A wireless design review method and system based on image recognition

Abstract

The application relates to the field of image recognition and understanding, and discloses a wireless design auditing method and system based on image recognition, which comprises the following steps: analyzing the layers and label information of a wireless design vector drawing, constructing a spatial ontology constraint graph, positioning a to-be-detected area in a survey photo based on the spatial ontology constraint graph, extracting background texture features by using a reverse mask mechanism to generate a topological persistent barcode, mapping the topological persistent barcode to a preset material feature space, combining attribute requirements in the spatial ontology constraint graph to generate a multimodal evidence reasoning tensor by orthogonal projection calculation, analyzing conflict uncertainty dimension values in the multimodal evidence reasoning tensor, judging installation carrier compliance according to a preset safety threshold, and outputting rectification strategies, so that hidden defects of correct position but illegal carrier attribute in a non-standardized survey scene are solved, and automatic verification ability of bearing safety constraints and conclusion explainability are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition and understanding, and in particular to a wireless design review method and system based on image recognition. Background Technology

[0002] In the field of equipment installation auditing, image recognition technology is often used to analyze images collected on-site to verify whether the equipment is installed according to design requirements. In existing technologies, target detection algorithms determine the existence and location information of equipment by identifying foreground targets in images. Their key characteristic is that they typically focus on detecting foreground equipment targets, paying less attention to the background environmental features of the equipment.

[0003] During actual project acceptance, engineers may find that while equipment has been installed according to the drawings, the actual wall or structure on which it is mounted may not be the load-bearing structure required by the design. Existing image recognition technology struggles to effectively extract and understand the physical properties of the background area in an image, making it impossible to correlate and verify the background features in the on-site image with the load-bearing properties specified in the design drawings. In such cases, relying solely on image comparison may result in a misjudgment of compliance, while insufficient structural strength of the actual installation carrier will directly affect the safety of the equipment installation, potentially leading to equipment detachment or structural damage, necessitating secondary construction and rectification. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides the following technical solution:

[0005] A wireless design review method based on image recognition, comprising:

[0006] S1, analyze the layers and annotation information of the wireless design vector paper, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram;

[0007] S2, based on the spatial ontology constraint map, locates the area to be detected in the survey photo, uses the inverse masking mechanism to extract background texture features, and generates a topologically persistent barcode through homology group calculation;

[0008] S3 maps the topologically persistent barcode to a preset material feature space, and generates a multimodal evidence reasoning tensor by orthogonal projection, in combination with the attribute requirements in the spatial ontology constraint diagram.

[0009] S4 analyzes the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determines the compliance of the installation carrier based on the preset safety threshold, and outputs the rectification strategy.

[0010] Furthermore, the process of analyzing the layers and annotation information of the wireless design vector drawing, extracting the spatial relationships between device primitives and carrier primitives, and constructing a spatial ontology constraint diagram includes the following steps:

[0011] S11, acquire wireless design vector graphics and perform vector entity deconstruction processing, and obtain vector entity set and layer index table through vector entity deconstruction processing;

[0012] S12, perform layer semantic classification processing on the vector entity set based on the layer index table, and obtain the device candidate primitive set and the carrier candidate primitive set through the layer semantic classification processing;

[0013] S13, Perform device element instantiation processing on the device candidate element set, and obtain the device element set and device attribute table through device element instantiation processing;

[0014] S14, Perform carrier primitive instantiation processing on the carrier candidate primitive set, and obtain the carrier primitive set and carrier attribute table through carrier primitive instantiation processing;

[0015] S15, perform mounting relationship inference processing based on the device element set and the carrier element set, and obtain the mounting relationship set through the mounting relationship inference processing;

[0016] S16. Based on the set of device primitives, the set of carrier primitives, the device attribute table, the carrier attribute table and the set of mounting relationships, perform graph structure encapsulation processing to obtain the spatial ontology constraint graph.

[0017] Furthermore, the carrier candidate primitive set is instantiated to obtain the carrier primitive set and carrier attribute table through the carrier primitive instantiation process, which includes the following steps:

[0018] S141, perform closure determination and boundary reconstruction processing on the polyline entities in the carrier candidate primitive set, and obtain the carrier boundary candidate set through closure determination and boundary reconstruction processing;

[0019] S142, Perform component fusion processing on the candidate set of carrier boundaries and the reference entity of structural component blocks to obtain the set of carrier primitives through component fusion processing;

[0020] S143, perform carrier attribute extraction processing on the text annotation objects in the wireless design vector paper, and obtain the carrier attribute table through carrier attribute extraction processing.

[0021] Furthermore, based on the spatial ontology constraint map, the region to be detected in the survey photograph is located. Background texture features are extracted using an inverse masking mechanism, and a topologically persistent barcode is generated via homology group calculation, comprising the following steps:

[0022] S21, acquire survey photos and perform image preprocessing to obtain a standardized set of survey photos and camera geometric parameters through image preprocessing;

[0023] S22, Receive the spatial ontology constraint diagram and perform audit task extraction processing on the spatial ontology constraint diagram, and obtain the list of device nodes to be audited and the list of carrier nodes to be audited through the audit task extraction processing;

[0024] S23, Based on the standardized survey photos, perform equipment visual positioning processing on the list of equipment nodes to be reviewed, and obtain the set of equipment detection boxes and the set of equipment segmentation masks through the equipment visual positioning processing;

[0025] S24, Perform reverse masking based on the device segmentation mask set to obtain the carrier background region set through reverse masking;

[0026] S25, perform grayscale topological filtering on the carrier background region set to obtain a carrier grayscale complex set through grayscale topological filtering;

[0027] S26, homology group calculation is performed on the carrier gray-level complex set to obtain the Betti number sequence set through homology group calculation;

[0028] S27, Perform persistent digest generation processing based on the set of Betty number sequences, and obtain a topological persistent barcode through persistent digest generation processing.

[0029] Furthermore, based on the standardized survey photos, the device visual positioning process is performed on the list of device nodes to be reviewed. The process of obtaining the device detection box set and the device segmentation mask set through the device visual positioning process includes the following steps:

[0030] S231, map the list of devices to be reviewed and the device type text in the device attribute table to a set of device detection categories, and drive the target detection network to perform device detection on the standardized survey photos through the set of device detection categories to obtain an initial set of device detection boxes;

[0031] S232, Perform candidate box cleaning processing based on nonmaximum suppression on the initial device detection box set, and obtain the device detection box set through the candidate box cleaning processing;

[0032] S233, Perform instance segmentation processing on standardized survey photos based on the set of equipment detection frames, and obtain a set of equipment segmentation masks through instance segmentation processing.

[0033] Furthermore, mapping the topologically persistent barcode to a predefined material feature space, and combining the attribute requirements in the spatial ontology constraint diagram, generates a multimodal evidence reasoning tensor through orthogonal projection, including the following steps:

[0034] S31, Obtain the topological persistent barcode and perform persistent vectorization processing to obtain a persistent feature vector set;

[0035] S32, Receive the spatial ontology constraint graph and perform attribute requirement extraction processing on the spatial ontology constraint graph, and obtain the carrier requirement semantic vector set through attribute requirement extraction processing;

[0036] S33, obtain the preset material feature space and perform feature space alignment processing, and obtain the alignment transformation parameter set through feature space alignment processing;

[0037] S34, based on the persistent feature vector set, the carrier requirement semantic vector set and the alignment transformation parameter set, perform orthogonal projection calculation processing to obtain the projection evidence pair set through orthogonal projection calculation processing;

[0038] S35, perform evidence tensor construction processing on the set based on projected evidence, and obtain multimodal evidence reasoning tensor through evidence tensor construction processing.

[0039] Further, receiving the spatial ontology constraint graph and performing attribute requirement extraction processing on the spatial ontology constraint graph, and obtaining the carrier requirement semantic vector set through attribute requirement extraction processing includes the following steps:

[0040] S321, Read the parent carrier element identifier associated with the device node identifier from the attribute dictionary set of the spatial ontology constraint diagram, and extract the material attribute text and mechanical grade mark from the carrier attribute table to obtain the carrier requirement text set.

[0041] S322, Perform normalization encoding on the carrier requirement text set to obtain the carrier requirement semantic token set through normalization encoding;

[0042] S323, perform semantic embedding processing on the set of semantic tokens required by the carrier, and obtain the set of semantic vectors required by the carrier through semantic embedding processing.

[0043] Furthermore, the steps involved in analyzing the conflict uncertainty dimension of the multimodal evidence reasoning tensor, determining the compliance of the installation carrier based on a preset security threshold, and outputting rectification strategies include:

[0044] S41, obtain the multimodal evidence reasoning tensor and perform evidence component unpacking processing, and obtain the compliance judgment input table through evidence component unpacking processing;

[0045] S42, obtain the preset security threshold and perform threshold consistency verification processing, and obtain the threshold set through threshold consistency verification processing;

[0046] S43, perform carrier compliance judgment processing based on the compliance judgment input table and threshold set, and obtain the carrier compliance result table through carrier compliance judgment processing;

[0047] S44, Based on the carrier compliance result table and the spatial ontology constraint diagram, perform rectification strategy generation processing to obtain a set of rectification strategies;

[0048] S45, based on the set of rectification strategies, performs audit result output processing, and obtains the audit report data packet through the audit result output processing.

[0049] Furthermore, based on the compliance judgment input table and the threshold set, the carrier compliance judgment process is performed, and the carrier compliance result table is obtained through the carrier compliance judgment process, including the following steps:

[0050] S431, Perform a matching degree threshold comparison process on each record in the compliance judgment input table, and obtain a matching degree conclusion marker through the matching degree threshold comparison process;

[0051] S432, perform uncertainty threshold comparison processing on each record in the compliance judgment input table, and obtain the uncertainty conclusion label through uncertainty threshold comparison processing;

[0052] S433, perform rule fusion processing on the matching degree conclusion mark and the uncertainty conclusion mark, and obtain the compliance mark and the non-compliance type mark through rule fusion processing and write them into the carrier compliance result table.

[0053] A wireless design review system based on image recognition, used to implement the aforementioned wireless design review method based on image recognition, the system comprising:

[0054] Spatial ontology constraint diagram construction module: used to parse the layers and annotation information of wireless design vector drawings, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram;

[0055] Topological persistent barcode generation module: used to locate the area to be detected in the survey photo based on the spatial ontology constraint map, extract background texture features using the inverse masking mechanism, and generate a topological persistent barcode through homology group calculation;

[0056] Multimodal Evidence Reasoning Tensor Generation Module: This module maps the topologically persistent barcode to a preset material feature space, combines the attribute requirements in the spatial ontology constraint diagram, and generates a multimodal evidence reasoning tensor through orthogonal projection calculation.

[0057] Carrier Compliance Judgment and Rectification Strategy Output Module: Used to parse the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determine the compliance of the installation carrier based on the preset security threshold, and output the rectification strategy.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention utilizes image recognition and understanding technologies to address the hidden defects in non-standardized survey scenarios where equipment is correctly positioned but carrier attributes violate regulations. It constructs a closed-loop processing chain from design constraints to on-site image evidence. At the image understanding level, an inverse masking mechanism is used to remove the main equipment area, forcing computational resources to be concentrated on the carrier background area. This avoids the attention bias of the target detection network towards foreground equipment, allowing image analysis to focus on the surface texture features of the carrier to which the equipment is attached. Furthermore, homology group computation is employed to convert the carrier background grayscale image into a topologically persistent barcode, stably recording the connectivity and porosity evolution processes caused by the microstructures of different materials. This overcomes the interference of lighting, shooting angle changes, and surface contamination on pixel-level texture comparison, achieving robust characterization of the carrier's texture structure. At the multimodal fusion level, orthogonal projection calculations are performed on the topologically persistent features and design semantic requirements in a preset material feature space, generating a multimodal evidence inference tensor that includes visual feature confidence, semantic requirement matching degree, and conflict uncertainty. This explicitly separates the compatible parts and conflict sources of visual evidence and semantic constraints, solving the problem of the difficulty in directly comparing the symbolic attributes of vector paper with on-site pixel textures. By analyzing the conflict uncertainty dimension value in the tensor and combining it with the preset safety threshold, the system can accurately determine the compliance of the carrier material. It can also refine the output of non-compliance types such as carrier material conflict, insufficient evidence, and local anomalies, and map them into actionable rectification strategies. This allows the audit conclusions to not only point out the problems, but also locate the source of the conflict. It provides traceable image evidence and targeted supplementary evidence paths for the engineering site, which significantly improves the application depth and reliability of image recognition and understanding technology in engineering carrier safety verification scenarios. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of a wireless design review method based on image recognition in this invention;

[0062] Figure 2 This is a schematic diagram of the spatial body constraint diagram in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram illustrating the extraction of the carrier background region set in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the carrier grayscale complex construction in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram illustrating the generation of a topology-persistent barcode in an embodiment of the present invention;

[0066] Figure 6 This is a schematic diagram of the multimodal evidence reasoning tensor in an embodiment of the present invention;

[0067] Figure 7 This is a schematic diagram illustrating the generation of rectification strategies in an embodiment of the present invention;

[0068] Figure 8 This is a functional block diagram of a wireless design review system based on image recognition in this invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1:

[0071] Please see Figure 1 As shown, this embodiment provides a wireless design review method based on image recognition, including:

[0072] S1: Analyze the layers and annotation information of the wireless design vector drawing, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram;

[0073] Furthermore, the process of analyzing the layers and annotation information of the wireless design vector drawing, extracting the spatial relationship between device primitives and carrier primitives, and constructing a spatial ontology constraint diagram includes the following steps:

[0074] S11: Obtain the wireless design vector paper and perform vector entity deconstruction processing to obtain a vector entity set and a layer index table. The data structure of the wireless design vector paper is a vector file containing three types of objects: layers, geometric primitives, and text annotations. The data structure of the vector entity set is a record list, with record fields including at least entity identifier, entity type, geometric parameters, layer identifier, and entity bounding box. The data structure of the layer index table is a key-value mapping table, where the key is the layer identifier and the value is the layer name and layer visibility attribute. Specifically, the vector entity deconstruction processing is performed as follows: lines, polylines, arcs, block references, and text objects are parsed according to entity type, and an entity bounding box is generated for each entity. The entity bounding box is used for subsequent geometric proximity calculations. For example, performing vector entity deconstruction on a wireless design vector drawing yields 325 line entities, 180 polyline entities, 42 arc entities, 56 block reference entities, and 98 text objects, totaling 701 vector entities. A bounding box is generated for each entity. For instance, if the insertion point coordinates of a block reference entity are (12500, 8300) and the block definition size is (600, 400), the generated bounding box will be from the lower left corner (12200, 8100) to the upper right corner (12800, 8500). Simultaneously, a layer index table containing 12 layers is obtained, where the layer identifier "Layer-03" corresponds to the layer name "Communication Equipment - Antenna," and the layer visibility attribute is set to visible.

[0075] S12, perform layer semantic classification processing on the vector entity set based on the layer index table to obtain a set of candidate device elements and a set of candidate carrier elements. The data structure of the device candidate element set is a subset of vector entities, including block reference entities and their associated leader entities in the device-related layers. The data structure of the carrier candidate element set is also a subset of vector entities, including polyline entities, section line entities, and structural component block reference entities in the building structure-related layers. Specifically, the layer semantic classification processing is performed as follows: read the industry-standard keywords in the layer name and combine them with the distribution of entity types within the layer for judgment. When the layer name contains device type keywords and the proportion of block reference entities is higher than a preset proportion threshold, the layer is marked as a device-related layer and imported into the device candidate element set. When the layer name contains structural component keywords and the proportion of polyline entities is higher than a preset proportion threshold, the layer is marked as a building structure-related layer and imported into the carrier candidate element set. The preset percentage threshold is determined as follows: The percentage of entity types in different layers is statistically analyzed in historically confirmed wireless design vector drawing samples, and the percentage value corresponding to the inflection point that minimizes the misclassification rate is taken as the preset percentage threshold. For example, in historically confirmed wireless design vector drawing samples, assuming 120 samples are collected, the average percentage distribution of block reference entities in the device-related layer is 78%, and the average percentage distribution of polyline entities in the building structure-related layer is 65%. By plotting the curve of misclassification rate changing with the percentage threshold, the inflection point for the device-related layer is found to be 60%, and the inflection point for the building structure-related layer is 50%. Therefore, 60% is taken as the preset percentage threshold for the device-related layer, and 50% is taken as the preset percentage threshold for the building structure-related layer.

[0076] S13, perform device element instantiation processing on the device candidate element set to obtain the device element set and device attribute table. The data structure of the device element set is a node list, where each node contains a device element identifier, a device geometric center point, a device orientation vector, and a device bounding box. The data structure of the device attribute table is a table, which at least contains a device element identifier, device type text, device specification text, and installation method text. Specifically, the device element instantiation processing is as follows: expand the block reference entity into sub-entities within the block and calculate the transformation matrix of the block reference. Obtain the device geometric center point and device orientation vector from the insertion point of the block reference and the transformation matrix. When the device type text exists in the form of a text object, associate the text object with the corresponding device element identifier according to the leader association relationship and write it into the device attribute table. The method for determining the lead wire association relationship is as follows: the shortest distance from the end point of the lead wire to the device bounding box is used as the association cost, and the device graphic element with the lowest association cost and below the preset distance threshold is selected as the association target of the text object; the method for determining the preset distance threshold is as follows: the upper limit of the allowable annotation offset is calculated based on the annotation scale and line width standard of the wireless design vector drawing, and this upper limit is used as the preset distance threshold. For example, assuming that the annotation scale of the wireless design vector drawing is 1:100 and the line width standard is 0.35mm, the upper limit of the allowable annotation offset is calculated as follows: in the drawing space, the allowable deviation between the end of the lead wire and the device bounding box is 3 times the line width, that is, 0.35×3=1.05mm; converted to the model space, it is 1.05×100=105mm, so 105mm is used as the preset distance threshold.

[0077] S14, perform carrier primitive instantiation processing on the carrier candidate primitive set to obtain a carrier primitive set and a carrier attribute table. The data structure of the carrier primitive set is a node list, where each node contains a carrier primitive identifier, a carrier geometric center point, a carrier principal direction vector, a carrier boundary polygon, and a carrier bounding box. The data structure of the carrier attribute table is a table, which at least contains a carrier primitive identifier, carrier type text, material attribute text, and mechanical grade mark. Further, S14 includes the following steps:

[0078] S141, perform closure determination and boundary reconstruction processing on the polyline entities in the carrier candidate primitive set, and obtain the carrier boundary candidate set through closure determination and boundary reconstruction processing; the data structure of the carrier boundary candidate set is a polygon list, and each polygon contains a vertex sequence and a topological closure marker; the closure determination method is to determine whether the distance between the first and last endpoints of the polyline entity is less than a preset endpoint closure threshold and to determine whether there is self-intersection; the preset endpoint closure threshold is determined by determining the endpoint merging tolerance according to the drawing unit system and the minimum drawing accuracy, and using the endpoint merging tolerance as the preset endpoint closure threshold. For example, assuming the drawing unit is millimeters and the minimum drawing accuracy is 0.01mm, the endpoint merging tolerance is taken as 10 times the minimum drawing accuracy, i.e., 0.01×10=0.1mm. Therefore, 0.1mm is used as the preset endpoint closure threshold. The coordinates of the first endpoint of a certain polyline entity are (5000.00, 3000.00) and the coordinates of the last endpoint are (5000.08, 3000.02). The distance between the first and last endpoints is 0.082mm, which is less than the preset endpoint closure threshold of 0.1mm. Therefore, it is determined to be topologically closed.

[0079] S142, perform component fusion processing on the carrier boundary candidate set and the structural component block reference entity to obtain a carrier primitive set through component fusion processing; the component fusion processing method is as follows: when the bounding box of the structural component block reference entity overlaps with the bounding box of a certain carrier boundary candidate in area and the overlap ratio is higher than a preset overlap threshold, the semantic tag of the structural component block reference entity is merged into the carrier primitive identifier corresponding to the carrier boundary candidate, and the carrier boundary polygon is generated from the carrier boundary candidate; the preset overlap threshold is determined by: statistically analyzing the typical overlap ratio distribution between component block references and boundary lines in historical drawings, and selecting the quantile points that can distinguish between the same component and adjacent components as the preset overlap threshold. For example, the typical overlap ratio distribution between component block references and boundary lines is statistically analyzed in historical drawings. Assuming that the 25th quantile of the overlap ratio distribution is 72% for the same component and the 75th quantile of the overlap ratio distribution for adjacent components is 35%, the median value that can distinguish between the two is selected as approximately 53.5%, and after rounding, 55% is used as the preset overlap threshold. The bounding box area of ​​a structural component block reference entity is 48000mm², and the overlap area with the bounding box of a candidate boundary of a carrier is 31200mm², with an overlap ratio of 31200 / 48000=65%, which is higher than the preset overlap threshold of 55%. Therefore, semantic tag merging is performed.

[0080] S143, perform carrier attribute extraction processing on the text annotation objects in the wireless design vector drawing to obtain a carrier attribute table. The carrier attribute extraction processing method is as follows: first, perform character recognition on the text annotation objects to obtain an annotation text set; then, perform engineering dictionary matching on the annotation text set to obtain material attribute text and mechanical grade markers; and finally, write the material attribute text and mechanical grade markers into the carrier attribute table according to the geometric proximity relationship between the text objects and the carrier bounding boxes. The engineering dictionary refers to a searchable dictionary containing material terms, load-bearing category terms, and structural component terms. The engineering dictionary is constructed by extracting terms from communication base station civil engineering specifications, common structural drawing sets, and enterprise construction method texts, and then solidifying them into a set of entries after manual review. For example, 68 terms such as load-bearing concrete wall, non-load-bearing partition wall, and steel beam were extracted from the civil engineering specifications for communication base stations; 95 terms such as C30 concrete, HRB400 steel bar, and MU10 brick were extracted from common structural drawings; and 43 terms such as expansion bolt fixing and chemical anchor bolt rebar were extracted from enterprise construction method texts. After manual review and merging of synonyms, these terms were solidified into an engineering dictionary containing 186 terms.

[0081] S15, perform mounting relationship inference processing based on the device primitive set and the carrier primitive set to obtain a mounting relationship set. The data structure of the mounting relationship set is an edge list, where each edge contains a subordinate device primitive identifier, a parent carrier primitive identifier, a geometric proximity cost, and a structural compatibility mark. Specifically, the mounting relationship inference processing is as follows: for each device primitive identifier, calculate the shortest distance from the device bounding box to each carrier boundary polygon and calculate the angle between the device orientation vector and the carrier main direction vector. Normalize the shortest distance and angle linearly to obtain the geometric proximity cost, and select the carrier primitive identifier with the smallest geometric proximity cost that is lower than a preset proximity threshold as the parent carrier primitive identifier. When the installation method text in the device attribute table indicates wall mounting or pole mounting, further verify whether the normal projection distance of the device's geometric center point on the outside of the carrier boundary polygon meets the installation gap constraint, thereby generating a structural compatibility mark. The preset proximity threshold is determined as follows: the maximum allowable mounting offset is calculated based on the drawing scale, the outer dimensions of common mounting parts, and the upper limit of the marked offset, and the maximum allowable mounting offset is converted into units and used as the preset proximity threshold.

[0082] In a preferred embodiment, the formula for calculating the geometric proximity cost is as follows:

[0083] ;

[0084] in, Represents the geometric proximity cost, with a value range of 1. ; This represents the shortest distance from the device bounding box to the carrier boundary polygon, in millimeters. The normalized baseline value represents the shortest distance, and its value is the maximum allowable mount offset corresponding to the preset proximity threshold. This represents the angle between the device's orientation vector and the carrier's principal direction vector, expressed in degrees. The normalized reference value for the included angle is 90 degrees. and Let be the weighting coefficient, satisfying Preferably , This indicates that the contribution of the shortest distance to the geometric proximity cost is greater than the contribution of the included angle.

[0085] For example, assuming the drawing scale is 1:50, the outer dimensions of common mounting components (such as pole clamps and wall mounts) are 150mm, and the upper limit of the offset is 50mm, then the maximum allowable mounting offset is calculated as 150 + 50 = 200mm; For a device element identified as "DEV-008", the shortest distance from its bounding box to the carrier boundary polygon corresponding to the carrier element identifier "CAR-015" is calculated to be 85mm, and the angle between the device orientation vector and the carrier principal direction vector is 12°. Substituting these values ​​into the geometric proximity cost calculation formula: The value is approximately 0.32, which is lower than the normalized value of 0.5 corresponding to the preset proximity threshold. Therefore, "CAR-015" is used as the parent carrier primitive identifier.

[0086] S16, based on the device primitive set, carrier primitive set, device attribute table, carrier attribute table, and mounting relationship set, a graph structure encapsulation process is performed to obtain a spatial ontology constraint graph. The data structure of the spatial ontology constraint graph is a graph data object, containing a node set, an edge set, and an attribute dictionary set. The node set is obtained by merging the device primitive set and the carrier primitive set, and the node type field is used to distinguish between device nodes and carrier nodes. The edge set is transformed from the mounting relationship set, and the edge direction points from device nodes to carrier nodes. The attribute dictionary set is obtained by merging the device attribute table and the carrier attribute table, with primitive identifiers as keys. The spatial ontology constraint graph refers to a machine-readable graph structure used to express the topological dependencies between devices and carriers, as well as the material and mechanical property constraints of the carrier nodes. The material attribute text and mechanical grade markers of the carrier nodes constitute the installation constraints on the device nodes. See also... Figure 2 This is a schematic diagram of spatial ontology constraints provided in an embodiment of this application. Figure 2As shown in the diagram, this figure illustrates the result of transforming unstructured vector lines into a graph data structure with logical dependencies. The square carrier node at the top acts as the parent, explicitly encapsulating the rigid constraints of the design, such as concrete walls and load-bearing concrete. The circular device nodes connected below are bound to the carrier via directed arrows (mounting relationships). The dashed box emphasizes that these attributes constitute the source of constraints for subsequent review. In non-standardized survey scenarios, the review pain point often lies in the ease of position alignment but the difficulty of attribute verification, because simple image detection cannot determine what material the device should be installed on. This embodiment constructs this spatial ontology constraint graph, forcibly establishing a parent-child inheritance relationship between devices and carriers in the data flow. This allows subsequent visual analysis to accurately retrieve material attribute text and mechanical grade markings from the design drawings using the parent carrier ID as an index, thereby avoiding review blind spots caused by data link breaks.

[0087] Specifically, S1 elevates device and carrier primitives in wireless design vector drawings from pure geometric lines to computable nodes and edges, retaining mounting relationships and structural compatibility markers on the edges. This allows subsequent reviews to move beyond simply relying on whether the device coordinates fall within a certain area, providing a clear constraint that the parent carrier must meet material attribute text and mechanical grade markings. In non-standardized survey scenarios, devices are often photographed, but carrier details are complex. Without a spatial ontology constraint diagram, reviews can only perform positional alignment and cannot establish the dependency relationship between the device and the carrier, thus failing to trigger verification of carrier attributes. With the spatial ontology constraint diagram output by S1, subsequent steps can use the parent carrier primitive identifier as an index to stably transmit the design-side carrier requirements to the background feature extraction and evidence reasoning stages on the image side, preventing the loss of carrier constraints during processing from a data flow perspective.

[0088] S2: Based on the spatial ontology constraint map, locate the area to be detected in the survey photo, extract the background texture features using the inverse masking mechanism, and generate a topologically persistent barcode through homology group calculation;

[0089] Furthermore, the process of locating the region to be detected in the survey photograph based on the spatial ontology constraint graph, extracting background texture features using an inverse masking mechanism, and generating a topologically persistent barcode via homology group calculation includes the following steps:

[0090] S21, acquire survey photos and perform image preprocessing to obtain standardized survey photos and a set of camera geometric parameters. The data structure of the survey photos is a pixel matrix, including color channels and shooting timestamp fields. The data structure of the standardized survey photos is a pixel matrix with unified resolution and color space. The data structure of the camera geometric parameter set is a key-value mapping table, including focal length information, principal point information, and distortion coefficient information. Specifically, the image preprocessing method is as follows: resample the survey photos at higher resolutions and perform color space conversion to obtain standardized survey photos. When shooting metadata is available, the camera geometric parameter set is obtained by parsing the shooting metadata. When shooting metadata is missing, a distortion estimation method based on line-preserving constraints is used to estimate distortion coefficient information from the standardized survey photos and write it into the camera geometric parameter set. The line-preserving constraint means that the edges of walls and beams in the building structure should be straight under ideal imaging. Distortion coefficient information is obtained by minimizing the detected line segment curvature.

[0091] Specifically, the distortion estimation method based on line preservation constraints is implemented by combining Hough transform line detection with radial distortion optimization. The processing steps are as follows: First, Canny edge detection is performed on the standardized survey photo to obtain the edge image. Then, a probabilistic Hough transform is performed on the edge image to detect the set of line segments. The input of the probabilistic Hough transform is the binary pixel matrix of the edge image, and the output is a list of coordinates of the endpoints of the line segments. Next, the curvature index is calculated for each detected line segment. The curvature index is the root mean square deviation of the sampled point of the line segment from the ideal line. Finally, the radial distortion coefficient is used as the optimization variable, and the sum of the curvature indices of all line segments is used as the objective function. The Levenberg-Marquardt nonlinear least squares optimization algorithm is used to iteratively solve for the radial distortion coefficients K1 and K2 that minimize the objective function.

[0092] For example, the original resolution of a survey photo is 4032×3024 pixels, and the color space is sRGB. After resolution resampling and unification to 1920×1440 pixels, and color space conversion to linear RGB space, a standardized survey photo is obtained. The focal length information of 4.25mm (equivalent to 26mm in 35mm format), principal point information of (960, 720), and radial distortion coefficients K1=-0.0256 and K2=0.0089 are obtained from the shooting metadata and written into the camera geometric parameter set.

[0093] S22, Receive the spatial ontology constraint graph and perform an audit task extraction process on the spatial ontology constraint graph to obtain a list of device nodes to be audited and a list of carrier nodes to be audited through the audit task extraction process; The data structure of the list of device nodes to be audited is a sequence of device node identifiers; The data structure of the list of carrier nodes to be audited is a sequence of carrier node identifiers, and the carrier node identifiers and the list of device nodes to be audited correspond one-to-one through the edge set of the spatial ontology constraint graph.

[0094] S23, based on the standardized survey photos, perform equipment visual positioning processing on the list of equipment nodes to be reviewed, obtaining a set of equipment detection boxes and a set of equipment segmentation masks through the equipment visual positioning processing; the data structure of the equipment detection box set is a list of rectangles, each rectangle containing the coordinates of the upper left corner, the coordinates of the lower right corner, and the equipment node identifier; the data structure of the equipment segmentation mask set is a list of binary masks, each binary mask containing pixel-level foreground markers and equipment node identifiers. Further, S23 includes the following steps:

[0095] S231, the list of devices to be reviewed and the device type text in the device attribute table are mapped to a set of device detection categories. The target detection network is driven by the set of device detection categories to perform device detection on the standardized survey photos to obtain an initial set of device detection boxes. The data structure of the set of device detection categories is a list of category identifiers, and the category identifiers are obtained by encoding the device type text using a category dictionary.

[0096] The target detection network specifically employs the YOLOv8 target detection model. The input data for the YOLOv8 target detection model is a pixel matrix of standardized survey photographs, with an input data structure of a three-dimensional tensor of shape (H, W, 3), where H is the image height in pixels, W is the image width in pixels, and 3 represents the RGB three channels. The output data of the YOLOv8 target detection model is a list of detection results, with an output data structure of a record list. Each record contains six fields: the X-coordinate of the top-left corner of the detection box, the Y-coordinate of the top-left corner of the detection box, the X-coordinate of the bottom-right corner of the detection box, the Y-coordinate of the bottom-right corner of the detection box, a category identifier, and a confidence score. The YOLOv8 target detection model adopts the YOLOv8m specification and is put into use after fine-tuning based on the labeled samples corresponding to the device detection category set.

[0097] The category dictionary refers to a mapping table from device type text to model category identifiers. It is established by grouping synonymous device type texts into a unified category identifier based on the enterprise's device coding system. For example, a device attribute table may contain device type texts such as "RRU radio frequency unit," "AAU active antenna unit," and "GPS antenna." After being encoded using the category dictionary, these are mapped to a set of device detection categories [CLS_001, CLS_002, CLS_005]. Synonyms such as "RRU radio frequency unit," "RRU," and "radio remote unit" are all grouped into the category identifier CLS_001.

[0098] S232, perform candidate box cleaning processing based on non-maximum suppression on the initial device detection box set to obtain the device detection box set. The candidate box cleaning processing adopts a greedy non-maximum suppression algorithm, which works as follows: sort the detection boxes from high to low confidence, take the detection box with the highest confidence as the benchmark, calculate the intersection-union ratio (IUR) of the remaining detection boxes with the benchmark box, and remove the corresponding detection box when the IUR is greater than a preset IUR threshold. Then, take the second highest confidence of the remaining detection boxes as the new benchmark and repeat the above process until all detection boxes have been processed. The IUR is a common metric in the field of object detection, used to measure the degree of overlap of rectangular boxes, and is calculated by dividing the intersection area of ​​two rectangular boxes by the union area.

[0099] The preset intersection-union ratio (IURR) threshold is set to 0.5. The preset IURR threshold is determined by statistically analyzing the detection precision and recall under different IURR thresholds on the labeled validation set, and selecting the IURR threshold that maximizes the F1 score. The confidence threshold is set to 0.25. The confidence threshold is determined by statistically analyzing the false positive rate and false negative rate under different confidence thresholds on the labeled validation set, and selecting the confidence threshold that results in a false negative rate of less than 5% and a false positive rate of less than 10%.

[0100] S233, Perform instance segmentation processing on standardized survey photos based on the device detection box set, and obtain a device segmentation mask set through instance segmentation processing; The instance segmentation processing method is as follows: take the area defined by the device detection box set as the segmentation input area, the segmentation network outputs the pixel-level foreground marker corresponding to the device node identifier, and write the pixel-level foreground marker into the device segmentation mask set.

[0101] The segmentation network specifically employs the Mask R-CNN instance segmentation model. The input data of this model includes a pixel matrix of standardized survey photos and a set of device detection boxes. The input data structure of the pixel matrix is ​​a 3D tensor with a shape of (H, W, 3), and the input data structure of the device detection box set is a list of rectangular box coordinates. The output data of the Mask R-CNN instance segmentation model is a list of instance segmentation masks, with a binary mask structure. Each binary mask has a shape of (H, W), where a pixel value of 0 represents the background and a pixel value of 1 represents the foreground device region. The Mask R-CNN instance segmentation model uses ResNet-50-FPN as its backbone network and was fine-tuned based on a communication device instance segmentation annotation dataset before being put into use.

[0102] S24. Perform reverse masking based on the device segmentation mask set to obtain a carrier background region set. The data structure of the carrier background region set is a region list, where each region contains a set of background pixels, a region bounding box, and a device node identifier. The core logic of the reverse masking is to shield the foreground device body region in the survey photo and extract the carrier background neighborhood where the device is actually mounted, so as to eliminate the interference of device pixels on carrier feature analysis and ensure that subsequent carrier material and structural feature analysis can focus on the carrier background region that is strongly related to the safety of the device.

[0103] Furthermore, the reverse masking process includes the following steps:

[0104] S241, perform morphological dilation processing on each binary mask in the device segmentation mask set to obtain a device body extended mask; the data structure of the device body extended mask is a binary mask; the structuring element used in the morphological dilation processing is a circular structuring element, and the scale of the structuring element is determined by the scale of the corresponding rectangular frame in the device detection box set. The method for determining the scale of the structuring element is: calculate the dilation radius based on the minimum value of the width and height of the rectangular frame and take an integer pixel radius that matches the pixel resolution. Specifically, multiply the minimum value of the width and height of the rectangular frame by the dilation coefficient 0.15 and then round it to obtain the radius pixel value of the circular structuring element. For example, assuming the top-left corner coordinates of a rectangle in the device detection frame set are (520, 380) and the bottom-right corner coordinates are (640, 465), the rectangle's width is 120 pixels and its height is 85 pixels. Taking the minimum value of 85 pixels, the expansion radius is calculated. According to the expansion coefficient of 0.15, the expansion radius is calculated to be 85 × 0.15 = 12.75 pixels. Taking the integer pixel radius that matches the pixel resolution, it is 13 pixels. Therefore, a circular structuring element with a radius of 13 pixels is used to perform morphological expansion processing.

[0105] S242, using the device body extended mask, performs pixel zeroing on the standardized survey photo and clips the annular neighborhood outside the device body extended mask. The pixel zeroing and clipping processes yield a set of carrier background areas. The annular neighborhood refers to the area extending outward from the rectangular frame of the device detection frame set, and its outer radius is determined by the visible carrier range in the device installation specifications. The outer radius is determined by reading the typical dimension from the visible boundary of the device base to the fixed surface from the device installation construction standard drawing set and converting it to the image resolution to obtain the outer radius pixel value. For example, if the typical dimension from the base of a certain type of RRU device to the visible boundary of the fixed surface is 350mm from the device installation construction standard drawing set, and assuming that each pixel corresponds to an actual size of 2.5mm after image resolution calibration, the outer radius pixel value is calculated as 350 / 2.5 = 140 pixels. Therefore, the outer radius of the annular neighborhood is set to 140 pixels, and the inner radius is determined by the outer boundary of the device body extended mask. See also... Figure 3 This is a schematic diagram of the extraction of the carrier background region set provided in the embodiments of this application. For example... Figure 3 As shown in the figure, this visually demonstrates the operation process of the reverse masking mechanism. In the original survey photo, the algorithm first generates a dark-colored extended mask for the main device, completely eliminating (zeroing out) the foreground device pixels (such as the antenna). Then, it expands outward from the device detection box to extract a ring-shaped neighborhood. The highlighted parts retained within this neighborhood constitute the set of background areas of the carrier. This processing logic reflects the unique decentralized visual attention mechanism of this invention: in conventional engineering reviews, AI models often overemphasize the existence of the device itself, ignoring the surfaces it is attached to. This embodiment, by physically obscuring the main device, forces the subsequent feature extraction algorithm to perform calculations only on the mounting surface (such as the wall or pole surface), effectively preventing the diversity of device appearance from interfering with the judgment of carrier material, and ensuring that the evidence chain focuses on the background texture features related to load-bearing safety.

[0106] S25, perform gray-level topological filtering on the carrier background region set to obtain a carrier gray-level complex set. The data structure of the carrier gray-level complex set is a list of complex objects. Each complex object contains a vertex set, edge set, face set, and a filtering threshold sequence, and is associated with a device node identifier. The carrier gray-level complex refers to constructing a combination of mathematical objects, such as simple complexes or cubic complexes, from the gray-level images in the carrier background region set according to pixel adjacency relationships, so that they can be used for homology group calculations. Specifically, the gray-level topological filtering process is as follows: the background pixel set in the carrier background region set is mapped to a gray-level scalar field, and down-filtering or up-filtering is performed on the gray-level scalar field according to a preset filtering threshold sequence. Under each filtering threshold, a corresponding complex object is constructed, thereby forming the carrier gray-level complex set. The filtering threshold sequence is determined by sampling the gray-level histogram of the background pixel set at quantiles and taking the quantile sequence that can cover the change process from low gray-level to high gray-level as the filtering threshold sequence. For example, quantile sampling is performed on the grayscale histogram of a certain carrier background region set. Assuming the grayscale value range is 0-255, nine quantiles are selected: 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%. The corresponding grayscale values ​​are 25, 48, 72, 95, 118, 142, 168, 195, and 223, respectively. The filter threshold sequence is then [25, 48, 72, 95, 118, 142, 168, 195, 223]. The complex object constructed when the filter threshold is 72 contains a set of 2847 vertices, 5623 edges, and 2776 faces. See also... Figure 4 This is a schematic diagram of the carrier grayscale complex construction provided in the embodiments of this application. Figure 4 As shown in the figure, this illustrates the process of transforming continuous image pixels into discrete algebraic topological structures. The left side is the extracted local grayscale image of the carrier background, and the right side is the grayscale complex of the carrier constructed after filtering. The figure highlights connected components (blue areas) formed by similar grayscale pixels and the enclosed holes (orange areas) with different colors. This step is crucial for addressing the complexities of on-site lighting: the pixel values ​​of material features such as micropores on concrete surfaces and gaps in brick walls vary greatly under different lighting conditions, but their topological structures (such as the number of holes and connectivity) remain relatively stable. By constructing grayscale complexes, the system no longer relies on light-sensitive color or texture statistics, but instead transforms the material surface into a geometric object robust to changes in lighting, laying the foundation for subsequent extraction of mathematically meaningful material fingerprints.

[0107] S26. Homology group calculation is performed on the carrier grayscale complex set to obtain a set of Betti number sequences. The data structure of the Betti number sequence set is a sequence list, where each sequence contains a filter threshold sequence and a corresponding Betti number vector. The Betti number vector contains at least two types of topological counts: the number of connected components and the number of holes. The homology group calculation is a common calculation process in topological data analysis, used to obtain topological invariants of different dimensions from the complex object. Specifically, the homology group calculation is performed as follows: for each complex object in the carrier grayscale complex set, the generation and merging events of connected components are calculated in ascending order of the filter threshold sequence, and the changes in the number of connected components are counted. At the same time, the generation and filling events of holes are calculated, and the changes in the number of holes are counted, thereby forming a Betti number sequence aligned with the filter threshold sequence. For example, performing homology group calculation on a certain carrier grayscale complex set according to the filtering threshold sequence [25,48,72,95,118,142,168,195,223] yields the following Betti number sequence: at a filtering threshold of 25, the number of connected components is 312 and the number of holes is 8; at a filtering threshold of 72, the number of connected components is 156 and the number of holes is 23; at a filtering threshold of 118, the number of connected components is 67 and the number of holes is 45; at a filtering threshold of 195, the number of connected components is 12 and the number of holes is 38; and at a filtering threshold of 223, the number of connected components is 3 and the number of holes is 15.

[0108] S27. A persistent digest generation process is performed based on the Betty number sequence set to obtain a topologically persistent barcode. The data structure of the topologically persistent barcode is a set of intervals, each interval containing a topological dimension identifier, a birth threshold, a death threshold, and a device node identifier. The birth threshold and death threshold correspond to the generation and extinction thresholds of topological features in the Betty number sequence set. Specifically, the persistent digest generation process involves traversing the Betty number sequence set, recording the birth and death thresholds of each connected component and hole, and mapping the birth threshold to the death threshold as intervals to be written into the topologically persistent barcode. For example, performing persistent digest generation processing on a set of Betty number sequences yields a topologically persistent barcode containing the following intervals: 89 intervals with a topological dimension of 0 (connected components), such as (birth threshold=25, death threshold=72, persistence=47) and (birth threshold=48, death threshold=195, persistence=147); and 42 intervals with a topological dimension of 1 (holes), such as (birth threshold=72, death threshold=168, persistence=96) and (birth threshold=95, death threshold=223, persistence=128). See also... Figure 5 This is a schematic diagram of topology persistent barcode generation provided in an embodiment of this application. For example... Figure 5As shown in the figure, this image vividly illustrates the dimensionality reduction process from image texture to topological fingerprint. The middle stripe demonstrates the dynamic evolution of connected components and pores on the carrier surface as the filtering threshold changes, showing their continuous "birth" and "death." The lower stripe represents the generated topologically persistent barcode, with the length of each line segment representing the persistence of a topological feature. The physical significance of this image lies in the fact that different carrier materials (such as smooth plasterboard and rough concrete) have drastically different microscopic topological evolution patterns. The persistent barcode, as a "shape summary," can compress these complex material surface structures into a set of intervals insensitive to translation, rotation, and illumination. In harsh engineering environments with varying shooting angles, this method based on topological data analysis (TDA) provides a more stable material characterization capability than traditional texture features, enabling the system to identify the essential structural properties of the carrier even through surface contamination.

[0109] Specifically, S2 removes the device from the visual input through inverse masking, forcing subsequent feature extraction to be calculated only on the carrier background region set. This eliminates the attention bias of the target detection network towards foreground targets, shifting the focus of the review from the presence of the device to whether the surface on which the device is attached meets the load-bearing requirements. Simultaneously, S2 uses carrier grayscale complex sets and homology group calculations to transform the textured surface into comparable topological invariants, avoiding direct reliance on pixel differences caused by color, lighting, and shooting angle. In non-standardized survey photographs, concrete, plasterboard, wood veneer, and metal pipes may be similar in color, but their microstructures lead to different evolutions in porosity and connectivity. Topologically persistent barcodes can stably record this evolution process, allowing for the extraction of material-related evidence even under complex lighting or surface contamination conditions. This provides a scene-dependent mathematical representation for determining the compliance of carrier attributes.

[0110] S3: Map the topologically persistent barcode to the preset material feature space, combine the attribute requirements in the spatial ontology constraint diagram, and generate a multimodal evidence reasoning tensor through orthogonal projection calculation;

[0111] Furthermore, the step of mapping the topologically persistent barcode to a preset material feature space, and generating a multimodal evidence reasoning tensor through orthogonal projection, in conjunction with the attribute requirements in the spatial ontology constraint graph, includes the following steps:

[0112] S31, acquire the topologically persistent barcode and perform persistent vectorization processing to obtain a persistent feature vector set. The data structure of the persistent feature vector set is a vector list, where each vector contains a fixed-length numerical feature and is associated with a device node identifier. The persistent vectorization processing refers to converting the interval set of the topologically persistent barcode into a numerical vector representation that can be used for measurement and projection operations. Specifically, the persistent vectorization processing is performed as follows: the interval length distribution of the topologically persistent barcode is accumulated according to the topological dimension and encoded into a persistent image raster; then, the persistent image raster is expanded by rows and columns to obtain persistent feature vectors. The raster resolution of the persistent image raster is determined by the value span of the birth threshold and death threshold in the topologically persistent barcode. The raster resolution is determined by performing equal-frequency binning on the joint distribution of the birth threshold and death threshold, and using the number of bins as the raster side length. For example, the joint distribution of birth and death thresholds in a certain topological persistent barcode is binned at equal frequencies. Assuming the birth threshold range is [0, 255] and the death threshold range is [0, 255], and the number of bins is 20, then the grid side length is 20, and the size of the persistent image grid is 20 × 20 = 400 grid units. The 0-dimensional topology and 1-dimensional topology are encoded separately and then concatenated. After unfolding, a persistent feature vector with a length of 800 is obtained, where the first 400 dimensions correspond to the persistent distribution of the 0-dimensional topology (connected components), and the last 400 dimensions correspond to the persistent distribution of the 1-dimensional topology (holes).

[0113] S32: Receive the spatial ontology constraint graph and perform attribute requirement extraction processing on the spatial ontology constraint graph. The attribute requirement extraction processing yields a carrier requirement semantic vector set. The data structure of the carrier requirement semantic vector set is a vector list, where each vector contains a fixed-length semantic embedding and is associated with a device node identifier. Further, S32 includes the following steps:

[0114] S321, read the parent carrier primitive identifier associated with the device node identifier from the attribute dictionary set of the spatial ontology constraint diagram, and extract the material attribute text and mechanical grade mark from the carrier attribute table to obtain the carrier requirement text set; the data structure of the carrier requirement text set is a list of text records, which at least includes the device node identifier, material attribute text and mechanical grade mark.

[0115] S322, perform normalized encoding processing on the carrier requirement text set to obtain a carrier requirement semantic token set; the data structure of the carrier requirement semantic token set is a token sequence list. Specifically, the normalized encoding processing method is as follows: map the material attribute text and mechanical grade mark to standard entries in the engineering controlled vocabulary, and convert the standard entries into token sequences; the engineering controlled vocabulary refers to a standardized vocabulary formed by supplementing the engineering dictionary with synonym merging rules and negation rules, used to eliminate ambiguity caused by near-synonyms such as concrete and reinforced concrete. For example, if the material attribute text in a carrier requirement text set is "C30 reinforced concrete", after normalized encoding processing, firstly, according to the synonym merging rules, "reinforced concrete" is merged into the standard entry "reinforced concrete", and "C30" is merged into the standard entry "C30 strength grade", and then converted into a token sequence [TOK_MAT_RC,TOK_STR_C30], which is written into the carrier requirement semantic token set.

[0116] S323, perform semantic embedding processing on the set of semantic tokens required by the carrier, and obtain the set of semantic vectors required by the carrier through semantic embedding processing; the semantic embedding processing method is as follows: use a text embedding model oriented to engineering terms to map the token sequence into a fixed-length semantic vector, and perform linear normalization on the semantic vector to make it satisfy the numerical stability of subsequent orthogonal projection calculation.

[0117] The text embedding model for engineering terminology specifically employs the Sentence-BERT text embedding model. The input data of the Sentence-BERT text embedding model consists of a sequence of tokens from a semantic token set required by the carrier. The input data structure is an integer sequence, where each integer is the index number of the token in the vocabulary. The maximum sequence length is 128 tokens. The output data of the Sentence-BERT text embedding model is a fixed-length semantic vector. The output data structure is a one-dimensional floating-point vector with a vector dimension of 256. The Sentence-BERT text embedding model is trained using engineering specifications, construction methods, and enterprise drawing annotations as corpora for domain-adaptive fine-tuning. Its output vector maintains the separability of material categories and load-bearing levels in the semantic space.

[0118] For example, semantic embedding processing is performed on the token sequence [TOK_MAT_RC,TOK_STR_C30], and a 256-dimensional semantic vector is output by the Sentence-BERT text embedding model. After L2 normalization of the vector, a carrier requirement semantic vector with a modulus of 1 is obtained and written into the carrier requirement semantic vector set.

[0119] S33, obtain a preset material feature space and perform feature space alignment processing to obtain an alignment transformation parameter set; the data structure of the preset material feature space is a vector space definition object, including a material category prototype vector set, a metric function definition, and a dimension description; the data structure of the material category prototype vector set is a vector list, where each vector corresponds to a material category and carries a material category identifier. The preset material feature space refers to a feature space that unifies the topological texture representation and semantic representation of common installation carrier materials into the same comparison domain, used to measure the degree to which the carrier's visual evidence supports the design attribute requirements.

[0120] Specifically, the construction method of the material category prototype vector set is as follows: collect the carrier background region set of confirmed material categories in the sample library, obtain the corresponding persistent feature vector set according to S25 to S31, and calculate the center vector of the persistent feature vector set for each material category as the material category prototype vector. For example, a set of carrier background regions with confirmed material categories is collected in the sample library. Assume 150 samples are collected for the concrete category, 120 for the brick wall category, 80 for the gypsum board category, and 95 for the metal surface category. After processing according to S25 to S31, the corresponding persistent feature vector set is obtained. The center vector of the 150 persistent feature vectors for the concrete category is calculated to obtain an 800-dimensional prototype vector for the concrete material category, and the material category is labeled as MAT_CONCRETE. The center vector of the 120 persistent feature vectors for the brick wall category is calculated to obtain a prototype vector for the brick wall material category, and the material category is labeled as MAT_BRICK. The center vector of the 80 persistent feature vectors for the gypsum board category is calculated to obtain a prototype vector for the gypsum board material category, and the material category is labeled as MAT_GYPSUM. The center vector of the 95 persistent feature vectors for the metal surface category is calculated to obtain a prototype vector for the metal surface material category, and the material category is labeled as MAT_METAL.

[0121] The feature space alignment process is implemented using a linear affine transformation. The data structure of the alignment transformation parameter set is a parameter record object, containing two fields: a transformation matrix and an offset vector. The data structure of the transformation matrix is ​​a two-dimensional floating-point matrix with the shape (D_out, D_in), where D_in is the original dimension of the persistent feature vector and D_out is the target dimension of the material feature space. The data structure of the offset vector is a one-dimensional floating-point vector with a length of D_out. The alignment transformation parameter set is determined by using the material category prototype vector set as an anchor point and employing the Procrustes analysis method to calculate the linear transformation parameters that optimally align the persistent feature vector set with the material category prototype vector set in the Euclidean distance sense.

[0122] S34, based on the persistent feature vector set, the carrier requirement semantic vector set and the alignment transformation parameter set, perform orthogonal projection calculation processing to obtain a projection evidence pair set; the data structure of the projection evidence pair set is a record list, and each record contains a device node identifier, visual projection coefficient, semantic projection coefficient and projection residual.

[0123] Specifically, the orthogonal projection calculation process is as follows: First, the persistent feature vectors are mapped to a preset material feature space according to the alignment transformation parameter set. Then, the mapped persistent feature vectors are projected onto the constraint subspaces corresponding to the set of carrier requirement semantic vectors to obtain visual projection coefficients. Simultaneously, the carrier requirement semantic vectors are projected onto the material subspace spanned by the set of material category prototype vectors to obtain semantic projection coefficients. Finally, the energy of the persistent feature vectors on the orthogonal complement of the constraint subspace is calculated as the projection residual. The constraint subspace refers to the subspace spanned by the carrier requirement semantic vectors and their nearest neighbor semantic vectors, used to express the material semantic neighborhood allowed by the design requirements. The nearest neighbor semantic vectors are selected by searching for standard terms that have a hierarchical relationship with the material attribute text in the engineering controlled vocabulary and taking their corresponding embedding vectors as the nearest neighbor semantic vectors.

[0124] In a preferred embodiment, the constraint subspace is constructed as follows: Let the carrier require a semantic vector of... The nearest neighbor semantic vector is ,Will Arranged in columns to form a matrix ,in Let k be the vector dimension and k be the number of nearest neighbor semantic vectors; for the matrix Performing QR decomposition yields orthogonal basis matrices. ,in To constrain the rank of the subspace, we construct the constrained subspace accordingly. .

[0125] In a preferred embodiment, the formula for calculating the visual projection coefficient is as follows:

[0126] ;

[0127] in, Represents the visual projection coefficient, with a value range of 100%. ; This represents the persistent feature vector after alignment transformation; Denotes the orthogonal basis matrix of the constrained subspace. Represents an orthogonal basis matrix Transpose of; This represents the L2 norm of a vector.

[0128] In a preferred embodiment, the formula for calculating the projection residual is as follows:

[0129] ;

[0130] in, Represents the projection residual, and represents the energy of the persistent eigenvector on the orthogonal complement of the constrained subspace; This represents the persistent feature vector after alignment transformation; This represents the orthogonal projection operator onto the constrained subspace.

[0131] For example, a device node is identified as "DEV-008", and its corresponding persistent feature vector is an 800-dimensional vector v. The carrier requires a 256-dimensional semantic vector u (corresponding to the material attribute text "load-bearing concrete wall"). After alignment transformation, v is mapped to a preset material feature space to obtain v'. Standard terms with hierarchical relationships to "load-bearing concrete wall" are retrieved from the engineering controlled vocabulary, including "reinforced concrete wall" and "C30 concrete". Their corresponding embedded vectors are taken as nearest neighbor semantic vectors u1 and u2. A matrix U is constructed from u, u1, and u2, and QR decomposition is performed to obtain an orthogonal basis matrix Q, which spans a constraint subspace S. The visual projection coefficient is calculated to be 0.78 according to the visual projection coefficient calculation formula. The semantic projection coefficient is calculated to be 0.92 when the projection of u on the material subspace spanned by the material category prototype vector set is calculated. The projection residual is calculated to be 0.35 according to the projection residual calculation formula.

[0132] S35, based on projected evidence, perform evidence tensor construction processing on the set to obtain a multimodal evidence inference tensor; the data structure of the multimodal evidence inference tensor is a three-dimensional tensor object, whose first dimension is the visual feature confidence dimension, the second dimension is the semantic requirement matching degree dimension, and the third dimension is the conflict uncertainty dimension, with the device node identifier as the index key. Further, S35 includes the following steps:

[0133] S351, the confidence level of the visual projection coefficients in the projection evidence set is calibrated to obtain the visual feature confidence component, the semantic projection coefficients are calibrated to obtain the semantic requirement matching component, and the projection residuals are calibrated to obtain the conflict uncertainty component, thereby obtaining the evidence triple set; the data structure of the evidence triple set is a record list, and each record contains a device node identifier, a visual feature confidence component, a semantic requirement matching component, and a conflict uncertainty component.

[0134] The confidence calibration, matching degree calibration, and uncertainty calibration are all implemented using a parameterized Sigmoid mapping function. The parameterized Sigmoid mapping function is determined as follows: collect manually confirmed compliance and non-compliance results from historical audit samples, and use the results as a supervision signal to fit a monotonic mapping function. This results in a higher visual feature confidence component when the visual projection coefficient is closer to the region corresponding to the material category prototype vector, a higher semantic requirement matching degree component when the semantic projection coefficient is more concentrated in the semantic neighborhood corresponding to the design requirements, and a higher conflict uncertainty component when the projection residual is larger.

[0135] In a preferred embodiment, the confidence scaling function for the visual feature confidence component is specifically:

[0136] ;

[0137] in, This represents the confidence component of visual features, with a value range of [value range missing]. ; Represents the visual projection coefficient, with a value range of 100%. ; The slope parameter represents the confidence level calibration function and controls the steepness of the function curve; The center point parameter of the confidence scaling function represents the value at which the visual projection coefficient reaches a confidence component of 0.5. (·) is an exponential function.

[0138] In a preferred embodiment, the matching degree scaling function of the semantic requirement matching degree component is specifically as follows:

[0139] ;

[0140] in, This represents the semantic requirement matching degree component, with a value range of [value range missing]. ; Represents the semantic projection coefficients, with a value range of 1. ; The slope parameter represents the matching degree calibration function; This represents the center point parameter of the matching degree calibration function.

[0141] In a preferred embodiment, the uncertainty calibration function for the conflicting uncertainty component is specifically:

[0142] ;

[0143] in, Represents the conflict uncertainty component, with a value range of . ; This represents the projection residual, with a value range of [value range missing]. ; It represents the attenuation coefficient of the uncertainty calibration function, which controls the sensitivity of the projected residual to conflicting uncertainty components.

[0144] For example, 500 manually verified compliant samples and 200 non-compliant samples are collected from historical audit samples. A parameterized Sigmoid mapping function is fitted using the compliance and non-compliance results as monitoring signals; the parameters of the fitted confidence scaling function are... , The parameters of the matching degree calibration function are: , The parameters of the uncertainty calibration function are For the device node identifier "DEV-008", the visual projection coefficient of 0.78 is substituted into the confidence scaling function. The confidence component of the visual features is 0.86, and the semantic projection coefficient is 0.92. These values ​​are then substituted into the matching score function. The semantic matching degree component is 0.97, and the projection residual is 0.35. Substituting these values ​​into the uncertainty calibration function, we obtain the following: The conflict uncertainty component is 0.41.

[0145] S352, the set of evidence triples is written into the corresponding slice of the three-dimensional tensor object according to the device node identifier, and a multimodal evidence reasoning tensor is obtained through writing; the writing method is: locating the tensor slice index by the device node identifier, and filling values ​​in the visual feature confidence dimension, semantic requirement matching degree dimension, and conflict uncertainty dimension. See Figure 6 This is a schematic diagram of the multimodal evidence reasoning tensor provided in the embodiments of this application. For example... Figure 6 As shown, the diagram illustrates the high-dimensional spatial structure of evidence fusion using a three-dimensional cube model. The three coordinate axes of the cube represent visual feature confidence, semantic requirement matching degree, and conflict uncertainty, respectively, while the data points inside represent the review status of different device nodes. Combined with the table view on the right, the distribution of evidence components for each node can be clearly seen. Unlike traditional two-dimensional similarity matching, this embodiment introduces the crucial dimension of conflict uncertainty (i.e., the residual component in orthogonal projection). This means that the system can not only determine whether a match exists but also quantify the intensity of the conflict through a third dimension. For example, when the visual evidence is extremely certain that the material is lightweight brick, but the design semantic requirement is concrete, the tensor model will not simply output a low matching degree but will instead present a high value on the conflict dimension. This tensor structure provides an interpretable mathematical basis for subsequently distinguishing between insufficient evidence and substantive violations, greatly reducing the risk of missed judgments in engineering reviews.

[0146] Specifically, S3 uses a pre-defined material feature space to place the persistent feature vector set and the carrier requirement semantic vector set within the same measurable domain. It then explicitly separates the compatible and conflicting parts of visual evidence and semantic constraints through orthogonal projection calculations, overcoming the difficulty of directly comparing the symbolic attributes of vector paper with the pixel textures on-site. If only simple similarity calculations are performed, visual noise or semantic synonyms can mask conflicts in the average value, leading to situations where the location is correct but the carrier attributes are non-compliant, potentially resulting in misjudgment. However, the multimodal evidence reasoning tensor output by S3 simultaneously retains the visual feature confidence component, the semantic requirement matching component, and the conflict uncertainty component. This allows subsequent judgments to not only see whether there is a match, but also why there is a mismatch and whether the conflict intensity originates from the visual or semantic side, thus providing interpretable and traceable evidence for engineering review.

[0147] S4: Analyze the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determine the compliance of the installation carrier based on the preset security threshold, and output the rectification strategy.

[0148] Furthermore, the process of analyzing the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determining the compliance of the installation carrier based on a preset security threshold, and outputting a rectification strategy includes the following steps:

[0149] S41, obtain the multimodal evidence reasoning tensor and perform evidence component unpacking processing to obtain a compliance judgment input table. The data structure of the compliance judgment input table is a table, and the table fields include at least the device node identifier, visual feature confidence component, semantic requirement matching degree component, and conflict uncertainty component. Specifically, the evidence component unpacking processing is performed as follows: traverse the slices of the multimodal evidence reasoning tensor according to the device node identifier, and write the values ​​of the visual feature confidence dimension, semantic requirement matching degree dimension, and conflict uncertainty dimension into the compliance judgment input table.

[0150] S42, obtain a preset safety threshold and perform threshold consistency verification processing to obtain a threshold set; the data structure of the threshold set is a record object, which includes at least a matching degree compliance threshold and a conflict uncertainty threshold. The preset safety threshold refers to the threshold rule used to map evidence components to qualified or unqualified conclusions under the premise of meeting the engineering bearing safety and the controllable risk of audit misjudgment. Specifically, the preset safety threshold is determined as follows: collect carrier violation samples that have undergone rectification and compliant samples that have passed acceptance from historical projects, use the upper quantile of the violation sample on the semantic requirement matching degree component as the lower bound of the matching degree compliance threshold, use the upper quantile of the compliant sample on the conflict uncertainty component as the upper bound of the conflict uncertainty threshold, and select a threshold combination between the lower bound and the upper bound that satisfies the requirement that the violation sample can be covered; the upper quantile is determined by the enterprise risk preference rule, which is formulated by the safety management department according to the structural failure consequence level and solidified into a configuration file. For example, 180 samples of carrier violations that have undergone rectification and 620 samples of compliant carriers that have passed acceptance were collected from historical projects. The distribution of the violation samples on the semantic requirement matching degree component was statistically analyzed, and its 90th quantile (upper quantile) was 0.35. Therefore, the lower bound of the matching degree compliance threshold was set to 0.35. The distribution of the compliant samples on the conflict uncertainty component was statistically analyzed, and its 90th quantile (upper quantile) was 0.42. Therefore, the upper bound of the conflict uncertainty threshold was set to 0.42. Combined with the enterprise risk preference rule (the safety management department sets the structural failure consequence level as level two, corresponding to a 5% tolerance for missed detection rate), the final matching degree compliance threshold was determined to be 0.40, and the conflict uncertainty threshold was set to 0.45.

[0151] S43, based on the compliance judgment input table and threshold set, perform carrier compliance judgment processing to obtain a carrier compliance result table; the data structure of the carrier compliance result table is a table, and the fields include at least the device node identifier, compliance flag, and non-compliance type flag. Further, S43 includes the following steps:

[0152] S431, Perform a matching degree threshold comparison process on each record in the compliance judgment input table, and obtain a matching degree conclusion mark through the matching degree threshold comparison process; the matching degree threshold comparison process is as follows: compare the semantic requirement matching degree component with the matching degree compliance threshold. When the semantic requirement matching degree component is less than the matching degree compliance threshold, output the matching degree conclusion mark as failing. When the semantic requirement matching degree component is greater than or equal to the matching degree compliance threshold, output the matching degree conclusion mark as passing.

[0153] S432, perform uncertainty threshold comparison processing on each record in the compliance judgment input table, and obtain uncertainty conclusion label through uncertainty threshold comparison processing; the uncertainty threshold comparison processing method is as follows: compare the conflict uncertainty component with the conflict uncertainty threshold, when the conflict uncertainty component is greater than the conflict uncertainty threshold, output uncertainty conclusion label as conflict significant, when the conflict uncertainty component is less than or equal to the conflict uncertainty threshold, output uncertainty conclusion label as conflict controllable.

[0154] S433, perform rule fusion processing on the matching degree conclusion mark and the uncertainty conclusion mark to obtain compliance mark and non-compliance type mark and write them into the carrier compliance result table; the rule fusion processing method is as follows: when the matching degree conclusion mark is passed and the uncertainty conclusion mark is controllable conflict, output compliance mark as qualified and output non-compliance type mark as empty; when the matching degree conclusion mark is failed and the uncertainty conclusion mark is significant conflict, output compliance mark as unqualified and output non-compliance type mark as carrier material conflict; when the matching degree conclusion mark is failed and the uncertainty conclusion mark is controllable conflict, output compliance mark as unqualified and output non-compliance type mark as insufficient carrier evidence; when the matching degree conclusion mark is passed and the uncertainty conclusion mark is significant conflict, output compliance mark as unqualified and output non-compliance type mark as local carrier anomaly. For example, for device node identifier "DEV-012", the semantic requirement matching degree component recorded in its compliance judgment input table is 0.28 and the conflict uncertainty component is 0.68. The semantic requirement matching degree component 0.28 is compared with the matching degree compliance threshold 0.40. Since 0.28 < 0.40, the matching degree conclusion is marked as unqualified. The conflict uncertainty component 0.68 is compared with the conflict uncertainty threshold 0.45. Since 0.68 > 0.45, the uncertainty conclusion is marked as significant conflict. After performing rule fusion processing, the matching degree conclusion is marked as unqualified and the uncertainty conclusion is marked as significant conflict. Therefore, the compliance is marked as unqualified and the unqualified type is marked as carrier material conflict.

[0155] S44, based on the carrier compliance result table and the spatial ontology constraint diagram, performs rectification strategy generation processing to obtain a rectification strategy set; the data structure of the rectification strategy set is a record list, and each record contains a device node identifier, a parent carrier element identifier, a rectification text instruction, and a risk warning marker. Further, S44 includes the following steps:

[0156] S441, filter the carrier compliance result table to obtain the equipment node identifiers marked as non-compliant, read the corresponding parent carrier element identifier and material attribute text based on the attribute dictionary set of the spatial ontology constraint graph, and obtain the rectification constraint source table by reading; the data structure of the rectification constraint source table is a table, and the fields include at least the equipment node identifier, the parent carrier element identifier, and the material attribute text.

[0157] S442, the rectification constraint source table and non-conformity type mark are templated and generated. The rectification text instruction is obtained through the templated generation process and written into the rectification strategy set. The templated generation process is as follows: when the non-conformity type mark is a carrier material conflict, the rectification text instruction includes the need to verify the material attribute text of the structure corresponding to the parent carrier element identifier and instructs that the equipment shall not be installed on the surface of the component that does not meet the material attribute text; when the non-conformity type mark is insufficient carrier evidence, the rectification text instruction includes the need to take additional photos of the installation surface details corresponding to the carrier background area set and instructs that the annular neighborhood outside the extended mask of the main body of the equipment should be covered; when the non-conformity type mark is a local carrier anomaly, the rectification text instruction includes the need to conduct on-site verification of the structure corresponding to the parent carrier element identifier and instructs that there be any hollow areas, surface obstruction, or non-load-bearing component coverage.

[0158] S443, based on the conflict uncertainty component in the multimodal evidence reasoning tensor, risk classification processing is performed on the rectification strategy set. Risk warning tags are obtained through risk classification processing and written into the rectification strategy set. The data structure of the risk warning tags is an enumerated value set, including high-risk warnings and general risk warnings. The risk classification processing method is as follows: when the conflict uncertainty component exceeds the conflict uncertainty threshold and the material category prototype vector pointed to by the visual feature confidence component has a mutually exclusive relationship with the material attribute text, a high-risk warning is output; otherwise, a general risk warning is output. The mutually exclusive relationship is determined by the mutually exclusive word pair rules in the engineering controlled vocabulary, and the mutually exclusive word pair rules are obtained by organizing the material substitution relationships explicitly prohibited in the structural safety specifications. See also... Figure 7 This is a schematic diagram illustrating the generation of the rectification strategy provided in an embodiment of this application. For example... Figure 7As shown in the diagram, the final decision output of the review system is presented in the form of a work order card. The card clearly lists the problematic device node (RRU-01), a comparison of the measured values ​​of the parent carrier with design requirements (lightweight partition wall vs. load-bearing concrete wall), and clear rectification instructions and risk warnings. This diagram demonstrates the application value of this closed-loop technical solution: based on the spatial ontological constraints and multimodal reasoning conclusions established in the preceding steps, the generated strategy is no longer a general non-compliance label, but can accurately trace back to the original constraints of the design drawings, indicating "what it should have been installed on" and "what it is now installed on." In particular, the high-risk warning function can automatically issue early warnings for serious hidden dangers such as load-bearing structural failure. This visualized rectification strategy card can be directly distributed to construction personnel terminals, guiding them to conduct targeted on-site verification or dismantling and reinstallation, thereby effectively ensuring the structural safety of the communication base station construction.

[0159] S45, based on the rectification strategy set, perform audit result output processing to obtain an audit report data package. The data structure of the audit report data package is a serializable data object, containing index information for the carrier compliance result table, the rectification strategy set, the equipment detection frame set, and the carrier background area set. Specifically, the audit result output processing method is as follows: superimpose the bounding boxes of the equipment detection frame set and the carrier background area set to generate visual annotation information, and encapsulate it together with the rectification text instructions into the audit report data package, so as to locate the image area with carrier violation risk and the corresponding design constraint source in the audit system interface.

[0160] Specifically, S4 transforms the multimodal evidence reasoning tensor into executable engineering review decisions. The key lies in simultaneously reading the semantic requirement matching degree component and the conflict uncertainty component and fusing them into rules, thereby avoiding missed judgments caused by only considering the matching degree and ignoring the conflict intensity. On construction sites, equipment may be installed on components that look similar but have completely different load-bearing logic. For example, a lightweight partition wall covered with a decorative layer and a load-bearing wall may have similar textures in certain areas. If only binary conclusions are output, the system will be forced to make arbitrary judgments when evidence is insufficient. S4 further subdivides non-compliance types into carrier material conflicts, insufficient carrier evidence, and local carrier anomalies, mapping them to different rectification text instructions and supplementary evidence paths. This allows the review loop to not only point out non-compliance but also to indicate that the problem stems from design requirement conflicts, insufficient on-site evidence, or local carrier anomalies, thus better aligning with on-site handling procedures and responsibility interfaces.

[0161] The above S1 to S4 form a closed-loop processing link from design vector constraints to on-site image evidence. The spatial ontology constraint graph ensures that the dependency relationship between the device and the parent carrier, as well as the material attribute text and mechanical grade mark, are not lost in the data flow. The topological persistent barcode ensures that the texture structure of the carrier background area set is stably represented under the reverse masking mechanism. The multimodal evidence reasoning tensor ensures that visual evidence and semantic constraints can be measured simultaneously in the same tensor container and the source of conflict can be explicitly separated. The preset security threshold and non-compliance type mark ensure that the review conclusion can be implemented into a set of rectification strategies. For the hidden defect of correct location but non-compliant carrier attributes in non-standard survey scenarios, traditional methods focus on aligning the main body of the equipment with the two-dimensional position while ignoring the carrier background, making it difficult to expose carrier conflicts during the review stage. This solution concentrates computational resources on the mounting surface through a reverse masking mechanism, transforms the mounting surface from pixel appearance into comparable topological invariants through homology group calculation, and then separates the semantic neighborhood allowed by design requirements and the orthogonal residuals of visual evidence through orthogonal projection calculation. This allows carrier violations to be amplified and presented in the conflict uncertainty component even when the equipment is correctly positioned, and is then stably intercepted by threshold rules to generate actionable rectification text instructions. Logically, this improves the automatic verification capability and interpretability of the conclusions regarding carrier safety constraints.

[0162] For example, after receiving the wireless design vector drawing, the audit system outputs a spatial ontology constraint diagram according to S1 and obtains the constraint that a certain device node identifier and its parent carrier element identifier and material attribute text are load-bearing concrete walls; after receiving the survey photos, the system outputs the topology persistent barcode corresponding to the device node identifier according to S2; the system converts the topology persistent barcode into a persistent feature vector set according to S3 and performs orthogonal projection calculation with the carrier requirement semantic vector set in the preset material feature space to obtain a multimodal evidence inference tensor; the system parses the multimodal evidence inference tensor according to S4 and determines that the conflict uncertainty component is significant and the semantic requirement matching degree component fails according to the threshold set, thereby outputting non-compliance in the carrier compliance result table, and outputting a rectification text instruction in the rectification strategy set that the device needs to be reinstalled to the structural surface corresponding to the parent carrier element identifier that meets the requirements of load-bearing concrete walls and indicating high risk. At the same time, the equipment detection box set and the carrier background area set are superimposed in the audit report data package to assist in on-site verification and rectification positioning.

[0163] Example 2:

[0164] This embodiment, based on Embodiment 1, provides a wireless design review system based on image recognition, such as... Figure 8 As shown, it includes:

[0165] Spatial ontology constraint diagram construction module: used to parse the layers and annotation information of wireless design vector drawings, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram;

[0166] Topological persistent barcode generation module: used to locate the area to be detected in the survey photo based on the spatial ontology constraint map, extract background texture features using the inverse masking mechanism, and generate a topological persistent barcode through homology group calculation;

[0167] Multimodal Evidence Reasoning Tensor Generation Module: This module maps the topologically persistent barcode to a preset material feature space, combines the attribute requirements in the spatial ontology constraint diagram, and generates a multimodal evidence reasoning tensor through orthogonal projection calculation.

[0168] Carrier Compliance Judgment and Rectification Strategy Output Module: Used to parse the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determine the compliance of the installation carrier based on the preset security threshold, and output the rectification strategy.

Claims

1. A wireless design review method based on image recognition, characterized in that, The method includes: S1: Analyze the layers and annotation information of the wireless design vector drawing, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram; S2: Based on the spatial ontology constraint map, locate the area to be detected in the survey photo, extract the background texture features using the inverse masking mechanism, and generate a topologically persistent barcode through homology group calculation; S3: Map the topologically persistent barcode to the preset material feature space, combine the attribute requirements in the spatial ontology constraint diagram, and generate a multimodal evidence reasoning tensor through orthogonal projection calculation; S4: Analyze the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determine the compliance of the installation carrier based on the preset security threshold, and output the rectification strategy.

2. The wireless design review method based on image recognition according to claim 1, characterized in that, The process of analyzing the layers and annotations of wireless design vector graphics, extracting the spatial relationships between device elements and carrier elements, and constructing a spatial ontology constraint diagram includes the following steps: Obtain the wireless design vector graphics and perform vector entity deconstruction processing to obtain a vector entity set and a layer index table through vector entity deconstruction processing; Based on the layer index table, perform layer semantic classification processing on the vector entity set to obtain the device candidate primitive set and the carrier candidate primitive set through layer semantic classification processing; Perform device element instantiation processing on the device candidate element set, and obtain the device element set and device attribute table through device element instantiation processing; Carrier candidate primitives are instantiated, and the carrier primitive set and carrier attribute table are obtained through the instantiation process. Based on the set of device primitives and the set of carrier primitives, the mounting relationship inference process is performed, and the mounting relationship set is obtained through the mounting relationship inference process. Based on the set of device primitives, the set of carrier primitives, the device attribute table, the carrier attribute table and the set of mounting relationships, a graph structure encapsulation process is performed to obtain a spatial ontology constraint graph.

3. The wireless design review method based on image recognition according to claim 2, characterized in that, The process of instantiating carrier primitives into a candidate primitive set and obtaining the carrier primitive set and carrier attribute table through instantiation includes the following steps: The closure determination and boundary reconstruction processes are performed on the polyline entities in the candidate primitive set of the carrier, and the candidate set of carrier boundaries is obtained through the closure determination and boundary reconstruction processes. The candidate set of carrier boundaries and the reference entity of structural component blocks are subjected to component fusion processing to obtain the set of carrier primitives. Perform carrier attribute extraction processing on the text annotation objects in the wireless design vector paper to obtain the carrier attribute table.

4. The wireless design review method based on image recognition according to claim 1, characterized in that, The process of locating the region to be detected in the survey photograph based on the spatial ontology constraint map, extracting background texture features using an inverse masking mechanism, and generating a topologically persistent barcode through homology group calculation includes the following steps: Acquire survey photos and perform image preprocessing to obtain a standardized set of survey photos and camera geometric parameters. Receive the spatial ontology constraint diagram and perform an audit task extraction process on the spatial ontology constraint diagram. The audit task extraction process yields a list of device nodes to be audited and a list of carrier nodes to be audited. Based on the standardized survey photos, the device visual positioning process is performed on the list of device nodes to be reviewed, and the device visual positioning process is used to obtain the set of device detection boxes and the set of device segmentation masks. Inverse masking is performed based on the device segmentation mask set to obtain the carrier background region set. A grayscale topological filtering process is performed on the carrier background region set to obtain a grayscale complex set of the carrier. Homology group calculation is performed on the carrier gray-level complex set to obtain the Betti number sequence set. Persistent digest generation is performed based on the set of Betty number sequences, and a topologically persistent barcode is obtained through persistent digest generation.

5. The wireless design review method based on image recognition according to claim 4, characterized in that, Based on standardized survey photos, visual positioning processing is performed on the list of equipment nodes to be reviewed. The process of obtaining the set of equipment detection boxes and the set of equipment segmentation masks through visual positioning processing includes the following steps: The list of devices to be reviewed and the device type text in the device attribute table are mapped to a set of device detection categories. The set of device detection categories drives the target detection network to perform device detection on standardized survey photos, resulting in an initial set of device detection boxes. A candidate box cleaning process based on nonmaximum suppression is performed on the initial set of device detection boxes to obtain the set of device detection boxes. Based on the set of equipment detection frames, instance segmentation processing is performed on standardized survey photos to obtain a set of equipment segmentation masks.

6. The wireless design review method based on image recognition according to claim 1, characterized in that, Mapping the topologically persistent barcode to a predefined material feature space, and combining the attribute requirements in the spatial ontology constraint diagram, the multimodal evidence reasoning tensor is generated through orthogonal projection, including the following steps: Obtain the topologically persistent barcode and perform persistent vectorization processing to obtain a set of persistent feature vectors; Receive the spatial ontology constraint graph and perform attribute requirement extraction processing on the spatial ontology constraint graph. The set of carrier requirement semantic vectors is obtained through the attribute requirement extraction processing. Obtain the preset material feature space and perform feature space alignment processing to obtain the alignment transformation parameter set through feature space alignment processing; Orthogonal projection calculation is performed based on the persistent feature vector set, the carrier requirement semantic vector set, and the alignment transformation parameter set to obtain the projection evidence pair set. Based on projected evidence, the set of evidence tensors is constructed and processed to obtain a multimodal evidence reasoning tensor.

7. The wireless design review method based on image recognition according to claim 6, characterized in that, Receiving the spatial ontology constraint graph and performing attribute requirement extraction processing on the spatial ontology constraint graph, and obtaining the carrier requirement semantic vector set through attribute requirement extraction processing includes the following steps: Read the parent carrier element identifier associated with the device node identifier from the attribute dictionary set of the spatial ontology constraint diagram, and extract the material attribute text and mechanical grade mark from the carrier attribute table to obtain the carrier requirement text set. The carrier requires a set of texts to be normalized and encoded, and the set of semantic tokens required by the carrier is obtained through the normalization and encoding process. The semantic embedding process is performed on the set of semantic tokens required by the carrier, and the set of semantic vectors required by the carrier is obtained through the semantic embedding process.

8. The wireless design review method based on image recognition according to claim 1, characterized in that, The steps involved in analyzing the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determining the compliance of the installation carrier based on a preset safety threshold, and outputting rectification strategies are as follows: Obtain the multimodal evidence reasoning tensor and perform evidence component unpacking processing to obtain the compliance judgment input table through evidence component unpacking processing; Obtain the preset security threshold and perform threshold consistency verification processing to obtain the threshold set through threshold consistency verification processing; Based on the compliance judgment input table and threshold set, the carrier compliance judgment process is performed, and the carrier compliance result table is obtained through the carrier compliance judgment process; Based on the carrier compliance result table and the spatial ontology constraint diagram, the rectification strategy generation process is performed to obtain a set of rectification strategies. The audit results are processed based on the set of rectification strategies, and an audit report data package is obtained through the audit results output processing.

9. The wireless design review method based on image recognition according to claim 8, characterized in that, Based on the compliance judgment input table and threshold set, the carrier compliance judgment process is performed. The carrier compliance result table is obtained through the carrier compliance judgment process, including the following steps: Perform a matching degree threshold comparison process on each record in the compliance judgment input table, and obtain a matching degree conclusion marker through the matching degree threshold comparison process; For each record in the compliance judgment input table, perform uncertainty threshold comparison processing to obtain an uncertainty conclusion label. The matching degree conclusion marker and the uncertainty conclusion marker are fused according to rules. The compliance marker and the non-compliance type marker are obtained through the rule fusion process and written into the carrier compliance result table.

10. A wireless design review system based on image recognition, used to implement the wireless design review method based on image recognition as described in any one of claims 1-9, characterized in that, The system includes: Spatial ontology constraint diagram construction module: used to parse the layers and annotation information of wireless design vector drawings, extract the spatial relationship between device elements and carrier elements, and construct a spatial ontology constraint diagram; Topological persistent barcode generation module: used to locate the area to be detected in the survey photo based on the spatial ontology constraint map, extract background texture features using the inverse masking mechanism, and generate a topological persistent barcode through homology group calculation; Multimodal Evidence Reasoning Tensor Generation Module: This module maps the topologically persistent barcode to a preset material feature space, combines the attribute requirements in the spatial ontology constraint diagram, and generates a multimodal evidence reasoning tensor through orthogonal projection calculation. Carrier Compliance Judgment and Rectification Strategy Output Module: Used to parse the value of the conflict uncertainty dimension in the multimodal evidence reasoning tensor, determine the compliance of the installation carrier based on the preset security threshold, and output the rectification strategy.

Images