An agent-based roof quality determination method, system and intelligent terminal

Abstract

The application discloses a roof quality determination method and system based on an intelligent agent and an intelligent terminal, and the method comprises the following steps: constructing a roof quality rule base and dividing the roof quality rule base into A-type rules and B-type rules; acquiring and expanding a roof IFC model; IFC model analysis and feature dimension reduction; constructing a cascade prompt word chain; intelligent semantic matching and task scheduling; rule engine deterministic determination; A-type rule automatic inspection and B-type rule manual auxiliary inspection collaborative output; result summarization and report generation. The application can consider the roof quality automatic determination of the AI large model semantic understanding ability and the rule engine deterministic execution ability, so that the AI large model is only responsible for rule intention recognition, component semantic matching and function scheduling, and the rule engine completes traceable and reproducible numerical and logical determination, thereby improving the accuracy, efficiency and standard traceability of roof quality inspection.

Description

Technical Field

[0001] This invention relates to the field of building engineering quality inspection, and in particular to a method, system and intelligent terminal for judging roof quality based on intelligent agents. Background Technology

[0002] Roof quality acceptance inspection involves a variety of items, including slope, waterproofing layer thickness, roll material overlap width, flashing termination height, rainwater inlet connection, vent spacing, insulation layer thickness, detailed joint construction, and surface appearance quality. Traditional inspection methods typically rely on construction personnel, quality inspectors, or supervisors to verify each item against the specifications, combining paper drawings, BIM, on-site measurement records, and experience to form an inspection conclusion.

[0003] Existing technologies also include solutions for using BIM in engineering quality inspection, but these solutions often employ hard-coded rules or fixed scripts, directly writing quality specifications into the software source code. This approach suffers from problems such as difficulty in rule maintenance, high cost of specification updates, difficulty in aligning model semantics with rule semantics, matching failures due to differences in component naming, and insufficient ability to determine complex spatial relationships.

[0004] In recent years, large-scale AI models have demonstrated strong capabilities in understanding regulatory clauses, semantic extraction, rule reconstruction, and natural language interaction. However, in roof quality assessment scenarios, directly outputting "qualified" or "unqualified" conclusions by large-scale AI models is susceptible to the influence of generative model probabilistic prediction mechanisms, leading to uncontrollable errors in numerical calculations, threshold comparisons, spatial geometric relationship judgments, and symbolic reasoning. For example, for rigid indicators such as "flashover termination height not less than 250mm" and "structural slope not less than 3%", large-scale AI models may give incorrect conclusions due to contextual omissions, unit misunderstandings, or reasoning illusions. Summary of the Invention

[0005] The purpose of this invention is to provide a roof quality assessment method, system, and intelligent terminal based on intelligent agents, in order to solve the problems of existing roof quality inspections relying on human experience, difficulties in maintaining BIM automatic review rules, and inaccurate judgment results due to the risk of illusion caused by direct judgment by large AI models.

[0006] To address the aforementioned technical problems, this invention provides a roof quality assessment method based on an intelligent agent, comprising:

[0007] Step 1: Construct a roof quality rule library and divide it into Class A rules and Class B rules;

[0008] Step 2: Obtain and expand the roof IFC model;

[0009] Step 3: IFC model analysis and feature dimensionality reduction;

[0010] Step 4: Construct a cascading suggestion chain;

[0011] Step 5: Intelligent semantic matching and task scheduling;

[0012] Step 6: Rule engine deterministic determination;

[0013] Step 7: Collaborative output of automated check for Category A rules and manual check for Category B rules;

[0014] Step 8: Results summary and report generation.

[0015] Furthermore, the roof quality assessment method based on intelligent agents provided by this invention, in step 1, includes the following method for constructing a roof quality rule base and dividing it into Class A rules and Class B rules:

[0016] Step 1.1: Obtain roof quality acceptance specifications, enterprise standards, project-specific quality requirements, and historical inspection experience to form a set of original rules and regulations;

[0017] Step 1.2: Input each original rule in the original rule text set into the AI ​​large model or manual review module, and extract the rule identifier, rule text, inspection type, severity, main target entity, secondary target entity, attributes, operators, threshold, unit, allowable error and inspection logic fields of each original rule to form a structured semantic rule text;

[0018] Step 1.3: Divide the structured semantic rule texts into Class A rules and Class B rules, where Class A rules are automated judgment rules and Class B rules are manually assisted inspection rules;

[0019] Step 1.4: Save the structured semantic rules after classification as a JSON file to form a roof quality rule library that can be dynamically read, maintained, and expanded.

[0020] Furthermore, in the roof quality assessment method based on intelligent agents provided by this invention, step 3, the method for IFC model parsing and feature dimensionality reduction includes:

[0021] Step 3.1: Analyze the IFC model to obtain the roof components and their attribute sets;

[0022] Step 3.2: Generate an object filter based on the structured semantic rule base, retaining only the components, attributes and topological relationships related to roof quality inspection, and removing redundant geometric coordinates, display styles and complex nested relationships that are difficult for large AI models to process directly and are irrelevant to rule judgment;

[0023] Step 3.3: Expand the multilayer material structure relationship in the original IFC model into lightweight key-value pairs to form an IFC property flow;

[0024] Step 3.4: Divide the large roof IFC attribute flow model into blocks;

[0025] Step 3.5: Output a standardized IFC attribute stream in JSON format.

[0026] Furthermore, in the roof quality determination method based on intelligent agents provided by the present invention, step 4, the method for constructing a cascading prompt word chain includes:

[0027] Step 4.1: Construct the first layer of prompts, which are role and task constraint prompts, used to limit the AI ​​large model's identity as a roofing project quality supervision expert, a data analysis expert, and a rule engine scheduling expert;

[0028] Step 4.2: Construct the second layer of prompt words, which are prompt words for rule intent recognition, used to identify rule intent and extract rule elements;

[0029] Step 4.3: Construct the third-level prompt words, which are component semantic mapping prompt words, used to map the fuzzy component names in the IFC model to the standard IFC component types or the standard types defined in the project;

[0030] Step 4.4: Construct the fourth layer of prompt words, which are function call prompt words used to convert rule checking logic into rule engine function handles and parameter packages;

[0031] Step 4.5: Construct the fifth layer of prompt words, which are result summary prompt words. This is used to allow only the structured execution results returned by the rule engine to be summarized, sorted and reported in a language-based manner, and to prevent modification of the underlying function's judgment results.

[0032] Furthermore, in the roof quality assessment method based on intelligent agents provided by the present invention, step 5, the intelligent semantic matching and task scheduling method includes:

[0033] Step 5.1: Read the rules to be checked from the rule base through the intelligent agent and establish a check task queue;

[0034] Step 5.2: For each Class A rule, the agent performs semantic retrieval and candidate component localization in the IFC attribute stream to determine the target component set;

[0035] Step 5.3: The agent performs integrity verification on the key attributes of the target component; if the required attribute exists, a function call is initiated; if the attribute is missing but can be calculated from the associated component, a derived attribute calculation task is generated; if the attribute cannot be obtained, a data missing warning is generated.

[0036] Step 5.4: The agent introduces topological environment awareness, incorporating surrounding components and their relative positions into the matching context;

[0037] Step 5.5: When multiple rules apply to the same component or there are multiple candidate matching results for the same component, the agent calculates the intent weight based on the rule severity level, entity confidence, attribute completeness and context consistency, and prioritizes scheduling tasks with high weights.

[0038] Furthermore, in the roof quality determination method based on intelligent agents provided by the present invention, step 6, the method of deterministic determination by the rule engine includes:

[0039] Step 6.1: The rule engine reads the function call parameter package generated by the agent and calls the preset function library according to the function name;

[0040] Step 6.2: For basic logic class rules, call the basic logic class function to determine the relationship between attribute values ​​and thresholds;

[0041] Step 6.3: For spatial geometry rules, call the spatial geometry function to calculate the intersection, net distance, elevation difference, and slope between entities;

[0042] Step 6.4: For topology relationship rules, call the topology relationship function to determine the component connection status, hierarchical stacking order, and drainage link relationship;

[0043] Step 6.5: For engineering-specific rules, call the specific calculation functions to complete specific quality checks such as waterproofing, insulation, drainage, and flashing;

[0044] Step 6.6: The rule engine outputs the judgment result using a deterministic algorithm.

[0045] Furthermore, in the roof quality assessment method based on intelligent agents provided by the present invention, step 7, the method for collaborative output of automated inspection of type A rules and manual assisted inspection of type B rules includes:

[0046] Step 7.1: For Class A rules, the agent automatically generates non-compliance record; when the measured value does not meet the requirement value or the deviation exceeds the allowable error, it is recorded as a non-compliance item; when key data is missing but the rule is detectable, it is recorded as a warning item.

[0047] Step 7.2: For Class B rules, the agent generates structured manual-assisted inspection guidelines, including inspection objects, floors / areas, component locations, inspection points, photo requirements, evidence upload fields, and review suggestions;

[0048] Step 7.3: For photos, measured data or verification results uploaded on-site, the intelligent agent binds them with unit identifiers and rule identifiers to form traceable quality closed-loop data.

[0049] Furthermore, in the roof quality assessment method based on intelligent agents provided by the present invention, step 8, the method for summarizing results and generating a report includes:

[0050] Step 8.1: Based on the results returned by the rule engine, the agent sorts and summarizes the non-compliant items according to severity, floor, roof zone, component type, and rule identifier;

[0051] Step 8.2: Generate a report on non-compliance of roof quality. The report should include at least the project name, model version, inspection time, inspection rule version, unit identifier, component name, rule clause, measured value, required value, deviation value, severity, and rectification suggestions.

[0052] Step 8.3: Generate a manual for manual inspection, which will be used by on-site personnel to review Category B rules;

[0053] Step 8.4: Write the inspection results back to the BIM or project quality management platform, mark the inspection status on the corresponding components, and realize visual positioning, problem tracking and rectification closed loop.

[0054] To address the aforementioned technical problems, this invention also provides a roof quality assessment system based on an intelligent agent, employing the aforementioned roof quality assessment method based on an intelligent agent, comprising:

[0055] The data input layer is used to input the expanded IFC model of the roof to be inspected, to input prompt words, and to upload on-site photos of Class B rules;

[0056] The model preprocessing layer is used to parse and reduce the dimensionality of the input IFC model.

[0057] The rule knowledge layer is used to execute the construction of the roof quality rule base;

[0058] The cascading prompt chain layer is used to perform the construction of the cascading prompt chain;

[0059] The intelligent scheduling layer is used to perform intelligent semantic matching and task scheduling;

[0060] The rule engine execution layer is used to perform deterministic decisions by the rule engine.

[0061] The results output layer is used to perform collaborative output of Class A automated inspections and Class B manual assisted inspections, as well as to summarize the results and generate reports.

[0062] To address the aforementioned technical problems, the present invention also provides an intelligent terminal for judging roof quality based on intelligent agents, which is embedded with the aforementioned method for judging roof quality based on intelligent agents.

[0063] Compared with existing technologies, the beneficial effects of the roof quality assessment method, system, and smart terminal based on intelligent agents provided by this invention are as follows:

[0064] Improve judgment accuracy: By decoupling semantic understanding from logical execution, the large AI model does not directly output a numerically qualified conclusion. Instead, the rule engine performs deterministic calculations, which significantly reduces the risk of misjudgment caused by the illusion of generative artificial intelligence.

[0065] Improve inspection efficiency: By using an IFC parser and feature dimensionality reduction, the massive IFC model is transformed into a lightweight attribute stream, reducing the interference of irrelevant geometric information on the reasoning process and improving the speed of rule matching and inspection execution.

[0066] Improve rule maintainability: Rules are stored independently in JSON structured files. When the rules change, there is no need to modify the underlying program code. Only the rule base needs to be updated to dynamically generate the judgment logic.

[0067] Improve model adaptability: By using agent semantic mapping and component name keyword extraction, it can adapt to component naming differences caused by different projects, different modeling habits, and different IFC export software.

[0068] Improve the traceability of results: Each judgment retains the rule identifier, unit identifier, measured value, required value and function execution path, which facilitates quality review, responsibility identification and rectification closure.

[0069] Balancing automation and manual verification: Automated judgment of hard indicators such as slope, thickness, height, and spacing, while outputting manual inspection manuals for soft indicators such as visual quality, avoiding the agent's overreach in judging matters that cannot be directly proven by the model.

[0070] A scalable platform is formed: the rule base, prompt word chain, agent scheduler, and rule function library can all be extended to other engineering quality inspection scenarios such as exterior walls, basements, decoration and renovation, and electromechanical installation. Attached Figure Description

[0071] Figure 1 This is a flowchart of an automated roof quality assessment method based on intelligent agents;

[0072] Figure 2 This is a structural diagram of constructing a cascading prompt word chain and a brief illustrative explanation thereof;

[0073] Figure 3 This is a structural diagram illustrating the construction of Class A rules, the rule engine function library, and the function call mapping layer;

[0074] Figure 4 This is a block diagram of an automated roof quality assessment system based on intelligent agents. Detailed Implementation

[0075] The present invention will now be described in detail with reference to the accompanying drawings. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0076] Please refer to Figures 1 to 3 This invention provides an automated roof quality assessment method based on an intelligent agent, comprising:

[0077] Step 1: Build a roof quality rule library.

[0078] Step 1.1: Obtain roof quality acceptance specifications, enterprise standards, project-specific quality requirements, and historical inspection experience to form a set of original rules and regulations.

[0079] Step 1.2: Input each original rule in the original rule text set into the AI ​​large model or manual review module, and extract the rule identifier (RuleID), rule text (RuleText), check type (CheckType), severity (Severity), main target entity (MainTargetEntity), secondary target entity (SecondaryTargetEntity), property (Property), operator (Operator), threshold (Threshold), unit (Unit), allowabletolerance (AllowableTolerance), and check logic (CheckLogic) fields of each original rule to form a structured semantic rule text.

[0080] Step 1.3: Based on the degree of recognition of Building Information Modeling (BIM), the structured semantic rules can be divided into Class A rules and Class B rules.

[0081] Among them, Category A rules are automated judgment rules, including check items such as slope, thickness, height, overlap width, spacing, and connection status, which can be calculated from the attributes or geometric relationships of the IFC model.

[0082] Among them, Category B rules are manual auxiliary inspection rules, including inspection items such as surface flatness, hollowness, cracks, contamination, and damage that require on-site sensory or image-assisted confirmation.

[0083] Step 1.4: Save the structured semantic rules after categorization as a JSON file to form a dynamically readable, maintainable, and scalable roof quality rule library. JSON is a lightweight data exchange format.

[0084] Step 2: Obtain and expand the roof IFC model.

[0085] Step 2.1: In the roof modeling or data integration stage, establish the roof components and their spatial topological relationships, including adjacent, intersecting, containing, connecting, supporting, hierarchical stacking, upstream and downstream drainage relationships, etc., and establish the BIM model of the roof to be inspected to provide a data basis for subsequent spatial geometry and topological relationship determination.

[0086] Step 2.2: Establish a binary attribute representation for the key quality parameters of the roof components, including design value and actual value. For example, the overlap width of the waterproof layer is recorded simultaneously as both the design overlap width and the actual overlap width measured on site, and the thickness of the insulation layer is recorded simultaneously as both the design thickness and the actual thickness.

[0087] Step 2.3 involves semantic extension of components in the BIM model, such as roof panels, waterproofing layers, insulation layers, slope-finding layers, leveling layers, parapet walls, rainwater outlets, roof penetration pipes, and vent holes, to give the BIM model a unified structured semantic representation of component categories, component names, component identifiers (GUIDs), spatial locations, material levels, design values, and measured values.

[0088] Step 2.5: Obtain the BIM model of the roof and export it as an IFC format file to form an IFC model, which is also called an IFC model file or IFC file.

[0089] Step 3: IFC model analysis and feature dimensionality reduction.

[0090] Step 3.1: Call the IFC parser to read the IFC model and parse the components (i.e., entities) and their attribute sets, such as roof (IfcRoof), floor slab (IfcSlab), covering layer (IfcCovering), material layer (IfcMaterialLayer), rainwater outlet (IfcDrain), pipe (IfcPipe), and building component unit (IfcBuildingElementProxy).

[0091] Step 3.2: Generate an object filter based on the main target entity (MainTargetEntity), secondary target entity (SecondaryTargetEntity), and property fields in the structured semantic rule base. Only retain components, attributes, and topological relationships related to roof quality inspection, and remove redundant geometric coordinates, display styles, and complex nested relationships that are difficult for large AI models to process directly and are irrelevant to rule determination.

[0092] Step 3.3: Expand the multilayer material structure relationships in the original IFC model into lightweight key-value pairs to form a lightweight property flow IFC model, i.e., reduce the dimensionality of the original IFC model. The lightweight property flow IFC model is also known as IFC property flow. Figure 1 The pure component attribute flow is used to represent the IFC attribute flow. For example, the multi-layer material structure layer (IfcMaterialLayerSetUsage) relationship is transformed into an IFC attribute flow of "layer name - material - design thickness - measured thickness - structure layer number".

[0093] Step 3.4: Divide the large roof model into blocks. Divide the IFC attribute stream into multiple data blocks according to floors, roof zones, structural layers, component types, or context window sizes, and create a block address identifier (BlockID) and component index table for each data block. By dividing the IFC attribute stream into multiple data blocks, the data volume of the original IFC model is prevented from exceeding the AI ​​large model inference limit. The large roof IFC attribute stream model refers to a building information model whose data volume exceeds the AI ​​large model inference limit.

[0094] Step 3.5: Output a standardized IFC property stream JSON format file, including the following fields: Element ID, EntityType, Name, PropertySet, GeometrySummary, TopologySummary, DesignValue, ActualValue, and SourceLocation.

[0095] Step 4: Construct a cascading chain of prompts. Please refer to [link / reference]. Figure 2 Specifically, it includes:

[0096] Step 4.1: Construct the first layer of prompts L1, namely role and task constraint prompts, to limit the AI ​​large model's identity to roofing project quality supervision expert, BIM / IFC data analysis expert and rule engine scheduling expert, and prohibit it from directly making the final qualification judgment on the numerical results.

[0097] Step 4.2: Construct the second layer of prompts (L2), namely the rule intent recognition prompts, which are used to identify the rule intent and extract rule elements, such as extracting components (i.e., entities), properties, operation logic, and thresholds.

[0098] Step 4.3: Construct the third-level prompt term L3, namely the component semantic mapping prompt term, which is used to map the fuzzy component names in the IFC model to the standard IFC component types or the standard types defined in the project. For example, map "parapet wall waterproof termination" and "rolled material flashing node" to the finishing layer (IfcCovering) or the corresponding custom type.

[0099] Step 4.4: Construct the fourth-level prompt (L4), i.e., the function call prompt, which is used to convert the rule checking logic into a rule engine function handle and parameter package. The parameter package includes the target, secondary target, field, operator, threshold, unit, allowableTolerance, and severity.

[0100] Step 4.5: Construct the fifth-level prompt word L5, which is the result summary prompt word. It is used to allow only the structured execution results returned by the rule engine to be summarized, sorted, grouped and reported in a language-based manner, and to prevent modification of the underlying function's judgment results.

[0101] Step 5: Intelligent semantic matching and task scheduling.

[0102] Step 5.1: Read the rules to be checked from the rule base through the agent, and establish a queue of check tasks according to severity, check type and component category.

[0103] Step 5.2: For each Class A rule, the Agent performs semantic retrieval and candidate component location in the IFC attribute stream, and determines the target component set by combining component type (EntityType), component name (Name), attribute set (PropertySet), topology summary (TopologySummary) and historical mapping table.

[0104] Step 5.3: The Agent performs integrity verification on the key attributes of the target component. If the required attribute exists, a function call is initiated; if the attribute is missing but can be calculated from related components, a derived attribute calculation task is generated; if the attribute cannot be obtained, a data missing warning is generated.

[0105] Step 5.4: The Agent introduces topology environment awareness, incorporating surrounding components and their relative positions into the matching context. For example, flashing height checks require simultaneously identifying the waterproofing layer termination location and the roof finish elevation, while rainwater inlet connection checks require simultaneously identifying the connection relationship between the rainwater inlet (IfcDrain) and the pipe (IfcPipe).

[0106] Step 5.5: When multiple rules apply to the same component or there are multiple candidate matching results for the same component, the Agent calculates the intent weight based on the severity of the rule, entity confidence, attribute completeness, and contextual consistency, and prioritizes scheduling tasks with high weights.

[0107] Step 6: Rule engine deterministic determination. Please refer to [link / reference]. Figure 1 and Figure 3 Specifically, it includes:

[0108] Step 6.1: The rule engine reads the function call parameter package generated by the Agent and calls the preset function library according to the function name.

[0109] Step 6.2: For basic logic rules, call the basic logic functions such as CompareAttribute, GreaterThanOrEquals, LessThanOrEquals, Equals, and Contains to determine the relationship between the attribute value and the threshold. Among them, CompareAttribute, GreaterThanOrEquals, LessThanOrEquals, and Equals functions perform the size determination, and the Contains function reflects the completeness.

[0110] Step 6.3: For spatial geometry rules, call spatial geometry functions such as Intersect, GetVerticalDistance, GetHorizontalDistance, GetSlope, and GetClearance to calculate the intersection, net distance, height difference, and slope between components.

[0111] Step 6.4: For topology relationship rules, call topology relationship functions such as IsConnected, IsAdjacent, VerifyLayerOrder, and IsWithinDrainagePath to determine the component connection status, layer stacking order, and drainage link relationship.

[0112] Step 6.5: For engineering-specific rules, call the specific calculation functions such as VerifyMaterial, VerifyOverlapWidth, VerifyUpstandHeight, VerifyInsulationThickness, and VerifyExhaustHoleSpacing to complete specific quality checks for waterproofing, insulation, drainage, and flashing.

[0113] Step 6.6: The rule engine outputs a judgment result of pass (PASS), fail (FAIL), warning (WARNING), or manual verification (MANUAL_CHECK) based on the deterministic algorithm, and returns the element ID, rule ID, actual value, required value, deviation, severity, and evidence path.

[0114] Step 7: Collaborate between automated checks of Category A rules and manual checks of Category B rules to output results.

[0115] Step 7.1: For Class A rules, the Agent automatically generates non-compliance records. When ActualValue does not meet RequiredValue or the deviation exceeds AllowableTolerance, it is recorded as FAIL; when critical data is missing but the rule is detectable, it is recorded as WARNING.

[0116] Step 7.2: For Category B rules, the Agent does not make a direct pass / fail determination, but generates a structured manual auxiliary inspection guide, including the inspection object, floor / area, component location, inspection points, photo requirements, evidence upload fields, and review suggestions.

[0117] Step 7.3: For photos, measured data or verification results uploaded on-site, Agent binds them with the original ElementID and RuleID to form a traceable quality closed-loop data.

[0118] Step 8: Results summary and report generation.

[0119] Step 8.1: The Agent sorts and summarizes the FAIL items according to Severity, floor, roof zone, component type and RuleID based on the results returned by the rule engine, which is the LLM result summary.

[0120] Step 8.2: Generate a report on non-compliance of roof quality. The report should include at least the project name, model version, inspection time, inspection rule version, element ID, component name, rule clause, measured value, required value, deviation value, severity, and rectification suggestions.

[0121] Step 8.3: Generate a manual for manual inspection, which is used by on-site personnel to review Category B rules.

[0122] Step 8.4: Write the inspection results back to the BIM or project quality management platform, mark the inspection status on the corresponding components, and realize visual positioning, problem tracking and rectification closed loop.

[0123] Please refer to Figure 4 This invention also provides an automated roof quality assessment system based on an intelligent agent, comprising:

[0124] The data input layer is used to input the expanded IFC model of the roof to be inspected, to input prompt words, and to upload on-site photos of Class B rules;

[0125] The model preprocessing layer is used to parse and reduce the dimensionality of the input IFC model.

[0126] The rule knowledge layer is used to execute the construction of the roof quality rule base;

[0127] The cascading prompt chain layer is used to perform the construction of the cascading prompt chain;

[0128] The intelligent scheduling layer is used to perform intelligent semantic matching and task scheduling;

[0129] The rule engine execution layer is used to perform deterministic decisions by the rule engine.

[0130] The results output layer is used to perform collaborative output of Class A automated inspections and Class B manual assisted inspections, as well as to summarize the results and generate reports.

[0131] The following example, using the pre-inspection of the roofing project of a public building as a case study, provides an illustrative explanation.

[0132] Step 101: After the detailed design of the roof project is completed, obtain the roof BIM model with the file name RoofModel.rvt, and confirm that the BIM includes components such as roof panels, slope layer, leveling layer, insulation layer, waterproof layer, protective layer, parapet wall, rainwater outlet, vent hole and pipe through the roof.

[0133] Step 102: Export the BIM file as an IFC file to form an IFC model, which can be denoted as Roof.ifc; at the same time, obtain the project's actual measurement data table Roof_Actual.xlsx, which records the on-site measured data such as the overlap width of the waterproof layer, the thickness of the insulation layer, the height of the flashing, and the spacing of the vent holes.

[0134] Step 103: In the model data integration module, the measured data table is matched back to the IFC model attribute set according to ElementID or component name to form an extended IFC model containing DesignValue and ActualValue.

[0135] Step 104: Read the roof quality semantic rule library RoofRule.json. The rule library includes Class A rules and Class B rules, for example: RuleID=GB50207-5.2.1-A, the inspection item is the insulation layer thickness deviation; RuleID=GB50207-6.2.7-A, the inspection item is the flashing termination height; RuleID=GB50207-4.1.3-B, the inspection item is the surface appearance of the waterproof layer.

[0136] Step 105: The Agent starts the IFC parser, extracts the entities (and components) of IfcRoof, IfcSlab, IfcCovering, IfcMaterialLayer, IfcDrain, IfcPipe and IfcBuildingElementProxy, and establishes an entity (i.e. component) index.

[0137] Step 106: The agent reduces the dimensionality of the model based on the target entity filter in the semantic rule base, retaining only the attributes required for quality checks, such as Slope, Thickness, Height, OverlapWidth, Coordinates, Connectivity, MaterialName, and LayerOrder.

[0138] Step 107: The agent divides the parsing results into several attribute stream data blocks. Each data block includes no more than the number of components in the preset context window, and establishes a mapping table between BlockID and ElementID.

[0139] Step 108: The agent invokes the first-level prompt, sets the agent as a roofing project quality supervision expert, and explicitly prohibits the agent from directly outputting numerical conclusions of "qualified" or "unqualified".

[0140] Step 109: The agent invokes the second-layer prompts to perform intent recognition for each rule. For example, for "The slope of the structure should not be less than 3%", the agent identifies Entity=IfcSlab or slope layer, Property=Slope, Operator=>=, Threshold=3%, and Function=VerifySlope.

[0141] Step 110: The agent invokes the third-layer prompts to perform semantic mapping on the model components. For example, components whose names contain "parapet flashing", "waterproof termination", or "roll-up membrane" are mapped as candidate target components for flashing height inspection.

[0142] Step 111: The agent retrieves the attribute stream, locates the target component set, and verifies the attribute completeness of the candidate components. If a component lacks UpstandHeightActual but has a waterproof layer termination elevation and a roof finishing surface elevation, a derived calculation task is generated to calculate UpstandHeightActual based on the height difference between the two.

[0143] Step 112: The agent invokes the fourth-level prompt to generate a function call parameter package. For example: Function=VerifyUpstandHeight, target=IfcCovering#E1025, secondary_target=IfcSlab#R001, field=UpstandHeightActual, operator=>=, threshold=250, unit=mm, severity=High.

[0144] Step 113: The rule engine executes the VerifyUpstandHeight function to read the waterproofing layer termination elevation and the roof finishing surface elevation, calculates the actual flashing height, and compares it with the 250mm threshold. If the actual height is 230mm, it returns FAIL with Deviation = -20mm.

[0145] Step 114: Check the insulation layer thickness. The rule engine filters IfcMaterialLayer entities whose names contain "insulation", reads LayerThicknessDesigned and LayerThicknessActual, and calculates whether it exceeds the allowable negative deviation according to the rule AllowableTolerance=5%.

[0146] Step 115: Check the drainage slope. The rule engine extracts the coordinates of feature points on the surface of the slope-finding layer, calculates the slope aspect and elevation difference, and calls the GetSlope function to obtain the actual slope. If the actual slope is less than the rule threshold, FAIL is output.

[0147] Step 116: Check the connection status of the rainwater inlet. The rule engine reads the topology relationship between IfcDrain and IfcPipe and calls the IsConnected function to verify the connection link. If the connection relationship is missing, WARNING or FAIL is output.

[0148] Step 117: For Category B rules, the Agent does not perform automatic compliance determination, but instead generates on-site inspection guidelines. For example: "Please inspect the appearance of the parapet wall waterproofing layer in Area A of the third floor of the roof, confirm whether there are any hollow areas, cracks, damage or contamination, and upload on-site photos."

[0149] Step 118: The rule engine returns all A-class rule execution results to the result summary module, including PASS, FAIL, WARNING, and items requiring manual inspection.

[0150] Step 119: Under the constraint of the fifth layer of prompt words, the AI ​​big model only performs language summarization on the structured results and generates a report on non-compliance items in roof quality. The report lists the problems in order of severity: High, Medium, and Low.

[0151] Step 120: The Agent writes the inspection results back to the BIM model or quality management platform, marks non-conforming components in red, marks warning components in yellow, and creates a closed-loop rectification log.

[0152] This invention provides a smart terminal for an automated roof quality assessment method based on intelligent agents. This smart terminal includes, but is not limited to, inspection robots and tablet computers embedded with the automated roof quality assessment method based on intelligent agents.

[0153] This invention provides an automated roof quality assessment method and its intelligent terminal based on an intelligent agent. It uses an IFC parser to reduce the complexity of the IFC model to a pure attribute stream required for quality inspection. A structured semantic rule base expresses roof engineering quality specifications. A cascading prompt word chain drives an intelligent agent to automatically match "rules—components—attributes—functions." A deterministic rule engine performs numerical comparisons, spatial geometry, topological relationships, and engineering-specific judgments, ultimately generating a non-compliance report and a manual-assisted inspection manual. This automated roof quality assessment method balances the semantic understanding capabilities of a large AI model with the deterministic execution capabilities of a rule engine. The AI ​​model is only responsible for rule intent recognition, component semantic matching, and function scheduling, while the rule engine performs traceable and reproducible numerical and logical judgments, thereby improving the accuracy, efficiency, and traceability of roof quality inspections. The resulting technical effects are as follows:

[0154] Improve judgment accuracy: By decoupling semantic understanding from logical execution, the large AI model does not directly output a numerically qualified conclusion. Instead, the rule engine performs deterministic calculations, which significantly reduces the risk of misjudgment caused by the illusion of generative artificial intelligence.

[0155] Improve inspection efficiency: By using an IFC parser and feature dimensionality reduction, the massive IFC model is transformed into a lightweight attribute stream, reducing the interference of irrelevant geometric information on the reasoning process and improving the speed of rule matching and inspection execution.

[0156] Improve rule maintainability: Rules are stored independently in JSON structured files. When the rules change, there is no need to modify the underlying program code. Only the rule base needs to be updated to dynamically generate the judgment logic.

[0157] Improve model adaptability: By using agent semantic mapping and component name keyword extraction, it can adapt to component naming differences caused by different projects, different modeling habits, and different IFC export software.

[0158] Improve the traceability of results: Each judgment retains RuleID, ElementID, ActualValue, RequiredValue and function execution path, which facilitates quality review, responsibility identification and rectification closure.

[0159] Balancing automation and manual verification: Automated judgment of hard indicators such as slope, thickness, height, and spacing, while outputting manual inspection manuals for soft indicators such as visual quality, avoiding the agent's overreach in judging matters that cannot be directly proven by the model.

[0160] A scalable platform is formed: the rule base, prompt word chain, agent scheduler, and rule function library can all be extended to other engineering quality inspection scenarios such as exterior walls, basements, decoration and renovation, and electromechanical installation.

[0161] This invention is not limited to the specific embodiments described above. Obviously, the embodiments described above are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the described embodiments of this invention are within the scope of protection of this invention. Those skilled in the art can make other modifications and variations to this invention. Therefore, if these modifications and variations of this invention fall within the scope of the claims of this invention, then this invention also intends to include these modifications and variations.

Claims

1. A method for determining roof quality based on intelligent agents, characterized in that, include: Step 1: Construct a roof quality rule library and divide it into Class A rules and Class B rules; Step 2: Obtain and expand the roof IFC model; Step 3: IFC model analysis and feature dimensionality reduction; Step 4: Construct a cascading suggestion chain; In step 4, the method for constructing the cascading cue word chain includes: Step 4.1: Construct the first layer of prompts, which are role and task constraint prompts, used to limit the AI ​​large model's identity as a roofing project quality supervision expert, a data analysis expert, and a rule engine scheduling expert; Step 4.2: Construct the second layer of prompt words, which are prompt words for rule intent recognition, used to identify rule intent and extract rule elements; Step 4.3: Construct the third-level prompt words, which are component semantic mapping prompt words, used to map the fuzzy component names in the IFC model to the standard IFC component types or the standard types defined in the project; Step 4.4: Construct the fourth layer of prompt words, which are function call prompt words used to convert rule checking logic into rule engine function handles and parameter packages; Step 4.5: Construct the fifth layer of prompt words, which are result summary prompt words. This is used to allow only the structured execution results returned by the rule engine to be summarized, sorted and reported in a language-based manner, and to prevent modification of the underlying function's judgment results. Step 5: Intelligent semantic matching and task scheduling; Step 6: Deterministic determination by the rule engine. The method of deterministic determination by the rule engine includes the rule engine reading the function call parameter package generated by the agent and calling the preset function library according to the function name; the rule engine outputs the determination result by a deterministic algorithm. Step 7: Collaborative output of automated check for Category A rules and manual check for Category B rules; Step 8: Results summary and report generation.

2. The roof quality determination method based on intelligent agents according to claim 1, characterized in that, In step 1, the method for constructing the roof quality rule base and dividing it into Class A and Class B rules includes: Step 1.1: Obtain roof quality acceptance specifications, enterprise standards, project-specific quality requirements, and historical inspection experience to form a set of original rules and regulations; Step 1.2: Input each original rule text in the original rule set into the AI ​​large model or manual review module, and extract the rule identifier, rule text, inspection type, severity, main target entity, secondary target entity, attributes, operators, thresholds, units, allowable error and inspection logic fields of each original rule to form a structured semantic rule text; Step 1.3: Divide the structured semantic rule texts into Class A rules and Class B rules, where Class A rules are automated judgment rules and Class B rules are manually assisted inspection rules; Step 1.4: Save the structured semantic rules after classification as a JSON file to form a roof quality rule library that can be dynamically read, maintained, and expanded.

3. The roof quality assessment method based on intelligent agents according to claim 1, characterized in that, In step 3, the method for IFC model parsing and feature dimensionality reduction includes: Step 3.1: Analyze the IFC model to obtain the roof components and their attribute sets; Step 3.2: Generate an object filter based on the structured semantic rule base, retaining only the components, attributes and topological relationships related to roof quality inspection, and removing redundant geometric coordinates, display styles and complex nested relationships that are difficult for large AI models to process directly and are irrelevant to rule judgment; Step 3.3: Expand the multilayer material structure relationship in the original IFC model into lightweight key-value pairs to form an IFC property flow; Step 3.4: Divide the large roof IFC attribute flow model into blocks; Step 3.5: Output a standardized IFC attribute stream in JSON format.

4. The roof quality determination method based on intelligent agents according to claim 1, characterized in that, In step 5, the intelligent semantic matching and task scheduling method includes: Step 5.1: Read the rules to be checked from the rule base through the intelligent agent and establish a check task queue; Step 5.2: For each Class A rule, the agent performs semantic retrieval and candidate component localization in the IFC attribute stream to determine the target component set; Step 5.3: The agent performs integrity verification on the key attributes of the target component; if the required attribute exists, a function call is initiated; if the attribute is missing but can be calculated from the associated component, a derived attribute calculation task is generated; if the attribute cannot be obtained, a data missing warning is generated. Step 5.4: The agent introduces topological environment awareness, incorporating surrounding components and their relative positions into the matching context; Step 5.5: When multiple rules apply to the same component or there are multiple candidate matching results for the same component, the agent calculates the intent weight based on the rule severity level, entity confidence, attribute completeness and context consistency, and prioritizes scheduling tasks with high weights.

5. The roof quality determination method based on intelligent agents according to claim 1, characterized in that, In step 6, the method for deterministic determination by the rule engine includes: Step 6.1: The rule engine reads the function call parameter package generated by the agent and calls the preset function library according to the function name; Step 6.2: For basic logic class rules, call the basic logic class function to determine the relationship between attribute values ​​and thresholds; Step 6.3: For spatial geometry rules, call the spatial geometry function to calculate the intersection, net distance, elevation difference, and slope between entities; Step 6.4: For topology relationship rules, call the topology relationship function to determine the component connection status, hierarchical stacking order, and drainage link relationship; Step 6.5: For engineering-specific rules, call the specific calculation function to complete the specific quality check, which includes waterproofing, insulation, drainage and flashing; Step 6.6: The rule engine outputs the judgment result using a deterministic algorithm.

6. The roof quality determination method based on intelligent agents according to claim 1, characterized in that, In step 7, the method for collaborative output of automated inspection of type A rules and manual inspection of type B rules includes: Step 7.1: For Class A rules, the agent automatically generates non-compliance record; when the measured value does not meet the requirement value or the deviation exceeds the allowable error, it is recorded as a non-compliance item; when key data is missing but the rule is detectable, it is recorded as a warning item. Step 7.2: For Class B rules, the agent generates structured manual-assisted inspection guidelines, including inspection objects, floors / areas, component locations, inspection points, photo requirements, evidence upload fields, and review suggestions; Step 7.3: For photos, measured data or verification results uploaded on-site, the intelligent agent binds them with unit identifiers and rule identifiers to form traceable quality closed-loop data.

7. The roof quality determination method based on intelligent agents according to claim 1, characterized in that, In step 8, the method for summarizing the results and generating the report includes: Step 8.1: Based on the results returned by the rule engine, the agent sorts and summarizes the non-compliant items according to severity, floor, roof zone, component type, and rule identifier; Step 8.2: Generate a report on non-compliance of roof quality. The report should include at least the project name, model version, inspection time, inspection rule version, unit identifier, component name, rule clause, measured value, required value, deviation value, severity, and rectification suggestions. Step 8.3: Generate a manual for manual inspection, which will be used by on-site personnel to review Category B rules; Step 8.4: Write the inspection results back to the BIM or project quality management platform, mark the inspection status on the corresponding components, and realize visual positioning, problem tracking and rectification closed loop.

8. A roof quality assessment system based on intelligent agents, characterized in that, The roof quality assessment method based on intelligent agents according to any one of claims 1 to 7 includes: The data input layer is used to input the expanded IFC model of the roof to be inspected, to input prompt words, and to upload on-site photos of Class B rules; The model preprocessing layer is used for parsing and feature dimensionality reduction of the IFC model; The rule knowledge layer is used to execute the construction of the roof quality rule base; The cascading prompt chain layer is used to perform the construction of the cascading prompt chain; The intelligent scheduling layer is used to perform intelligent semantic matching and task scheduling; The rule engine execution layer is used to perform deterministic decisions by the rule engine. The results output layer is used to perform collaborative output of Class A automated inspections and Class B manual assisted inspections, as well as to summarize the results and generate reports.

9. A smart terminal for roof quality assessment based on intelligent agents, characterized in that, The roof quality assessment method based on intelligent agents, as described in any one of claims 1 to 7, is embedded and installed.

Images