Chemical accident risk analysis and causal network construction method based on text mining and GNN-Apriori association rule

Abstract

The invention belongs to the field of industrial safety and artificial intelligence, and discloses a chemical accident risk analysis method and system based on text mining and GNN-Apriori association rules. The method comprises the following steps: cleaning and structuring a chemical accident text, fusing TextRank, BM25 and BERT algorithms, and extracting a high-quality keyword set through a weighted model; constructing a graph structure based on a keyword co-occurrence relationship, and learning node semantic embedding by using GNN; inputting the embedded vector into an improved Apriori algorithm, and mining a semantic enhanced association rule in combination with semantic similarity, support degree, confidence and lifting degree; and finally, constructing a chemical accident causal complex network, and evaluating key nodes by using node centrality by taking the improvement degree as an edge weight to realize risk propagation path analysis and visual display. According to the method, chemical accident causes can be intelligently identified, semantic association is mined, a causal network is modeled, and interpretable decision support is provided for accident prevention and the like.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of chemical safety and artificial intelligence information processing, specifically involving a method for chemical accident risk analysis and causal network construction based on text mining and GNN-Apriori association rules. This method comprehensively applies Natural Language Processing (NLP), Graph Neural Networks (GNN), multi-strategy keyword extraction (TextRank, BM25, BERT), and an improved Apriori association rule algorithm to achieve intelligent information extraction, deep semantic learning, and complex causal relationship identification from chemical accident texts.

[0002] This invention is primarily applied to scenarios such as automated risk identification, causal chain reasoning, accident mode discovery, and safety decision support for chemical safety accidents. By introducing a GNN-Apriori algorithm model that combines semantic embedding and structural learning, it can perform in-depth analysis of large-scale chemical accident report texts, uncovering potential high-risk factors, implicit multi-level causal chains, and complex coupling relationships. Simultaneously, this invention also incorporates complex network theory, utilizing the mined association rules to construct a directed weighted network for chemical accident causation, enabling key node identification, risk propagation path analysis, and quantitative evaluation of accident prevention and control strategies.

[0003] Therefore, this invention is widely applicable to safety management departments of chemical enterprises, emergency management agencies, accident investigation agencies, hazardous chemical supervision systems, and scientific research units, and can provide intelligent technical support for early warning decision-making, risk assessment, and safety improvement of chemical accidents. Background Technology

[0004] With the rapid development of the chemical industry and the increasing complexity of its processes, chemical accidents are occurring frequently and in various forms, including leaks, explosions, poisoning, and fires. These accidents often exhibit characteristics of multiple intertwined causes, chain-like evolution, and high uncertainty, posing serious threats to human safety, the ecological environment, and the sustainable operation of enterprises. Research into the causes of these accidents and risk identification are crucial topics in the field of chemical safety management.

[0005] Existing accident causation analysis methods mainly include Fault Tree Analysis (FTA), Failure Mode and Effects Analysis (FMEA), and Hazard and Operability Study (HAZOP). These traditional methods are widely used in engineering safety analysis, but they generally suffer from the following problems:

[0006] (1) Reliance on expert experience: Most methods require experts to determine risk factors and causal relationships through experience or manual reasoning, which is highly subjective and makes it difficult to guarantee consistency and comprehensiveness;

[0007] (2) Limited processing capacity: It is difficult to efficiently process massive amounts of unstructured accident text data and cannot extract general patterns from a large number of historical accident reports;

[0008] (3) Lack of semantic understanding: Traditional statistical methods are unable to capture the deep semantic relationships and potential multi-level causal chains in accident texts, resulting in incomplete risk identification;

[0009] (4) Shallow rule mining: The classic Apriori algorithm is based on the frequency mining of itemsets to mine association rules. It only considers co-occurrence relationships and ignores the semantic relevance and structural features between words, which easily leads to associations that are "high frequency but have no causal significance".

[0010] In recent years, with the development of Natural Language Processing (NLP) and machine learning technologies, researchers have begun to explore the application of text mining techniques in accident analysis. For example, they use methods such as LDA topic modeling, TFIDF, and TextRank to extract accident keywords and reveal risk factors through clustering or co-occurrence analysis. However, these methods still primarily focus on word frequency statistics and shallow structural analysis, lacking deep semantic representation and contextual understanding capabilities.

[0011] In association rule mining, the improved Apriori algorithm can discover multi-level and multi-dimensional potential relationships, but its calculation is still based on frequency and cannot distinguish potential key factors with high semantic similarity but low frequency of occurrence. In addition, the existing Apriori model lacks the use of network structure information and cannot reflect the complex topological dependencies and semantic coupling relationships between accident factors.

[0012] To address the aforementioned issues, there is an urgent need for a novel analytical framework that combines semantic understanding, structural modeling, and rule mining. This framework should be able to fully explore the semantic features of chemical accident texts, automatically learn the deep relationships between accident causes, and achieve intelligent identification and visual representation of high-risk chains.

[0013] Therefore, it is necessary to propose an association rule mining method based on multi-strategy keyword extraction and graph neural network enhancement. By integrating TextRank, BM25, BERT, GNN and Apriori algorithms, a unified semantic structure model of chemical accident causes is constructed. Combined with complex network theory, the intelligent construction and quantitative analysis of accident causal networks are realized to make up for the shortcomings of existing methods in big data semantic mining and complex causal reasoning. Summary of the Invention

[0014] This invention addresses the technical problems of existing chemical accident risk analysis methods when processing large-scale unstructured text, such as difficulties in information extraction, unclear semantic relationships, shallow rule mining, and insufficient visualization of causal relationships. It proposes a method for chemical accident risk analysis and causal network construction based on text mining and GNN-Apriori association rules.

[0015] Specifically, the present invention aims to solve the following key problems:

[0016] 1. Chemical accident text information is unstructured and semantically ambiguous: Traditional methods rely on manual reading or shallow statistical models, which makes it difficult to accurately extract technical terms and potential risk factors from accident texts and lacks a deep understanding of semantic context.

[0017] 2. Insufficient singularity and precision in keyword extraction: A single model (such as TextRank or TFIDF) is difficult to take into account word frequency importance, contextual semantics and structural relationships, resulting in low keyword quality and affecting the accuracy of subsequent causal analysis.

[0018] 3. The Apriori algorithm cannot reflect semantic associations and structural features: The traditional Apriori algorithm only relies on the frequency of itemset occurrences to mine rules, ignoring the semantic embedding information between accident factors and their position in the co-occurrence network, making it difficult to identify high-value low-frequency factors.

[0019] 4. The complexity of accident causal relationships and the lack of network representation: Existing research usually stays at the level of linear causal reasoning, and cannot characterize the multi-path propagation and multi-level interaction between accident causes, and lacks a quantifiable risk chain network model.

[0020] 5. Lack of a visualized intelligent decision support system: Traditional risk analysis results are presented in text or table form, which is not conducive to intuitively revealing key nodes, paths and risk propagation mechanisms, thus limiting the auxiliary decision-making capabilities of safety management departments.

[0021] Therefore, the main technical objective of this invention is to construct an intelligent system that integrates multi-strategy text feature extraction, graph neural network semantic embedding, improved Apriori association mining, and complex network visualization analysis, so as to realize the automatic identification, rule mining, and network presentation of the causes of chemical accidents, thereby providing chemical enterprises and regulatory authorities with an efficient, scientific, and interpretable risk analysis tool.

[0022] To address the aforementioned technical problems, this invention proposes a method for chemical accident risk analysis and causal network construction based on text mining and GNN-Apriori association rules. This method achieves systematic identification and quantitative modeling of the causes of chemical accidents through multi-strategy keyword extraction, semantic embedding learning, structure-enhanced association mining, and network-based visualization analysis. Its core steps and system module structure are as follows:

[0023] 1. Text data preprocessing module:

[0024] Used for data cleaning, word segmentation, and format standardization of unstructured texts such as chemical accident reports and investigation conclusions.

[0025] This module includes:

[0026] Chinese word segmentation submodule: Based on the Python-based Jieba word segmentation tool, it performs accurate word segmentation on accident texts;

[0027] Domain dictionary construction submodule: Introduces chemical engineering terminology, chemical substance catalog, safety terminology, and operational scenario vocabulary to improve word segmentation accuracy;

[0028] Stop word filtering submodule: Utilizes the stop word list from Harbin Institute of Technology and industry-specific word lists to remove invalid words (such as "company" and "accident").

[0029] The text structuring transformation submodule converts the word segmentation results into vectorized form, providing input for subsequent algorithm analysis.

[0030] 2. Multi-strategy fusion keyword extraction module:

[0031] This module combines the advantages of TextRank, BM25, and BERT algorithms to build a multi-dimensional keyword fusion model:

[0032] TextRank structural feature extraction: Construct a word co-occurrence graph and evaluate the importance of words by iteratively calculating node weights;

[0033] BM25 Statistical Weighting: Introduces word frequency, document length, and inverse document frequency correction to achieve keyword importance measurement;

[0034] BERT semantic embedding: It uses a bidirectional Transformer model to generate contextual semantic vectors, capturing synonyms and deep semantics;

[0035] Fusion weighting mechanism: By normalizing and weighted summing, the weights of the three factors are combined to obtain a set of keywords with high semantic quality, forming a feature vector representation of chemical accidents.

[0036] 3. GNN Semantic Structure Embedding Module:

[0037] This module uses Graph Convolutional Network (GCN) to perform graph structure modeling and feature learning for keyword co-occurrence relationships.

[0038] The extracted keywords are used as nodes, and word co-occurrence relationships are used as edges to construct a weighted undirected graph;

[0039] Each node's feature input graph is convolutional layer, which aggregates neighborhood information to generate a high-dimensional embedding vector;

[0040] Joint encoding of semantic and structural information is achieved through multi-layer convolution and non-linear activation;

[0041] The output node embedding matrix serves as the input feature space for the subsequent Apriori algorithm.

[0042] 4. Improved GNN-Apriori Association Rule Mining Module:

[0043] Based on the traditional Apriori algorithm, a semantic similarity metric for GNN embedding vectors is introduced to achieve semantically enhanced association rule mining:

[0044] By calculating the cosine similarity of node embeddings and the support threshold, potential highly semantically relevant itemsets are identified.

[0045] Use the support, confidence, and lift metrics to filter strongly correlated rules;

[0046] Itemsets with low frequency but high semantic similarity are retained to discover implicit high-risk relationships;

[0047] The output is a set of multi-level causal rules in the form of "factor A → factor B".

[0048] 5. Causal Network Construction and Visualization Module:

[0049] This module constructs a complex network of chemical accident causations based on the rule results output by GNN-Apriori:

[0050] Nodes represent accident risk factors, and the weight of edges is described by lift.

[0051] Forming a directed weighted graph structure;

[0052] The size of a node is determined by its centrality (such as degree centrality or eigenvector centrality);

[0053] Analyze system coupling using complex network metrics (clustering coefficient, path length, connectivity);

[0054] Visualize causal networks using Gephi or Python NetworkX to show key causal chains and high-risk paths.

[0055] Method and steps:

[0056] S1: Obtain the text of the chemical accident report, and perform word segmentation, stop word removal, and formatting to form a structured corpus;

[0057] S2: Employ TextRank, BM25, and BERT algorithms for multi-strategy fusion keyword extraction to obtain a semantically enhanced keyword set;

[0058] S3: Construct a word graph based on keyword co-occurrence relationships, and use GNN for node embedding and semantic structure representation learning;

[0059] S4: Input the GNN embedding vector into the Apriori algorithm and mine semantically enhanced association rules based on support, confidence and lift thresholds;

[0060] S5: Based on the mining results, construct a complex causal network of chemical accidents, and use the lift degree as the edge weight to complete the visualization of the causal chain and the identification of key nodes.

[0061] System implementation method:

[0062] This invention can be implemented using the Python platform. The core algorithm uses PyTorch Geometric for GNN training, and the Apriori algorithm is implemented based on the mlxtend.frequent_patterns module. It can be combined with Pandas, NetworkX, and Gephi for data processing and visualization. The system can be deployed on a local server or a cloud-based data analysis platform for automated processing and dynamic risk identification of accident data in chemical enterprises.

[0063] The present invention has the following significant beneficial effects:

[0064] 1. Integrating multiple strategies to improve keyword extraction quality:

[0065] This invention innovatively combines TextRank, BM25, and BERT algorithms to comprehensively extract keywords from three levels: statistical features, structural relationships, and semantic information.

[0066] Compared to keyword extraction methods using a single model, this invention significantly improves semantic accuracy and domain adaptability while maintaining the representativeness of word frequency, ensuring that the extracted keywords are both professional and explanatory.

[0067] 2. Introduce graph neural networks to enhance semantic structure representation:

[0068] By constructing a keyword co-occurrence graph and using a GNN model for node embedding learning, joint modeling of semantic and structural information was achieved.

[0069] This technology can automatically learn the complex topological relationships between different accident factors, discover potential semantic couplings and implicit causal chains, and significantly improve the depth of accident factor identification and relationship modeling.

[0070] 3. Improve the Apriori algorithm to achieve semantically enhanced rule mining:

[0071] This invention incorporates semantic similarity and frequency information into the association analysis index system by introducing node embedding vectors generated by GNN.

[0072] This method can identify low-frequency but high-value potential causal relationships, overcome the dependence of the traditional Apriori algorithm on high-frequency itemsets, and improve the semantic rationality and application guidance value of the rules.

[0073] 4. Construct quantifiable causal complex network models:

[0074] This invention utilizes the results of GNN-Apriori rules to construct a causal network for chemical accidents, using lift as the edge weight and node centrality as the risk indicator to form a directed weighted graph.

[0075] This model can reveal the multi-path, multi-level accident evolution mechanism, realize the identification of high-risk nodes and the tracking of critical paths, and provide quantitative basis for accident prevention and risk management.

[0076] 5. Achieve intelligent and visualized risk decision support:

[0077] By combining complex network analysis and visualization tools (such as Gephi), the network structure of accident causal relationships and risk propagation patterns can be intuitively displayed.

[0078] Managers can quickly identify key causal nodes and high-risk chains, thereby formulating targeted prevention and control measures and significantly improving the scientific nature and timeliness of safety management in chemical enterprises.

[0079] 6. Strong technical versatility and scalability:

[0080] The text mining and association analysis framework proposed in this invention is not only applicable to chemical accidents, but can also be extended to accident analysis and safety assessment in high-risk industries such as mining, construction, transportation, and energy, and has broad industrial application prospects.

[0081] In summary, this invention, through algorithm fusion and structural innovation, breaks through the bottlenecks of semantic understanding and rule mining in traditional accident analysis, enabling intelligent conversion from text to knowledge and providing an efficient and reliable decision support system for chemical safety management.

[0082] This invention has the following innovative features:

[0083] Algorithm integration and innovation: For the first time, the weights of three algorithms, TextRank, BM25 and BERT, are integrated to achieve multi-dimensional and multi-semantic level keyword extraction, effectively improving the semantic recognition quality of accident texts.

[0084] Innovative Semantic Enhancement Association Mining: By introducing graph neural network embedding results into the Apriori algorithm, association rule mining is transformed from frequency-driven to a dual-driven mechanism of "semantics + structure".

[0085] Innovative Causal Network Modeling: A directed weighted graph is constructed using lift as the edge weight, and key causal nodes are identified through network centrality measurement, thereby achieving quantitative modeling of the accident risk chain.

[0086] Intelligent visualization innovation: Combining complex network theory with Gephi visualization tools, it enables an intuitive presentation of the causal relationships of chemical accidents, facilitating safety managers to quickly identify key risk paths.

[0087] Cross-industry applicability innovation: This framework can be extended to accident text analysis in high-risk industries such as mining, construction, transportation, and energy, and has strong versatility and scalability. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the overall process of the chemical accident risk analysis method based on text mining and GNN-Apriori association rules described in this invention.

[0089] Figure 2 Flowchart for preprocessing text data related to chemical accidents;

[0090] Figure 3 A schematic diagram of the multi-strategy fusion keyword extraction model (including the fusion structure of TextRank, BM25 and BERT).

[0091] Figure 4 The diagram shows the embedding structure of a keyword co-occurrence network based on GNN.

[0092] Figure 5 A flowchart illustrating the working principle of the GNN-Apriori fusion algorithm;

[0093] Figure 6 A framework diagram for constructing and visualizing the complex network of causes of chemical accidents;

[0094] Figure 7 This is a schematic diagram of the overall system architecture of the present invention (including a three-layer structure of data layer, algorithm layer, and application layer). Detailed Implementation

[0095] The entire process enables end-to-end analysis from unstructured accident text data to structured knowledge networks. For example... Figure 1 As shown, the embodiments of the present invention mainly include seven stages:

[0096] Text input → Data preprocessing → Multi-strategy keyword extraction → GNN embedding learning → GNN-Apriori rule mining → Causal network construction → Visualization output.

[0097] Step 1: Text data preprocessing as follows Figure 2

[0098] In this embodiment, the input data comes from 566 publicly available chemical accident investigation reports and emergency management announcements. The data preprocessing module performs the following operations in sequence:

[0099] 1. Chinese word segmentation: The Jieba word segmentation library in Python is used, and a custom dictionary is loaded to cover chemical substance names, reaction devices, operation types, safety terms, etc., to improve word segmentation accuracy.

[0100] 2. Stop word removal: The stop word list of Harbin Institute of Technology is used, and a stop word set is constructed by combining industry vocabulary. Words that have no actual analytical significance, such as "company", "occurrence", and "accident", are deleted.

[0101] 3. Text structuring transformation: The word segmentation results of each accident text are stored in list form to form a corpus to be analyzed, providing input for subsequent keyword extraction and network modeling.

[0102] Step 2: Multi-strategy keyword extraction, such as Figure 3

[0103] This embodiment employs a multi-strategy keyword extraction algorithm that integrates TextRank, BM25, and BERT.

[0104] 1. TextRank structural feature calculation:

[0105] Construct a word co-occurrence graph (G = (V,E)), where nodes (V_i) represent words and edge weights (w_{ij}) represent the number of co-occurrences.

[0106] The formula for iteratively updating the weights of each node is:

[0107]

[0108] Where (d = 0.85) is the damping coefficient.

[0109] 2. BM25 statistical weighted calculation:

[0110] Calculate the importance of each keyword:

[0111]

[0112] Where (k_1 = 1.2, b = 0.75) and (f(q,D)) are word frequencies.

[0113] 3. BERT semantic embedding:

[0114] Using a pre-trained Chinese BERT model, the accident text is segmented and encoded to extract the bidirectional semantic vector of each word.

[0115] 4. Calculation of fusion weights:

[0116] Combining the results of the three algorithms, a normalization and weighted fusion strategy is adopted:

[0117]

[0118] Among them (α: β: γ = 0.3 : 0.3 : 0.4), semantic features are dominant.

[0119] Take the first few keywords to form a keyword feature vector matrix. ).

[0120] IV. GNN Semantic Embedding Learning Figure 4

[0121] Based on keyword co-occurrence relationships, an undirected weighted graph (G = (V,E)) is constructed, where the edge weights are determined by the frequency of keyword occurrences in the same text.

[0122] Node embedding is implemented using a Graph Convolutional Neural Network (GCN), and its layer propagation formula is as follows:

[0123]

[0124] in:

[0125] H(l)∈R(N×d) is the node feature matrix of the l-th layer, where N is the number of nodes and d is the feature dimension.

[0126] Â = A + I: Adjacency matrix with added self-loops, where A is the original adjacency matrix and I is the identity matrix;

[0127] D: Degree matrix, ΣjÂij;

[0128] W(l): The learnable weight matrix of the l-th layer

[0129] σ: Nonlinear activation function (such as ReLU)

[0130] The output node embedding matrix is ​​obtained after two convolutional layers. Each keyword obtains a structure-enhanced semantic representation vector.

[0131] V. GNN-Apriori Association Rule Mining Figure 5

[0132] The embedding matrix (Z) is incorporated into the Apriori algorithm, and an itemset similarity (sin(A,B) = cos(ZA,ZB)) is defined, while a semantic similarity threshold (θ) is considered when generating frequent itemsets.

[0133] The specific excavation process is as follows:

[0134] 1. Candidate itemset generation: Semantic similarity filtering is used to retain itemsets with (sin(A,B) > θ).

[0135] 2. Support and confidence level calculation:

[0136]

[0137]

[0138] 3. Lift Calculation:

[0139]

[0140] 4. Rule filtering and sorting: Strongly associated rules are filtered based on a support ≥ 0.01 and a confidence ≥ 0.2 threshold.

[0141] For example, the confidence level of the rule "system defect → violation" is 0.52 and the lift is 5.76, indicating that system management defects significantly increase the probability of violation.

[0142] VI. Construction and Analysis of Causal Networks for Chemical Accidents Figure 6

[0143] Import the mined association rule results into a complex network model:

[0144] Nodes represent the causative factors of an accident;

[0145] Directed edges represent causal relationships;

[0146] Edge weights are assigned using lift values;

[0147] Node size is determined based on eigenvector centrality;

[0148] Network metrics (degree centrality, betweenness centrality, clustering coefficient) are used to identify key causal chains.

[0149] Visual output is generated using Gephi software, forming something like... Figure 6 The causal network model shown has core nodes including "system defects", "violations of regulations", "equipment aging", and "high-temperature mixing", revealing high-risk paths in the accident chain.

[0150] VII. System Implementation and Operation Results Figure 7

[0151] The system is implemented using Python:

[0152] Text preprocessing and keyword extraction: based on the jieba, transformers and rank_bm25 libraries;

[0153] GNN modeling: using PyTorch Geometric;

[0154] Rule mining: using the mlxtend.frequent_patterns.apriori module;

[0155] Network analysis and visualization: using NetworkX and Gephi.

Claims

1. A method for chemical accident risk analysis based on text mining and GNN-Apriori association rules, characterized in that, Includes the following steps: Text data preprocessing: Data cleaning, word segmentation, stop word filtering, and formatting were performed on chemical accident report texts to construct a structured corpus; The word segmentation uses a domain dictionary containing chemical engineering terminology, and after removing invalid words, a corpus matrix is ​​formed. Multi-strategy fusion keyword extraction: Keywords are extracted using the TextRank algorithm, the BM25 weighted algorithm, and the BERT semantic model, and then fused and weighted according to the following formula: Where (α, β, γ = 0.3 : 0.3 : 0.4), to obtain a set of keywords with high semantic quality; GNN semantic embedding learning: A graph structure (G=(V,E)) is constructed based on keyword co-occurrence relationships, and a graph convolutional neural network (GNN) model is used to calculate the semantic embedding of nodes. The semantically enhanced node embedding matrix (Z) is obtained; GNN-Apriori Association Rule Mining: Input the embedding vector (Z) into the Apriori algorithm, define the semantic similarity (sin(A,B) = \cos(Z_A,Z_B)), and generate a semantically enhanced association rule set; Calculate the support, confidence, and lift of the rule: Filter strong association rules based on thresholds; Causal network construction and visualization: A directed weighted graph is constructed using keywords as nodes and lift degree as edge weights. The node size is determined by centrality. The risk chain structure is analyzed using complex network indicators (such as degree centrality, clustering coefficient, and path length), and a causal network graph is output through visualization tools.

2. The method according to claim 1, characterized in that, The text data preprocessing module further includes: a Chinese word segmentation submodule; a domain dictionary construction submodule; and a stop word filtering submodule.

3. The method according to claim 1, characterized in that, The GNN model adopts a two-layer graph convolutional neural network structure with 128 hidden layer nodes, ReLU activation function, Adam optimizer, and learning rate set to 0.

01.

4. The method according to claim 1, characterized in that, The Apriori algorithm has a minimum support threshold of 0.01, a confidence threshold of 0.2, and a semantic similarity threshold (θ > 0.7).

5. The method according to claim 1, characterized in that, The node centrality indices of the causal network include degree centrality, betweenness centrality, and eigenvector centrality, which are used to identify key causal nodes.

6. A system for implementing the method according to any one of claims 1 to 5, characterized in that, include: Data layer: used for text collection and preprocessing of chemical accident data; Algorithm layer: including Multi-strategy keyword extraction module, GNN semantic embedding module, GNN-Apriori association rule mining module; Application layer: includes causal network construction module and visualization analysis module.

Images