Environment analysis report semantic generation method and system based on knowledge graph and deep learning

Abstract

The application provides a kind of environment analysis report semantic generation method and system based on knowledge graph and deep learning, it is related to natural language processing technical field, obtains the original record text set of environment monitoring to be handled and pre-set knowledge graph structure;According to the monitoring point entity node set and pollution diffusion mode semantic subgraph network respectively generates monitoring point entity anchor point positioning distribution sequence and pollution diffusion semantic mode matching response sequence, then input time series causal reasoning network carries out cross time series causal reasoning processing, generates pollution causal chain feature representation;Based on pollution causal chain feature representation and original monitoring data sentence unit, sentence-level semantic fusion processing is carried out, and pollution causal enhancement semantic representation is generated;According to pollution causal enhancement semantic representation and pollutant index attribute relationship edge set, semantic generation path planning processing is carried out, and environment analysis report semantic text sequence is generated.The application realizes the causal logic driven high-quality automatic generation of environment analysis report.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to a method and system for generating semantic information for environment analysis reports based on knowledge graphs and deep learning. Background Technology

[0002] Environmental monitoring is a core foundational task for ecological environmental protection and pollution prevention decision-making. After completing on-site sampling and data recording, monitoring personnel need to organize, summarize, and compile a large amount of raw monitoring data into standardized environmental analysis reports to convey key information such as environmental quality status, causes of pollution incidents, and development trends to environmental management departments and the public. The quality of environmental analysis reports directly affects the scientific rigor and timeliness of environmental management decisions. However, traditional manual writing methods are not only time-consuming, labor-intensive, and inefficient, but also result in significant differences in the depth of understanding and expression of the same batch of monitoring data among different writers, making it difficult to effectively guarantee the professionalism, logical consistency, and information completeness of the report content.

[0003] The format of environmental monitoring reports is quite standardized compared to other written reports. Because they must be prepared in accordance with both national standards and government regulations governing discharge permit holders, each environmental monitoring report often only records the results of a single test, indicating whether or not the standards were exceeded. It does not conduct longitudinal or monthly analyses, nor does it perform in-depth calculations or predict future trends. Consequently, the data in each environmental monitoring report is often isolated, discrete, and difficult to compare or summarize to derive patterns.

[0004] Therefore, more professional and comprehensive environmental analysis reports can be generated based on environmental monitoring reports. This involves extracting, statistically analyzing, classifying, and performing in-depth calculations on various data from environmental monitoring reports prepared according to standardized national standards. This allows for the generation of analytical conclusions based on historical monitoring data results and effective predictions of future monitoring data ranges. By providing new functions such as tracking and analyzing environmental management, environmental analysis reports offer value-added services to companies holding discharge permits, ensuring daily compliance with emission standards and environmental protection management.

[0005] In recent years, with the rapid development of natural language generation technology, deep learning-based automatic text generation methods have been gradually introduced into the field of assisted writing of environmental analysis reports. Existing technologies typically employ a sequence-to-sequence encoder-decoder architecture, using raw monitoring records as input and directly generating report text through attention mechanisms or recurrent neural networks. While these methods alleviate the burden of manual writing to some extent, the generated reports are essentially just superficial restates and simple sentence rearrangements of the raw monitoring data, lacking the exploration and expression of deep logical relationships between the monitoring data. Especially in paragraphs requiring rigorous logical reasoning, such as pollution causal analysis and descriptions of pollution diffusion trends, the text generated by existing methods often suffers from broken causal chains and ambiguous logical relationships, failing to reflect the deeper semantics behind the monitoring data, such as pollution source-sink relationships, pollution transmission pathways, and concentration evolution patterns.

[0006] Furthermore, existing technologies have significant shortcomings in utilizing domain knowledge. Most methods rely solely on implicit knowledge from general pre-trained language models or employ simple keyword matching to assist text generation. They fail to construct a structured knowledge system specific to the environmental monitoring field, and even more so, they fail to systematically model and integrate the spatial relationships between monitoring point entities, the attribute relationships between pollutant indicators, and the causal evolution relationships of pollution events over time. This results in generated report texts that cannot reflect the unique spatiotemporal causal characteristics and diffusion evolution patterns of environmental monitoring data, making it difficult to meet the comprehensive requirements of professional environmental analysis reports for semantic integrity, logical rigor, and domain professionalism. This severely restricts the application value of automatic environmental analysis report generation technology in practical environmental monitoring work. Summary of the Invention

[0007] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for semantic generation of environment analysis reports based on knowledge graphs and deep learning, the method comprising:

[0008] Obtain the set of original environmental monitoring record texts to be processed and a preset knowledge graph structure for the environmental monitoring field. The preset knowledge graph structure includes a set of entity nodes of monitoring points, a set of relationship edges of pollutant index attributes, and a semantic subgraph network of pollution diffusion patterns.

[0009] Based on the set of entity nodes of the monitoring points, the set of original environmental monitoring records to be processed is subjected to entity anchor point location processing to generate a distribution sequence of entity anchor points of the monitoring points. Based on the semantic subgraph network of the pollution diffusion pattern, the set of original environmental monitoring records to be processed is subjected to pollution diffusion semantic pattern matching processing to generate a pollution diffusion semantic pattern matching response sequence.

[0010] The monitoring point entity anchor point location distribution sequence and the pollution diffusion semantic pattern matching response sequence are input into a preset temporal causal reasoning network for cross-temporal causal reasoning processing to generate a pollution causal chain feature representation. The cross-temporal causal reasoning processing includes pollution source-sink causal reasoning processing, pollution path causal reasoning processing, and pollution concentration causal reasoning processing.

[0011] Based on the pollution causal chain feature representation and the original monitoring data sentence units in the original environmental monitoring record text set to be processed, sentence-level semantic fusion processing is performed to generate a pollution causal enhanced semantic representation.

[0012] Based on the enhanced semantic representation of pollution causality and the set of pollutant indicator attribute relationship edges in the preset knowledge graph structure, semantic generation path planning is performed to generate a semantic text sequence of environmental analysis report. The semantic text sequence of environmental analysis report includes a pollution cause analysis paragraph, a pollution diffusion trend description paragraph, and a monitoring conclusion summary paragraph.

[0013] Furthermore, embodiments of the present invention also provide a semantic generation system for environmental analysis reports based on knowledge graphs and deep learning, characterized in that it includes:

[0014] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned semantic generation method for environment analysis reports based on knowledge graphs and deep learning by executing the machine-executable instructions.

[0015] Based on the above, by acquiring the set of original environmental monitoring records to be processed and combining it with a pre-defined knowledge graph structure containing a set of monitoring point entity nodes, a set of pollutant indicator attribute relationship edges, and a semantic subgraph network of pollution diffusion patterns, the set of monitoring point entity nodes is used to perform monitoring point entity anchor point localization processing on the original record text, and the semantic subgraph network of pollution diffusion patterns is used to perform pollution diffusion semantic pattern matching processing on the original record text. This generates a monitoring point entity anchor point localization distribution sequence and a pollution diffusion semantic pattern matching response sequence, respectively. Thus, the spatial positioning information and diffusion evolution pattern information implicit in the original text are explicitly extracted in the form of structured sequences. On this basis, the above two types of sequences are input into a pre-defined temporal causal reasoning network to perform cross-temporal causal reasoning covering pollution source-sink causal reasoning processing, pollution path causal reasoning processing, and pollution concentration causal reasoning processing, generating a pollution causal chain feature representation. This allows the pollution source-sink evolution relationship, pollution transmission path relationship, and pollutant concentration causal relationship between different monitoring times to be systematically modeled and encoded in a unified high-dimensional feature space.

[0016] Subsequently, based on the pollution causal chain feature representation and the original monitoring data sentence units, sentence-level semantic fusion processing is performed to generate a pollution causal enhanced semantic representation. This achieves deep interweaving and complementary enhancement of domain knowledge-driven causal logic and original data semantics at the sentence granularity. The resulting semantic representation retains the factual accuracy of the original monitoring data while incorporating the logical completeness and professional depth endowed by causal reasoning. Finally, based on the pollution causal enhanced semantic representation and the set of pollutant indicator attribute relationships, semantic generation path planning processing is performed to generate a semantic text sequence of environmental analysis report containing paragraphs on pollution cause analysis, pollution diffusion trend description, and monitoring conclusion summary. This results in an environmental analysis report with a rigorous and complete causal logical chain in content, a complete professional paragraph system covering cause analysis, trend description, and conclusion summary in structure, and fully reflects the spatiotemporal causal relationship characteristics and diffusion evolution laws unique to the field of environmental monitoring. This significantly improves the professional depth, logical consistency, and decision support capabilities of the environmental analysis report. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the execution flow of the semantic generation method for environment analysis reports based on knowledge graphs and deep learning provided in an embodiment of the present invention.

[0018] Figure 2 This is a logical schematic diagram of the semantic generation method for environment analysis reports based on knowledge graphs and deep learning provided in this embodiment of the invention.

[0019] Figure 3 This is a schematic diagram of exemplary hardware and software components of the semantic generation system for environment analysis reports based on knowledge graphs and deep learning provided in an embodiment of the present invention. Detailed Implementation

[0020] Figure 1 This is a flowchart illustrating a semantic generation method for environment analysis reports based on knowledge graphs and deep learning, provided in one embodiment of the present invention. The following is a detailed explanation in conjunction with... Figure 2 A detailed introduction is provided. This embodiment applies the entire technical solution to a scenario of automatically generating semantic environmental analysis reports from surface water environmental monitoring data surrounding an industrial park. The industrial park has 15 water quality monitoring points numbered M01 to M15, monitoring indicators including chemical oxygen demand (COD), ammonia nitrogen, total phosphorus, and heavy metals. All monitoring data collection, transmission, and processing comply with environmental monitoring data management regulations, and data transmission uses an encryption protocol, ensuring no personal privacy information is involved. A pre-built knowledge graph structure for the environmental monitoring field has been pre-constructed, comprising a set of entity nodes for monitoring points, a set of edge relationships for pollutant indicator attributes, and a semantic subgraph network for pollution diffusion patterns.

[0021] Step S110: Obtain the set of original environmental monitoring records to be processed and the preset knowledge graph structure for the environmental monitoring field. The preset knowledge graph structure includes a set of entity nodes of monitoring points, a set of relationship edges of pollutant index attributes, and a semantic subgraph network of pollution diffusion patterns.

[0022] The system retrieves a set of raw environmental monitoring records for the current month's surface water monitoring in the industrial park from the environmental monitoring data management platform. This set contains monitoring records from 15 monitoring points. Each record includes a monitoring timestamp, monitoring point name, monitoring indicator value, monitoring equipment number, and a description of the monitoring method. Simultaneously, a pre-defined knowledge graph structure for the environmental monitoring field is loaded. This pre-defined knowledge graph structure is stored in a graph database as a directed attribute graph. The set of monitoring point entity nodes includes 15 monitoring point entity nodes numbered M01 to M15. Each monitoring point entity node stores the monitoring point name, spatial coordinates (latitude and longitude), and the watershed zoning attribute to which the point belongs. The set of pollutant indicator attribute relationship edges includes positive correlation edges between chemical oxygen demand and ammonia nitrogen, negative correlation edges between dissolved oxygen and ammonia nitrogen, and positive correlation edges between total phosphorus and chlorophyll. Each attribute relationship edge carries relationship type and relationship strength weight attributes. The pollution diffusion pattern semantic subgraph network comprises a set of pollution source nodes, a set of pollution propagation path edges, and a set of pollution receptor nodes. The pollution source node set includes industrial park discharge outlet nodes, the pollution propagation path edge set includes river flow direction path edges and surface runoff path edges, and the pollution receptor node set includes downstream monitoring section nodes. This pollution diffusion pattern semantic subgraph network describes the complete diffusion process of pollutants from pollution sources to pollution receptors via propagation paths using a directed graph topology. The aforementioned knowledge graph data was constructed by domain experts based on the actual geographical information and hydrological data of the industrial park, following knowledge graph construction specifications.

[0023] Step S120: Based on the set of entity nodes of the monitoring points, perform entity anchor point localization processing on the set of original environmental monitoring records to be processed to generate a distribution sequence of entity anchor points of the monitoring points. Based on the semantic subgraph network of the pollution diffusion pattern, perform semantic pattern matching processing on the set of original environmental monitoring records to be processed to generate a semantic pattern matching response sequence of pollution diffusion.

[0024] Using the set of entity nodes of monitoring points and the semantic subgraph network of pollution diffusion patterns loaded in step S110, the set of original environmental monitoring record texts to be processed is subjected to two parallel processing methods: entity localization and semantic pattern matching.

[0025] Step S121: Extract the text character stream sequence from the original environmental monitoring record text set to be processed, and perform a character-level bitwise comparison between the text character stream sequence and the monitoring point entity node names in the monitoring point entity node set to obtain a character-level bitwise comparison matching position sequence.

[0026] All text content is extracted from the original environmental monitoring records and concatenated into a continuous text character stream sequence, where each element is a Unicode character. The names of the 15 monitoring points in the monitoring point entity node set are used as pattern strings, and the Aho-Corasick multi-pattern string matching algorithm is employed to perform character-by-character comparison of the text character stream sequence. This algorithm constructs a trie containing all monitoring point entity node names and creates a failure jump pointer for each node in the trie. During a single scan of the text character stream sequence, the algorithm simultaneously detects the occurrence positions of all monitoring point entity node names. When the scanning pointer locates the terminating character of a monitoring point entity node name in the text character stream sequence, the matching monitoring point entity node name, the starting character position, and the ending character position are recorded, generating a character-by-character comparison matching position sequence.

[0027] Step S122: Based on the character-level bit-by-bit comparison and matching position sequence, perform monitoring point entity boundary anchoring processing on the original record text set of the environment monitoring to be processed, determine the start character position and end character position of the monitoring point entity name in the text, and generate a monitoring point entity anchor start and end position pair sequence.

[0028] The character-level bit-by-bit matching position sequence generated in step S121 is processed for monitoring point entity boundary anchoring. For each matching record in the character-level bit-by-bit matching position sequence, the starting character position of the match is extracted as the starting character position of the monitoring point entity name in the text, and the ending character position of the match is extracted as the ending character position. If the same monitoring point entity name appears multiple times in the text, each appearance is recorded independently. The monitoring point entity name, starting character position, and ending character position of each appearance are combined into a triplet. All triplets are arranged in ascending order of the starting character position to generate a sequence of monitoring point entity anchor point start and end position pairs.

[0029] Step S123: Map the names of each monitoring point entity in the sequence of the start and end positions of the monitoring point entity anchor points to the unique identifiers of the monitoring point entity nodes in the set of monitoring point entity nodes, and generate a mapping sequence of monitoring point entity node identifiers.

[0030] For each triplet in the sequence of start and end positions of monitoring point entity anchors generated in step S122, the monitoring point entity name is precisely matched with the monitoring point entity node names in the set of monitoring point entity nodes to find the corresponding monitoring point entity node. The unique identifier of the monitoring point entity node, i.e., the node number, is extracted. The monitoring point entity name in the triplet is replaced with the corresponding unique identifier of the monitoring point entity node, generating a monitoring point entity node identifier mapping sequence. Each element in the monitoring point entity node identifier mapping sequence contains the unique identifier of the monitoring point entity node, the start character position, and the end character position.

[0031] Step S124: Sort the mapping sequence of entity node identifiers of the monitoring points according to the order in which the entities of the monitoring points appear in the text, and record the text span distance between adjacent entities of the monitoring points to generate a monitoring point entity anchor point positioning distribution sequence. The monitoring point entity anchor point positioning distribution sequence includes a unique identifier sequence of entity nodes of the monitoring points and a text span distance sequence between adjacent entities.

[0032] Sort the monitoring point entity node identifier mapping sequence generated in step S123 in ascending order of the starting character position. Calculate the text span distance between every two adjacent monitoring point entities, where the text span distance equals the starting character position of the next monitoring point entity minus the ending character position of the previous monitoring point entity. Arrange the sorted monitoring point entity node unique identifiers sequentially to form a monitoring point entity node unique identifier sequence, and arrange all adjacent entity text span distances sequentially to form an adjacent entity text span distance sequence. Combine these two to form the monitoring point entity anchor point location distribution sequence.

[0033] Step S125: Extract the set of pollution source nodes, the set of pollution propagation path edges, and the set of pollution receptor nodes from the pollution diffusion pattern semantic subgraph network, convert the set of pollution source nodes into a set of pollution source semantic templates, convert the set of pollution propagation path edges into a set of pollution propagation path semantic templates, and convert the set of pollution receptor nodes into a set of pollution receptor semantic templates.

[0034] A set of pollution source nodes is extracted from the semantic subgraph network of pollution diffusion patterns. Each pollution source node contains a pollution source name, pollution source category, and emission characteristic attributes. The pollution source name and pollution source category of each pollution source node are converted into a pollution source semantic template. The pollution source semantic template is represented in the form of a regular expression and matches text fragments containing synonyms of pollution source name or pollution source category. Similarly, the propagation path name and propagation medium type of each edge in the pollution propagation path edge set are converted into a pollution propagation path semantic template. The receptor name and receptor type of each pollution receptor node in the pollution receptor node set are converted into a pollution receptor semantic template.

[0035] Step S126: Use the pollution source semantic template set, pollution propagation path semantic template set, and pollution receptor semantic template set to perform sliding window semantic template matching processing on the original environmental monitoring record text set to be processed, and record the matching hit position and matching template type identifier of each semantic template in the text.

[0036] For the text character stream sequence of the raw environmental monitoring records to be processed, the sliding window length is set to the length of the text character stream sequence, and the sliding step is 1 character. At each sliding window position, the text fragment within the window is subjected to regular expression matching against the sets of semantic templates for pollution sources, pollution propagation paths, and pollution receptors, respectively. When a text fragment successfully matches a semantic template, the starting character position and ending character position of the matched text fragment in the original text, as well as the matching template type identifier, are recorded. The matching template type identifier can be a pollution source template, a propagation path template, or a pollution receptor template.

[0037] Step S127: Based on the matching hit location and the matching hit template type identifier, construct a set of pollution diffusion semantic matching path instances, and perform path instance confidence scoring and filtering on the set of pollution diffusion semantic matching path instances to serialize and generate a pollution diffusion semantic pattern matching response sequence.

[0038] Using the matching results recorded in step S126, a complete semantic matching path for pollution diffusion is constructed, and the confidence level of the path is evaluated and filtered.

[0039] Step S1271: Based on the matching hit position and the matching hit template type identifier, combine the matching hit sequences that satisfy the condition of going from the set of pollution source nodes through the set of pollution propagation path edges to the set of pollution receptor nodes into pollution diffusion semantic matching path instances, and obtain a set of pollution diffusion semantic matching path instances.

[0040] From the matching results recorded in step S126, extract all matching items with the template type identifier of "pollution source template" as path start point candidates, extract all matching items with the template type identifier of "propagation path template" as path intermediate node candidates, and extract all matching items with the template type identifier of "pollution receptor template" as path end point candidates. For each path start point candidate, find the matching item in the text that follows it and has the template type identifier of "propagation path template", then find the matching item that follows the "propagation path" matching item and has the template type identifier of "pollution receptor template". If the three satisfy the order relationship of increasing starting character position, combine the three into a pollution diffusion semantic matching path instance. Traverse all combinations to generate a set of pollution diffusion semantic matching path instances.

[0041] Step S1272: Perform path instance confidence scoring on each pollution diffusion semantic matching path instance in the pollution diffusion semantic matching path instance set, calculate the consistency of the matching template type identifier sequence with the direction of the edges in the pollution diffusion pattern semantic subgraph network and the continuity of the text span of the path instance, and obtain the path instance confidence score set.

[0042] For each pollution diffusion semantic matching path instance, extract the pollution source category of its pollution source template, the propagation medium type of its propagation path template, and the receptor type of its pollution receptor template. In the pollution diffusion pattern semantic subgraph network, search for a path from a node of the pollution source category through an edge of the propagation medium type to a node of the receptor type. If a path exists, the directional consistency value is set to 1; otherwise, it is set to 0. Simultaneously, calculate the text span continuity of the path instance. The continuity value is equal to 1 minus the ratio of the text span distance between adjacent matches in the path instance to the total length of the text character stream sequence. The path instance confidence score is equal to the directional consistency value multiplied by the continuity value. Traverse all pollution diffusion semantic matching path instances to obtain the path instance confidence score set.

[0043] Step S1273: Based on the path instance confidence score set, perform path instance filtering on the pollution diffusion semantic matching path instance set, retain pollution diffusion semantic matching path instances whose path instance confidence scores are higher than the preset path instance confidence filtering threshold, and obtain the filtered pollution diffusion semantic matching path instance set.

[0044] The path instance confidence score of each pollution diffusion semantic matching path instance is compared with the preset path instance confidence screening threshold, and pollution diffusion semantic matching path instances with path instance confidence scores higher than the preset path instance confidence screening threshold are retained.

[0045] Step S1274: Serialize the pollution diffusion semantic matching path instances in the filtered set into a pollution diffusion semantic pattern matching response sequence. The pollution diffusion semantic pattern matching response sequence includes a path instance identifier sequence, a sequence of monitoring point entity nodes contained in the path instance, and a semantic template type sequence corresponding to the path instance.

[0046] After filtering pollution diffusion semantic matching path instances, each instance in the set is assigned a unique path instance identifier. The sequence of monitoring point entity nodes traversed by this pollution diffusion semantic matching path instance in the text is extracted; that is, the unique identifiers of the monitoring point entity nodes obtained through the mapping between the starting and ending matching items in step S123. The semantic template type sequence corresponding to the pollution diffusion semantic matching path instance is extracted; that is, the sequential list of pollution source template type, propagation path template type, and pollution receptor template type. The above information is serialized to generate a pollution diffusion semantic pattern matching response sequence.

[0047] Step S130: Input the monitoring point entity anchor point location distribution sequence and the pollution diffusion semantic pattern matching response sequence into a preset temporal causal reasoning network for cross-temporal causal reasoning processing to generate a pollution causal chain feature representation. The cross-temporal causal reasoning processing includes pollution source-sink causal reasoning processing, pollution path causal reasoning processing, and pollution concentration causal reasoning processing.

[0048] Using the monitoring point entity anchor point location distribution sequence generated in step S124 and the pollution diffusion semantic pattern matching response sequence generated in step S1274, cross-temporal causal reasoning is performed through a preset temporal causal reasoning network.

[0049] Step S131: Input the sequence of unique identifiers of entity nodes of monitoring points in the entity anchor point location distribution sequence of the monitoring points into the entity embedding layer of the temporal causal inference network for entity embedding encoding processing, map each unique identifier of entity node of monitoring points into an entity embedding vector, and generate a sequence of entity embedding vectors of monitoring points.

[0050] The pre-defined temporal causal inference network's entity embedding layer is a learnable embedding matrix. The number of rows in the embedding matrix equals the total number of entity nodes at the monitoring points, which is 15. The number of columns in the embedding matrix equals the pre-defined entity embedding vector dimension De. For each entity node in the monitoring point anchor location distribution sequence, a unique identifier is assigned. A lookup table is used to extract the corresponding entity embedding vector from the embedding matrix, generating a monitoring point entity embedding vector sequence. Each entity embedding vector in the monitoring point entity embedding vector sequence has a dimension of De.

[0051] Step S132: Input the path instance identifier sequence and semantic template type sequence in the pollution diffusion semantic pattern matching response sequence into the path embedding layer of the temporal causal reasoning network for path embedding encoding processing, map the path instance identifier to the path instance embedding vector, map the semantic template type to the semantic template embedding vector, and generate the path instance embedding vector sequence and the semantic template embedding vector sequence.

[0052] The pre-defined path embedding layer of the temporal causal inference network contains two embedding matrices: a path instance embedding matrix and a semantic template embedding matrix. For each path instance identifier in the path instance identifier sequence, a path instance embedding vector is obtained by looking up a table in the path instance embedding matrix; the path instance embedding vector has a dimension of Dp. For each semantic template type in the semantic template type sequence, a semantic template embedding vector is obtained by looking up a table in the semantic template embedding matrix; the semantic template embedding vector has a dimension of Dt. This generates both the path instance embedding vector sequence and the semantic template embedding vector sequence.

[0053] Step S133: Perform time encoding fusion processing on the timestamp information of each text statement in the monitoring point entity embedding vector sequence and the original record text set of the environment monitoring to be processed, and add time position encoding to the monitoring point entity embedding vector according to the order of the timestamps to generate a monitoring point entity embedding vector sequence with enhanced time position encoding.

[0054] The timestamp information of each text statement is extracted from the raw monitoring record text set to be processed, and the timestamp information is converted into Unix timestamp values. For the p-th time position, a time position encoding vector PEp is generated, and the dimension of PEp is the same as the dimension of the entity embedding vector De. The element value of the 2i-th dimension of PEp is sin(p / (10000^(2i / De))), and the element value of the 2i+1-th dimension of PEp is cos(p / (10000^(2i / De))), where i is an integer index from 0 to De / 2 minus 1. The time position encoding vector PEp is added element-wise to the corresponding monitoring point entity embedding vector to generate a time position encoded enhanced monitoring point entity embedding vector sequence.

[0055] Step S134: Input the time location-enhanced monitoring point entity embedding vector sequence, path instance embedding vector sequence, and semantic template embedding vector sequence into the causal graph attention layer of the temporal causal reasoning network for causal graph attention aggregation processing. Use the pollution source entity embedding vector as the query vector, the pollution propagation path embedding vector as the key vector, and the pollution receptor entity embedding vector as the value vector to calculate the attention weight distribution.

[0056] The pre-defined temporal causal inference network's causal graph attention layer receives three types of input vector sequences. The entity embedding vector of the pollution source node in the temporal location-enhanced monitoring point entity embedding vector sequence is used as the query vector Q. The path instance embedding vector sequence and the semantic template embedding vector sequence are concatenated to form the key vector K. The entity embedding vector of the pollution receptor node in the temporal location-enhanced monitoring point entity embedding vector sequence is used as the value vector V. The attention weight distribution is calculated, and the attention weight matrix A is calculated as: A = softmax((Q × K^T) / sqrt(Dk)), where Dk is the dimension of the key vector K, K^T is the transpose of the key vector K, and softmax is the flexible maximum transfer function.

[0057] Step S135: Based on the attention weight distribution, perform weighted aggregation calculation on the causal intensity of the diffusion of pollutants from the pollution source entity to the pollution receiver entity via the pollution propagation path, and generate the pollution source entity aggregated feature vector, the pollution propagation path aggregated feature vector, and the pollution receiver entity aggregated feature vector.

[0058] Using the attention weight matrix A calculated in step S134, the value vector V is weighted and aggregated, and the aggregated feature vector of the pollution receptor entity is equal to A multiplied by V. Similarly, the query vector Q is weighted and aggregated with the attention weight matrix A, and the aggregated feature vector of the pollution source entity is equal to A^T multiplied by Q. The key vector K is weighted and aggregated with the attention weight matrix A, and the aggregated feature vector of the pollution propagation path is equal to A multiplied by K. These three types of aggregated feature vectors respectively store the pollution source information, propagation path information, and pollution receptor information after causal relationship aggregation.

[0059] Step S136: Input the aggregated feature vector of the pollution source entity, the aggregated feature vector of the pollution propagation path, and the aggregated feature vector of the pollution receptor entity into the temporal causal convolution module of the temporal causal inference network, and perform causal convolution operation along the time dimension to generate the temporal causal convolution feature vector of the pollution source, the temporal causal convolution feature vector of the pollution propagation path, and the temporal causal convolution feature vector of the pollution receptor.

[0060] The pre-defined temporal causal inference network's temporal causal convolution module employs a dilated causal convolution structure, containing multiple dilated convolution layers with increasing dilation coefficients layer by layer. For the aggregated feature vector sequence of pollution source entities, a one-dimensional causal convolution operation is performed along the time dimension. The convolution kernel size is 3, and the dilation coefficients increase layer by layer from 1, 2, and 4. During convolution, only information from the current time step and previous time steps is used, without introducing future information leakage. Batch normalization and the ReLU nonlinear activation function are applied after the convolution operation. After multiple layers of dilated causal convolution, a temporal causal convolution feature vector for pollution sources is generated. Similarly, the aggregated feature vectors for pollution propagation paths and pollution receptor entities are processed to generate temporal causal convolution feature vectors for pollution propagation paths and pollution receptors.

[0061] Step S137: Perform cross-entity gated fusion processing on the pollution source temporal causal convolutional feature vector, pollution propagation path temporal causal convolutional feature vector, and pollution receptor temporal causal convolutional feature vector. Use the gating mechanism to perform element-wise fusion weight allocation on the three types of temporal causal convolutional feature vectors to generate cross-entity gated fusion feature vector.

[0062] The temporal causal convolutional feature vectors Hs (pollution source), Hp (pollution propagation path), and Hr (pollution receptor) are concatenated to obtain a joint feature vector. This joint feature vector is then input into a fully connected layer. The output dimension of the fully connected layer is 3 times the feature dimension. The output is reshaped into three gating weight matrices Gs, Gp, and Gr, each with the same dimension as the corresponding feature vector. Softmax normalization is applied to the three gating weight matrices at corresponding positions, ensuring that Gs + Gp + Gr equals 1 at each position. The cross-entity gating fusion feature vector is calculated as: Hfused = Gs⊙Hs + Gp⊙Hp + Gr⊙Hr, where ⊙ denotes element-wise multiplication.

[0063] Step S138: Input the cross-entity gated fusion feature vector into the causal chain decoding layer of the temporal causal inference network, perform causal chain sequence decoding and temporal alignment and feature connection processing with the original monitoring data statement units, and generate pollution causal chain feature representation.

[0064] The cross-entity gated fusion feature vector generated in step S137 is input into the causal chain decoding layer of the temporal causal inference network for decoding.

[0065] Step S1381: Input the cross-entity gated fusion feature vector into the causal chain decoding layer of the temporal causal inference network to perform causal chain sequence decoding processing, and generate the causal transmission intensity sequence and causal transmission direction sequence between the pollution source and the pollution receptor.

[0066] The causal chain decoding layer employs a recurrent neural network decoder, with each decoder unit being a gated recurrent unit. The cross-entity gated fusion feature vector (Hfused) is sequentially input into the gated recurrent unit at each time step. At each time step, the gated recurrent unit outputs a causal decoding hidden state vector. This causal decoding hidden state vector is then input into two fully connected heads: the causal strength regression head outputs a scalar value, mapped to a causal transmission strength value between 0 and 1 using a sigmoid function; the causal direction classification head outputs a two-dimensional vector, mapped to probability values ​​for forward and reverse causality using softmax, with the direction showing the higher probability value taken as the causal transmission direction. By traversing all time steps, a causal transmission strength sequence and a causal transmission direction sequence between the pollution source and the pollution receptor are generated.

[0067] Step S1382: Perform time-series alignment processing on the timestamp information of the original monitoring data statement units in the original environmental monitoring record text set to be processed, based on the causal transmission intensity sequence and causal transmission direction sequence, to generate local feature representations of the pollution causal chain corresponding to each original monitoring data statement unit.

[0068] The timestamp information of each raw monitoring data statement unit is extracted from the set of raw environmental monitoring records to be processed. The timestamps of each time step in the causal transmission strength sequence and causal transmission direction sequence are time-aligned and matched with the timestamp information of the raw monitoring data statement unit using nearest neighbor time matching. After matching, each raw monitoring data statement unit corresponds to a causal transmission strength value and a causal transmission direction value. These two values ​​are concatenated with the causal decoding hidden state vector of the corresponding time step to generate the local feature representation of the pollution causal chain corresponding to that raw monitoring data statement unit.

[0069] Step S1383: Perform feature connection processing on the local feature representation of the pollution causal chain corresponding to each original monitoring data statement unit according to the order of the statements to generate a pollution causal chain feature representation. The pollution causal chain feature representation includes the sequence of local feature representations of the pollution causal chain corresponding to each statement and the sequence of pollution causal transmission features between statements.

[0070] The local feature representations of the pollution causal chain corresponding to each original monitoring data statement unit are arranged sequentially according to their order in the text, forming a sequence of local feature representations of the pollution causal chain for each statement. The difference vectors of the local feature representations of the pollution causal chain between adjacent original monitoring data statement units are calculated, and all difference vectors are arranged sequentially to form a sequence of pollution causal transmission features between statements. These two sequences are combined to generate the pollution causal chain feature representation.

[0071] Step S140: Perform sentence-level semantic fusion processing based on the pollution causal chain feature representation and the original monitoring data sentence units in the original environmental monitoring record text set to be processed, and generate a pollution causal enhanced semantic representation.

[0072] The pollution causal chain feature representation generated in step S1383 is combined with the original monitoring data statement units in the original environmental monitoring record text set to be processed, and the original semantics are integrated into the causal chain information.

[0073] Step S141: Perform sentence segmentation processing on the original monitoring data sentence units in the original environmental monitoring record text set to be processed, and divide each original monitoring data sentence unit into an original monitoring data word sequence.

[0074] For each raw monitoring data sentence unit in the set of raw environmental monitoring records to be processed, a pre-trained word segmenter in the environmental monitoring domain is used for sentence segmentation. This word segmenter, based on lexical matching and a conditional random field model, divides each raw monitoring data sentence unit into a sequence of raw monitoring data tokens, where each token is a word or punctuation mark. The raw monitoring data token sequence corresponding to the u-th raw monitoring data sentence unit is denoted as Tu, and Tu contains Lu tokens.

[0075] Step S142: Input the original monitoring data lexical sequence into the lexical embedding layer of the pre-trained environmental monitoring domain language representation model, map each original monitoring data lexical into a lexical semantic embedding vector, and generate the original monitoring data lexical semantic embedding vector sequence.

[0076] The pre-trained language representation model for environmental monitoring employs a BERT-based pre-trained model, which has been pre-trained on a large corpus of environmental monitoring texts. The lexical embedding layer comprises a lexical embedding matrix, a positional embedding matrix, and a segmental embedding matrix. For each lexical in the original monitoring data lexical sequence Tu, the lexical embedding vector is obtained by looking up a table in the lexical embedding matrix, the positional embedding vector is obtained by looking up a table in the positional embedding matrix, and the segmental embedding vector is obtained by looking up a table in the segmental embedding matrix. These three are then added element-wise to obtain the lexical semantic embedding vector. The lexical semantic embedding vectors corresponding to all lexicals in the original monitoring data lexical sequence Tu are arranged in order to generate the original monitoring data lexical semantic embedding vector sequence Eu, where the shape of Eu is Lu multiplied by the lexical embedding dimension Dw.

[0077] Step S143: Perform sentence-level pooling on the original monitoring data semantic embedding vector sequence, and perform mean pooling on the semantic embedding vector within each original monitoring data sentence unit to generate the original monitoring data sentence-level semantic vector sequence.

[0078] For the u-th original monitoring data statement unit, its original monitoring data semantic embedding vector sequence Eu is a matrix of shape Lu multiplied by the word embedding dimension Dw. First, the concentration value word, pollutant indicator name word, and exceedance judgment word contained in the original monitoring data statement unit are identified. These words are marked as key information words and assigned a first preset weight value Wk. The remaining words are assigned a second preset weight value Wo, where Wk is greater than Wo. Then, a weighted average is calculated on the word semantic embedding vector using the weight values ​​corresponding to each word, i.e., a weighted arithmetic mean is obtained along the corresponding dimension, resulting in an original monitoring data statement-level semantic vector Su with dimension Dw. This weighted pooling operation ensures that core information such as concentration values, pollutant indicator names, and exceedance judgments dominate the statement-level semantic vector, thereby accurately reflecting the exceedance judgment and toxicity assessment semantics carried by the statement, effectively avoiding semantic dilution and misjudgment caused by traditional mean pooling of key information. The above operation is performed on all original monitoring data statement units to generate the original monitoring data statement-level semantic vector sequence.

[0079] Step S144: Extract the pollution causal chain local feature representation sequence corresponding to each original monitoring data statement unit from the pollution causal chain feature representation, and map each pollution causal chain local feature representation in the pollution causal chain local feature representation sequence to the same vector dimension space as the original monitoring data statement-level semantic vector sequence to generate a pollution causal chain local feature mapping vector sequence.

[0080] Extract the local feature representation sequence of each statement from the pollution causal chain feature representation. The dimension of each local feature representation is Dc. Map the local feature representation of the pollution causal chain to a vector space of dimension Dw through a fully connected layer. The weight matrix of the fully connected layer has a dimension of Dc multiplied by Dw. After mapping, the sequence of local feature mapping vectors of the pollution causal chain is obtained.

[0081] Step S145: Perform sentence-by-sentence concatenation processing on the original monitoring data statement-level semantic vector sequence and the pollution causal chain local feature mapping vector sequence. Perform vector concatenation operation on the statement-level semantic vector of each original monitoring data statement unit and the corresponding pollution causal chain local feature mapping vector to generate a concatenated statement semantic vector sequence. Input the concatenated statement semantic vector sequence into a preset inter-statement bidirectional attention interaction module to perform context-weighted aggregation and residual normalization processing to generate a pollution causal enhanced semantic representation.

[0082] The original monitoring data statement-level semantic vector Su of the u-th original monitoring data statement unit is concatenated with the corresponding pollution causal chain local feature mapping vector Cu in the vector dimension. The concatenated statement semantic vector is [Su;Cu], with a dimension of 2 times Dw. All concatenated statement semantic vectors are arranged in order to generate a sequence of concatenated statement semantic vectors. The sequence of concatenated statement semantic vectors is then input into a preset bidirectional attention interaction module between statements for processing.

[0083] Step S1451: Input the concatenated sentence semantic vector sequence into a preset inter-sentence bidirectional attention interaction module, calculate the bidirectional attention weight between each concatenated sentence semantic vector and the other concatenated sentence semantic vectors, generate an inter-sentence bidirectional attention weight matrix, use the inter-sentence bidirectional attention weight matrix to perform context-weighted aggregation processing on the concatenated sentence semantic vector sequence, and perform weighted summation of each concatenated sentence semantic vector according to the inter-sentence bidirectional attention weight matrix to generate a context-enhanced concatenated sentence semantic vector sequence.

[0084] The pre-defined bidirectional attention interaction module between sentences employs a multi-layer Transformer encoder structure. It takes the sequence of concatenated sentence semantic vectors as input and calculates the attention weight between any two concatenated sentence semantic vectors using multiple attention heads. For the i-th and j-th concatenated sentence semantic vectors, the attention weight is calculated as follows: the i-th concatenated sentence semantic vector is mapped to a query vector using a query weight matrix, and the j-th concatenated sentence semantic vector is mapped to a key vector using a key weight matrix. The dot product of the query vector and the key vector is calculated, divided by the square root of the key vector's dimension, and then normalized using softmax to obtain the attention weight value. The attention weight matrix is ​​then used to perform a weighted sum of the attention weights on the concatenated sentence semantic vector sequence. The output at each position is the weighted sum of the attention weights of all concatenated sentence semantic vectors at all positions, generating a context-enhanced concatenated sentence semantic vector sequence.

[0085] Step S1452: Perform residual join processing on the context-enhanced concatenated sentence semantic vector sequence, add the original monitoring data sentence-level semantic vector sequence and the context-enhanced concatenated sentence semantic vector sequence element by element to generate a residual-enhanced concatenated sentence semantic vector sequence, perform layer normalization processing on the residual-enhanced concatenated sentence semantic vector sequence, and perform mean and variance normalization operations on each dimension of each residual-enhanced concatenated sentence semantic vector to generate a normalized concatenated sentence semantic vector sequence.

[0086] The original monitoring data sentence-level semantic vector sequence is element-wise added to the corresponding vectors in the context-enhanced concatenated sentence semantic vector sequence to complete the residual join processing. Each vector after residual join is then subjected to layer normalization. The mean and variance of all dimensions of the vector are calculated. The mean is subtracted from each dimension value, and the result is divided by the variance plus the square root of the smallest positive number. This result is then multiplied by a learnable scaling parameter plus a learnable offset parameter to generate a normalized concatenated sentence semantic vector sequence.

[0087] Step S1453: Determine the normalized concatenated sentence semantic vector sequence as the pollution causal enhanced semantic representation, which includes the normalized concatenated sentence semantic vector sequence corresponding to each original monitoring data sentence unit and the context-enhanced semantic association matrix between sentences.

[0088] The normalized concatenated sentence semantic vector sequence generated in step S1452 is determined as the pollution causal enhanced semantic representation. At the same time, the inter-sentence bidirectional attention weight matrix output by the last layer of the inter-sentence bidirectional attention interaction module is output as the inter-sentence context enhanced semantic association matrix.

[0089] Step S150: Based on the enhanced semantic representation of pollution causality and the set of pollutant indicator attribute relationship edges in the preset knowledge graph structure, perform semantic generation path planning processing to generate a semantic text sequence of environmental analysis report. The semantic text sequence of environmental analysis report includes a pollution cause analysis paragraph, a pollution diffusion trend description paragraph, and a monitoring conclusion summary paragraph.

[0090] Using the pollution causal enhanced semantic representation generated in step S1453 and the pollutant indicator attribute relationship edge set in the preset knowledge graph structure loaded in step S110, a semantic generation path is planned and an environmental analysis report text is generated.

[0091] Step S151: Extract the attribute relationship type sequence and attribute relationship direction sequence between pollutant indicators from the pollutant indicator attribute relationship edge set of the preset knowledge graph structure. The attribute relationship type sequence includes positive correlation relationship edges and negative correlation relationship edges of pollutant indicators.

[0092] Extract all attribute relationship edges from the set of pollutant indicator attribute relationship edges within the predefined knowledge graph structure. Each attribute relationship edge has an attribute relationship type of "positive correlation" or "negative correlation," and the attribute relationship direction points from the starting pollutant indicator node to the ending pollutant indicator node. Summarize the attribute relationship types and attribute relationship directions of all attribute relationship edges to generate an attribute relationship type sequence and an attribute relationship direction sequence.

[0093] Step S152: Convert the attribute relationship type sequence and attribute relationship direction sequence into a set of pollutant indicator relationship semantic path templates. Each pollutant indicator relationship semantic path template contains a starting pollutant indicator node, an ending pollutant indicator node, and a relationship semantic label.

[0094] Traverse the sequence of attribute relationship types and the sequence of attribute relationship directions. For each attribute relationship edge, construct a pollutant indicator relationship semantic path template. The structure of the pollutant indicator relationship semantic path template is in the form of a triple, with three fields: the starting pollutant indicator node name, the ending pollutant indicator node name, and the relationship semantic label. The relationship semantic label is generated according to the attribute relationship type; the relationship semantic label for positive correlation is "increase leads to increase," etc., and the relationship semantic label for negative correlation is "increase leads to decrease," etc. All pollutant indicator relationship semantic path templates constitute a pollutant indicator relationship semantic path template set.

[0095] Step S153: Perform pollutant index mention detection processing on the normalized concatenated statement semantic vector sequence in the pollution causal enhanced semantic representation, identify the pollutant index mention identifier corresponding to each normalized concatenated statement semantic vector, and generate a pollutant index mention identifier sequence.

[0096] For each normalized concatenated semantic vector in the normalized concatenated semantic vector sequence, a pollutant indicator mention detection classifier is used. This classifier consists of a fully connected layer followed by a sigmoid function, with the output dimension of the fully connected layer equal to the total number of pollutant indicators. The classifier outputs a multi-label probability vector, where each dimension corresponds to a pollutant indicator, and the probability value indicates whether the statement mentions that pollutant indicator. Pollutant indicators with probability values ​​exceeding a preset threshold are used as pollutant indicator mention identifiers for that statement. All pollutant indicator mention identifiers for all statements are aggregated to generate a pollutant indicator mention identifier sequence.

[0097] Step S154: Match the pollutant indicator mention identifier sequence with the starting pollutant indicator node and the ending pollutant indicator node in the pollutant indicator relation semantic path template set to construct a statement relation semantic path graph network with the original monitoring data statement unit as the node and the pollutant indicator relation semantic path template as the edge.

[0098] Each original monitoring data statement unit is used as a statement node in the statement relation semantic path graph network. For each template in the pollutant indicator relation semantic path template set, if the pollutant indicator mention identifier of a statement node contains the starting pollutant indicator node of the template, and the pollutant indicator mention identifier of another statement node contains the ending pollutant indicator node of the template, then a directed edge is established between these two statement nodes. The edge direction is from the statement node containing the starting pollutant indicator node to the statement node containing the ending pollutant indicator node, and the attribute of the edge is the relation semantic label of that template. All statement nodes and directed edges constitute the statement relation semantic path graph network.

[0099] Step S155: Perform graph traversal path search processing on the statement relation semantic path graph network. Starting from the starting statement node containing the pollution source entity mention identifier, perform depth-first path traversal along the edge direction of the pollutant index relation semantic path template to generate a set of statement semantic path candidate sequences.

[0100] In the statement relation semantic path graph network, statement nodes containing pollution source entity mention identifiers are identified as starting statement nodes. Starting from each starting statement node, a depth-first path traversal is performed along the edge directions. During the traversal, the sequence of statement nodes traversed and the sequence of relation semantic labels of the edges are recorded. Traversal of a path terminates when it reaches a preset maximum depth or cannot be extended further. All generated statement node sequences are collected as candidate sequences of statement semantic paths, and a set of candidate sequences of statement semantic paths is generated.

[0101] Step S156: Perform path semantic coherence scoring on each statement semantic path candidate sequence in the statement semantic path candidate sequence set, calculate the semantic similarity of the normalized concatenated statement semantic vectors between adjacent statement nodes in the path and the overall pollution causal logic consistency score of the path, and obtain the path semantic coherence score set.

[0102] For each candidate semantic path sequence in the set of candidate semantic paths, normalized concatenated semantic vectors corresponding to two adjacent statement nodes on the path are extracted. The causal semantic correlation is calculated as follows: the normalized concatenated semantic vectors corresponding to two adjacent statement nodes are input into a pre-trained causal semantic correlation discriminant network. This discriminant network is used to determine whether there is a causal succession relationship between the two statements and outputs the causal correlation confidence score. The causal correlation confidence score is used as the causal semantic correlation score between adjacent statements. The average of the causal semantic correlation scores of all adjacent statements on the path is used as a component of the path's semantic coherence score. Simultaneously, the attribute sequence, i.e., the relational semantic label sequence, is extracted from the path and matched with the corresponding pollution causal transmission feature sequence between statements in the pollution causal chain feature representation. The consistency rate between the relational semantic labels and the causal transmission direction is calculated as the pollution causal logical consistency score. The path semantic coherence score is equal to the weighted sum of the semantic coherence score component and the pollution causal logical consistency score. After scoring all candidate semantic path sequences, a path semantic coherence score set is obtained.

[0103] Step S157: Based on the path semantic coherence score set, perform path optimization processing on the statement semantic path candidate sequence set, and select the statement semantic path candidate sequence with the highest path semantic coherence score as the semantic path sequence for pollution cause analysis.

[0104] From the set of candidate semantic path sequences for statements, the candidate semantic path sequence with the highest path semantic coherence score is selected as the semantic path sequence for pollution cause analysis.

[0105] Step S158: Input the normalized concatenated semantic vectors corresponding to each statement node in the pollution cause analysis semantic path sequence into a preset paragraph generation recurrent neural network decoder, perform pollution cause analysis paragraph decoding and generation, and pollution diffusion trend and monitoring conclusion paragraph generation based on path extension, to generate an environmental analysis report semantic text sequence.

[0106] Using the semantic path sequence of pollution cause analysis selected in step S157, the text of each paragraph is generated by a preset paragraph generation recurrent neural network decoder.

[0107] Step S1581: Input the normalized concatenated semantic vectors corresponding to each statement node in the pollution cause analysis semantic path sequence into a preset paragraph generation recurrent neural network decoder to perform pollution cause analysis paragraph decoding and generation processing, and generate a pollution cause analysis paragraph text sequence.

[0108] The pre-defined paragraph generation recurrent neural network decoder employs a language generation model based on a long short-term memory network. The normalized concatenated semantic vectors corresponding to each sentence node in the pollution cause analysis semantic path sequence are sequentially input into the decoder. At each time step, the decoder generates a word unit, and the word unit generation probability is calculated from the hidden state of the long short-term memory network through a fully connected layer and softmax. A beam search strategy is used for decoding, with the beam width set to a preset value, generating multiple candidate pollution cause analysis paragraph text sequences. The candidate with the highest overall probability is selected as the pollution cause analysis paragraph text sequence.

[0109] Step S1582: Perform path extension and expansion processing on the semantic path sequence of pollution cause analysis. Based on the set of pollution propagation path edges in the semantic subgraph network of pollution diffusion mode, extend and expand the semantic path of the statement outward to generate a semantic path sequence of pollution diffusion trend and a semantic path sequence of monitoring conclusion summary.

[0110] Starting from the terminal statement node of the pollution cause analysis semantic path sequence, the system extends outward along the pollution propagation path edge in the statement relationship semantic path graph network. During this extension, statement nodes related to the pollution receptor node and containing descriptions of pollutant concentration change trends are selected to form a pollution diffusion trend semantic path sequence. Based on this sequence, statement nodes containing summary descriptions such as monitoring conclusions and exceedance determinations are further extracted to form a monitoring conclusion summary semantic path sequence.

[0111] Step S1583: Input the pollution diffusion trend semantic path sequence and the monitoring conclusion summary semantic path sequence into the paragraph generation recurrent neural network decoder for paragraph decoding and generation processing, generate the pollution diffusion trend description paragraph text sequence and the monitoring conclusion summary paragraph text sequence, and combine the pollution cause analysis paragraph text sequence, the pollution diffusion trend description paragraph text sequence and the monitoring conclusion summary paragraph text sequence into the environmental analysis report semantic text sequence.

[0112] The semantic path sequence of pollution diffusion trend is input into the paragraph generation recurrent neural network decoder, and the same decoding process as in step S1581 is used to generate the text sequence of the pollution diffusion trend description paragraph. The semantic path sequence of monitoring conclusion summary is input into the paragraph generation recurrent neural network decoder to generate the text sequence of the monitoring conclusion summary paragraph. The text sequences of pollution cause analysis paragraph, pollution diffusion trend description paragraph, and monitoring conclusion summary paragraph are concatenated in sequence, with paragraph separators inserted in between, to generate the semantic text sequence of the environmental analysis report. The semantic text sequence of the environmental analysis report is the final output environmental monitoring semantic report text.

[0113] After step S150 is completed, the generated semantic text sequence of the environmental analysis report is used as the base report, and subsequent enhancement processing steps are performed to enrich the report content and improve the report's professionalism and readability.

[0114] For example, the method may further include: step S210: performing semantic constraint processing based on the semantic text sequence of the environmental analysis report and the feature representation of the pollution causal chain to generate a semantic constraint vector sequence, and inputting the semantic constraint vector sequence and the set of original environmental monitoring records to be processed into a preset semantic refinement network for semantic refinement processing to generate a refined report text sequence.

[0115] The semantic text sequence of the environmental analysis report generated in step S150 and the pollution causal chain feature representation generated in step S130 are subjected to semantic constraint processing. The semantic constraint processing employs a gating constraint mechanism, multiplying the semantic embedding vector of each word in the environmental analysis report semantic text sequence element-wise with the pollution causal chain local feature representation at the corresponding time step in the pollution causal chain feature representation to generate a semantic constraint vector sequence. This semantic constraint vector sequence enhances the correlation with the pollution causal relationship while preserving the original semantics. Next, the semantic constraint vector sequence and the original environmental monitoring record text set to be processed are input into a preset semantic refinement network. The preset semantic refinement network adopts a text generation model based on an encoder-decoder architecture. The encoder reads the original text features of the original environmental monitoring record text set to be processed, and the decoder uses the semantic constraint vector sequence as the initial hidden state to generate refined text word-by-word. During the refinement process, the decoder calculates the attention weight between the current decoding hidden state and the encoder output at each generation step, copying or generating words from the original text to generate a refined report text sequence with more fluent sentences and more standardized terminology.

[0116] Step S220: Perform concentration value marking processing on the refined report text sequence, extract concentration value units and concentration unit units from the original environmental monitoring record text set to be processed, and bind the concentration value units and concentration unit units to the indicator mention positions in the refined report text sequence to generate a value-bound report sequence.

[0117] Concentration numerical units and concentration unit units are extracted from the raw environmental monitoring record text set using regular expression matching. The regular expression pattern is: number followed by an optional decimal point, then a number followed by the concentration unit. The extracted concentration numerical units and concentration unit units are then paired. The pollutant indicator mention positions are located in the refined report text sequence, identified by the pollutant indicator mention detection classifier in step S153. The concentration numerical pairs are then bound to the nearest pollutant indicator mention position by inserting the concentration numerical pair text after the pollutant indicator name in the refined report text sequence. After binding all concentration numerical pairs, a numerical binding report sequence is generated.

[0118] Step S230: Perform grade inference processing on the numerical binding report sequence, call the preset grade inference rule library, and perform grade category inference based on the comparison relationship between the concentration numerical unit and the preset concentration threshold to generate a grade label sequence.

[0119] The system invokes a pre-defined level inference rule library, which stores the national surface water environmental quality standard concentration thresholds for various pollutant indicators. For each concentration value pair in the numerical binding report sequence, the pollutant indicator name and concentration value are extracted, and the corresponding pre-defined concentration threshold for that pollutant indicator is searched in the pre-defined level inference rule library. The concentration value is compared with the concentration thresholds of various water quality standards to determine the corresponding water quality level category, such as Class I, Class II, Class III, Class IV, Class V, or worse than Class V. The water quality level category is converted into level-labeled text fragments, and all level-labeled text fragments are arranged in the order of their appearance in the text to generate a level-labeled sequence.

[0120] Step S240: Perform hierarchical fusion processing on the level marker sequence and the numerical binding report sequence, and inject the level marker sequence into the report paragraph level and the statement level to generate a level-enhanced report sequence.

[0121] At the paragraph level, the water quality grade of the paragraph topic is determined based on the paragraph content, and the corresponding grade marker text fragment is inserted as the paragraph's opening introductory text. At the statement level, for each statement containing concentration value pairs, the corresponding grade marker text fragment is appended to the end of the statement. After completing the injection at the paragraph and statement levels, a grade enhancement report sequence is generated.

[0122] Step S250: Perform structural reorganization processing on the grade enhancement report sequence. Based on the order of the causal analysis paragraph, trend description paragraph, and conclusion summary paragraph, and combined with the grade tag sequence, optimize the sorting of statements within the paragraphs to generate a structural reorganization sequence.

[0123] In the graded enhancement report sequence, the order of statements within each paragraph is optimized. The sorting priority is as follows: statements marked as high pollution level are sorted first, followed by statements marked as low pollution level; within the same pollution level, they are sorted in descending order of the causal transmission strength in the pollution causal chain feature representation. Simultaneously, logical connectors are added between paragraphs to ensure a natural and coherent transition from the causal analysis paragraph to the trend description paragraph to the conclusion summary paragraph. After completing the sorting and connector processing, a structural reorganization sequence is generated.

[0124] Step S260: Perform spatial association processing on the structural reorganization sequence, extract the spatial coordinate information of the point entity node set in the preset knowledge graph structure, convert the spatial coordinate information into spatial orientation description fragments, and embed the spatial orientation description fragments into the point mention positions in the structural reorganization sequence to generate a spatial enhancement report sequence.

[0125] Spatial coordinate information (longitude and latitude) of each monitoring point entity node is extracted from the set of monitoring point entity nodes within a pre-defined knowledge graph structure. The spatial orientation conversion module is then invoked to convert the spatial coordinate information into spatial orientation description fragments relative to a reference point within the industrial park, such as "northwest of the park" or "approximately several kilometers downstream," using natural language descriptions. In the structural reconstruction sequence, the location reference for each monitoring point entity node is identified, and the corresponding spatial orientation description fragment is embedded after that location reference as supplementary spatial information. A spatial enhancement report sequence is then generated.

[0126] Step S270: Perform cross-media propagation processing on the spatial enhancement report sequence, extract the cross-media propagation edge set from the diffusion pattern subgraph network of the preset knowledge graph structure, convert the cross-media propagation edge set into cross-media description fragments, and append the cross-media description fragments to the corresponding diffusion description statements to generate the cross-media enhancement report sequence.

[0127] A set of cross-media propagation edges is extracted from the semantic subgraph network of pollution diffusion patterns within a predefined knowledge graph structure. Each edge in the set represents the migration path of pollutants from one medium to another, such as "water to sediment" or "atmosphere to water." Each cross-media propagation edge is converted into a cross-media description fragment, with the template for generating the fragment being "pollutants migrate from the source medium to the target medium via propagation." In the spatially enhanced reporting sequence, diffusion description statements are located, and the corresponding cross-media description fragments are appended after them. This generates the cross-media enhanced reporting sequence.

[0128] Step S280: Perform coherent smoothing processing on the cross-media enhanced report sequence, predict the semantic breakage degree of paragraph connection boundary through a pre-trained environmental text language model, generate transitional semantic text fragments at the semantic breakage boundary, and generate a coherent smooth report sequence.

[0129] A pre-trained environmental text language model is loaded. This model employs a unidirectional language model structure based on a Transformer decoder and has been fine-tuned on environmental monitoring report corpora. For the connection boundaries between each pair of adjacent paragraphs in the cross-media enhanced report sequence, the text fragments before and after the boundary are input into the pre-trained environmental text language model, and the perplexity of the next word at the boundary is calculated. Higher perplexity indicates a greater degree of semantic breakage. For semantic break boundaries with perplexity exceeding a preset threshold, the pre-trained environmental text language model is invoked to generate transitional semantic text fragments. The generation method involves using the last few words of the previous paragraph as context, generating transitional sentences word by word until the generation probability of the first few words of the next paragraph exceeds the threshold. The generated transitional semantic text fragments are then inserted at the semantic break boundaries. This generates a coherent and smooth report sequence.

[0130] Step S290: Perform timeliness annotation processing on the coherent and smooth report sequence, extract the monitoring timestamp information and monitoring frequency information of the original environmental monitoring record text set to be processed, convert the monitoring timestamp information and monitoring frequency information into timeliness description text fragments, insert the timeliness description text fragments into the report header area, and generate a timeliness-annotated report sequence.

[0131] Extract monitoring timestamp information (start and end times of data collection) and monitoring frequency information (number of daily monitoring sessions or sampling intervals) from the raw environmental monitoring record text set to be processed. Convert the monitoring timestamp and frequency information into time-definite description text fragments, using the template "This report is generated based on monitoring data from [date] to [date], with a monitoring frequency of [number] times per day." Insert these time-definite description text fragments into the header area of ​​a coherent and smooth report sequence as a description of the report's timeliness. Generate a time-definite labeled report sequence.

[0132] Step S2100: Determine the timeliness-labeled report sequence as the final output environmental monitoring report semantic sequence. The environmental monitoring report semantic sequence includes a causal analysis paragraph, a trend description paragraph, a conclusion summary paragraph, a grade label, a spatial orientation description, a cross-media description, and a timeliness description.

[0133] The timeliness labeling report sequence generated in step S290 is used as the final output environmental monitoring report semantic sequence, which contains all the semantic components required by the claims.

[0134] Step S310: Extract the time-series numerical sequence of indicators from the original text set of environmental monitoring records to be processed. The time-series numerical sequence of indicators contains concentration numerical records at continuous monitoring time points. Perform attribute association processing on the time-series numerical sequence of indicators and the set of indicator attribute edges in the preset knowledge graph structure to generate an attribute-associated numerical sequence.

[0135] The time-series numerical sequences of indicators for each monitoring point are extracted from the original environmental monitoring record text set in timestamp order. These time-series numerical sequences are a two-dimensional data structure, with rows corresponding to monitoring time points and columns corresponding to pollutant indicator types. The time-series numerical sequences are then associated with the pollutant indicator attribute relationship edge set in a pre-defined knowledge graph structure. Specifically, for each pair of pollutant indicators connected by an attribute relationship edge, the time-series numerical pairs of these two pollutant indicators are extracted from the time-series numerical sequences to form an attribute-associated numerical sequence.

[0136] Step S320: Perform trend analysis processing on the attribute-related numerical sequence, segment the attribute-related numerical sequence by using a preset time-series segmentation algorithm to obtain time-series segment intervals, and extract the change slope features and fluctuation amplitude features of the time-series segment intervals to generate a segmented trend feature set.

[0137] A pre-defined time-series segmentation algorithm—a time-series segmented linear representation algorithm based on inflection points—is used to segment the attribute-related numerical sequences. This algorithm divides the time-series data into several linear trend segments by detecting local extrema and trend change points. For each time-series segment interval, a linear regression line is fitted using the least squares method, and the slope of the line is used as the slope characteristic of that time-series segment interval. The standard deviation of the values ​​within each time-series segment interval is calculated as the fluctuation amplitude characteristic. The slope characteristics and fluctuation amplitude characteristics of all time-series segment intervals are summarized to generate a set of segmented trend features.

[0138] Step S330: Perform trend fusion processing on the segmented trend feature set and the pollution causal chain feature representation, and use a preset trend attention network to perform attention weighting on the change slope feature and fluctuation amplitude feature of the segmented trend feature set to generate a trend fusion feature vector.

[0139] The trend attention network is pre-defined as a multi-head attention layer. The causal transmission strength sequence in the pollution causal chain feature representation is used as the query vector, and the concatenated vectors of the slope and fluctuation amplitude features of each segment in the segmented trend feature set are used as the key and value vectors, respectively. Attention weights are then calculated. These attention weights are used to weighted aggregate the segmented trend features to generate a trend fusion feature vector. This trend fusion feature vector integrates causal chain information and trend statistics.

[0140] Step S340: Perform trend semantic injection processing on the trend fusion feature vector and the semantic text sequence of the environmental analysis report, and inject the trend description text fragment corresponding to the trend fusion feature vector into the trend description paragraph of the semantic text sequence of the environmental analysis report to generate a trend-enhanced report sequence.

[0141] The trend fusion feature vector is input into a trend description text generator, which is a text decoder based on a Long Short-Term Memory (LSTM) network. This decoder uses the trend fusion feature vector as its initial hidden state to generate trend description text fragments. These fragments contain trend descriptions such as "the indicator is trending upwards or downwards" and "the rate of change is fast or slow." Within the trend description paragraphs of the semantic text sequence in the environmental analysis report, the insertion points for these trend descriptions are located, and the trend description text fragments are injected. This generates a trend-enhanced report sequence.

[0142] Step S350: Perform causal explanation supplementation processing on the trend description text fragments in the trend enhancement report sequence, extract causal explanation statement fragments from the pollution causal chain feature representation, and concatenate the causal explanation statement fragments to the trend description text fragments to generate the causal explanation enhancement report sequence.

[0143] From the local feature representation sequence of the pollution causal chain corresponding to each statement in the pollution causal chain feature representation, extract the text content corresponding to the statements with the highest causal transmission strength values ​​as causal explanation statement fragments. In the trend enhancement report sequence, for each trend description text fragment, concatenate the causal explanation statement fragment that matches its time to the trend description text fragment to form a causal explanation structure of "trend description, cause lies in causal explanation". Generate the causal explanation enhancement report sequence.

[0144] Step S360: Perform abnormal fluctuation annotation processing on the causal explanation enhancement report sequence, identify abnormal segment intervals whose fluctuation amplitude features exceed the preset fluctuation threshold from the segmented trend feature set, generate abnormal annotation text fragments for the abnormal segment intervals, and attach the abnormal annotation text fragments to the corresponding time series description statements to generate an abnormal annotation report sequence.

[0145] Set a preset fluctuation threshold. Iterate through the fluctuation amplitude features of each time series segment in the segmented trend feature set, marking time series segment intervals whose fluctuation amplitude features exceed the preset fluctuation threshold as abnormal segment intervals. For abnormal segment intervals, generate anomaly annotation text fragments, containing descriptions such as "data fluctuations are abnormal during this period" or "fluctuation amplitude is significantly larger than normal". In the causal explanation enhancement report sequence, locate the time series description statement corresponding to the abnormal segment interval, and append the anomaly annotation text fragment to the end of that time series description statement. Generate anomaly annotation report sequence.

[0146] Step S370: Perform point comparison analysis on the anomaly annotation report sequence, extract the time series numerical sequence of indicators of entity nodes at different points in the preset knowledge graph structure, perform comparison difference calculation on the segmented trend feature set of entity nodes at different points, generate point comparison difference text fragments, insert the point comparison difference text fragments into the conclusion summary paragraph, and generate point comparison report sequence.

[0147] Extract the time-series numerical sequences of indicators for different monitoring point entity nodes from the set of monitoring point entity nodes in the preset knowledge graph structure. Perform trend analysis processing in step S320 on each monitoring point entity node to obtain the segmented trend feature set of each monitoring point entity node. Calculate the difference in the slope characteristics of the same indicator for the same time period between any two monitoring point entity nodes, and select several comparison pairs with the largest absolute difference values. Convert the comparison differences into text fragments of point comparison differences, with the conversion template being "The rate of increase of chemical oxygen demand at monitoring point M01 is higher than that at monitoring point M05 by a certain amount". Summarize all the text fragments of point comparison differences and insert them into the conclusion summary paragraph. Generate a sequence of point comparison report sequences.

[0148] Step S380: Perform historical comparison processing on the point comparison report sequence, extract the historical indicator time series value sequence from the historical environmental monitoring original record text set, calculate the year-on-year deviation between the historical indicator time series value sequence and the indicator time series value sequence, generate historical comparison text fragments, and attach the historical comparison text fragments to the trend description paragraph to generate a historical comparison report sequence.

[0149] Extract the historical environmental monitoring raw record text set for the same period last year from the historical database of the environmental monitoring data management platform, and execute step S310 to extract the historical indicator time series numerical sequence. For each pollutant indicator, calculate the year-on-year deviation between the average value of the current indicator time series numerical sequence and the average value of the historical indicator time series numerical sequence. The year-on-year deviation is equal to the current average value minus the historical average value divided by the historical average value. Convert the year-on-year deviation into a historical comparison text fragment, with the conversion template being "Compared with the same period last year, the chemical oxygen demand concentration increased or decreased by how many percent?" Append the historical comparison text fragment to the trend description statement of the corresponding pollutant indicator in the trend description paragraph. Generate the historical comparison report sequence.

[0150] Step S390: Perform indicator association interpretation processing on the historical comparison report sequence, extract indicator association rules from the indicator attribute edge set of the preset knowledge graph structure, convert the indicator association rules into association interpretation text fragments, and embed the association interpretation text fragments into the causal analysis paragraph to generate an association interpretation report sequence.

[0151] Extract all attribute relationship edges from the pollutant indicator attribute relationship edge set of the pre-defined knowledge graph structure. Each attribute relationship edge represents an indicator association rule. Convert these association rules into association explanation text fragments, using the template: "Generally, an increase in the initial pollutant indicator will lead to an increase or decrease in the final pollutant indicator, with a positive or negative correlation between the two." In the causal analysis paragraphs of historical comparison report sequences, find statements involving the initial or final pollutant indicators in the association rules and embed the association explanation text fragments near these statements. Generate an association explanation report sequence.

[0152] Step S3100: Perform comprehensive analysis on the associated interpretation report sequence, extract comprehensive analysis keywords from the grade marker sequence, point comparison difference text fragment, historical comparison text fragment, and associated interpretation text fragment, combine the comprehensive analysis keywords into a comprehensive analysis conclusion text fragment, and append the comprehensive analysis conclusion text fragment to the end of the conclusion summary paragraph to generate a comprehensive analysis report sequence.

[0153] The most frequently occurring grade marker keywords are extracted from the grade marker sequence. Keywords of pollutant indicators and monitoring points with the most significant differences are extracted from the text fragments comparing site differences. Keywords of pollutant indicators with the largest year-on-year deviation are extracted from the text fragments comparing historical data from the same period. Keywords describing the most significant relationships are extracted from the text fragments explaining the relationships. These extracted keywords are then combined according to a preset template to form a comprehensive assessment conclusion text fragment. The template is: "Comprehensive analysis shows that the pollutant indicator name of the monitoring site has the highest pollution level, and the increase or decrease compared to the same period last year is mainly attributed to the relationship description." This comprehensive assessment conclusion text fragment is appended to the end of the conclusion summary paragraph. A comprehensive assessment report sequence is generated. This comprehensive assessment report sequence is an enhanced environmental monitoring report with added indicator trend analysis, anomaly annotation, site comparison, historical comparison, and indicator relationship explanation.

[0154] Step S410: Extract sensory description text fragments from the original environmental monitoring record text set to be processed. The sensory description text fragments include water color description text, water odor description text, and water transparency description text. Perform semantic normalization processing on the sensory description text fragments to generate a normalized sensory description sequence.

[0155] Sensory description text fragments were extracted from the original environmental monitoring records to be processed using a pattern recognition method based on keyword matching and rules. A sensory description keyword dictionary was constructed, including keywords related to water color such as "light yellow," "yellow-green," and "dark brown," keywords related to water odor such as "odorless," "fishy smell," and "irritating smell," and keywords related to water transparency such as "clear," "turbid," and "transparent to the bottom." The entire text was scanned using this keyword dictionary to locate sentence fragments containing the aforementioned keywords, and these fragments were extracted as sensory description text fragments. The extracted sensory description text fragments underwent semantic normalization, mapping colloquial and non-standardized sensory description terms to standardized environmental monitoring terminology. For example, "black" was mapped to "deep black," "very smelly" to "has a distinctly foul odor," and "very low visibility" to "extremely low transparency." The mapping process was based on a pre-compiled standardized sensory description terminology mapping table. The standardized sensory description sequence was arranged in the order of its appearance in the original text.

[0156] Step S420: Extract the set of pollution source industry category nodes from the diffusion pattern subgraph network of the preset knowledge graph structure, perform industry feature association processing on the set of pollution source industry category nodes and the standardized sensory description sequence, and match the standardized sensory description sequence with the pollution source industry category through the preset industry feature matching library to generate a suspected pollution source labeling sequence.

[0157] A set of industry category nodes for pollution sources is extracted from the semantic subgraph network of pollution diffusion patterns within a pre-defined knowledge graph structure. Each industry category node contains an industry category name and typical sensory characteristic description attributes. A pre-defined industry feature matching library stores the correspondence between each pollution source industry category and typical sensory description features. For example, "dyeing and printing industry" corresponds to "water color is bluish or reddish, with a dye odor," and "livestock and poultry farming industry" corresponds to "water is yellowish-brown, with a distinct fecal odor." The sensory description features in each set of standardized sensory descriptions are matched with the typical sensory features of each pollution source industry category in the pre-defined industry feature matching library. The similarity is calculated using the Jakarta similarity coefficient, which is the intersection of the two sets divided by the union. The pollution source industry category with the highest similarity is selected as the matching result, generating a suspected pollution source labeling sequence. Each element in this suspected pollution source labeling sequence contains a standardized sensory description fragment and its matched pollution source industry category name.

[0158] Step S430: Perform causal source tracing reasoning processing on the industry suspected pollution source labeling sequence and the pollution causal chain feature representation. Validate the causal transmission path between the suspected pollution sources and the pollution causal chain feature representation in the industry suspected pollution source labeling sequence through a preset source tracing reasoning network, and generate a pollution source tracing path verification sequence.

[0159] The pre-defined source tracing inference network is a graph attention network-based inference model. The pollution source nodes corresponding to the suspected pollution source industry categories in the pollution diffusion pattern semantic subgraph network from the industry suspected pollution source labeling sequence are used as starting nodes, and the pollution receptor node with the highest causal transmission strength in the pollution causal chain feature representation is used as the ending node. All possible paths from the starting node to the ending node are searched in the pollution diffusion pattern semantic subgraph network, with paths connected by pollution propagation path edges. For each path, the source tracing confidence is calculated, which is equal to the product of the edge weights of all propagation edges on the path. Paths with source tracing confidence exceeding a pre-defined source tracing verification threshold are considered verified pollution source tracing paths, generating a pollution source tracing path verification sequence. This sequence contains a sequence of verified path nodes and the path source tracing confidence.

[0160] Step S440: Perform source tracing information embedding processing on the pollution source tracing path verification sequence and the environmental analysis report semantic text sequence, embed the source tracing description text fragment corresponding to the pollution source tracing path verification sequence in the cause analysis paragraph, and generate a source tracing enhancement report sequence.

[0161] The verified pollution source tracing paths in the verification sequence are converted into source tracing description text fragments. The conversion template is: "Based on sensory characteristics and pollution diffusion path analysis, the pollution source is initially traced back to the industry category of the pollution source. The pollutants diffused from the pollution source name to the monitoring point name via the transmission medium." In the causal analysis paragraph of the semantic text sequence of the environmental analysis report, the sentences related to source tracing are located, and the source tracing description text fragments are embedded into those locations. An enhanced source tracing report sequence is then generated.

[0162] Step S450: Perform risk level assessment processing on the source tracing enhancement report sequence, classify the risk level according to the path confidence score of the level label sequence and the pollution source tracing path verification sequence, generate risk level assessment text fragments, insert the risk level assessment text fragments into the conclusion summary paragraph, and generate a risk assessment report sequence.

[0163] The highest pollution level marker for each monitoring point is extracted from the level marker sequence, and the highest path source tracing confidence value is extracted from the pollution source tracing path verification sequence. Risk level classification uses a weighted comprehensive scoring method. The risk comprehensive score R1 is calculated as: R1 = A1 × L1 + A2 × C1, where A1 is the pollution level weight, L1 is the numerical score of the pollution level, A2 is the source tracing weight, and C1 is the path source tracing confidence. Risk levels are classified according to the numerical range of R1, and the risk levels are converted into risk level assessment text fragments. The conversion template is "Comprehensive assessment determines that the water environment pollution risk level of this monitoring area is a risk level." The risk level assessment text fragments are inserted into the conclusion summary paragraph of the source tracing enhancement report sequence. A risk assessment report sequence is then generated.

[0164] Step S460: Process the risk assessment report sequence to generate governance suggestions. Call the preset governance measures knowledge base, match the corresponding governance measures text template according to the risk level assessment text fragment and the industry suspected pollution source labeling sequence, generate governance suggestion text fragments, and attach the governance suggestion text fragments to the end of the conclusion summary paragraph to generate the governance suggestion report sequence.

[0165] A pre-defined governance measures knowledge base stores text templates for governance measures corresponding to different risk levels and different industry categories of pollution sources. Based on the risk level determined in step S450 and the industry category of the suspected pollution source identified in step S420, a matching search is performed in the pre-defined governance measures knowledge base, using a combined query method of risk level and industry category. The matched governance measures text templates are extracted, and the parameter placeholders in the templates are filled to generate governance suggestion text fragments. These fragments include sub-descriptions of source control suggestions, process interruption suggestions, and end-of-pipe treatment suggestions. The governance suggestion text fragments are appended to the end of the conclusion summary paragraph of the risk assessment report sequence. A governance suggestion report sequence is then generated.

[0166] Step S470: Perform regulatory clause association processing on the governance suggestion report sequence, call the preset environmental protection law knowledge base, match the corresponding regulatory clause text according to the industry suspected pollution source labeling sequence and grade labeling sequence, generate regulatory clause reference text fragments, and embed the regulatory clause reference text fragments into the cause analysis paragraph to generate a regulatory association report sequence.

[0167] A pre-defined environmental regulations knowledge base stores the texts of environmental laws and regulations corresponding to various pollution source industry categories and water pollution levels. Based on the pollution source industry category in the suspected pollution source labeling sequence and the pollution level label in the level labeling sequence, a matching search is performed in the pre-defined environmental regulations knowledge base. The matching regulatory clause texts are extracted, including the regulation name, clause number, and specific content of the clause. The regulatory clause texts are converted into regulatory clause reference text fragments, with the conversion template being "According to Article X of the Regulation Name, a summary of the clause content." In the causal analysis section of the governance recommendation report sequence, the sentences related to the regulatory basis are located, and the regulatory clause reference text fragments are embedded. A regulatory association report sequence is generated.

[0168] Step S480: Process the monitoring frequency adjustment suggestion for the regulatory-related report sequence. Based on the fluctuation amplitude characteristics and risk level assessment text fragments of the segmented trend feature set, calculate the frequency adjustment of the monitoring frequency information, generate a monitoring frequency adjustment suggestion text fragment, and insert the monitoring frequency adjustment suggestion text fragment into the timeliness description area to generate a frequency adjustment report sequence.

[0169] The fluctuation amplitude characteristics of each pollutant indicator at each monitoring point are extracted from the segmented trend feature set, with the fluctuation amplitude characteristic value taken as the standard deviation. Risk levels are extracted from the risk level assessment text fragment. The monitoring frequency adjustment calculation uses a lookup table method: a pre-defined monitoring frequency adjustment rule table defines the recommended monitoring frequency multiples corresponding to different risk levels and different fluctuation amplitude ranges. For example, monitoring points with high risk levels and large fluctuation amplitudes are recommended to have their frequency adjusted to a certain multiple of the baseline frequency. The current monitoring frequency is multiplied by the recommended monitoring frequency multiple to obtain the adjusted monitoring frequency. The monitoring frequency adjustment result is converted into a monitoring frequency adjustment suggestion text fragment, with the conversion template being "It is recommended to adjust the monitoring frequency of the monitoring point name from the current frequency to the adjusted frequency." The monitoring frequency adjustment suggestion text fragment is inserted into the timeliness description area of ​​the regulatory-related report sequence. A frequency adjustment report sequence is generated.

[0170] Step S490: Perform report summary generation processing on the frequency adjustment report sequence, extract key sentences from the cause analysis paragraph, trend description paragraph and conclusion summary paragraph, combine the key sentences into report summary text fragments, and place the report summary text fragments in the report header area to generate a summary report sequence.

[0171] An extractive summarization algorithm based on text ranking is employed. Statements from the causal analysis, trend description, and conclusion summary paragraphs of the frequency adjustment report sequence are selected as candidate summary statements. For each candidate summary statement, an importance score is calculated. This score equals the semantic similarity between the candidate summary statement and the statement with the highest causal transmission strength in the pollution causal chain feature representation, plus the weighted score of the rank-marked keywords contained in the candidate summary statement, plus the weighted score of the statement's position within the paragraph. Candidate summary statements are sorted in descending order of importance score, and the top-ranked statements are selected as key statements. These key statements are then combined into report summary text fragments in the original text order. These report summary text fragments are placed at the beginning of the frequency adjustment report sequence, i.e., after the report title and before the main text paragraphs, to generate the summary report sequence.

[0172] Step S4100: Determine the summary report sequence as the final output complete environmental monitoring report semantic sequence. The complete environmental monitoring report semantic sequence includes a report summary, causal analysis paragraph, trend description paragraph, conclusion summary paragraph, level label, source tracing description, risk assessment description, governance recommendation description, regulatory reference description, and frequency adjustment recommendation description.

[0173] The summary report sequence generated in step S490 is used as the final output semantic sequence of the complete environmental monitoring report. This complete environmental monitoring report semantic sequence integrates the output information of all enhancement modules, including sensory analysis, source tracing reasoning, risk assessment, remediation recommendations, regulatory relevance, frequency adjustment recommendations, and report summary.

[0174] Step S510: Extract monitoring equipment identification information and monitoring method identification information from the original environmental monitoring record text set to be processed, match the monitoring equipment identification information and monitoring method identification information with the preset equipment method knowledge base, and generate a sequence of equipment method specification descriptions.

[0175] From the raw environmental monitoring records to be processed, named entity recognition technology is used to extract monitoring equipment identification information and monitoring method identification information. The monitoring equipment identification information includes the equipment name, equipment model, and equipment number; the monitoring method identification information includes the testing method standard number and testing method name. The extracted monitoring equipment identification information and monitoring method identification information are matched against a pre-set equipment and method knowledge base, which stores the standardized expressions of equipment standard names and testing method standard names. The matching process uses fuzzy matching and alias mapping techniques to map non-standard equipment names and method names to standardized names. A sequence of equipment and method specification descriptions is generated, where each element contains the equipment specification name and the method specification name.

[0176] Step S520: Perform equipment anomaly reasoning processing on the equipment method specification description sequence and the pollution causal chain feature representation. Use a preset equipment anomaly reasoning rule base to perform equipment anomaly causal reasoning on the abnormal fluctuation features of the monitoring equipment identification information and the pollution causal chain feature representation to generate an equipment anomaly reasoning result sequence.

[0177] A pre-defined equipment anomaly inference rule base stores causal inference rules between equipment failure modes and monitoring data anomaly modes. The rule format is "If the monitoring equipment experiences a certain type of failure, the monitoring data will exhibit a certain type of abnormal characteristic." The equipment type information in the equipment method specification description sequence is correlated with the abnormal fluctuation characteristics of the pollution causal transmission characteristic sequence between statements in the pollution causal chain feature representation. These abnormal fluctuation characteristics include data spikes, continuous data drift, and data exceeding detection limits. A forward linking inference algorithm is used, starting from the known equipment type, matching rules that meet the conditions in the pre-defined equipment anomaly inference rule base to deduce possible causes of equipment anomalies. An equipment anomaly inference result sequence is generated, containing possible equipment anomaly types and their corresponding inference confidence levels.

[0178] Step S530: Perform device anomaly semantic embedding processing on the device anomaly inference result sequence and the environmental analysis report semantic text sequence, embed the device anomaly description text fragment corresponding to the device anomaly inference result sequence in the cause analysis paragraph, and generate a device anomaly report sequence.

[0179] The device anomaly type with the highest confidence in the device anomaly inference result sequence is converted into a device anomaly description text fragment. The conversion template is "Based on equipment status analysis, the monitored device name may have an anomaly type, which may lead to abnormal characteristics in the monitoring data." Within the causal analysis paragraph of the environmental analysis report's semantic text sequence, the positions of statements related to monitoring data quality are located, and the device anomaly description text fragment is embedded. This generates a device anomaly report sequence.

[0180] Step S540: Perform point layout optimization reasoning processing on the equipment abnormality report sequence. Based on the monitoring point entity anchor point location distribution sequence and pollution diffusion semantic pattern matching response sequence, identify coverage blind spots in the spatial distribution of monitoring points through a preset spatial coverage analysis algorithm, generate point layout optimization suggestion text fragments, and attach the point layout optimization suggestion text fragments to the conclusion summary paragraph to generate a layout optimization report sequence.

[0181] Using the unique identifier sequence of monitoring point entity nodes in the monitoring point entity anchor point location distribution sequence, the spatial coordinates of each monitoring point are extracted from the monitoring point entity node set of a preset knowledge graph structure. A preset spatial coverage analysis algorithm—a Voronoi diagram-based spatial coverage analysis method—is employed to construct Voronoi polygon segments using the spatial coordinates of each monitoring point as generation points, and the area of ​​each Voronoi polygon is calculated. Voronoi polygon areas with areas exceeding a preset area threshold are identified as coverage blind spots. Simultaneously, by combining pollution diffusion semantic pattern matching information from the response sequence, it is determined whether the coverage blind spot is located in a critical segment of the pollution diffusion path. If so, a text fragment suggesting optimized monitoring point placement is generated, recommending the addition of monitoring points in the coverage blind spot. This optimized monitoring point placement text fragment is appended to the conclusion summary paragraph of the equipment anomaly report sequence, generating a deployment optimization report sequence.

[0182] Step S550: Perform data quality assessment processing on the deployment optimization report sequence. Based on the degree of missing data in the equipment anomaly inference result sequence and the indicator time series value sequence, conduct data quality level assessment, generate data quality assessment text fragments, insert the data quality assessment text fragments into the timeliness description area, and generate a quality assessment report sequence.

[0183] The equipment anomaly types and quantities are extracted from the equipment anomaly inference result sequence. The data missing rate is calculated from the time-series numerical sequence of indicators; the data missing rate equals the number of missing data points divided by the total number of data points. Data quality level assessment uses a rule-based inference method: if the equipment anomaly type includes a severe fault type or the data missing rate exceeds a preset missing rate threshold, it is rated as "poor data quality"; if the data missing rate is moderate but there are no severe equipment anomalies, it is rated as "moderate data quality"; all other cases are rated as "good data quality". The data quality level is converted into a data quality assessment text fragment. This text fragment is inserted into the time-description area of ​​the deployment optimization report sequence. A quality assessment report sequence is then generated.

[0184] Step S560: Perform pollution source analysis processing on the quality assessment report sequence, extract the pollution source analysis path set from the diffusion pattern subgraph network of the preset knowledge graph structure, perform path matching between the pollution source analysis path set and the pollution causal chain feature representation, generate pollution source analysis text fragments, and embed the pollution source analysis text fragments into the causal analysis paragraph to generate the source analysis report sequence.

[0185] From the semantic subgraph network of pollution diffusion patterns within a pre-defined knowledge graph structure, all paths from pollution source nodes to monitoring point entity nodes are extracted as a pollution source analysis path set. The node and edge sequences of each path in this set are matched against the causal transmission intensity and causal transmission direction sequences in the pollution causal chain feature representation. The matching score M2 is calculated as: M2 = R1 × I1, where R1 is the directional consistency rate and I1 is the average intensity. The pollution source analysis path with the highest matching score is selected as the optimal analysis path. The optimal analysis path is converted into a pollution source analysis text fragment, with the template "The pollutants at the monitoring point mainly originate from the pollution source name and diffuse to this location via the propagation path name." This pollution source analysis text fragment is embedded into the causal analysis paragraph of the quality assessment report sequence. A source analysis report sequence is then generated.

[0186] Step S570: Perform pollution flux estimation processing on the source analysis report sequence. Based on the time series numerical sequence of the indicators and the spatial coordinate information of the points in the set of entity nodes of the monitoring points, estimate the transmission flux of pollutants between the monitoring points, generate a pollution flux description text fragment, and attach the pollution flux description text fragment to the trend description paragraph to generate a flux estimation report sequence.

[0187] The spatial distance between monitoring points is calculated by combining the concentration values ​​of adjacent monitoring points in the time series of indicators with the spatial coordinate information of the monitoring point entity node set. Pollution flux estimation adopts the discrete approximation form of Fick's diffusion law. The flux estimation value F1 is calculated as: F1 = D2 × (U1 - U2) / D3, where D2 is the diffusion coefficient, U1 is the upstream concentration value, U2 is the downstream concentration value, and D3 is the spatial distance between adjacent points. The diffusion coefficient is obtained from a preset diffusion coefficient table based on the pollutant type. The calculated flux estimation value is converted into a pollution flux description text fragment, with the conversion template being "Pollutant from monitoring point A to monitoring point B: Pollutant Name, Transmission Flux, Approximately Estimated Flux Value". This pollution flux description text fragment is appended to the trend description paragraph of the source apportionment report sequence. A flux estimation report sequence is then generated.

[0188] Step S580: Perform pollution hotspot identification processing on the flux estimation report sequence. Use a preset spatial clustering algorithm to perform spatial clustering on the monitoring point entity anchor point location distribution sequence and grade label sequence to identify the spatial range of the pollution hotspot, generate a pollution hotspot description text fragment, and insert the pollution hotspot description text fragment into the trend description paragraph to generate a hotspot identification report sequence.

[0189] A pre-defined spatial clustering algorithm—a density-based spatial clustering algorithm—is employed. The spatial coordinates of each monitoring point in the monitoring point anchor location distribution sequence are used as cluster sample points, and the pollution level values ​​in the level label sequence are used as the sample point weights. The clustering algorithm groups monitoring points that are spatially close and have consistently high pollution levels into one cluster. The cluster with the most sample points or the cluster with the highest average pollution level is identified as a pollution hotspot. The spatial envelope coordinates of the pollution hotspots are extracted and converted into a descriptive text fragment, with the template "Pollution hotspots are identified as mainly located within the spatial range description area, where pollutant concentrations are generally high." This descriptive text fragment is inserted into the trend description paragraph of the flux estimation report sequence, generating a hotspot identification report sequence.

[0190] Step S590: Perform multi-media collaborative analysis on the hot zone identification report sequence, extract the multi-media collaborative pollution path set from the diffusion pattern subgraph network of the preset knowledge graph structure, convert the multi-media collaborative pollution path set into multi-media collaborative description text fragments, and attach the multi-media collaborative description text fragments to the cross-media description region to generate a multi-media collaborative report sequence.

[0191] From the semantic subgraph network of pollution diffusion patterns within a predefined knowledge graph structure, paths involving two or more propagation media are extracted as a set of multi-media collaborative pollution paths. A characteristic of multi-media collaborative pollution paths is that they contain cross-media propagation edges. Each path in the set is converted into a multi-media collaborative description text fragment, with the conversion template being "The pollutant name diffuses simultaneously to the receptor region through both media A and media B, exhibiting a multi-media collaborative pollution effect." This multi-media collaborative description text fragment is then appended to the cross-media description region of the hotspot identification report sequence. A multi-media collaborative report sequence is then generated.

[0192] Step S5100: Determine the multi-media collaborative report sequence as the final output semantic sequence of the comprehensive analysis environmental monitoring report. The semantic sequence of the comprehensive analysis environmental monitoring report includes a causal analysis paragraph, a trend description paragraph, a conclusion summary paragraph, an equipment anomaly description, a deployment optimization description, a quality assessment description, a source analysis description, a flux estimation description, a hot zone identification description, and a multi-media collaborative description.

[0193] The multi-media collaborative report sequence generated in step S590 is used as the final output semantic sequence of the comprehensive analysis environmental monitoring report.

[0194] Step S610: Extract meteorological parameter description fragments and hydrological parameter description fragments from the original environmental monitoring record text set to be processed, perform parameter semantic parsing processing on the meteorological parameter description fragments and hydrological parameter description fragments, and generate meteorological and hydrological parameter semantic sequences.

[0195] An information extraction method based on dictionaries and regular expressions is employed. The meteorological parameter dictionary contains keywords such as "temperature," "wind speed," "wind direction," and "precipitation," while the hydrological parameter dictionary contains keywords such as "water level," "flow velocity," "discharge rate," and "water temperature." From the raw environmental monitoring records to be processed, sentences containing these keywords are scanned, and complete meteorological and hydrological parameter description fragments are extracted. The extracted fragments undergo parameter semantic parsing, extracting the numerical values ​​and units from the text descriptions to form structured parameter records. A semantic sequence of meteorological and hydrological parameters is generated, where each element includes the parameter name, parameter value, parameter unit, and monitoring timestamp.

[0196] Step S620: Perform environmental-driven association processing on the semantic sequence of meteorological and hydrological parameters and the feature representation of the pollution causal chain. Infer the driving relationship between meteorological and hydrological parameters and the pollution diffusion causal chain through a preset environmental-driven inference network to generate an environmental driving factor weight sequence.

[0197] The pre-defined environment-driven inference network is a Bayesian network inference model. Its network structure predefines the probabilistic dependencies between meteorological and hydrological parameter nodes and pollution diffusion causal nodes. Each meteorological and hydrological parameter value in the semantic sequence of meteorological and hydrological parameters is input into the pre-defined environment-driven inference network as evidence nodes. The posterior probability of the pollution diffusion causal nodes in the network is updated through a belief propagation algorithm. The contribution weight W1 of each meteorological and hydrological parameter to the causal transmission intensity in the pollution causal chain feature representation is calculated as follows: W1 = M3 / S2, where M3 is the mutual information value between the parameter and the causal transmission intensity, and S2 is the sum of the mutual information values ​​between all parameters and the causal transmission intensity. The contribution weights of all meteorological and hydrological parameters are summarized to generate an environment-driven factor weight sequence.

[0198] Step S630: Perform semantic embedding processing of driving factors based on the environmental driving factor weight sequence and the environmental analysis report semantic text sequence, embed the meteorological and hydrological driving description text fragments corresponding to the environmental driving factor weight sequence into the trend description paragraph, and generate the environmental driving report sequence.

[0199] From the environmental driving factor weight sequence, select the meteorological and hydrological parameters with the highest weight values ​​as the main environmental driving factors. Convert these main environmental driving factors into meteorological and hydrological driving descriptive text fragments, using the template "Pollution diffusion trends are significantly affected by the main environmental driving factors, where the contribution weight of parameter names is the weight value." Locate the descriptive positions of the trend influencing factors in the trend description paragraph of the environmental analysis report semantic text sequence and embed the meteorological and hydrological driving descriptive text fragments. Generate the environmental driving report sequence.

[0200] Step S640: Perform seasonal change pattern comparison processing on the environmental driven report sequence, call the preset seasonal pollution pattern library, compare the similarity of the indicator time series value sequence with the historical seasonal patterns in the preset seasonal pollution pattern library, generate seasonal pattern comparison text fragments, and attach the seasonal pattern comparison text fragments to the trend description paragraph to generate a seasonal comparison report sequence.

[0201] A pre-defined seasonal pollution pattern library stores typical pollutant concentration variation patterns for each season in the same historical period. These patterns are represented as mean curves plus fluctuation ranges. The time-series numerical sequences of indicators are compared with the historical seasonal patterns corresponding to the current season, and a distance value D1 is calculated using a dynamic time warping algorithm. The similarity score S1 is calculated as: S1 = 1 / (1+D1). If S1 is greater than a preset threshold, a seasonal pattern matching text fragment is generated; if S1 is less than or equal to the preset threshold, a seasonal pattern difference text fragment is generated; if it is lower than the preset similarity threshold, a seasonal pattern comparison text fragment indicating "significant difference from historical seasonal patterns" is generated. The seasonal pattern comparison text fragment is appended to the trend description paragraph of the environmentally driven report sequence. A seasonal comparison report sequence is then generated.

[0202] Step S650: Perform emission inventory association processing on the seasonal comparison report sequence, call the preset pollution emission inventory database, verify the sequence to match the corresponding emission inventory entries according to the pollution source tracing path, generate emission inventory association text fragments, and embed the emission inventory association text fragments into the cause analysis paragraph to generate emission inventory report sequence.

[0203] The pre-defined pollution emission inventory database stores annual pollution emission declaration data for each polluting entity within the industrial park, including information such as the entity name, types of pollutants emitted, and emission amounts. The names of pollution source nodes verified in the pollution source tracing path verification sequence are matched with the names of polluting entities in the pre-defined pollution emission inventory database. Emission inventory entries for the matching entities are extracted and converted into emission inventory-related text fragments, with the conversion template being "Based on pollution emission inventory data, the pollution source name, pollutant name, emission amount, and emission amount value for this year." These emission inventory-related text fragments are embedded into the causal analysis section of the seasonal comparison report sequence. An emission inventory report sequence is then generated.

[0204] Step S660: Perform health risk warning processing on the emission inventory report sequence, call the preset health risk assessment knowledge base, generate health risk warnings based on the toxicity parameter information in the set of pollutant indicator attribute relationships of the level label sequence, generate health risk warning text fragments, insert the health risk warning text fragments into the conclusion summary paragraph, and generate a health warning report sequence.

[0205] A pre-defined health risk assessment knowledge base stores health toxicity parameters for various pollutant indicators, including carcinogenicity slope factors and reference doses. Pollutant indicators with higher pollution levels are extracted from the grade label sequence, and their health toxicity parameters are extracted from the pollutant indicator attribute relationship edge set within the pre-defined knowledge graph structure. Health risk warnings are assessed using the risk quotient method. First, based on the exposure scenario (e.g., drinking water) and exposure parameters (e.g., daily water intake, body weight), the exposure concentration (unit: mg / L) is converted into a daily intake per kilogram of body weight (unit: mg / (kg·day)). Then, the risk quotient is calculated as this intake divided by the reference dose (unit: mg / (kg·day)). The exposure concentration is taken as the median of the concentration range corresponding to the grade label. Based on the risk quotient range, different levels of health risk warning text fragments are generated. These health risk warning text fragments are inserted into the conclusion summary paragraph of the emission inventory report sequence. A health warning report sequence is then generated.

[0206] Step S670: Perform ecological impact description processing on the health alert report sequence, extract the ecological receptor node set from the diffusion pattern subgraph network of the preset knowledge graph structure, associate the ecological receptor node set with the level label sequence to generate ecological impact description text fragments, and attach the ecological impact description text fragments to the conclusion summary paragraph to generate the ecological impact report sequence.

[0207] An ecological receptor node set is extracted from the semantic subgraph network of pollution diffusion patterns within a predefined knowledge graph structure. This set includes ecological receptor nodes such as fish, algae, and benthic organisms. The pollutant tolerance concentration threshold for each ecological receptor in the set is compared with the concentration values ​​of each pollutant indicator in the graded label sequence. If the concentration value exceeds the tolerance concentration threshold, a potential ecological impact is identified for that ecological receptor. The names of ecological receptors with potential ecological impacts and their corresponding pollutant indicator names are converted into ecological impact description text fragments, with the conversion template being "The concentration of pollutant indicator name in the monitoring area may have an adverse impact on the ecological receptor name." These ecological impact description text fragments are appended to the conclusion summary paragraph of the health advisory report sequence. An ecological impact report sequence is then generated.

[0208] Step S680: Perform public communication adaptation processing on the ecological impact report sequence, call the preset public communication semantic template library, replace the professional terms in the causal analysis paragraph, trend description paragraph and conclusion summary paragraph with semantic expressions that the public can understand, and generate a public communication semantic report sequence.

[0209] A pre-defined public communication semantic template library stores the mapping relationship between environmental monitoring technical terms and publicly understandable semantic expressions. For example, "chemical oxygen demand" maps to "aquatic organic pollution level index," and "ammonia nitrogen" maps to "nitrogenous nutrients." The text of each paragraph in the ecological impact report sequence is scanned for terms, and the matched technical terms are replaced with their corresponding publicly understandable semantic expressions. Simultaneously, complex causal reasoning is simplified into simple causal statements. This generates a public communication semantic report sequence.

[0210] Step S690: Perform multi-report format conversion processing on the public communication semantic report sequence, and perform format mapping on the public communication semantic report sequence according to the preset report format template to generate a multi-format report sequence that can be adapted to different report recipients.

[0211] The preset report format template library contains various report format templates, such as formal document templates for government management departments, popular science briefing templates for the general public, and professional report templates for technical personnel. Select the corresponding preset report format template based on the target report recipient type. Fill in the content of each paragraph in the public communication semantic report sequence according to the template's chapter structure, adjusting format attributes such as font, paragraph spacing, and heading levels. Generate a multi-format report sequence adaptable to different report recipients.

[0212] Step S6100: Determine the multi-format report sequence as the final output multi-scenario adapted environmental monitoring report semantic sequence, which includes environmental driving description, seasonal comparison description, emission inventory association description, health risk warning description, ecological impact description, and public communication semantic expression.

[0213] The multi-format report sequence generated in step S690 is used as the final output semantic sequence of the multi-scenario adapted environmental monitoring report. This semantic sequence of the multi-scenario adapted environmental monitoring report further enhances the semantic adaptation capabilities for meteorological and hydrological driven analysis, seasonal comparison, emission inventory verification, health and ecological risk alerts, and public-facing functions, building upon the original comprehensive report.

[0214] In one exemplary embodiment, a semantic generation system for environment analysis reports based on knowledge graphs and deep learning is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, it can be comprised of a processor, memory, input / output interfaces, communication interfaces, a display unit, and input devices. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input devices are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a semantic generation method for environmental analysis reports based on knowledge graphs and deep learning. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the knowledge graph and deep learning-based environmental analysis report semantic generation system, or an external keyboard, touchpad, or mouse, etc.

[0215] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A semantic generation method for environmental analysis reports based on knowledge graphs and deep learning, characterized in that, The method includes: The system obtains a set of original environmental monitoring record texts to be processed and a preset knowledge graph structure for the environmental monitoring field. The preset knowledge graph structure includes a set of entity nodes of monitoring points, a set of relationship edges of pollutant index attributes, and a semantic subgraph network of pollution diffusion patterns. Based on the set of entity nodes of the monitoring points, the set of original environmental monitoring records to be processed is subjected to entity anchor point location processing to generate a distribution sequence of entity anchor points of the monitoring points. Based on the semantic subgraph network of the pollution diffusion pattern, the set of original environmental monitoring records to be processed is subjected to pollution diffusion semantic pattern matching processing to generate a pollution diffusion semantic pattern matching response sequence. The monitoring point entity anchor point location distribution sequence and the pollution diffusion semantic pattern matching response sequence are input into a preset temporal causal reasoning network for cross-temporal causal reasoning processing to generate a pollution causal chain feature representation. Based on the pollution causal chain feature representation and the original monitoring data sentence units in the original environmental monitoring record text set to be processed, sentence-level semantic fusion processing is performed to generate a pollution causal enhanced semantic representation. Based on the enhanced semantic representation of pollution causality and the set of pollutant indicator attribute relationship edges in the preset knowledge graph structure, semantic generation path planning is performed to generate a semantic text sequence of environmental analysis report. The semantic text sequence of environmental analysis report includes a pollution cause analysis paragraph, a pollution diffusion trend description paragraph, and a monitoring conclusion summary paragraph.

2. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 1, characterized in that, The step of performing monitoring point entity anchor point localization processing on the original environmental monitoring record text set to be processed based on the monitoring point entity node set to generate a monitoring point entity anchor point localization distribution sequence, and performing pollution diffusion semantic pattern matching processing on the original environmental monitoring record text set to be processed based on the pollution diffusion pattern semantic subgraph network to generate a pollution diffusion semantic pattern matching response sequence, includes: Extract the text character stream sequence from the original environmental monitoring record text set to be processed, and perform a character-level bit-by-bit comparison between the text character stream sequence and the monitoring point entity node name in the monitoring point entity node set to obtain the character-level bit-by-bit comparison matching position sequence. Based on the character-level bit-by-bit comparison and matching position sequence, the monitoring point entity boundary anchoring processing is performed on the original record text set of the environment monitoring to be processed to determine the start character position and end character position of the monitoring point entity name in the text, and a monitoring point entity anchor start and end position pair sequence is generated. Map the name of each monitoring point entity in the sequence of the start and end positions of the monitoring point entity anchor points to the unique identifier of the monitoring point entity node in the set of monitoring point entity nodes, and generate a monitoring point entity node identifier mapping sequence. The mapping sequence of entity node identifiers of the monitoring points is sorted according to the order in which the entities appear in the text, and the text span distance between adjacent entities is recorded to generate a monitoring point entity anchor point positioning distribution sequence. The monitoring point entity anchor point positioning distribution sequence includes a unique identifier sequence of entity nodes of the monitoring points and a text span distance sequence between adjacent entities. Extract the pollution source node set, pollution propagation path edge set, and pollution receptor node set from the pollution diffusion pattern semantic subgraph network, convert the pollution source node set into a pollution source semantic template set, the pollution propagation path edge set into a pollution propagation path semantic template set, and the pollution receptor node set into a pollution receptor semantic template set. Using the pollution source semantic template set, pollution propagation path semantic template set, and pollution receptor semantic template set, a sliding window semantic template matching process is performed on the original environmental monitoring record text set to be processed, and the matching hit position and matching template type identifier of each semantic template in the text are recorded. Based on the matching hit location and the matching hit template type identifier, a set of pollution diffusion semantic matching path instances is constructed, and the set of pollution diffusion semantic matching path instances is subjected to path instance confidence scoring and filtering processing, and a pollution diffusion semantic pattern matching response sequence is generated by serialization.

3. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 2, characterized in that, The step involves constructing a set of pollution diffusion semantic matching path instances based on the matching hit location and the matching hit template type identifier, performing path instance confidence scoring and filtering on the pollution diffusion semantic matching path instance set, and serializing and generating a pollution diffusion semantic pattern matching response sequence, including: Based on the matching hit position and the matching hit template type identifier, the matching hit sequences that satisfy the condition of going from the set of pollution source nodes through the set of pollution propagation path edges to the set of pollution receptor nodes are combined into pollution diffusion semantic matching path instances, thus obtaining a set of pollution diffusion semantic matching path instances. For each pollution diffusion semantic matching path instance in the pollution diffusion semantic matching path instance set, a path instance confidence score is processed. The consistency degree of the matching hit template type identifier sequence with the direction of the edge in the pollution diffusion pattern semantic subgraph network and the continuity degree of the text span of the path instance are calculated to obtain a path instance confidence score set. Based on the path instance confidence score set, the pollution diffusion semantic matching path instance set is subjected to path instance filtering processing, and pollution diffusion semantic matching path instances with path instance confidence scores higher than the preset path instance confidence filtering threshold are retained to obtain the pollution diffusion semantic matching path instance filtered set. After filtering the pollution diffusion semantic matching path instances, the pollution diffusion semantic matching path instances in the set are serialized into a pollution diffusion semantic pattern matching response sequence. The pollution diffusion semantic pattern matching response sequence includes a path instance identifier sequence, a monitoring point entity node sequence contained in the path instance, and a semantic template type sequence corresponding to the path instance.

4. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 1, characterized in that, The step of inputting the location distribution sequence of the monitoring point entity anchor points and the response sequence of the pollution diffusion semantic pattern into a preset temporal causal reasoning network for cross-temporal causal reasoning processing to generate a pollution causal chain feature representation includes: The sequence of unique identifiers of entity nodes of monitoring points in the distribution sequence of entity anchor points of monitoring points is input into the entity embedding layer of the temporal causal inference network for entity embedding encoding processing. Each unique identifier of entity node of monitoring point is mapped to an entity embedding vector to generate a sequence of entity embedding vectors of monitoring points. The path instance identifier sequence and semantic template type sequence in the pollution diffusion semantic pattern matching response sequence are input into the path embedding layer of the temporal causal inference network for path embedding encoding processing. The path instance identifier is mapped to the path instance embedding vector, and the semantic template type is mapped to the semantic template embedding vector, generating the path instance embedding vector sequence and the semantic template embedding vector sequence. The timestamp information of each text statement in the monitoring point entity embedding vector sequence and the original monitoring record text set of the environment to be processed is fused with time encoding. Time position encoding is added to the monitoring point entity embedding vector according to the order of the timestamps to generate a monitoring point entity embedding vector sequence with enhanced time position encoding. The time-location encoded enhanced monitoring point entity embedding vector sequence, path instance embedding vector sequence, and semantic template embedding vector sequence are input into the causal graph attention layer of the temporal causal reasoning network for causal graph attention aggregation processing. The pollution source entity embedding vector is used as the query vector, the pollution propagation path embedding vector is used as the key vector, and the pollution receptor entity embedding vector is used as the value vector to calculate the attention weight distribution. Based on the attention weight distribution, the causal intensity of the diffusion of pollutants from the pollution source entity to the pollution receiver entity via the pollution propagation path is weighted and aggregated to generate the pollution source entity aggregated feature vector, the pollution propagation path aggregated feature vector, and the pollution receiver entity aggregated feature vector. The aggregated feature vectors of pollution source entities, pollution propagation paths, and pollution receptor entities are input into the temporal causal convolution module of the temporal causal inference network, and causal convolution operations are performed along the time dimension to generate temporal causal convolution feature vectors of pollution source entities, pollution propagation paths, and pollution receptors. Cross-entity gated fusion processing is performed on the pollution source temporal causal convolutional feature vector, pollution propagation path temporal causal convolutional feature vector and pollution receptor temporal causal convolutional feature vector. The gating mechanism is used to perform element-wise fusion weight allocation on the three types of temporal causal convolutional feature vectors to generate cross-entity gated fusion feature vector. The cross-entity gated fusion feature vector is input into the causal chain decoding layer of the temporal causal inference network to perform causal chain sequence decoding and temporal alignment and feature connection processing with the original monitoring data statement units, thereby generating a pollution causal chain feature representation.

5. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 4, characterized in that, The process of inputting the cross-entity gated fusion feature vector into the causal chain decoding layer of the temporal causal inference network, performing causal chain sequence decoding and temporal alignment and feature concatenation with the original monitoring data statement units, and generating a pollution causal chain feature representation includes: The cross-entity gated fusion feature vector is input into the causal chain decoding layer of the temporal causal inference network for causal chain sequence decoding processing, generating the causal transmission intensity sequence and causal transmission direction sequence between the pollution source and the pollution receptor; The timestamp information of the original monitoring data statement units in the original environmental monitoring original record text set to be processed is time-aligned with the causal transmission intensity sequence and causal transmission direction sequence to generate local feature representations of pollution causal chains corresponding to each original monitoring data statement unit; The local feature representations of the pollution causal chain corresponding to each original monitoring data statement unit are processed by feature connection according to the order of the statements to generate a pollution causal chain feature representation. The pollution causal chain feature representation includes the sequence of local feature representations of the pollution causal chain corresponding to each statement and the sequence of pollution causal transmission features between statements.

6. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 1, characterized in that, The step of performing sentence-level semantic fusion processing based on the pollution causal chain feature representation and the original monitoring data sentence units in the original environmental monitoring record text set to be processed, to generate a pollution causal enhanced semantic representation, includes: The original monitoring data sentence units in the original environmental monitoring record text set to be processed are processed by sentence word segmentation, and each original monitoring data sentence unit is divided into an original monitoring data word sequence. The original monitoring data word sequence is input into the word embedding layer of a pre-trained environmental monitoring domain language representation model, and each original monitoring data word is mapped to a word semantic embedding vector to generate an original monitoring data word semantic embedding vector sequence. The original monitoring data word semantic embedding vector sequence is subjected to sentence-level pooling processing. The mean pooling operation is performed on the word semantic embedding vector within each original monitoring data sentence unit to generate the original monitoring data sentence-level semantic vector sequence. Extract the pollution causal chain local feature representation sequence corresponding to each original monitoring data statement unit from the pollution causal chain feature representation, and map each pollution causal chain local feature representation in the pollution causal chain local feature representation sequence to the same vector dimension space as the original monitoring data statement-level semantic vector sequence to generate a pollution causal chain local feature mapping vector sequence. The original monitoring data statement-level semantic vector sequence and the pollution causal chain local feature mapping vector sequence are processed by concatenating statement by statement. The statement-level semantic vector of each original monitoring data statement unit is concatenated with the corresponding pollution causal chain local feature mapping vector to generate a concatenated statement semantic vector sequence. The concatenated statement semantic vector sequence is input into a preset inter-statement bidirectional attention interaction module to perform context-weighted aggregation and residual normalization processing to generate a pollution causal enhanced semantic representation.

7. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 6, characterized in that, The step of inputting the concatenated semantic vector sequence into a preset inter-statement bidirectional attention interaction module, performing context-weighted aggregation and residual normalization processing, and generating a pollution causal enhanced semantic representation includes: The concatenated sentence semantic vector sequence is input into a preset inter-sentence bidirectional attention interaction module. The bidirectional attention weight between each concatenated sentence semantic vector and the other concatenated sentence semantic vectors is calculated to generate an inter-sentence bidirectional attention weight matrix. The concatenated sentence semantic vector sequence is then subjected to context-weighted aggregation processing using the inter-sentence bidirectional attention weight matrix. Finally, each concatenated sentence semantic vector is weighted and summed according to the inter-sentence bidirectional attention weight matrix to generate a context-enhanced concatenated sentence semantic vector sequence. The context-enhanced concatenated sentence semantic vector sequence is subjected to residual join processing. The original monitoring data sentence-level semantic vector sequence is added element-wise to the context-enhanced concatenated sentence semantic vector sequence to generate a residual-enhanced concatenated sentence semantic vector sequence. The residual-enhanced concatenated sentence semantic vector sequence is subjected to layer normalization processing. The mean and variance normalization operations are performed on each dimension of each residual-enhanced concatenated sentence semantic vector to generate a normalized concatenated sentence semantic vector sequence. The normalized concatenated sentence semantic vector sequence is determined as the pollution causal enhanced semantic representation, which includes the normalized concatenated sentence semantic vector sequence corresponding to each original monitoring data sentence unit and the context-enhanced semantic association matrix between sentences.

8. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 1, characterized in that, The step of generating a semantic text sequence for an environmental analysis report by performing semantic generation path planning based on the enhanced semantic representation of pollution causality and the set of pollutant indicator attribute relationships in the preset knowledge graph structure includes: Extract the attribute relationship type sequence and attribute relationship direction sequence between pollutant indicators from the set of attribute relationship edges of the pollutant indicator in the preset knowledge graph structure. The attribute relationship type sequence includes positive correlation relationship edges and negative correlation relationship edges of pollutant indicators. The attribute relationship type sequence and attribute relationship direction sequence are converted into a set of pollutant indicator relationship semantic path templates. Each pollutant indicator relationship semantic path template contains a starting pollutant indicator node, an ending pollutant indicator node, and a relationship semantic label. The normalized concatenated statement semantic vector sequence in the pollution causal enhanced semantic representation is subjected to pollutant index mention detection processing to identify the pollutant index mention identifier corresponding to each normalized concatenated statement semantic vector and generate a pollutant index mention identifier sequence. The pollutant indicator mention identifier sequence is matched with the starting pollutant indicator node and the ending pollutant indicator node in the pollutant indicator relation semantic path template set to construct a statement relation semantic path graph network with the original monitoring data statement unit as the node and the pollutant indicator relation semantic path template as the edge. The graph traversal path search process is performed on the statement relation semantic path graph network. Starting from the starting statement node containing the pollution source entity mention identifier, a depth-first path traversal is performed along the edge direction of the pollutant index relation semantic path template to generate a set of statement semantic path candidate sequences. For each statement semantic path candidate sequence in the statement semantic path candidate sequence set, perform path semantic coherence scoring processing, calculate the semantic similarity of the normalized concatenated statement semantic vectors between adjacent statement nodes in the path and the overall pollution causal logic consistency score of the path, and obtain the path semantic coherence score set. Based on the path semantic coherence score set, the statement semantic path candidate sequence set is subjected to path optimization processing, and the statement semantic path candidate sequence with the highest path semantic coherence score is selected as the semantic path sequence for pollution cause analysis. The normalized concatenated semantic vectors corresponding to each statement node in the pollution cause analysis semantic path sequence are input into a preset paragraph generation recurrent neural network decoder. The process of generating pollution cause analysis paragraphs and generating pollution diffusion trend and monitoring conclusion paragraphs based on path extension is performed to generate a semantic text sequence of environmental analysis report.

9. The semantic generation method for environment analysis reports based on knowledge graphs and deep learning according to claim 8, characterized in that, The normalized concatenated semantic vectors corresponding to each statement node in the pollution cause analysis semantic path sequence are input into a preset paragraph generation recurrent neural network decoder. This decodes and generates pollution cause analysis paragraphs and generates paragraphs based on path extension and expansion of pollution diffusion trends and monitoring conclusions, generating a semantic text sequence for an environmental analysis report. This includes: The normalized concatenated semantic vectors corresponding to each statement node in the pollution cause analysis semantic path sequence are input into a preset paragraph generation recurrent neural network decoder to perform pollution cause analysis paragraph decoding and generation processing, thereby generating a pollution cause analysis paragraph text sequence. The semantic path sequence for pollution cause analysis is extended and expanded by path extension processing. Based on the set of pollution propagation path edges in the semantic subgraph network of pollution diffusion mode, the semantic path of the statement is extended and expanded outward to generate a semantic path sequence for pollution diffusion trend and a semantic path sequence for monitoring conclusion summary. The semantic path sequence of pollution diffusion trend and the semantic path sequence of monitoring conclusion summary are respectively input into the paragraph generation recurrent neural network decoder for paragraph decoding and generation processing, generating a paragraph text sequence describing pollution diffusion trend and a paragraph text sequence summarizing monitoring conclusion. The paragraph text sequence of pollution cause analysis, the paragraph text sequence describing pollution diffusion trend and the paragraph text sequence summarizing monitoring conclusion are then combined into a semantic text sequence of environmental analysis report.

10. A semantic generation system for environmental analysis reports based on knowledge graphs and deep learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the semantic generation method for environmental analysis reports based on knowledge graphs and deep learning according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images