adaptively fusing the normalized attention scores and feature weighted combination; The enhanced feature map is input into a dense connection network, a standard convolution is decomposed into a depth convolution and a point-wise convolution through a depth separable convolution to reduce the calculation amount, a dense connection is used to make the nodes of the later layers simultaneously receive the features of all previous layers to perform layer feature reuse, and a water level region segmentation map and a water level measurement value are obtained by adaptively learning the importance weight of the feature channel through a compression excitation mechanism; According to a Bayesian inference framework, in a three-layer probability model including an observation layer, a hidden variable layer and a priori layer, the observation data and physical constraints are input as priori knowledge, the parameters of the approximate posterior distribution are calculated by iteratively optimizing the approximate posterior distribution through a variational inference method to obtain the posterior probability distribution of the water level measurement, and based on the posterior probability distribution, the prediction mean and variance are calculated by randomly inactivating part of the neurons and sampling multiple times in the network forward calculation to estimate the measurement uncertainty value and generate a reliability score table. The high-confidence water level data set obtained by screening is subjected to spectrum analysis, time-frequency analysis is realized by continuous wavelet transform, the time series is decomposed and reconstructed by using spectrum decomposition method, a plurality of intrinsic characteristic functions are calculated by using an adaptive signal decomposition algorithm based on an extreme point envelope, and a periodic characteristic vector and a trend characteristic vector are extracted by using amplitude threshold control decomposition process.

5. The method of claim 4, wherein, The enhanced feature map is input into a dense connection network, the standard convolution is decomposed into a depth convolution and a point-by-point convolution by depth separation convolution to reduce the calculation amount, the dense connection is used to make the nodes of the rear layer simultaneously receive the features of all the previous layers to perform layer feature reuse, and the water level region segmentation map and the water level measurement value are obtained by adaptively learning the importance weight of the feature channel through the compression excitation mechanism. The enhanced feature map is subjected to feature extraction by depth separation convolution, the standard convolution is decomposed into a depth convolution and a point-by-point convolution, the depth convolution extracts spatial correlation features by independently performing convolution operation on each input channel, and the point-by-point convolution obtains reorganized features by combining channel information of 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, a plurality of depth separation convolution layers are arranged in each dense connection block, the first layer of the first dense connection block extracts features from the reorganized features to obtain first layer output features, the second layer extracts features from the reorganized features and the first layer output features after concatenation to obtain second layer output features, and the third layer extracts features from the reorganized features, the first layer output features and the second layer output features after concatenation to obtain third layer output features; The features between adjacent dense connection blocks are reduced in spatial size and increased in channel number by convolution operation and feature sampling operation to obtain multi-scale features, the channel description vector is obtained by globally statistically analyzing the feature maps output by the dense connection blocks, the compressed feature vector is obtained by inputting the channel description vector into the first full connection layer, the excitation feature vector is obtained by inputting the compressed feature vector into the second full connection layer, the channel weight coefficient is obtained by normalizing the excitation feature vector, and the weighted features are obtained by multiplying the channel weight coefficient with the multi-scale 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 features of the third dense connection block are up-sampled and fused with the weighted features of the second dense connection block to obtain first fusion features, the first fusion features are up-sampled and fused with the weighted features of the first dense connection block to obtain a fusion feature map, and the fusion feature map is input into a segmentation branch and a measurement branch, respectively, the water level region segmentation map with the same size as the enhanced feature map is output by the multi-layer convolution operation of the segmentation branch, and the water level measurement value is output by the feature pooling operation and full connection operation of the measurement branch.

6. The method of claim 1, wherein, constructing a nonlinear state space model based on the periodic feature vector and the trend feature vector, combining particle filtering to obtain a state estimation sequence and constructing a causal discovery network, calculating a causal relationship matrix of water level changes based on the causal discovery network and decomposing to obtain a plurality of sub-matrices, constructing a hierarchical attention network on different spatial scales and executing to obtain a water level feature tensor, obtaining an encrypted model parameter update amount by combining a homomorphic encryption algorithm through a pre-set multi-agent collaborative decision system, calculating an optimal sampling strategy vector and performing abnormal analysis on the water level to obtain an abnormal feature vector, adding the optimal sampling strategy vector and the abnormal feature vector to a hierarchical reinforcement learning network to obtain a detection resource allocation scheme and an execution parameter matrix, and analyzing the abnormal propagation path graph and the emergency monitoring parameter sequence based on the causal relationship matrix include: collecting water level data at each monitoring site according to a pre-set sampling period, performing wavelet decomposition on the water level data to obtain periodic components and trend components, constructing the periodic components into a periodic feature vector, and constructing the trend components into a trend feature vector; constructing a nonlinear state space model by taking the periodic feature vector as an observation variable and the trend feature vector as a state variable, generating a pre-set number of particle samples for each state variable, and combining particle filtering to iteratively calculate the particle samples through importance sampling, calculating the likelihood probability value of each particle sample as a particle weight based on observation data, and resampling the particle samples to obtain a state estimation sequence; constructing 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, and dividing the causal relationship matrix into a plurality of regions and decomposing it into a plurality of sub-matrices according to the spatial positions of the monitoring points; constructing a hierarchical attention network on different spatial scales, the first layer attention network inputting a sub-matrix representing the causal relationship within a region, the second layer attention network inputting a sub-matrix representing the causal relationship between adjacent regions, and the third layer attention network inputting a sub-matrix representing the causal relationship between remote regions, and obtaining a water level feature tensor by weighting and fusing features at different spatial scales through attention weights; designing a federated optimization objective function, taking each monitoring site as a participant in federated learning, using a homomorphic encryption algorithm to encrypt model parameters, using local sampling data to calculate model parameter gradients and encrypting them by each participant, aggregating encrypted model parameter gradients on a server to obtain encrypted model parameter updates, and obtaining global model parameters through pre-set iteration optimization rounds; An optimal sampling strategy is calculated by a pre-configured multi-agent collaborative decision system, wherein the multi-agent collaborative decision system comprises a plurality of agents, the agents correspond to a plurality of monitoring areas respectively, the agents interact with each other through a communication network to exchange water level feature tensor information, a state value function is calculated based on the water level feature tensor, and an optimal sampling strategy vector is obtained by using a collaborative decision algorithm, abnormal analysis is performed on collected water level data, time series features and spatial distribution features are extracted to obtain an abnormal feature vector; The optimal sampling strategy vector and the abnormal feature vector are added to a hierarchical reinforcement learning network, a state evaluation subnetwork in the hierarchical reinforcement learning network outputs a state vector, an action generation subnetwork outputs a detection resource allocation scheme matrix, a value estimation subnetwork predicts long-term returns of the detection resource allocation scheme to obtain a return vector, and an execution parameter matrix is obtained through policy optimization; Based on the execution parameter matrix and the causal relationship matrix, an abnormal propagation path is analyzed, an abnormal event location is located based on the execution parameter matrix, an abnormal propagation link is traced back through the causal relationship matrix, an abnormal influence range is analyzed and an abnormal propagation path graph is drawn, a key monitoring area is determined and a sampling frequency is increased, and an emergency monitoring parameter sequence containing a monitoring area and a monitoring frequency is generated.

7. The method of claim 6, wherein, Based on the execution parameter matrix and the causal relationship matrix, an abnormal propagation path is analyzed, an abnormal event location is located based on the execution parameter matrix, an abnormal propagation link is traced back through the causal relationship matrix, an abnormal influence range is analyzed and an abnormal propagation path graph is drawn, a key monitoring area is determined and a sampling frequency is increased, and an emergency monitoring parameter sequence containing a monitoring area and a monitoring frequency is generated, which comprises: Based on the execution parameter matrix and the causal relationship matrix, water level data is segmented by using a sliding time window, and a transfer entropy value of a water level data sequence between each pair of monitoring points is calculated as a causal relationship strength, wherein the transfer entropy value is calculated by calculating a conditional probability and accumulating, and the transfer entropy value is filled into a corresponding position of the causal relationship matrix; An execution parameter matrix is constructed, a row vector of the execution parameter matrix corresponds to a monitoring point, and a column vector corresponds to an execution parameter, the execution parameter includes a mean deviation, a variance ratio, a trend slope and a fluctuation period obtained based on historical data statistical analysis, a multi-level abnormality discrimination rule is established based on the execution parameter, and a fuzzy inference mechanism is combined to evaluate the abnormality degree of real-time data of the monitoring point; A depth-first search strategy is used for backtracking analysis in the causal relationship matrix, a search direction is determined based on a causal relationship strength threshold from an abnormal point, neighboring nodes with the strongest causal relationship are preferentially accessed, an access mark table is maintained to record accessed nodes, and a complete propagation link from an abnormal source to an impact end is obtained through iterative search. Based on the complete propagation link, the abnormal propagation path is analyzed by a hierarchical processing strategy, the correlation between the abnormal point and the surrounding monitoring point is analyzed in the local area to obtain the local propagation characteristics, the analysis range is expanded to the adjacent area to study the propagation mode at a larger spatial scale, and the abnormal propagation network structure is obtained by comprehensively analyzing the results of each level at a global scale; According to the abnormal propagation network structure, the spatial arrangement of the monitoring points and the propagation path is performed by an adaptive layout algorithm, a base map is constructed based on geographic information data, and the monitoring points are projected onto the base map, the propagation path is drawn by using a Bezier curve to achieve smooth transition, the size and color of the arrow are dynamically adjusted according to the propagation intensity, the key monitoring area is determined by combining a clustering analysis method, the nodes in the abnormal propagation network are clustered according to the position and abnormal degree to identify the abnormal concentrated area, the importance score of each clustering area is calculated, the abnormal degree, influence range and propagation speed are considered in the importance score, the monitoring area is divided into grades according to the importance score, and the monitoring frequency is determined; A dynamic programming method is used to generate an emergency monitoring parameter sequence, a monitoring resource allocation optimization model is established with the maximum abnormal monitoring effect as the objective function, the number of monitoring devices and the communication bandwidth are used as resource constraints, the optimal monitoring parameter configuration scheme is solved, the monitoring parameters are adjusted online based on the abnormal development trend, and the emergency monitoring parameter sequence including the monitoring area and the monitoring frequency is obtained.

8. A water level distributed monitoring system based on a visual sensor network for implementing the method according to any one of the preceding claims 1-7, characterized in that, Comprise: The first unit is used for collecting water level image samples of a monitoring area and constructing a wave spectrum decomposition network to extract frequency domain features of the water level image samples, constructing a multi-scale frequency spectrum feature matrix, constructing an energy matrix based on the multi-scale frequency spectrum feature matrix and calculating an energy distribution density, generating an energy mask image corresponding to the water level image samples, guiding the attention mechanism to perform self-supervised decomposition on the image through the energy mask image, obtaining a water level scale feature tensor, a water surface texture feature tensor and an environment background feature tensor, inputting the feature tensors obtained by decomposition into a multi-modal contrast learning network, calculating a discriminative feature vector group through mutual information maximization, constructing a random field model based on the spatial distribution characteristics of the visual sensing nodes, calculating the spatial dependence relationship between the nodes to obtain a node correlation probability matrix, inputting the node correlation probability matrix and the discriminative feature vector group into a graph neural network, and obtaining a parameter matrix of the water level propagation prediction model through message passing iteration; The second unit is configured to receive a real-time water level image, input a 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 by sparse coding the multi-resolution feature representation, input the feature code stream into a decoupling representation network, obtain a water level feature map by separation and calculate a complementarity measure value based on the water level feature map, perform adaptive fusion based on the complementarity measure value to obtain an enhanced feature map, input the enhanced feature map into a dense connection network, obtain a water level region segmentation map and a water level measurement value by cross-layer feature reuse, calculate a posterior probability distribution of the water level measurement according to a Bayesian inference framework combined with prior knowledge and observation data, calculate a measurement uncertainty value based on the posterior probability distribution to generate a reliability score table, screen a high-confidence water level dataset and perform spectral analysis to extract a periodic feature vector and a trend feature vector; The third unit is configured to construct a nonlinear state space model based on the periodic feature vector and the trend feature vector, combine particle filtering to calculate a state estimation sequence and construct a causal discovery network, calculate a causal relationship matrix of the water level change based on the causal discovery network and decompose the causal relationship matrix to obtain a plurality of sub-matrices, construct a hierarchical attention network on different spatial scales and execute to obtain a water level feature tensor, combine a homomorphic encryption algorithm to obtain an encrypted model parameter update amount by a federated optimization objective function, calculate an optimal sampling strategy vector by a pre-set multi-agent collaborative decision system and perform 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 to obtain a detection resource allocation scheme and an execution parameter matrix, and analyze an abnormal propagation path map and an emergency monitoring parameter sequence based on the causal relationship matrix.

Images