Construction method of expressway maintenance knowledge graph

Abstract

The invention discloses a construction method of an expressway maintenance knowledge graph, relates to the technical field of graph construction, and solves the technical problems that most knowledge exists in a scattered and disordered form and lacks effective integration and utilization, and the requirements for rapid query, accurate application and intelligent decision support of expressway maintenance knowledge are difficult to meet. According to the method, multiple entity extraction and relation determination methods are adopted, entities and relations can be more comprehensively and accurately recognized from text data, richer, more detailed and logical knowledge extraction results are generated through extraction of entity attributes and secondary classification based on the attributes, and for a reliability evaluation mechanism of newly-added knowledge, the reliability of the newly-added knowledge is improved. Comprehensive judgment is performed from multiple aspects of source, content integrity and timeliness, and complementary fusion with old knowledge is considered when newly added knowledge is processed, so that the knowledge system can be ensured to always keep advancement, integrity and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of graph construction, and specifically to a method for constructing a highway maintenance knowledge graph. Background Art

[0002] Highway maintenance is a complex and systematic project involving numerous factors and data. Traditional maintenance decisions often rely on experience and manual judgment, making it difficult to comprehensively and accurately assess maintenance needs and effectiveness. A knowledge-graph-based highway maintenance model can integrate and analyze vast amounts of maintenance data, engineering knowledge, and expert experience to form a knowledge base and model library for maintenance decision-making.

[0003] According to publication number CN112347265B, a method for constructing a knowledge graph is provided, comprising: performing word segmentation on a sentence to be processed to obtain multiple individual words; identifying entities in the multiple individual words to obtain entity pairs in groups of two; obtaining a sentence vector for each entity pair; extracting representation features of the sentence based on the sentence vector; performing feature screening on the representation features to shield noise features therein; predicting entity relationships of the entity pairs based on the representation features; and constructing a knowledge graph based on the entity pairs and corresponding entity relationships. The present invention can effectively filter out useless noise features in sentences during the process of constructing a knowledge graph, thereby improving the accuracy of entity relationship prediction.

[0004] However, when some existing knowledge graphs are constructed, knowledge mostly exists in a scattered and disordered form, lacking effective integration and utilization. Traditional knowledge management methods are unable to meet the needs of rapid query, accurate application and intelligent decision-making support for highway maintenance knowledge. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a method for constructing a highway maintenance knowledge graph, which solves the problem that most knowledge exists in a scattered and disordered form, lacks effective integration and utilization, and is difficult to meet the needs of rapid query, accurate application and intelligent decision-making support for highway maintenance knowledge.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for constructing a highway maintenance knowledge graph, the method specifically comprising the following steps: Step S1: Acquire highway maintenance information and pre-process the acquired maintenance information to obtain pre-processed data; Step S2: extract knowledge from the obtained pre-processed data, generate knowledge extraction results by processing the text data and analyzing the relationship between entities in combination with historical data; Step S3: Analyze the obtained knowledge extraction results, fuse the relationships corresponding to the knowledge extraction results, and obtain the fused knowledge results based on similarity analysis; Step S4: Select a suitable atlas model based on the obtained fusion results, and use domain adaptation technology to adjust the model parameters and structure. At the same time, optimize the atlas model in combination with the expert database to generate an optimized model.

[0007] As a further solution of the present invention, the specific method of pre-processing the maintenance information in step S1 is: Data cleaning: Clean the collected data to remove noise, duplication and errors; Data classification: classify data according to its type and source; Data annotation: Labeling of text data that requires machine learning or deep learning.

[0008] As a further solution of the present invention, the specific method of extracting knowledge from the preprocessed data in step S2 is: Obtain preprocessed data and text data from the preprocessed data, identify entities in the text data, and then label all acquired entities as i, where i = 1, 2, ..., j, where j represents the number of entities, and mark entities with relationships as associated entities; Then the acquired related entities are analyzed, and the entity attributes corresponding to the associated entities are obtained. The associated entities are secondary classified according to the entity attributes, and the associated entities with the same entity attributes are classified to obtain related entities with the same attributes, and the knowledge extraction results are generated.

[0009] As a further solution of the present invention, the entities in step S2 include maintenance technology, disease entities, maintenance equipment and materials, and the method for identifying the entities is a deep learning algorithm; At the same time, entity attributes include the applicable scope of maintenance technology, the characteristics of diseases, and the performance parameters of equipment, and entity attributes are obtained through text parsing.

[0010] As a further solution of the present invention, the specific method of analyzing the knowledge extraction result in step S3 is: Obtain all knowledge extraction results and determine whether there is new knowledge in the knowledge extraction results. If there is new knowledge, mark the corresponding extraction results as extraction results to be analyzed. Otherwise, if there is no new knowledge, do not process the extraction results. Obtain the marked extraction results to be analyzed, and obtain any group of extraction results to be analyzed as the analysis object, and at the same time obtain the new knowledge corresponding to the analysis object, and then analyze the reliability of the new knowledge, and the reliability analysis includes the source of the new knowledge, the content integrity of the new knowledge and the timeliness of the new knowledge. Based on the reliability analysis, determine whether to retain the new knowledge. If it needs to be retained, the new knowledge will replace the knowledge extraction result. If it does not need to be retained, the new knowledge will be eliminated. After analyzing and sorting the new knowledge, the processing results are obtained, and the processing results are processed and analyzed.

[0011] As a further solution of the present invention, the specific method of judging based on reliability analysis in step S3 is: If the source of new knowledge, the content integrity of new knowledge and the timeliness of new knowledge are all normal, it means that the new knowledge needs to be retained. Conversely, if the source of new knowledge, the content integrity of new knowledge and the timeliness of new knowledge are all normal and any one of them has an abnormality, it will not be processed.

[0012] As a further solution of the present invention, the specific method of analyzing the processing results in step S3 is: The processing results are obtained and labeled as n, and n=1, 2, ..., m, where m represents the number of processing results. Then, any group of processing results is obtained as the target object, and the similarity value between the target object and the remaining processing results is calculated. The remaining processing results corresponding to the similarity value greater than the threshold are fused to obtain the fusion result, and the specific value of the threshold is set by the operator.

[0013] As a further solution of the present invention, the specific method of generating the optimization model in step S4 is: Select an open-source pre-trained language model based on the Transformer architecture, and then use domain adaptation techniques to adjust the model parameters and structure; First, we conduct expert interviews and questionnaire surveys to collect feedback on the model's performance and suggestions for improvement. We organize the collected expert knowledge and feedback into structured data, build an expert knowledge base, and then construct a training dataset based on the expert knowledge base that includes the feedback and suggestions. Based on the collected real-time feedback, the model is further trained using a reinforcement learning algorithm to optimize the output results and obtain an optimized model.

[0014] The present invention provides a method for constructing a highway maintenance knowledge graph. Compared with existing technologies, it has the following advantages: By employing multiple entity extraction and relationship determination methods, this invention can more comprehensively and accurately identify entities and relationships from text data. Furthermore, by extracting entity attributes and performing attribute-based secondary classification, it generates richer, more detailed, and logical knowledge extraction results. Compared with traditional manual knowledge organization methods, this method greatly improves the efficiency and accuracy of knowledge extraction and can uncover more complex relationships and knowledge details hidden in the text. The reliability assessment mechanism for new knowledge comprehensively judges its source, content integrity, and timeliness, and considers its complementary integration with existing knowledge when processing new knowledge, ensuring that the knowledge system always remains advanced, complete, and accurate. Traditional technologies often lack a systematic assessment and effective integration mechanism for new knowledge, which can easily lead to delayed knowledge updates or a chaotic knowledge system. By building an open-source pre-trained language model based on the Transformer architecture and optimizing it using domain adaptation techniques, combined with an expert database and reinforcement learning algorithms, the resulting optimized model can better understand and process instructions and tasks in the highway maintenance field. Its performance is continuously improved through multiple rounds of feedback and training. Compared to traditional fixed models or models lacking domain-specific optimization, this model can provide more intelligent and accurate decision support and efficient knowledge query services for highway maintenance, significantly improving the efficiency and quality of maintenance work and adapting to the ever-evolving and changing needs of the highway maintenance field. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a diagram of the steps of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] For example 1, please refer to Figure 1 , this application provides a method for constructing a highway maintenance knowledge graph, which specifically includes the following steps: Step S1: Acquire highway maintenance information, including maintenance documents and maintenance records. Preprocess the acquired maintenance information to obtain preprocessed data. The preprocessing includes data cleaning, classification, and labeling. Natural language processing (NLP) tools and regular expressions are used to clean the data.

[0018] Specifically, data cleaning: clean the collected data to remove noise, duplication, and errors. For example, maintenance records may contain some erroneous data, such as incorrect records of disease locations and irregular repair date formats. Data classification: Classify data according to its type (such as maintenance technology, disease type, maintenance equipment, etc.) and source. For example, all data on bridge maintenance technology can be classified into one category, and data on pavement disease types can be classified into another category. Data annotation: For some text data that requires machine learning or deep learning, annotation work is performed, for example, marking the names of diseases mentioned in the text, the disease categories corresponding to maintenance measures, etc.

[0019] Step S2: Perform knowledge extraction on the obtained pre-processed data, generate knowledge extraction results by processing the text data and analyzing the relationship between entities in combination with historical data.

[0020] Obtain preprocessed data, and at the same time obtain text data in the preprocessed data, and identify entities in the text data. The entities here include maintenance techniques (micro-surfacing, grouting, etc.), disease entities (cracks, potholes, etc.), maintenance equipment (sweepers, rollers, etc.) and materials (asphalt, cement, etc.). For example, rule-based methods, machine learning methods (such as support vector machines) or deep learning methods (such as Bi-LSTM+CRF) can be used for entity extraction. Taking a maintenance record text as an example, "Micro-surfacing technology is used to treat road rutting on XX road section", the maintenance technology entity "micro-surfacing" and the disease entity "rutting" can be identified through entity extraction, and then all the acquired entities are labeled. Denoted as i, and i = 1, 2, ..., j, where j represents the number of entities. At the same time, the relationship between entities is determined based on the naive Bayes classifier, and entities with existing relationships are marked as associated entities. Here, associated entities represent the existence of two or more entities, such as the relationship between maintenance technology and disease (a certain maintenance technology is used to treat a certain disease) and the relationship between equipment and maintenance tasks (a certain equipment is used to perform a certain maintenance task). For example, when determining the relationship between the two entities "road roller" and "road compaction", relevant features are first extracted (such as verbs and descriptive words related to the two in the text), and then the probability of belonging to the "equipment and maintenance task" relationship category is calculated based on the trained naive Bayes model; The obtained associated entities are then analyzed, and their corresponding entity attributes are extracted. These entity attributes represent, for example, the applicable scope of the maintenance technology, the characteristics of the defect (length, width, depth, etc.), and the performance parameters of the equipment. Specific attribute extraction can be achieved through methods such as text parsing and regular expression matching. For example, for the maintenance technology entity "micro-surfacing," attributes such as "applicable to early-stage asphalt pavement defect prevention" and "thickness generally between 5-10mm" can be extracted. Based on these entity attributes, the associated entities are then secondary classified. Specifically, secondary classification involves grouping associated entities with the same entity attributes to generate related entities with the same attributes, and generating knowledge extraction results. For example, among the numerous associated entities, all maintenance technology entities applicable to cement pavement maintenance with a thickness requirement of 3-5mm are grouped together; and all bridge crack defect entities with a length of 1-3 meters, a width of 0.1-0.3 meters, and located on bridge piers are grouped together.

[0021] Step S3: Analyze the obtained knowledge extraction results, fuse the relationships corresponding to the knowledge extraction results, and obtain the fused knowledge results based on similarity analysis.

[0022] Obtain all knowledge extraction results and determine whether there is new knowledge in the knowledge extraction results. The new knowledge here is represented as knowledge that is different from the current knowledge extraction results. If there is new knowledge, the corresponding extraction result will be marked as the extraction result to be analyzed. Otherwise, if there is no new knowledge, the extraction result will not be processed. For example, in the highway maintenance knowledge extraction results, if the current knowledge system only has traditional repair technology for a certain disease, and a new biomaterial-based repair technology appears in the extraction results, then this extraction result has new knowledge.

[0023] Obtain the marked extraction results to be analyzed, and take any group of extraction results to be analyzed as the analysis object, and at the same time obtain the new knowledge corresponding to the analysis object, and then analyze the reliability of the new knowledge. The reliability analysis includes the source of the new knowledge, the content integrity of the new knowledge, and the timeliness of the new knowledge. For example, in evaluating the source of the new knowledge, it is carefully examined whether its source comes from an authoritative organization, a professional scientific research team, or experienced industry experts. For example, if the new knowledge is about the performance introduction of a new type of highway pavement maintenance material, if its source is a research report obtained through a large number of experiments by a well-known transportation research institute, then the reliability of its source is relatively high; on the contrary, if it is just a personal sharing on an unknown website, the reliability is greatly reduced; Regarding the analysis of content completeness, taking a new bridge defect detection technology as an example, complete knowledge content should include detection principles, scope of application, detection equipment requirements, operating procedures, and data interpretation. If the new knowledge only mentions the detection principles and lacks other important parts, the content is not complete enough. Regarding the analysis of timeliness, for example, in the current context of advocating green and environmentally friendly transportation infrastructure maintenance, if the newly added knowledge is about a highly polluting but reasonably efficient maintenance method, then its timeliness is poor; Based on the reliability analysis, it is judged whether to retain the newly added knowledge. If it needs to be retained, the source of the newly added knowledge, the content integrity of the newly added knowledge and the timeliness of the newly added knowledge are all normal, then the new knowledge will replace the knowledge extraction result. If it does not need to be retained, the source of the newly added knowledge, the content integrity of the newly added knowledge and the timeliness of the newly added knowledge are all normal. If there is an abnormality in any of the groups, the new knowledge will be eliminated. When eliminating, the content in the newly added knowledge that is complementary to the current knowledge can be obtained. For example, the old knowledge contains rich operating experience of a certain maintenance equipment, and the new knowledge provides a new maintenance technology for the equipment. The two can be combined to form a more comprehensive knowledge system about the equipment. After analyzing and sorting the new knowledge, the processing results are obtained, and the processing results here include those that have not been The processed knowledge extraction results and the knowledge after replacement by the new knowledge are labeled as n at the same time, and n=1, 2, ..., m, where m represents the number of processing results. Then, any group of processing results is obtained as the target object, and the similarity value between the target object and the remaining processing results is calculated. The similarity value here is represented by cosine similarity. The cosine similarity between the target object and the processing result is calculated. At the same time, the processing results are fused according to the calculated similarity value to obtain a fusion result. The fusion result obtained here indicates that the corresponding meanings between the processing results are similar. For example, one data source mentions that "thermal regeneration technology can repair aging diseases of asphalt pavement", and another data source also has a similar statement but the wording is slightly different. These relationships need to be fused and unified.

[0024] For example, suppose we have two sets of vector representations of highway maintenance technology. Vector A represents the "micro-surfacing" technology, and its vector is (0.8, 0.6, 0.7). The three dimensions here may represent the applicability of the technology (such as the degree of applicability to different road types), cost-effectiveness (such as the ratio of cost to benefit), and environmental friendliness (such as the degree of environmental friendliness). Vector B represents the "filling" technology, and its vector is (0.6, 0.7, 0.4). First, calculate the dot product A*B=(0.8×0.6)+(0.6×0.7)+(0.7×0.4)=1.18, then calculate the modulus of vector A |A|=1.22, the modulus of vector B |A|=1.005, and further calculate the cosine similarity cos =0.96, which shows that the “micro-surfacing” technology and the “grouting” technology are similar in these three dimensions.

[0025] Step S4: Select a suitable atlas model based on the obtained fusion results, and use domain adaptation technology to adjust the model parameters and structure. At the same time, optimize the atlas model in combination with the expert database to generate an optimized model.

[0026] Select an open-source pre-trained language model based on the Transformer architecture, such as Qwen and Llama. Then, use domain adaptation technology to adjust the model parameters and structure. Specific methods include instruction fine-tuning and human feedback reinforcement learning. Specifically, instruction fine-tuning refers to fine-tuning using domain-specific instruction data, enabling the model to follow instructions in the highway maintenance field without explicit examples. Human feedback reinforcement learning refers to reinforcement learning through feedback from human experts on the model output to continuously improve the model's performance. First, conduct expert interviews and questionnaires, design a scientific and reasonable interview and questionnaire survey process, collect expert feedback on model performance and improvement suggestions, then organize the collected expert knowledge and feedback, convert them into structured data, build an expert knowledge base, and build a training dataset containing expert feedback and suggestions based on the expert knowledge base. Based on the real-time feedback collected, the model is further trained using a reinforcement learning algorithm to optimize the output results. Based on the task characteristics of the highway maintenance field, model evaluation indicators such as accuracy, recall rate, F1 value, etc. are set to comprehensively evaluate the model performance. In the process of multiple rounds of feedback and training, the performance and accuracy of the model are continuously improved.

[0027] Example 2, as Example 2 of the present invention, is implemented on the basis of Example 1, and differs from Example 1 in the following aspects: When calculating the similarity value, this embodiment uses Euclidean distance to analyze the similarity. The specific analysis method is as follows: Vector A represents the "micro-surfacing" technology, and its vector is (0.8, 0.6, 0.7). Vector B represents the "filling" technology, and its vector is (0.6, 0.7, 0.4). First, the Euclidean distance d(A, B) = 0.37 is calculated, and then the similarity s = 1 / 0.37 ≈ 0.73 is calculated.

[0028] Embodiment 3, as the third embodiment of the present invention, focuses on combining the implementation processes of embodiment 1 and embodiment 2.

[0029] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0030] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for constructing a highway maintenance knowledge graph, characterized in that: The method specifically comprises the following steps: Step S1: obtaining highway maintenance information, and preprocessing the obtained maintenance information to obtain preprocessed data; Step S2: extract knowledge from the obtained preprocessed data, generate knowledge extraction results by processing the text data and analyzing the relationship between entities in combination with historical data; Step S3: Analyze the obtained knowledge extraction results, fuse the relationships corresponding to the knowledge extraction results, and obtain the fused knowledge results based on similarity analysis; Step S4: Select a suitable graph model based on the obtained fusion results, and use domain adaptation technology to adjust the model parameters and structure. At the same time, optimize the graph model in combination with the expert database to generate an optimized model.

2. The method for constructing a highway maintenance knowledge graph according to claim 1, characterized in that: The specific method of preprocessing the maintenance information in step S1 is: Data cleaning: Clean the collected data to remove noise, duplication and erroneous data; Data classification: classify data according to its type and source; Data labeling: Labeling of text data that requires machine learning or deep learning.

3. The method for constructing a highway maintenance knowledge graph according to claim 1, characterized in that: The specific method of extracting knowledge from the preprocessed data in step S2 is: Obtain preprocessed data, and simultaneously obtain text data in the preprocessed data, and identify entities in the text data, then label all the acquired entities as i, where i=1, 2, ..., j, where j represents the number of entities, and mark entities with relationships as associated entities; Then, the acquired related entities are analyzed, and the entity attributes corresponding to the related entities are acquired. The related entities are secondary classified according to the entity attributes, and the related entities with the same entity attributes are classified to obtain related entities with the same attributes, and the knowledge extraction results are generated.

4. The method for constructing a highway maintenance knowledge graph according to claim 3, characterized in that: The entities in step S2 include maintenance technology, disease entities, maintenance equipment and materials, and the method for identifying the entities is a deep learning method; At the same time, entity attributes include the applicable scope of maintenance technology, the characteristics of diseases, and the performance parameters of equipment, and entity attributes are obtained through text parsing.

5. The method for constructing a highway maintenance knowledge graph according to claim 1, characterized in that: The specific method of analyzing the knowledge extraction result in step S3 is: Obtain all knowledge extraction results and determine whether there is new knowledge in the knowledge extraction results. If there is new knowledge, mark the corresponding extraction results as extraction results to be analyzed. Otherwise, if there is no new knowledge, do not process the extraction results. Obtain the marked extraction results to be analyzed, and obtain any group of extraction results to be analyzed as the analysis object, and at the same time obtain the new knowledge corresponding to the analysis object, and then analyze the reliability of the new knowledge, and the reliability analysis includes the source of the new knowledge, the content integrity of the new knowledge and the timeliness of the new knowledge. Based on the reliability analysis, determine whether to retain the new knowledge. If it needs to be retained, replace the knowledge extraction result with the new knowledge. If it does not need to be retained, the new knowledge will be eliminated. After analyzing and sorting the new knowledge, the processing result is obtained, and the processing result is processed and analyzed.

6. A method for constructing a highway maintenance knowledge graph according to claim 5, characterized in that: The specific method of judging based on reliability analysis in step S3 is: If the source of the new knowledge, the content integrity of the new knowledge, and the timeliness of the new knowledge are all normal, it means that the new knowledge needs to be retained. Conversely, if the source of the new knowledge, the content integrity of the new knowledge, and the timeliness of the new knowledge are all normal and any one of them is abnormal, it will not be processed.

7. The method for constructing a highway maintenance knowledge graph according to claim 5, characterized in that: The specific method of analyzing the processing result in step S3 is: The processing results are obtained and labeled as n, where n=1, 2, ..., m, where m represents the number of processing results. Then, any group of processing results is obtained as the target object, and the similarity value between the target object and the remaining processing results is calculated. The remaining processing results corresponding to the similarity values ​​greater than the threshold are fused to obtain the fusion result, and the specific value of the threshold is set by the operator.

8. The method for constructing a highway maintenance knowledge graph according to claim 1, characterized in that: The specific method of generating the optimization model in step S4 is: Select an open-source pre-trained language model based on the Transformer architecture, and then use domain adaptation technology to adjust the model parameters and structure; First, we conduct expert interviews and questionnaire surveys to collect experts’ feedback on model performance and suggestions for improvement. We organize the collected expert knowledge and feedback and convert them into structured data to build an expert knowledge base. Based on the expert knowledge base, we build a training data set that includes expert feedback and suggestions. Based on the real-time feedback collected, the model is further trained using a reinforcement learning algorithm to optimize the output results and obtain an optimized model.

Images