An artificial intelligence-based soil pollutant monitoring method and system

Abstract

The application relates to the technical field of environmental intelligent monitoring, and discloses a soil pollutant monitoring method and system based on artificial intelligence. The method comprises the following steps: receiving a data stream containing description information and parameter readings of a monitoring node, first converting the data stream into an information embedding vector and a state coding vector through multi-level feature mapping to form a primary feature set. Then, the set is input into an iterative evolution network for multiple rounds of feature refinement: the network dynamically adjusts the feature extraction dimension of the next round according to the comparison result of each round output and a historical knowledge base, and finally generates a high-level pollution feature set. According to this, the matching degree with a preset pollution source template is calculated, a judgment result of marking a potential pollution source category and a spatial range is generated, and a regulation instruction is output by driving a strategy engine. Through deep feature fusion of heterogeneous data and dynamic feature extraction based on feedback, the application improves the adaptive recognition accuracy of complex pollution modes.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental monitoring technology, specifically to a method and system for monitoring soil pollutants based on artificial intelligence. Background Technology

[0002] Current automatic monitoring technologies for soil pollutants are typically based on predetermined data processing workflows. These methods directly concatenate or simply normalize observational data of different natures, such as discrete descriptive text and continuous time-series readings. This processing method obscures the physical and semantic differences between the data, making it difficult for subsequent analysis to effectively capture key typological characteristics and dynamic evolution information in pollution events. The representativeness of the model input features is thus limited.

[0003] Most existing monitoring models employ a fixed structure, and their parameters and feature extraction logic remain unchanged after deployment. Faced with complex and ever-changing pollution scenarios, these static models struggle to adaptively adjust to new patterns reflected in real-time data. Their analysis process lacks continuous interaction and reference with historical experience, resulting in limited ability to identify unknown or mutated pollution patterns and insufficient accuracy and robustness in long-term monitoring.

[0004] A monitoring method is needed that can effectively integrate multi-source heterogeneous monitoring data and possess online self-optimization capabilities. This method should be able to deeply mine the underlying information from both textual descriptions and physical readings, and construct an effective joint representation of both. Simultaneously, the core analysis module of this method should be able to proactively adjust its feature focus direction based on dynamic comparisons between real-time analysis results and historical knowledge bases, thereby improving the accuracy of perception and identification of complex pollution situations. Summary of the Invention

[0005] The purpose of this invention is to provide a soil pollutant monitoring method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a soil pollutant monitoring method based on artificial intelligence, the method comprising:

[0007] Receive raw soil observation data stream from monitoring nodes, the raw soil observation data stream containing pollutant description information sequence and environmental parameter time series readings;

[0008] Multi-level feature mapping processing is performed on the raw soil observation data stream to obtain a primary feature set associated with pollution events. The multi-level feature mapping processing refers to the process of converting the descriptive information sequence into an information embedding vector and the environmental parameter time series readings into a state coding vector.

[0009] inputting the primary feature set into an iterative evolution network for multi-round feature refinement processing, the iterative evolution network taking the comparison result of the refined feature set output by the previous round and the historical pollution pattern knowledge base as feedback to dynamically adjust the feature extraction dimension of the next round, and generating a senior pollution feature set through the multi-round feature refinement processing;

[0010] calculating a pollution state judgment result of the current monitoring period according to the matching degree of the senior pollution feature set and a preset pollution source feature template, the pollution state judgment result being used to mark a potential pollution source category and its spatial influence range;

[0011] driving a governance strategy calculation engine based on the pollution state judgment result to output a parameterized regulation instruction set for the potential pollution source category and its spatial influence range.

[0012] Preferably, the multi-level feature mapping processing on the soil original observation data stream to obtain a primary feature set associated with a pollution event includes the following steps:

[0013] performing pollution event word segmentation and entity recognition processing on the pollution material description information sequence to obtain a pollution entity word sequence and a pollution behavior word sequence;

[0014] calling a pre-constructed pollution semantic knowledge network, mapping each word in the pollution entity word sequence into an entity node, and mapping each word in the pollution behavior word sequence into a behavior node;

[0015] in the pollution semantic knowledge network, creating a semantic relationship edge between the entity node and the behavior node according to a pre-defined pollution field logical relationship, and forming a preliminary pollution semantic feature network;

[0016] performing topological structure coding processing on the preliminary pollution semantic feature network to convert the relationship between the nodes and edges in the network into a fixed-dimension information embedding vector;

[0017] In parallel, performing abnormal segment interception and regularization processing on the environmental parameter time series readings to identify and extract non-stationary change intervals in the readings;

[0018] performing multi-granularity convolution coding processing on the data in the non-stationary change intervals to capture change patterns at different time scales, and generating the state coding vector;

[0019] splicing the information embedding vector and the state coding vector to constitute the primary feature set.

[0020] Preferably, the primary feature set is input into an iterative evolution network for multi-round feature refinement processing, the iterative evolution network dynamically adjusts the feature extraction dimension of the next round according to the comparison result of the refined feature set output by the previous round and the historical pollution pattern knowledge base, and generates a senior pollution feature set after the multi-round feature refinement processing, including the following steps:

[0021] In the first round of processing of the iterative evolution network, the primary feature set is decomposed into semantic branch features and environmental branch features;

[0022] The semantic branch features and the environmental branch features are respectively input into different feature refinement channels of the network, each feature refinement channel comprising a plurality of serially connected feature transformation layers for extracting channel-specific deep features;

[0023] The channel-specific deep features output from different feature refinement channels are subjected to inter-channel feature cross and compensation operations to generate a first-round refined feature set;

[0024] The first-round refined feature set is subjected to similarity retrieval and comparison with the typical pollution feature patterns stored in the historical pollution pattern knowledge base to obtain a feature difference vector;

[0025] In the subsequent rounds of processing of the iterative evolution network, the feature difference vector generated in the previous round is weighted and fused with the derived features of the primary feature set to be processed in the current round to adjust the attention weight distribution of each feature refinement channel;

[0026] The processes of feature refinement, cross compensation and comparison with the historical knowledge base are repeatedly performed until a preset iteration round or a feature convergence condition is reached, and finally the senior pollution feature set is output.

[0027] Preferably, the pollution state determination result of the current monitoring period is calculated according to the matching degree of the senior pollution feature set and a preset pollution source feature template, including the following steps:

[0028] A pollution source feature template library containing multiple known pollution source categories is constructed, each pollution source feature template being composed of a group of discriminative feature vectors;

[0029] The multidimensional similarity measure between the senior pollution feature set and each pollution source feature template in the pollution source feature template library is calculated to obtain a similarity measure spectrum;

[0030] The similarity measure spectrum is subjected to peak detection and pattern analysis to screen out candidate pollution source categories exceeding a confidence threshold;

[0031] For each candidate pollution source category, a feature component related to a pollution diffusion physical model is inversed from the high-level pollution feature set to estimate the spatial influence intensity and range of the candidate pollution source category on the monitoring point;

[0032] The candidate pollution source category and its corresponding spatial influence intensity and range are integrated to generate a structured pollution state judgment result.

[0033] Preferably, after generating the pollution state judgment result of the current monitoring period, a pollution traceability path generation step is further included:

[0034] The candidate pollution source category and spatial influence range in the pollution state judgment result are analyzed;

[0035] Starting from the candidate pollution source category and ending at the monitoring point, a reverse path search is performed in a pre-set geographical information and pollution propagation knowledge graph;

[0036] Combining the wind direction and hydrological data in the environmental parameter time series readings, the multiple potential propagation paths searched are subjected to feasibility assessment and weight assignment;

[0037] The highest-weighted pre-set number of potential propagation paths are integrated to generate a visual pollution traceability path graph, which is used to describe the possible propagation trajectory of the pollutant from the source to the monitoring point.

[0038] Preferably, the pollution state judgment result is used to drive a governance strategy calculation engine to output a parameterized regulation instruction set for the potential pollution source category and its spatial influence range, including the following steps:

[0039] The pollution state judgment result and the pollution traceability path graph are submitted to a governance strategy calculation engine as joint inputs;

[0040] The governance strategy calculation engine includes a governance case reasoning network, which, after receiving the joint inputs, retrieves a case set with the highest scenario similarity from the historical successful governance cases stored therein;

[0041] The governance intervention measures adopted by the retrieved case set with the highest scenario similarity are decomposed and reorganized to form a basic intervention measure set;

[0042] The spatial influence range in the pollution state judgment result is used to perform spatial adaptability adjustment on each measure in the basic intervention measure set, and the pollution traceability path graph is used to perform time sequence planning on the implementation nodes and intensity of the measures to generate a preliminary governance strategy sequence;

[0043] inputting the preliminary governance strategy sequence into a strategy simulation module to predict the pollution concentration variation trend after the strategy execution based on a simplified pollution migration model;

[0044] if the predicted trend does not meet the preset governance target, automatically optimizing and adjusting the measure parameters in the preliminary governance strategy sequence until the simulation prediction result meets the standard, and then converting the adjusted strategy sequence into the parameterized regulation instruction set.

[0045] Preferably, the governance case reasoning network in the governance strategy calculation engine performs the steps of case retrieval and strategy generation, including:

[0046] building a case representation model taking the pollution scene feature vector, the governance measure vector and the governance effect vector as a triple;

[0047] converting the joint input into a current scene feature vector, and calculating the Euclidean distance between the current scene feature vector and the pollution scene feature vectors of all historical cases in the case representation model;

[0048] selecting a plurality of historical cases with the smallest Euclidean distance to form a neighboring case set;

[0049] extracting the governance measure vector of each historical case in the neighboring case set, and analyzing the commonalities and differences of the governance measure vector in different environmental dimensions;

[0050] based on the commonalities and differences, using a weighted voting and feature interpolation method to synthesize a new governance measure vector for the current scene;

[0051] decoding the new governance measure vector into a specific executable governance action description as the basic intervention measure set.

[0052] Preferably, after the output of the parameterized regulation instruction set for the potential pollution source category and its spatial influence range, a governance process closed-loop learning step is further included:

[0053] collecting a new round of soil original observation data stream fed back by the monitoring nodes after the execution of the parameterized regulation instruction set;

[0054] performing the same multi-level feature mapping and iterative evolution processing on the new round of soil original observation data stream to obtain a pollution feature set after governance;

[0055] calculating a feature change vector between the pollution feature set after governance and the high-level pollution feature set before governance;

[0056] taking the pollution state judgment result corresponding to the current governance task, the parameterized regulation instruction set and the feature change vector together as a new governance experience sample;

[0057] Add the new governance experience sample to the historical pollution pattern knowledge base and governance case reasoning network for updating network parameters and knowledge base content, realizing self-evolution of the monitoring and governance system.

[0058] Preferably, before receiving the soil original observation data stream from the monitoring node, the method further comprises a step of adaptive deployment planning of the monitoring node:

[0059] Obtain historical pollution distribution maps, geographic information data and meteorological statistical data of the target monitoring area;

[0060] Analyze the historical pollution distribution maps using a spatial clustering algorithm to identify potential high-incidence pollution areas;

[0061] Simulate possible diffusion paths of pollutants in the potential high-incidence pollution areas in combination with the geographic information data and meteorological statistical data;

[0062] Calculate optimal monitoring node layout coordinates at key nodes of the simulated diffusion paths and core locations of the potential high-incidence pollution areas;

[0063] Generate a monitoring network deployment scheme based on the optimal monitoring node layout coordinates, which is used to guide the installation location and density of physical monitoring equipment to ensure the coverage and representativeness of the soil original observation data stream.

[0064] Preferably, the application further comprises an artificial intelligence-based soil pollutant monitoring system, which comprises a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to realize the artificial intelligence-based soil pollutant monitoring method as described above.

[0065] Compared with the prior art, the application has the following beneficial effects:

[0066] By constructing an iterative evolution network, the output of each round of feature refinement is compared with the historical pollution pattern knowledge base in real time, and the feature extraction dimension of the next round of network is dynamically adjusted according to the similarity difference. This makes the feature extraction process no longer a fixed calculation of forward transmission, but forms a closed-loop feedback system of "analysis-comparison-adjustment". The system can focus on more relevant or more abnormal feature dimensions according to the potential pattern implied by the current data stream, thereby optimizing its internal representation in real time during the reasoning process. This mechanism improves the recognition accuracy and self-adaptive ability of the model to complex, hidden or new pollution patterns, so that the monitoring system can continuously optimize itself with experience accumulation.

[0067] The pollution description information sequence and the environmental parameter time sequence reading are branched and isomodulated by using multi-level feature mapping processing. The text type description information is converted into information embedding vectors by semantic understanding, and the abstract attributes such as the category and the form of the pollutants are captured. The continuous environmental parameter reading is converted into state coding vectors by time sequence modeling, and the dynamic change law is described. After the high-level semantic and low-level signal are respectively deep-converted, the unified primary feature set is formed. The inherent characteristics of the heterogeneous data source are respected and utilized, so that the subsequent analysis can be established on the basis of the features with less information loss and more explicit physical meaning, and the high-quality joint representation is provided for accurately associating the pollution events and the causes. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The working principle diagram of the soil pollutant monitoring method based on artificial intelligence is described.

[0069] Figure 2 The flowchart for generating the pollution state determination result is described.

[0070] Figure 3 The flowchart for case retrieval and strategy generation of the governance case reasoning network is described.

[0071] Figure 4 The correlation analysis diagram of the adjacent case governance cycle and the cadmium concentration reduction rate is described.

[0072] Figure 5 The multi-dimensional feature value comparison diagram before and after the soil pollution governance is described. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described 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. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0074] Please refer to Figure 1The application provides a soil pollutant monitoring method and system based on artificial intelligence. The method comprises: receiving a soil original observation data stream from a monitoring node deployed in a monitoring area, the data stream containing a text information sequence describing a pollutant and time sequence readings of environmental parameters such as temperature, humidity, pH value, and specific ion concentration. The method first performs a multi-level feature mapping process on the soil original observation data stream. The process converts the pollutant description information sequence in text form into a machine-understandable information embedding vector through semantic analysis, and converts the environmental parameter time sequence readings into a state coding vector reflecting state changes through signal analysis. The two vectors together constitute a primary feature set associated with a pollution event.

[0075] The primary feature set is input into an iterative evolution network for multi-round feature refinement processing. After each round of processing, the network compares the refined feature set output by it with a pre-constructed historical pollution pattern knowledge base, uses the difference information generated by the comparison as feedback, dynamically adjusts the feature extraction focus and dimension of the network in the next round of processing, and generates a high-level pollution feature set that can more deeply reflect the nature of pollution after multiple iterations. According to the matching degree calculation of the high-level pollution feature set and a plurality of preset pollution source feature templates, the pollution state determination result of the current monitoring period is generated by analyzing the matching result. The result not only marks the potential pollution source category, but also estimates its spatial influence range. Based on the pollution state determination result, a governance strategy calculation engine is driven, which automatically generates and outputs a complete set of specific executable parameterized control instructions for the identified pollution source category and its influence range, thereby completing the closed loop from monitoring to governance suggestion.

[0076] Embodiment 1: Perform pollution event tokenization and entity recognition on the pollution description information sequence, distinguish the pollution entity word sequence describing the pollution substance and the pollution behavior word sequence describing the pollution behavior or state. Call a pre-constructed pollution semantic knowledge network containing the entity and behavior concepts and their relationships in the pollution field, map each word in the identified pollution entity word sequence to an entity node in the semantic network, and map each word in the pollution behavior word sequence to a behavior node. Within the pollution semantic knowledge network, according to the pre-defined logical relationship in the pollution field, create semantic relationship edges between related entity nodes and behavior nodes, thereby forming a preliminary structured pollution semantic feature network. Perform topology structure encoding processing on this preliminary network, for example using graph neural network technology, to convert the complex relationship between nodes and edges in the network into a fixed-dimensional information embedding vector. Process the environmental parameter time series readings, identify and intercept the non-stationary change intervals in the reading sequence through algorithm, which may correspond to the occurrence of pollution events. Perform multi-granularity convolutional coding processing on the data in the intercepted non-stationary change interval, use convolution kernels of different scales to capture the change patterns of the data at different time scales such as short-term and medium-term, and finally generate a state coding vector that condenses the time series change features. Perform splicing operation on the information embedding vector obtained by the semantic path and the state coding vector obtained by the environmental parameter path to form the primary feature set as the input for subsequent processing.

[0077] In a specific implementation, the step of performing multi-level feature mapping processing on the soil original observation data stream is described through an instance of a specific monitoring node. The soil original observation data stream uploaded by the monitoring node contains a sequence of pollution description information "abnormal increase in chromium ion concentration in soil on the east side of the factory area, accompanied by yellow precipitate", and a set of continuously monitored environmental parameter time series, including soil pH value sequence and temperature sequence. The pollution event word segmentation and entity recognition processing operates on the sequence of pollution description information, divides the sequence into lexical units and identifies the types, obtaining the sequence of pollution entity words "chromium ion" and "yellow precipitate", and the sequence of pollution behavior words "abnormal increase in concentration" and "accompanied by". The pre-constructed pollution semantic knowledge network is called, which contains entity nodes such as "heavy metal", "compound", "precipitation", and behavior nodes such as "concentration change" and "associated phenomenon". The "chromium ion" and "yellow precipitate" are mapped to the entity nodes in the pollution semantic knowledge network, and the "abnormal increase in concentration" and "accompanied by" are mapped to the behavior nodes. Within the pollution semantic knowledge network, according to the pre-defined logical relationship in the pollution field, a "performance" semantic relationship edge is created between the "chromium ion" entity node and the "abnormal increase in concentration" behavior node, and two "subject-accompanying" semantic relationship edges are created between the "chromium ion" entity node and the "yellow precipitate" entity node through the "accompanied by" behavior node, forming a preliminary pollution semantic feature network. The topological structure coding processing is performed on the preliminary pollution semantic feature network, and the node features and edge relationships are aggregated and propagated using a graph convolution network, finally outputting a 128-dimensional information embedding vector. The single-layer propagation formula of the graph convolution network is represented as:

[0078]

[0079] wherein the symbol represents the adjacency matrix of the pollution semantic feature network after adding the self-loop, the symbol represents the corresponding degree matrix, the symbol represents the feature matrix of all nodes in the layer, the symbol represents the trainable weight matrix of the layer, and the symbol ​represents a nonlinear activation function. In parallel, the pH value sequence in the environmental parameter time series reading accepts abnormal segment interception and regularization processing, and a mutation of the statistical quantity is calculated through a sliding window to identify a non-stationary change interval of 8 hours, in which the pH value drops from 7.2 to 5.8. The multi-granularity convolutional encoding processing acts on the pH value sequence and the temperature sequence in the non-stationary change interval, uses three one-dimensional convolution kernels with widths of 3, 6, and 12 to slide in the time dimension, extracts local and cross-period change patterns, and after the outputs of the three convolutional layers are pooled and spliced, a 64-dimensional state encoding vector is generated. The information embedding vector and the state encoding vector are directly spliced to form a 192-dimensional primary feature set.

[0080] In some embodiments, the pollution event word segmentation and entity recognition processing adopts a sequence labeling model based on the combination of a bidirectional long short-term memory network and a conditional random field. The model uses a large amount of labeled pollution event text during training, and can accurately identify pollutant entities and their state behaviors. Predefined pollution domain logical relationships are stored in a relationship rule base, which includes various relationship types such as “cause”, “show”, “suppress”, “transform”, and applicable entity and behavior combination conditions. The multi-granularity convolutional encoding processing uses dilated convolution to expand the receptive field, and convolutional layers with different dilation rates are set in parallel to capture the dependence relationship of environmental parameter time series readings at different intervals in the non-stationary change interval. In addition to the graph convolutional network, the topological structure encoding processing can also use the graph attention network to assign different aggregation weights to different adjacent nodes in the pollution semantic feature network through the attention mechanism. Optionally, the information embedding vector and the state encoding vector are each subjected to layer normalization processing before splicing to balance the numerical scales of the feature vectors from different sources. The dimension of the primary feature set is adjusted through a learnable linear projection layer to adapt to the input requirements of the subsequent iterative evolution network.

[0081] In the first round of processing of the iterative evolution network, the spliced primary feature set is decomposed into semantic branch features and environmental branch features. The semantic branch features and the environmental branch features are respectively input into different feature refining channels in the network, and each channel internally includes a plurality of serially connected feature transformation layers for in-depth extraction of specific deep features in the respective channel. Then, the channel-specific deep features output from the different feature refining channels are subjected to inter-channel feature cross and compensation operations, for example, through an attention mechanism or a feature fusion layer, so that the semantic information and the environmental information are mutually supplemented and corrected to generate a first round of refined feature set. The first round of refined feature set is subjected to similarity retrieval and comparison with the typical pollution feature patterns stored in the historical pollution pattern knowledge base, and a feature difference vector reflecting the difference between the current features and the historical patterns is calculated. In the subsequent rounds of processing of the iterative evolution network, the feature difference vector generated in the previous round is weighted and fused with the new feature representation derived from the primary feature set to be processed in the current round, so as to adjust the attention weight distribution of each feature refining channel in the network and guide the network to pay more attention to the feature dimensions with greater difference from the historical patterns in the next round. The above-mentioned feature refining, cross-compensation and comparison with the historical knowledge base are repeatedly performed until a preset iteration round or a feature convergence condition is reached, and finally a stable and highly discriminative advanced pollution feature set serving as the final criterion is output.

[0082] In a specific implementation, the process of inputting the primary feature set into the iterative evolution network for multi-round feature refining is described through a specific example, in which the primary feature set is a 192-dimensional feature vector. In the first round of processing of the iterative evolution network, the 192-dimensional primary feature set is decomposed into semantic branch features and environmental branch features, the semantic branch features corresponding to the first 128-dimensional information embedding vector and the environmental branch features corresponding to the last 64-dimensional state encoding vector. The semantic branch features are input into the semantic feature refining channel of the iterative evolution network, and the semantic feature refining channel includes three serially connected linear transformation layers and activation functions. The environmental branch features are input into the environmental feature refining channel of the iterative evolution network, and the environmental feature refining channel includes two serially connected one-dimensional convolution layers and pooling layers. These feature transformation layers are used to extract channel-specific deep features of the semantic branch and the environmental branch, respectively. The channel-specific deep features output from the semantic feature refining channel and the channel-specific deep features output from the environmental feature refining channel are sent to a feature cross module, which performs inter-channel feature cross and compensation operations by calculating a bilinear interaction matrix and applying soft attention weights to generate a fused 256-dimensional vector as the first round of refined feature set.

[0083] In a specific implementation, the first round of refined feature set is compared with the historical pollution pattern knowledge base, which stores the feature vectors of three typical pollution feature patterns, i.e., "electroplating wastewater chromium pollution", "leather tanning chromium pollution", and "natural background chromium anomaly". The cosine similarity between the first round of refined feature set and each typical pollution feature pattern vector in the knowledge base is calculated, obtaining the similarity sequence [0.65, 0.82, 0.31].

[0084]

[0085] wherein the symbol represents the calculated feature difference vector, the symbol represents the refined feature set vector output in the current round, the symbol represents the maximum value 0.82 in the similarity sequence, and the symbol represents the "leather tanning chromium pollution" typical pollution feature pattern vector corresponding to the maximum value. The physical meaning of is the weighted feature vector of the historical typical pollution pattern, the symbol is the weight coefficient, which is used to quantify the matching degree of the current monitoring scene and the historical typical pattern. The higher the matching degree, the greater the weight, and the stronger the contribution of the historical pattern to the calibration of the current feature. The product of the two is essentially the historical reference feature vector based on the matching degree correction. Therefore, The calculation logic of

[0086] is the deviation quantification of the current refined feature and the corrected historical reference feature. The deviation direction is determined by the positive and negative of the difference value of each dimension of the vector, and the deviation amplitude is reflected by the absolute value of the difference value. Finally, the goal of calibrating the current feature extraction direction through historical experience is achieved. The calculated feature difference vector reflects the deviation direction and amplitude of the current monitoring feature and the historical "leather tanning chromium pollution" pattern in the feature space.In a specific implementation, the iterative evolution network enters a second round of processing, and the feature difference vector generated in the first round is weighted and fused with the derived features of the primary feature set to be processed in the current round. The derived features of the primary feature set are 128-dimensional projection vectors obtained by a learnable linear projection layer. The weighted fusion is realized through a gating mechanism, and the feature difference vector controls the fusion weight to dynamically adjust the attention weight distribution within the semantic feature refinement channel and the environmental feature refinement channel, so that the network pays more attention to the feature dimensions with larger values in the feature difference vector in the second round. The adjusted feature refinement channel transforms the weighted input features again, and the feature cross and compensation operations are repeated to generate a second round of refined feature set. The second round of refined feature set is again subjected to similarity retrieval and comparison with the historical pollution pattern knowledge base to generate a new feature difference vector. The above iterative process is repeated, and after reaching the preset number of 5 iterations, the final output of the fifth round of refined feature set is the high-level pollution feature set.

[0087] In some embodiments, the inter-channel feature cross and compensation operation is realized by using a multi-head attention mechanism. The channel-specific deep features of the semantic branch and the environmental branch are respectively used as the query and key-value pairs, and the information is fully interacted and compensated through the calculation of multiple attention heads. The weighted fusion of the feature difference vector and the derived feature can also be realized by element-wise addition, then generating a modulation coefficient through a sigmoid function, and then scaling the original derived feature with the modulation coefficient. The convergence condition of the iterative evolution network is set as the L2 norm of the feature difference vector of the refined feature set being less than a set threshold. When the L2 norm of the feature difference vector calculated in a certain round is lower than the threshold 0.05, even if the maximum number of iterations has not been reached, the iterative evolution network is terminated in advance and outputs the current refined feature set as the high-level pollution feature set. Optionally, the historical pollution pattern knowledge base supports dynamic updating. When the high-level pollution feature set is confirmed to represent a new pollution pattern after subsequent judgment, this high-level pollution feature set can be added to the historical pollution pattern knowledge base as a new typical pollution feature pattern. Optionally, the feature refinement channels of each round of the iterative evolution network can share part of the weight parameters, and the residual connection is used to ensure the effective propagation of the gradient. This design can increase the depth of the network while maintaining the stability of the training.

[0088] Embodiment 3: see Figure 2, a pollution source feature template library containing multiple known pollution source categories is constructed, each template being composed of a set of feature vectors capable of effectively identifying the pollution source of that category. A multi-dimensional similarity measure between the advanced pollution feature set and each template in the pollution source feature template library is calculated, resulting in a similarity measure spectrum reflecting the matching degree with each category. Peak detection and pattern analysis are performed on the similarity measure spectrum, and those candidate pollution source categories corresponding to peaks exceeding a pre-set confidence threshold are screened out. For each screened candidate pollution source category, feature components related to the pollution diffusion physical model are retrieved from the advanced pollution feature set, and based on these components and a simple diffusion model, the spatial impact intensity and approximate range of the candidate pollution source category on the current monitoring point are estimated. The structured pollution state judgment result is generated by integrating all candidate pollution source categories and their respective spatial impact intensity and range information.

[0089] After generating the pollution state judgment result, the pollution traceability path generation step is performed. The candidate pollution source categories and estimated spatial impact range information contained in the pollution state judgment result are analyzed. Starting from the possible source location of the candidate pollution source category and ending at the current monitoring point, a reverse path search is performed in a pre-set geographical information and pollution propagation knowledge graph, which contains geographical correlation information such as terrain, water system, and pipe network. Combining the hydro-meteorological data such as wind direction and water flow speed recorded in the environmental parameter time series readings, the physical feasibility of the multiple potential propagation paths searched is evaluated, and each path is assigned a weight value. The pre-set number of potential propagation paths with the highest weight after evaluation are integrated and visually expressed, generating a pollution traceability path map that describes the possible propagation trajectory and direction of pollutants from the potential source to the monitoring point.

[0090] In a specific implementation, the process of generating pollution state determination results according to the matching degree calculation of the high-level pollution feature set and the preset pollution source feature template is described through a specific example, where the high-level pollution feature set is a 256-dimensional feature vector, and the preset pollution source feature template library contains the feature templates of three known pollution source categories: "chemical plant benzene series leakage", "gas station oil product leakage", and "pesticide overuse". Calculate the multidimensional similarity measure between the high-level pollution feature set and each pollution source feature template in the pollution source feature template library, and use a combination of weighted cosine similarity and Manhattan distance to calculate a three-dimensional similarity measure spectrum, with specific values of [0.76, 0.45, 0.89]. Perform peak detection and pattern analysis on the similarity measure spectrum, set the confidence threshold to 0.70, and select "chemical plant benzene series leakage" and "pesticide overuse" as candidate pollution source categories that exceed the confidence threshold. For the "pesticide overuse" candidate pollution source category, reverse the feature components related to the pollution diffusion physical model from the 256-dimensional high-level pollution feature set, specifically extract the feature dimensions corresponding to soil porosity and adsorption coefficient, estimate the spatial influence intensity and range of the "pesticide overuse" candidate pollution source category on the monitoring point, and the estimation result is output as an influence intensity of 0.8 (normalized value) and an influence radius of 125 meters. Integrate the "chemical plant benzene series leakage" and "pesticide overuse" candidate pollution source categories and their respective estimated spatial influence intensity and range to generate a structured pollution state determination result, which records the candidate pollution source category, confidence, influence intensity, and influence radius information in JSON format.

[0091] In a specific implementation, the pollution traceability path generation step is subsequently initiated to resolve the "pesticide overuse" candidate pollution source category recorded in the pollution state determination result with the 125-meter spatial impact range information. With the possible source location corresponding to the "pesticide overuse" candidate pollution source category, i.e., the center point of a piece of farmland in the upwind direction of the monitoring area, as the path starting point and the current monitoring point coordinates as the path endpoint, a reverse path search is performed in a pre-set geographical information and pollution propagation knowledge graph, which contains regional elevation models, surface water flow direction data, underground water flow field simulation data, and road pipe network information. In combination with the wind direction data in the environmental parameter time series readings, the average wind direction during the monitoring period is southeast, the wind speed data is 2.1 meters per second, and the soil moisture time series readings in the hydrological data show that there has been a recent irrigation event. The multiple potential propagation paths obtained by the reverse path search are subjected to feasibility assessment and weight assignment, and the assessment basis includes the angle between the path direction and the wind direction, the soil permeability data, and whether it passes through a known hydrological channel. Each path is assigned a feasibility weight value between 0 and 1. The two potential propagation paths with the highest weights after feasibility assessment are integrated, one path is described as an "atmospheric deposition dominated downwind diffusion path", and the other path is described as an "irrigation water lateral migration path", generating a visual pollution traceability path graph, which describes the two possible propagation trajectories of the pollutant from the farmland center point to the monitoring point in the form of arrows and contour lines superimposed on the geographical base map.

[0092] In some embodiments, the multi-dimensional similarity measure employs a metric learning based method, which maps the high-level pollution feature set and the pollution source feature template to the same metric space through a trainable small neural network, and then calculates the distance, with the reciprocal of the distance as the similarity measure value. The estimation formula of the spatial impact intensity and range can be represented as an analytical model based on feature components, for example:

[0093]

[0094] wherein the symbol represents the estimated spatial impact radius, with the dimension of length (meters), the symbol represents a diffusion attenuation coefficient related to soil type, with the unit of (meters), which is a pre-defined calibration coefficient according to the physical properties of the target region's soil texture, porosity, etc. Sandy soil has strong air permeability and fast pollutant diffusion, with a value range of 15-25m; loamy soil with a value range of 8-15m; clay soil has poor air permeability and slow diffusion, with a value range of 3-8m, which is obtained by statistical calibration of regional soil survey data to ensure the dimension, the symbol This represents the pollution intensity feature component derived from the advanced pollution feature set. It is dimensionless, ranging from 0 to 1, and is obtained by aggregating the pollutant concentration exceedance multiples, pollution duration, and other dimensions from the advanced pollution feature set. A larger value indicates higher pollution intensity. The symbol is... This represents the soil blocking characteristic component derived from the set of advanced pollution characteristics. It is dimensionless, ranging from 0.1 to 1.0, and is calculated from dimensions such as soil organic matter content and clay mineral ratio within the advanced pollution characteristic set. A higher value indicates a stronger adsorption and blocking effect of the soil on pollutants. The dimensional logic of the formula... For a dimensionless term, and for a term with a dimension of (meters) After multiplication, the final output is... The dimension of the graph is (meter), which conforms to the physical definition of the spatial influence radius. The construction of the geographic information and pollution propagation knowledge graph adopts multi-source data fusion technology, integrating satellite remote sensing imagery, geological survey reports, and municipal drainage network vector data. Nodes in the knowledge graph represent geographic entities or pollution sources, and edges represent potential pathways or influence relationships of material migration.

[0095] In some embodiments, an adaptive peak-finding algorithm is used for peak detection of the similarity metric spectrum. The algorithm automatically determines the peak position and boundary based on the local gradient changes of the similarity metric spectrum, avoiding the failure of fixed thresholds when the overall distribution of the metric spectrum shifts. The feasibility assessment and weighting process introduces multi-criteria decision analysis, assigning different importance weights to multiple assessment criteria such as wind direction matching, hydrological connectivity, and terrain slope, and then weighted summing to obtain the total weight value for each potential propagation path. Optionally, the structured output format of the pollution status determination results can be extended; in addition to JSON format, XML or Protocol Buffers formats can also be used to adapt to the integration needs of different downstream systems. The visualization rendering of the pollution tracing path map supports layered control; users can choose to show or hide different information layers such as terrain, water system, and path.

[0096] Optionally, the reverse path search algorithm can employ either the A* search algorithm or Dijkstra's algorithm. These algorithms assign cost values ​​related to pollutant migration resistance to the edges in the geographical information and pollution propagation knowledge graph, searching for the lowest cumulative cost path from the endpoint to the starting point. The weighting process can incorporate a random forest classifier, training it with path features and historically confirmed pollution propagation cases to predict the feasibility probability of the current path as the weights.

[0097] Example 4: See Figure 3The pollution state determination result and the pollution traceability path graph are submitted to a governance strategy calculation engine as joint inputs. The engine includes a governance case reasoning network that, after receiving the joint inputs, retrieves a case set with the highest similarity to the current input scenario from a historical successful governance case library stored therein. The governance intervention measures adopted by the retrieved high-similarity cases are analyzed, decomposed, and reorganized to extract their core measures to form a basic intervention measure set. The detailed spatial influence range data in the pollution state determination result are used to perform spatial adaptability adjustment on each measure in the basic intervention measure set to determine the area and intensity gradient of the measure implementation, and the pollution traceability path graph is used to perform timing planning on the implementation nodes and timing of the measures to generate a preliminary governance strategy sequence with spatial and temporal dimensions. The preliminary governance strategy sequence is input to a strategy simulation module, which predicts the change trend of the pollutant concentration in the monitoring area after the strategy sequence is executed based on a simplified pollution migration and reaction model. If the simulated change trend does not meet the preset governance target, the measure parameters in the preliminary governance strategy sequence are automatically optimized and adjusted by an optimization algorithm, and the simulation prediction is performed again until the simulation prediction result meets the governance target requirement, at which time the final strategy sequence after adjustment and optimization is converted into specific, parameterized control instruction sets that can be issued to the execution unit.

[0098] The specific steps of case retrieval and strategy generation by the governance case reasoning network in the governance strategy calculation engine include: constructing a case representation model with a pollution scene feature vector, a taken governance measure vector, and a final governance effect vector as a triple. The joint input is converted into a current scene feature vector, and the Euclidean distance between the current scene feature vector and the pollution scene feature vectors of all historical cases in the case representation model is calculated. The multiple historical cases with the smallest Euclidean distance are selected to form a neighboring case set. The governance measure vectors of each historical case in the neighboring case set are extracted, and the commonalities and differences of the measure vectors in different environmental dimensions are analyzed. Based on the analyzed commonalities and differences, a new governance measure vector for the current scene is synthesized by using the weighted voting and feature interpolation methods. The new governance measure vector is converted into a specific executable governance action description by a decoder, and the description set is used as the basic intervention measure set.

[0099] In a specific implementation, the process of driving the governance strategy calculation engine to output a set of parameterized regulation instructions based on the pollution status determination result is illustrated through a specific example, in which the pollution status determination result identifies the "soil cadmium pollution" candidate pollution source category, the spatial impact range is estimated as a circular area with a radius of 80 meters centered on the monitoring point, and the pollution trace path diagram depicts the trajectory of the pollutant spreading from the suspected electroplating workshop in the northwest direction along the groundwater flow. The pollution status determination result and the pollution trace path diagram are submitted to the governance strategy calculation engine as joint inputs. After receiving the joint inputs, the governance case reasoning network in the governance strategy calculation engine converts the joint inputs into a current scene feature vector containing pollution type, impact radius, hydrogeological conditions, and other attributes. The governance case reasoning network retrieves the case set with the highest scenario similarity from its stored historical successful governance case library. Each case in the historical successful governance case library is stored in the form of a pollution scene feature vector, a governance measure vector, and a governance effect vector triplet. The Euclidean distance between the current scene feature vector and the pollution scene feature vector of all historical cases in the historical successful governance case library is calculated, and the three historical cases with the smallest Euclidean distance are selected to form the neighboring case set. The governance measure vector of each historical case in the neighboring case set is extracted, and each governance measure vector is a multi-dimensional vector whose dimensions encode specific measure parameters such as "type of applied passivation agent", "amount of applied passivation agent", "irrigation control mode", "whether to conduct deep plowing", etc. Analyze the commonalities and differences of these governance measure vectors in different environmental dimensions, for example, in the soil pH value dimension, three cases all use the "apply lime" measure, but there are differences in the amount of application. Based on the analyzed commonalities and differences, a new governance measure vector is synthesized using the weighted voting and feature interpolation methods, weighted voting is used to determine discrete measure options, and feature interpolation is used to determine continuous measure parameters. The new governance measure vector is decoded into specific executable governance action descriptions, and the decoding process is completed through a predefined governance measure codebook, generating a set of basic intervention measures. Refer to Table 1.

[0100] Table 1: Governance measures and effects table of the neighboring case set

[0101]

[0102] Spatial adaptation adjustments were made to each measure in the basic intervention measure set using the spatial impact range determined by the pollution status assessment results. Within a circular impact area with a radius of 80 meters, differentiated measure intensities were implemented according to the gradient of distance from the pollution source, with the intensity of measures in the central area at 100% of the design value and the intensity in the peripheral area at 70% of the design value. A pollution causal path map was used to plan the implementation nodes and intensities of the measures in a time sequence. The plan was to begin applying a passivating agent on the first day after the pollution source was cut off, and to begin controlled irrigation on the third day, generating a preliminary treatment strategy sequence. This preliminary treatment strategy sequence was input into a strategy simulation module, which predicted the trend of available cadmium concentration changes at monitoring points and surrounding grid points over 30 days after the strategy implementation, based on a simplified pollution migration model. Simulation prediction results showed that on the 30th day, the predicted concentration value in the core area was still higher than the preset treatment target threshold. Automatic optimization adjustments were made to the measure parameters in the preliminary treatment strategy sequence, using a particle swarm optimization algorithm to optimize key parameters such as "lime application rate" and "irrigation days." The optimization process involves repeated simulation predictions until the simulation prediction results satisfy the condition that the concentration at all grid points is below the target threshold on day 30. At this point, the adjusted final strategy sequence is transformed into a set of parameterized control instructions, which includes specific material delivery coordinates, delivery amounts, irrigation on / off schedules, and other instructions.

[0103] In some embodiments, the process of synthesizing a new governance measure vector by weighted voting and feature interpolation can be represented by the following formula:

[0104]

[0105] Where: symbol Represents the vector of synthesized new governance measures, symbol This indicates the first value calculated based on the inverse Euclidean distance. The weights of the nearest neighbor cases, with symbols Indicates the first A vector of governance measures for each neighboring case, symbolic. Indicates the number of cases in the nearest neighbor set, symbol This represents a small random disturbance coefficient, with the sign... and These represent the maximum and minimum vectors of governance measures for all neighboring cases in each dimension. The retrieval process of the governance case inference network can employ locality-sensitive hashing to accelerate similarity retrieval in a massive historical database of successful governance cases. The simplified pollution migration model in the strategy simulation module can consider a one-dimensional convection-dispersion equation and linear adsorption equilibrium, with model parameters obtained from the feature components inverted from the pollution state determination results.

[0106] In some embodiments, the spatial adaptation can be combined with a geographic information system, superimposing the circular influence area on a soil type zoning map, fine-tuning the measure parameters for different soil type sub-regions. The timing plan takes into account the agricultural calendar and weather forecast data, avoiding the implementation of measures other than irrigation regulation during periods of high probability of precipitation. The objective function of the automatic optimization adjustment can be set as the weighted sum of multiple goals such as governance compliance time, governance cost, and ecological disturbance degree, for multi-objective optimization. Optionally, the output format of the parameterized set of control instructions is compatible with automated execution equipment, which can directly drive variable rate fertilizer applicators, irrigation control valves, and other equipment. The simulation prediction results of the governance strategy sequence can be visualized and output in the form of an animation showing the temporal and spatial changes in governance effect as the concentration contour map evolves over time. Optionally, the decoding process of the basic intervention measure set can have interpretability, mapping each dimension in the new governance measure vector to a human-readable natural language description, and attaching an explanation of which or which neighboring cases it originated from. The strategy simulation module can run Monte Carlo simulations to consider the uncertainty of model parameters and give confidence intervals for concentration prediction values.

[0107] Referring to Figure 4 The correspondence between the governance period (column, left axis, unit: day) and the cadmium concentration reduction rate (line, right axis, unit: %) of three neighboring governance cases (Case-23, Case-41, Case-67) is presented in the form of a two-axis chart. Specifically, the governance period of Case-23 is 30 days, corresponding to a cadmium concentration reduction rate of 60%; the governance period of Case-41 is 45 days, corresponding to a cadmium concentration reduction rate of 75%, which is the highest among the three cases; the governance period of Case-67 is 30 days, corresponding to a cadmium concentration reduction rate of 55%. This chart directly reveals the correlation between the governance period and the cadmium pollution governance effect. Within a certain range, the extension of the governance period can improve the cadmium concentration reduction rate, and also reflects the differences in governance efficiency of different combinations of governance measures.

[0108] After the output parameterized regulation instruction set is executed, the system collects a new round of soil original observation data stream monitored by the nodes after the execution of the instructions. The same multi-level feature mapping and iterative evolution process is performed on the new data stream to obtain a feature set reflecting the state of the soil after treatment. A feature change vector between the feature set after treatment and the high-level pollution feature set obtained before treatment is calculated, which quantifies the feature level change brought by the treatment measures. The pollution state judgment result corresponding to the complete treatment task, the output parameterized regulation instruction set and the calculated feature change vector are jointly used to form a new treatment experience sample. The new treatment experience sample is added to the historical pollution pattern knowledge base and the treatment case reasoning network as training data, which is used to update the internal parameters of the network and the storage content of the knowledge base, so as to realize the self-evolution and performance improvement of the whole monitoring and treatment system when dealing with new problems.

[0109] Before the system starts initial monitoring, a self-adaptive deployment planning step of the monitoring nodes is performed. The historical pollution distribution map of the target monitoring area, detailed geographic information data and long-term meteorological statistical data are obtained. The spatial clustering algorithm is used to analyze the historical pollution distribution map to identify potential areas with high pollution. Combined with the geographic information data and the meteorological statistical data, the possible diffusion paths of the pollutants under different conditions are simulated in the identified potential areas with high pollution. The optimal monitoring node deployment coordinates are calculated by an optimization algorithm at the key nodes on the simulated diffusion paths and the core positions of the potential areas with high pollution. A detailed monitoring network deployment scheme is generated according to the calculated optimal monitoring node deployment coordinates, which is used to guide the installation position and deployment density of the physical monitoring equipment in the actual environment, so as to ensure that the subsequent collected soil original observation data stream has good spatial coverage and regional representativeness from the source.

[0110] In a specific implementation, the governance process closed-loop learning step is started after the execution of the parameterized regulation instruction set is completed, and a new round of soil original observation data stream monitored by the monitoring node after the execution of the parameterized regulation instruction set is collected. The new round of soil original observation data stream includes the description information of the soil after the treatment, i.e., “the cadmium ion concentration tends to be stable after the passivation treatment”, and the new time sequence readings of the environmental parameters. The new round of soil original observation data stream is subjected to the same multi-level feature mapping processing and iterative evolution processing as the historical processing procedure. After the pollution event word segmentation, entity recognition, pollution semantic knowledge network mapping, topological structure coding, abnormal fragment interception, multi-granularity convolution coding, and multi-round refinement of the iterative evolution network, the pollution feature set after the treatment is obtained. The feature change vector between the pollution feature set after the treatment and the stored high-level pollution feature set before the treatment is calculated. The calculation method of the feature change vector is the difference between the corresponding dimension values of the two feature sets. The pollution state judgment result corresponding to the current treatment task, the parameterized regulation instruction set, and the calculated feature change vector are collectively used as a new governance experience sample, and the governance experience sample is packaged in a structured format. The new governance experience sample is added to the historical pollution mode knowledge base and the governance case reasoning network, used to update the internal parameters and storage content of the historical pollution mode knowledge base and the governance case reasoning network, realize the self-evolution of the monitoring and treatment system, and the increase of the historical pollution mode knowledge base makes the knowledge base have a new record for the “cadmium pollution passivation response” mode.

[0111] In a specific implementation, the adaptive deployment planning step of the monitoring nodes is performed in the initial deployment phase of the system, obtaining the historical pollution distribution map of the target monitoring area, geographic information data and meteorological statistical data. The historical pollution distribution map marks the location and intensity of seventeen times of soil heavy metal exceeding events in the past five years. The geographic information data includes a digital elevation model and a land use type map. The meteorological statistical data includes a ten-year average wind direction rose diagram and precipitation distribution data. The spatial clustering algorithm is used to analyze the historical pollution distribution map. The spatial clustering algorithm uses a noise-based density-based clustering method to identify three pollution-prone potential areas, each of which is composed of a group of historical pollution event points with adjacent geographical positions. Combined with the geographic information data and meteorological statistical data, the possible diffusion path of the pollutants in the pollution-prone potential area is simulated. The simulation process uses a particle diffusion model, takes the core point of the pollution-prone potential area as the release point, and combines the water flow direction derived from the digital elevation model with the ten-year average wind direction for multiple simulations. At the key nodes of the simulated diffusion path and the core positions of the pollution-prone potential area, the optimal monitoring node layout coordinates are calculated. The calculation process takes satisfying the maximum area coverage and the minimum device quantity as the objective function, and solves it through the greedy algorithm. The optimal monitoring node layout coordinates include eight specific geographical position points. According to the optimal monitoring node layout coordinates, a monitoring network deployment scheme is generated. The monitoring network deployment scheme is output as a planning document containing coordinate points, recommended device models and network topology diagrams, which is used to guide the installation location and layout density of physical monitoring devices in the target monitoring area.

[0112] In some embodiments, the feature change vector is used to fine-tune the model parameters of the iterative evolution network and the governance case reasoning network, and its parameter update process follows the gradient-based learning rule. The update formula can be expressed as:

[0113]

[0114] wherein the symbol represents the updated network parameters, the symbol represents the network parameters before updating, the symbol represents the learning rate hyperparameter, and the symbol represents the loss function constructed by the feature change vector and the expected improvement direction The gradient of the parameter set . In addition to adding new samples, the update of the historical pollution pattern knowledge base also triggers the recalculation of the cluster centers of the existing pattern vectors. When the spatial clustering algorithm analyzes the historical pollution distribution map, it can combine the intensity attribute of the pollution event for weighted clustering. High-intensity pollution event points have higher weight values in the clustering process.

[0115] In some embodiments, when calculating the optimal monitoring node deployment coordinates, the objective function can consider the coverage, risk weights of different sub-regions, and device deployment costs, and solve the Pareto optimal solution set through a multi-objective optimization algorithm, from which a deployment scheme that meets the budget constraint is selected. The network topology graph in the monitoring network deployment scheme needs to determine the communication links between the monitoring nodes and the location of the data aggregation point. The communication link planning needs to consider terrain obstruction and signal attenuation. Optionally, the addition of new governance experience samples in the governance process closed-loop learning step needs to go through a confidence evaluation link. Only when the improvement degree shown by the feature change vector after governance exceeds the preset threshold, and the subsequent continuous monitoring data is stable, the sample will be finally included in the historical pollution pattern knowledge base and the governance case reasoning network. The adaptive deployment planning step can be periodically re-executed to update the historical pollution distribution map using new monitoring data and dynamically optimize the monitoring network layout.

[0116] Referring to Figure 5 In the feature change vector analysis in the governance process closed-loop learning, the differences in feature values in the semantic features, environmental features, and spatio-temporal features before and after governance are presented. Specifically, the feature values before governance (purple column) are at a high level in each dimension, and the feature values after governance (orange column) are significantly reduced. The difference between the two constitutes the feature change (blue column). For example, the feature value before governance in the spatio-temporal feature dimension is close to 0.9, and the feature value after governance is reduced to about 0.6, corresponding to a feature change of about 0.3, which directly quantifies the improvement effect of the governance measures on each dimension of the feature. These feature change vectors will be the core part of the governance experience samples, used to update the historical pollution pattern knowledge base and the governance case reasoning network, supporting the self-evolution of the system.

[0117] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or inherent to such a process, method, article or device.

[0118] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring soil pollutants based on artificial intelligence, characterized in that, The steps are executed in the following order: Receive raw soil observation data stream from monitoring nodes, the raw soil observation data stream containing pollutant description information sequence and environmental parameter time series readings; Multi-level feature mapping processing is performed on the raw soil observation data stream to obtain a primary feature set associated with pollution events. The multi-level feature mapping processing refers to the process of converting the descriptive information sequence into an information embedding vector and the environmental parameter time series readings into a state coding vector. The initial feature set is input into the iterative evolution network for multiple rounds of feature refinement. The iterative evolution network uses the comparison result between the refined feature set output in the previous round and the historical pollution pattern knowledge base as feedback to dynamically adjust the feature extraction dimension in the next round. The first round of refined feature set is compared with the typical pollution feature patterns stored in the historical pollution pattern knowledge base to obtain the feature difference vector. In subsequent rounds of the iterative evolution network, the feature difference vector generated in the previous round is weighted and fused with the derived features of the primary feature set to be processed in the current round, thereby adjusting the attention weight distribution of each feature refinement channel. The advanced pollution feature set is generated after the aforementioned multiple rounds of feature refinement. Based on the matching degree between the advanced pollution feature set and the preset pollution source feature template, a pollution status determination result for the current monitoring period is generated. The pollution status determination result is used to mark the categories of potential pollution sources and their spatial influence range. The pollution status determination result drives the governance strategy calculation engine, which outputs a set of parameterized control instructions for the potential pollution source categories and their spatial influence range.

2. The soil pollutant monitoring method based on artificial intelligence according to claim 1, characterized in that, The process of performing multi-level feature mapping on the raw soil observation data stream to obtain a primary feature set associated with the pollution event includes the following steps: The pollutant description information sequence is processed by pollution event word segmentation and entity recognition to obtain pollution entity word sequence and pollution behavior word sequence; Invoke the pre-built pollution semantic knowledge network to map each word in the pollution entity word sequence to an entity node and each word in the pollution behavior word sequence to a behavior node; Within the pollution semantic knowledge network, semantic relationship edges are created between the entity nodes and the behavior nodes based on predefined pollution domain logical relationships, forming a preliminary pollution semantic feature network; The preliminary pollution semantic feature network is subjected to topological structure encoding processing to transform the relationship between nodes and edges in the network into a fixed-dimensional information embedding vector; In parallel, abnormal segment truncation and normalization processing is performed on the time series readings of the environmental parameters to identify and extract non-stationary variation intervals in the readings; Multi-granularity convolutional coding is performed on the data within the non-stationary change range to capture the change patterns at different time scales and generate the state coding vector. The information embedding vector and the state encoding vector are concatenated to form the primary feature set.

3. The artificial intelligence-based soil pollutant monitoring method according to claim 2, characterized in that, The process of inputting the initial feature set into an iterative evolution network for multiple rounds of feature refinement, with the comparison result between the refined feature set output from the previous round and the historical contamination pattern knowledge base serving as feedback, dynamically adjusts the feature extraction dimensions for the next round. This multi-round feature refinement process generates an advanced contamination feature set, comprising the following steps: In the first round of processing of the iterative evolution network, the primary feature set is decomposed into semantic branch features and environmental branch features; The semantic branch features and the environmental branch features are respectively input into different feature refinement channels of the network. Each feature refinement channel contains multiple cascaded feature transformation layers to extract channel-specific deep features. The channel-specific depth features from different feature refinement channels are subjected to inter-channel feature crossover and compensation operations to generate the first round of refined feature set; The process of feature refinement, cross-compensation, and comparison with the historical knowledge base is repeated until the preset number of iterations or feature convergence conditions are reached, and finally the advanced pollution feature set is output.

4. The soil pollutant monitoring method based on artificial intelligence according to claim 3, characterized in that, The step of calculating the pollution status determination result for the current monitoring period based on the matching degree between the advanced pollution feature set and the preset pollution source feature template includes the following steps: Construct a pollution source feature template library containing multiple known pollution source categories. Each pollution source feature template consists of a set of discriminative feature vectors. Calculate the multi-dimensional similarity measure between the advanced pollution feature set and each pollution source feature template in the pollution source feature template library to obtain the similarity measure spectrum; Peak detection and pattern analysis are performed on the similarity metric spectrum to screen out candidate pollution source categories that exceed the confidence threshold; For each candidate pollution source category, feature components related to the pollution diffusion physical model are retrieved from the high-level pollution feature set to estimate the spatial influence intensity and range of the candidate pollution source category on the monitoring point. By combining the candidate pollution source categories and their corresponding spatial influence intensity and range, a structured pollution status determination result is generated.

5. The artificial intelligence-based soil pollutant monitoring method according to claim 4, characterized in that, After generating the pollution status determination result for the current monitoring period, the process also includes a pollution tracing path generation step: Analyze the candidate pollution source categories and spatial impact ranges in the pollution status determination results; Starting from the candidate pollution source category and ending at the monitoring point, a reverse path search is performed in the preset geographic information and pollution transmission knowledge graph. Based on the wind direction and hydrological data in the time series readings of the environmental parameters, the feasibility of the searched potential propagation paths is evaluated and weighted. By integrating the preset number of potential propagation paths with the highest weight, a visualized pollution tracing path map is generated, which is used to describe the possible propagation trajectory of pollutants from the source to the monitoring point.

6. The artificial intelligence-based soil pollutant monitoring method according to claim 5, characterized in that, The pollution status determination result-driven governance strategy calculation engine outputs a set of parameterized control instructions for the potential pollution source categories and their spatial influence range, including the following steps: The pollution status determination result and the pollution cause tracing path diagram are used as joint inputs and submitted to the governance strategy calculation engine. The governance strategy calculation engine includes a governance case reasoning network. After receiving the joint input, the governance case reasoning network retrieves the set of cases with the highest scenario similarity from its stored historical successful governance cases. The governance intervention measures adopted by the set of cases with the highest scenario similarity retrieved are decomposed and reorganized to form a set of basic intervention measures; The spatial impact range in the pollution status determination results is used to adjust the spatial adaptability of each measure in the set of basic intervention measures, and the implementation nodes and intensity of the measures are planned in a time sequence using the pollution tracing path map to generate a preliminary governance strategy sequence. The preliminary governance strategy sequence is input into the strategy simulation module, and the trend of pollutant concentration change after the strategy is implemented is predicted based on a simplified pollution migration model. If the predicted trend does not meet the preset governance objectives, the parameters of the measures in the preliminary governance strategy sequence will be automatically optimized and adjusted until the simulation prediction results meet the targets. At this point, the adjusted strategy sequence will be converted into the parameterized control instruction set.

7. The artificial intelligence-based soil pollutant monitoring method according to claim 6, characterized in that, The governance case reasoning network in the governance strategy calculation engine performs the following steps for case retrieval and strategy generation: Construct a case representation model with a triplet consisting of a pollution scenario feature vector, a treatment measure vector, and a treatment effect vector; The joint input is transformed into a current scene feature vector, and the Euclidean distance between the current scene feature vector and the pollution scene feature vectors of all historical cases in the case representation model is calculated. Select the historical cases with the smallest Euclidean distance to form a neighboring case set; Extract the governance measure vector for each historical case in the neighboring case set, and analyze the commonalities and differences of the governance measure vector across different environmental dimensions; Based on the commonalities and differences mentioned above, a new governance measure vector for the current scenario is synthesized by using weighted voting and feature interpolation methods. The new governance measures vector is decoded into specific, executable governance action descriptions, which serve as the basic set of intervention measures.

8. The artificial intelligence-based soil pollutant monitoring method according to claim 1, characterized in that, Following the output of the parameterized control command set for the potential pollution source category and its spatial impact range, a closed-loop learning step for the governance process is also included: After collecting and executing the set of parameterized control instructions, a new round of raw soil observation data stream is fed back by the monitoring nodes; The same multi-level feature mapping and iterative evolution processing is performed on the new round of original soil observation data stream to obtain the pollution feature set after treatment; Calculate the feature change vector between the set of pollution features after treatment and the set of advanced pollution features before treatment; The pollution status determination results, parameterized control command set, and the feature change vector corresponding to this governance task will be used together as a new governance experience sample. The new governance experience samples are added to the historical pollution pattern knowledge base and governance case reasoning network to update network parameters and knowledge base content, thereby enabling the self-evolution of the monitoring and governance system.

9. The artificial intelligence-based soil pollutant monitoring method according to claim 1, characterized in that, Before receiving the raw soil observation data stream from the monitoring nodes, an adaptive deployment planning step for the monitoring nodes is also included: Obtain historical pollution distribution maps, geographic information data, and meteorological statistics for the target monitoring area; The historical pollution distribution map was analyzed using a spatial clustering algorithm to identify potential areas with high pollution incidence. By combining the aforementioned geographic information data and meteorological statistics, possible diffusion paths of pollutants are simulated in potentially high-risk pollution areas; Calculate the optimal coordinates for the deployment of monitoring nodes at key nodes along the simulated diffusion path and at the core locations of potential high-risk pollution areas. A monitoring network deployment scheme is generated based on the optimal monitoring node layout coordinates. This scheme guides the installation location and density of physical monitoring equipment to ensure the coverage and representativeness of the original soil observation data stream.

10. A soil pollutant monitoring system based on artificial intelligence, characterized in that, The soil pollutant monitoring system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the artificial intelligence-based soil pollutant monitoring method as described in any one of claims 1-9.

Images