A highway disease intelligent diagnosis system based on multi-agent cooperation

Abstract

This invention discloses an intelligent highway disease diagnosis system based on multi-agent collaboration, comprising: a data construction module for collecting and preprocessing multi-source highway business data and disease-related information; a preliminary retrieval module for generating a first query result data set; a structured query generation module for generating a second query result data set and a complete data set; a workflow generation module for generating a highway disease diagnosis workflow description text; a causal reasoning module for generating a disease state vector sequence and a disease causal relationship graph; a multi-agent execution module for generating a multi-agent diagnosis intermediate result set; a diagnosis reasoning module for generating a highway disease diagnosis result set; and a diagnosis output and optimization module for generating a diagnostic natural language report, structured diagnostic results, and a continuous optimization diagnostic task. This invention employs a large language model and multi-agent collaboration to achieve intelligent highway disease diagnosis, offering the advantage of high efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation infrastructure detection technology, and in particular to an intelligent diagnosis system for highway defects based on multi-agent collaboration. Background Technology

[0002] Current highway defect diagnosis relies heavily on manual inspections, human experience-based judgment, and semi-automated identification tools. There is a lack of unified integration methods among highway design data, construction data, inspection data, monitoring data, and maintenance data. Knowledge is not presented in a structured manner, resulting in fragmented diagnostic criteria and isolated information. Traditional diagnostic techniques generally employ single-model processing, failing to achieve collaborative reasoning between knowledge and data, and struggling to handle cross-table and multi-dimensional data query needs. This limits both accuracy and timeliness when facing complex highway scenarios.

[0003] Meanwhile, existing diagnostic systems lack multi-agent collaboration mechanisms, failing to decompose complex diagnostic tasks into dynamic task flows, resulting in rigid diagnostic processes and difficulty in adaptive scheduling. The construction of causal relationships for diseases typically relies on static rules, making it difficult to reflect the temporal changes and correlations of real diseases. Existing technologies generally lack workflow generation capabilities based on large language models, dynamic causal reasoning capabilities based on improved fuzzy cognitive graphs, and efficient mapping mechanisms between natural language and databases, hindering the automation, intelligence, and continuous optimization of diagnostic processes.

[0004] Therefore, how to provide an intelligent diagnosis system for highway defects based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent diagnosis system for highway defects based on multi-agent collaboration. This invention employs a large language model and multi-agent collaboration to achieve intelligent diagnosis of highway defects, which has the advantage of high efficiency.

[0006] According to an embodiment of the present invention, a highway defect intelligent diagnosis system based on multi-agent collaboration includes: The data construction module is used to collect and preprocess multi-source business data and information related to road defects. The preliminary retrieval module transforms the query requirements in the highway disease diagnosis workflow into structured retrieval requests, and performs preliminary data retrieval based on knowledge graphs and digital twin data to generate the first query result data set. The structured query generation module is used to generate a second query result data set and integrate it with the first query result data set to form a complete data set; The workflow generation module takes the complete dataset and inputs it into the large language model to generate a description text of the highway disease diagnosis workflow, a multi-agent task list, and task constraint information. The causal reasoning module takes the complete dataset as input into the improved fuzzy cognitive graph and generates a disease state vector sequence and a disease causal relationship graph. The multi-agent execution module is used to perform data processing, disease identification, and disease analysis, and generate a set of intermediate results for multi-agent diagnosis. The diagnostic reasoning module is used to perform disease reasoning, assessment and trend analysis, and generate a set of highway disease diagnosis results. The diagnostic output and optimization module is used to generate natural language reports and structured diagnostic results for highway defects, and to perform quality control and workflow updates based on the diagnostic results and execution status, so as to achieve continuous optimization of diagnostic tasks.

[0007] Optionally, modules can be integrated using the following methods: Collect and process data, as well as standards and literature related to road defects, perform preprocessing and correlation modeling, and generate a knowledge graph dataset of road defects and a digital twin dataset of road defects. The query requirements in the highway disease diagnosis workflow are transformed into structured search requests, a preliminary data retrieval is performed, and the first query result data set is generated. The NL2SQL natural language conversion model generates a second query result data set based on the natural language diagnostic needs and the first query result data set, and integrates it with the first query result data set to form a complete data set. The complete dataset is input into the large language model to generate a description text of the highway disease diagnosis workflow, a multi-agent task list, and task constraint information. The complete dataset is input into the improved fuzzy cognitive graph, and through factor modeling, causal weight construction and temporal state update, a disease state vector sequence and a disease causal relationship graph are generated. Based on the multi-agent task list, the disease causal relationship graph and the disease state vector sequence, the multi-agents collaboratively perform data processing, disease identification and disease analysis to generate a set of multi-agent diagnostic intermediate results. The multi-agent diagnostic intermediate result set, the disease causal relationship graph and the disease state vector sequence are input into the time-series causal discovery algorithm, the dynamic Bayesian risk assessment algorithm and the trend prediction algorithm based on echo state network to perform disease inference, disease assessment and disease development trend analysis, and generate a set of highway disease diagnosis results. Based on the highway defect diagnosis result set and the highway defect diagnosis workflow description text, natural language reports and structured diagnosis results are generated for highway defect diagnosis. Quality control and workflow updates are performed to achieve continuous optimization of the diagnosis task.

[0008] Optionally, the generation of the highway defect knowledge graph dataset and the highway defect digital twin dataset includes: Collect highway design data, construction data, testing data, monitoring data and maintenance data, and compile relevant standards and literature on road defects into a raw data set; Based on the data type, the original dataset is subjected to field extraction, feature extraction, and time-series alignment to generate a structured dataset containing design fields, construction fields, inspection fields, monitoring fields, and maintenance fields; Based on the field content in the structured data set, road entities, structural entities, defect entities, monitoring entities, and maintenance entities are generated to form an entity set; Based on the field associations between entity sets, structural relationships, temporal relationships, and disease-related relationships are generated to form a data relationship set; Based on the correspondence between the entity set and the data relationship set, a highway disease knowledge graph data set containing entity nodes, relationship edges and rule content is constructed; The monitoring, detection, and maintenance fields are extracted from the structured dataset and aligned with the temporal relationships in the data relationship set to form the input content for digital twin modeling. The digital twin modeling inputs are then mapped to a state and updated over time to generate a digital twin dataset of highway defects.

[0009] Optionally, the generation of the first query result data set includes: The query requirements in the highway disease diagnosis workflow are parsed into structured query requirements that include target fields, constraint fields, and related fields. The structured query requirements, together with the highway disease knowledge graph dataset and the highway disease digital twin dataset, are used as the retrieval input. Based on the search input, semantic encoding, vector representation and similarity matching are performed through the RAG search model, and a structured search request for knowledge graph search and digital twin data search is constructed based on the matching results; Based on the structured retrieval request, a preliminary data retrieval is performed using the RAG retrieval model. The entity fields, attribute fields, and relation fields of the highway disease knowledge graph dataset are matched and calculated to generate preliminary knowledge graph retrieval results. Based on the structured search request, a preliminary data search is performed using the RAG search model. Cross-table and multi-dimensional matching calculations are performed on the monitoring, detection, and maintenance fields of the digital twin data set of highway defects to generate preliminary search results for digital twin data. The preliminary search results from the knowledge graph and the preliminary search results from the digital twin data are aligned by fields and integrated by content to form the first query result data set.

[0010] Optionally, the generation of the complete dataset includes: The system obtains the user's input natural language diagnostic requirements, performs word segmentation, semantic intent recognition, and field association processing on the natural language diagnostic requirements, and forms natural language query semantic content that includes semantic intent fields, constraint fields, and target fields. The semantic content of a natural language query is matched with a set of query results data through field matching and semantic completion to form query mapping content for converting natural language into structured queries; The query mapping content is input into the NL2SQL natural language conversion model, and a structured query statement generation step including semantic parsing, field mapping inference and multi-table join inference is performed to obtain the structured query statement; Execute cross-table, multi-dimensional database queries based on structured query statements to form a second query result data set; The field content in the second query result data set is compared with the field content in the first query result data set to complete and ensure consistency, forming the complete data set required for diagnosis.

[0011] Optionally, the generation of the highway defect diagnosis workflow description text, multi-agent task list, and task constraint information includes: The complete dataset is input into a large language model, and a description text of the highway disease diagnosis workflow is generated through semantic reasoning. The structured workflow text content is also generated according to the workflow execution order. Based on the description text of the highway disease diagnosis workflow, the task requirements of each workflow link are converted into corresponding task execution items through a large language model, forming a multi-agent task list arranged in the execution order. Based on the multi-agent task list, task constraint information is generated through a large language model, which describes the input field requirements, output field requirements, dependency field requirements, and task order requirements for each task execution item, thus forming task constraint information.

[0012] Optionally, the generation of the disease state vector sequence and the disease causal relationship graph includes: The complete dataset is input into the improved fuzzy cognitive graph. The disease attribute field, disease monitoring field, and disease association field in the complete dataset are extracted in a structured manner. A disease factor set is formed by the feature division rules of disease-related factors. Based on the disease factor set, causal weights among disease factors are constructed through an improved fuzzy cognitive graph. The direction, intensity, and conditions of action between disease influencing factors are quantified. The causal weights are updated based on the correlation fields in the complete dataset to form an initial causal weight set. The initial causal weight set is combined with the time-series monitoring fields in the complete dataset. The time-series state update is performed through the improved fuzzy cognitive graph. The disease factors are updated step by step according to the time-series state change rules, generating a disease state vector sequence corresponding to each time step. Based on the initial causal weight set and the disease state vector sequence, a disease causal relationship graph is formed through causal structure generation rules.

[0013] Optionally, the generation of the multi-agent diagnostic intermediate result set includes: Obtain the multi-agent task list and task constraint information, match the multi-agent task list with the disease causal relationship graph and disease state vector sequence, and form the multi-agent execution input content according to the data dependency conditions required for the task execution item; Based on the multi-agent execution input, each agent executes the corresponding task execution item according to the input field requirements, output field requirements, dependency field requirements and task order requirements specified in the task constraint information, and independently processes the data processing task, disease identification task and disease analysis task in the task execution item. When executing each task, each agent calls an image feature extraction algorithm based on convolutional neural networks, a disease feature analysis algorithm based on decision trees, and a disease classification algorithm based on logistic regression, depending on the type of task. The causal structure information in the disease causal relationship graph is used as the feature association constraint between disease factors, and the temporal state information in the disease state vector sequence is used as the time series feature input. The agent performs feature extraction, feature filtering, and classification judgment on the data during the task execution process, and generates the task execution result content through the feature vector, classification label, and confidence score output by the model. The task execution results of each agent are recorded. The agent scheduler dynamically adjusts the task execution order in the multi-agent task list based on the execution status information, data dependency status information and task completion status information during the task execution process, thus forming an updated scheduling result of the multi-agent execution process. Based on the updated scheduling results, each agent continues to execute the remaining tasks, and the execution results of all tasks are integrated into a multi-agent diagnostic intermediate result set according to the task order.

[0014] Optionally, the generation of the highway defect diagnosis result set includes: The multi-agent diagnostic intermediate result set, the disease causal relationship graph, and the disease state vector sequence are used as inference input content. Based on the inference input content, disease event fields, causal structure fields, and temporal state fields are extracted, and the three types of fields are combined to form the field set required for inference calculation. According to the calculation requirements of the disease inference algorithm, the inference calculation fields for disease type inference and disease location inference are selected from the field set required for the inference calculation. The disease type inference result and the disease location inference result are generated by performing correlation operation between the causal structure field and the disease event field. The disease type inference results and disease location inference results are combined with the time-series state fields in the disease state vector sequence. Based on the calculation requirements of the disease assessment algorithm and the disease development trend analysis algorithm, the inference calculation fields for disease risk assessment and disease trend analysis are selected from the field set required for inference calculation. The assessment calculation and trend calculation are performed to generate the disease development trend analysis results and the disease risk level assessment results. The results of disease type inference, disease location inference, disease development trend analysis, and disease risk level assessment are integrated according to the disease diagnosis rules to form a set of highway disease diagnosis results.

[0015] Optionally, the continuous optimization of the diagnostic task includes: The highway disease diagnosis result set and the highway disease diagnosis workflow description text are simultaneously input into the large language model. The report is generated based on the disease type field, disease location field, disease development trend field, and disease risk level field in the highway disease diagnosis result set, as well as the workflow task field and workflow sequence field in the highway disease diagnosis workflow description text. Based on the report, the input content is generated, and the large language model generates a natural language report on highway disease diagnosis and structured diagnosis results as output. Based on the diagnostic field content in the highway disease diagnosis result set and the result field content in the structured diagnosis result output, quality control processing is performed on the diagnosis execution status field, task completion status field and process execution deviation information to generate quality control result information. Based on the quality control results, the multi-agent task list and the highway defect diagnosis workflow description text are updated to form an updated task list and updated workflow description text. The updated task list and updated workflow description text are then used as input for the next round of highway defect diagnosis tasks, thereby achieving continuous optimization and iterative improvement of the highway defect diagnosis process and completing intelligent diagnosis of highway defects.

[0016] The beneficial effects of this invention are: This invention achieves structured integration of multi-source data from highway design, construction, inspection, monitoring, and maintenance by constructing a knowledge graph dataset and a digital twin dataset of highway defects. This makes the data foundation required for diagnosis more unified, complete, and reasonable. The structured processing of query requirements based on the RAG retrieval model automatically maps query targets in the highway defect diagnosis workflow into structured retrieval requests, ensuring efficient and accurate initial retrieval of the knowledge graph and digital twin data. Natural language diagnostic requirements are semantically parsed and structured query statements are generated using the NL2SQL model, achieving automatic conversion from natural language to cross-table queries in the database. This eliminates the need for manually constructing complex queries for professional data retrieval, significantly improving data acquisition efficiency.

[0017] This invention utilizes a large language model to generate a workflow description text for highway defect diagnosis, a multi-agent task list, and task constraint information. This provides the entire diagnostic process with a clear execution structure and unified data support, solving the problems of traditional systems where task allocation relies on manual intervention and the process lacks consistency. The improved fuzzy cognitive graph generates a sequence of defect state vectors and a causal relationship graph through causal weight construction and temporal state updates. This enables dynamic reasoning in defect development trend analysis, combining temporal data and causal structure, thus improving the accuracy of defect identification and analysis. The multi-agent collaborative execution mechanism distributes data processing, defect identification, and defect analysis tasks based on the task list and constraint information, maintaining the continuity and efficiency of the diagnostic process under the dynamic scheduling of the agent scheduler.

[0018] Simultaneously, the intermediate results of multi-agent diagnosis are used for disease inference, risk assessment, and trend analysis, ensuring the diagnostic results are professional, interpretable, and adaptable to various scenarios. Combined with a large language model-based diagnostic report generation and quality control mechanism, the system can automatically output natural language diagnostic reports and continuously optimize diagnostic tasks, allowing the workflow and task list to be continuously improved through multiple rounds of diagnosis. Through these technologies, this invention significantly improves the automation level, inference accuracy, and process intelligence of highway disease diagnosis, overcoming the shortcomings of existing technologies in areas such as data integration difficulties, weak inference capabilities, and a lack of self-optimization mechanisms. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the overall process of a highway disease intelligent diagnosis system based on multi-agent collaboration proposed in this invention. Figure 2This is a schematic diagram of the reasoning process for generating disease state vector sequences and disease causal relationship graphs based on improved fuzzy cognitive graphs in a multi-agent collaborative intelligent diagnosis system for highway diseases proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 and Figure 2 A multi-agent collaborative intelligent diagnosis system for highway defects includes: The data construction module is used to collect and preprocess multi-source business data and information related to road defects. The preliminary retrieval module transforms the query requirements in the highway disease diagnosis workflow into structured retrieval requests, and performs preliminary data retrieval based on knowledge graphs and digital twin data to generate the first query result data set. The structured query generation module is used to generate a second query result data set and integrate it with the first query result data set to form a complete data set; The workflow generation module takes the complete dataset and inputs it into the large language model to generate a description text of the highway disease diagnosis workflow, a multi-agent task list, and task constraint information. The causal reasoning module takes the complete dataset as input into the improved fuzzy cognitive graph and generates a disease state vector sequence and a disease causal relationship graph. The multi-agent execution module is used to perform data processing, disease identification, and disease analysis, and generate a set of intermediate results for multi-agent diagnosis. The diagnostic reasoning module is used to perform disease reasoning, assessment and trend analysis, and generate a set of highway disease diagnosis results. The diagnostic output and optimization module is used to generate natural language reports and structured diagnostic results for highway defects, and to perform quality control and workflow updates based on the diagnostic results and execution status, so as to achieve continuous optimization of diagnostic tasks.

[0022] In this embodiment, the modules are interconnected using the following method: Collect highway design data, construction data, testing data, monitoring data, maintenance data, as well as highway distress-related standards and literature. Clean, label, and associate the data to generate a highway distress knowledge graph dataset and a highway distress digital twin dataset, which will serve as the basic data source for subsequent data query and reasoning analysis. Based on the highway disease knowledge graph dataset and the highway disease digital twin dataset, the query requirements in the highway disease diagnosis workflow are transformed into structured retrieval requests through the RAG retrieval model. Preliminary data retrieval is performed to generate the first query result dataset, providing a data foundation for subsequent natural language queries and reasoning. The system obtains the user's input natural language diagnostic requirements, inputs the natural language diagnostic requirements and the first query result data set into the NL2SQL natural language conversion model, generates a structured query statement and executes cross-table, multi-dimensional database queries to generate a second query result data set, and supplements and integrates the second query result data set with the first query result data set to form the complete data set required for diagnosis. The complete dataset is input into the large language model, and a description text of the highway disease diagnosis workflow is generated based on the complete dataset, which includes data collection, data processing, disease identification, disease analysis, diagnosis and assessment, report generation and quality control. Furthermore, a multi-agent task list and corresponding task constraint information are generated, so that subsequent multi-agent collaborative execution steps can be processed based on a unified task structure and data basis. The complete dataset is input into the improved fuzzy cognitive graph. By modeling disease-related factors, constructing causal weights, and updating the time-series state, a disease state vector sequence and a disease causal relationship graph are generated. These two are used as the reasoning basis for multi-agents to perform disease identification and disease analysis tasks. Based on the multi-agent task list, the disease causal relationship graph, and the disease state vector sequence, the multi-agents collaboratively execute data processing, disease identification, and disease analysis tasks. The agent scheduler dynamically adjusts the task execution order according to the task status, generating a set of multi-agent diagnostic intermediate results. The multi-agent diagnostic intermediate result set, the disease causal relationship graph and the disease state vector sequence are input into the time-series causal discovery algorithm, the dynamic Bayesian risk assessment algorithm and the trend prediction algorithm based on echo state network. The disease inference algorithm, disease assessment algorithm and disease development trend analysis algorithm are executed to generate a set of highway disease diagnostic results containing disease type, location, development trend and risk level. The set of highway defect diagnosis results and the description text of the highway defect diagnosis workflow are input into the large language model to generate a natural language report of highway defect diagnosis and a structured diagnosis result output. Quality control is performed based on the diagnosis results and execution status, and the task list and workflow description text are updated to achieve continuous optimization of the diagnosis task and complete intelligent diagnosis of highway defects.

[0023] In this embodiment, the generation of the highway defect knowledge graph dataset and the highway defect digital twin dataset includes: Collect highway design data, construction data, testing data, monitoring data and maintenance data, and compile relevant standards and literature on road defects into a raw data set; Based on the data type, the original dataset is subjected to field extraction, feature extraction, and time-series alignment to generate a structured dataset containing design fields, construction fields, inspection fields, monitoring fields, and maintenance fields; Based on the field content in the structured data set, road entities, structural entities, defect entities, monitoring entities, and maintenance entities are generated to form an entity set; Based on the field associations between entity sets, structural relationships, temporal relationships, and disease-related relationships are generated to form a data relationship set; Based on the correspondence between the entity set and the data relationship set, a highway disease knowledge graph data set containing entity nodes, relationship edges and rule content is constructed; The monitoring, detection, and maintenance fields are extracted from the structured dataset and aligned with the temporal relationships in the data relationship set to form the input content for digital twin modeling. By mapping the state and updating the time sequence of the input content for digital twin modeling, a set of digital twin data for highway defects is generated to describe the defect process.

[0024] In this embodiment, the generation of the first query result data set includes: The query requirements in the highway disease diagnosis workflow are parsed into structured query requirements that include target fields, constraint fields, and related fields. The structured query requirements, together with the highway disease knowledge graph dataset and the highway disease digital twin dataset, are used as the retrieval input. Based on the search input, semantic encoding, vector representation and similarity matching are performed through the RAG search model, and a structured search request for knowledge graph search and digital twin data search is constructed based on the matching results; The specific package library for constructing the structured retrieval request is as follows: Based on the structured query requirements, the entity field matching parameters, attribute field matching parameters, and relation field matching parameters for knowledge graph retrieval are determined; the monitoring field association parameters, detection field association parameters, and maintenance field association parameters for digital twin data retrieval are determined; and a set of retrieval control parameters for performing knowledge graph data matching and digital twin data matching is formed based on the matching parameters and association parameters. Based on the structured retrieval request, a preliminary data retrieval is performed using the RAG retrieval model. The entity fields, attribute fields, and relation fields of the highway disease knowledge graph dataset are matched and calculated to generate preliminary knowledge graph retrieval results. Based on the structured search request, a preliminary data search is performed using the RAG search model. Cross-table and multi-dimensional matching calculations are performed on the monitoring, detection, and maintenance fields of the digital twin data set of highway defects to generate preliminary search results for digital twin data. The preliminary search results from the knowledge graph and the preliminary search results from the digital twin data are aligned by fields and integrated by content to form the first query result data set.

[0025] In this embodiment, the generation of the complete data set includes: The system obtains the user's input natural language diagnostic requirements, performs word segmentation, semantic intent recognition, and field association processing on the natural language diagnostic requirements, and forms natural language query semantic content that includes semantic intent fields, constraint fields, and target fields. The semantic content of a natural language query is matched with a set of query results data through field matching and semantic completion to form query mapping content for converting natural language into structured queries; The query mapping content is input into the NL2SQL natural language conversion model, and a structured query statement generation step including semantic parsing, field mapping inference and multi-table join inference is performed to obtain the structured query statement; The generation of the structured query statement specifically includes: extracting target field information for constructing the query statement based on the query mapping content, selecting data table information for executing the query based on the field correspondence, determining the association condition information for connecting data tables based on the cross-table association requirements, and combining the target field information, data table information and association condition information to form an executable structured query statement. Execute cross-table, multi-dimensional database queries based on structured query statements to form a second query result data set; The field content in the second query result data set is compared with the field content in the first query result data set to complete and ensure consistency, forming the complete data set required for diagnosis.

[0026] In this embodiment, the generation of the highway defect diagnosis workflow description text, the multi-agent task list, and the task constraint information includes: The complete dataset is input into a large language model, and semantic reasoning is used to generate a description text of the highway disease diagnosis workflow, which includes data collection, data processing, disease identification, disease analysis, diagnosis and assessment, report generation, and quality control. The structured workflow text content is generated according to the workflow execution order. Based on the description text of the highway disease diagnosis workflow, the task requirements of each workflow link are converted into corresponding task execution items through a large language model, forming a multi-agent task list arranged in the execution order. The generation of the multi-agent task list specifically includes: extracting step information for task division based on the process step field in the description text of highway disease diagnosis workflow; determining the execution order relationship between tasks based on the field dependency field; determining the serial-parallel relationship of tasks based on the process order field; and forming a multi-agent task list containing task identifiers, corresponding process steps, and task execution order based on the step information, dependency relationship, and serial-parallel relationship. Based on the multi-agent task list, task constraint information is generated through a large language model, which describes the input field requirements, output field requirements, dependency field requirements, and task order requirements for each task execution item, forming task constraint information to guide task execution. The generation of the task constraint information specifically includes: determining the input field information required for each task based on the task fields in the multi-agent task list, determining the execution trigger condition information of the task based on the task dependency fields, determining the result usage condition information of the task based on the input and output fields, and forming task constraint information for constraining the task execution process based on the input field information required for the task, the execution trigger condition information and the result usage condition information.

[0027] In this embodiment, the generation of the disease state vector sequence and the disease causal relationship diagram includes: The complete dataset is input into the improved fuzzy cognitive graph. The disease attribute field, disease monitoring field, and disease association field in the complete dataset are extracted in a structured manner. A set of disease factors for causal reasoning is formed by the feature division rules of disease-related factors. Based on the disease factor set, causal weights among disease factors are constructed through an improved fuzzy cognitive graph. The direction, intensity, and conditions of action between disease influencing factors are quantified. The causal weights are updated based on the association fields in the complete dataset to form an initial causal weight set for performing temporal reasoning. The mechanism for constructing causal weights among disease factors specifically includes: dynamically updating causal weights based on disease association fields in the complete dataset; introducing structured association fields from the database into the weight calculation process by combining information on the direction, intensity, and conditions of action between disease factors; enabling causal weights to be updated based on real monitoring data, their own relationships, and disease development patterns; and realizing a dynamic weight mapping mechanism for highway disease scenarios. The initial causal weight set is combined with the time-series monitoring fields in the complete dataset. The time-series state update is performed through the improved fuzzy cognitive graph. The disease factors are updated step by step according to the time-series state change rules, generating a disease state vector sequence corresponding to each time step. Based on the initial causal weight set and the disease state vector sequence, a disease causal relationship graph is formed through causal structure generation rules.

[0028] In this embodiment, the generation of the multi-agent diagnostic intermediate result set includes: Obtain the multi-agent task list and task constraint information, match the multi-agent task list with the disease causal relationship graph and disease state vector sequence, and form the multi-agent execution input content according to the data dependency conditions required for the task execution item; Based on the multi-agent execution input, each agent executes the corresponding task execution item according to the input field requirements, output field requirements, dependency field requirements and task order requirements specified in the task constraint information, and independently processes the data processing task, disease identification task and disease analysis task in the task execution item. When executing each task, each agent calls an image feature extraction algorithm based on convolutional neural networks, a disease feature analysis algorithm based on decision trees, and a disease classification algorithm based on logistic regression, depending on the type of task. The causal structure information in the disease causal relationship graph is used as the feature association constraint between disease factors, and the temporal state information in the disease state vector sequence is used as the time series feature input. The agent performs feature extraction, feature filtering, and classification judgment on the data during the task execution process, and generates the task execution result content through the feature vector, classification label, and confidence score output by the model. The task execution results of each agent are recorded. The agent scheduler dynamically adjusts the task execution order in the multi-agent task list based on the execution status information, data dependency status information and task completion status information during the task execution process, thus forming an updated scheduling result of the multi-agent execution process. The formation of the updated scheduling result specifically includes: determining the execution completion status of each task execution item based on the execution status information reported by each agent; determining whether the task execution item meets its input field requirements based on the data dependency status information reported by the task execution item during execution; determining the set of task execution items that can currently enter the execution stage based on the task order requirements specified in the task constraint information; and reordering the task execution order in the multi-agent task list according to the execution completion status, data dependency satisfaction, and task order requirements of the task execution items; in the reordering process, a new execution order is formed by sorting the execution priority values ​​of the task execution items, wherein the execution priority value is a priority value calculated by weighting the number of dependency fields, execution completion status, and data dependency satisfaction status of the task execution item. The weighting method is to sum the weighted values ​​corresponding to the number of dependency fields, execution completion status, and data dependency satisfaction status according to a pre-set weight, and form an updated scheduling result for subsequent task execution processes based on the above reordering result; Based on the updated scheduling results, each agent continues to execute the remaining tasks, and the execution results of all tasks are integrated into a multi-agent diagnostic intermediate result set according to the task order.

[0029] In this embodiment, the generation of the highway defect diagnosis result set includes: The multi-agent diagnostic intermediate result set, the disease causal relationship graph, and the disease state vector sequence are used as inference input content. Based on the inference input content, disease event fields, causal structure fields, and temporal state fields are extracted, and the three types of fields are combined to form the field set required for inference calculation. According to the calculation requirements of the disease inference algorithm, the inference calculation fields for disease type inference and disease location inference are selected from the field set required for the inference calculation. The disease type inference result and the disease location inference result are generated by performing correlation operation between the causal structure field and the disease event field. The disease type inference results and disease location inference results are combined with the time-series state fields in the disease state vector sequence. Based on the calculation requirements of the disease assessment algorithm and the disease development trend analysis algorithm, the inference calculation fields for disease risk assessment and disease trend analysis are selected from the field set required for inference calculation. The assessment calculation and trend calculation are performed to generate the disease development trend analysis results and the disease risk level assessment results. The results of inference on disease type, inference on disease location, analysis of disease development trend, and assessment of disease risk level are integrated according to the disease diagnosis rules to form a set of highway disease diagnosis results that includes disease type, disease location, disease development trend, and disease risk level.

[0030] In this embodiment, the continuous optimization of the diagnostic task includes: The highway disease diagnosis result set and the highway disease diagnosis workflow description text are simultaneously input into the large language model. The report is generated based on the disease type field, disease location field, disease development trend field, and disease risk level field in the highway disease diagnosis result set, as well as the workflow task field and workflow sequence field in the highway disease diagnosis workflow description text. Based on the input content generated from the report, the large language model generates a natural language report and structured diagnostic results output for highway disease diagnosis. The natural language report includes descriptions of disease type, disease location, disease development trend, and disease risk level. The structured diagnostic results output includes field-based diagnostic structure content suitable for system processing. The generation of natural language reports and structured diagnostic results for highway defects specifically includes: semantically encoding the defect type, defect location, defect development trend, and defect risk level fields in the report generation input, converting each field content into a semantic vector representation for text generation; constructing a report generation template structure based on workflow task fields and workflow sequence fields, and organizing the semantic vectors into paragraphs according to the template structure; subsequently, executing natural language generation rules based on the semantic vectors and template structure, including sentence generation rules, paragraph organization rules, and terminology constraint rules, to generate diagnostic natural language text containing defect type descriptions, defect location descriptions, development trend descriptions, and risk level descriptions; further, structurally mapping the semantic vector content according to field names, and generating structured diagnostic results output containing defect type, defect location, defect trend, and defect risk level fields through field correspondence; wherein the natural language report generation and structured output generation are different results of the same semantic generation process, completed synchronously through text generation rules and field mapping rules; Based on the diagnostic field content in the highway disease diagnosis result set and the result field content in the structured diagnosis result output, quality control processing is performed on the diagnosis execution status field, task completion status field and process execution deviation information to generate quality control result information. Based on the quality control results, the multi-agent task list and the highway defect diagnosis workflow description text are updated to form an updated task list and updated workflow description text. The updated task list and updated workflow description text are then used as input for the next round of highway defect diagnosis tasks, thereby achieving continuous optimization and iterative improvement of the highway defect diagnosis process and completing intelligent diagnosis of highway defects.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a relatively long section of a highway in a certain province. This route traverses complex structural types, with a wide range of historical data sources, including design, construction, inspection, and monitoring archives accumulated over many years, as well as written records from inspection and maintenance units, and a large amount of industry standards and literature. Traditional diagnostic methods require manual review across multiple systems, making timely information integration difficult. Professionals often spend considerable time determining the nature, development trend, and risk level of road defects. The fragmented diagnostic basis and incomplete reasoning chain are prominent problems, severely impacting the efficiency of highway defect management.

[0032] In practical use, this invention first collects and integrates the historical data, real-time monitoring content, and disease-related standards and literature of the route. Through unified data cleaning, structured annotation, and relational modeling, it generates a highway disease knowledge graph data set and a highway disease digital twin data set, so that the data that were originally inconsistent in format and time sequence can be logically interconnected, providing a stable data foundation for subsequent diagnosis and reasoning.

[0033] During routine inspections, staff input diagnostic requests in natural language via voice or text, such as inquiring about recent structural stress anomalies or pavement damage development. The system automatically matches the natural language request with the first query result dataset, inputs it into the NL2SQL model, automatically generates structured query statements, and executes cross-database and cross-table queries. Subsequently, the system supplements and combines the second query result dataset with the first query result dataset to form a complete dataset covering the entire process of design, construction, testing, monitoring, and maintenance.

[0034] After obtaining the complete dataset, this invention uses a large language model to generate a description text of the highway disease diagnosis workflow, breaking down the diagnosis process into multiple stages such as data collection, processing, identification, analysis, evaluation, report generation, and quality control. Furthermore, it constructs a multi-agent task list and task constraint information, enabling different agents to work collaboratively according to a unified structured task system.

[0035] Subsequently, the present invention inputs the complete dataset into the improved fuzzy cognitive graph, models the disease factors, constructs and dynamically updates the causal weights between the factors, and performs time-series state updates in conjunction with time-series monitoring content, generating a disease state vector sequence and a disease causal relationship graph. This allows the disease development law to be explicitly described in a causal structure, enabling the system to be closer to the actual situation when analyzing the essence of the disease and avoiding the subjective bias of traditional experience-based judgment.

[0036] During the collaborative execution phase, each agent performs data processing, disease identification, and disease analysis based on the task list and task constraint information. The agent scheduler in this invention monitors task status, data dependencies, and execution progress in real time, dynamically adjusting the task execution order to improve overall execution efficiency and enable a more rational combination of on-site data and historical information. During the execution of each task, the agents collaboratively process data, using the disease causal relationship graph and disease state vector sequence as the basis for reasoning to generate a set of intermediate diagnostic results.

[0037] In the inference output stage, the system extracts core fields for inference based on the intermediate result set, the causal relationship diagram of the disease, and the disease state vector sequence. It then generates a set of diagnostic results containing disease type, location, development trend, and risk level through disease inference algorithms, evaluation algorithms, and trend analysis algorithms. This result set not only contains structured fields but also has clear causal basis and time series interpretation, making the diagnosis more transparent and auditable.

[0038] Subsequently, this invention inputs the diagnostic result set and the diagnostic workflow description text into a large language model to generate a natural language diagnostic report for management and maintenance personnel. The report includes information such as disease type, location, causal characteristics, development trend, and risk level, presenting the disease status in an easy-to-understand manner. Simultaneously, it generates structured diagnostic results for automatic system archiving and subsequent scheduling.

[0039] After the entire process is completed, this invention performs quality control based on the diagnostic content, execution status, and process deviation information, and updates the task list and workflow description text, enabling the system to continuously optimize the diagnostic logic based on actual performance, thereby gradually improving diagnostic accuracy and process intelligence during continuous operation.

[0040] Through application verification on highways, this invention has demonstrated good usability and stability in multi-source data integration, natural language query, causal reasoning, multi-agent collaboration, trend prediction, and automated diagnosis. It effectively overcomes the problems of long time consumption, incoherent information, and opaque reasoning in traditional manual diagnosis processes, making the disease diagnosis process more systematic, automated, and intelligent, and providing reliable technical support for large-scale highway disease management.

[0041] Table 1 Performance Comparison between Multi-Agent Collaborative Diagnosis System and Traditional Methods

[0042] Table 1 shows a comprehensive performance comparison between the method of the present invention and traditional manual diagnosis, traditional rule-based algorithms, and single-agent diagnosis systems. As can be seen from the data, the present invention achieves significant improvements in all core indicators.

[0043] In terms of diagnostic accuracy, this invention achieves 95.6%, which is a significant improvement over traditional manual methods and traditional rule-based algorithms, and is also significantly better than single-agent systems. The improved accuracy is mainly due to the fact that this invention uses knowledge graphs, digital twin data, and improved fuzzy cognitive graphs simultaneously, making the diagnostic basis more comprehensive, and the multi-agent collaboration reduces the problem of omissions in single-model reasoning.

[0044] In terms of diagnosis time, this invention only takes 9.1 minutes, which is far less than the 42 minutes of manual diagnosis, the 27.5 minutes of rule-based algorithms, and the 18.4 minutes of single-agent diagnosis. The reason for the efficiency improvement is that all tasks can be driven by an automatic workflow, and multiple agents can process disease identification, analysis and reasoning in parallel, which greatly reduces the time overhead caused by manual intervention and serial calculation.

[0045] In terms of stability in complex scenarios, this invention achieves 93.2%, significantly higher than the other three methods. The improved fuzzy cognitive graph can continuously update the temporal state of the disease, enabling the system to remain stable even under conditions of multiple disease types overlapping, noise interference, or data loss.

[0046] In terms of cross-database query success rate, this invention achieves 97.8%, a significant improvement compared to traditional methods. Thanks to the combined use of the RAG retrieval model and the NL2SQL model, the system more accurately translates workflow and natural language requirements, thus ensuring effective retrieval of cross-table and multi-source data.

[0047] In terms of disease trend prediction accuracy, this invention achieves 94.5%, representing a significant improvement. The intermediate results of multi-agent diagnosis, combined with disease causal relationship graphs and temporal state vector sequences, make the reasoning process more consistent with disease development patterns, thereby enhancing the reliability of trend and risk prediction.

[0048] In summary, this invention, through a combined mechanism of "knowledge and data enhancement + multi-agent collaboration + improved fuzzy cognitive graph + intelligent workflow", outperforms existing methods in terms of accuracy, efficiency, stability and usability, and the performance improvement is supported by clear data.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A highway defect intelligent diagnosis system based on multi-agent collaboration, characterized in that, include: The data construction module is used to collect and preprocess highway data and information related to road defects; The preliminary retrieval module transforms the query requirements in the highway disease diagnosis workflow into structured retrieval requests, and performs preliminary data retrieval based on knowledge graphs and digital twin data to generate the first query result data set. The structured query generation module is used to generate a second query result data set and integrate it with the first query result data set to form a complete data set; The workflow generation module takes the complete dataset and inputs it into the large language model to generate a description text of the highway disease diagnosis workflow, a multi-agent task list, and task constraint information. The causal reasoning module takes the complete dataset as input into the improved fuzzy cognitive graph and generates a disease state vector sequence and a disease causal relationship graph. The multi-agent execution module is used to perform data processing, disease identification, and disease analysis, and generate a set of intermediate results for multi-agent diagnosis. The diagnostic reasoning module is used to perform disease reasoning, assessment and trend analysis, and generate a set of highway disease diagnosis results. The diagnostic output and optimization module is used to generate natural language reports and structured diagnostic results for highway defects, and to perform quality control and workflow updates based on the diagnostic results and execution status, so as to achieve continuous optimization of diagnostic tasks.

2. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 1, characterized in that, The modules are connected in the following way: Collect and process data, as well as standards and literature related to road defects, perform preprocessing and correlation modeling, and generate a knowledge graph dataset of road defects and a digital twin dataset of road defects. The query requirements in the highway disease diagnosis workflow are transformed into structured search requests, a preliminary data retrieval is performed, and the first query result data set is generated. The NL2SQL natural language conversion model generates a second query result data set based on the natural language diagnostic needs and the first query result data set, and integrates it with the first query result data set to form a complete data set. The complete dataset is input into the large language model to generate a description text of the highway disease diagnosis workflow, a multi-agent task list, and task constraint information. The complete dataset is input into the improved fuzzy cognitive graph, and through factor modeling, causal weight construction and temporal state update, a disease state vector sequence and a disease causal relationship graph are generated. Based on the multi-agent task list, the disease causal relationship graph, and the disease state vector sequence, the multi-agents collaboratively perform data processing, disease identification, and disease analysis to generate a set of intermediate diagnostic results for the multi-agents. The agent scheduler then dynamically adjusts the task execution order based on task constraint information. The multi-agent diagnostic intermediate result set, the disease causal relationship graph and the disease state vector sequence are input into the time-series causal discovery algorithm, the dynamic Bayesian risk assessment algorithm and the trend prediction algorithm based on echo state network. Disease reasoning, disease assessment and disease development trend analysis are executed in sequence to generate a set of highway disease diagnosis results. Based on the highway defect diagnosis result set and the highway defect diagnosis workflow description text, natural language reports and structured diagnosis results are generated for highway defect diagnosis. Quality control and workflow updates are performed to achieve continuous optimization of the diagnosis task.

3. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the highway defect knowledge graph dataset and the highway defect digital twin dataset includes: Collect highway design data, construction data, testing data, monitoring data and maintenance data, and compile relevant standards and literature on road defects into a raw data set; Based on the data type, the original dataset is subjected to field extraction, feature extraction, and time-series alignment to generate a structured dataset containing design fields, construction fields, inspection fields, monitoring fields, and maintenance fields; Based on the field content in the structured data set, road entities, structural entities, defect entities, monitoring entities, and maintenance entities are generated to form an entity set; Based on the field associations between entity sets, structural relationships, temporal relationships, and disease-related relationships are generated to form a data relationship set; Based on the correspondence between the entity set and the data relationship set, a highway disease knowledge graph data set containing entity nodes, relationship edges and rule content is constructed; The monitoring, detection, and maintenance fields are extracted from the structured dataset and aligned with the temporal relationships in the data relationship set to form the input content for digital twin modeling. The digital twin modeling inputs are then mapped to a state and updated over time to generate a digital twin dataset of highway defects.

4. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the first query result data set includes: The query requirements in the highway disease diagnosis workflow are parsed into structured query requirements that include target fields, constraint fields, and related fields. The structured query requirements, together with the highway disease knowledge graph dataset and the highway disease digital twin dataset, are used as the retrieval input. Based on the search input, semantic encoding, vector representation and similarity matching are performed through the RAG search model, and a structured search request for knowledge graph search and digital twin data search is constructed based on the matching results; Based on the structured retrieval request, a preliminary data retrieval is performed using the RAG retrieval model. The entity fields, attribute fields, and relation fields of the highway disease knowledge graph dataset are matched and calculated to generate preliminary knowledge graph retrieval results. Based on the structured search request, a preliminary data search is performed using the RAG search model. Cross-table and multi-dimensional matching calculations are performed on the monitoring, detection, and maintenance fields of the digital twin data set of highway defects to generate preliminary search results for digital twin data. The preliminary search results from the knowledge graph and the preliminary search results from the digital twin data are aligned by fields and integrated by content to form the first query result data set.

5. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the complete dataset includes: The system obtains the user's input natural language diagnostic requirements, performs word segmentation, semantic intent recognition, and field association processing on the natural language diagnostic requirements, and forms natural language query semantic content that includes semantic intent fields, constraint fields, and target fields. The semantic content of a natural language query is matched with a set of query results data through field matching and semantic completion to form query mapping content for converting natural language into structured queries; The query mapping content is input into the NL2SQL natural language conversion model, and a structured query statement generation step including semantic parsing, field mapping inference and multi-table join inference is performed to obtain the structured query statement; Execute cross-table, multi-dimensional database queries based on structured query statements to form a second query result data set; The field content in the second query result data set is compared with the field content in the first query result data set to complete and ensure consistency, forming the complete data set required for diagnosis.

6. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the highway defect diagnosis workflow description text, multi-agent task list, and task constraint information includes: The complete dataset is input into a large language model, and a description text of the highway disease diagnosis workflow is generated through semantic reasoning. The structured workflow text content is also generated according to the workflow execution order. Based on the description text of the highway disease diagnosis workflow, the task requirements of each workflow link are converted into corresponding task execution items through a large language model, forming a multi-agent task list arranged in the execution order. Based on the multi-agent task list, task constraint information is generated through a large language model, which describes the input field requirements, output field requirements, dependency field requirements, and task order requirements for each task execution item, thus forming task constraint information.

7. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the disease state vector sequence and the disease causal relationship diagram includes: The complete dataset is input into the improved fuzzy cognitive graph. The disease attribute field, disease monitoring field, and disease association field in the complete dataset are extracted in a structured manner. A disease factor set is formed by the feature division rules of disease-related factors. Based on the disease factor set, causal weights among disease factors are constructed through an improved fuzzy cognitive graph. The direction, intensity, and conditions of action between disease influencing factors are quantified. The causal weights are updated based on the correlation fields in the complete dataset to form an initial causal weight set. The initial causal weight set is combined with the time-series monitoring fields in the complete dataset. The time-series state update is performed through the improved fuzzy cognitive graph. The disease factors are updated step by step according to the time-series state change rules, generating a disease state vector sequence corresponding to each time step. Based on the initial causal weight set and the disease state vector sequence, a disease causal relationship graph is formed through causal structure generation rules.

8. The intelligent diagnosis system for highway defects based on multi-agent collaboration according to claim 2, characterized in that, The generation of the multi-agent diagnostic intermediate result set includes: Obtain the multi-agent task list and task constraint information, and match the input field requirements and dependency field requirements in the task constraint information with the disease causal relationship diagram and disease state vector sequence to form the multi-agent execution input content that meets the task input conditions. Based on the multi-agent execution input content, each agent selects the data fields required for task execution according to the input field requirements in the task constraint information, verifies the preconditions for task execution according to the dependency field requirements, determines the output structure of the task execution item according to the output field requirements, determines its actual execution order according to the task order requirements, and executes the corresponding data processing, disease identification, or disease analysis tasks. When executing each task, each agent calls an image feature extraction algorithm based on convolutional neural networks, a disease feature analysis algorithm based on decision trees, and a disease classification algorithm based on logistic regression, depending on the type of task. The causal structure information in the disease causal relationship graph is used as the feature association constraint between disease factors, and the time-series state information in the disease state vector sequence is used as the time-series feature input. The agent performs feature extraction, feature filtering, and classification judgment on the data during the task execution process. Through the feature vector, classification label, and confidence score output by the model, the agent generates the task execution result content that meets the requirements of the output field in the task constraint information. The task execution results are reported to the agent scheduler. The agent scheduler verifies the execution status, data dependency satisfaction status, and output field integrity of the task execution items according to the task order requirements, dependency field requirements, and execution trigger conditions specified in the task constraint information. Based on this, the task execution order in the multi-agent task list is adjusted to generate an updated scheduling result that meets the requirements of the task constraint information. Based on the updated scheduling results, each agent continues to execute the remaining tasks. The execution results of all tasks are then integrated into a multi-agent diagnostic intermediate result set according to the output field requirements in the task constraint information.

9. A highway defect intelligent diagnosis system based on multi-agent collaboration according to claim 2, characterized in that, The generation of the highway defect diagnosis result set includes: The multi-agent diagnostic intermediate result set, the disease causal relationship graph, and the disease state vector sequence are used as inference input content. Based on the inference input content, disease event fields, causal structure fields, and temporal state fields are extracted, and the three types of fields are combined to form the field set required for inference calculation. According to the computational requirements of the temporal causal discovery algorithm, the inference calculation fields for disease type inference and disease location inference are selected from the field set required for the inference calculation. The disease type inference result and the disease location inference result are generated by performing correlation operations between the causal structure field and the disease event field. The disease type inference results and disease location inference results are combined with the time-series state fields in the disease state vector sequence. Based on the calculation requirements of the dynamic Bayesian risk assessment algorithm and the trend prediction algorithm based on echo state network, the inference calculation fields for disease risk assessment and disease trend analysis are selected from the field set required for inference calculation. The assessment calculation and trend calculation are performed to generate the disease development trend analysis results and disease risk level assessment results. The results of disease type inference, disease location inference, disease development trend analysis, and disease risk level assessment are integrated according to the disease diagnosis rules to form a set of highway disease diagnosis results.

10. A highway defect intelligent diagnosis system based on multi-agent collaboration according to claim 2, characterized in that, The continuous optimization of the diagnostic task includes: The highway disease diagnosis result set and the highway disease diagnosis workflow description text are simultaneously input into the large language model. The report is generated based on the disease type field, disease location field, disease development trend field, and disease risk level field in the highway disease diagnosis result set, as well as the workflow task field and workflow sequence field in the highway disease diagnosis workflow description text. Based on the report, the input content is generated, and the large language model generates a natural language report on highway disease diagnosis and structured diagnosis results as output. Based on the diagnostic field content in the highway disease diagnosis result set and the result field content in the structured diagnosis result output, quality control processing is performed on the diagnosis execution status field, task completion status field and process execution deviation information to generate quality control result information. Based on the quality control results, the multi-agent task list and the highway defect diagnosis workflow description text are updated to form an updated task list and updated workflow description text. The updated task list and updated workflow description text are then used as input for the next round of highway defect diagnosis tasks, thereby achieving continuous optimization and iterative improvement of the highway defect diagnosis process and completing intelligent diagnosis of highway defects.

Images