Multimodal data-driven risk prediction method for sudden water pollution events

Abstract

This invention provides a multimodal data-driven method for predicting the risk of sudden water pollution events, comprising: acquiring multimodal data related to sudden water pollution events in a target area, and obtaining multimodal feature vectors by encoding the multimodal data; constructing an end-to-end prediction model, wherein the prediction model includes a bidirectional constraint fusion interpretation layer and a prediction output layer; the bidirectional constraint fusion interpretation layer incorporates a coupled modal weight network and a cross-modal attention network, simultaneously performing feature fusion and interpretation generation; combining the fused features with the enhanced features to generate a comprehensive representation vector, and inputting it into the prediction output layer to obtain the risk prediction result; when predicting the risk of a target area, simultaneously acquiring the risk prediction result, the modal level contribution interpretation, and the fine-grained contribution interpretation from the forward propagation process of the prediction model. The prediction structure of this invention is accurate and interpretable.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to risk prediction of sudden water pollution events, specifically a multimodal data-driven method for predicting the risk of sudden water pollution events. Background Technology

[0002] Sudden water pollution events in river basins are characterized by strong randomness of outbreaks, complex causes, rapid spread, wide range of impact, and long duration of harm. In the risk prediction technology for sudden water pollution events, early engineering applications have focused on the hydrodynamic-water quality coupling mechanism model. By constructing mathematical and physical equations of watershed hydrological cycle, hydrodynamic process and pollutant migration and transformation, numerical simulation of the pollutant diffusion process in the river network can be achieved.

[0003] With the full coverage and continuous improvement of the watershed water environment monitoring sensor network, the volume of long-term monitoring data such as hydrology and water quality, hydrology and meteorology has grown rapidly. Statistical analysis models, represented by time series analysis, multiple linear regression, and Bayesian statistics, have been gradually applied on a large scale, enabling preliminary prediction of water pollution risk trends based on historical monitoring data.

[0004] In recent years, with the rapid development of artificial intelligence and big data technologies, data-driven models based on machine learning and deep learning have become a research hotspot and mainstream development direction in the field of sudden water pollution risk prediction. Although existing technologies for predicting the risk of sudden water pollution events have achieved phased research progress and application results, especially the exploration of multimodal data fusion technologies, which has provided a new technical path for improving the accuracy of risk prediction, the following technical problems still exist in practical engineering applications:

[0005] Most existing methods for predicting the risk of sudden water pollution rely on single types of data, such as water quality monitoring data, geographical environmental factors, or pollution source statistics. They lack the comprehensive utilization of multi-source heterogeneous information such as event text, dynamic hydrological processes, and river network propagation conditions, making it difficult for models to fully depict the entire process of the occurrence, diffusion, and impact of sudden water pollution events.

[0006] Most existing risk prediction methods focus on the static estimation of risk values, ignoring the directional and connectivity characteristics of the spread of sudden water pollution events along the main river, tributaries, lakes, reservoirs and urban water systems. They are difficult to depict the dynamic process of risk spreading from local pollution sources to downstream areas, river corridors or key nodes.

[0007] While existing deep learning methods have good nonlinear modeling capabilities in complex environmental problems, their black-box nature makes it difficult for the models to explain the reasons for the formation of high-risk areas. They cannot clearly point out what geographical background, hydrological conditions, event types, or pollutant characteristics have dominated the risk increase, thus limiting their widespread application in watershed risk prevention and emergency management. Summary of the Invention

[0008] The purpose of this invention is to provide a multimodal data-driven method for predicting the risk of sudden water pollution events, aiming to solve the technical problems mentioned in the background.

[0009] To achieve the above objectives, the present invention provides the following technical solution.

[0010] A multimodal data-driven method for predicting the risk of sudden water pollution events includes the following steps:

[0011] Acquire multimodal data related to sudden water pollution events in the target area, and obtain multimodal feature vectors by encoding the multimodal data;

[0012] An end-to-end prediction model is constructed, including a bidirectional constraint fusion interpretation layer and a prediction output layer; wherein, the bidirectional constraint fusion interpretation layer has built-in mutually coupled modal weight network and cross-modal attention network, which simultaneously perform feature fusion and interpretation generation;

[0013] In feature fusion, the modal weighting network uses adaptively generated modal weighting coefficients to weight the feature vectors of each modality to obtain fused features; the modal weighting coefficients are also directly output as an explanation of the modal contribution of each modality to the prediction result; the modal weighting network uses the generated modal weighting coefficients as prior constraint input to the cross-modal attention network to constrain the global distribution of attention weights;

[0014] In the interpretation generation, the cross-modal attention network generates an attention weight matrix to characterize the interactive dependencies between modalities and features through a physical causal mask-constrained attention mechanism. The attention weight matrix is ​​then used to interactively enhance the features of each modality to obtain enhanced features. The attention weight matrix is ​​also directly output as a fine-grained interpretation of the contribution of each input feature to the prediction result.

[0015] By combining fused features with enhanced features, a comprehensive representation vector is generated and input into the prediction output layer to obtain the risk prediction result. Specifically, when predicting the risk of the target region, the risk prediction result, modal level contribution interpretation, and fine-grained contribution interpretation are obtained simultaneously from the forward propagation process of the prediction model.

[0016] Furthermore, in the step of encoding multimodal data:

[0017] The encoding of the spatial feature vector includes: determining a local spatial window centered on the target location, extracting raster data of multiple geographic environmental variables within the spatial window, using a river network mask to mask the non-water body areas of the raster data, and then encoding the masked raster data through a convolutional neural network to obtain a spatial feature vector used to characterize the spatial heterogeneity of the environmental background.

[0018] The encoding of the time series feature vector includes: taking the current time as the reference, extracting the sequence data of multiple hydrological and meteorological variables within the historical time window before the current time and the future forecast time window after the current time, taking the current time as the time series reference zero point, encoding the sequence data by time position to mark the relative position of each time step with the current time, and then using a recurrent neural network to encode the sequence data to obtain a time series feature vector that represents the spatiotemporal dynamic changes related to risk.

[0019] The encoding of text feature vectors includes: inputting event description text into a pre-trained language model, extracting semantic features from the event description text, adapting and transforming the semantic features through a polluted event word vector fine-tuning layer, and obtaining text feature vectors that represent the semantic attributes of the event.

[0020] The encoding of topological feature vectors includes: using river cross sections, river confluence points, and lake / reservoir nodes as graph nodes, and hydraulic connectivity relationships as graph edges, constructing weighted graph edges based on river flow, length, and slope to build a river network graph structure that characterizes the connectivity and propagation characteristics of the water system; using a graph neural network to encode the river network graph structure to generate node embedding vectors for each graph node, and organizing the node embedding vectors into topological feature vectors according to the hydraulic upstream and downstream order.

[0021] Furthermore, in the modal weight network:

[0022] Learnable linear transformations are performed on the feature vectors of each modality to obtain the corresponding initial weight scores;

[0023] Calculate the cosine similarity between any two modal feature vectors, and generate a modal complementarity penalty term based on the sum of the absolute values ​​of each cosine similarity, so that the high redundancy of modal combinations is suppressed by weights.

[0024] The initial weight score is added to the modal complementarity penalty term to obtain the final weight score; the final weight score is then normalized to obtain the modal weight coefficient.

[0025] The fused features are generated by weighting and summing the feature vectors of each modality using the modality weight coefficients.

[0026] The modal weight coefficients are adaptively adjusted based on the content and complementarity of each modal feature under different event scenarios.

[0027] Furthermore, in cross-modal attention networks:

[0028] The feature vectors of each modality are mapped to a unified dimension and concatenated to form a joint feature matrix; a learnable linear projection is applied to the joint feature matrix to generate a query matrix, a key matrix, and a value matrix.

[0029] Construct a physical causal mask matrix for water pollution propagation; the physical causal mask matrix only applies to the part of the joint feature matrix corresponding to the graph node features contained in the topological feature vector, and is used to prohibit attention connections between downstream node features and upstream node features and non-connected node features. The physical causal mask is not applied to the feature dimensions of spatial feature vectors, temporal feature vectors, and text feature vectors, as well as their cross-modal attention with topological feature vectors.

[0030] Based on the query matrix and the key matrix, an initial attention score is obtained by scaling dot product operation. The initial attention score is then masked using the physical causal mask matrix and normalized to obtain the cross-modal attention weight matrix.

[0031] The value matrix is ​​weighted and aggregated using the cross-modal attention weight matrix to generate cross-modal enhanced features;

[0032] The element values ​​in the cross-modal attention weight matrix are used to characterize the interaction strength between different features and their fine-grained contribution to the prediction results.

[0033] Furthermore, in the step of generating the comprehensive representation vector:

[0034] Aggregate the attention weights between features belonging to the same modality in the attention weight matrix to obtain the inter-modal interaction strength vector; calculate the consistency score between the modal weight coefficients and the inter-modal interaction strength vector;

[0035] A dynamic gating unit is constructed to adaptively adjust the fusion ratio of the fused feature and the enhanced feature based on the consistency score;

[0036] According to the fusion ratio, the fusion feature and the enhanced feature are weighted and fused to generate the comprehensive representation vector.

[0037] Furthermore, it also includes the step of training the prediction model, including:

[0038] The method employs a joint optimization approach using prediction loss and hierarchical interpretation consistency loss, where:

[0039] The predicted loss is constructed based on the risk prediction results and the true labels;

[0040] The hierarchical interpretation consistency loss is used to constrain the distribution of feature contributions, including modality-level loss and feature-level loss:

[0041] The modal-level loss is constructed based on the matching degree between the modal weight coefficients and the prior known modal importance in the sudden water pollution event;

[0042] The feature-level loss is constructed based on the matching degree between the attention weight matrix and the pre-constructed physical causal constraint matrix; the physical causal constraint matrix is ​​a target matrix of the same dimension as the attention weight matrix, constructed according to known physical causal relationships, where the elements represent whether there is a known causal relationship or known causal strength between the corresponding feature pairs.

[0043] Furthermore, the cross-modal attention network employs a topology-guided two-level attention structure, including:

[0044] The intermodal attention layer is used to calculate the overall interaction relationship between different modalities and generate modal-level attention weights.

[0045] The intramodal attention layer is used to calculate local attention weights for different feature dimensions within each modality, generating feature-level attention weights. The intramodal attention layer of the topological modality takes the node embedding vector in the topological feature vector as input and introduces a river network topology guidance mechanism to make the attention weights decay exponentially along the preset water flow direction.

[0046] The cross-modal enhancement feature is obtained by fusing the inter-modal attention output and the intra-modal attention output, and simultaneously outputs multi-granularity interpretation results of modality-level interpretation and feature-level interpretation.

[0047] Furthermore, the modal weight network incorporates an online adaptive prior correction module, which performs the following steps:

[0048] The event category of the sudden water pollution event is identified based on the text feature vector;

[0049] The prior distribution of the dynamically updated modal weights corresponding to the event category is invoked. The prior distribution is generated based on the modal contribution statistics of historical event samples and is updated online as new samples are added.

[0050] The initial modal weight coefficients generated by the modal weight network are weighted and fused with the prior distribution to obtain the corrected modal weight coefficients, which are used as the final modal weight coefficients for generating fused features.

[0051] The weights of the prior distribution in the weighted fusion adaptively decrease as the number of model training iterations increases.

[0052] Furthermore, when the risk prediction results show the existence of high-risk spatial units with a risk probability exceeding a preset threshold, the following steps are performed:

[0053] Centered on the high-risk spatial unit, upstream and downstream associated nodes are determined along the topological connectivity of the river network structure;

[0054] The propagation weight coefficients of each associated node are calculated based on the river flow, length, and hydraulic gradient.

[0055] Based on the propagation weight coefficients, the modal contribution and fine-grained feature contribution of the associated nodes are recalculated.

[0056] Based on the weighted contribution changes of the associated nodes, a risk propagation explanatory chain is generated, which includes the risk origin node, key propagation nodes, main diffusion paths, and the dominant driving factors of each node.

[0057] Furthermore, in the step of generating the risk propagation explanation chain:

[0058] Based on the river flow velocity, flow rate and water level parameters in real-time hydrological monitoring data, and combined with the pollutant diffusion coefficient, under the assumption of default initial release parameters or the worst-case scenario, the estimated arrival time and peak concentration time of pollutants to each associated node are calculated if a pollution event occurs in the high-risk spatial unit.

[0059] Based on the pollutant type identified by the text feature vector, a pre-stored pollutant-specific correction coefficient library is called to obtain the migration attenuation coefficient and adsorption coefficient of the corresponding pollutant, and the propagation weight coefficient is corrected accordingly.

[0060] Based on the revised propagation weight coefficient, the risk contribution of each associated node is recalculated, and the expected arrival time, peak concentration time, and pollutant type-specific driving factors are integrated into the risk propagation explanation chain to generate a dynamic risk propagation explanation chain that includes the time dimension.

[0061] The pollutant-specific correction coefficient library is constructed based on pollutant migration and transformation data from historical sudden water pollution events and is continuously updated as new event samples are added.

[0062] Compared with existing technologies, the beneficial effects of the multimodal data-driven risk prediction method for sudden water pollution events of this invention are:

[0063] This invention acquires and encodes spatial feature vectors, temporal feature vectors, textual feature vectors, and topological feature vectors to comprehensively characterize sudden water pollution events from aspects such as geographical background, hydrological and meteorological dynamics, event semantic attributes, and river network connectivity and propagation characteristics. Furthermore, it constructs a bidirectional constraint fusion interpretation layer that coordinates a modal weight network and a cross-modal attention network. At the feature fusion level, the modal weight network adaptively generates dynamic modal weight coefficients based on the content of each modal feature, dynamically adjusting the contribution of each modality under different event scenarios, and utilizing weighted fusion features to characterize the overall discriminative ability of multimodal information. At the interaction modeling level, the cross-modal attention network, under the prior constraints of modal weights, generates an attention weight matrix through a physical causal mask constraint attention mechanism, structurally modeling the interaction dependencies between modal features in the joint features. This enables the model to fully capture the propagation dependencies under the conditions of river network topological connectivity, thereby effectively characterizing the dynamic propagation characteristics of risk upstream and downstream along the river network.

[0064] In the forward propagation process where feature fusion and interpretation generation are executed simultaneously, this invention directly outputs the modality weight coefficients as an explanation of the modality-level contribution of each modality to the prediction result, and simultaneously outputs the attention weight matrix as a fine-grained explanation of the contribution of each input feature to the prediction result. This allows the prediction model to provide the prediction result and its multi-level attribution explanation simultaneously without relying on additional post-analysis, clarifying the reasons for the increase in dominant risk and significantly enhancing the credibility of the model. Attached Figure Description

[0065] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0066] In the attached diagram:

[0067] Figure 1 This is a flowchart of the multimodal data-driven method for predicting the risk of sudden water pollution events according to the present invention;

[0068] Figure 2 This is an architecture diagram of the prediction model provided by the present invention. Detailed Implementation

[0069] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] The specific implementation of the present invention will be described below with reference to specific embodiments.

[0071] Please refer to Figure 1In one embodiment of the present invention, a multimodal data-driven method for predicting the risk of sudden water pollution events is provided, comprising the following steps:

[0072] Step S101: Obtain multimodal data related to the sudden water pollution event in the target area, and obtain multimodal feature vectors by encoding the multimodal data;

[0073] Among them, multimodal feature vectors include spatial feature vectors, temporal feature vectors, text feature vectors, and topological feature vectors;

[0074] In the encoding of spatial features, this invention determines a local spatial window centered on the target location, such as an area with a side length of 10km centered on the target point; by extracting raster data of multiple geographic environmental variables within this spatial window, such as including but not limited to land use type, soil type, elevation, slope, vegetation cover index, etc., each variable constitutes a band of raster data; then, using a pre-stored river network mask, the raster data is processed to mask non-water body areas, setting the raster cell values ​​outside the water body areas determined by the river network mask to zero, so that subsequent convolution operations only focus on the geographic features of the water body and its adjacent areas;

[0075] Next, the masked multi-band raster data is input into a convolutional neural network, which includes three convolutional layers and two pooling layers. Finally, a fixed-dimensional spatial feature vector V is obtained through global average pooling. s Spatial feature vector V s Used to characterize the spatial heterogeneity of the environmental context;

[0076] In one implementation of this invention, the parameters of the convolutional neural network are set as follows:

[0077] First convolutional layer: kernel size 3×3, stride 1, number of output channels 32, activation function is ReLU;

[0078] First-level pooling: max pooling, window size 2×2, stride 2;

[0079] The second convolutional layer has a kernel size of 3×3, a stride of 1, 64 output channels, and the activation function is ReLU.

[0080] Second-level pooling: max pooling, window size 2×2, stride 2;

[0081] The third convolutional layer has a kernel size of 3×3, a stride of 1, 128 output channels, and the activation function is ReLU.

[0082] Global average pooling: The feature map output from the third convolutional layer is subjected to global average pooling in the spatial dimension, resulting in a spatial feature vector V with a dimension of 128. s ;

[0083] In the encoding of temporal feature vectors, in order to solve the problem of inconsistency in time reference between the training and inference stages, the zero point is no longer the time when the event occurs, but the current time is uniformly used as the temporal reference zero point. For training samples, the current time is defined as the time point before the event occurs (e.g., 6 hours) to simulate the scenario where only historical and forecast data are available during real inference. For the inference stage, the current time is the time when the model executes the prediction.

[0084] After determining the current time, the present invention extracts sequence data of multiple hydro-meteorological variables within a historical time window (e.g., the previous 24 hours) before the current time and a future forecast time window (e.g., the next 12 hours, using numerical weather prediction and hydrological forecast model output) after the current time. The hydro-meteorological variables include, but are not limited to, rainfall, river level, flow rate, flow velocity, and water temperature.

[0085] After extraction, the sequence data is processed by time position encoding. The current time is marked as the time position zero of the time series feature. Each time step obtains a time position encoding vector based on its time difference with the current time (in hours). By concatenating the time position encoding vector with the variable value vector of the corresponding time step, the model can perceive the relative time relationship between each time step and the current time.

[0086] Furthermore, this invention inputs the sequence data with time position encoding into a two-layer Long Short-Term Memory (LSTM) network for encoding, and takes the hidden state of the last time step as the temporal feature vector V. t This yields a temporal feature vector V used to characterize the spatiotemporal dynamic changes related to risk. t ;

[0087] Specifically, in the two-layer long short-term memory network LSTM of the present invention, a Dropout layer is set after each LSTM layer, and the time position encoding adopts sine and cosine encoding method, which is concatenated with the variable value vector of each time step to form a time sequence input vector;

[0088] For the text feature vector, this invention inputs the event description text (e.g., "A chemical plant's storage tank leaked, and the pollutants are benzene compounds, which are expected to affect downstream water intakes") into a pre-trained language model, such as the BERT model, and extracts the hidden state vector corresponding to its [CLS] tag as the initial semantic feature. This initial semantic feature is then adapted and transformed through a pollution event word vector fine-tuning layer. The fine-tuning layer consists of a fully connected layer and a ReLU activation function, and the output is the text feature vector V. x Vector V x Semantic attributes used to characterize events;

[0089] In one specific implementation, the pre-trained language model uses the commonly used BERT model, fine-tuning the fully connected layers based on the word vectors of contamination events to finally obtain the text feature vector V. x .

[0090] For the topological feature vector, river cross-sections, river segment confluences, and lake / reservoir nodes within the target area are used as graph nodes, and hydraulic connectivity is used as graph edges. For constructing edge weights, considering river flow Q, river segment length L, and average river segment slope S, the edge weights are defined as follows:

[0091]

[0092] in, The average flow rate of the river section The length of the river segment The average slope of the river segment is used to construct a weighted directed graph as the structure of the river network.

[0093] Then, a three-layer graph convolutional network (GCN) is used to encode the river network graph structure, generating a node embedding vector for each graph node. The node embedding vectors are arranged in the order of upstream and downstream water flow. Starting from the upstream source node of the river network, the nodes on each branch are arranged sequentially along the water flow direction, ultimately forming a matrix-form topological feature vector V. g The rows of the matrix correspond to the nodes, and the columns correspond to the embedding dimensions.

[0094] In the three-layer graph convolutional network GCN:

[0095] The first layer of GCN has the following input dimensions: the original feature dimension of the nodes (default 16, including attributes such as river flow, length, and slope), the output dimension is 64, the activation function is ReLU, and batch normalization layer and Dropout layer are set.

[0096] The second GCN layer has an input dimension of 64 and an output dimension of 128. The activation function is ReLU, and batch normalization and dropout layers are set.

[0097] The third layer of GCN: input dimension is 128, output dimension is 128, and there is no activation function;

[0098] Ultimately, the node embedding vector of each graph node has a dimension of 128, and the topological feature vector V g It is an N×128 matrix, where N is the number of nodes in the river network graph.

[0099] Please continue to refer to Figure 1 and Figure 2 The risk prediction method of this invention further includes:

[0100] Step S102: Construct an end-to-end prediction model, which includes a bidirectional constraint fusion interpretation layer and a prediction output layer; the bidirectional constraint fusion interpretation layer has a built-in mutually coupled modal weight network and a cross-modal attention network, which simultaneously perform feature fusion and interpretation generation; wherein, feature fusion is used to generate fused features, which are then enhanced features in the interpretation generation.

[0101] In the feature fusion of the prediction model, the modal weight network receives the feature vectors V of each modality. s V t V x V g Dynamic modal weight coefficients are adaptively generated based on the modal feature content, and the modal weight coefficients are used to perform weighted summation on the modal feature vectors to obtain fused features;

[0102] Specifically, by performing a learnable linear transformation on each modality feature vector, mapping each vector to a one-dimensional scalar, we can obtain the initial weight score for each vector:

[0103]

[0104] In the formula, k represents the index of the modal eigenvector. The initial weight scores corresponding to each vector are expressed as follows: , , and ; Let T represent the learnable weight vector, and let T denote the transpose of the vector. This represents the learnable bias term. This represents the high-dimensional feature representation obtained by the feature extraction network for the corresponding mode.

[0105] Furthermore, for any two modal eigenvectors , First, global average pooling is performed along the feature dimension. Then, based on the principle of cosine similarity, feature vectors of any two modalities are calculated. , Cosine similarity between , , The summary vector is obtained by global average pooling of modal feature vectors. A modal complementarity penalty term is generated based on the sum of the absolute values ​​of each cosine similarity, so that the high redundancy of modal combinations is suppressed by weights.

[0106] In this embodiment, the modal complementarity penalty term P is expressed as:

[0107]

[0108] In the formula, This is an adjustable penalty intensity hyperparameter, with a default value of 0.1;

[0109] The modal complementarity penalty term P provided in this embodiment is used to characterize that the more consistent the general vector directions of any two modes are, the higher the redundancy of the information carried by the two modes. The larger the value of the penalty term, the more it suppresses the weight in subsequent calculations.

[0110] Finally, this invention adds the initial weight score to the modal complementarity penalty term to obtain the final weight score, and further performs Softmax normalization on the weight score to obtain the modal weight coefficient. This modal weight coefficient is directly used as the modal level contribution of each mode to the prediction result.

[0111] Furthermore, the present invention uses the obtained modal weight coefficients to perform a weighted summation of the modal feature vectors, thereby obtaining the desired fusion features;

[0112] In a preferred embodiment, the modal weighting coefficients are adaptively adjusted based on the content and complementarity of each modal feature under different event scenarios.

[0113] It should be noted that the modal weight coefficients of this invention are also directly output as an explanation of the modal contribution of each modality to the prediction result; in addition, the modal weight network uses the generated modal weight coefficients as prior constraints input to the cross-modal attention network to constrain the global distribution of attention weights.

[0114] In the interpretation generation of step S102 of this embodiment, the cross-modal attention network receives the feature vectors of each modality and the prior constraints of the modality weights, generates an attention weight matrix to characterize the interactive dependencies between modalities and features through the attention mechanism of physical causal mask constraints, and uses the attention weight matrix to interactively enhance the features of each modality to obtain enhanced features; the attention weight matrix is ​​also directly output as a fine-grained interpretation of the contribution of each input feature to the prediction result.

[0115] Specifically, in cross-modal attention networks:

[0116] First, the feature vectors of each modality are mapped to a unified dimension d=128 through their respective fully connected projection layers. Then, they are concatenated along the feature dimensions to form a joint feature matrix. ,in, Let N represent the set of real numbers, where N is the total number of feature dimensions across the four modalities.

[0117] Then, a learnable linear projection is applied to the joint feature matrix to generate the query matrix. Key matrix Sum matrix ,in, , , All are d×d learnable parameter matrices;

[0118] Furthermore, a physical causal mask matrix M for water pollution propagation is constructed. The physical causal mask matrix only applies to the part of the joint feature matrix corresponding to the graph node features contained in the topological feature vector. It is used to prohibit attention connections between downstream node features and upstream node features and non-connected node features. The physical causal mask is not applied to the feature dimensions of spatial feature vectors, temporal feature vectors, and text feature vectors, as well as their cross-modal attention with topological feature vectors.

[0119] Specifically, the dimension of the mask matrix M is the same as that of the attention score matrix, which is N×N. The definition rules of the elements in the mask M are as follows: For the parts that do not involve topological modal features (i.e., attention connections within and between spatial, temporal, and textual features), the corresponding mask value is set to 0 (no masking).

[0120] For the part involving topological modal features, if the feature corresponding to the query position belongs to a graph node j and the feature corresponding to the key position belongs to a graph node i, then the determination is made based on the hydraulic connectivity and upstream-downstream relationship between node i and node j in the river network graph structure. If node i is located downstream of node j (i.e., the water flows from i to downstream rather than to j), or there is no hydraulic connectivity path between node i and node j, then the element value at the corresponding position in M ​​is set to -∞ (i.e., masking is applied), otherwise it is set to 0.

[0121] The masking mechanism in this embodiment of the invention forces the attention within the topological mode to flow only along the direction of water flow (upstream attention to downstream) and the connectivity path, which conforms to the physical laws of water pollution diffusion;

[0122] Based on the query matrix and the key matrix, an initial attention score is obtained by scaling dot product operation. The initial attention score is then masked using the physical causal mask matrix and normalized to obtain the cross-modal attention weight matrix.

[0123] The initial attention score matrix is ​​calculated based on the query matrix and the key matrix. The physical causality mask matrix M is superimposed on the initial attention score matrix, represented as: S masked =S+M, where the masked position approaches zero in the subsequent Softmax operation due to the addition of -∞, and for S masked Each row is Softmax normalized to obtain the cross-modal attention weight matrix A;

[0124] In the calculation process, the modal weighting coefficient α is introduced as a priori constraint: when generating the joint feature matrix, each modal eigenvector is multiplied by its corresponding modal weighting coefficient α. k Then, the features are concatenated to give the modal features with higher modal weights a stronger initial representation in the attention calculation.

[0125] Finally, the cross-modal attention weight matrix A is used to perform weighted aggregation on the value matrix V to obtain the cross-modal enhancement feature F. enhanced =A·V;

[0126] The attention weight matrix A is directly used as the output to interpret the fine-grained contribution of each input feature to the prediction result; the element values ​​in matrix A are used to characterize the interaction strength between different features and their fine-grained contribution to the prediction result.

[0127] Please continue to refer to Figure 1 and Figure 2 The risk prediction method of this invention further includes:

[0128] Step S103: Combine the fused features and enhanced features to generate a comprehensive representation vector, and input it into the prediction output layer to obtain the risk prediction result; when performing risk prediction on the target region, simultaneously obtain the risk prediction result, the modal level contribution explanation and the fine-grained contribution explanation from the forward propagation process of the prediction model.

[0129] In the step of generating the comprehensive representation vector:

[0130] Aggregate the attention weights between features belonging to the same modality in the attention weight matrix to obtain the inter-modal interaction strength vector; calculate the consistency score between the modal weight coefficients and the inter-modal interaction strength vector;

[0131] Specifically, when calculating the modality-level consistency score between the modality weight coefficients and the attention weight matrix, the attention weights between features belonging to the same modality in the attention weight matrix A are aggregated, including:

[0132] For all feature pairs between mode i and mode j, sum and normalize their attention weights to obtain a 4×4 inter-modal interaction strength matrix A. modal The maximum value of each column in the intermodal interaction intensity matrix is ​​used to construct a 4-dimensional intermodal interaction intensity vector h. attn The modal weight coefficient vector is: Furthermore, the cosine similarity calculation method is used to obtain the consistency score c=cos(α,h) for the two modes i and j. attn );

[0133] After calculating the score c, a dynamic gating unit is constructed to adaptively adjust the fusion ratio of the fused features and the enhanced features according to the score c. Finally, based on the fusion ratio, the fused features and the enhanced features are weighted and fused to obtain the comprehensive representation vector described in this embodiment. Through this fusion method, the present invention enables the model to rely more on the result of global weighted fusion when the modal weight coefficients and attention interaction intensity are highly consistent; when there is a discrepancy between the two, the model tends to adopt the fine-grained features enhanced by cross-modal interaction.

[0134] The gate value g is obtained by mapping the consistency score c, and is expressed as:

[0135]

[0136] Where σ is the Sigmoid function. and The gating value g is a learnable parameter, ranging from 0 to 1;

[0137] Finally, the present invention inputs the obtained comprehensive representation vector Z into the prediction output layer, which consists of two fully connected layers and a Sigmoid output unit, and finally outputs a scalar value between 0 and 1 as the probability of a sudden water pollution event occurring in the target area within a preset time period in the future.

[0138] The prediction output layer specifically includes:

[0139] The first fully connected layer has an input dimension of 128, an output dimension of 64, a ReLU activation function, and a Dropout layer.

[0140] The second fully connected layer has an input dimension of 64, an output dimension of 1, and no activation function.

[0141] Sigmoid output unit: Maps the output of the second fully connected layer to the [0,1] interval to obtain the risk probability value.

[0142] Furthermore, it also includes the step of training the prediction model, including:

[0143] The method employs a joint optimization approach using prediction loss and hierarchical interpretation consistency loss, where:

[0144] The predicted loss is constructed based on the risk prediction results and the true labels, and the predicted loss adopts binary cross-entropy loss.

[0145] The hierarchical interpretation consistency loss is used to constrain the distribution of feature contributions, including modality-level loss and feature-level loss:

[0146] The modal-level loss is constructed based on the matching degree between the modal weight coefficients and the prior known modal importance in the sudden water pollution event;

[0147] The feature-level loss is constructed based on the matching degree between the attention weight matrix and the pre-constructed physical causal constraint matrix; the physical causal constraint matrix is ​​a target matrix of the same dimension as the attention weight matrix, constructed according to known physical causal relationships, where the elements represent whether there is a known causal relationship or known causal strength between the corresponding feature pairs.

[0148] Specifically, in the specific implementation of this invention, the layered interpretation consistency loss Including modal level loss and feature-level loss ;

[0149] Modal level loss is based on modal weight coefficients Prior importance of known modes The matching degree between them is constructed, where β is a 4-dimensional vector. The relative importance of each modality is pre-defined by domain experts based on the event type, and the difference between them is measured using KL divergence, expressed as:

[0150]

[0151] In the formula, This represents the prior weight of the k-th mode, where k represents the mode index. This represents the model learning weights for the k-th modality. Indicates KL divergence based on the baseline;

[0152] Feature-level loss The model is constructed based on the matching degree between the attention weight matrix A and the pre-constructed physical causal constraint matrix C. Matrix C is a target matrix with the same dimension as A, and its construction method is as follows: domain experts label whether there is a causal relationship between feature pairs based on known physical causal knowledge—if there is a known physical causal relationship between feature i and feature j (for example, there is a hydrodynamic causal relationship between upstream flow characteristics and downstream water level characteristics), then... Set to 1; otherwise set to 0.

[0153] Furthermore, this invention uses element-wise binary cross-entropy to represent the difference between A and C:

[0154]

[0155] The total loss function is expressed as: ,in, and To balance the hyperparameters of the various loss weights, all are set to 0.1 by default; N represents the total number of features. Represents the model attention weights for feature pairs (i,j); The causal constraint label represents the feature pair (i,j);

[0156] During the training of the prediction model, the Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and 200 training iterations.

[0157] During training, L2 regularization was applied to the weights of all convolutional layers, fully connected layers, and GCN layers, with a weight decay coefficient of 1e-4. In addition, the weights of convolutional and fully connected layers were initialized using He normality, and the bias was initialized to 0. The weights of GCN layers were initialized using Xavier uniformity. A cosine annealing learning rate decay strategy was adopted, with an initial learning rate of 0.001, a minimum learning rate of 1e-6, and an annealing period of 200 rounds.

[0158] Furthermore, an early stopping strategy is adopted during training to prevent overfitting. The monitoring metric is the validation set AUC, and the patience value is 20. That is, if the validation set AUC does not improve for 20 consecutive rounds, the training is terminated early, and the model weight with the highest validation set AUC is saved.

[0159] Furthermore, in step S103, three types of results are simultaneously acquired during the forward propagation process of the model:

[0160] Predict the risk probability value of the output layer;

[0161] The modal weight coefficients output by the modal weight network represent the modal contribution of each mode to the current prediction. The higher the weight of a mode, the more critical that type of information is in the current risk assessment.

[0162] The attention weight matrix A output by the cross-modal attention network can be used to analyze which specific feature interactions contribute most to the risk prediction results by retrieving the feature pairs with the largest weights in the matrix, thus achieving fine-grained feature-level attribution explanation.

[0163] In one implementation of the present invention, the cross-modal attention network adopts a topology-guided two-layer attention structure, which includes an inter-modal attention layer and an intra-modal attention layer; wherein, the former is used to calculate the overall interaction relationship between different modalities and generate modal-level attention weights; the latter is used to calculate local attention weights for different feature dimensions within each modality and generate feature-level attention weights.

[0164] The internal attention layer of the topological modality takes the node embedding vector in the topological feature vector as input and introduces a river network topology guidance mechanism to make the attention weight decay exponentially along the preset water flow direction. The topology-guided two-level attention structure adopts a serial connection method, which includes first calculating the global modal interaction weight through the inter-modal attention layer, and then inputting the inter-modal attention output into the corresponding intra-modal attention layer of each modality to calculate the local feature interaction weight.

[0165] The final cross-modal enhancement feature is the element-wise sum of the inter-modal attention output and the intra-modal attention output;

[0166] The exponential decay mechanism of water flow direction in this embodiment of the invention is expressed as follows:

[0167]

[0168] In the formula, These are the original attention weights output by the attention layer within the topological mode. Let be the hydraulic distance between node i and node j (i.e., the path length along the direction of water flow). This is the attenuation coefficient, with a default value of 0.01 / km; in this embodiment, the attention weight between nodes that are farther apart is lower;

[0169] In this embodiment, the cross-modal enhancement features are obtained by fusing inter-modal attention outputs and intra-modal attention outputs, and simultaneously output multi-granularity interpretation results of modal-level interpretation and feature-level interpretation.

[0170] Furthermore, in a preferred embodiment of the present invention, the modal weight network incorporates an online adaptive prior correction module, which is used to perform the following steps:

[0171] The event category of the sudden water pollution event is identified based on the text feature vector;

[0172] The prior distribution of the dynamically updated modal weights corresponding to the event category is invoked. The prior distribution is generated based on the modal contribution statistics of historical event samples and is updated online as new samples are added.

[0173] The initial modal weight coefficients generated by the modal weight network are weighted and fused with the prior distribution to obtain the corrected modal weight coefficients, which are used as the final modal weight coefficients for generating fused features.

[0174] The weights of the prior distribution in the weighted fusion adaptively decrease as the number of model training iterations increases.

[0175] Furthermore, when the risk probability output in step 103 of this invention exceeds a preset threshold (the threshold value can be 0.7), the dynamic risk propagation explanation chain generation process is initiated; specifically including:

[0176] First, identify high-risk spatial units, such as the local spatial windows extracted from the spatial features in step S101; then, with the high-risk spatial units as the center, determine the upstream associated node set and the downstream associated node set along the topological connectivity of the river network structure.

[0177] In addition, this invention takes the river section where the high-risk unit is located as the starting point, traverses upstream along the connecting edge until the source node, and includes all nodes on the traversal path into the upstream associated node set; traverses downstream along the connecting edge until the sink or estuary, and includes all nodes on the traversal path into the downstream associated node set.

[0178] Furthermore, the propagation weight coefficients of each associated node are calculated based on the river flow, length, and hydraulic gradient. For example, for any associated node j, the propagation weight coefficient between it and the central node c of the high-risk spatial unit is defined. , represented as:

[0179]

[0180] in, Let this be the average flow rate on the path from node c to node j. This is the total path length. The average slope of the path;

[0181] Next, based on the propagation weight coefficient Recalculate the modal contribution and fine-grained feature contribution of each associated node. Specifically, for each associated node j, calculate its corresponding topological feature vector V. g The portion belonging to that node is weighted and combined with the modal weight coefficient α and attention weight matrix A output by the model to obtain the risk contribution score of that node.

[0182] Finally, based on the weighted contribution changes of each associated node, a risk propagation explanation chain is generated along the upstream and downstream directions, including the risk origin node, key propagation nodes, main diffusion paths, and the dominant driving factors of each node. Ultimately, the explanation chain in this embodiment is represented in text form, such as: the risk of high-risk unit A mainly originates from pollutant input from upstream node B (contribution 0.35), and the river conditions at key propagation node C accelerate pollutant migration (contribution 0.28). It is expected to mainly propagate along the path B→C→A, with the dominant driving factors being high river velocity and low slope.

[0183] Based on the generated risk propagation explanation chain, this embodiment further incorporates time dimension information;

[0184] In this step, since the actual initial release time and amount of pollutants cannot be obtained in the actual reasoning stage, this embodiment sets a default initial release parameter, such as assuming that a conservative pollutant release at a unit concentration occurs instantaneously in a high-risk spatial unit at the current moment. Based on this assumption, the current moment is taken as the initial moment of pollutant release, and the unit concentration is taken as the initial concentration boundary condition. Using the river network hydrodynamic parameters and the pollutant convection and diffusion equation, the time series concentration distribution of pollutants propagating to each associated node along each connected path is calculated. Then, the expected arrival time and peak concentration time are extracted from it. The risk spatial distribution can be characterized by dynamic propagation time series information with clear physical time meaning, thereby incorporating the time dimension into the risk propagation explanation chain.

[0185] Subsequently, based on the river flow velocity v, flow rate Q, and water level H parameters from real-time hydrological monitoring data, combined with the longitudinal diffusion coefficient of pollutants... Under the assumption of default initial release parameters or worst-case scenario, the estimated arrival time and peak concentration time of pollutants to each associated node are calculated if a pollution event occurs in the high-risk spatial unit. Specifically, the estimated arrival time of pollutants to each associated node is calculated using the analytical solution of the one-dimensional convection-diffusion equation. and peak concentration time , respectively represented as:

[0186]

[0187]

[0188] Where L is the propagation distance and v is the average flow velocity of the river cross section;

[0189] Based on the pollutant type (such as benzene series compounds) identified by the text feature vector in step S1, a pre-stored pollutant-specific correction coefficient library is called. This coefficient library is stored in key-value pair format, where the key is the pollutant type and the value includes the migration attenuation coefficient. and adsorption coefficient ;

[0190] After obtaining the corresponding coefficients, the propagation weight coefficients are corrected. Represented as:

[0191]

[0192] In the formula, L represents the propagation distance. This represents the propagation weight coefficient before correction.

[0193] The risk contribution of each associated node is recalculated based on the corrected propagation weight coefficient, and the expected arrival time, peak concentration time, and pollutant type-specific driving factors are integrated into the risk propagation explanation chain to generate a dynamic risk propagation explanation chain that includes the time dimension. For example, if a benzene series leak occurs in high-risk unit A, it is expected to reach the downstream water intake node D after 2.3 hours and reach the peak concentration after 5.1 hours. The high mobility and low adsorption of benzene series significantly aggravate the exposure risk of downstream nodes.

[0194] The pollutant-specific correction coefficient library is constructed based on monitoring data of pollutant migration and transformation from historical sudden water pollution events. When a new water pollution event sample is generated, the system extracts key parameters from the post-event assessment report and updates the attenuation and adsorption coefficients of the corresponding pollutants in the coefficient library using an exponential moving average method, thus achieving continuous updates.

[0195] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0196] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0197] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0198] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-modal data-driven risk prediction method for sudden water pollution events, characterized in that, Includes the following steps: Acquire multimodal data related to the sudden water pollution event in the target area, and obtain multimodal feature vectors by encoding the multimodal data; in the step of encoding the multimodal data: The encoding of spatial feature vectors includes: determining a local spatial window centered on the target location, extracting raster data of multiple geographic environmental variables within the spatial window, using a river network mask to mask non-water body areas of the raster data, and then encoding the masked raster data through a convolutional neural network to obtain a spatial feature vector used to characterize the spatial heterogeneity of the environmental background. The encoding of the time series feature vector includes: taking the current time as the reference, extracting the sequence data of multiple hydrological and meteorological variables within the historical time window before the current time and the future forecast time window after the current time, taking the current time as the time series reference zero point, encoding the sequence data by time position to mark the relative position of each time step with the current time, and then using a recurrent neural network to encode the sequence data to obtain a time series feature vector that represents the spatiotemporal dynamic changes related to risk. The encoding of text feature vectors includes: inputting event description text into a pre-trained language model, extracting semantic features from the event description text, adapting and transforming the semantic features through a polluted event word vector fine-tuning layer, and obtaining text feature vectors that characterize the semantic attributes of the event. The encoding of topological feature vectors includes: using river cross-sections, river confluence points, and lake / reservoir nodes as graph nodes, and hydraulic connectivity relationships as graph edges, constructing weighted graph edges based on river flow, length, and slope to build a river network graph structure that characterizes the connectivity and propagation characteristics of the water system; using a graph neural network to encode the river network graph structure to generate node embedding vectors for each graph node, and organizing the node embedding vectors into topological feature vectors according to the hydraulic upstream and downstream order; An end-to-end prediction model is constructed, including a bidirectional constraint fusion interpretation layer and a prediction output layer; wherein, the bidirectional constraint fusion interpretation layer has built-in mutually coupled modal weight network and cross-modal attention network, which simultaneously perform feature fusion and interpretation generation; In feature fusion, the modal weighting network receives feature vectors from each modality, adaptively generates dynamic modal weighting coefficients based on the content of each modality feature, and uses the modal weighting coefficients to perform a weighted summation of each modality feature vector to obtain the fused features. The modal weighting coefficients are also directly output as an explanation of the modal contribution of each modality to the prediction result. The modal weighting network uses the generated modal weighting coefficients as prior constraints input to the cross-modal attention network to constrain the global distribution of attention weights. In the interpretation generation, the cross-modal attention network receives the feature vectors of each modality and the prior constraints of the modality weights. It generates an attention weight matrix to characterize the interactive dependencies between modalities and features through an attention mechanism with physical causal mask constraints. The attention weight matrix is ​​then used to interactively enhance the features of each modality to obtain enhanced features. The attention weight matrix is ​​also directly output as a fine-grained explanation of the contribution of each input feature to the prediction result. By combining fused features with enhanced features, a comprehensive representation vector is generated and input into the prediction output layer to obtain the risk prediction result. Specifically, when predicting the risk of the target region, the risk prediction result, modal level contribution interpretation, and fine-grained contribution interpretation are obtained simultaneously from the forward propagation process of the prediction model.

2. The multi-modal data driven burst water pollution event risk prediction method of claim 1, wherein, In modal weighted networks: A learnable linear transformation is performed on each modality feature vector to obtain the corresponding initial weight score, expressed as: ； In the formula, k represents the index of the modal eigenvector. The initial weight scores corresponding to each vector are expressed as follows: , , and ; Let T represent the learnable weight vector, and let T denote the transpose of the vector. This represents the learnable bias term. This represents the high-dimensional feature representation obtained by the feature extraction network for the corresponding mode. Calculate the cosine similarity between any two modal feature vectors, generate a modal complementarity penalty term based on the sum of the absolute values ​​of each cosine similarity, add the initial weight score to the modal complementarity penalty term to obtain the final weight score, and perform normalization to obtain the modal weight coefficients. The modal weight coefficients are used to weight and sum the feature vectors of each modality to generate fused features; the modal weight coefficients are adaptively adjusted according to the content and complementarity of each modal feature under different event scenarios.

3. The multi-modal data-driven burst water pollution event risk prediction method of claim 2, wherein, In cross-modal attention networks: The feature vectors of each modality are concatenated to form a joint feature matrix; a learnable linear projection is then introduced into the joint feature matrix to generate a query matrix, a key matrix, and a value matrix. A physical causal mask matrix for water pollution propagation is constructed. The mask matrix is ​​only applied to the part of the joint feature matrix corresponding to the graph node features contained in the topological feature vector. It is used to prohibit attention connections between downstream node features and upstream node features, as well as between features of non-connected nodes. The physical causal mask is not applied to the feature dimensions of spatial feature vectors, temporal feature vectors, and text feature vectors, as well as their cross-modal attention with topological feature vectors. Based on the query matrix and key matrix, the initial attention score is obtained by scaling the dot product operation. The initial attention score is then masked using the physical causal mask matrix to obtain the cross-modal attention weight matrix. The cross-modal attention weight matrix is ​​used to weight and aggregate the value matrix to generate cross-modal enhanced features. The element values ​​in the cross-modal attention weight matrix are used to characterize the interaction strength between different features and their fine-grained contribution to the prediction results.

4. The multi-modal data-driven flash water pollution event risk prediction method of claim 3, wherein, In the step of generating the comprehensive representation vector: Aggregate the attention weights among features belonging to the same modality in the attention weight matrix to obtain the intermodal interaction strength vector; Calculate the consistency score between the modal weight coefficients and the intermodal interaction strength vector; A dynamic gating unit is constructed, which adaptively adjusts the fusion ratio of the fused features and the enhanced features based on the consistency score. Based on the adjusted fusion ratio, the fused features and the enhanced features are weighted and fused to generate a comprehensive representation vector.

5. The multimodal data-driven method for predicting the risk of sudden water pollution events according to any one of claims 2 to 4, characterized in that, It also includes the step of training the prediction model, including: The method employs a joint optimization approach using prediction loss and hierarchical interpretation consistency loss, where: The predicted loss is constructed based on risk prediction results and real labels; The hierarchical interpretation consistency loss is used to constrain the distribution of feature contributions, including modal-level loss and feature-level loss; wherein, the modal-level loss is constructed based on the matching degree between the modal weight coefficients and the prior known modal importance in the sudden water pollution event; the feature-level loss is constructed based on the matching degree between the attention weight matrix and the pre-constructed physical causal constraint matrix.

6. The multi-modal data-driven flash water pollution event risk prediction method of claim 5, wherein, The cross-modal attention network employs a topology-guided two-level attention structure, including: The intermodal attention layer is used to calculate the overall interaction relationship between different modalities and generate modal-level attention weights. The intramodal attention layer is used to calculate local attention weights for different feature dimensions within each modality and generate feature-level attention weights. Among them, the intramodal attention layer of the topological modality takes the node embedding vector in the topological feature vector as input and introduces a river network topology guidance mechanism to make the attention weights decay exponentially along the preset water flow direction.

7. The multi-modal data-driven burst water pollution event risk prediction method of claim 6, wherein, The modality weighting network incorporates an online adaptive prior correction module, which performs the following steps: The event category of the sudden water pollution event is identified based on the text feature vector; The prior distribution of the dynamically updated modal weights corresponding to the event category is invoked. The prior distribution is generated based on the modal contribution statistics of historical event samples and is updated online as new samples are added. The initial modal weight coefficients generated by the modal weight network are weighted and fused with the prior distribution to obtain the corrected modal weight coefficients, which are used as the final modal weight coefficients for generating fused features. Among them, the weights of the prior distribution in the weighted fusion decrease as the number of model training iterations increases.

8. The multi-modal data-driven flash water pollution event risk prediction method of claim 7, wherein, When the risk prediction results show that there are high-risk spatial units with a risk probability exceeding a preset threshold, the following steps are performed: Centered on the high-risk spatial unit, upstream and downstream associated nodes are determined along the topological connectivity of the river network structure; The propagation weight coefficients of each associated node are calculated based on the river flow, length, and hydraulic gradient. Based on the propagation weight coefficient, the modal contribution and fine-grained feature contribution of the associated nodes are recalculated; Based on the weighted contribution changes of the associated nodes, a risk propagation explanatory chain is generated, which includes the risk origin node, key propagation nodes, main diffusion paths, and the dominant driving factors of each node.

9. The multi-modal data-driven flash water pollution event risk prediction method of claim 8, wherein, In the steps of generating the risk propagation explanation chain: Based on the river flow velocity, flow rate and water level parameters in real-time hydrological monitoring data, and combined with the pollutant diffusion coefficient, the estimated arrival time and peak concentration time of pollutants to each associated node are calculated if a pollution event occurs in the high-risk spatial unit. Based on the pollutant type identified by the text feature vector, a pre-stored pollutant-specific correction coefficient library is called to obtain the migration attenuation coefficient and adsorption coefficient of the corresponding pollutant, and the propagation weight coefficient is corrected accordingly. Based on the revised propagation weight coefficient, the risk contribution of each associated node is recalculated, and the expected arrival time, peak concentration time, and pollutant type-specific driving factors are integrated into the risk propagation explanation chain to generate a dynamic risk propagation explanation chain that includes the time dimension.

Images