Building structure service period safety prediction method, system and device

Abstract

The invention provides a building structure service period safety prediction method, system and device, and relates to the technical field of building structure service period safety prediction. The method mainly comprises the steps of 1, collecting multi-modal data of a building structure, and inputting the multi-modal data into a large model for data processing and feature recognition to obtain key features related to the service period safety of the building structure; step 2, converting the key features into a standardized numerical value format through a conversion method to obtain feature vectors; and step 3, inputting the feature vector into a safety prediction model, and predicting the safety service period. According to the scheme, multi-source data can be automatically integrated in real time through a large model, and key features are identified; and the safe service period is objectively and reliably predicted through the safe prediction model.

Description

Technical Field

[0001] This invention relates to the field of building structure service life safety prediction technology, specifically a building structure service life safety prediction method, system and device. Background Technology

[0002] The service life of a building structure, also known as the design service life or lifespan, refers to the period during which a building structure can be used normally for its intended purpose under predetermined maintenance conditions.

[0003] Traditional methods for assessing the service life of building structures primarily rely on periodic manual inspections and experience-based judgments. While some quantitative methods have emerged, they generally suffer from the following drawbacks: Disadvantage 1: It relies heavily on historical experience data, and for new structures or structures in special environments, the lack of historical experience data leads to inaccurate assessments. Disadvantage 2: The data throughout the entire lifecycle of the structure is scattered across different units with different formats, making it difficult to integrate and utilize. Disadvantage 3: It heavily relies on the engineer's personal experience and subjective judgment; Disadvantage 4: The degradation model has too many simplification conditions and assumptions, and can only be evaluated statically. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, and apparatus for predicting the safety of building structures during their service life, so as to solve at least one technical problem existing in the background art.

[0005] In a first aspect, to achieve the above objectives, the present invention provides a method for predicting the safety of building structures during their service life, comprising the following steps: Step 1: Collect multimodal data of the building structure, input it into a large model for data processing and feature recognition, and obtain key features related to the safety of the building structure during its service life; Step 2: Convert the key features into a standardized numerical format using a transformation method to obtain the feature vector; Step 3: Input the feature vector into the safety prediction model to predict the safe service life.

[0006] In one feasible implementation, the multimodal data in step 1 includes text data, chart data, and formula data; The text data originates from testing and identification reports, industry standard documents, accident investigation reports, and laws and regulations, etc.; the data processing method for the text data includes natural language processing (NLP) methods using large models, including entity recognition, relation extraction, and semantic analysis, to obtain structured text information, including material properties, damage descriptions, and standard requirements; The chart data comes from line graphs, bar charts, and time series graphs in the testing and appraisal report, as well as design charts in industry standard documents; the data processing method for the chart data includes extracting data points, trend lines, and labels through large model image processing methods to obtain numerical data, including stress values, deformation amounts, and time series values. The formula data includes design formulas in industry standard documents and analysis formulas in accident investigation reports; the data processing method for the formula data includes identifying variables and constants through analytical formula methods of large models to obtain formula parameters, including stress, strain and safety factor.

[0007] In one feasible implementation, the key features include material characteristics, environmental characteristics, load characteristics, damage characteristics, and specification characteristics; The material characteristics include concrete compressive strength, steel bar yield strength, and elastic modulus, etc. The environmental characteristics include ambient temperature, relative humidity, chloride ion concentration, and carbonization coefficient, etc. The load characteristics include maximum load, number of load cycles, and stress range, etc. The damage characteristics include crack width, corrosion depth, deformation amount, and damage index, etc. The specified features include target reliability index, safety factor, and design life.

[0008] In one feasible implementation, the training data preprocessing optimization method for the large model specifically includes the following steps: Step a1: Collect multimodal raw data in the field of building structure service life. The raw data includes inspection and appraisal reports, industry standard documents, accident investigation reports and laws and regulations. The multimodal data includes three types of modalities: text, charts and formulas. Step a2: Use optical character recognition technology to identify and extract the raw data in step a1, perform preprocessing operations on the raw data to generate standardized data, and decompose and supplement the chart information using a three-level description system of basic layer, technical layer and application layer to build the original database. Step a3: Based on the standardized data generated in step a2, use Neo4j graph database technology to establish the interrelationships between building code clauses, form a metadata lineage network, and classify the data based on evaluation indicators to construct a metadata lineage graph database. Step a4: Perform quality filtering, redundancy removal, semantic enhancement, and structured representation processing on the data in the original database. Combine the metadata lineage from step a3 to extract keywords and generate text block-word vector pairs to construct a dedicated database for the service life of building structures. Step a5: Combine retrieval enhancement generation technology with a dedicated database of building structure service life and a metadata lineage graph database to optimize the training results of the large model; Step a6: Align image information with text information using multimodal fusion technology, design cross-modal cue learning templates, and enhance the semantic understanding of chart information by the large model.

[0009] Preferably, the optical character recognition technology in step a2 specifically includes: performing noise reduction on the image in the original data, performing region separation and character segmentation on the text, charts, and formulas, performing recognition on the segmented characters, performing spelling correction and format correction on the recognized characters, performing structured representation on the multimodal data and forming a preset data format, wherein the multimodal data includes text, charts, and formula information, and the preset data format is arranged according to serial number, problem description, problem photo, and rectification suggestions.

[0010] Preferably, the specific process of constructing the metadata lineage graph database in step a3 includes: based on historical data, theory and experience, clarifying the metadata lineage relationship between data information, associating various data information according to their mutual relationships, and dividing them into multiple indicators according to workers, operations and equipment to form a metadata lineage graph database.

[0011] Preferably, the specific process of constructing a dedicated database for the service life of building structures in step a4 includes: first, extracting the text information and problem description information from the original database; then, calling an open-source word segmentation model to achieve semantic segmentation, cutting the text into multiple text blocks to replace the original complete text paragraphs; extracting and verifying whether the text keywords contain illegal sensitive information based on metadata lineage; inputting the processed text blocks into the open-source word segmentation model to obtain the word segmentation vectors of the corresponding text blocks, forming text block-word segmentation vector pairs; and storing the formed text block-word segmentation vector pairs into a vector database to obtain a dedicated database for the service life of building structures.

[0012] Preferably, the retrieval enhancement generation technology in step a5 specifically involves: retrieving data information related to the input question from the proprietary database of the building structure service life field by combining sparse retrieval and dense retrieval, inputting the retrieval results into the generation model, and generating a coherent answer based on the retrieval content. The sparse retrieval is used to accurately match professional terms, and the dense retrieval is used to capture semantic similarity.

[0013] Preferably, the multimodal fusion technique in step a6 specifically involves: designing a cross-modal cue learning template based on the image-grounded text branch of the BLIP model. The cross-modal cue learning template includes three levels: basic description, technical specifications, and engineering decisions. Using all text query features in the data as context, the attention scores between the fusion features and each text query feature are calculated. The context information is adaptively expanded to enhance the fusion features. The correlation between image features and text features is calculated through an attention mechanism to generate enhanced fusion features. The contribution of image and text modalities is balanced using a learnable gated query function. The updated feature chain after balancing is calculated first, and then the contribution of text query features in the fusion features is quantified through the gated query function, thereby achieving the alignment processing of multimodal data.

[0014] Preferably, the establishment of the interrelationships in step a3 is based on expert experience and historical data, and the interrelationships include reference relationships, supplementary relationships, and substitution relationships.

[0015] Preferably, the evaluation indicators in step a3 are classified according to worker activities, operations, and equipment, and each category of evaluation indicators is divided into several sub-indicators and weighted accordingly.

[0016] Preferably, the open-source word segmentation model adopts one or more of the jieba model and the HanLP model.

[0017] Preferably, the retrieval algorithm of the retrieval enhancement generation technology adopts one or more of the TF-IDF algorithm and the Dense PassageRetrieval algorithm.

[0018] In one feasible implementation, the conversion method in step 2 specifically includes unit conversion, missing value filling, and normalization processing.

[0019] In one feasible implementation, the safe service period in step 3 refers to the time from the current time until the structural reliability reaches the allowable threshold.

[0020] In one feasible implementation, the security prediction model includes a reliability index model, a degradation model, and a safe service life model; The reliability index model is used to assess the structural safety level based on stress-intensity interference theory. The specific formulas include: ; in, Indicates the reliability index; This represents the mean resistance (e.g., bending moment capacity), in units of 1. or ; It represents the standard deviation of resistance, used to reflect the uncertainty of resistance; and It can be calculated from material characteristics and damage characteristics; This represents the mean value of the load effect (such as the maximum bending moment), in units of... or ; Indicates the standard deviation of the loading effect; and It can be calculated from load characteristics and environmental characteristics; The degradation model includes a concrete carbonation model and a steel corrosion model; The specific calculation formulas for the concrete carbonation model include: ; in, Indicates carbonization depth, in millimeters; The carbonization coefficient is represented and can be estimated from environmental characteristics; Indicates time, in years; The specific calculation formulas for the steel corrosion model include: ; in, Indicates the corrosion depth, in millimeters; This represents the corrosion rate, which can be estimated from environmental characteristics.

[0021] In one feasible implementation, the specific calculation formula for the safe service life model includes: ; in, Indicates the remaining safe service life; Indicates the current reliability index; The target reliability index can be obtained from the specification features; The reliability index represents the rate of degradation, which can be obtained from the degradation model.

[0022] In one feasible implementation, the safe service life model further includes a damage basis model, the specific calculation formula of which includes: ; in, This represents the critical damage value, which can be obtained from the standard features; This indicates the current damage value, which can be obtained from the damage characteristics; The damage rate is represented and can be obtained from the degradation model.

[0023] In one feasible implementation, step 3, when predicting the safe service life, obtains the confidence interval of the safe service life using a probabilistic method (such as Monte Carlo simulation), specifically including: Step b1: Define a probability distribution for each feature vector; Step b2: Randomly select a value from the probability distribution of each feature vector to form random training data; Step b3: Input the random training data into the safety prediction model to calculate the safe service life; Step b4: Iterate through step b2 until the preset iteration termination condition is met to obtain random training results; Step b5: Perform statistical analysis on the random training results to obtain the confidence interval for the safe service life.

[0024] In one feasible implementation, the confidence interval is a 90% confidence interval, with the lower limit being the value of the 5th percentile and the upper limit being the value of the 95th percentile.

[0025] Secondly, based on the same inventive concept, this application also provides a building structure service life safety prediction system, including a data acquisition module, a data processing module and a result generation module; The data acquisition module is used to collect multimodal data of the building structure; The data processing module includes an identification unit, a conversion unit, and a prediction unit; The identification unit is used to input the multimodal data into a large model for data processing and feature recognition to obtain key features related to the safety of the building structure during its service life. The conversion unit is used to convert key features into a standardized numerical format through a conversion method to obtain a feature vector; The prediction unit inputs the feature vector into the safety prediction model to predict the safe service life. The result generation module is used to distribute the safe service period externally.

[0026] Thirdly, based on the same inventive concept, this application also provides a building structure service life safety prediction device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the building structure service life safety prediction method as described above. The bus connects the functional components for transmitting information.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: This application provides a method, system, and device for predicting the service life safety of building structures, which can integrate multi-source data in real time and automatically through a large model to identify key features; and objectively and reliably predict the safe service life through a safety prediction model.

[0028] When training large models, this approach improves the accuracy of data semantic understanding through standardized processing of multimodal data and a three-level description system, ensuring that the model can capture complex information in the field of building structure service life. By using optical character recognition technology combined with noise reduction, region separation, and spell correction operations, it effectively solves noise and format problems in the original data, improves the reliability of data extraction, constructs a metadata lineage network, clarifies the reference, supplement, and substitution relationships between normative clauses, and enables the model to generate logically rigorous output results based on expert experience and historical data. Through retrieval-enhanced generation technology, the model can accurately match professional terms and capture semantic similarity. By adopting cross-modal prompting learning templates and attention mechanisms, it achieves effective alignment of image and text information, enhances the model's semantic understanding of charts and formulas, and improves the performance of large models. Attached Figure Description

[0029] Figure 1 A flowchart of a method for predicting the safety of a building structure during its service life, provided by an embodiment of the present invention; Figure 2 A flowchart of the training data preprocessing optimization method for a large model provided in an embodiment of the present invention; Figure 3 This is an interaction diagram of the optical character recognition technology and the retrieval enhancement generation technology provided in the embodiments of the present invention; Figure 4 A flowchart of the probability method provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0031] It should also be noted that the specific embodiments or implementation methods described below are a series of optimized settings listed by the present invention to further explain the specific content of the invention, and these settings can be combined or used in conjunction with each other.

[0032] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for predicting the safety of a building structure during its service life, including the following steps: Step 1: (Collect multimodal data of building structure by web crawling or manual input) Input the data into a large model for data processing and feature recognition to obtain key features related to the safety of building structure during service life; Step 2: Convert the key features into a standardized numerical format using a transformation method to obtain the feature vector; Step 3: Input the feature vector into the safety prediction model to predict the safe service life.

[0033] Furthermore, the multimodal data in step 1 includes text data, chart data, and formula data; The text data originates from testing and identification reports, industry standard documents, accident investigation reports, and laws and regulations, etc.; the data processing method for the text data includes natural language processing (NLP) methods using large models, including entity recognition, relation extraction, and semantic analysis, to obtain structured text information, including material properties, damage descriptions, and standard requirements; The chart data comes from line graphs, bar charts, and time series graphs in the testing and appraisal report, as well as design charts in industry standard documents; the data processing method for the chart data includes extracting data points, trend lines, and labels through large model image processing methods to obtain numerical data, including stress values, deformation amounts, and time series values. The formula data includes design formulas in industry standard documents and analysis formulas in accident investigation reports; the data processing method for the formula data includes identifying variables and constants through analytical formula methods of large models to obtain formula parameters, including stress, strain and safety factor.

[0034] Furthermore, the key features include material characteristics, environmental characteristics, load characteristics, damage characteristics, and specification characteristics; The material characteristics include concrete compressive strength, steel bar yield strength, and elastic modulus, etc. The environmental characteristics include ambient temperature, relative humidity, chloride ion concentration, and carbonization coefficient, etc. The load characteristics include maximum load, number of load cycles, and stress range, etc. The damage characteristics include crack width, corrosion depth, deformation amount, and damage index, etc. The specified features include target reliability index, safety factor, and design life.

[0035] In this embodiment, as Figure 2-3 As shown, the training data preprocessing optimization method for the large model specifically includes the following steps: Step a1: Collect multimodal raw data in the field of building structure service life. The raw data includes inspection and appraisal reports, industry standard documents, accident investigation reports and laws and regulations. The multimodal data includes three types of modalities: text, charts and formulas. Specifically, the inspection and appraisal reports include periodic inspection reports, safety assessment reports, and damage assessment reports for building structures, covering common problems such as concrete cracks, steel corrosion, and foundation settlement. Industry standard documents include national standards (GB), industry standards (JGJ), and local standards, such as the "Standard for Identification of Dangerous Buildings" (JGJ125-2016) and the "Code for Design of Building Structures" (GB 50009). Accident investigation reports are used to record detailed investigation reports of building structure accidents, including accident causes, responsibility analysis, and rectification measures, such as structural assessments after building collapses or fires. Laws and regulations include legal provisions and policy documents related to building safety, such as the "Regulations on the Administration of Construction Project Quality" and the "Regulations on the Administration of Dangerous Buildings in Cities." Textual data includes textual descriptions, standard clauses, and legal provisions in the reports, ensuring the completeness and readability of the text. Chart data includes structural design drawings, crack distribution diagrams, and load distribution tables, clearly indicating technical parameters and engineering background. Formula data includes mathematical formulas involving structural mechanics and material performance calculations, such as differential equations and matrix operations, which must be saved in standard formats such as LaTeX. Step a2: Use optical character recognition technology to identify and extract the raw data in step a1, perform preprocessing operations on the raw data to generate standardized data, and decompose and supplement the chart information using a three-level description system of basic layer, technical layer and application layer to build the original database. Specifically, Optical Character Recognition (OCR) technology is a technique that converts text in images (such as text in scanned documents or photos) into editable and searchable digital text. Utilizing OCR's region detection algorithm, an attention-based fractional recognition module is developed to establish a detection model for special mathematical characters. Symbolic calculations are performed and mathematical formulas are recognized through a computational software interface. The computational software can be one or more of Wolfram Alpha, Mathematica, or Matlab. A standardized typesetting format is automatically generated using LaTeX tools. An example is shown below. The derivation is as follows: ,in The differential equations involved in the service life analysis of building structures, and the detection model is a typical mathematical model describing the vibration characteristics of structures, which is the existing technology; This is the general solution, representing the structural displacement response; Indicates position coordinates; Indicates the wavenumber parameter; Indicate boundary conditions; The optical character recognition technology specifically involves: denoising the image in the original data by using median filtering, Gaussian filtering, or deep learning denoising algorithms to eliminate noise, stains, shadows, and other interference in the scanned document, thereby improving image quality; performing region separation and character segmentation on text, charts, and formulas; recognizing the segmented characters; correcting spelling errors and formatting the recognized characters; and structurally representing multimodal data and forming a preset data format. The multimodal data includes text, charts, and formula information, and the preset data format is arranged according to serial number, problem description, problem photo, and rectification suggestions. Specifically, the base layer is used to objectively describe the information in the charts, for example: a tensile crack with a width greater than 1.0 mm appears on the slab; the technical layer is used to associate the information with standards and regulations, for example: it meets the requirements of Article 5.4.3 of JGJ125-2016; the application layer is used to describe the engineering significance of the information in the charts, for example: the width of the crack exceeds the requirements of the "Standard for Identification of Dangerous Buildings" JGJ125-2016, and it is assessed as a dangerous component. Compared with traditional OCR technology, this solution semantically interprets the information in the charts through a three-level description system, improving the accuracy of understanding large models. Step a3: Based on the standardized data generated in step a2, use Neo4j graph database technology to establish the interrelationships between building code clauses, form a metadata lineage network, and classify the data based on evaluation indicators to construct a metadata lineage graph database. Specifically, Neo4j is a graph-based database that stores data using nodes (entities), edges (relationships), and attributes. The specific process for constructing a metadata lineage graph database includes: based on historical data, theory, and experience, clarifying the metadata lineage relationships between data information. Metadata lineage relationships are a graph structure network based on clause references, supplements, and substitutions. Nodes represent specification clauses, such as clause 5.4.3 of JGJ 125-2016, and edges represent mutual relationships, such as GB 50009 → Reference → JGJ 125. Metadata lineage relationships support dynamic updates. When new standard clauses are added, the system automatically links historical data and, through manual confirmation of relationship weights, associates various data information according to their interrelationships. Simultaneously, it categorizes data by worker, operation, and equipment using multiple indicators, forming a metadata lineage relationship database. The establishment of interrelationships is based on expert experience and historical data, including citation, supplementary, and substitution relationships. An example is: GB50108-2008 Clause 3.1.2 → Citation → GB55030-2022 Clause 3.3.10. Evaluation indicators are categorized according to worker activities, operations, and equipment. Each category of evaluation indicators is further divided into several sub-indicators with weighted classifications, constructed based on the principle of "full process, full chain, and hierarchical classification." Multiple evaluation indicators, taking the inherent risk of operation in the work indicators as an example, inherent risk of operation = A+B+C+D+E+F+G+H+I+J+K+L, where inherent risk is the sum of the scores of each item, and the weights can be selected according to the needs of the implementation environment. In this example, the weight of each item is set to 1. Item A is the key work activity, item B is the number of high-impact risk activities involved in the work activities, item C is the sensitive work activity, item D is the non-routine work activity, item E is the back-office work activity, item F is the top 5 safety risks identified by the project team, item G is the proportion of subcontractor work activities with a work completion rate of ≥80%, item H is the construction plan, item I is the foundation pit construction operation, item J is the cross-operation of multiple trades, item K is the impact of the work environment, and item L is the risk of the work time period. Step a4: Perform quality filtering, redundancy removal, semantic enhancement, and structured representation processing on the data in the original database. Combine the metadata lineage from step a3 to extract keywords and generate text block-word vector pairs to construct a domain-specific database for the service life of building structures. Specifically, the process of constructing a dedicated database for the service life of building structures includes: first, extracting text information and problem description information from the original database; then, calling an open-source word segmentation model to perform semantic segmentation, cutting the text into multiple text blocks to replace the original complete text paragraphs; extracting and verifying whether there is any illegal sensitive information in the text keywords based on the metadata lineage; inputting the processed text blocks into the open-source word segmentation model to obtain the word segmentation vectors of the corresponding text blocks, forming text block-word segmentation vector pairs; and storing the formed text block-word segmentation vector pairs into a vector database to obtain a dedicated database for the service life of building structures. The open-source word segmentation model uses one or more of the jieba model and the HanLP model. Specifically, quality filtering is used to remove incomplete, erroneous, or low-quality data; redundancy removal is used to eliminate duplicate or highly similar data, such as identical rectification suggestions in multiple reports, merging duplicate descriptions of "beam support cracks" in two inspection reports; semantic enhancement is used to supplement the implicit semantics of text through domain dictionaries or context analysis, for example, expanding "GB 50009" to "Code for Design of Building Structures" GB 50009-2012; and structured representation is used to convert unstructured text into a machine-processable format, for example, converting the text "crack width 1.2mm" into a structured field: {defect type: crack, width (mm): 1.2}; A specific process example is as follows: The input data is a test report: "The width of the diagonal crack in the beam support of a certain factory building is 1.5mm, which exceeds the limit (0.3mm) of Article 5.4.3 of the 'Standard for Identification of Dangerous Buildings' (JGJ 125-2016), and immediate reinforcement is required." This text information is subjected to quality filtering and semantic segmentation. After semantic segmentation, two text blocks are obtained: text block 1 is "the width of the diagonal crack in the beam support is 1.5mm", and text block 2 is "exceeds the limit of Article 5.4.3 of JGJ 125-2016". Keywords are extracted based on the metadata lineage. Keywords such as "diagonal crack", "limit" and "dangerous component" are extracted from JGJ 125-2016 5.4.3. Fixed-dimensional word segmentation vectors are generated to capture the semantic features of the text and obtain the word segmentation vectors of the corresponding text blocks, forming text block-word segmentation vector pairs. The formed text block-word segmentation vector pairs are stored in a vector database to obtain a dedicated database for the service life of building structures. This database can quickly respond to retrieval requests. Step a5: Combine retrieval enhancement generation technology with a dedicated database of building structure service life and a metadata lineage graph database to optimize the training results of the large model; Specifically, the retrieval-enhanced generation technique involves combining sparse and dense retrieval methods to retrieve data related to the input question from a specialized database on the service life of building structures. The retrieval results are then input into a generative model to generate a coherent answer based on the retrieved content. Keyword retrieval or semantic retrieval is used, allowing sparse retrieval to precisely match technical terms and dense retrieval to capture semantic similarity. The retrieval algorithm employs one or more of the TF-IDF and Dense Passage Retrieval algorithms, and the generation model is one of the GPT-3, BERT, or T5 models. Figure 2 As shown, the application process of Retrieval Enhanced Generation (RAG) technology in the field of building structure service life is demonstrated, showing the complete processing flow from multi-source data input to structured output; A specific process example is as follows: A user asks a question such as "How should I handle the diagonal cracks in a concrete beam?". The system uses sparse retrieval (TF-IDF algorithm) to precisely match technical terms (such as "diagonal crack" and "concrete beam"), returning relevant standard clauses: "JGJ125-2016 5.4.3: Diagonal cracks wider than 0.3mm require immediate reinforcement." Then, dense retrieval (Dense PassageRetrieval algorithm) captures semantically similar content (such as "shear crack" and "support crack"), returning similar cases: "Case A: A factory beam has a 1.2mm diagonal crack at the support, which is reinforced with carbon fiber cloth." The sparse and dense retrieval results are merged, sorted by relevance, and input into a generative model (such as GPT-3). Using the retrieval results as context, a coherent answer is generated: "According to Article 5.4.3 of the 'Standard for Identification of Dangerous Buildings' (JGJ 125-2016), beams with diagonal cracks wider than 0.3mm require immediate reinforcement. It is recommended to use carbon fiber cloth for reinforcement (refer to Case A), and verify the stress condition of the support." Step a6: Align image information with text information using multimodal fusion technology, design cross-modal cue learning templates, and enhance the semantic understanding of chart information by the large model; Specifically, based on the image-grounded text branch of the BLIP model (Bootstrapped Language-Image Pre-training), a cross-modal prompt learning template is designed. This template includes three levels: basic description, technical specifications, and engineering decisions. All text query features in the data are used as context. The attention scores between the fused features and each text query feature are calculated. The context information is adaptively expanded to enhance the fused features. The correlation between image features and text features is calculated through an attention mechanism to generate enhanced fused features. The contribution of image and text modalities is balanced using a learnable gated query function. The updated feature chain after balancing is calculated first, and then the contribution of text query features in the fused features is quantified through the gated query function, thereby achieving the alignment processing of multimodal data. The gated query function is a weight allocation model based on a gating mechanism. The specific operation is as follows: given the fusion features of the image-grounded text branch. Text query features of self-attention modules Utilizing batch data volume All dimensions are The query features are used as context to enhance the calculation. Attention score between each text query feature Calculated using the formula ,in, This refers to the batch size, i.e., the number of samples processed at one time. For feature dimensions; Indicates the first One fusion feature For the Text query features The higher the attention weight, the stronger the semantic association between the fused feature and the text query feature, and the more information the fused feature will absorb from the text query feature to enhance itself. Used to calculate the One fusion feature With the Text query features Dot product similarity; Used to calculate batch data volume Features of all batch text queries ( (representing batch indexes) and fusion features The dot product result can be used to improve the semantic quality of multimodal data fusion, providing higher-quality features for large model training or inference. Examples of cross-modal cue learning templates are shown in Table 1: Table 1

[0036] Taking a certain region's building structure construction drawing review AI system as an example, the model, optimized by a domain-specific database and metadata lineage graph database for the service life of building structures, is installed in the system. The input image information is that a concrete column has vertical cracks, peeling of the protective layer, and exposed and corroded main reinforcement. The output cross-modal prompt learning template is shown in Table 2. Table 2

[0037] Based on the above cross-modal prompting learning template, the suggested treatment is as follows: first reinforce the structure, then unload the supports, clear the area of ​​personnel, set up warning signs, and monitor settlement, stress, and deformation. Conduct testing and assessment, analyze the cause of the column's compression cracking, guide subsequent reinforcement design and construction, and hire a qualified design and construction unit to handle the situation. The retrieved data was compared with the traditional BLIP model, as shown in Table 3: Table 3

[0038] In summary, this method improves the semantic understanding accuracy of charts and formulas by employing OCR technology combined with a three-level description system. The metadata lineage network, based on the Neo4j graph database, constructs reference, supplement, and substitution relationships between standard clauses, enhancing the logical rigor of the model. The professionalism of the model output is optimized through retrieval-enhanced generation technology. Efficient alignment of images and text is achieved through the BLIP model and cross-modal prompting learning templates, thereby improving the reliability and practicality in scheme analysis. It is suitable for professional scenarios such as construction problem diagnosis and has good application prospects.

[0039] Furthermore, the conversion method in step 2 specifically includes conventional unit conversion, missing value imputation, and normalization processing.

[0040] Furthermore, the safe service period in step 3 refers to the time from the current time until the structural reliability reaches the allowable threshold.

[0041] Furthermore, the safety prediction model includes a reliability index model, a degradation model, and a safe service life model; The reliability index model is used to assess the structural safety level based on stress-intensity interference theory. The specific formulas include: ; in, Indicates the reliability index; This represents the mean resistance (e.g., bending moment capacity), in units of 1. or ; It represents the standard deviation of resistance, used to reflect the uncertainty of resistance; and It can be calculated from conventional material characteristics and damage characteristics; This represents the mean value of the load effect (such as the maximum bending moment), in units of... or ; Indicates the standard deviation of the loading effect; and It can be calculated from conventional load characteristics and environmental characteristics; The degradation model includes a concrete carbonation model and a steel corrosion model; The specific calculation formulas for the concrete carbonation model include: ; in, Indicates carbonization depth, in millimeters; Indicates time, in years; The carbonization coefficient, which can be estimated from environmental characteristics, includes: ; in, The reference carbonization coefficient is expressed as follows: ; This represents an empirical constant, which can take the value 3. This represents the standard value of the compressive strength of a concrete cube, expressed in megapascals (MPa). The carbon dioxide concentration influence coefficient reflects the impact of carbon dioxide in the environment on the carbonization rate. Its specific expression is: ; Indicates the volumetric concentration of carbon dioxide in the ambient environment; Indicates the baseline concentration; The temperature influence coefficient reflects the effect of ambient temperature on the carbon dioxide diffusion rate. Its specific expression includes: ; This represents the average absolute temperature of the environment. Indicates the reference absolute temperature; It represents the apparent activation energy, used to reflect the sensitivity of the carbonization reaction to temperature; Represents the ideal gas constant; The relative humidity influence coefficient is represented by the following formulas: ; in, This indicates the average relative humidity of the environment; The specific calculation formulas for the steel corrosion model include: ; in, Indicates the corrosion depth, in millimeters; The corrosion rate, which can be estimated from environmental characteristics, includes: ; in, Indicates the rate of deep corrosion, measured in millimeters per year; This indicates the concentration of free chloride ions in the pore water of concrete. The humidity function can be expressed in the form of: ; All represent coefficients, which can be obtained from industry standard documents.

[0042] Furthermore, the specific calculation formula for the safe service life model includes: ; in, Indicates the remaining safe service life; Indicates the current reliability index; This represents the target reliability index, which can be obtained from the specification features and can take a value of 3. The reliability index represents the rate of degradation, which can be obtained from the degradation model.

[0043] Furthermore, the safe service life model also includes a damage basis model, the specific calculation formula of which includes: ; in, The critical damage value (such as the maximum allowable crack width) can be obtained from the specification features; This indicates the current damage value, which can be obtained from the damage characteristics; The damage rate can be represented by a degradation model (e.g., for steel corrosion, a value can be taken as...). ).

[0044] Furthermore, such as Figure 4 As shown, in step 3, when predicting the safe service life, the confidence interval of the safe service life is obtained through a probabilistic method (such as Monte Carlo simulation), specifically including: Step b1: Define a probability distribution for each feature vector (e.g., material strength follows a normal distribution, and steel corrosion rate follows a log-normal distribution). Step b2: Randomly select a value from the probability distribution of each feature vector to form random training data; Step b3: Input the random training data into the safety prediction model to calculate the safe service life; Step b4: Iterate through step b2 until the preset iteration termination condition is met to obtain random training results; Step b5: Perform statistical analysis on the random training results to obtain the confidence interval for the safe service life.

[0045] Furthermore, the confidence interval is a 90% confidence interval, with the lower limit being the value of the 5th percentile (i.e., there is a 5% probability that the lifespan is shorter than this value, which can be regarded as a pessimistic estimate), and the upper limit being the value of the 95th percentile (i.e., there is a 95% probability that the lifespan is shorter than this value, which can be regarded as an optimistic estimate). For example, if there is a 90% probability that the predicted safe service life will fall between 18 and 35 years, 18 years can be taken as the final safe service life to improve the prediction safety.

[0046] Example 2: This embodiment provides a building structure service life safety prediction system, including a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect multimodal data of the building structure; The data processing module includes an identification unit, a conversion unit, and a prediction unit; The identification unit is used to input the multimodal data into a large model for data processing and feature recognition to obtain key features related to the safety of the building structure during its service life. The conversion unit is used to convert key features into a standardized numerical format through a conversion method to obtain a feature vector; The prediction unit inputs the feature vector into the safety prediction model to predict the safe service life. The result generation module is used to distribute the safe service period externally.

[0047] Example 3: This embodiment provides a building structure service life safety prediction device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the building structure service life safety prediction method as described above. The bus connects the various functional components for information transmission.

[0048] In another embodiment, this solution can also be implemented using an integrated device, which may include corresponding modules that perform one or more steps in the various embodiments described above. A module may be one or more hardware modules specifically configured to perform the corresponding step, or implemented by a processor configured to perform the corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented through some combination thereof.

[0049] The processor executes the various methods and processes described above. For example, the method implementations in this scheme can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some implementations, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other implementations, the processor can be configured to execute one of the methods described above by any other suitable means (e.g., by means of firmware).

[0050] This device can be implemented using a bus architecture. A bus architecture can include any number of interconnect buses and bridges, depending on the specific application of the hardware and overall design constraints. The bus connects various circuits, including one or more processors, memory, and / or hardware modules. The bus can also connect various other circuits such as peripherals, voltage regulators, power management circuitry, external antennas, etc.

[0051] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Component (EISA) buses, etc. Buses can be divided into address buses, data buses, control buses, etc.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the safety of building structures during their service life, characterized in that, include: Step 1: Collect multimodal data of the building structure, input it into a large model for data processing and feature recognition, and obtain key features related to the safety of the building structure during its service life; Step 2: Convert the key features into a standardized numerical format using a transformation method to obtain the feature vector; Step 3: Input the feature vector into the safety prediction model to predict the safe service life.

2. The prediction method according to claim 1, characterized in that, The training data preprocessing optimization method for the large model includes: Step a1: Collect multimodal raw data in the field of building structure service life. The raw data includes inspection and appraisal reports, industry standard documents, accident investigation reports and laws and regulations. The multimodal data includes three types of modalities: text, charts and formulas. Step a2: Use optical character recognition technology to identify and extract the raw data in step a1, perform preprocessing operations on the raw data to generate standardized data, and decompose and supplement the chart information using a three-level description system of basic layer, technical layer and application layer to build the original database. Step a3: Based on the standardized data generated in step a2, use Neo4j graph database technology to establish the interrelationships between building code clauses, form a metadata lineage network, and classify the data based on evaluation indicators to construct a metadata lineage graph database. Step a4: Perform quality filtering, redundancy removal, semantic enhancement, and structured representation processing on the data in the original database. Combine the metadata lineage from step a3 to extract keywords and generate text block-word vector pairs to construct a dedicated database for the service life of building structures. Step a5: Combine retrieval enhancement generation technology with a dedicated database of building structure service life and a metadata lineage graph database to optimize the training results of the large model; Step a6: Align image information with text information using multimodal fusion technology, design cross-modal cue learning templates, and enhance the semantic understanding of chart information by the large model.

3. The prediction method according to claim 2, characterized in that, The optical character recognition technology in step a2 specifically involves: performing noise reduction on the images in the original data; performing region separation and character segmentation on the text, charts, and formulas; performing recognition on the segmented characters; performing spelling correction and format correction on the recognized characters; and performing structured representation on the multimodal data to form a preset data format. The multimodal data includes text, charts, and formula information, and the preset data format is arranged according to serial number, problem description, problem photo, and rectification suggestions.

4. The prediction method according to claim 2, characterized in that, The specific process of constructing the metadata lineage graph database in step a3 includes: based on historical data, theory and experience, clarifying the metadata lineage relationship between data information, associating various data information according to their mutual relationships, and dividing them into multiple indicators according to workers, operations and equipment to form a metadata lineage graph database.

5. The prediction method according to claim 2, characterized in that, The specific process of constructing a dedicated database for the service life of building structures in step a4 includes: first, extracting the text information and problem description information from the original database; then, calling an open-source word segmentation model to achieve semantic segmentation, cutting the text into multiple text blocks to replace the original complete text paragraphs; extracting and verifying whether there is any illegal sensitive information in the text keywords based on the metadata lineage; inputting the processed text blocks into the open-source word segmentation model to obtain the word segmentation vectors of the corresponding text blocks, forming text block-word segmentation vector pairs; and storing the formed text block-word segmentation vector pairs into a vector database to obtain a dedicated database for the service life of building structures.

6. The prediction method according to claim 2, characterized in that, The retrieval enhancement generation technique in step a5 specifically involves: retrieving data information related to the input question from the proprietary database of the building structure service life field by combining sparse retrieval and dense retrieval, inputting the retrieval results into the generation model, and generating a coherent answer based on the retrieval content. The sparse retrieval is used to accurately match professional terms, and the dense retrieval is used to capture semantic similarity.

7. The prediction method according to claim 2, characterized in that, The multimodal fusion technique in step a6 is as follows: Based on the image-grounded text branch of the BLIP model, a cross-modal prompt learning template is designed. The cross-modal prompt learning template includes three levels: basic description, technical specifications, and engineering decisions. All text query features in the data are used as context. The attention scores between the fusion features and each text query feature are calculated. The context information is adaptively expanded to enhance the fusion features. The correlation between image features and text features is calculated through the attention mechanism to generate enhanced fusion features. The gatedquery function is used to balance the contributions of image and text modalities. The updated feature chain after balance is calculated first, and then the contribution of text query features in the fusion features is quantified through the gatedquery function.

8. The prediction method according to claim 1, characterized in that, The security prediction model includes a reliability index model, a degradation model, and a safe service life model; The reliability index model includes the following specific formulas: ； in, Indicates the reliability index; Indicates the mean resistance; Indicates the standard deviation of resistance; This represents the mean of the load effect; Indicates the standard deviation of the loading effect; The degradation model includes a concrete carbonation model and a steel corrosion model; The specific calculation formulas for the concrete carbonation model include: ； in, Indicates the depth of carbonization; Indicates the carbonization coefficient; Indicates time; The specific calculation formulas for the steel corrosion model include: ； in, Indicates the depth of corrosion; Indicates the corrosion rate; The specific calculation formulas for the safe service life model include: ； in, Indicates the remaining safe service life; Indicates the current reliability index; Indicates the target reliability index; This indicates the rate of reliability degradation.

9. A building structure service life safety prediction system employing the prediction method described in any one of claims 1-8, characterized in that, It includes a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect multimodal data of the building structure; The data processing module includes an identification unit, a conversion unit, and a prediction unit; The identification unit is used to input the multimodal data into a large model for data processing and feature recognition to obtain key features related to the safety of the building structure during its service life. The conversion unit is used to convert key features into a standardized numerical format through a conversion method to obtain a feature vector; The prediction unit inputs the feature vector into the safety prediction model to predict the safe service life. The result generation module is used to distribute the safe service period externally.

10. A building structure service life safety prediction device, characterized in that, It includes a processor, a memory, and a bus. The memory stores instructions and data read by the processor. The processor is used to call the instructions and data in the memory to execute the prediction method as described in any one of claims 1-8. The bus connects the functional components for transmitting information.

Images