Large language model road disease diagnosis method based on three-layer causal reasoning mechanism

Abstract

The application discloses a large language model pavement disease diagnosis method based on a three-layer causal reasoning mechanism, comprising the following steps: establishing a multi-source knowledge acquisition and structured representation path for pavement disease diagnosis; extracting knowledge triples by using a large language model, and completing entity normalization and relationship standardization; constructing a pavement disease domain knowledge graph, and establishing a three-layer causal network of a "reason layer-mechanism layer-disease layer"; performing causal probability propagation and candidate cause sorting, and establishing an explanation path of a target disease; and fusing text retrieval and knowledge graph retrieval to generate a structured diagnosis result. The method takes a large language model as the core, combines knowledge extraction, knowledge graph construction and retrieval enhancement generation technology, and performs structured representation and intelligent reasoning on pavement disease related knowledge, so that the accuracy, explainability and engineering application value of pavement disease diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of road engineering defect detection and intelligent diagnosis technology, and in particular to a method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism. Background Technology

[0002] Over long-term use, road surfaces are susceptible to various defects such as cracks, ruts, potholes, loosening, subsidence, and bleeding due to repeated traffic loads, changes in the natural environment, material performance degradation, and differences in construction quality. These defects not only reduce road comfort and service performance but also shorten pavement lifespan, increase maintenance costs, and in severe cases, even affect driving safety. Therefore, accurate, timely, and reasonable diagnosis of pavement defects is a crucial technical issue in road maintenance and operation management. Current pavement defect diagnosis mainly relies on manual experience, standardized queries, or traditional machine learning methods. Manual diagnosis, typically involving engineering technicians combining on-site inspection results, defect appearance characteristics, and existing engineering experience, offers some flexibility but heavily depends on the professional experience and knowledge of the diagnostic personnel, leading to problems such as knowledge fragmentation, low diagnostic efficiency, inconsistent judgment standards, and poor consistency in diagnostic results. Existing technical solutions employ rule-based, statistical analysis, or traditional machine learning methods to assist in the diagnosis of pavement defects. While these methods can improve identification efficiency to some extent, they typically rely on pre-defined rule systems or limited sample training results. They focus more on the direct mapping between defect manifestations and diagnostic results, making it difficult to effectively address the issues of multi-factor coupling, multi-level transmission, and complex interaction mechanisms in the formation of pavement defects. Furthermore, they lack the ability to explain the mechanisms of defect formation, making it difficult to meet the needs of engineering practice for interpretability and causal analysis capabilities in diagnosis. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a pavement distress diagnosis method based on a three-layer causal reasoning mechanism using a large language model. This method takes a large language model as its core and combines knowledge extraction, knowledge graph construction, and retrieval enhancement generation techniques to perform structured representation and intelligent reasoning on pavement distress-related knowledge, thereby improving the accuracy, interpretability, and engineering application value of pavement distress diagnosis.

[0004] The technical solution adopted in this invention is: a method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism, comprising the following steps:

[0005] P1: Establishing a multi-source knowledge acquisition and structured representation path for pavement distress diagnosis:

[0006] Collect multi-source heterogeneous data related to pavement distress diagnosis, including academic literature, industry standards, engineering cases, inspection records, and maintenance experience. Perform text cleaning, noise reduction, paragraph segmentation, and semantic blockization on the raw data to form a standardized text set that can be used for knowledge extraction. Further construct a knowledge ontology in the field of pavement distress diagnosis, and uniformly define distress entities, causal entities, mechanism entities, and engineering condition entities to provide a standard semantic framework for subsequent knowledge extraction and knowledge graph construction.

[0007] P2: Extract knowledge triples using a large language model, and complete entity normalization and relation standardization:

[0008] The preprocessed text is input into the large language model Qwen2.5:7B to extract semantic relationships between disease types, influencing factors, mechanisms of action, and engineering conditions, forming knowledge triples in the form of Head-Relation-Tail. The confidence of the extracted knowledge triples is evaluated by combining domain rules, consistency checks, and manual verification. Synonymous entities, near-synonyms, and duplicate records from different knowledge sources are normalized to improve the consistency and accuracy of knowledge representation.

[0009] P3: Construct a knowledge graph for the field of road surface defects and establish a three-layer causal network of "cause layer - mechanism layer - defect layer":

[0010] The normalized knowledge triples are imported into the Neo4j graph database to construct a knowledge graph for the pavement disease domain, which includes disease nodes, cause nodes, mechanism nodes, and engineering condition nodes. The weights of the relation edges are set according to the reliability of the knowledge source, the frequency of relation occurrence, and the confidence of the triples. Based on the knowledge graph, a three-layer causal network of "cause layer - mechanism layer - disease layer" is further established to characterize the causal transmission process of external factors evolving into target diseases through intermediate mechanisms.

[0011] P4: Perform causal probability propagation and candidate cause ranking to establish an explanatory path for the target disease:

[0012] For the target disease, the influence of each candidate cause on the target disease is calculated based on the three-layer causal network. Through conditional probability propagation from the cause layer to the mechanism layer and from the mechanism layer to the disease layer, the comprehensive contribution of different potential causes to the target disease is obtained. Combining path weight, relationship strength and evidence support, the candidate causes are ranked to obtain the Top-N cause set of the target disease and its corresponding dominant mechanism path, thereby realizing the interpretable inversion from disease manifestation to potential causes.

[0013] P5: Integrating text retrieval and knowledge graph retrieval to generate structured diagnostic results:

[0014] For the disease description information or query questions input by users, relevant evidence is retrieved from the text knowledge base and knowledge graph, respectively. The retrieval results are then fused, reordered, and causal enhancement is performed. The sorted text evidence, graph relationships, candidate causes, and causal paths are input into the large language model to generate a structured diagnostic report that includes at least the disease identification results, key cause set, diagnostic conclusion, causal explanation chain, and recommended treatment measures. This achieves a complete technical process from disease description input to disease identification, cause analysis, evidence association, and treatment suggestions output.

[0015] Furthermore, step P1 is implemented through the following steps:

[0016] Let the set of original knowledge sources related to pavement distress diagnosis be:

[0017] (1.1)

[0018] in, Represents a set of multi-source knowledge sources. Indicates the first A source of knowledge This indicates the total number of knowledge sources; these knowledge sources include academic literature, industry technical standards, typical engineering cases, testing records, and related online data.

[0019] (1) The original knowledge source set Text cleaning, paragraph segmentation, semantic block division, and manual proofreading are performed to obtain a preprocessed set of text fragments.

[0020] (1.2)

[0021] in, This represents a set of preprocessed text fragments. Indicates the first A text fragment, Indicates the total number of text fragments;

[0022] (2) To evaluate the usability of each text segment, the text segment effectiveness evaluation function is defined as follows:

[0023] (1.3)

[0024] in, Represents a text fragment The effectiveness score; Indicates the cleanliness index of the text; Indicates semantic completeness index; Indicators representing domain relevance; , , This represents the corresponding adjustment coefficient; the formula uses a compression mapping method to limit the score to the range of 0 to 1, which is used to filter high-quality knowledge fragments.

[0025] (3) After completing data preprocessing, a semantic framework for pavement defect diagnosis is constructed and defined as follows:

[0026] (1.4)

[0027] in, Represents a domain semantic framework; Represents a set of entity units; Represents a set of relational constraints; Represents a collection of attribute descriptions; This represents a set of hierarchical mapping rules.

[0028] (4) Assemble the entity unit set Further expressed as:

[0029] (1.5)

[0030] in, Represents a collection of diseased entities. Represents the set of causal entities. Represents a set of mechanism entities. This represents the set of engineering condition entities; this entity partitioning method provides a unified semantic foundation for subsequent knowledge unit extraction and three-layer causal modeling.

[0031] Furthermore, step P2 is implemented through the following steps:

[0032] (1) Using a large language model to perform semantic understanding and structured information extraction on the preprocessed text fragments, the unstructured text is converted into knowledge units; let the first... The knowledge units are:

[0033] (1.6)

[0034] in, Indicates the first Each knowledge unit; Represents the source entity; Semantic links Represents the target entity; This indicates the addition of context markers; unlike the conventional triple structure, this formula, by introducing context markers, allows knowledge representation to simultaneously preserve semantic relationships and contextual information.

[0035] (2) To ensure the reliability of the knowledge extraction results, a credibility function is constructed for each knowledge unit:

[0036] (1.7)

[0037] in, Indicates the first The overall credibility of each knowledge unit; Indicates the credibility of the language model generation; Indicates the credibility of rule matching Indicates the credibility of knowledge consistency; This represents the adjustment factor for each confidence level; the formula uses a complementary probability product structure, which differs from the conventional linear weighted summation form.

[0038] (3) Perform unified mapping on synonymous entities from different sources; assume two candidate entities and The merge strength is:

[0039] (1.8)

[0040] in Indicates candidate entities and Merge strength; Indicates literal similarity; Indicates the semantic proximity of vectors; Indicates the proximity of co-occurrence in context;

[0041] (4) Standardize the semantic links in the knowledge unit and map similar relationships in different texts into a unified semantic link category to improve the consistency and stability of the subsequent knowledge association network construction.

[0042] Furthermore, step P3 is implemented through the following steps:

[0043] (1) Import the standardized knowledge units into the graph database to construct a knowledge association network for pavement defects; let the knowledge association network be:

[0044] (1.9)

[0045] in, Represents a knowledge association network; Represents a set of node clusters; Represents a collection of links; Represents the set of link strengths;

[0046] (2) To enhance the reasoning ability of the network, for any link Setting the association strength is represented as:

[0047] (1.10)

[0048] in, Represents a node With nodes The strength of the correlation between them; This represents the credibility mapping value of a knowledge unit; Indicates the frequency coefficient across documents; This represents the source reliability coefficient; the formula uses logarithmic compression to make the strength growth of high-frequency, high-reliability relationships more stable.

[0049] (3) Based on the aforementioned knowledge association network, a three-layer causal structure is constructed:

[0050] (1.11)

[0051] in, This represents a three-layered causal structure. Represents the set of nodes in the causal layer; Represents the set of nodes in the mechanism layer; Represents the set of disease layer nodes; This represents the transfer mapping from the causal layer to the mechanistic layer; This represents the transfer mapping from the mechanism layer to the disease layer;

[0052] (4) The set of nodes in the causal layer is represented as:

[0053] (1.12)

[0054] The set of nodes in the mechanism layer is represented as follows:

[0055] (1.13)

[0056] The set of disease layer nodes is represented as follows:

[0057] (1.14)

[0058] in, This is a set of nodes in the causation layer, used to represent various external causations that lead to the occurrence of the target disease; This is a set of nodes at the mechanism layer, used to represent the intermediate mechanisms triggered by factors and acting on the disease formation process; This is a set of disease layer nodes used to represent various identifiable pavement defects. Indicates the first in the causative layer 1 node, Represents the first in the mechanism layer One node; Indicates the first layer of disease Each node.

[0059] Furthermore, step P4 is implemented through the following steps:

[0060] (1) For any target disease Define the trigger Its contribution to the original dissemination is as follows:

[0061] (1.15)

[0062] in, Indicates the trigger Diseases The original contribution to the spread; Indicates the trigger To mechanism The transmission strength; Representation mechanism To disease The transmission strength; Indicates the trigger arrive The transmission strength of each mechanism node; Representation mechanism To the The transmission strength of each disease node; k represents the node index of the mechanism layer; For the normalized summation index of the mechanism layer; Normalized summation index for the disease layer.

[0063] (2) To further consider the path length and supporting evidence, a revised propagation contribution is introduced:

[0064] (1.16)

[0065] in, Indicates the revised propagation contribution; Indicates the contribution of the original propagation; Indicates the trigger To disease The effective path length; Indicates the path decay factor; Indicates evidence compensation;

[0066] (3) The modified transmission contributions of all inducing factors are ranked to obtain the disease. Key triggers set:

[0067] (1.17)

[0068] in, Indicates disease The corresponding set of the first K key triggers; Indicates the revised propagation contribution; This indicates the number of key triggers selected.

[0069] (4) The set of key incentives after sorting Further tracing its corresponding dominant mechanism chain and related evidence will help to identify the target disease. The explanation path.

[0070] Furthermore, step P5 is implemented through the following steps:

[0071] (1) Query problem based on user input Combined with domain literature datasets and structured knowledge sets Construct a mapping relationship for diagnostic tasks:

[0072] (1.18)

[0073] in, Represents the diagnostic mapping function; Indicates the diagnostic conclusion; This represents the set of key triggers; Represents the set of interpretation paths;

[0074] (2) Retrieve candidate evidence related to the query question from both the text knowledge base and the knowledge association network, and define the dual-channel coupling retrieval value as:

[0075] (1.19)

[0076] in, Indicates candidate evidence fragments Dual-channel coupled retrieval values; This indicates the query question entered by the user; This represents the text semantic retrieval response value; This indicates the response value for the map path retrieval;

[0077] (3) Based on the dual-channel coupled retrieval values, and further considering the causal chain support capability, the evidence priority value is defined as:

[0078] (1.20)

[0079] in, This indicates evidence fragments. The priority value; This indicates the query question entered by the user; Indicates the causal enhancement coefficient; This indicates evidence fragments. The strength of the support for the corresponding causal chain;

[0080] (4) Input the sorted evidence set, key cause set, and dominant mechanism chain into the large language model, and define the first... The final diagnostic output probability for this type of disease is:

[0081] (1.21)

[0082] in, Indicates the first The final diagnostic output probability of the disease type; Indicates the disease identification response value; This indicates the key factors supporting the response value; Indicates the response value for evidence consistency; , , Indicates the corresponding adjustment parameter; This represents the disease category index variable used when summing the denominators; Indicates the total number of candidate disease categories; Indicates the first Disease identification response values ​​for different types of diseases; Indicates the first Key inducing factors supporting response values ​​for this type of disease; Indicates the first Consistency response value of evidence for similar diseases.

[0083] (5) Target disease The structured diagnostic results are represented as follows:

[0084] (1.22)

[0085] in, Indicates disease Structured diagnostic results This represents the set of key triggers; Indicates the dominant mechanism chain; This indicates the set of supporting evidence; This represents a collection of treatment suggestions;

[0086] (6) Based on the structured diagnostic results It outputs a diagnostic report that includes disease identification results, key causes, dominant mechanism chain, supporting evidence, and recommended treatment measures, thus realizing a complete technical process from disease description input to disease identification, cause analysis, mechanism explanation, and treatment recommendations.

[0087] Beneficial effects:

[0088] (1) This invention constructs a knowledge association network for pavement defect diagnosis by uniformly acquiring, structurally extracting, and standardizing knowledge related to pavement defects from academic literature, industry standards, etc., thereby solving the problems of scattered sources, inconsistent expressions, and difficulty in integration and reuse of defect knowledge in the prior art, and improving the uniformity, manageability, and reusability of pavement defect diagnosis knowledge. This framework realizes the complete process from multi-source knowledge acquisition and structured representation to intelligent diagnostic output, and improves the structured expression ability of the relationship between defect types, causal factors, and engineering conditions.

[0089] (2) Based on the knowledge association network, this invention constructs a three-layer causal structure of "inducement layer - mechanism layer - disease layer", which improves the diagnosis of pavement diseases from simple surface correlation matching to explicit causal propagation analysis. It can characterize the transmission path of external inducements evolving into target diseases through intermediate action mechanisms, thereby overcoming the shortcomings of existing technologies in describing the formation mechanism of diseases and the coupling relationship of multiple factors, and improving the accuracy of key inducement identification and the depth of diagnostic analysis. As shown in Table 1, the experimental results show that the method of this invention achieves 80% in the Explanation Chain Coverage (ECC) index, which is higher than 70% for Dense-RAG, 72% for KG-RAG and 75% for Hybrid-RAG; and reaches 0.91 in the Mean Reverse Rank (MRR) index, which is higher than 0.84 for Dense-RAG, 0.86 for KG-RAG and 0.88 for Hybrid-RAG. This shows that the method of this invention can not only generate a more complete causal explanation chain, but also return key evidence more quickly.

[0090] Table 1 Evaluation Results

[0091]

[0092] (3) This invention integrates text knowledge retrieval and knowledge association network retrieval, and combines the strength of causal support to enhance the ranking of evidence. This overcomes the problems of insufficient evidence recall and unstable ranking caused by relying solely on keyword matching, text similarity matching or single knowledge query methods in the prior art, and improves the retrieval ability, association ability and ranking quality of evidence related to the target disease. Specifically, as shown in Table 1, the method of the present invention achieves 0.89, 0.87, 0.89, 0.88, and 85% in Precision, Recall, F1 score, NDCG@5, and P@5, respectively. These values ​​are superior to those of Dense-RAG (0.82, 0.75, 0.78, 0.76, and 78%), KG-RAG (0.84, 0.80, 0.82, 0.79, and 80%), and Hybrid-RAG (0.86, 0.83, 0.84, 0.82, and 83%), indicating that the present invention has superior performance in evidence recall, evidence prioritization, and diagnostic accuracy.

[0093] (4) When outputting diagnostic results, this invention not only provides disease type identification results, but also simultaneously outputs the key inducing factor set, dominant mechanism chain, and supporting evidence, thereby forming a traceable interpretation path. This solves the problems of insufficient interpretability of diagnostic results, unclear evidence chain, and difficulty for engineers to verify the diagnostic basis in the prior art, and enhances the interpretability, verifiability, and engineering credibility of the diagnostic results. The method of this invention has achieved a further transformation from "related information retrieval" to "causal evidence organization," and in complex semantic expressions and multi-factor coupling scenarios, it has better interpretability and ranking ability than KG-RAG and Hybrid-RAG.

[0094] (5) This invention integrates disease identification, cause analysis, mechanism explanation, evidence association, and treatment suggestion generation into a complete closed-loop process from disease description input to diagnostic conclusion output. This overcomes the problem of the separation between identification, analysis, and maintenance suggestions in the prior art, and can provide road maintenance engineers with more complete, continuous, and practical decision support, thus improving the application value of this invention in actual engineering scenarios. Engineering case studies show that, as shown in Table 2, this invention selected 10 typical pavement disease descriptions for testing. The examples cover a variety of common diseases such as longitudinal cracks, transverse cracks, network cracks, rutting, potholes, loosening, bulging, misalignment, and bleeding. The model can automatically identify the disease type based on the input disease phenomenon and further provide Top 3 cause analysis, comprehensive diagnostic conclusions, and recommended treatment measures, indicating that this invention has good diagnostic adaptability and engineering practical value.

[0095] Table 2 Output results of the intelligent pavement distress diagnosis system

[0096]

[0097] (6) This invention also demonstrates good feasibility in adapting to the underlying large language model. For example... Figure 6 As shown, this invention further compares the performance of five basic large language models in the task of generating answers for pavement distress diagnosis. Qwen2.5:7B achieved BLEU-4 and ROUGE-L scores of 67.5 and 78.5 respectively, while Qwen2 achieved 66.3 and 77.2 respectively. These scores are significantly better than ChatGLM3's 61.1 and 73.4, InternLM2.5's 52.5 and 67.5, and Yi1.5's 28.4 and 41.8. This demonstrates that the diagnostic framework proposed in this invention can form a good synergy with high-performance large language models, thereby further improving the quality of generated and expressed diagnostic results.

[0098] In summary, this invention not only has significant advantages in knowledge organization, causal modeling, retrieval ranking, and interpretation chain generation, but also verifies its comprehensive performance in terms of diagnostic accuracy, evidence retrieval quality, interpretation completeness, and engineering application value through experimental results. It can effectively improve the reliability and interpretability of intelligent diagnosis of pavement defects. Attached Figure Description

[0099] Figure 1 This is a schematic diagram of the method framework of the present invention;

[0100] Figure 2 This is a schematic diagram illustrating the relationship between disease types and causes in this invention;

[0101] Figure 3 This is a schematic diagram of the clustering of pavement distress research topics based on knowledge graph triples in this invention;

[0102] Figure 4 This is a schematic diagram of the three-layer causal reasoning structure of the present invention;

[0103] Figure 5 This is a schematic diagram illustrating the calculation of causal propagation and causal contribution in this invention;

[0104] Figure 6 This is a schematic diagram showing the performance evaluation results of the large language model for pavement distress diagnosis of the present invention. Detailed Implementation

[0105] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0106] like Figure 1 As shown, a method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism includes the following steps:

[0107] P1: Establishing a multi-source knowledge acquisition and structured representation path for pavement distress diagnosis:

[0108] Collect multi-source heterogeneous data related to pavement distress diagnosis, including academic literature, industry standards, engineering cases, inspection records, and maintenance experience. Perform text cleaning, noise reduction, paragraph segmentation, and semantic blockization on the raw data to form a standardized text set that can be used for knowledge extraction. Further construct a knowledge ontology in the field of pavement distress diagnosis, and uniformly define distress entities, causal entities, mechanism entities, and engineering condition entities to provide a standard semantic framework for subsequent knowledge extraction and knowledge graph construction.

[0109] The P1 process is implemented through the following steps:

[0110] Let the set of original knowledge sources related to pavement distress diagnosis be:

[0111] (1.1)

[0112] in, Represents a set of multi-source knowledge sources. Indicates the first A source of knowledge This indicates the total number of knowledge sources; these knowledge sources include academic literature, industry technical standards, typical engineering cases, testing records, and related online data.

[0113] (1) The original knowledge source set Text cleaning, paragraph segmentation, semantic block division, and manual proofreading are performed to obtain a preprocessed set of text fragments.

[0114] (1.2)

[0115] in, This represents a set of preprocessed text fragments. Indicates the first A text fragment, Indicates the total number of text fragments;

[0116] (2) To evaluate the usability of each text segment, the text segment effectiveness evaluation function is defined as follows:

[0117] (1.3)

[0118] in, Represents a text fragment The effectiveness score; Indicates the cleanliness index of the text; Indicates semantic completeness index; Indicators representing domain relevance; , , This represents the corresponding adjustment coefficient; the formula uses a compression mapping method to limit the score to the range of 0 to 1, which is used to filter high-quality knowledge fragments.

[0119] (3) After completing data preprocessing, a semantic framework for pavement defect diagnosis is constructed and defined as follows:

[0120] (1.4)

[0121] in, Represents a domain semantic framework; Represents a set of entity units; Represents a set of relational constraints; Represents a collection of attribute descriptions; This represents a set of hierarchical mapping rules.

[0122] (4) Assemble the entity unit set Further expressed as:

[0123] (1.5)

[0124] in, Represents a collection of diseased entities. Represents the set of causal entities. Represents a set of mechanism entities. This represents the set of engineering condition entities; this entity partitioning method provides a unified semantic foundation for subsequent knowledge unit extraction and three-layer causal modeling.

[0125] P2: Extract knowledge triples using a large language model, and complete entity normalization and relation standardization:

[0126] The preprocessed text is input into the large language model Qwen2.5:7B to extract semantic relationships between disease types, influencing factors, mechanisms of action, and engineering conditions, forming knowledge triples in the form of Head-Relation-Tail. The confidence of the extracted knowledge triples is evaluated by combining domain rules, consistency checks, and manual verification. Synonymous entities, near-synonyms, and duplicate records from different knowledge sources are normalized to improve the consistency and accuracy of knowledge representation.

[0127] The P2 process is implemented through the following steps:

[0128] (1) Using a large language model to perform semantic understanding and structured information extraction on the preprocessed text fragments, the unstructured text is converted into knowledge units; let the first... The knowledge units are:

[0129] (1.6)

[0130] in, Indicates the first Each knowledge unit; Represents the source entity; Semantic links Represents the target entity; This indicates the addition of context markers; unlike the conventional triple structure, this formula, by introducing context markers, allows knowledge representation to simultaneously preserve semantic relationships and contextual information.

[0131] (2) To ensure the reliability of the knowledge extraction results, a credibility function is constructed for each knowledge unit:

[0132] (1.7)

[0133] in, Indicates the first The overall credibility of each knowledge unit; Indicates the credibility of the language model generation; Indicates the credibility of rule matching Indicates the credibility of knowledge consistency; This represents the adjustment factor for each confidence level; the formula uses a complementary probability product structure, which differs from the conventional linear weighted summation form.

[0134] (3) Perform unified mapping on synonymous entities from different sources; assume two candidate entities and The merge strength is:

[0135] (1.8)

[0136] in Indicates candidate entities and Merge strength; Indicates literal similarity; Indicates the semantic proximity of vectors; Indicates the proximity of co-occurrence in context;

[0137] (4) Standardize the semantic links in the knowledge unit and map similar relationships in different texts into a unified semantic link category to improve the consistency and stability of the subsequent knowledge association network construction.

[0138] like Figure 2 As shown, this invention establishes the correlation between different types of pavement defects and their candidate causes based on the results of knowledge extraction and normalization. This is used to characterize the distribution characteristics of the main inducing factors corresponding to various defects, thereby providing basic support for subsequent identification and ranking of key causes.

[0139] like Figure 3 As shown, this invention further organizes knowledge in the field of pavement defects by topic clustering based on knowledge graph triples, and expresses crack-type defects, water damage-related mechanisms, and vehicle and structural deformation factors in a graph-based manner to support subsequent semantic retrieval, knowledge association, and reasoning analysis.

[0140] P3: Construct a knowledge graph for the field of road surface defects and establish a three-layer causal network of "cause layer - mechanism layer - defect layer":

[0141] The normalized knowledge triples are imported into the Neo4j graph database to construct a knowledge graph for the pavement disease domain, which includes disease nodes, cause nodes, mechanism nodes, and engineering condition nodes. The weights of the relation edges are set according to the reliability of the knowledge source, the frequency of relation occurrence, and the confidence of the triples. Based on the knowledge graph, a three-layer causal network of "cause layer - mechanism layer - disease layer" is further established to characterize the causal transmission process of external factors evolving into target diseases through intermediate mechanisms.

[0142] The P3 process is implemented through the following steps:

[0143] (1) Import the standardized knowledge units into the graph database to construct a knowledge association network for pavement defects; let the knowledge association network be:

[0144] (1.9)

[0145] in, Represents a knowledge association network; Represents a set of node clusters; Represents a collection of links; Represents the set of link strengths;

[0146] (2) To enhance the reasoning ability of the network, for any link Setting the association strength is represented as:

[0147] (1.10)

[0148] in, Represents a node With nodes The strength of the correlation between them; This represents the credibility mapping value of a knowledge unit; Indicates the frequency coefficient across documents; This represents the source reliability coefficient; the formula uses logarithmic compression to make the strength growth of high-frequency, high-reliability relationships more stable.

[0149] (3) Based on the aforementioned knowledge association network, a three-layer causal structure is constructed:

[0150] (1.11)

[0151] in, This represents a three-layered causal structure. Represents the set of nodes in the causal layer; Represents the set of nodes in the mechanism layer; Represents the set of disease layer nodes; This represents the transfer mapping from the causal layer to the mechanistic layer; This represents the transfer mapping from the mechanism layer to the disease layer;

[0152] (4) The set of nodes in the causal layer is represented as:

[0153] (1.12)

[0154] The set of nodes in the mechanism layer is represented as follows:

[0155] (1.13)

[0156] The set of disease layer nodes is represented as follows:

[0157] (1.14)

[0158] in, This is a set of nodes in the causation layer, used to represent various external causations that lead to the occurrence of the target disease; This is a set of nodes at the mechanism layer, used to represent the intermediate mechanisms triggered by factors and acting on the disease formation process; This is a set of disease layer nodes used to represent various identifiable pavement defects. Indicates the first in the causative layer 1 node, Represents the first in the mechanism layer One node; Indicates the first layer of disease Each node.

[0159] like Figure 4 As shown, this invention constructs a three-layer causal reasoning structure of "cause-mechanism-disease form" based on a knowledge association network. The cause layer is used to represent external inducements, the mechanism layer is used to represent the intermediate evolution mechanism in the process of disease formation, and the disease form layer is used to represent the final observable road surface disease type. The causal transmission model from inducement to disease is realized through the mapping relationship between each layer.

[0160] P4: Perform causal probability propagation and candidate cause ranking to establish an explanatory path for the target disease:

[0161] For the target disease, the influence of each candidate cause on the target disease is calculated based on the three-layer causal network. Through conditional probability propagation from the cause layer to the mechanism layer and from the mechanism layer to the disease layer, the comprehensive contribution of different potential causes to the target disease is obtained. Combining path weight, relationship strength and evidence support, the candidate causes are ranked to obtain the Top-N cause set of the target disease and its corresponding dominant mechanism path, thereby realizing the interpretable inversion from disease manifestation to potential causes.

[0162] The P4 process is implemented through the following steps:

[0163] (1) For any target disease Define the trigger Its contribution to the original dissemination is as follows:

[0164] (1.15)

[0165] in, Indicates the trigger Diseases The original contribution to the spread; Indicates the trigger To mechanism The transmission strength; Representation mechanism To disease The transmission strength; Indicates the trigger arrive The transmission strength of each mechanism node; Representation mechanism To the The transmission strength of each disease node; k represents the node index of the mechanism layer; For the normalized summation index of the mechanism layer; Normalized summation index for the disease layer.

[0166] (2) To further consider the path length and supporting evidence, a revised propagation contribution is introduced:

[0167] (1.16)

[0168] in, Indicates the revised propagation contribution; Indicates the contribution of the original propagation; Indicates the trigger To disease The effective path length; Indicates the path decay factor; Indicates evidence compensation;

[0169] (3) The modified transmission contributions of all inducing factors are ranked to obtain the disease. Key triggers set:

[0170] (1.17)

[0171] in, Indicates disease The corresponding set of the first K key triggers; Indicates the revised propagation contribution; This indicates the number of key triggers selected.

[0172] (4) The set of key incentives after sorting Further tracing its corresponding dominant mechanism chain and related evidence will help to identify the target disease. The explanation path.

[0173] like Figure 5 As shown, this invention uses a constructed three-layer causal network to propagate the conditional probability between candidate causes and target diseases, and uses a matrix-based approach to characterize the contribution intensity of different cause nodes to different disease nodes, thereby realizing key cause identification, contribution ranking, and generation of dominant explanatory paths.

[0174] P5: Integrating text retrieval and knowledge graph retrieval to generate structured diagnostic results:

[0175] For the disease description information or query questions input by users, relevant evidence is retrieved from the text knowledge base and knowledge graph, respectively. The retrieval results are then fused, reordered, and causal enhancement is performed. The ordered text evidence, graph relationships, candidate causes, and causal paths are input into the large language model to generate a structured diagnostic report that includes at least the disease identification results, key cause set, diagnostic conclusion, causal explanation chain, and recommended treatment measures. This achieves a complete technical process from disease description input to disease identification, cause analysis, evidence association, and treatment suggestions output.

[0176] The P5 process is implemented through the following steps:

[0177] (1) Query problem based on user input Combined with domain literature datasets and structured knowledge sets Construct a mapping relationship for diagnostic tasks:

[0178] (1.18)

[0179] in, Represents the diagnostic mapping function; Indicates the diagnostic conclusion; This represents the set of key triggers; Represents the set of interpretation paths;

[0180] (2) Retrieve candidate evidence related to the query question from both the text knowledge base and the knowledge association network, and define the dual-channel coupling retrieval value as:

[0181] (1.19)

[0182] in, Indicates candidate evidence fragments Dual-channel coupled retrieval values; This indicates the query question entered by the user; This represents the text semantic retrieval response value; This indicates the response value for the map path retrieval;

[0183] (3) Based on the dual-channel coupled retrieval values, and further considering the causal chain support capability, the evidence priority value is defined as:

[0184] (1.20)

[0185] in, This indicates evidence fragments. The priority value; This indicates the query question entered by the user; Indicates the causal enhancement coefficient; This indicates evidence fragments. The strength of the support for the corresponding causal chain;

[0186] (4) Input the sorted evidence set, key cause set, and dominant mechanism chain into the large language model, and define the first... The final diagnostic output probability for this type of disease is:

[0187] (1.21)

[0188] in, Indicates the first The final diagnostic output probability of the disease type; Indicates the disease identification response value; This indicates the key factors supporting the response value; Indicates the response value for evidence consistency; , , Indicates the corresponding adjustment parameter; This represents the disease category index variable used when summing the denominators; Indicates the total number of candidate disease categories; Indicates the first Disease identification response values ​​for different types of diseases; Indicates the first Key inducing factors supporting response values ​​for this type of disease; Indicates the first Consistency response value of evidence for similar diseases.

[0189] (5) Target disease The structured diagnostic results are represented as follows:

[0190] (1.22)

[0191] in, Indicates disease Structured diagnostic results This represents the set of key triggers; Indicates the dominant mechanism chain; This indicates the set of supporting evidence; This represents a collection of treatment suggestions;

[0192] (6) Based on the structured diagnostic results It outputs a diagnostic report that includes disease identification results, key causes, dominant mechanism chain, supporting evidence, and recommended treatment measures, thus realizing a complete technical process from disease description input to disease identification, cause analysis, mechanism explanation, and treatment recommendations.

[0193] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism, characterized in that: Includes the following steps: P1: Establishing a multi-source knowledge acquisition and structured representation path for pavement distress diagnosis: Collect multi-source heterogeneous data related to pavement distress diagnosis, including academic literature, industry standards, engineering cases, inspection records, and maintenance experience. Perform text cleaning, noise reduction, paragraph segmentation, and semantic blockization on the raw data to form a standardized text set that can be used for knowledge extraction. Further construct a knowledge ontology in the field of pavement distress diagnosis, and uniformly define distress entities, causal entities, mechanism entities, and engineering condition entities to provide a standard semantic framework for subsequent knowledge extraction and knowledge graph construction. P2: Extract knowledge triples using a large language model, and complete entity normalization and relation standardization: The preprocessed text is input into the large language model Qwen2.5:7B to extract semantic relationships between disease types, influencing factors, mechanisms of action, and engineering conditions, forming knowledge triples in the form of Head-Relation-Tail. The confidence of the extracted knowledge triples is evaluated by combining domain rules, consistency checks, and manual verification. Synonymous entities, near-synonyms, and duplicate records from different knowledge sources are normalized to improve the consistency and accuracy of knowledge representation. P3: Construct a knowledge graph for the field of road surface defects and establish a three-layer causal network of "cause layer - mechanism layer - defect layer": The normalized knowledge triples are imported into the Neo4j graph database to construct a knowledge graph for the pavement disease domain, which includes disease nodes, cause nodes, mechanism nodes, and engineering condition nodes. The weights of the relation edges are set according to the reliability of the knowledge source, the frequency of relation occurrence, and the confidence of the triples. Based on the knowledge graph, a three-layer causal network of "cause layer - mechanism layer - disease layer" is further established to characterize the causal transmission process of external factors evolving into target diseases through intermediate mechanisms. P4: Perform causal probability propagation and candidate cause ranking to establish an explanatory path for the target disease: For the target disease, the influence of each candidate cause on the target disease is calculated based on the three-layer causal network. Through conditional probability propagation from the cause layer to the mechanism layer and from the mechanism layer to the disease layer, the comprehensive contribution of different potential causes to the target disease is obtained. Combining path weight, relationship strength and evidence support, the candidate causes are ranked to obtain the Top-N cause set of the target disease and its corresponding dominant mechanism path, thereby realizing the interpretable inversion from disease manifestation to potential causes. P5: Integrating text retrieval and knowledge graph retrieval to generate structured diagnostic results: For the disease description information or query questions input by users, relevant evidence is retrieved from the text knowledge base and knowledge graph, respectively. The retrieval results are then fused, reordered, and causal enhancement is performed. The sorted text evidence, graph relationships, candidate causes, and causal paths are input into the large language model to generate a structured diagnostic report that includes at least the disease identification results, key cause set, diagnostic conclusion, causal explanation chain, and recommended treatment measures. This achieves a complete technical process from disease description input to disease identification, cause analysis, evidence association, and treatment suggestions output.

2. The method for diagnosing pavement defects based on a large language model using a three-layer causal reasoning mechanism as described in claim 1, characterized in that: The process in step P1 is implemented through the following steps: Let the set of original knowledge sources related to pavement distress diagnosis be: （1.1）； in, Represents a set of multi-source knowledge sources. Indicates the first A source of knowledge This indicates the total number of knowledge sources; these knowledge sources include academic literature, industry technical standards, typical engineering cases, testing records, and related online data. (1) The original knowledge source set Text cleaning, paragraph segmentation, semantic block division, and manual proofreading are performed to obtain a preprocessed set of text fragments. （1.2）； in, This represents a set of preprocessed text fragments. Indicates the first A text fragment, Indicates the total number of text fragments; (2) To evaluate the usability of each text segment, the text segment effectiveness evaluation function is defined as follows: （1.3）； in, Represents a text fragment The effectiveness score; Indicates the cleanliness index of the text; Indicates semantic completeness index; Indicators representing domain relevance; , , This represents the corresponding adjustment coefficient; the formula uses a compression mapping method to limit the score to the range of 0 to 1, which is used to filter high-quality knowledge fragments. (3) After completing data preprocessing, a semantic framework for pavement defect diagnosis is constructed and defined as follows: （1.4）； in, Represents a domain semantic framework; Represents a set of entity units; Represents a set of relational constraints; Represents a collection of attribute descriptions; Represents a set of hierarchical mapping rules; (4) Assemble the entity unit set Further expressed as: （1.5）； in, Represents a set of entity units. Represents a collection of diseased entities. Represents the set of causal entities. Represents a set of mechanism entities. This represents the set of engineering condition entities; this entity partitioning method provides a unified semantic foundation for subsequent knowledge unit extraction and three-layer causal modeling.

3. The method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism according to claim 2, characterized in that: The process in step P2 is implemented through the following steps: (1) Using a large language model to perform semantic understanding and structured information extraction on the preprocessed text fragments, the unstructured text is converted into knowledge units; let the first... The knowledge units are: （1.6）； in, Indicates the first Each knowledge unit; Represents the source entity; Indicates semantic linking; Represents the target entity; This indicates the addition of context markers; unlike the conventional triple structure, this formula, by introducing context markers, allows knowledge representation to simultaneously preserve semantic relationships and contextual information. (2) To ensure the reliability of the knowledge extraction results, a credibility function is constructed for each knowledge unit: （1.7）； in, Indicates the first The overall credibility of each knowledge unit; Indicates the credibility of the language model generation; Indicates the reliability of rule matching; Indicates the credibility of knowledge consistency; This represents the adjustment factor for each confidence level; the formula uses a complementary probability product structure, which differs from the conventional linear weighted summation form. (3) Perform unified mapping on synonymous entities from different sources; assume two candidate entities and The merge strength is: （1.8）； in Indicates candidate entities and Merge strength; Indicates literal similarity; Indicates the semantic proximity of vectors; Indicates the proximity of co-occurrence in context; (4) Standardize the semantic links in the knowledge unit and map similar relationships in different texts into a unified semantic link category to improve the consistency and stability of the subsequent knowledge association network construction.

4. The method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism according to claim 3, characterized in that: The process in step P3 is implemented through the following steps: (1) Import the standardized knowledge units into the graph database to construct a knowledge association network for pavement defects; let the knowledge association network be: （1.9）； in, Represents a knowledge association network; Represents a set of node clusters; Represents a collection of links; Represents the set of link strengths; (2) To enhance the reasoning ability of the network, for any link Setting the association strength is represented as: （1.10）； in, Represents a node With nodes The strength of the correlation between them; This represents the credibility mapping value of a knowledge unit; Indicates the frequency coefficient across documents; This represents the source reliability coefficient; the formula uses logarithmic compression to make the strength growth of high-frequency, high-reliability relationships more stable. (3) Based on the aforementioned knowledge association network, a three-layer causal structure is constructed: （1.11）； in, This represents a three-layered causal structure. Represents the set of nodes in the causal layer; Represents the set of nodes in the mechanism layer; Represents the set of disease layer nodes; This represents the transfer mapping from the causal layer to the mechanistic layer; This represents the transfer mapping from the mechanism layer to the disease layer; (4) The set of nodes in the causal layer is represented as: （1.12）； The set of nodes in the mechanism layer is represented as follows: （1.13）； The set of disease layer nodes is represented as follows: （1.14）； in, This is a set of nodes in the causation layer, used to represent various external causations that lead to the occurrence of the target disease; This is a set of nodes at the mechanism layer, used to represent the intermediate mechanisms triggered by factors and acting on the disease formation process; This is a set of disease layer nodes used to represent various identifiable pavement defects. Indicates the first in the causative layer 1 node, Represents the first in the mechanism layer One node; Indicates the first layer of disease Each node.

5. The method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism according to claim 4, characterized in that: The process in step P4 is implemented through the following steps: (1) For any target disease Define the trigger Its contribution to the original dissemination is as follows: （1.15）； in, Indicates the trigger Diseases The original contribution to the spread; Indicates the trigger To mechanism The transmission strength; Representation mechanism To disease The transmission strength; Indicates the trigger arrive The transmission strength of each mechanism node; Representation mechanism To the The transmission strength of each disease node; k represents the node index of the mechanism layer; For the normalized summation index of the mechanism layer; Normalized summation index for the disease layer; (2) To further consider the path length and supporting evidence, a revised propagation contribution is introduced: （1.16）； in, Indicates the revised propagation contribution; Indicates the contribution of the original propagation; Indicates the trigger To disease The effective path length; Indicates the path decay factor; Indicates evidence compensation; (3) The modified transmission contributions of all inducing factors are ranked to obtain the disease. Key triggers set: （1.17）； in, Indicates disease The corresponding set of the first K key triggers; Indicates the revised propagation contribution; This indicates the number of key triggers selected; (4) The set of key incentives after sorting Further investigation is needed to trace the corresponding dominant mechanism chain and related evidence to form the target disease. The explanation path.

6. The method for diagnosing pavement defects based on a large language model with a three-layer causal reasoning mechanism according to claim 5, characterized in that: The process in step P5 is implemented through the following steps: (1) Query problem based on user input Combined with domain literature datasets and structured knowledge sets Construct a mapping relationship for diagnostic tasks: （1.18）； in, Represents the diagnostic mapping function; Indicates the diagnostic conclusion; This represents the set of key triggers; Represents the set of interpretation paths; (2) Retrieve candidate evidence related to the query question from both the text knowledge base and the knowledge association network, and define the dual-channel coupling retrieval value as: （1.19）； in, Indicates candidate evidence fragments Dual-channel coupled retrieval values; This indicates the query question entered by the user; This represents the text semantic retrieval response value; This indicates the response value for the map path retrieval; (3) Based on the dual-channel coupled retrieval values, and further considering the causal chain support capability, the evidence priority value is defined as: （1.20）； in, This indicates evidence fragments. The priority value; This indicates the query question entered by the user; Indicates the causal enhancement coefficient; This indicates evidence fragments. The strength of the support for the corresponding causal chain; (4) Input the sorted evidence set, key cause set, and dominant mechanism chain into the large language model, and define the first... The final diagnostic output probability for this type of disease is: （1.21）； in, Indicates the first The final diagnostic output probability of the disease type; Indicates the disease identification response value; This indicates the key factors supporting the response value; Indicates the response value for evidence consistency; , , Indicates the corresponding adjustment parameter; This represents the disease category index variable used when summing the denominators; Indicates the total number of candidate disease categories; Indicates the first Disease identification response values ​​for different types of diseases; Indicates the first Key inducing factors supporting response values ​​for this type of disease; Indicates the first Evidence consistency response value for similar diseases; (5) Target disease The structured diagnostic results are represented as follows: （1.22）； in, Indicates disease Structured diagnostic results; This represents the set of key triggers; Indicates the dominant mechanism chain; This indicates the set of supporting evidence; This represents a collection of treatment suggestions; (6) Based on the structured diagnostic results It outputs a diagnostic report that includes disease identification results, key causes, dominant mechanism chain, supporting evidence, and recommended treatment measures, thus realizing a complete technical process from disease description input to disease identification, cause analysis, mechanism explanation, and treatment recommendations.

Images