Prefabricated part quality evaluation and decision-making method fusing ontology and computer vision

Abstract

The invention relates to the field of visual inspection methods of engineering quality, in particular to a precast concrete slab component size quality evaluation and decision-making method fusing ontology and computer vision, which comprises the following steps: establishing a machine visual inspection frame comprising a multi-target recognition model and a pixel-level segmentation model; an SPDConv module and a CSP-OmniKernel module are introduced into the multi-target recognition model to reserve detail features of the small target, a detection frame of the small target of the component is output, and target parameters are output according to information of the detection frame; carrying out fine segmentation on the detection frame, generating a pixel-level mask and outputting a corresponding size; and detecting the number and spacing of the distribution ribs, and outputting a pixel-level size parameter as a measurement result. And inputting a measurement result into the ontology model for matching and reasoning, and automatically outputting a size quality evaluation result of the component, a root cause of a problem, and a responsibility entity and a repair suggestion for tracing a size quality generation stage. The method aims at automatic detection, interpretable responsibility traceability and decision support of the size quality of the precast concrete member in the whole production cycle.

Description

Technical Field

[0001] This invention relates to the field of visual inspection methods for engineering quality, and more specifically to a method for quality assessment and decision-making of prefabricated components that integrates ontology and computer vision. Background Technology

[0002] With the popularization of prefabricated buildings, the application of precast concrete slabs is becoming increasingly widespread. However, the dimensional quality control of these slabs still heavily relies on manual sampling inspection, which has drawbacks such as low efficiency, strong subjectivity, inconsistent standards, and difficulty in knowledge preservation and reuse. Although computer vision technology has been used for component dimensional inspection, existing research is mostly limited to defect identification and lacks high-level semantic understanding capabilities such as severity assessment of quality problems, cause analysis, and maintenance decisions, resulting in a "semantic gap."

[0003] The rich expert experience accumulated during the production of prefabricated components remains difficult to systematically and structurally preserve and apply to automated decision-making. However, ontology technology, as a method capable of semantically representing structured knowledge and making it machine-readable, has the potential to transform this implicit experience into reasonable and reusable explicit knowledge models. While ontology has demonstrated value in knowledge management within the traditional construction field, its in-depth application in the quality control of prefabricated buildings remains extremely limited, and a systematic and scalable solution has yet to be formed. More importantly, existing research focuses only on computer vision methods or ontology technology; most methods are fragmented, lacking integration, and computer vision methods have low accuracy in detecting small targets in prefabricated components such as electrical boxes and holes, failing to accurately pinpoint the root cause and stage of quality problems when they are detected. Summary of the Invention

[0004] The present invention aims to provide a method for quality assessment and decision-making of prefabricated components that integrates ontology and computer vision, in order to solve the problems of low detection accuracy of small target prefabricated components and the inability to accurately locate the root cause of quality problems.

[0005] The prefabricated component quality assessment and decision-making method that integrates ontology and computer vision in this solution includes the following steps: Step 1: Obtain original images of precast components at different production stages, as well as relevant dimensional and quality domain knowledge of precast concrete slab components; Also includes: Step 2: Construct a size and quality ontology model of the prefabricated component and establish a machine vision inspection framework. The machine vision inspection framework includes a multi-target recognition model and a pixel-level segmentation model. The size and quality ontology model defines inference rules for quality assessment and accountability. Step 3: In the multi-target recognition model, the SPDConv module is introduced to retain the detailed features of small targets, and the CSP-OmniKernel module is introduced to efficiently capture the features of targets at different scales and enhance key detailed features while controlling computational overhead. The detection boxes of the precast concrete slab outline, reserved holes, embedded junction boxes, and truss reinforcement are output. Based on the detection box information, the number of each small target, the coordinates of the embedded junction boxes, the center coordinates of the truss reinforcement and their spacing are output. The detection box areas of the precast concrete slab outline, reserved holes and reserved holes identified by the multi-target recognition model are further refined by the pixel-level segmentation model to generate a pixel-level mask. Based on the mask information, the outline size of the precast concrete slab, the outline size of the reserved hole, and the diameter and center coordinates of the reserved hole are output. Based on the bidirectional scanning adaptive algorithm, the number and spacing of the distributed reinforcement in the precast concrete slab component are detected by the physical driven scanning strategy and dynamic length threshold, and the pixel-level accurate size parameters are output as the detection results. Step 4: Input the detection results of all small targets into the dimensional quality ontology model, perform matching and reasoning through pre-set rules, and automatically output the overall dimensional quality assessment results of the precast concrete slab component, the root cause of the quality problem, the responsible entity for tracing the specific stage of dimensional quality occurrence, and specific repair suggestions.

[0006] Furthermore, in step 2, inference rules are defined using the SWRL language.

[0007] Furthermore, in step 2, the method for constructing the size-mass ontology model includes: Sub-step 2.1: Determine the boundary range of the ontology's management, reasoning on the knowledge topics of precast concrete component dimensions and quality, and parametric reasoning; Sub-step 2.2 involves reusing, expanding, or associating standard ontologies or terminologies within the domain. Sub-step 2.3: Extract core knowledge about dimensional quality from domain experts and standards, analyze and structure it, and enumerate the key nouns, verbs and phrases that need to be defined in the ontology; Sub-step 2.4 organizes the terms in sub-step 2.3 into a classification system, determines the classes, and establishes the inheritance relationship between subclasses and parent classes; Sub-step 2.5 defines the relationships between classes to form object attributes, and defines the data characteristics of a class to form data attributes; Sub-step 2.6 adds logical rules to the attributes defined in sub-step 2.5. The logical rules include domain, value range, cardinality constraint and other logical constraints. Sub-step 2.7: Populate the ontology with specific individual data; Sub-step 2.8 verifies whether the constructed ontology conforms to the goals and scope set in step 1.

[0008] Furthermore, in step 3, the CSP-OmniKernel module performs local feature extraction to capture details and global feature fusion through multi-scale feature fusion combined with dynamic attention. The process is as follows: The input data is convolved by Conv, split and processed in parallel by Split, and features are extracted by combining Omnikernel multi-kernel function and DConv depthwise convolution. Finally, Concat is used to concatenate the data to enhance its expressive power. Local details are captured through the local branch of 1×1DConv, global context is captured through the global branches of 31×1 and 31×31DConv, and local details and global context are weighted and fused through the FSAM feature space attention module. After the input is processed by Conv1×1, global features are extracted through GAP global average pooling, and the spatial correlation of features is enhanced by combining SCA spatial context attention and FFFT fast Fourier transform. Global dependencies are captured through global average pooling and fast Fourier transform, and then multiplied with the input features after Conv 1×1 adjustment to dynamically enhance key features.

[0009] Furthermore, the FSAM feature space attention module performs global average pooling and max pooling on the input feature map in the channel dimension to obtain two feature maps, which are then concatenated and processed by convolution and the Sigmoid activation function to generate spatial attention weights. The original feature map is multiplied by the spatial weights and then weighted and fused.

[0010] Furthermore, in step 4, the Drools rule engine is used for reasoning. The reasoning process of the Drools rule engine is as follows: The semantic web rules and language rules are converted into Drools rules, and the ontology knowledge and recognition results are converted into ontology individuals and Drools facts. The Drools rule domain is matched with Drools facts, and the reasoning result is derived based on the matching result.

[0011] Furthermore, in step 4, the detection results and reasoning decision results of each test are stored in a structured manner to form a case library, and then associated with the size and quality ontology model through semantic tags.

[0012] Compared with existing technologies, the beneficial effects of this solution are: By integrating ontology models and machine vision, an end-to-end solution covering the entire production cycle of precast concrete slabs was constructed, encompassing detection, evaluation, and decision-making. Implicit expert knowledge was made explicit through structured ontology models and inference rules, enabling the system to store, retrieve, and iteratively optimize this knowledge. The proposed STDE-YSC model significantly improved the detection accuracy for small targets (electrical boxes, holes); the RG-BS method, designed for rebar detection in complex environments, offers high efficiency and robustness. Finally, powered by the Drools rule engine, defects in precast concrete components were automatically detected, and the root causes were automatically inferred, providing intelligent decision-making results and a reliable basis for maintenance. This not only improved the detection accuracy for small targets but also provided automated inspection of the dimensional quality of precast concrete components throughout the entire production cycle, as well as interpretable accountability and decision support. Attached Figure Description

[0013] Figure 1 This is a flowchart of an embodiment of the prefabricated component quality assessment and decision-making method that integrates ontology and computer vision; Figure 2 This is a flowchart of the ontology construction method in an embodiment of the prefabricated component quality assessment and decision-making method that integrates ontology and computer vision. Figure 3 This is a structural diagram of the STDE-YSC model in an embodiment of the prefabricated component quality assessment and decision-making method that integrates ontology and computer vision. Figure 4 This is a structural diagram of the CSP-OmniKernel module in an embodiment of the prefabricated component quality assessment and decision-making method that integrates ontology and computer vision. Figure 5 This is a schematic diagram illustrating the implementation principle of Drools inference in an embodiment of the prefabricated component quality assessment and decision-making method that integrates ontology and computer vision. Figure 6 This is a schematic diagram of target classification for precast concrete slabs in an embodiment of the precast component quality assessment and decision-making method that integrates ontology and computer vision. Detailed Implementation

[0014] The following detailed description provides further details on specific implementation methods.

[0015] A prefabricated component quality assessment and decision-making method that integrates ontology and computer vision, such as Figure 1 As shown, it includes the following steps: Step 1: Obtain original images of precast components at different production stages, such as assembly, production, and completion. Also obtain relevant dimensional and quality domain knowledge of the precast concrete slab components. Domain knowledge includes the types of quality defects (such as dimensional deviations), the stage and location of defects, the causes of defects, the responsible entity (such as the work team), quality assessment standards, and repair methods. Knowledge sources include national standards, specifications, literature, project experience, and expert interviews.

[0016] Step 2: Construct a dimensional quality ontology model of the precast components and establish a machine vision inspection framework. This framework includes a multi-target recognition model and a pixel-level segmentation model. The ontology modeling method uses the Protégé tool to construct the dimensional quality ontology model of the precast concrete slab components, defining classes, attributes, and rules. Core concepts such as Component, QualityIssue, Cause, ResponsibleParty, and RepairSuggestion are formally defined using the OWL language, along with object and data attributes between these core concepts. The dimensional quality ontology model defines inference rules for quality assessment and accountability. These rules are defined using the SWRL language. For example, quality assessment might be evaluated as 'reject' if a target dimension deviation exceeds a threshold. Accountability accountability might be defined as 'the mold team' if the defect is caused by inaccurate mold positioning. For instance, rule judgments might include: "Contour deviation 7mm > threshold 5mm" -> triggering "Contour quality assessment 'reject'"; "Electrical box coordinate deviation 12mm > threshold 10mm" -> triggering "Electrical box quality assessment 'reject'".

[0017] like Figure 2 As shown, the methods for constructing the size-mass ontology model include: Sub-step 2.1: Determine the boundary range of the ontology's management, reasoning on the knowledge topics of precast concrete component (PCS) dimensions and quality, and parametric reasoning; Sub-step 2.2 involves reusing, expanding, or associating standard ontologies or terminologies in the query domain, such as IFC and BOT ontologies in the construction domain, IFC information conversion, and literature reviews, indicating that researchers have analyzed Industrial Foundation (IFC) standards. Sub-step 2.3 extracts core knowledge about dimensional quality from domain experts and standards (such as drawings and construction specifications), analyzes and structures it, and enumerates key nouns, verbs and phrases that need to be defined in the ontology, such as: "component", "dimension", "tolerance", "deviation", "inspection point", "qualified", "unqualified" etc. Sub-step 2.4 organizes the terms in sub-step 2.3 into a classification system, determines the classes, which are concept categories, and establishes the subclass-parent class inheritance relationship. For example, "dimensional deviation" may be a subclass of "quality defect"; "positional deviation" is another subclass of "dimensional deviation". Sub-step 2.5 defines the relationships between classes to form object attributes, such as "component hasDefect" and "inspection point measures". It also defines the data characteristics of the classes to form data attributes, such as "component hasID (data type: string)" and "dimension hasNominalValue (data type: floating point)". Sub-step 2.6 adds logical rules to the attributes defined in sub-step 2.5 to make their meaning more precise. The logical rules include domain, range, cardinality constraint and other logical constraints. For example, the domain of the attribute "measures" is "inspection point" and the range is "size". The cardinality constraint is such that "an inspection report must have one and only one inspection date". Other logical constraints are such as "if a component 'has a serious positional deviation', then its 'installation status' must be 'uninstallable'". Use Protégé to specifically define attributes, relations, and constraints, and formally encode theoretical knowledge into ontology files, usually in .owl format; Sub-step 2.7: Populate the ontology with specific individual data. For example, create an instance of "Component_A" which belongs to the "Beam" class; create an instance of "Deviation_001" which belongs to the "Position Deviation" class and associate it with "Component_A" through the hasDefect property. Sub-step 2.8 verifies whether the constructed ontology conforms to the goals and scope set in step 1, and evaluates its consistency (no logical contradictions), completeness (whether it covers important terms and relationships), clarity, and usability.

[0018] Step 3: In the multi-object recognition model STDE-YSC, the SPDConv module is introduced to preserve the detailed features of small objects, and the CSP-OmniKernel module is introduced to efficiently capture the features of objects at different scales and enhance key detailed features while controlling computational overhead. The output includes detection boxes for the outline of the precast concrete slab, reserved holes, pre-embedded junction boxes, and truss reinforcement. Based on the detection box information, the number of each small object, the coordinates of the embedded junction boxes, the center coordinates of the truss reinforcement, and their spacing are output. The spacing of the truss reinforcement is calculated based on the center coordinates of the truss reinforcement. By introducing the SPDConv and CSP-OmniKernel modules into the multi-object recognition model, the detailed features of small objects can be preserved as much as possible without excessively increasing the computational complexity of the model. The number of small objects, such as precast concrete... The number of precast concrete slab outlines, reserved openings, reserved holes, embedded junction boxes, and truss reinforcement bars is determined. The detection bounding boxes of the precast concrete slab outlines, reserved openings, and reserved holes identified by the multi-object recognition model are further refined using the pixel-level segmentation model HQ-SAM. A pixel-level mask is generated, and based on the mask information, the dimensions of the precast concrete slab outline (length and width), the dimensions of the reserved opening outline (length and width), and the diameter and center coordinates of the reserved opening are output. Based on a bidirectional scanning adaptive algorithm, using a physically driven scanning strategy and dynamic length threshold, the number and spacing of distributed reinforcement bars in the precast concrete slab component are detected, and pixel-level precision dimensional parameters are output as the detection results. The detection results include the detected dimensional data, such as the length and width corresponding to the component outline, and the number and coordinates corresponding to the embedded junction boxes. The multi-object recognition model STDE-YSC is as follows. Figure 3 As shown, within the PyTorch deep learning framework, the STDE-YSC model is trained and deployed based on the YOLOv11 architecture, and the open-source HQ-SAM model is used for instance segmentation. Figure 6 As shown, the obtained small target areas include contours, electrical boxes, holes, openings, truss reinforcement, etc. The pixel-level segmentation model HQ-SAM is used to accurately calculate the dimensions of the contours, holes, and openings. The reinforcement mesh detection module RG-BS is used for further processing; the logic of the RG-BS algorithm is implemented using Java or Python, enabling efficient and robust detection of the quantity and spacing of distributed reinforcement, thus improving the accuracy of the measurement results.

[0019] For example, during actual inspection, measurement results are quantified as follows: contour length and width deviation is +7mm, electrical box coordinate deviation is -12mm, etc. The STDE-YSC model identifies contours, electrical boxes, holes, openings, and truss reinforcement in the image, simultaneously counting the number of each target. The coordinates of the electrical box and truss reinforcement are calculated using the generated detection box information (the spacing of the truss reinforcement is calculated using these coordinates). Based on the STDE-YSC model's detection results, the HQ-SAM model performs fine segmentation of the contours, holes, and openings, and calculates the dimensions of their corresponding targets. The RG-BS method is used to detect the number and spacing of the distributed reinforcement.

[0020] like Figure 4 As shown, the CSP-OmniKernel module uses multi-scale feature fusion combined with dynamic attention to extract local features to capture details and fuse global features. The process is as follows: The input data is convolved by Conv, split and processed in parallel by Split, and features are extracted by combining Omnikernel multi-kernel function and DConv depthwise convolution. Finally, Concat is used to concatenate the data to enhance its expressive power. Local details are captured through a 1×1 DConv local branch, and global context is captured through 31×1 and 31×31 DConv global branches. The FSAM feature space attention module performs weighted fusion of local details and global context. The FSAM feature space attention module obtains two feature maps by performing global average pooling and max pooling on the input feature map in the channel dimension. After concatenation, spatial attention weights are generated through convolution and the Sigmoid activation function. The original feature map is then multiplied with the spatial weights for weighted fusion. After the input is processed by Conv1×1, global features are extracted through global average pooling (GAP), and spatial context attention (SCA) and fast Fourier transform (IFFT) are combined to enhance the spatial correlation of features. Global dependencies are captured using GAP and IFFT, and after being adjusted by Conv 1×1, they are multiplied with the input features to dynamically enhance key features.

[0021] Step 4, as follows Figure 5As shown, the Drools rule engine is used to input the detection results of all small targets into the dimensional quality ontology model. The detection results are the detected dimensional data. For example, the component outline corresponds to length and width, and the embedded junction box corresponds to quantity and coordinates. Through pre-set rules, matching and reasoning are performed to automatically output the overall dimensional quality assessment results of the precast concrete slab component, the root cause of the quality problem, the responsible entity for tracing the specific stage of dimensional quality problems, and specific repair suggestions. The production of precast concrete slab components goes through three stages: assembly, molding, and completion. Dimensional quality problems may occur at each stage, and the responsible entity may be different at each stage. Therefore, by reasoning and determining the responsible entity, it is possible to pinpoint each responsible entity. The assessment results include acceptance and rejection. Each detection result and reasoning decision result are structured and stored to form a case library, which is associated with the dimensional quality ontology model through semantic tags. The case library periodically feeds new knowledge to the data layer to expand ontology concepts and optimize the computer vision model dataset, realizing continuous iteration of system capabilities and closed-loop management of knowledge. The reasoning process of the Drools rule engine is as follows: The semantic web rules and language rules are converted into Drools rules, and the ontology knowledge and recognition results are converted into ontology individuals and Drools facts. The Drools rule domain is matched with Drools facts, and the reasoning result is derived based on the matching result.

[0022] This embodiment achieves full-process automation and intelligence, combining the automatic detection capabilities of computer vision with the intelligent reasoning capabilities of ontology. It automates the entire process from "image input" to "maintenance decision" output, significantly improving efficiency. It enhances the objectivity and standardization of decision-making by unifying evaluation standards through an ontology model and utilizing a rule engine for logical reasoning, effectively overcoming the inconsistency of subjective human judgment and reducing quality disputes. It enables the accumulation and reuse of domain knowledge: tacit expert knowledge is made explicit into a structured ontology model and reasoning rules, allowing the system to save, call, and iteratively optimize knowledge, solving the problem of knowledge inheritance. Algorithm optimizations were performed for industrial scenarios: the proposed STDE-YSC model significantly improves the detection accuracy of small targets (electrical boxes, holes); the RG-BS method is highly efficient and robust for rebar detection in complex backgrounds. It supports full lifecycle quality management: the system design covers the three stages of assembly, production, and completion, enabling the detection and handling of quality problems at the earliest stage, reducing rework costs, and achieving true closed-loop quality management.

[0023] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A prefabricated component quality assessment and decision-making method integrating ontology and computer vision, comprising the following steps: Step 1: Obtain original images of precast components at different production stages, as well as relevant dimensional and quality domain knowledge of precast concrete slab components; Its characteristic is that it further includes: Step 2: Construct a size and quality ontology model of the prefabricated component and establish a machine vision inspection framework. The machine vision inspection framework includes a multi-target recognition model and a pixel-level segmentation model. The size and quality ontology model defines inference rules for quality assessment and accountability. Step 3: In the multi-target recognition model, the SPDConv module is introduced to retain the detailed features of small targets, and the CSP-OmniKernel module is introduced to efficiently capture the features of targets at different scales and enhance key detailed features while controlling computational overhead. The detection boxes of the precast concrete slab outline, reserved holes, embedded junction boxes, and truss reinforcement are output. Based on the detection box information, the number of each small target, the coordinates of the embedded junction boxes, the center coordinates of the truss reinforcement and their spacing are output. The detection box areas of the precast concrete slab outline, reserved holes and reserved holes identified by the multi-target recognition model are further refined by the pixel-level segmentation model to generate a pixel-level mask. Based on the mask information, the outline size of the precast concrete slab, the outline size of the reserved hole, and the diameter and center coordinates of the reserved hole are output. Based on the bidirectional scanning adaptive algorithm, the number and spacing of the distributed reinforcement in the precast concrete slab component are detected by the physical driven scanning strategy and dynamic length threshold, and the pixel-level accurate size parameters are output as the detection results. Step 4: Input the detection results of all small targets into the dimensional quality ontology model, perform matching and reasoning through pre-set rules, and automatically output the overall dimensional quality assessment results of the precast concrete slab component, the root cause of the quality problem, the responsible entity for tracing the specific stage of the dimensional quality problem, and specific repair suggestions.

2. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision as described in claim 1, characterized in that: In step 2, the inference rules are defined using the SWRL language.

3. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision according to claim 2, characterized in that: In step 2, the method for constructing the size-mass ontology model includes: Sub-step 2.1: Determine the boundary range of the ontology's management, reasoning on the knowledge topics of precast concrete component dimensions and quality, and parametric reasoning; Sub-step 2.2 involves reusing, expanding, or associating standard ontologies or terminologies within the domain. Sub-step 2.3: Extract core knowledge about dimensional quality from domain experts and standards, analyze and structure it, and enumerate the key nouns, verbs and phrases that need to be defined in the ontology; Sub-step 2.4 organizes the terms in sub-step 2.3 into a classification system, determines the classes, and establishes the inheritance relationship between subclasses and parent classes; Sub-step 2.5 defines the relationships between classes to form object attributes, and defines the data characteristics of a class to form data attributes; Sub-step 2.6 adds logical rules to the attributes defined in sub-step 2.

5. The logical rules include domain, value range, cardinality constraint and other logical constraints. Sub-step 2.7: Populate the ontology with specific individual data; Sub-step 2.8 verifies whether the constructed ontology conforms to the goals and scope set in step 1.

4. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision according to claim 3, characterized in that: In step 3, the CSP-OmniKernel module performs local feature extraction to capture details and global feature fusion by combining multi-scale feature fusion with dynamic attention. The process is as follows: The input data is convolved by Conv, split and processed in parallel by Split, and features are extracted by combining Omnikernel multi-kernel function and DConv depthwise convolution. Finally, Concat is used to concatenate the data to enhance its expressive power. Local details are captured through the local branch of 1×1DConv, global context is captured through the global branches of 31×1 and 31×31DConv, and local details and global context are weighted and fused through the FSAM feature space attention module. After the input is processed by Conv1×1, global features are extracted through global average pooling, and spatial context attention and fast Fourier transform are combined to enhance the spatial correlation of features. Global dependencies are captured through global average pooling and fast Fourier transform, and then multiplied with the input features after Conv 1×1 adjustment to dynamically enhance key features.

5. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision according to claim 4, characterized in that: The FSAM feature space attention module obtains two feature maps by performing global average pooling and max pooling on the input feature map in the channel dimension. After concatenation, spatial attention weights are generated by convolution and sigmoid activation function. The original feature map is multiplied by the spatial weights and then weighted and fused.

6. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision according to claim 1, characterized in that: In step 4, the Drools rule engine is used for reasoning. The reasoning process of the Drools rule engine is as follows: The semantic web rules and language rules are converted into Drools rules, and the ontology knowledge and recognition results are converted into ontology individuals and Drools facts. The Drools rule domain is matched with Drools facts, and the reasoning result is derived based on the matching result.

7. The method for quality assessment and decision-making of prefabricated components integrating ontology and computer vision according to claim 6, characterized in that: In step 4, the detection results and reasoning decision results of each test are stored in a structured manner to form a case library, and then associated with the size quality ontology model through semantic tags.

Images