Intelligent pavement maintenance scheme generation method based on retrieval enhancement generation

Abstract

The invention relates to the technical field of intelligent information processing, and discloses an intelligent pavement maintenance scheme generation method based on retrieval enhancement generation, which comprises the following steps of: acquiring a disease text, an image parameter, a structure hierarchy, a material parameter and traffic and climate data, and constructing a multi-domain intersection problem vector; rough retrieval is completed based on semantics and the adjacent distance between the structure and the load, fine retrieval is executed in combination with the disease stage, the cross-season and the traffic level, and a candidate data set is formed; screening an effective data set according to the structure thickness, the meteorological sequence, the cause chain and the material compatibility; extracting cause, extension, process and material fragments to construct a data fragment group, and inputting the data fragment group and a multi-domain cross problem vector into a generation model to form a preliminary maintenance scheme; and carrying out material, process and environment consistency checking on the maintenance scheme, adding constraint for regeneration when the verification is not passed, and finally outputting the maintenance scheme. According to the invention, the whole-process structured control of the pavement maintenance scheme from data selection to content generation is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, and more specifically, to an intelligent method for generating road maintenance schemes based on retrieval enhancement. Background Technology

[0002] As highways age, various pavement defects are showing a trend of diversified causes, more complex expansion paths, and increased difficulty in material adaptation. Existing methods for generating pavement maintenance plans mainly rely on a single type of input, such as semantic descriptions of defects, image features of defects, or manual experience parameters, and generate maintenance plans by retrieving fixed templates or relying on simple rule combinations.

[0003] However, such methods have the following shortcomings: First, existing technologies typically only perform shallow feature extraction on the text or images of the damage, lacking the ability to perform multi-domain cross-correlation analysis between the semantics of the damage, the geometric features of the damage, the structural hierarchy parameters, and the traffic load features. This prevents the construction of a joint problem description reflecting the essential mechanism of the damage, resulting in a mismatch between the search results and the actual causes of the damage. Second, traditional search methods often rely on a single distance metric or label matching, roughly comparing the damage samples with historical data, ignoring multi-dimensional fine-grained differences such as semantic feature density, differences in the number of structural hierarchy levels, and load peak sequences. This leads to the inclusion of a large amount of data in the search results that is inconsistent with the stage of the damage, the seasonal environment, or the traffic level, resulting in insufficient accuracy in the subsequent scheme generation. Furthermore, when generating maintenance schemes, existing technologies often use direct text splicing or fixed template filling, lacking a mechanism to decompose historical valid data into structured inputs such as causal fragments, extended path fragments, construction procedure fragments, and material fragments. This often results in generated content with problems such as incorrect procedure order, incompatible material parameters, or inconsistencies with the target environment. Furthermore, existing technologies lack calculation methods based on constraints such as material strength range, temperature window overlap ratio, and construction equipment compatibility range, making it difficult to guarantee the logical consistency and feasibility of the generated solutions.

[0004] In summary, existing technologies generally suffer from problems such as incomplete descriptions of road defects, inaccurate multidimensional searches, and a lack of logical verification of generated results, making it difficult to meet the intelligent generation needs in complex road defect scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent method for generating road maintenance schemes based on retrieval enhancement, so as to solve the above-mentioned problems.

[0006] This invention provides an intelligent method for generating road maintenance plans based on retrieval enhancement, comprising:

[0007] S1: Collect disease description text, disease image parameters, road structure level parameters, road section material parameters, traffic operation parameters, and climate fluctuation parameters;

[0008] Based on lexical segmentation, syntactic dependency classification, and structural position encoding, a disease semantic matrix, a geometric feature matrix, a structural hierarchy matrix, and a load action sequence are generated, respectively.

[0009] Based on the semantic principal axis in the disease semantic matrix, the damage direction vector in the geometric feature matrix, the layer index in the structural hierarchy matrix, and the peak sequence table in the load action sequence, a multi-domain cross problem vector is constructed.

[0010] S2: Use multi-domain cross-question vectors to perform a coarse search on the historical maintenance database, filter candidate data entries based on the combined distance of semantic adjacency distance, structural adjacency distance and load adjacency distance, and perform a fine search based on disease stage information, cross-seasonal information and traffic level information to form a candidate data set;

[0011] S3: For each data entry in the candidate dataset, perform an exclusion step based on the road structure layer thickness range, cumulative rainfall sequence, diurnal temperature range sequence, high temperature exposure duration sequence, completeness of the semantic chain of disease causes, and compatibility of material parameter groups to obtain the effective dataset;

[0012] S4: Extract text fragments of disease causes, disease expansion paths, construction procedures, and material parameters from the effective data set to form a set of cause fragments, an expansion fragment set, a process fragment set, and a material fragment set, respectively.

[0013] Establish fragment connection relationships based on the causal order, process order, and material usage order between fragments to construct data fragment groups;

[0014] S5: Input the multi-domain cross problem vector and data fragment group into the generation model to generate a preliminary maintenance plan that includes the cause segment, extension segment, process segment and material segment;

[0015] S6: Perform material parameter consistency verification, process logic consistency verification, and environmental condition consistency verification on the preliminary maintenance plan in sequence;

[0016] If all consistency checks pass, the final maintenance plan will be output.

[0017] If at least one fails, the corresponding inconsistency item will be constructed as a generation constraint and steps S5 and S6 will be re-executed until all consistency checks pass.

[0018] Furthermore, the generation of the disease semantic matrix in S1 includes:

[0019] The disease description text is divided into a sequence of semantic units according to the subject-verb-object structure;

[0020] Calculate the semantic stability value for each semantic unit and sort them according to semantic stability to form a semantic axis;

[0021] Calculate the gradient distribution at the edge of the damaged area based on the parameters of the disease image and generate the damage direction vector;

[0022] Generate layer indexes corresponding to surface layer, base layer, subbase layer and subgrade based on road structure layer parameters;

[0023] The semantic axis, damage direction vector, layer index, and load action sequence are concatenated according to their temporal positions to form a structured semantic code.

[0024] Furthermore, the coarse search performed in S2 includes:

[0025] For multi-domain intersection problems, vector normalization is performed, and the vectors are split into semantic vector subsets, structural vector subsets, and load vector subsets.

[0026] Calculate the semantic adjacency distance, structural adjacency distance, and load adjacency distance respectively;

[0027] Arrange the distance values ​​in ascending order to form a distance arrangement sequence;

[0028] When the minimum adjacent distance corresponds to a semantic vector subset, match data entries whose difference between the number of semantic nodes and the target number of semantic nodes is within a preset range; when the minimum adjacent distance corresponds to a structural vector subset, match data entries whose difference between the number of structural levels and the target number of structural levels is within a preset range; when the minimum adjacent distance corresponds to a load vector subset, match data entries whose difference between the number of load peaks and the target number of load peaks is within a preset range.

[0029] A coarse candidate set is formed based on the data entries that meet the criteria;

[0030] When the number of coarse candidate sets exceeds the set range, data entries are removed in reverse order of the distance sorting sequence until the number of coarse candidate sets is within the threshold range.

[0031] Furthermore, performing a fine search in S2 includes:

[0032] Extract the disease stage marker for each data entry in the coarse candidate set and calculate the stage difference with the target disease stage marker;

[0033] Generate seasonal vectors from the temperature, rainfall, and diurnal temperature range sequences over three consecutive months;

[0034] Compare seasonal vector differences and exclude data entries whose differences exceed the set range;

[0035] Compare the differences in traffic levels and exclude data entries with inconsistent traffic levels;

[0036] The remaining data items are sorted by stage difference, and the data items with the smallest stage difference are retained to form a candidate data set.

[0037] Furthermore, the exclusion of material parameter groups in S3 includes:

[0038] The minimum strength, maximum strength, and room temperature compressive strength of each material in the material combination are encoded as strength ranges;

[0039] The construction temperature window is divided into a lower limit temperature, a target temperature, and an upper limit temperature, forming a ternary temperature window.

[0040] Compare whether the intensity range covers the intensity range required for the target disease; if it does not, exclude the corresponding data entry.

[0041] Compare the overlap ratio between the ternary temperature window and the temperature fluctuation sequence of the target segment. If the overlap ratio is less than the specified threshold, the corresponding data entry is excluded.

[0042] Compare the compatibility range of construction equipment. If there are items in the material combination that are unavailable for equipment, then exclude the corresponding data entries.

[0043] Furthermore, the construction of data fragment groups in S4 includes:

[0044] Calculate the number of causal nodes for each causal segment and sort them by the number of nodes;

[0045] Calculate the process dependency for each construction process segment and sort them by dependency.

[0046] A segment mapping relationship table is constructed based on the difference between the number of causal nodes and the process dependence.

[0047] Based on the fragment mapping relationship table, fragments from the causal fragment set, extended fragment set, process fragment set, and material fragment set are selected in sequence to form a data fragment group arranged by number.

[0048] Furthermore, the preliminary maintenance plan generated in S5 includes:

[0049] The causal fragments, extended fragments, process fragments, and material fragments in the data fragment group are respectively formed into four input sequences according to their serial numbers;

[0050] The four input sequences are used as four auxiliary input segments of the generative model, and the multi-domain cross problem vector is used as the main input segment.

[0051] Insert the four input segments into the generated model as generation prompts according to their sequential positions;

[0052] The generated model outputs a preliminary maintenance plan, including the cause segment, extension segment, process segment, and material segment, in the order of insertion.

[0053] The generative model is an encoder-decoder model based on the Transformer architecture, which generates a preliminary textualized maintenance plan based on the input vector and data fragment groups.

[0054] Furthermore, the material parameter consistency check in S6 includes:

[0055] Compare the strength codes of each material in the preliminary maintenance plan with the corresponding strength codes of the materials in the valid data set;

[0056] Compare the overlap ratio between the temperature windows of each material in the preliminary maintenance plan and the temperature windows of the corresponding materials in the effective data set to see if they meet the preset overlap range.

[0057] Compare whether the construction equipment used in the preliminary maintenance plan is within the equipment compatibility range of the valid data set;

[0058] If all three comparisons meet the consistency criteria, then the material parameter consistency check is passed.

[0059] Furthermore, the process logic consistency check in S6 includes:

[0060] Calculate the sequential number of each construction procedure in the preliminary maintenance plan and compare it with the sequential numbering pattern of the corresponding procedure in the valid data set;

[0061] Compare whether the dependencies between procedures in the preliminary maintenance plan are complete and without duplication;

[0062] Calculate the time interval between adjacent processes in the preliminary maintenance plan and compare it with the corresponding time interval range in the valid data set;

[0063] When all three comparisons meet the consistency conditions, the process logic consistency check is performed.

[0064] Furthermore, the environmental condition consistency check in S6 includes:

[0065] Extract construction temperature sequence, rainfall interval sequence, and traffic load sequence from the preliminary maintenance plan;

[0066] The three types of sequences are compared with the temperature sequence, rainfall sequence and traffic load sequence of the target road segment respectively, and the minimum value of the three types of overlap is calculated and denoted as the minimum overlap.

[0067] The minimum overlap is compared with the environmental matching range of the effective dataset;

[0068] When the minimum overlap is within the environmental matching range, environmental condition consistency is checked.

[0069] The present invention has the following technical effects:

[0070] This application provides an intelligent generation method for road maintenance schemes based on retrieval enhancement. By constructing multi-domain cross-problem vectors, it achieves deep integration of semantic features, geometric morphology, structural layer, and load response of road defects. It improves the matching accuracy of historical data by combining distance filtering and multi-dimensional fine retrieval mechanism, ensures the logical coherence of the scheme content by using structured fragment connection relationship, and guarantees the engineering feasibility of the scheme by relying on multi-level consistency verification of material parameters, process logic, and environmental conditions. It has the advantages of accurately reflecting the internal mechanism of road defects, improving the accuracy of historical data retrieval, and ensuring the logical self-consistency and engineering feasibility of the maintenance scheme in terms of material selection, process connection, and environmental adaptation. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 The flowchart illustrates an intelligent method for generating road maintenance schemes based on retrieval enhancement, as provided in an embodiment of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] See Figure 1 As shown, this embodiment of the invention provides an intelligent generation method for road maintenance schemes based on retrieval enhancement, including:

[0076] In the traditional process of generating road maintenance schemes, the reliance on a single type of input for disease description leads to a lack of multi-domain cross-correlation analysis capabilities among disease semantics, damage geometry features, structural hierarchy parameters, and traffic load characteristics, making it impossible to construct a joint problem description that reflects the essential mechanism of the disease. Specifically, existing technologies only perform shallow feature extraction on disease text or images, lacking a deep fusion mechanism for semantic axes, damage direction vectors, layer indices, and peak sequences, resulting in problem vectors that cannot accurately represent the multi-domain dynamic characteristics of the disease. Furthermore, the multi-dimensional retrieval process relies on a single distance metric or label matching, ignoring fine-grained quantification of semantic feature density, differences in the number of structural hierarchy levels, and load peak sequences, leading to systematic deviations in retrieval results from disease stages, seasonal environments, or traffic levels. In addition, scheme generation uses direct text splicing or fixed template filling, failing to structurally decompose historical data into causal segments, extended path segments, construction procedure segments, and material segments, resulting in problems such as incorrect procedure sequences, incompatible material parameters, or inconsistent environmental conditions. Therefore, the aforementioned defects directly affect the accuracy, logical consistency, and feasibility of the maintenance plan, thereby limiting the plan's adaptability to complex pavement defects.

[0077] For example, in a heavily trafficked section of a highway, the damage manifests as longitudinal cracks, with a measured crack width of 3 mm and a depth extending to the base layer. Simultaneously, road structural layer parameters were collected, including a surface layer thickness of 100 mm and a base layer thickness of 200 mm. Traffic operation parameters showed an average daily number of heavily trafficked vehicles of 5000. Existing technology only performs retrieval based on the damage image features, failing to integrate the layer index in the structural layer matrix and the peak sequence in the load action sequence. This results in retrieval results containing maintenance data applicable to lightly trafficked, seasonal road sections. In this case, the construction procedures required low-temperature construction, while actual climate fluctuation parameters indicate that it is currently a hot summer, with diurnal temperature variations exceeding the material's applicable range. Consequently, the material parameter segments in the generated solution do not match the temperature window, the construction procedure segments are incorrectly arranged, and the material strength range is incompatible with environmental conditions.

[0078] If the above problems are not resolved, the generated maintenance plan will not be able to accurately match the causes of the disease with the target environmental conditions, resulting in the failure of material parameter group compatibility and the break of process logic dependencies. This will lead to ineffective maintenance measures, accelerated degradation of road structure performance, and may increase the risk of traffic disruption caused by repeated maintenance operations.

[0079] For this, please refer to Figure 1 As shown, this application proposes an intelligent generation method for road maintenance schemes based on retrieval enhancement, including:

[0080] S1.1: Collect disease description text, disease image parameters, road structure level parameters, road section material parameters, traffic operation parameters, and climate fluctuation parameters;

[0081] S1.2: Generate the disease semantic matrix, geometric feature matrix, structural hierarchy matrix and load action sequence based on lexical segmentation, syntactic dependency classification and structural position encoding respectively;

[0082] S1.3: Construct a multi-domain cross problem vector based on the semantic principal axis in the disease semantic matrix, the damage direction vector in the geometric feature matrix, the layer index in the structural hierarchy matrix, and the peak sequence table in the load action sequence;

[0083] S2: Use multi-domain cross-question vectors to perform a coarse search on the historical maintenance database, filter candidate data entries based on the combined distance of semantic adjacency distance, structural adjacency distance and load adjacency distance, and perform a fine search based on disease stage information, cross-seasonal information and traffic level information to form a candidate data set;

[0084] S3: For each data entry in the candidate dataset, perform an exclusion step based on the road structure layer thickness range, cumulative rainfall sequence, diurnal temperature range sequence, high temperature exposure duration sequence, completeness of the semantic chain of disease causes, and compatibility of material parameter groups to obtain the effective dataset;

[0085] S4.1: Extract text fragments of disease causes, disease expansion paths, construction procedures, and material parameters from the effective data set to form a set of cause fragments, an expansion fragment set, a process fragment set, and a material fragment set, respectively.

[0086] S4.2: Establish the connection relationship between fragments based on the causal order, process order, and material usage order, and construct a data fragment group;

[0087] S5: Input the multi-domain cross problem vector and data fragment group into the generation model to generate a preliminary maintenance plan that includes the cause segment, extension segment, process segment and material segment;

[0088] S6: Perform material parameter consistency verification, process logic consistency verification, and environmental condition consistency verification on the preliminary maintenance plan in sequence;

[0089] If all consistency checks pass, the final maintenance plan will be output.

[0090] If at least one fails, the corresponding inconsistency item will be constructed as a generation constraint and steps S5 and S6 will be re-executed until all consistency checks pass.

[0091] This application provides an intelligent method for generating pavement maintenance plans based on retrieval enhancement, aiming to solve the problems of incomplete description of pavement defects, inaccurate multidimensional retrieval, and lack of logical verification of the generated results during the pavement maintenance plan generation process. The multi-domain cross-problem vector refers to a problem representation vector that integrates semantic, geometric, structural, and load features. In practical applications, this can be achieved through feature concatenation or weighted averaging, for example, by aligning feature vectors of different dimensions according to time series and then concatenating them, or by combining them into a single vector after dimensionality reduction through principal component analysis. Furthermore, the combined distance of semantic adjacency distance, structural adjacency distance, and load adjacency distance can be calculated as a weighted sum of each sub-distance, or the weight allocation can be optimized using a distance learning model based on support vector machines. Specifically, defect stage information, cross-seasonal information, and traffic level information can be precisely retrieved based on timestamp sequence matching or expert rule bases. For example, the stage difference can be determined by comparing the defect occurrence time with the time tags of historical data, or seasonal feature vectors can be generated using meteorological data for similarity comparison. Therefore, exclusion conditions such as the thickness range of road structure layers, cumulative rainfall sequence, diurnal temperature range sequence, high temperature exposure duration sequence, integrity of the semantic chain of disease causes, and compatibility of material parameter groups can be set as predefined threshold ranges. Filtering is achieved through parameter comparison, such as checking whether the thickness value is within the allowable range or whether the calculated cumulative rainfall value exceeds the threshold. As a preferred implementation, the construction of data fragment groups can be based on text similarity to calculate the causal order, process order, and material usage order between fragments, or by using a predefined logical rule graph to determine the connection relationship. The generative model can adopt a sequence-to-sequence architecture, such as an encoder-decoder structure based on a recurrent neural network, receiving multi-domain cross-question vectors and data fragment groups as input, and outputting a preliminary maintenance plan containing causal segments, extension segments, process segments, and material segments. Finally, material parameter consistency verification, process logic consistency verification, and environmental condition consistency verification are performed through parameter range verification and logical rule checks, such as comparing whether the material strength is within the allowable range or verifying whether the process dependency relationship is complete. If they fail, the inconsistencies are constructed as generation constraints, and the scheme generation and verification steps are re-executed, thereby ensuring that the final scheme meets all consistency conditions. This embodiment systematically improves the completeness of disease description, the accuracy of retrieval, and the logical consistency of solution generation by integrating multi-dimensional parameters to construct a joint problem description, implementing multi-dimensional fine-grained retrieval, and a structured fragment generation and iterative verification mechanism.

[0092] In the intelligent generation of pavement maintenance plans, deterioration description text, deterioration image parameters, road structure hierarchy parameters, road segment material parameters, traffic operation parameters, and climate fluctuation parameters are collected simultaneously to cover multi-dimensional information including semantics, geometry, structure, load, and environment, avoiding the one-sidedness of descriptions caused by single inputs. Lexical segmentation, syntactic dependency classification, and structural position encoding are applied to generate the deterioration semantic matrix, geometric feature matrix, structural hierarchy matrix, and load action sequence. These processing methods deeply extract the deep structural features of the text and images, rather than relying solely on surface keywords or simple features. Furthermore, the semantic principal axis in the deterioration semantic matrix, the damage direction vector in the geometric feature matrix, the layer index in the structural hierarchy matrix, and the peak sequence table in the load action sequence are integrated to construct a multi-domain cross-problem vector. This vector, by fusing the stability of the semantic principal axis, the geometric orientation of the damage direction vector, the structural hierarchy of the layer index, and the load dynamics of the peak sequence, forms a joint problem description reflecting the essential mechanism of the deterioration. For example, when the description text of a road defect on a certain urban expressway section is "longitudinal crack, length 30 meters, width 8 millimeters", the defect image parameters show that the gradient distribution of the crack edge is concentrated in the east-west direction, the road structure layer parameters include a surface layer thickness of 10 centimeters and a base layer thickness of 20 centimeters, the traffic operation parameters record an average daily heavy-load vehicle traffic volume of 5,000 vehicles, and the climate fluctuation parameters provide a cumulative rainfall of 50 millimeters in the past three days, lexical segmentation generates a subject-verb-object semantic unit sequence, semantic stability sorting forms the semantic main axis, the damage direction vector is calculated to be dominated by the east-west direction, the layer index is marked as surface layer and base layer, and the peak of the load action sequence is identified as the morning and evening peak hours, thus forming a structured semantic code.

[0093] Multi-domain cross-problem vectors were used to perform a coarse search on the historical maintenance database. Combined distances of semantic adjacency, structural adjacency, and load adjacency were calculated to filter candidate data entries. This process avoided coarse comparisons based on a single distance metric by quantifying semantic feature density, differences in the number of structural levels, and fine-grained differences in load peak sequences. Specifically, damage stage information, cross-seasonal information, and traffic level information were used for a fine search. Damage stage markers were compared, and seasonal vector differences and traffic level differences were evaluated to exclude data entries inconsistent with the actual damage stage, seasonal environment, or traffic level, forming a candidate dataset. Each data entry in the candidate dataset underwent an exclusion process based on road structure layer thickness range, cumulative rainfall sequence, diurnal temperature range sequence, high-temperature exposure duration sequence, completeness of the damage cause semantic chain, and compatibility of material parameter sets. Environmental parameter thresholds and material logic constraints were applied to filter invalid entries, resulting in a valid dataset. From the effective dataset, fragments of text about disease causes, disease propagation paths, construction procedures, and material parameters are extracted and formed into a fragment set. The causal order, procedure order, and material usage order between fragments are used to establish the connection relationship between fragments and construct a data fragment group. This structured decomposition avoids the problems of incorrect procedure order or material incompatibility caused by direct text splicing.

[0094] Multi-domain cross-problem vectors and data fragment groups are input into the generative model to generate a preliminary maintenance plan containing causal segments, extension segments, process segments, and material segments. The multi-domain context of the problem vectors and the structured order of the fragment groups are used to guide the generation process, ensuring logical coherence. The preliminary maintenance plan is sequentially checked for material parameter consistency, process logic consistency, and environmental condition consistency. Material strength coding, temperature window overlap ratio, and construction equipment compatibility range are compared to verify material parameter consistency; process sequence number, dependency integrity, and time interval range are compared to verify process logic consistency; and the sequence overlap of construction temperature sequence, rainfall interval sequence, and traffic load sequence is calculated to verify environmental condition consistency. If all consistency checks pass, the final maintenance plan is output; if at least one fails, the corresponding inconsistency is constructed as a generation constraint, and steps S5 and S6 are re-executed until all consistency checks pass. This iterative mechanism forces the plan to satisfy constraints such as material strength range, process dependency, and environmental sequence overlap.

[0095] This method integrates multi-source data to construct a joint problem description, achieving accurate capture of the essential mechanism of the disease and solving the problem of incomplete disease description. Through coarse retrieval based on combined distance and fine retrieval based on disease stage, it ensures that the retrieval results strictly match the actual disease stage, seasonal environment, and traffic level, solving the problem of inaccurate multi-dimensional retrieval. Through structured fragment generation and iterative verification mechanism, it forces the solution to meet the requirements of material logic constraints, process dependencies, and environmental sequence overlap, avoiding problems such as incorrect process order, incompatible material parameters, or inconsistent environment, thereby ensuring the logical consistency and feasibility of the generated solution.

[0096] In some of the embodiments described above in this application, a disease semantic matrix is ​​proposed to construct a multi-domain cross-question vector. However, in its implementation, it only relies on lexical segmentation and syntactic dependency classification for shallow feature extraction, and fails to deeply analyze the intrinsic stability and structural correlation of semantic units. As a result, the semantic matrix cannot effectively form a semantic axis that reflects the core mechanism of the disease, resulting in an incomplete construction of the question vector, which in turn affects the matching accuracy between the subsequent retrieval results and the actual causes of the disease.

[0097] In response, this application further proposes the following steps for generating the disease semantic matrix in S1: splitting the disease description text into a sequence of semantic units according to the subject-verb-object structure; calculating the semantic stability value for each semantic unit and sorting them according to semantic stability to form a semantic principal axis; calculating the edge gradient distribution of the damaged area based on the disease image parameters and generating a damage direction vector; generating layer indices corresponding to the surface layer, base layer, subbase layer and subgrade based on the road structure layer parameters; and concatenating the semantic principal axis, damage direction vector, layer index and load action sequence according to temporal position to form a structured semantic code.

[0098] The semantic unit sequence refers to the set of logical units formed by decomposing the disease description text according to the subject-verb-object grammatical structure. It can be realized using dependency parsing tools or rule-based parsers. The purpose is to accurately capture the core logical relationships between the disease subject, behavior, and object, providing a structured foundation for deep semantic analysis. The semantic stability value can be understood as an indicator that quantifies the reliability and criticality of semantic units in the disease description. It can be calculated based on word frequency statistics, context consistency analysis, or expert scoring models in historical disease reports. The purpose is to highlight core semantic information and filter redundant interference to ensure the stability of the semantic axis. The damage direction vector refers to vector data that represents the geometric direction and severity of the road damage area. It can be generated using the Canny edge detection algorithm or the Sobel gradient operator combined with directional clustering methods. The purpose is to integrate image geometric features into semantic analysis and make up for the limitations of pure text description. The layer index refers to the encoded information that identifies the layer position of the road structure. It can be mapped to the surface layer, base layer, subbase layer, and subgrade as discrete index values ​​according to road design specifications. The purpose is to establish a computable association between the disease and the road structure layer and support multi-domain feature fusion.

[0099] Specifically, the proposed solution first decomposes the disease description text using a subject-verb-object structure to form a sequence of semantic units to extract the core logic of the sentences. Then, it calculates and sorts the stability values ​​of each semantic unit, constructing a semantic axis reflecting the core mechanism of the disease to ensure the priority of key information. Simultaneously, it generates a damage direction vector based on disease image parameters to quantify geometric features, and generates a layer index based on road structure layer parameters to associate structural levels. Finally, it aligns and concatenates the semantic axis, damage direction vector, layer index, and load action sequence according to time series to form a structured semantic code. This process, through the synergy of syntactic structure parsing, semantic stability quantification, geometric feature extraction, and structural layer encoding, achieves the organic integration of semantic, geometric, and structural features, enabling the problem vector to fully represent the essential mechanism of the disease.

[0100] As a preferred embodiment, the specific implementation of this application is as follows: For the description text of the defect "longitudinal cracks appear in the asphalt surface layer", the Stanford Parser is used to split the subject, verb, and object to obtain the semantic unit sequence "asphalt surface layer", "appears", and "longitudinal cracks"; by statistically analyzing the occurrence frequency of each semantic unit in historical maintenance data, the stability value is calculated, and "longitudinal cracks" is ranked first on the semantic axis; edge gradient analysis is performed on the defect image using the OpenCV library to generate a damage direction vector representing the direction of the cracks; according to the road design drawings, the surface layer, base layer, subbase layer and subgrade are mapped to the layer index 0, 1, 2, 3; finally, the semantic axis, damage direction vector, layer index and traffic load peak sequence are aligned and spliced ​​according to the timestamp to form a structured semantic code.

[0101] Through the above scheme, this application can accurately construct a semantic axis that reflects the core mechanism of the disease, effectively integrate text semantics, image geometry and road structure features, significantly improve the integrity of the multi-domain cross question vector, and thus enhance the matching accuracy between the subsequent search results and the actual causes of the disease.

[0102] In practical applications, some of the embodiments described above in this application propose coarse retrieval to screen historical maintenance data. However, in its implementation, the screening mechanism that relies solely on the combination distance fails to fully consider the fine-grained differences in specific parameters such as the number of semantic nodes, the number of structural levels, and the number of load peaks. This results in a large number of data entries in the coarse candidate set that do not match the target disease stage, seasonal environment, or traffic level, and the size of the set may get out of control, thereby reducing the accuracy of subsequent fine retrieval and the reliability of the generated scheme.

[0103] In this regard, this application further proposes that the coarse search performed in S2 includes:

[0104] For multi-domain intersection problems, vector normalization is performed, and the vectors are split into semantic vector subsets, structural vector subsets, and load vector subsets.

[0105] Calculate the semantic adjacency distance, structural adjacency distance, and load adjacency distance respectively;

[0106] Arrange the distance values ​​in ascending order to form a distance arrangement sequence;

[0107] When the minimum adjacent distance corresponds to a semantic vector subset, match data entries whose difference between the number of semantic nodes and the target number of semantic nodes is within a preset range; when the minimum adjacent distance corresponds to a structural vector subset, match data entries whose difference between the number of structural levels and the target number of structural levels is within a preset range; when the minimum adjacent distance corresponds to a load vector subset, match data entries whose difference between the number of load peaks and the target number of load peaks is within a preset range.

[0108] A coarse candidate set is formed based on the data entries that meet the criteria;

[0109] When the number of coarse candidate sets exceeds the set range, data entries are removed in reverse order of the distance sorting sequence until the number of coarse candidate sets is within the threshold range.

[0110] Among these, normalization refers to standardizing the multi-domain cross-problem vectors to eliminate dimensional differences, which can be achieved using the L2 norm normalization algorithm to ensure the comparability of different feature dimensions; the number of semantic nodes refers to the number of semantic units in the disease description text, which can be determined by generating a sequence of semantic units based on a subject-verb-object structure segmentation algorithm, aiming to quantify the density of semantic features; the number of structural levels refers to the total number of levels in the road structure, including surface layer, base layer, subbase layer, and subgrade, which can be obtained directly from the road structure level parameters, aiming to reflect the structural complexity; the number of peak load occurrences refers to the frequency of peak occurrences in the traffic load sequence, which can be extracted by analyzing the load action sequence using a peak detection algorithm, aiming to capture the dynamic changes in load characteristics; the preset range refers to the allowable threshold range of parameter differences, which can be dynamically set according to the statistical distribution of historical maintenance data, aiming to balance the strictness and flexibility of the screening; the threshold range refers to the upper limit of the size of the coarse candidate set, which can be adaptively adjusted based on the real-time processing capability of the system, aiming to prevent the set size from expanding.

[0111] Specifically, the proposed solution normalizes and splits the multi-domain cross-problem vectors into independent subsets, ensuring comparability of semantic, structural, and load features across their respective dimensions and preventing the combined distance from masking key differences. Subsequently, the adjacent distances of each subset are calculated to accurately quantify semantic similarity, structural matching, and load matching, thereby capturing multi-dimensional fine-grained features. After sorting by distance values ​​to form a distance sequence, key parameters are dynamically matched based on the subsets corresponding to the minimum adjacent distance. For example, when the minimum distance points to a load vector subset, priority is given to focusing on differences in the number of load peaks for screening, ensuring the screening process aligns with the most relevant dimension. A coarse candidate set is formed strictly according to matching conditions, effectively excluding entries inconsistent with the target disease features. When the set size exceeds the limit, entries with larger distances are removed in reverse order of the distance sequence, dynamically controlling the candidate set size. This preserves highly relevant data while preventing uncontrolled growth, providing a high-quality input foundation for subsequent refined retrieval.

[0112] As a specific implementation method, the solution of this application is implemented as follows: In the coarse search stage, the multi-domain cross-problem vector is first normalized and split into semantic vector subsets, structural vector subsets, and load vector subsets; the adjacent distances of the three types of subsets are calculated and sorted to form a distance arrangement sequence; assuming that the adjacent distance of the load is the smallest, data entries whose difference between the load peak frequency and the target value is within a preset range are matched, for example, the difference is small and conforms to the historical data distribution; after forming the coarse candidate set, if the number exceeds the system set range, entries are removed from the end according to the distance arrangement sequence until the set size falls back to a reasonable threshold.

[0113] Through the above technical solutions, the number of data entries in the coarse candidate set that do not match the target disease stage, seasonal environment, or traffic level is significantly reduced, and the size of the set is effectively controlled, thereby improving the accuracy of subsequent fine retrieval and the reliability of maintenance plan generation.

[0114] Specifically, in some of the embodiments described above in this application, a fine search is proposed to screen candidate data based on disease stage information, cross-seasonal information, and traffic level information. However, in the coarse search process, due to the lack of sufficient consideration of the dynamic evolution of disease development stages, the continuous characteristics of seasonal environmental changes, and the real-time differences in traffic levels, a large number of data entries that do not match the target disease stage, seasonal environment, or traffic level are mixed into the coarse candidate set. This makes the candidate data set unable to accurately reflect the fine-grained requirements of the actual disease scenario, which in turn leads to problems such as process logic errors, material parameter conflicts, or insufficient environmental adaptability in the subsequent maintenance plan generation.

[0115] In this regard, this application further proposes the following steps for a precise search:

[0116] Extract the disease stage marker for each data entry in the coarse candidate set and calculate the stage difference with the target disease stage marker;

[0117] Generate seasonal vectors from the temperature, rainfall, and diurnal temperature range sequences over three consecutive months;

[0118] Compare seasonal vector differences and exclude data entries whose differences exceed the set range;

[0119] Compare the differences in traffic levels and exclude data entries with inconsistent traffic levels;

[0120] The remaining data items are sorted by stage difference, and the data items with the smallest stage difference are retained to form a candidate data set.

[0121] Among them, disease stage marking refers to the quantitative identifier characterizing the development process of a disease. It can be implemented using discrete labels or continuous numerical sequences based on disease type and severity, with the aim of accurately quantifying the differences in the development process of a disease from its initial stage to its deterioration. Seasonal vectors can be understood as a method of converting time-series data of climate parameters into feature vectors. They can be implemented using time series analysis techniques such as sliding window statistics or frequency domain transformation, with the aim of fully capturing the dynamic correlation of temperature, rainfall, and temperature difference parameters during seasonal transitions. Seasonal vector difference degree refers to a metric for measuring the similarity between two seasonal vectors. It can be implemented using Euclidean distance or dynamic time warping algorithms, with the aim of filtering out environmentally incompatible data by quantifying the differences in the overall fluctuation patterns of climate sequences. Traffic level difference degree can be understood as a comparative indicator of traffic load levels. It can be implemented based on the differences in traffic flow classification or axle load data, with the aim of ensuring the compatibility of the selected data with the traffic conditions of the target road segment. Stage difference sorting refers to the process of arranging data entries according to the stage difference value from smallest to largest. It can be implemented using standard sorting algorithms, with the aim of prioritizing the data entries whose disease development stage is closest to the target.

[0122] Specifically, the precise retrieval mechanism of this application sequentially performs disease stage difference calculation, seasonal vector generation, difference degree comparison, and sorting operations. First, it quantifies the differences in disease development progress to filter data matching the development stage. Then, it generates dynamic seasonal vectors based on climate parameter sequences to characterize seasonal features. Next, it excludes entries incompatible with the seasonal environment by using a difference degree threshold, and simultaneously filters data with mismatched traffic conditions based on traffic level differences. Finally, it selects the entries with the smallest stage difference from the remaining entries to form a candidate dataset. This multi-dimensional dynamic screening process ensures that the candidate dataset is highly consistent with the target scenario in the three key dimensions of disease development stage, seasonal environment, and traffic level, thus providing an accurate and reliable input basis for subsequent maintenance plan generation.

[0123] As a specific implementation method, the solution of this application is implemented as follows: For asphalt pavement cracking defects appearing on a certain highway section, the system extracts the defect stage markers of each data entry from the coarse candidate set, such as "early", "middle" or "late" markers determined based on crack width and depth, and calculates the stage difference with the target defect stage marker; generates a seasonal vector based on the historical data of temperature, rainfall and diurnal temperature range of the road section for the past three months to characterize the current climate characteristics; compares the seasonal vector difference degree to exclude data entries with low winter temperature characteristics; compares the traffic level difference degree to retain only the data that is consistent with the high traffic load level of the target road section; sorts the remaining data entries according to the stage difference, and selects the data with the smallest stage difference to form a candidate data set.

[0124] Through the above-mentioned approach, this application effectively avoids data screening bias caused by misalignment of disease development stages, seasonal environmental mismatch, or inconsistent traffic levels, ensuring that the candidate dataset accurately reflects the fine-grained requirements of the actual disease scenario, thereby significantly improving the logical consistency, material compatibility, and environmental adaptability of subsequent maintenance plans.

[0125] Specifically, in some of the embodiments described above in this application, the exclusion of material parameter groups is proposed to ensure the compatibility of material parameters with the target disease conditions. However, the candidate data set formed during the fine search process may contain data entries where the material strength does not cover the target requirements, the temperature window does not match the actual climate fluctuations, and the construction equipment is unavailable. This leads to inconsistent material parameters in the subsequently generated maintenance plan, affecting the feasibility of the plan.

[0126] In this regard, this application further proposes that the exclusion of material parameter groups in S3 includes:

[0127] The minimum strength, maximum strength, and room temperature compressive strength of each material in the material combination are encoded as strength ranges;

[0128] The construction temperature window is divided into a lower limit temperature, a target temperature, and an upper limit temperature, forming a ternary temperature window.

[0129] Compare whether the intensity range covers the intensity range required for the target disease; if it does not, exclude the corresponding data entry.

[0130] Compare the overlap ratio between the ternary temperature window and the temperature fluctuation sequence of the target segment. If the overlap ratio is less than the specified threshold, the corresponding data entry is excluded.

[0131] Compare the compatibility range of construction equipment. If there are items in the material combination that are unavailable for equipment, then exclude the corresponding data entries.

[0132] Among them, strength range refers to the technical characterization method of integrating discrete material strength parameters into a continuous range. It can be achieved by statistical distribution or envelope fitting of minimum strength, maximum strength and room temperature compressive strength, with the aim of transforming the basic performance of materials into quantifiable coverage comparison conditions; ternary temperature window refers to the structured representation of the construction temperature window into key temperature nodes, which can be achieved by fitting temperature sensor data or extrapolating climate models, with the aim of accurately describing the matching relationship between the construction window and dynamic climate conditions; comparing strength range coverage refers to the technical operation of verifying whether the material strength adaptability range meets the requirements of disease, which can be achieved by interval inclusion algorithms or boundary value verification, with the aim of eliminating material combinations with insufficient or excessive strength; comparing overlap ratio refers to the calculation process of quantifying the matching degree between temperature window and climate sequence, which can be achieved by time series overlap integral or sliding window matching algorithm, with the aim of ensuring the feasibility of construction conditions in actual climate fluctuations; comparing construction equipment compatibility refers to the technical link of checking the availability of equipment required for material construction, which can be achieved by comparing equipment parameter databases or querying resource scheduling systems, with the aim of ensuring the operational continuity of the construction process.

[0133] Specifically, the proposed solution first encodes the minimum strength, maximum strength, and room-temperature compressive strength of each material in the material combination as strength ranges, transforming discrete parameters into continuous ranges and establishing a quantitative benchmark for the basic performance of the materials. Then, the construction temperature window is divided into a lower limit temperature, a target temperature, and an upper limit temperature, forming a ternary temperature window to provide a precise descriptive framework for temperature feasibility. Based on this, by comparing whether the strength ranges cover the strength range required by the target disease, the matching between material strength and disease requirements is directly verified, eliminating data entries with insufficient coverage. Simultaneously, the overlap ratio between the ternary temperature window and the temperature fluctuation sequence of the target section is compared; when the overlap ratio is below a specified threshold, the corresponding entry is excluded, ensuring the coordination between the construction window and dynamic climate changes. Finally, the compatibility range of construction equipment is compared to identify and exclude material combinations for which the equipment is unusable, forming a complete material parameter verification chain. This serialized exclusion mechanism, through the collaborative verification of strength, temperature, and equipment dimensions, transforms ambiguous material descriptions into calculable constraints, effectively filtering out historical data that strictly matches the conditions of the target disease.

[0134] As a specific implementation method, the scheme of this application is implemented as follows: For asphalt concrete material combination, the minimum strength of 20 MPa, the maximum strength of 40 MPa and the compressive strength at room temperature of 30 MPa can be encoded as a strength range of [20, 40] MPa; the construction temperature window is represented as a ternary structure of a lower limit temperature of 130 degrees Celsius, a target temperature of 150 degrees Celsius and an upper limit temperature of 170 degrees Celsius; when comparing, if the strength range required for the target defect is [25, 35] MPa, then [20, 40] MPa covers this range; the overlap ratio between the ternary temperature window and the temperature fluctuation sequence of the target section is calculated to be 70%, which is greater than the specified threshold of 60%; at the same time, the equipment compatibility check confirms that the paver model is in the list of available equipment, thereby retaining the data entry. If the strength range [15,25] MPa of a certain cement-stabilized crushed stone material does not cover the target range [20,30] MPa, then the corresponding data entry is excluded; if the overlap ratio between the ternary temperature window of a certain material and the climate sequence is 50%, which is lower than the threshold, then the entry is excluded; if the roller model required by the material combination is not within the equipment compatibility range, then data entries indicating unusable equipment are excluded.

[0135] Through the above-mentioned scheme, this application effectively eliminates historical data entries where the material strength does not meet the requirements of the disease, the construction temperature window is out of sync with the actual climate fluctuations, and the construction equipment is unusable. It ensures that the material parameters in the effective data set are strictly matched with the target disease conditions in three dimensions: strength adaptability, temperature feasibility, and equipment operability. This avoids the risk of maintenance plan implementation failure due to inconsistent material parameters and significantly improves the feasibility and reliability of the generated plan.

[0136] Specifically, in some of the embodiments described above in this application, a method for constructing data fragment groups is proposed to establish fragment connection relationships based on causal order, process order, and material usage order. However, in the implementation process, due to the lack of quantitative sorting and systematic mapping of causal fragments and process fragments, the fragment connection relationships are difficult to accurately reflect the intrinsic relationship between the causes of diseases and construction logic, leading to problems such as incorrect process order, mismatched material parameters, or broken causal chains, which in turn affects the logical consistency and feasibility of subsequent maintenance plans.

[0137] In this regard, this application further proposes that the construction of data fragment groups in S4 includes:

[0138] Calculate the number of causal nodes for each causal segment and sort them by the number of nodes;

[0139] Calculate the process dependency for each construction process segment and sort them by dependency.

[0140] A segment mapping relationship table is constructed based on the difference between the number of causal nodes and the process dependence.

[0141] Based on the fragment mapping relationship table, fragments from the causal fragment set, extended fragment set, process fragment set, and material fragment set are selected in sequence to form a data fragment group arranged by number.

[0142] In practical applications, the number of causal nodes refers to the number of key event points in the description of disease causes. Specifically, it can be achieved by using natural language processing technology to extract semantic units through dependency parsing and counting them. The purpose is to quantify the complexity of disease causes and provide a basis for ranking.

[0143] Among them, process dependency can be understood as the strength index of the logical dependency relationship between construction processes. Specifically, it can be calculated by analyzing the in-degree or out-degree of nodes in the process network diagram. Its purpose is to ensure that key processes with high dependency are processed first and to maintain the continuity of construction logic.

[0144] Specifically, the fragment mapping relationship table refers to a data structure built based on the difference between the number of causal nodes and the process dependency. It can be implemented using a two-dimensional matrix or an association list. Its purpose is to dynamically align the intrinsic relationship between causal fragments and process fragments and eliminate the risk of fragment disconnection.

[0145] In addition, selecting data fragment groups sequentially can refer to selecting fragments from each fragment set according to the sequential index of the mapping table. Specifically, this can be achieved by traversing the rows or columns of the mapping table. The purpose is to seamlessly connect the cause, extension, process, and material fragments in a logical order to form a data fragment group with a rigorous structure.

[0146] Specifically, the proposed solution first sorts the causal fragments in ascending order by the number of nodes, prioritizing core causal fragments with fewer nodes to ensure the hierarchical and complete nature of the causal description. Second, it sorts the process fragments in descending order by dependency, prioritizing key processes with higher dependency to maintain the coherence of the construction logic. Then, it constructs a mapping table based on the difference between the number of nodes and dependency, using this difference as a matching metric to dynamically align causal fragments and process fragments, ensuring a precise correspondence between the causal chain and construction steps. Finally, it selects fragments sequentially from the causal fragment set, extended fragment set, process fragment set, and material fragment set according to the sequential index of the mapping table, forming a group of consecutively numbered data fragments. This quantitative sorting and dynamic mapping mechanism effectively avoids problems such as reversed processes, conflicting material parameters, or broken causal chains by systematically processing the inherent relationships between fragments, laying the foundation for generating logically consistent maintenance schemes.

[0147] As a specific embodiment, the solution of this application is implemented as follows: For the network crack disease of asphalt pavement, the number of causal nodes in the causal segment "cracks caused by freeze-thaw cycles" is calculated as a smaller value, and the number of nodes in "caused by aging of base materials" is calculated as a larger value. After sorting by the number of nodes in ascending order, the former is treated first. The process dependence of the process segment "clean the cracks first and then fill" is calculated as a higher value, and the dependence of "direct filling" is calculated as a lower value. After sorting by dependence in descending order, the process with high dependence is treated first. A mapping relationship table is constructed based on the difference between the number of nodes and the dependence. When the difference is in a small range, a matching relationship is established. Then, according to the mapping relationship table, the causal segment "caused by freeze-thaw cycles", the extension segment "cracks extend to the base layer", the process segment "clean and then fill" and the material segment "use epoxy resin" are selected in sequence to form a logically coherent data segment group.

[0148] The above technical solutions effectively solve the problem of inaccurate segment connection logic, ensuring that the construction process of data segment groups closely matches the actual relationship between the causes of diseases and construction procedures, avoiding errors in the sequence of procedures, mismatched material parameters, or broken causal chains, thereby improving the logical consistency and feasibility of subsequent maintenance plans.

[0149] The above-mentioned scheme proposes to construct data fragment groups to structure the input of historical maintenance data. However, in this process, the static organization of the fragment groups lacks a dynamic integration mechanism with the problem context, which leads to logical breaks between the cause segment, extension segment, process segment and material segment when generating maintenance schemes. Specifically, this manifests as incorrect process sequence dependence, disconnect between material parameters and disease characteristics, and lack of environmental condition adaptability, making it difficult to output a coherent scheme that meets the actual maintenance needs.

[0150] In this regard, this application further proposes that the preliminary maintenance plan generated in S5 includes:

[0151] The causal fragments, extended fragments, process fragments, and material fragments in the data fragment group are respectively formed into four input sequences according to their serial numbers;

[0152] The four input sequences are used as four auxiliary input segments of the generative model, and the multi-domain cross problem vector is used as the main input segment.

[0153] Insert the four input segments into the generated model as generation prompts according to their sequential positions;

[0154] The generated model outputs a preliminary maintenance plan, including the cause segment, extension segment, process segment, and material segment, in the order of insertion.

[0155] The generative model is an encoder-decoder model based on the Transformer architecture, which generates a preliminary textualized maintenance plan based on the input vector and data fragment groups.

[0156] The process of organizing the causal, extended, procedural, and material fragments from the data fragment group into four input sequences according to their numbers refers to organizing the structured fragments into a computable sequence based on logical temporal order. This can be achieved using serialization encoding techniques, such as converting text fragments into word embedding vector sequences, to ensure strict inheritance of causal order and procedural dependency logic. Using these four input sequences as four auxiliary input segments for the generative model, and using the multi-domain cross-question vector as the main input segment, means using the question context vector as the primary input source and the fragment sequences as auxiliary information flows. This can be configured as a multi-channel input structure for the model, aiming to enhance the accurate matching depth between question features and historical data. Finally, the four input segments are inserted into the generative model according to their numerical positions. The generation prompt refers to fixing the input order based on the temporal position of the segment group number. This can be achieved through prompting engineering methods, with the aim of ensuring that the model decoding process strictly follows the logical chain from cause to material. The generation model outputs a preliminary maintenance plan containing cause segment, extension segment, process segment, and material segment in the insertion order. This means that the output segment order is forced to be consistent with the input sequence order. This can be achieved through constrained decoding strategies, with the aim of ensuring the inherent consistency of the scheme segment structure. The generation model is an encoder-decoder model based on the Transformer architecture, which refers to a neural network architecture using a self-attention mechanism. It can be fine-tuned using pre-trained language models such as BART, with the aim of dynamically capturing the long-range dependencies between the problem vector and the segments.

[0157] Specifically, the solution in this application ensures that the input data strictly follows the causal order and process dependency logic by forming the data fragment groups into an input sequence according to their numbers; it combines the four input sequences as auxiliary input segments with the main input segment of the multi-domain cross-question vector, so that the context features of the question accurately guide the matching of historical data fragments; it fixes the temporal relationship between causes, extensions, processes and material fragments by inserting generation prompts according to their positions, so that the model decoding process inherits the complete logical chain; the generation model outputs according to the insertion order, forcing the paragraph structure to be consistent with the input order, avoiding process reversal or material misplacement; finally, the encoder-decoder model based on the Transformer architecture dynamically integrates the question vector and fragment group information through a self-attention mechanism to output a professional and coherent textual solution.

[0158] As a specific implementation method, the solution of this application is implemented as follows: The generation model can be specifically implemented as an encoder-decoder model based on the T5 architecture. The input sequence is formed by converting text fragments into word vector sequences. In the generation prompt, the cause fragment sequence is placed at the beginning of the prompt according to the sequence number. Then, the extended fragment sequence, the process fragment sequence, and the material fragment sequence are inserted in sequence. When the model decodes, it strictly generates the text content of the cause segment, extended segment, process segment, and material segment according to the insertion order.

[0159] Through the above technical solution, this application effectively solves the problem of logical break between the cause segment, extension segment, process segment and material segment when generating maintenance plan, ensuring the correct process sequence, matching of material parameters with disease characteristics and adaptation of environmental conditions, thereby outputting a coherent plan that meets actual maintenance needs.

[0160] Specifically, in some of the embodiments described above in this application, a preliminary curing plan containing material segments is proposed to describe the material parameters. However, in this process, the generated material parameters may not be consistent with the actual conditions, resulting in mismatched material strength, non-overlapping temperature windows, or incompatibility of construction equipment, making the curing plan impossible to implement.

[0161] In this regard, this application further proposes that the material parameter consistency check in S6 includes:

[0162] Compare the strength codes of each material in the preliminary maintenance plan with the corresponding strength codes of the materials in the valid data set;

[0163] Compare the overlap ratio between the temperature windows of each material in the preliminary maintenance plan and the temperature windows of the corresponding materials in the effective data set to see if they meet the preset overlap range.

[0164] Compare whether the construction equipment used in the preliminary maintenance plan is within the equipment compatibility range of the valid data set;

[0165] If all three comparisons meet the consistency criteria, then the material parameter consistency check is passed.

[0166] Among them, strength coding refers to the structured representation of material strength parameters, which can be achieved by combining minimum strength value, maximum strength value, and room temperature compressive strength. The purpose is to accurately quantify the material strength range to match the needs of disease treatment. Temperature window can be understood as the mechanism for defining the suitable temperature range for material construction. It can be achieved by using a triplet coding consisting of lower limit temperature, target temperature, and upper limit temperature. The purpose is to ensure the performance stability of the material under the temperature fluctuation sequence of the target range. Equipment compatibility range refers to the set of constraints on the applicable range of construction equipment. It can be achieved by combining a list of equipment types and operating parameter thresholds. The purpose is to verify the actual usability of the equipment in historical successful cases.

[0167] Specifically, the proposed solution employs a multi-dimensional parameter comparison mechanism to form a closed-loop verification logic. First, the strength coding comparison stage verifies the coverage of the material strength range in the preliminary solution against historical strength ranges in the valid data set, ensuring that the minimum strength required for the target defect is met and avoiding excessive strength. Second, the temperature window overlap ratio calculation stage assesses the matching degree between the construction temperature window and the actual environment based on the temperature fluctuation sequence of the target section, preventing material performance degradation due to temperature deviations. Third, the equipment compatibility verification stage eliminates the risk of equipment unavailability by comparing the construction equipment parameters with the equipment compatibility ranges of historical cases. These three verification stages are linked together using logical AND relationships. Material parameters are only considered consistent when all three comparison results simultaneously meet preset conditions, thus constructing a collaborative verification system encompassing strength, temperature, and equipment dimensions, avoiding the limitations of single-dimensional verification.

[0168] As a specific implementation method, the solution of this application is implemented as follows: For asphalt mixture materials, the strength code can be represented as a combination of minimum strength value, maximum strength value, and compressive strength at room temperature. The temperature window can be defined as a ternary structure including a lower limit temperature, a target temperature, and an upper limit temperature. The equipment compatibility range can include a list of paver models and the operating parameter range of compaction equipment. During the verification process, the system automatically extracts the material parameters from the preliminary maintenance plan, compares the strength code of the asphalt mixture with the strength codes of the corresponding materials in the valid dataset for interval coverage, calculates the overlap ratio between the temperature window and the temperature fluctuation sequence of the target section, and verifies whether the selected paver model is within the equipment compatibility range. Finally, the consistency of the material parameters is determined based on a comprehensive assessment of the three comparison results.

[0169] Through the above technical solution, this application effectively solves the obstacles to the implementation of maintenance schemes caused by the inconsistency between material parameters and actual conditions, ensuring that the material strength covers the needs of the defects, the temperature window adapts to environmental fluctuations, and the availability of construction equipment, thereby improving the feasibility and reliability of the maintenance scheme in complex pavement defect scenarios.

[0170] Specifically, in some of the embodiments described above in this application, a material parameter consistency check is proposed to ensure the matching of material parameters. However, in the process of its implementation, the logical consistency of the construction procedures is not fully checked, which may lead to problems such as incorrect procedure sequence, missing dependencies, or unreasonable time intervals in the generated maintenance plan, affecting the feasibility and logical integrity of the plan.

[0171] In response, this application further proposes that the process logic consistency check in S6 includes:

[0172] Calculate the sequential number of each construction procedure in the preliminary maintenance plan and compare it with the sequential numbering pattern of the corresponding procedure in the valid data set;

[0173] Compare whether the dependencies between procedures in the preliminary maintenance plan are complete and without duplication;

[0174] Calculate the time interval between adjacent processes in the preliminary maintenance plan and compare it with the corresponding time interval range in the valid data set;

[0175] When all three comparisons meet the consistency conditions, the process logic consistency check is performed.

[0176] Among them, sequence number comparison refers to calculating and comparing the execution sequence of construction procedures. This can be achieved using sequence similarity algorithms such as longest common subsequence or dynamic time warping. Its purpose is to verify the sequential logic of the generated scheme based on historically verified procedure execution sequence patterns, avoiding the problem of procedure reversal. Dependency comparison refers to verifying the completeness and uniqueness of logical dependencies between procedures. This can be achieved using dependency graph construction and topology sorting checks or logical rule engines. Its purpose is to ensure that there are no missing or redundant steps in the procedure chain, preventing process breakage. Time interval comparison refers to verifying whether the time difference between adjacent procedures is within a feasible range. This can be achieved using time window overlap calculation or sequence matching algorithms. Its purpose is to use historical engineering data to verify the rationality of the time arrangement and eliminate construction risks. Consistency condition judgment refers to confirming the pass when all verification dimensions meet the requirements. This can be achieved using logical AND operations or conditional gate circuits. Its purpose is to force the three elements of sequence, dependency, and time to simultaneously meet historical experience standards, avoiding the omission of potential vulnerabilities in single-dimensional verification.

[0177] Specifically, the proposed solution matches the execution sequence of the generated solution with the verification patterns in historical valid data through sequential number comparison, ensuring that the construction process conforms to actual engineering experience; it verifies the completeness of the logical chain of the generated solution based on the process dependency relationships recorded in the valid data through dependency comparison, ensuring tight connection between processes; it verifies the time arrangement of adjacent processes by using the actual feasible time constraints in historical data through time interval comparison, ensuring timeliness and feasibility; and it simultaneously satisfies the construction of a multi-dimensional joint verification mechanism through these three types of verification, so that the three core logical elements of sequence, dependency, and time work together to form a complete process logic verification system, thereby comprehensively ensuring the logical consistency and actual operability of the maintenance solution at the construction process level.

[0178] As a specific implementation method, the solution of this application is implemented as follows: In the scenario of road crack repair, the system first extracts the construction sequence number of the preliminary maintenance plan, such as the sequence of crack detection, damaged area cleaning, base layer repair, material filling and surface treatment, and compares it with the sequence pattern of similar cases in the effective database to confirm that the sequence conforms to historical experience; then, it checks the process dependencies, such as material filling must be performed only once after base layer repair to ensure that the dependency relationship is complete and without duplication; then, it calculates the time interval between adjacent processes, such as base layer repair and material filling, and compares it with the historical time range of the corresponding processes in the effective database to verify that the interval is within a reasonable range; finally, when all comparisons meet the consistency conditions, it confirms that the process logic consistency check has passed.

[0179] Through the above solution, this application effectively solves the problem of inconsistent construction procedures, ensuring that the sequence of procedures in the generated maintenance plan is correct, the dependencies are complete, and the time arrangement is reasonable, thereby improving the feasibility and logical integrity of the plan.

[0180] Specifically, in some of the embodiments described above in this application, an environmental condition consistency check is proposed to verify the suitability of the maintenance plan with the target environment. However, in its implementation, the existing check method only performs static comparison of single environmental factors such as temperature, rainfall or traffic load, ignoring the time-series dynamic change characteristics of multidimensional environmental parameters. This results in the inability to capture the sequence overlap difference between the actual road section environment and the plan requirements. The check results are one-sided and easily affected by local matching interference, making it difficult to fully guarantee the feasibility of the plan under complex climate and traffic conditions.

[0181] In this regard, this application further proposes that the environmental condition consistency check in S6 includes:

[0182] Extract construction temperature sequence, rainfall interval sequence, and traffic load sequence from the preliminary maintenance plan;

[0183] The three types of sequences are compared with the temperature sequence, rainfall sequence and traffic load sequence of the target road segment respectively, and the minimum value of the three types of overlap is calculated and denoted as the minimum overlap.

[0184] Compare the minimum overlap with the environmental matching range of the effective dataset;

[0185] When the minimum overlap is within the environmental matching range, environmental condition consistency is checked.

[0186] Among them, the construction temperature sequence refers to the time series data of dynamic temperature changes during the construction period preset in the maintenance plan. It can be achieved by interpolating historical meteorological data or collecting data from real-time monitoring equipment, with the aim of reflecting the temperature fluctuation patterns during the construction process. The rainfall interval sequence can be understood as the time interval sequence of rainfall events during the construction period. It can be constructed based on meteorological forecast models or regional rainfall records, with the aim of quantifying the impact of rainfall on the construction window period. The traffic load sequence is specifically a dynamic sequence of traffic flow intensity during the construction period, such as real-time traffic flow data obtained from roadside sensing equipment, with the aim of assessing the impact of traffic conditions on the construction window. The interference level of maintenance operations; sequence overlap comparison refers to the calculation process of quantifying the dynamic matching degree of two time series sequences, which can be achieved by dynamic time warping algorithm or sequence correlation coefficient analysis method, with the aim of accurately capturing the temporal correlation characteristics of environmental parameters; minimum overlap refers to the minimum value among the three types of overlap calculation results, which can be understood as the key bottleneck indicator of environmental adaptability, with the aim of focusing on the weakest dimension to ensure overall compatibility; environmental matching range is the threshold range of environmental parameters of historical successful cases in the effective data set, which can be determined based on statistical clustering method, with the aim of providing an objective verification benchmark that is deeply bound to engineering practice.

[0187] Specifically, the proposed solution extracts a multi-dimensional environmental sequence from the preliminary maintenance plan, dynamically compares it with the actual environmental sequence of the target road segment, calculates the sequence overlap, and takes the minimum value as the core indicator. This is then compared with the environmental matching range determined by historical experience, thereby achieving a systematic verification of the consistency of environmental conditions. The process first extracts construction temperature, rainfall interval, and traffic load sequences based on the dynamic change data of environmental parameters preset during the plan generation stage, providing input data closely related to actual construction for verification. Then, the overlap of these three sequences is compared with the corresponding environmental sequence of the target road segment. By quantifying the degree of dynamic temporal matching, the real-time status of temperature fluctuations, rainfall cycles, and traffic loads is accurately captured. Furthermore, the minimum value of the three overlaps is calculated as the minimum overlap, ensuring that the verification focuses on the weakest environmental dimension and preventing a good match of a single factor from masking the overall risk. Finally, the minimum overlap is compared with the environmental matching range of the effective dataset, introducing the adaptation threshold of historical successful cases as an objective benchmark, mandating that all key environmental dimensions meet the minimum matching requirements, thereby comprehensively verifying the feasibility of the solution in complex environments.

[0188] As a specific implementation method, in the process of generating a maintenance plan for a city's expressway, the system extracts the following from the preliminary maintenance plan: a construction temperature sequence showing continuous high-temperature construction windows; a rainfall interval sequence showing the construction period during the dry season with no rainfall interference; and a traffic load sequence showing the work arrangements during low-traffic periods at night. The three sequences are compared with the temperature, rainfall, and traffic load sequences monitored concurrently on the target road section for sequence overlap. The calculated overlap is 0.65 for temperature, 0.85 for rainfall, and 0.75 for traffic load, with a minimum overlap of 0.65. The effective environmental matching range of the dataset, determined through historical case statistics, is 0.70-0.95. Since the minimum overlap of 0.65 is below the lower threshold, the system determines that the environmental condition consistency check fails and constructs the temperature mismatch item as a constraint to re-optimize the plan.

[0189] Through the above technical solution, this application can systematically capture the temporal dynamic change characteristics of multidimensional environmental parameters, avoid the verification results being affected by local matching, ensure the feasibility of the maintenance plan under complex climate and traffic conditions, and effectively reduce the risk of construction failure due to environmental incompatibility.

[0190] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligently generating road maintenance plans based on retrieval enhancement, characterized in that, include: S1: Collect disease description text, disease image parameters, road structure level parameters, road section material parameters, traffic operation parameters, and climate fluctuation parameters; Based on lexical segmentation, syntactic dependency classification, and structural position encoding, a disease semantic matrix, a geometric feature matrix, a structural hierarchy matrix, and a load action sequence are generated, respectively. Based on the semantic principal axis in the disease semantic matrix, the damage direction vector in the geometric feature matrix, the layer index in the structural hierarchy matrix, and the peak sequence table in the load action sequence, a multi-domain cross problem vector is constructed. S2: Use multi-domain cross-question vectors to perform a coarse search on the historical maintenance database, filter candidate data entries based on the combined distance of semantic adjacency distance, structural adjacency distance and load adjacency distance, and perform a fine search based on disease stage information, cross-seasonal information and traffic level information to form a candidate data set; S3: For each data entry in the candidate dataset, perform an exclusion step based on the road structure layer thickness range, cumulative rainfall sequence, diurnal temperature range sequence, high temperature exposure duration sequence, completeness of the semantic chain of disease causes, and compatibility of material parameter groups to obtain the effective dataset; S4: Extract text fragments of disease causes, disease expansion paths, construction procedures, and material parameters from the effective data set to form a set of cause fragments, an expansion fragment set, a process fragment set, and a material fragment set, respectively. Establish fragment connection relationships based on the causal order, process order, and material usage order between fragments to construct data fragment groups; S5: Input the multi-domain cross problem vector and data fragment group into the generation model to generate a preliminary maintenance plan that includes the cause segment, extension segment, process segment and material segment; S6: Perform material parameter consistency verification, process logic consistency verification, and environmental condition consistency verification on the preliminary maintenance plan in sequence; If all consistency checks pass, the final maintenance plan will be output. If at least one fails, the corresponding inconsistency item will be constructed as a generation constraint and steps S5 and S6 will be re-executed until all consistency checks pass.

2. The method according to claim 1, characterized in that, The disease semantic matrix generated in S1 includes: The disease description text is divided into a sequence of semantic units according to the subject-verb-object structure; Calculate the semantic stability value for each semantic unit and sort them according to semantic stability to form a semantic axis; Calculate the gradient distribution at the edge of the damaged area based on the parameters of the disease image and generate the damage direction vector; Generate layer indexes corresponding to surface layer, base layer, subbase layer and subgrade based on road structure layer parameters; The semantic axis, damage direction vector, layer index, and load action sequence are concatenated according to their temporal positions to form a structured semantic code.

3. The method according to claim 1, characterized in that, The coarse search performed in S2 includes: For multi-domain intersection problems, vector normalization is performed, and the vectors are split into semantic vector subsets, structural vector subsets, and load vector subsets. Calculate the semantic adjacency distance, structural adjacency distance, and load adjacency distance respectively; Arrange the distance values ​​in ascending order to form a distance arrangement sequence; When the minimum adjacent distance corresponds to a semantic vector subset, match data entries whose difference between the number of semantic nodes and the target number of semantic nodes is within a preset range; when the minimum adjacent distance corresponds to a structural vector subset, match data entries whose difference between the number of structural levels and the target number of structural levels is within a preset range; when the minimum adjacent distance corresponds to a load vector subset, match data entries whose difference between the number of load peaks and the target number of load peaks is within a preset range. A coarse candidate set is formed based on the data entries that meet the criteria; When the number of coarse candidate sets exceeds the set range, data entries are removed in reverse order of the distance sorting sequence until the number of coarse candidate sets is within the threshold range.

4. The method according to claim 3, characterized in that, Performing a fine search in S2 includes: Extract the disease stage marker for each data entry in the coarse candidate set and calculate the stage difference with the target disease stage marker; Generate seasonal vectors from the temperature, rainfall, and diurnal temperature range sequences over three consecutive months; Compare seasonal vector differences and exclude data entries whose differences exceed the set range; Compare the differences in traffic levels and exclude data entries with inconsistent traffic levels; The remaining data items are sorted by stage difference, and the data items with the smallest stage difference are retained to form a candidate data set.

5. The method according to claim 4, characterized in that, The exclusion of material parameter groups in S3 includes: The minimum strength, maximum strength, and room temperature compressive strength of each material in the material combination are encoded as strength ranges; The construction temperature window is divided into a lower limit temperature, a target temperature, and an upper limit temperature, forming a ternary temperature window. Compare whether the intensity range covers the intensity range required for the target disease; if it does not, exclude the corresponding data entry. Compare the overlap ratio between the ternary temperature window and the temperature fluctuation sequence of the target segment. If the overlap ratio is less than the specified threshold, the corresponding data entry is excluded. Compare the compatibility range of construction equipment. If there are items in the material combination that are unavailable for equipment, then exclude the corresponding data entries.

6. The method according to claim 5, characterized in that, The construction of data fragment groups in S4 includes: Calculate the number of causal nodes for each causal segment and sort them by the number of nodes; Calculate the process dependency for each construction process segment and sort them by dependency. A segment mapping relationship table is constructed based on the difference between the number of causal nodes and the process dependence. Based on the fragment mapping relationship table, fragments from the causal fragment set, extended fragment set, process fragment set, and material fragment set are selected in sequence to form a data fragment group arranged by number.

7. The method according to claim 6, characterized in that, The initial maintenance plan generated in S5 includes: The causal fragments, extended fragments, process fragments, and material fragments in the data fragment group are respectively formed into four input sequences according to their serial numbers; The four input sequences are used as four auxiliary input segments of the generative model, and the multi-domain cross problem vector is used as the main input segment. Insert the four input segments into the generated model as generation prompts according to their sequential positions; The generated model outputs a preliminary maintenance plan, including the cause segment, extension segment, process segment, and material segment, in the order of insertion. The generative model is an encoder-decoder model based on the Transformer architecture, which generates a preliminary textualized maintenance plan based on the input vector and data fragment groups.

8. The method according to claim 7, characterized in that, The material parameter consistency check in S6 includes: Compare the strength codes of each material in the preliminary maintenance plan with the corresponding strength codes of the materials in the valid data set; Compare the overlap ratio between the temperature windows of each material in the preliminary maintenance plan and the temperature windows of the corresponding materials in the effective data set to see if they meet the preset overlap range. Compare whether the construction equipment used in the preliminary maintenance plan is within the equipment compatibility range of the valid data set; If all three comparisons meet the consistency criteria, then the material parameter consistency check is passed.

9. The method according to claim 8, characterized in that, The process logic consistency check in S6 includes: Calculate the sequential number of each construction procedure in the preliminary maintenance plan and compare it with the sequential numbering pattern of the corresponding procedure in the valid data set; Compare whether the dependencies between procedures in the preliminary maintenance plan are complete and without duplication; Calculate the time interval between adjacent processes in the preliminary maintenance plan and compare it with the corresponding time interval range in the valid data set; When all three comparisons meet the consistency conditions, the process logic consistency check is performed.

10. The method according to claim 9, characterized in that, The environmental condition consistency check in S6 includes: Extract construction temperature sequence, rainfall interval sequence, and traffic load sequence from the preliminary maintenance plan; The three types of sequences are compared with the temperature sequence, rainfall sequence and traffic load sequence of the target road segment respectively, and the minimum value of the three types of overlap is calculated and denoted as the minimum overlap. The minimum overlap is compared with the environmental matching range of the effective dataset; When the minimum overlap is within the environmental matching range, environmental condition consistency is checked.

Images