Hollow slab girder bridge maintenance decision-oriented hierarchical knowledge graph construction method

Abstract

The invention discloses a hollow slab girder bridge maintenance decision-oriented hierarchical knowledge graph construction method, which comprises the following steps of: constructing a hollow slab girder bridge disease classification and maintenance decision general knowledge graph, combing disease types and evaluation indexes of a hollow slab girder bridge through cross-standard feature alignment and rule fusion, and defining each disease and the index thereof in a variable manner so as to construct a hollow slab girder bridge disease classification and maintenance decision-oriented hierarchical knowledge graph. The weight of each disease index is determined based on a fuzzy-Bayesian hybrid reasoning algorithm, and a complete general knowledge graph structure is formed; extracting disease information based on a detection report of a single bridge, endowing each disease with a maintenance suggestion by using a general knowledge graph, and forming a maintenance decision knowledge graph special for the bridge; and dynamically updating the special knowledge graph, and adjusting disease information and maintenance suggestions according to a subsequent detection report of the bridge to ensure timeliness and accuracy of the graph. According to the method, the problems that knowledge is dispersed and systematization is difficult in traditional hollow slab girder bridge maintenance decision making are solved, and scientificity and efficiency of bridge maintenance management are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge engineering maintenance decision-making, and specifically relates to a hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making, especially an intelligent decision-making technology for hollow slab girder bridge disease classification, maintenance rule quantitative modeling and dynamic knowledge graph construction. Background Art

[0002] Hollow slab girder bridges, a predominant structural form of highway bridges in my country, are prone to defects such as cracks and concrete spalling during long-term operation, requiring regular inspection and maintenance to ensure structural safety. However, existing maintenance decisions rely on manual experience and regulatory requirements, resulting in redundant and inefficient rule systems, subjective indicator weighting, insufficient dynamic adaptability, and low semantic matching accuracy. In recent years, knowledge graph technology has provided new insights into bridge maintenance decision-making, but its application in bridge engineering remains limited. Existing methods often employ general knowledge graph frameworks, failing to consider the specific characteristics of hollow slab girder bridge defects and the specialized maintenance rules. Furthermore, they lack a dynamic update mechanism for knowledge graphs tailored to the characteristics of individual bridges. Therefore, a hierarchical knowledge graph approach for hollow slab girder bridges that integrates quantitative rule screening, dynamic weight allocation, and multi-source semantic matching is urgently needed to achieve precise and intelligent maintenance decision-making. Summary of the Invention

[0003] Purpose of the invention: The present invention proposes a hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making, which realizes intelligent decision-making of hollow slab girder bridge disease classification, quantitative modeling of maintenance rules and dynamic knowledge graph construction.

[0004] Technical Solution: The method for constructing a hierarchical knowledge graph for hollow slab girder bridge maintenance decision-making according to the present invention includes the following steps:

[0005] S1. Constructing a general knowledge graph layer: Based on the defect diagnosis and treatment content of bridge maintenance specifications, we construct a classification of hollow slab girder bridge defects and maintenance rules. We use a fuzzy matrix weight allocation method, combining Bayesian reasoning and Beta distribution to quantify the importance of indicators. We also integrate the TF-IDF and BERT models to calculate the confidence level of defect information and indicators, forming a complete general knowledge graph layer.

[0006] S2. Build a dedicated knowledge graph layer: Extract the disease entity and its indicator text based on the inspection report of a single bridge j Based on the general knowledge graph, the knowledge graph reasoning and rule calculation are performed on the special knowledge graph to calculate the belief distribution of disease classification. Based on the type and severity of the damage, maintenance recommendations are assigned to each damage and maintenance measure nodes are added to the dedicated knowledge graph to form a maintenance decision-making knowledge graph dedicated to bridges.

[0007] S3. Dynamic update of special knowledge graph: construct spatiotemporal characteristic parameters of disease through component number, relative position and detection time to realize the timeliness judgment of updated data; for historical diseases, based on the disease severity index value δ' j The change of Δδ' j Evaluate development trends and output warnings of different levels; for new diseases, trigger the general knowledge graph reasoning engine to generate treatment plans, and dynamically associate the disease evolution data in the detection report with maintenance recommendations through the incremental update mechanism to achieve accurate iteration of graph information and synchronization of risk status.

[0008] Furthermore, the implementation process of step S1 is as follows:

[0009] S11. Extract IF-THEN structure rules based on bridge maintenance specifications and construct a set of rules C containing disease classification R and maintenance decision rule set M R The composite rule system R is specifically defined as R=(R F ,{R F →C R ,(R' F ,D)→M R}); R F represents the feature set used to represent disease classification, R' F Represents the feature set used to represent the severity of the disease; for rules that describe the same conclusion but have different features, the disjunction operation is used to merge them to effectively reduce rule redundancy; based on the universal Score applicablity (r) and discrimination Score discrimination (r) Screening out the optimal rule set with high universality and discrimination;

[0010] S12. Quantifying different indicators f in the classification and maintenance decision rules of typical hollow slab girder bridge diseases based on the Bayesian-Beta framework i and f j The relative importance of and mapped to triangular fuzzy number a i,j By comparing the relative importance of each index, we can get the triangular fuzzy number and build a complete fuzzy judgment matrix A=(a ij ) n×n The fuzzy judgment matrix establishes the fuzzy judgment relationship between the indicator layer and the disease classification target layer in the evaluation rules. The weight distribution is solved by the fuzzy comprehensive operator. Combined with entropy optimization, accurate weighting is achieved under multi-criteria decision-making, and a data-driven hollow slab girder bridge indicator weight system is output.

[0011] S13, integrating TF-IDF high-frequency features with BERT deep semantic representation, first calculate the rule index value δ j and real disease index text d in bridge inspection jTheir respective TF-IDF vectors and Obtain surface feature matching degree S based on cosine similarity TF-IDF (δ j ,d j ), and use the pre-trained BERT model to set the rule index value δ j and real disease index text d in bridge inspection j Encode into semantic vector and calculate deep semantic similarity S BERT (δ j ,d j ), realize the multi-dimensional collaborative matching of rules and disease characteristics; by fusing these two similarity scores, realize the multi-dimensional collaborative matching of the rule base in the general knowledge graph and the disease instance data in the special knowledge graph and the indicator confidence C rule comprehensive assessment.

[0012] Furthermore, the implementation process of step S2 is as follows:

[0013] S21, extracting disease information based on the inspection report of a single bridge, based on the weights of different indicators in the general knowledge graph ω i and the confidence index c i Weighted aggregation drives the adaptive matching and activation of rule chains to achieve knowledge reasoning and rule calculation for dedicated knowledge graphs; each disease category D k Supported by multiple related indicators, the disease classification belief distribution of each disease manifestation is calculated

[0014] S22. For each disease of a single bridge, combined with the disease classification results D k and disease severity index f' j Provide maintenance measures j , the decision result is expressed as <disease entity, recommended_measure, m j >The triple form of the dedicated knowledge graph is updated to support explainable traceability.

[0015] Furthermore, the implementation process of step S3 is as follows:

[0016] S31. Fusion of multi-period inspection data drives the iteration of a dedicated knowledge graph to build a spatiotemporal correlation-indicator coupling discrimination framework: The three parameters of component number, relative position, and inspection time are generated into a sparse vector V through TF-IDF. TF-IDF , and input the BERT model to obtain the dense semantic vector V BERT ; Calculate the spatiotemporal similarity between historical diseases and newly diagnosed diseases, and determine them as historical diseases when the similarity exceeds the threshold;

[0017] S32, for historical diseases, based on the adjacent detection cycle change value Δδ' of the disease severity index value j Set static thresholds to divide the warning into three levels; when δ' j When the preset severity threshold is exceeded, a severe warning is triggered, indicating that the disease is accelerating; when Δδ' j When the threshold is between the mild threshold, a mild warning is output, prompting the need for continuous monitoring; when Δδ' j When it is less than the mild threshold or shows negative growth, it is judged to be in a stable state; for new diseases, general knowledge graph reasoning and rule calculation are triggered to generate treatment plans; through the incremental update mechanism, the disease evolution data in the inspection report is dynamically linked with the maintenance recommendations to achieve accurate iteration of the graph information and synchronization of the risk status.

[0018] Furthermore, the implementation process of step S11 is as follows:

[0019] S11-1. Extract IF-THEN structure rules from bridge maintenance specifications and construct a set of rules containing disease classification C R and maintenance decision rule set M R The composite rule system is defined as follows:

[0020] R=(R F ,{R F →C R ,(R' F ,D)→M R})

[0021] Among them, R F ={f j =δ j |j=1,2,...,m} represents the index set f used to represent disease classification j is the indicator name, δ j is the indicator value, m is the total number of indicators; R' F ={f' j =δ' j |j=1,2,...,n} represents the feature set used to represent the severity of the disease; represents the disease classification rule set; k represents the total number of rules, usually k>|D|, J (i) Represents the index set of indicators involved in the i-th rule; represents the auxiliary maintenance decision rule set; k2 represents the total number of rules, J (v) Represents the feature index set involved in the vth rule; M={m1,m2,...,m p} represents the maintenance decision set;

[0022] S11-2. Use disjunction operation to merge to unify the description of rules, reduce redundancy and improve the simplicity of the rule base; the specific merging formula is:

[0023] R=(R F ,{R F →C Rmerged ,(R' F ,D)→M Rmerged})

[0024] For all results d n The rule index set, For all results m p The rule index set of ;

[0025] S11-3. Score for measuring the applicability of rules in multiple standards:

[0026]

[0027] Where S is the set of all specifications; δ(r∈Γ) is an indicator function that counts 1 if rule r appears in specification Γ and 0 otherwise;

[0028] Rule discrimination evaluates the specificity of the rule to the disease type:

[0029]

[0030] Among them, Conflict(r) refers to the number of conflicts of rule r, that is, the number of times the rule feature combination is the same but points to different disease types or maintenance measures; Total(r) refers to the total number of times rule r appears in all summarized rule sets; the rules with an applicability score greater than 80% and a discrimination greater than 50% are retained, and the rules with an applicability score not exceeding 80% or a discrimination not exceeding 50% need to be optimized or removed.

[0031] Furthermore, the implementation process of step S12 is as follows:

[0032] S12-1. By calculating the probability of disease occurrence P(D k |f m ) and P(D k |f n ), and introduce Beta distribution for interval reasoning to obtain the credibility interval of disease occurrence:

[0033]

[0034] Where W m,n Representation rule f m Relative to rule f nThe importance of is used to measure the difference in their impact on the disease; Indicates the index f m When the disease occurs, k The probability of occurrence, D k represents the different mechanisms of disease D, and there are K possible disease mechanisms; P(f m |D k ) is the disease D k Occurrence indicator f m Observed probability; P(D k ) is the disease D k Prior probability; interval reasoning is performed through Beta distribution to obtain the credibility interval of disease occurrence;

[0035] S12-2. Based on the expert knowledge in the field of bridge maintenance, combined with the probability ratio a i,j , a fuzzy judgment matrix is established for the key evaluation rules, and the fuzzy judgment matrix is constructed using triangular fuzzy numbers. The form is as follows:

[0036] A=[a ij ] n×n

[0037] a ij =(l ij ,m ij ,u ij )

[0038] Among them, A is the fuzzy judgment matrix, each element a i,j It is calculated by the relative importance of rules; for each fuzzy judgment matrix, calculate the possibility that the i-th rule is more important than other rules:

[0039]

[0040] After normalization, we get the weight vector W = Normalize(d(A1), d(A2), ..., d(A n )) T .

[0041] Furthermore, the implementation process of step S13 is as follows:

[0042] S13-1. First, the rule index value δ j and real disease index text d in bridge inspection j , perform word segmentation and calculate their respective TF-IDF vectors and Then, the cosine similarity is used to calculate the matching degree between the two vectors. The calculation formula is as follows:

[0043]

[0044] S13-2. In the BERT calculation stage, the pre-trained BERT model is used to convert the rule indicator f and the disease indicator text D into high-dimensional vector representations; the cosine similarity is used to calculate the matching degree between the two. The calculation formula is as follows:

[0045]

[0046] Indicator confidence C rule The calculation of is done using the hierarchical matching formula as follows:

[0047]

[0048] Based on indicator confidence, multi-dimensional collaborative matching of general knowledge graph rule base and special knowledge graph disease instance data is achieved.

[0049] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are as follows: the present invention effectively solves the problems of low efficiency, redundant rule system and strong subjectivity in traditional manual maintenance, and realizes accurate classification of bridge defects and intelligent recommendation of maintenance plans through general knowledge graph reasoning and multi-layer rule modeling, combined with expert knowledge fusion; through deep semantic matching technology, the problem of inconsistency between defect feature description and rule indicator description is solved, and the matching accuracy of maintenance decision and the utilization efficiency of detection data are significantly improved; a special knowledge graph dynamic update framework of spatiotemporal correlation-indicator coupling is innovatively constructed, which overcomes the problem of collaborative analysis of single bridge characteristics and multi-period data, and forms an adaptive maintenance plan with three-level linkage of "monitoring-maintenance-reinforcement", which realizes the full life cycle intelligent decision-making of hollow slab beam bridge defect evolution tracing and iterative optimization of maintenance strategy; the present invention provides a new idea for intelligent bridge maintenance, significantly improves the timeliness, accuracy and engineering applicability of bridge maintenance, and provides a scientific basis for accurate maintenance decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 It is a schematic diagram of a general knowledge graph model;

[0052] Figure 3 It is a dedicated knowledge graph instance graph;

[0053] Figure 4 It is a knowledge graph reasoning instance diagram for disease type identification and maintenance measure decision-making. DETAILED DESCRIPTION

[0054] The present invention is further described in detail below with reference to the accompanying drawings.

[0055] like Figure 1As shown, the present invention proposes a hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making, including the following steps:

[0056] S1. Construct a general knowledge graph layer. Based on the defect diagnosis and treatment content of bridge maintenance specifications, construct a classification and maintenance rules for hollow slab girder bridge defects. A fuzzy matrix weight allocation method is used, combining Bayesian reasoning and Beta distribution to quantify the importance of indicators. The TF-IDF and BERT models are integrated to calculate the confidence level of defect information and indicators, forming a complete general knowledge graph construction process and formula expression.

[0057] like Figure 2 As shown in the figure, the general knowledge graph is a structured knowledge system with bridge defect characteristics, maintenance rules and dynamic data as its core. By integrating standard texts, multi-source monitoring data and semantic analysis technology, a dynamic reasoning framework between defects and maintenance strategies is established.

[0058] S11. Extract IF-THEN structure rules based on bridge maintenance specifications and construct a set of rules C containing disease classification R and maintenance decision rule set M R The composite rule system R. For rules that describe the same conclusion but have different characteristics, the disjunction operation is used to merge them, effectively reducing the redundancy of rules. Based on two key indicators, universality Score applicablity (r) and discrimination Score discrimination (r) Screen out the optimal rule set with high universality and discriminability.

[0059] Extract IF-THEN structure rules from bridge maintenance specifications and construct a set of rules C containing disease classification R and maintenance decision rule set M R The composite rule system is defined as follows:

[0060] R=(R F ,{R F →C R ,(R' F ,D)→M R})

[0061] Among them, R F ={f j =δ j |j=1,2,...,m} represents the index set used to represent disease classification, f j is the indicator name, δ j is the indicator value, m is the total number of indicators; R' F ={f' j =δ' j |j=1,2,...,n} represents the feature set used to represent the severity of the disease; represents the disease classification rule set; k represents the total number of rules, usually k>|D|, J (i) Represents the index set of indicators involved in the i-th rule; represents the auxiliary maintenance decision rule set; k2 represents the total number of rules, J (v) Represents the feature index set involved in the vth rule; M={m1,m2,...,m p} represents the maintenance decision set.

[0062] For rules that describe the same disease type or maintenance decision but have different characteristics, this invention uses a disjunction (OR) operation to merge them to unify the description rules, reduce redundancy and improve the simplicity of the rule base. The specific merging formula is:

[0063] R=(R F ,{R F →C Rmerged ,(R' F ,D)→M Rmerged})

[0064] For all results d n The rule index collection. For all results m p The rule index collection.

[0065] To ensure that the rules are both universal and discriminative, the present invention proposes the following two indicators:

[0066] Measuring the universality of rules across multiple norms:

[0067]

[0068] where S is the set of all specifications; δ(r∈Γ) is an indicator function that counts 1 if rule r appears in specification Γ and 0 otherwise.

[0069] Rule discrimination: Evaluates the specificity of the rule to the disease type:

[0070]

[0071] Conflict(r) refers to the number of conflicts within rule r, i.e., the number of times rule features share the same combination but point to different disease types or maintenance measures. Total(r) refers to the total number of times rule r appears in all summarized rule sets. Rules with an applicability score >80% and a discrimination score >50% are retained; rules with low applicability or low discrimination are optimized or removed.

[0072] S12. Quantify different indicators f in the classification and maintenance decision rules of 12 typical defects of hollow slab girder bridges based on the Bayesian-Beta framework i and f j The relative importance of and mapped to triangular fuzzy number a i,j By comparing the relative importance of each index, we can get the triangular fuzzy number and build a complete fuzzy judgment matrix A=(a ij ) n×n This matrix establishes a fuzzy judgment relationship between the indicator layer and the disease classification target layer in the evaluation rules. The weight distribution is solved by using a fuzzy comprehensive operator, and combined with entropy optimization, precise weighting is achieved under multi-criteria decision-making, outputting a data-driven indicator weight system for hollow slab girder bridges.

[0073] Specifically, by calculating the probability of disease occurrence P(D k |R m ) and P(D k |R n ) and introduce the Beta distribution for interval reasoning, thereby obtaining the confidence intervals for disease occurrence. These confidence intervals reflect the range of probability of disease occurrence under given rules.

[0074]

[0075] Where W m,n Representation rule R m Relative to rule R n The importance of is used to measure the difference in their impact on the disease; Representation rule R m When the disease occurs, k The probability of occurrence, where D k represents the different mechanisms of disease D, and there are K possible disease mechanisms. m |D k ) is the disease D k When rule R occurs m Observed probability; P(D k ) is the disease D k In addition, in order to model uncertainty, interval inference is performed using the Beta distribution to obtain the confidence interval of the disease occurrence.

[0076] According to the expert knowledge in the field of bridge maintenance, the probability ratio a i,j , a fuzzy judgment matrix is established for key evaluation rules. The fuzzy judgment matrix is constructed using triangular fuzzy numbers (TFN). The core elements of the fuzzy judgment matrix are calculated by rule causal weights (RCW) and are as follows:

[0077] A=[aij ] n×n

[0078] a ij =(l ij ,m ij ,u ij )

[0079] Where: A is the fuzzy judgment matrix, each element a i,j Calculated by the relative importance of rules.

[0080] Because the influence relationships between different rules are uncertain, a triangular fuzzy number (l,m,n) is used to represent them, where l and u represent the lower and upper bounds, respectively, and m is the median. This matrix construction not only ensures the rationality of rule weight calculation but also ensures robustness and adaptability of weight allocation in the presence of fuzzy information and uncertainty.

[0081] For each fuzzy judgment matrix, calculate the probability that the i-th rule is more important than the other rules:

[0082]

[0083] After normalization, we get the weight vector W = Normalize(d(A1), d(A2), ..., d(A n )) T .

[0084] S13, integrating TF-IDF high-frequency features with BERT deep semantic representation, first calculate the rule index value δ j and real disease index text d in bridge inspection j Their respective TF-IDF vectors and Obtain surface feature matching degree S based on cosine similarity TF-IDF (δ j ,d j ), and use the pre-trained BERT model to set the rule index value δ j and real disease index text d in bridge inspection j Encode into semantic vector and calculate deep semantic similarity S BERT (δ j ,d j ) to achieve multi-dimensional coordinated matching of rules and disease characteristics.

[0085] Specifically, first extract the feature description in the rule text and construct the disease classification rule feature set R F , and the disease feature text set D extracted from the knowledge graph FThen, TF-IDF is used to calculate the similarity between the rule features and the disease feature text. Specifically, first, the rule features and the disease feature text D F , perform word segmentation and calculate their respective TF-IDF vectors and Then, the cosine similarity is used to calculate the matching degree between the two vectors. The calculation formula is as follows:

[0086]

[0087] In the BERT calculation stage, in order to calculate the semantic similarity between the rule feature and the disease feature text, the pre-trained BERT model is first used to convert the rule feature R and the disease feature text D into a high-dimensional vector representation. and Then, the cosine similarity is also used to calculate the matching degree between the two. The calculation formula is as follows:

[0088]

[0089] Finally, the rule confidence C rule The calculation adopts a hierarchical matching strategy:

[0090]

[0091] Based on this formula, multi-dimensional collaborative matching of the general knowledge graph rule base and the special knowledge graph disease instance data is achieved.

[0092] S2. Build a dedicated knowledge graph layer, extract disease information based on the inspection report of a single bridge, use the general knowledge graph to assign maintenance recommendations for each disease, and form a bridge-specific maintenance decision-making knowledge graph.

[0093] S21, extracting disease information based on the inspection report of a single bridge, based on the weights of different indicators in the general knowledge graph ω i and the confidence index c i Weighted aggregation drives adaptive matching and activation of rule chains, for each disease category D k Supported by multiple related indicators, the disease classification belief distribution of each disease manifestation is calculated like Figure 3As shown in the figure, based on the inspection report of a single bridge to be maintained, the system first extracts and integrates the disease information and constructs a dedicated knowledge graph. This graph integrates the attribute information of the disease node, including key indicators such as morphology, surrounding concrete conditions, whether there is water seepage and whitening, defective component number, reference point distance, and inspection time. Taking Bridge A as an example, the system extracts various crack types such as vertical cracks, diagonal cracks, longitudinal cracks, and transverse cracks, as well as their occurrence locations (such as webs, bottom plates, main beams, etc.). It also detects typical diseases such as concrete falling off and bridge whitening, forming a complete disease knowledge graph, which provides structured and visual decision support for the formulation of subsequent precise maintenance measures.

[0094] S22. For each disease of a single bridge, combined with the disease classification results D k and disease severity index f' j Provide maintenance measures j , the decision result is expressed as <disease entity, recommended_measure, m j > triples form to update the dedicated knowledge graph and support explainable traceability. Figure 4 As shown in the figure, the reasoning model outputs the belief distribution of the disease type based on the reasoning and rule activation matching of the knowledge graph. For each disease of a single bridge, the system generates maintenance decisions through a three-level reasoning mechanism: 1) Disease type identification reasoning: Based on the input disease feature data such as morphological parameters and location information, the disease type belief distribution is calculated through the reasoning model driven by the knowledge graph, and the probabilistic disease type inference result is output, such as Figure 4 The defect type is marked outside the box. 2) Maintenance Measure Matching Reasoning: The defect type and severity index are combined and input into the rule engine, executing a compound matching conditional "IF defect type > AND severity level = Level THEN recommended measure." 3) Knowledge Graph Dynamic Update: The final maintenance measure is updated to the dedicated knowledge graph as a triplet of <disease entity, recommended_measure, specific measure>, enabling explainable traceability of the decision-making process. Through knowledge graph reasoning, the system comprehensively considers defect type, location, and severity to provide differentiated maintenance plans for different components, achieving precise maintenance decisions.

[0095] S3. Dynamically update the dedicated knowledge graph and adjust the disease information and maintenance recommendations based on subsequent bridge inspection reports to ensure the timeliness and accuracy of the graph.

[0096] S31, integrate multi-period inspection data to drive the iteration of the dedicated knowledge graph and build a spatiotemporal correlation-indicator coupling discrimination framework. The three parameters of component number, relative position and inspection time are generated into a sparse vector V through TF-IDF. TF-IDF , and input the BERT model to obtain the dense semantic vector VBERT ; Calculate the spatiotemporal similarity between historical diseases and new diseases. When the similarity exceeds the threshold of 0.8, it is determined to be a historical disease.

[0097] S32, for historical diseases, based on the adjacent detection cycle change value Δδ' of the disease severity index value j Set static thresholds to divide the warning into three levels. When δ' j When the preset severity threshold is exceeded, a severe warning is triggered, indicating that the disease is accelerating; when Δδ' j When the threshold is between the mild threshold, a mild warning is output, prompting the need for continuous monitoring; when Δδ' j A stable state is determined when the severity is below the mild threshold or shows negative growth. For new diseases, a similar process to step S2 triggers the general knowledge graph inference engine to generate a treatment plan. An incremental update mechanism dynamically links disease evolution data from inspection reports with maintenance recommendations, ultimately achieving precise iteration of graph information and synchronization of risk status.

[0098] The above content merely illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of this technical solution belong to the technical idea proposed by the present invention and fall within the protection scope of the claims of the present invention.

Claims

1. A hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making, characterized by: The steps include: S1. Constructing a general knowledge graph layer: Based on the defect diagnosis and treatment content of bridge maintenance specifications, we construct a classification of hollow slab girder bridge defects and maintenance rules. We use a fuzzy matrix weight allocation method, combining Bayesian reasoning and Beta distribution to quantify the importance of indicators. We also integrate the TF-IDF and BERT models to calculate the confidence level of defect information and indicators, forming a complete general knowledge graph layer. S2. Build a dedicated knowledge graph layer: Extract the disease entity and its indicator text based on the inspection report of a single bridge j Based on the general knowledge graph, the knowledge graph reasoning and rule calculation are performed on the special knowledge graph to calculate the belief distribution of disease classification. Based on the type and severity of the damage, maintenance recommendations are assigned to each damage and maintenance measure nodes are added to the dedicated knowledge graph to form a maintenance decision-making knowledge graph dedicated to bridges. S3. Dynamic update of special knowledge graph: construct spatiotemporal characteristic parameters of disease through component number, relative position and detection time to realize the timeliness judgment of updated data; for historical diseases, based on the disease severity index value δ' j The change of Δδ' j Evaluate development trends and output warnings of different levels; for new diseases, trigger the general knowledge graph reasoning engine to generate treatment plans, and dynamically associate the disease evolution data in the detection report with maintenance recommendations through the incremental update mechanism to achieve accurate iteration of graph information and synchronization of risk status.

2. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 1 is characterized in that: The implementation process of step S1 is as follows: S11. Extract IF-THEN structure rules based on bridge maintenance specifications and construct a set of rules C containing disease classification R and maintenance decision rule set M R The composite rule system R is specifically defined as R=(R F ,{R F →C R ,(R' F ,D)→M R }); R F represents the feature set used to represent disease classification, R' F Represents the feature set used to represent the severity of the disease; for rules that describe the same conclusion but have different features, a disjunction operation is used to merge them to effectively reduce rule redundancy; Based on universal Score applicablity (r) and discrimination Score discrimination (r) Screening out the optimal rule set with high universality and discrimination; S12. Quantifying different indicators f in typical disease classification and maintenance decision rules for hollow slab girder bridges based on Bayesian-Beta framework i and f j The relative importance of and mapped to triangular fuzzy number a i,j By comparing the relative importance of each index, we can get the triangular fuzzy number and build a complete fuzzy judgment matrix A=(a ij ) n×n The fuzzy judgment matrix establishes the fuzzy judgment relationship between the indicator layer and the disease classification target layer in the evaluation rules. The weight distribution is solved by the fuzzy comprehensive operator. Combined with entropy optimization, accurate weighting is achieved under multi-criteria decision-making, and a data-driven hollow slab girder bridge indicator weight system is output. S13, integrating TF-IDF high-frequency features with BERT deep semantic representation, first calculate the rule index value δ j and real disease index text d in bridge inspection j Their respective TF-IDF vectors and Obtain surface feature matching degree S based on cosine similarity TF-IDF (δ j ,d j ), and use the pre-trained BERT model to set the rule index value δ j and real disease index text d in bridge inspection j Encode into semantic vector and calculate deep semantic similarity S BERT (δ j ,d j ), realize the multi-dimensional collaborative matching of rules and disease characteristics; by fusing these two similarity scores, realize the multi-dimensional collaborative matching of the rule base in the general knowledge graph and the disease instance data in the special knowledge graph and the indicator confidence C rule comprehensive assessment.

3. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 1 is characterized in that: The implementation process of step S2 is as follows: S21, extracting disease information based on the inspection report of a single bridge, based on the weights of different indicators in the general knowledge graph ω i and the confidence index c i Weighted aggregation drives the adaptive matching and activation of rule chains to achieve knowledge reasoning and rule calculation for dedicated knowledge graphs; each disease category D k Supported by multiple related indicators, the disease classification belief distribution of each disease manifestation is calculated S22. For each disease of a single bridge, combined with the disease classification results D k and disease severity index f' j Provide maintenance measures j , the decision result is expressed as <disease entity, recommended_measure, m j >The triple form of the dedicated knowledge graph is updated to support explainable traceability.

4. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 1 is characterized in that: The implementation process of step S3 is as follows: S31. Fusion of multi-period inspection data drives the iteration of a dedicated knowledge graph to build a spatiotemporal correlation-indicator coupling discrimination framework: The three parameters of component number, relative position, and inspection time are generated into a sparse vector V through TF-IDF. TF-IDF , and input the BERT model to obtain the dense semantic vector V BERT ; Calculate the spatiotemporal similarity between historical diseases and newly diagnosed diseases, and determine them as historical diseases when the similarity exceeds the threshold; S32, for historical diseases, based on the adjacent detection cycle change value Δδ' of the disease severity index value j Set static thresholds to divide the warning into three levels; when δ' j A severe warning is triggered when the preset severity threshold is exceeded, indicating that the disease is accelerating; When Δδ' j When the threshold is between the mild and high thresholds, a mild warning is output, prompting the need for continuous monitoring. When Δδ' j When it is less than the mild threshold or shows negative growth, it is judged to be in a stable state; for new diseases, general knowledge graph reasoning and rule calculation are triggered to generate treatment plans; through the incremental update mechanism, the disease evolution data in the inspection report is dynamically linked with the maintenance recommendations to achieve accurate iteration of the graph information and synchronization of the risk status.

5. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 2 is characterized in that: The implementation process of step S11 is as follows: S11-1. Extract IF-THEN structure rules from bridge maintenance specifications and construct a set of rules containing disease classification C R and maintenance decision rule set M R The composite rule system is defined as follows: R=(R F ,{R F →C R ,(R' F ,D)→M R }) Among them, R F ={f j =δ j |j=1,2,...,m} represents the index set f used to represent disease classification j is the indicator name, δ j is the indicator value, m is the total number of indicators; R' F ={f' j =δ' j |j=1,2,...,n} represents the feature set used to represent the severity of the disease; represents the disease classification rule set; k represents the total number of rules, usually k>|D|, J (i) Represents the index set of indicators involved in the i-th rule; represents the auxiliary maintenance decision rule set; k2 represents the total number of rules, J (v) Represents the feature index set involved in the vth rule; M={m1,m2,...,m p } represents the maintenance decision set; S11-2. Use disjunction operation to merge to unify the description of rules, reduce redundancy and improve the simplicity of the rule base; the specific merging formula is: R=(R F ,{R F →C Rmerged ,(R' F ,D)→M Rmerged }) For all results d n The rule index set of For all results m p The rule index set of ; S11-3. Score for measuring the applicability of rules in multiple standards: Where S is the set of all specifications; δ(r∈Γ) is an indicator function that counts 1 if rule r appears in specification Γ and 0 otherwise; Rule discrimination evaluates the specificity of the rule to the disease type: Among them, Conflict(r) refers to the number of conflicts of rule r, that is, the number of times the rule feature combination is the same but points to different disease types or maintenance measures; Total(r) refers to the total number of times rule r appears in all summarized rule sets.

6. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 2 is characterized in that: The implementation process of step S12 is as follows: S12-1. By calculating the probability of disease occurrence P(D k |f m ) and P(D k |f n ), and introduce Beta distribution for interval reasoning to obtain the credibility interval of disease occurrence: Where W m,n Representation rule f m Relative to rule f n The importance of is used to measure the difference in their impact on the disease; Indicates the index f m When the disease occurs, k The probability of occurrence, D k represents the different mechanisms of disease D, and there are K possible disease mechanisms; P(f m |D k ) is the disease D k Occurrence indicator f m Observed probability; P(D k ) is the disease D k Prior probability; interval reasoning is performed through Beta distribution to obtain the credibility interval of disease occurrence; S12-2. Based on the expert knowledge in the field of bridge maintenance, combined with the probability ratio a i,j , a fuzzy judgment matrix is established for the key evaluation rules, and the fuzzy judgment matrix is constructed using triangular fuzzy numbers. The form is as follows: A=[a ij ] n×n a ij =(l ij ,m ij ,u ij ) Among them, A is the fuzzy judgment matrix, each element a i,j It is calculated by the relative importance of rules; for each fuzzy judgment matrix, calculate the possibility that the i-th rule is more important than other rules: After normalization, we get the weight vector W = Normalize(d(A1), d(A2), ..., d(A n )) T .

7. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 2 is characterized in that: The implementation process of step S13 is as follows: S13-1. First, the rule index value δ j and real disease index text d in bridge inspection j , perform word segmentation and calculate their respective TF-IDF vectors and Then, the cosine similarity is used to calculate the matching degree between the two vectors. The calculation formula is as follows: S13-2. In the BERT calculation stage, the pre-trained BERT model is used to convert the rule indicator f and the disease indicator text D into high-dimensional vector representations; the cosine similarity is used to calculate the matching degree between the two. The calculation formula is as follows: Indicator confidence C rule The calculation of is done using the hierarchical matching formula as follows: Based on indicator confidence, multi-dimensional collaborative matching of general knowledge graph rule base and special knowledge graph disease instance data is achieved.

8. The hierarchical knowledge graph construction method for hollow slab girder bridge maintenance decision-making according to claim 5 is characterized in that: The rules with applicability scores greater than 80% and discrimination greater than 50% are retained, and the rules with applicability scores not exceeding 80% or discrimination not exceeding 50% need to be optimized or removed.

Images