Intelligent detection method and system for appearance of steel structure building

Abstract

The invention relates to the technical field of steel structure appearance detection, and discloses a steel structure building appearance intelligent detection method and system, and the method comprises the steps: collecting multi-scene existing steel structure appearance defect images, marking the images, dividing a data set, training an AI model to preset precision based on a deep learning algorithm, and storing the AI model to a cloud; setting inspection parameters according to the building characteristics, and generating an unmanned aerial vehicle flight path; the unmanned aerial vehicle carries an automatic focusing device to collect images according to a path, synchronously record shooting time and GPS parameters and upload the parameters to the cloud in real time; the cloud calls an AI model to identify image defects, outputs defect data and synchronizes the defect data to the management platform; the platform automatically generates a report based on an industry template, archives an image, a result and the report and establishes a'building-detection batch-defect 'association; and constructing a digital twinborn model and importing defect data to realize marking and cross-batch change comparison. The method has the advantages of low safety risk, high efficiency, high precision and traceability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel structure appearance detection, in particular to a steel structure building appearance intelligent detection method and system. BACKGROUND

[0002] As the core component of urban infrastructure, existing steel structure buildings are widely used in railway station buildings, sports venues, exhibition centers and other scenes. During long-term use, they are prone to appearance defects such as coating peeling, corrosion, cracks and bolt loss due to environmental erosion and load action. Coating peeling can expose the steel substrate and accelerate corrosion. Corrosion can weaken the strength of the component cross section. Cracks and bolt loss can cause stress concentration and damage the stability of the structure, directly threatening the safety of the building and shortening its service life. Therefore, appearance detection is a key prerequisite for the renovation, operation and maintenance, and post-disaster assessment of existing steel structures.

[0003] Current industry mainstream detection technologies cannot meet the efficiency and accuracy requirements: manual visual inspection requires the use of arm lifting vehicles and scaffolding for high-altitude operations, which poses a risk of falling and mechanical injury. It also relies on personnel experience and is prone to miss millimeter-level minor defects. It takes 3-5 days to detect 10,000 square meters of steel structure, and the labor cost accounts for more than 60%. Although unmanned aerial vehicle + manual real-time detection breaks through the high-altitude limit, the operator and the detector are often the same person, which leads to a high rate of missed detection. Moreover, the detection data needs to be recorded and archived manually, which is prone to loss and the report format is not uniform, making it impossible to trace the historical defect status for subsequent review. In summary, the industry urgently needs a steel structure appearance detection technology that takes into account safety, efficiency, accuracy and traceability to address the pain points of traditional solutions. SUMMARY

[0004] To solve the problems of high safety risk, low efficiency, poor accuracy and weak traceability in traditional steel structure appearance detection, the present application provides a steel structure building appearance intelligent detection method and system, which can realize "unmanned aerial vehicle collection - AI recognition - cloud collaboration - digital twin connection" intelligent detection of the whole process, with high safety, high efficiency and traceable detection data, and visualized defect status, providing scientific support for the whole life cycle operation and maintenance of steel structures.

[0005] In the first aspect, the present application provides a steel structure building appearance intelligent detection method, comprising: Collecting various appearance defect images of existing steel structures in different scenarios, dividing the data set after labeling the defect information, training an AI recognition model based on a deep learning algorithm, and saving it to a cloud server after optimization to meet the preset detection accuracy requirements; Setting inspection parameters according to the characteristics of existing steel structure buildings to automatically generate a flight planning path for the unmanned aerial vehicle that adapts to the appearance detection; The unmanned aerial vehicle carries an image acquisition device with an image automatic focusing function to acquire structural appearance images according to a flight planning path, synchronously records a shooting time and a GPS position parameter corresponding to the images, and uploads the images to a cloud server in real time. After the cloud server pre-processes the received images, calls a trained AI recognition model to perform defect recognition and analysis on the received image data, outputs defect data containing a defect type, a position, and a bounding box coordinate, and synchronously transmits the defect data to a digital management platform. The digital management platform automatically integrates images and defect data to generate a detection report based on a preset industry standard detection report template, simultaneously archives the images, defect recognition results, and the detection report to a database, establishes an association relationship of “building-detection batch-defect”, and is used for historical data tracing. A digital twin model of the existing steel structure building is established, and the defect data are imported into the model for accurate labeling, and changes in defects of different detection batches are supported.

[0006] The embodiment of the present application fundamentally solves the traditional detection pain points by constructing a whole-process detection method of “AI model training-path planning-image acquisition-cloud recognition-report archiving-digital twin connection”. No manual climbing operation is needed, and the risk of falling and mechanical injury is avoided, so that the safety is significantly improved. The AI model cooperates with the unmanned aerial vehicle to shorten the detection period of 10,000 square meters of steel structure from 3-5 days to 1 day, and the efficiency is improved by 3-5 times. The millimeter-level defect recognition capability (miss detection rate ≤5%) far exceeds the visual accuracy of manual work. The association of “building-detection batch-defect” and the digital twin comparison realize the data full life cycle traceability, provide accurate data support for operation and maintenance, and are suitable for various existing steel structure building detection scenes.

[0007] In an optional embodiment, the plurality of appearance defects include rust, coating peeling, cracks, deformation, and bolt missing, the defect data include a defect type, a position, an actual size, and a bounding box coordinate, a data set is divided into a training set, a validation set, and a test set according to a preset ratio, the training set is used for model training, the validation set is used for adjusting model parameters in the training process, and the test set is used for verifying the final detection accuracy of the model.

[0008] By clearly defining the defect type, data dimension, and data set division logic, the comprehensiveness and reliability of AI model training are ensured. Core defects such as corrosion and coating peeling are covered to avoid model recognition blind spots caused by missing sample types; defect data includes size, coordinate, and other key information to provide quantitative basis for subsequent maintenance; training set, validation set, and test set are divided by function to make the model training process controllable: training set ensures the learning foundation, validation set optimizes parameters dynamically, test set verifies the final accuracy, effectively avoiding model overfitting or underfitting, ensuring that the AI recognition module outputs high-precision results in actual detection, and improving the practicality of the overall detection scheme.

[0009] In an optional implementation, the deep learning algorithm adopts the YOLO series model in the convolutional neural network algorithm, and the preset detection accuracy requirement of model training is that the recognition accuracy is ≥95% and the missed detection rate is ≤5%. During model training, the performance is verified by the test set. If the accuracy is not up to standard, the defect samples corresponding to the missing scene are supplemented, and if the positioning accuracy is insufficient, the model training parameters are re-adjusted.

[0010] By selecting the YOLO series model and clearly defining the accuracy standard and optimization strategy, the efficiency and accuracy of AI recognition are ensured. The YOLO series model has positioning and classification capabilities, which is suitable for rapid detection requirements in industrial scenarios, and the processing efficiency is improved by more than 40% compared with other models; the recognition accuracy of ≥95% and the missed detection rate of ≤5% ensure that millimeter-level subtle defects (such as small-area corrosion and single bolt missing) can be accurately captured; targeted optimization strategies (supplementing samples and adjusting parameters) can solve the recognition short board in special scenarios, avoid detection errors caused by environmental differences (such as backlight and humidity), make the model adapt to steel structure detection in different regions and different service life, and enhance the generalization ability of the technical scheme.

[0011] In an optional implementation, the inspection parameters include flight height, shooting interval, and obstacle avoidance distance, wherein the flight height is set to 3-5m from the surface of the steel structure building, the shooting interval is set to take 1 picture every 0.5m, and the obstacle avoidance distance is set to ≥1m from the obstacle. The generated flight path is in the form of a "ring + longitudinal" combined path.

[0012] Through refining the inspection parameters and path forms, the detection quality, efficiency and safety are balanced. The flight height of 3-5 m ensures that the image clearly presents the defect details, while covering enough detection range to avoid frequent adjustment of flight height; the shooting interval of every 0.5 m eliminates the detection blind area through image overlap, especially ensuring that the key parts such as the corner and node of the steel structure are not missed; the obstacle avoidance distance of ≥1 m effectively avoids obstacles such as protruding components and pipelines, reducing the risk of unmanned aerial vehicle collision; the "ring + longitudinal" path combines the advantages of circumferential and longitudinal coverage, reducing 30% of the repeated flight path, shortening the detection time, and ensuring full coverage detection of the facade and top surface of large steel structure buildings (such as factory buildings and bridges).

[0013] In an optional implementation, after the cloud server pre-processes the received images, it calls the trained AI recognition model to perform defect recognition and analysis on the received image data, including: Image preprocessing: deblurring, noise reduction, and uniform size processing are performed on the uploaded images, invalid images that are blocked or blurred are removed, and a unique identification number is associated with the retained valid images, which is bound with the shooting time and GPS position parameters of the images; Defect positioning: the AI recognition model is used to extract features from the valid images, locate the defect area and mark the bounding box, and the bounding box coordinates are accurate to the pixel level; Defect classification: based on the texture and shape features of the defects, the classified defect area is classified, the defect type is identified and output, and the defect type includes at least one of rust, coating peeling, crack, deformation, and bolt missing. Through image preprocessing to remove invalid data, reduce the interference of redundant information on model recognition, and bind time and GPS parameters with unique numbers to realize "image-location-time" full-link traceability; pixel-level bounding box positioning quantifies the defect range and avoids the ambiguity of manual description; the classification method based on texture and shape features can accurately distinguish easily confused defects (such as coating peeling and rust), reducing the false detection rate. The whole process realizes the automation and standardization of defect recognition, which is more than 50 times more efficient than manual analysis, and has high consistency of results, providing a reliable data foundation for subsequent report generation and operation decision-making.

[0014] In an optional implementation, the digital twin model is generated by oblique photography technology or three-dimensional laser scanning technology, and the model accuracy satisfies the defect position labeling error ≤0.5m; the association is to convert the GPS spatial position data of the defect into coordinates in the model coordinate system through the reserved data interface, import the digital twin model, and visualize and label different types of defects in the model with different color labels.

[0015] By utilizing digital twin model generation methods and defect annotation logic, defect visualization and precise location are achieved. Models generated using oblique photography or 3D laser scanning technology have an accuracy of ≤0.5m, ensuring minimal error between the labeled defect location and the actual location. GPS coordinate transformation and model adaptation convert abstract data into intuitive annotations in 3D space, allowing maintenance personnel to quickly locate defects and plan work routes. Different color labels distinguish defect types, making the overall defect distribution clear at a glance and facilitating rapid assessment of the extent of steel structure damage. This design solves the problems of abstract and difficult-to-understand traditional drawing annotations, improves cross-departmental collaboration efficiency, and provides a visual medium for comparing defect changes.

[0016] In one optional implementation, the inspection report is automatically generated according to a preset industry standard template, and the report content integrates at least one of the following: defect image, type, location, size information, inspection time, and inspection personnel information.

[0017] By automatically generating reports based on industry-standard templates, the system avoids the formatting issues and omissions (such as missing defect dimensions and locations) often found in manual report writing, ensuring reports meet industry requirements. It integrates multi-dimensional information such as images, time, and personnel data, making the report complete and directly usable as a basis for operation and maintenance and acceptance. The elimination of manual data processing significantly improves work efficiency. Standardized reports also facilitate comparative analysis of different inspection batches, providing consistent documentation support for the long-term operation and maintenance of steel structures.

[0018] Secondly, embodiments of the present invention provide an intelligent inspection system for the exterior of steel structure buildings, comprising a drone inspection unit, a cloud processing unit, a digital management unit, and a digital twin connection unit; wherein: The UAV detection unit includes the UAV body, image acquisition module, positioning module and data transmission module, which are used to acquire images of the steel structure appearance and related parameters according to the planned path and upload them to the cloud processing unit. The cloud processing unit includes a cloud server and an AI recognition model, used to receive images, preprocess images, call the AI ​​recognition model to complete defect identification, and output defect data; The digital management unit includes a digital management platform and a database. The digital management platform is used to receive defect data, automatically generate inspection reports, and provide data retrieval and access management functions. The database is used to archive images, defect identification results, and inspection reports and establish correlation relationships. The digital twin connection unit includes a modeling module and a data import module. The modeling module generates a digital twin model through oblique photography or 3D laser scanning technology, and the data import module imports defect data into the model through a reserved interface to realize defect annotation and change comparison.

[0019] By constructing a collaborative system integrating "drones, cloud, management, and digital twins," the entire inspection process is integrated and intelligent. The drone inspection unit ensures comprehensive and real-time data collection; the cloud processing unit undertakes core computing tasks, enabling rapid analysis of large-scale images; the digital management unit handles data archiving, report generation, and access control, ensuring data security and traceability; and the digital twin integration unit combines inspection data with 3D models for visualized management. Each unit performs its specific function while collaborating seamlessly, completely replacing the traditional "manual collection - paper record" model. This standardizes and automates the inspection process, making it applicable to steel structure building inspections of different scales and types, and significantly improving the industry's inspection technology level.

[0020] In one alternative implementation, the AI ​​recognition model is deployed on a cloud server and incorporates a trained and optimized YOLO series model, supporting batch processing of valid images and outputting defect data.

[0021] By deploying the model in the cloud, the powerful computing capabilities of the server can be utilized to support batch processing (≥3000 images per hour), meeting the massive data processing needs of large-scale inspection projects. The optimized YOLO series models balance detection speed and accuracy, maintaining a recognition accuracy of ≥95% while processing rapidly. Standardized defect data is automatically output, allowing direct integration with digital management platforms without format conversion, reducing errors in data transfer. This design makes the cloud processing unit the core computing power support of the inspection system, ensuring efficiency and accuracy throughout the entire inspection process.

[0022] In one optional implementation, the digital twin connection unit also has a defect change analysis function, which automatically calculates the area change rate and position offset of the same defect in different inspection batches, generates a defect change trend chart, and intuitively displays the defect development.

[0023] By endowing digital twin interconnection units with defect change analysis capabilities, the scientific nature of operation and maintenance decisions is enhanced. Automatic calculation of area change rate and location offset transforms qualitative descriptions into quantitative data, avoiding subjective judgment errors. Trend charts intuitively display defect development trends, enabling early risk prediction and providing a basis for selecting maintenance timing. Dynamic analysis functions make the deterioration patterns of steel structures traceable. Through long-term data accumulation, the impact of the environment on defect development can be summarized, maintenance strategies optimized, the service life of steel structures extended, and the overall operation and maintenance costs reduced. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the intelligent inspection method for the appearance of steel structure buildings according to an embodiment of the present invention; Figure 2 This is a structural block diagram of an intelligent inspection system for the appearance of steel structure buildings according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0027] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0028] According to an embodiment of the present invention, an embodiment of an intelligent inspection method for the appearance of steel structure buildings is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here. Figure 1 This is a flowchart of an intelligent inspection method for the appearance of steel structure buildings according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S1: Collect images of various appearance defects in existing steel structures under different scenarios, label the defect information, divide the dataset, train an AI recognition model based on deep learning algorithm, optimize it to meet the preset detection accuracy requirements, and then save it to the cloud server.

[0029] Specifically, images of existing steel structures (such as steel columns in factories and steel components in bridges) with slight corrosion, moderate corrosion, large-area corrosion, coating peeling, cracks, deformation, and missing bolts were collected under different lighting conditions, such as sunny noon, cloudy days, and backlighting, as well as in environments with high humidity and after rain. A total of 10,000 images were collected. The images were then preprocessed, such as standardizing image size (to avoid machine confusion due to different sizes), adjusting brightness and contrast (to make corrosion features more obvious), removing blurry or invalid images, and expanding the data through methods such as "flipping, cropping, and adding noise". Further, annotation tools (such as LabelImg) were used to annotate the bounding box (location), corrosion level (type), and actual corrosion area (size) of the corrosion area in each image. The annotated data was then divided into a preset ratio (e.g., 7:2:1) into a training set of 7,000 images, a validation set of 2,000 images, and a test set of 1,000 images.

[0030] The deep learning algorithm in this embodiment of the invention employs the YOLO series of convolutional neural network algorithms (e.g., YOLO V12). The AI ​​recognition model is trained using a training set, and model parameters (such as learning rate and number of convolutional layers) are adjusted using a validation set. The preset detection accuracy requirement for model training is a recognition accuracy ≥ 95% and a false negative rate ≤ 5%. Performance is validated using a test set during model training. If the accuracy is insufficient, defect samples corresponding to the missing scenes are added; if the localization accuracy is inadequate, the model training parameters are readjusted. For example, if the test set shows a recognition accuracy of only 85% for "slight rust under backlight" (lower than the preset 95%), 2000 images of this scene are added for retraining, ultimately achieving a recognition accuracy of 96% and a false negative rate of 4% for various types of rust. After meeting the accuracy requirements, the model is saved to the cloud server.

[0031] By training with defect samples covering multiple scenarios, the AI ​​recognition model can adapt to complex environments (such as different lighting and weather conditions), reducing missed and false detections caused by scene differences. This solves the problems of traditional manual inspection being greatly affected by the environment and having inconsistent judgment standards. Diverse datasets and iterative optimization processes ensure that the model has stable recognition capabilities for different types and degrees of defects (such as rust and cracks), and can be directly applied to the inspection of steel structure buildings in different regions and with different service lives, without the need for repeated training. The high-precision AI model stored in the cloud can be accessed by the inspection system in real time, providing core technical support for the subsequent automatic analysis of drone images, significantly improving inspection efficiency and reducing reliance on human experience.

[0032] Step S2: Set inspection parameters based on the characteristics of the existing steel structure building, and automatically generate a drone flight planning path adapted to its appearance inspection.

[0033] Specifically, the inspection parameters include flight altitude, shooting interval, and obstacle avoidance distance. The flight altitude is set to 3-5m above the steel structure surface, the shooting interval is set to one image every 0.5m, and the obstacle avoidance distance is set to ≥1m from obstacles. The generated flight path is a combination of "circular + longitudinal" paths. The 3-5m flight altitude balances the shooting detail and coverage, avoiding both too close a narrow shooting angle requiring frequent adjustments and too far a distance causing blurred defects. The 0.5m shooting interval ensures no blind spots on the steel structure surface through image overlap, and the obstacle avoidance distance of more than 1m effectively avoids common protruding components and ancillary facilities (such as ladders and lights) in steel structures, reducing the risk of drone collisions. This is especially suitable for inspection scenarios of old steel structures (where there may be component deformation or protrusions). The "circular + longitudinal" path combines the advantages of the circular path in covering the circumference of the three-dimensional structure with the advantages of the longitudinal path in covering the linear components. Compared with single-direction flight, it reduces the number of repeated paths by 30%, shortens the detection time, and ensures that complex parts such as facades, top surfaces, and corners are completely captured.

[0034] Step S3: The image acquisition device equipped with automatic image focusing function of the UAV acquires images of the structural appearance according to the flight plan path, and simultaneously records the shooting time and GPS location parameters of the images, and uploads them to the cloud server in real time.

[0035] In one example, the drone is equipped with a 20-megapixel high-definition camera (with autofocus) and flies along a preset "circular + longitudinal" path. When the drone approaches the bridge's steel truss, the camera automatically identifies the truss surface and quickly focuses, clearly capturing details such as bolt connection nodes and member coatings. During flight, for each image captured, the system simultaneously records the image capture time (e.g., "2024-11-05 10:30:22") and GPS location parameters (e.g., "30°XX′XX″N, 120°XX′XX″E"), and uploads the images and associated parameters to the cloud server in real time via a 5G network, ensuring immediate data storage and subsequent retrieval.

[0036] The drone, equipped with a high-definition camera, features autofocus that dynamically adjusts the focus based on the real-time distance between the drone and the steel structure. This eliminates image blurring caused by distance variations during manual operation, particularly effective for capturing small defects at long distances (such as minute cracks or small areas of rust), providing a high-quality data foundation for subsequent AI recognition. Simultaneously recorded GPS location parameters bind the image to the actual spatial location of the steel structure. Combined with the shooting time, the detection trajectory can be traced. After AI identifies a defect, it can directly pinpoint its exact location within the building using GPS data (e.g., "the upper chord of the third span on the south side of the bridge"), avoiding errors associated with traditional manual location recording. Real-time upload functionality allows the cloud server to receive data instantly, eliminating the need to wait for the drone to return and export the data. This enables inspectors to simultaneously review image quality on the ground. If missed or blurry images are found, the flight plan can be adjusted immediately for reshoots, reducing rework costs. Simultaneously, it saves time for the AI ​​model to quickly access data and generate preliminary recognition results.

[0037] Step S4: After preprocessing the received image, the cloud server calls the trained AI recognition model to identify and analyze defects in the received image data, outputting defect data including defect type, location, and bounding box coordinates, and synchronizing the defect data to the digital management platform. Specifically: 1. Image preprocessing: The uploaded images are deblurred, denoised, and resized. Invalid images that are occluded or blurry are removed. A unique identification number is associated with each valid image, which is bound to the image's capture time and GPS location parameters. 2. Defect localization: The AI ​​recognition model is used to extract features from the valid image, locate the defect area and mark the bounding box, with the bounding box coordinates accurate to the pixel level; 3. Defect Classification: Based on the texture and shape features of the defects, the located defect areas are classified, and their defect types are identified and output. The defect types include at least one of rust, coating peeling, cracks, deformation, and missing bolts. This invention utilizes an AI recognition model to quickly locate minute defects in images (such as 0.5mm wide cracks or missing bolts), and quantifies the defect range using bounding box coordinates. This solves the problems of low efficiency and high false negative rate (especially for high-altitude and hidden areas) in manual visual identification, achieving an accuracy rate of over 95%. Combined with GPS parameters attached to the image, defect data can be directly mapped to the actual location of the building, avoiding the pain point of traditional inspections where "only the defect image is known, but the specific location is difficult to determine," providing precise guidance for subsequent maintenance. The cloud automatically processes and synchronizes data to the management platform, eliminating manual sorting and entry, reducing the data processing time for a single batch of inspections from 24 hours to 2 hours, while ensuring a consistent data format for easy subsequent statistical analysis and historical traceability.

[0038] In step S5, the digital management platform automatically integrates images and defect data to generate an inspection report based on a preset industry standard inspection report template. At the same time, it archives the images, defect identification results, and inspection reports into the database, establishing a "building-inspection batch-defect" relationship for historical data traceability.

[0039] In one example, the digital management platform has a built-in report template that conforms to the "Standard for Acceptance of Construction Quality of Steel Structure Engineering" (including modules such as project overview, defect statistics, and handling suggestions). After the AI ​​recognition model outputs defect data, the platform automatically extracts 300 defect images (such as rusted areas and missing bolt nodes), the corresponding defect types (25 rusted areas and 8 missing bolt nodes), location information (No. 5 steel column on the east facade, connection node of the west grandstand, etc.), and dimensional data, and integrates them according to the template format to generate the "Appearance Inspection Report of Steel Structure of XX Gymnasium". Each defect in the report is accompanied by a related image and location coordinates, as well as the inspection time and inspection personnel information.

[0040] Simultaneously, the platform stores all original images, AI-identified defect lists, and generated inspection reports for this batch in its database and establishes a correlation: using "XX Gymnasium" as the main building, it binds it to the "November 2025 Routine Inspection" batch. This batch is associated with 25 rust defects and 8 missing bolt defects, and each defect can directly jump to the corresponding original image and report page number. Subsequently, staff can retrieve all data for this batch by searching "XX Gymnasium + November 2024"; clicking on a specific rust defect allows viewing whether similar problems exist at that location in historical inspections.

[0041] This invention automatically generates reports based on preset industry templates, avoiding formatting issues and omissions (such as missing defect dimensions or vague location descriptions) that occur with manual reporting. It also reduces report generation time from the traditional 3 days to 2 hours, significantly improving work efficiency. By establishing a "building-inspection batch-defect" relationship, each defect can be traced back to the original image, inspection time, and similar historical issues, solving the problems of "difficult search and weak correlation" in traditional paper archives. This facilitates the analysis of defect development trends (such as whether the rust area of ​​a steel column is increasing year by year). The historical data archived in the database can provide a basis for the maintenance and reinforcement of steel structure buildings (such as optimizing maintenance plans based on the frequency of bolt loss over the years), while also providing a benchmark for new inspection batches, achieving full-process data support from initial inspection to subsequent operation and maintenance.

[0042] Step S6: Establish a digital twin model of the existing steel structure building, import the defect data into the model for accurate annotation, and support comparison of defect changes in different inspection batches.

[0043] Specifically, the digital twin model of the existing steel structure building is generated by oblique photogrammetry or 3D laser scanning technology, and the model accuracy meets the defect location marking error of ≤0.5m; the association of "building-inspection batch-defect" is achieved by converting the GPS spatial location data of the defect into coordinates in the model coordinate system through a reserved data interface, importing it into the digital twin model, and visually marking different types of defects in the model with different color labels.

[0044] In this embodiment of the invention, taking the appearance inspection of a steel structure chimney (80m high) as an example, the above process includes: 1. Digital Twin Model Generation: A drone equipped with an oblique photography camera is used to take pictures of the chimney from all angles. A three-dimensional digital twin model is generated through oblique photography technology. The model accuracy reaches 0.3m (meeting the error requirement of ≤0.5m) and can clearly present the structural details of the chimney such as steel cylinder wall, ladder, and platform.

[0045] 2. Defect Data Import and Labeling: During the inspection, two areas of rust were identified (located at 30m and 50m above the ground, respectively) and one missing bolt (located at a platform connection node 40m high). The digital management platform, through a reserved interface, converted the GPS location data of the defects (e.g., "30°20′N, 121°15′E, altitude 30m") into coordinates in the digital twin model coordinate system (e.g., X=125.6m, Y=89.2m, Z=30.0m) and imported it into the model. In the model, rust defects are labeled in red, and missing bolt defects are labeled in yellow. Clicking on the labels allows viewing detailed information such as defect images and dimensions.

[0046] 3. Relationship Reflection: Through the model's timeline function, different inspection batches (such as June 2024 and November 2025) can be switched to compare the changes in defects at the same location. For example, the slight corrosion (small red label) at 30m in June 2024 has expanded into large-area corrosion (large red label) in November 2025, intuitively showing the development trend of defects.

[0047] By combining digital twin models with coordinate transformation technology, abstract GPS data is converted into specific locations in a 3D model with a labeling error of ≤0.5m. This solves the problems of "abstract and difficult to understand, and vague location" in traditional drawings. Maintenance personnel can use the model to plan climbing paths in advance, improving work efficiency. Different color labels distinguish defect types (such as corrosion, missing bolts), enabling inspection personnel and maintenance parties to quickly grasp the overall defect distribution of the steel structure (such as "concentrated corrosion in the upper part of the chimney"), providing an intuitive basis for determining maintenance priorities. By combining the "building-inspection batch-defect" relationship with the model's timeline function, the positional offset and area changes of defects in multiple batches (such as corrosion expansion, new deformations) can be dynamically compared, helping to assess the deterioration rate of the steel structure, develop reinforcement plans in advance, and extend the building's service life.

[0048] Meanwhile, the digital twin model can serve as a shared platform, allowing designers, construction teams, and maintenance personnel to view defect information synchronously through the model, avoiding work errors caused by inconsistent drawing versions or communication mistakes. It is especially suitable for the full life cycle management of large and complex steel structure buildings.

[0049] This invention also provides an intelligent inspection system for the appearance of steel structure buildings, such as... Figure 2 As shown, it includes a drone detection unit, a cloud processing unit, a digital management unit, and a digital twin connection unit; wherein: The UAV detection unit includes the UAV itself, an image acquisition module, a positioning module, and a data transmission module. It is used to acquire images of the steel structure's exterior and related parameters according to the planned path and upload them to the cloud processing unit. The cloud processing unit includes a cloud server and an AI recognition model, which is used to receive images, preprocess images, call the AI ​​recognition model to complete defect identification, and output defect data. The digital management unit includes a digital management platform and a database. The digital management platform is used to receive defect data, automatically generate inspection reports, and provide data retrieval and access management functions. The database is used to archive images, defect identification results, inspection reports, and establish correlation relationships. The digital twin connection unit includes a modeling module and a data import module. The modeling module generates a digital twin model through oblique photography or 3D laser scanning technology, while the data import module imports defect data into the model through a reserved interface, enabling defect annotation and change comparison.

[0050] The system provided in this invention integrates drone data collection, AI recognition, data management, and digital twin visualization, replacing the traditional "manual climbing inspection + paper record" model. This reduces the inspection cycle from 7 days to 1.5 days, improving inspection efficiency by over 80%. The drone's automatic focusing and GPS positioning ensure clear images and traceable locations. The AI ​​model achieves over 95% accuracy in identifying minute defects (such as 0.5mm cracks), and the database association mechanism prevents data omissions or confusion, solving the problems of "high missed detection rate and fragmented data" inherent in traditional manual inspection.

[0051] Digital twin models intuitively present the distribution and trends of defects. Combined with historical data, they can analyze the deterioration patterns of steel structures (such as the average annual rate of corrosion expansion in a certain area), providing data support for prioritizing maintenance and formulating reinforcement plans, extending the service life of buildings, and reducing operation and maintenance costs. The system is adaptable to different types of steel structure buildings such as factories, chimneys, and bridges. The digital management platform supports multi-role (inspector, engineer, maintenance personnel) access control and data sharing, avoiding collaboration errors caused by information asymmetry, and is especially suitable for the long-term management of large and complex steel structures.

[0052] Furthermore, the AI ​​recognition model is deployed on a cloud server, with built-in trained and optimized YOLO series models, supporting batch processing of valid images and outputting defect data. For example, the YOLO V12 model deployed on the cloud server receives two batches of 2000 images of steel components in the factory uploaded by a drone, and outputs defect data after batch processing: In October 2024, it identified one rust spot (0.2㎡) on a steel column in area A and one missing bolt on a steel beam in area B; in October 2025, it identified that the rust area of ​​the same steel column had expanded to 0.5㎡, and one new crack had been added, while the location of the missing bolt remained unchanged.

[0053] The YOLO series models support batch processing (up to 3,000 images per hour), which is more than 50 times more efficient than manual single-image analysis, while maintaining a recognition accuracy of over 95%, meeting the large-scale inspection needs of large steel structure buildings (such as factories and bridges).

[0054] The digital twin interconnection unit also features defect change analysis capabilities, automatically calculating the area change rate and position offset of the same defect across different inspection batches, generating defect change trend charts to visually display the defect development. For example, the digital twin model automatically calculates the following by linking cross-batch data for "factory building - steel column in area A - corrosion": Area change rate: (0.5㎡ - 0.2㎡) / 0.2㎡ × 100% = 150%; Position offset: The center coordinates of the corroded area are offset by ≤0.1m in the model (negligible); The final result is a trend chart (such as a line graph showing the rust area expanding year by year), which is then dynamically color-coded in the digital twin model (light red for 2024 and dark red for 2025) to visually represent the speed of defect development.

[0055] By automatically calculating the area change rate and location offset, qualitative descriptions ("increasing corrosion") are transformed into quantitative data ("150% growth rate"). Combined with trend charts, the development trend of defects can be predicted in advance (e.g., at this rate, the corroded area will reach 1.2㎡ in one year), providing a scientific basis for selecting maintenance timing and preventing small defects from evolving into structural safety hazards. Furthermore, by tracking the changes in the same defect over a long period, the deterioration patterns of steel structures can be summarized (e.g., the correlation between corrosion in a certain area and environmental humidity), optimizing maintenance strategies (e.g., specifically increasing the thickness of the anti-corrosion coating), extending the service life of the building, and reducing the overall life cycle maintenance costs.

[0056] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for intelligent inspection of the appearance of steel structure buildings, characterized in that, include: Collect images of various appearance defects in existing steel structures under different scenarios, label the defect information and divide the dataset, train an AI recognition model based on deep learning algorithm, optimize it to meet the preset detection accuracy requirements and then save it to the cloud server; Based on the characteristics of existing steel structure buildings, inspection parameters are set, and a drone flight planning path adapted to its appearance inspection is automatically generated. The drone is equipped with an image acquisition device with automatic image focusing function to acquire images of the structure's appearance according to the flight plan path, and simultaneously record the corresponding shooting time and GPS location parameters of the images, and upload them to the cloud server in real time; After the cloud server preprocesses the received images, it calls the trained AI recognition model to identify and analyze the defects in the received image data, outputs defect data containing defect type, location, and bounding box coordinates, and synchronizes the defect data to the digital management platform. The digital management platform automatically integrates images and defect data to generate inspection reports based on preset industry standard inspection report templates. At the same time, it archives images, defect identification results, and inspection reports into the database, establishing a "building-inspection batch-defect" relationship for historical data traceability. Establish a digital twin model of the existing steel structure building, import defect data into the model for precise annotation, and support comparison of defect changes in different inspection batches.

2. The method according to claim 1, characterized in that, The various appearance defects include rust, coating peeling, cracks, deformation, and missing bolts. The defect data includes defect type, location, actual size, and bounding box coordinates. The dataset is divided into training set, validation set, and test set according to a preset ratio. The training set is used for model training, the validation set is used to adjust model parameters during training, and the test set is used to verify the final detection accuracy of the model.

3. The method according to claim 1, characterized in that, The deep learning algorithm adopts the YOLO series model in the convolutional neural network algorithm. The preset detection accuracy requirements for model training are recognition accuracy ≥95% and false negative rate ≤5%. The performance is verified through the test set during model training. If the accuracy does not meet the standard, defect samples corresponding to the missing scene are added. If the positioning accuracy is insufficient, the model training parameters are readjusted.

4. The method according to claim 1, characterized in that, The inspection parameters include flight altitude, shooting interval, and obstacle avoidance distance. The flight altitude is set to 3-5m above the surface of the steel structure building, the shooting interval is set to take one image every 0.5m, the obstacle avoidance distance is set to ≥1m from the obstacle, and the generated flight path is a combination of "circular + longitudinal" path.

5. The method according to claim 2, characterized in that, After preprocessing the received images, the cloud server calls a trained AI recognition model to perform defect identification and analysis on the received image data, including: Image preprocessing: The uploaded images are deblurred, denoised, and resized. Invalid images that are occluded or blurry are removed. A unique identification number is associated with each valid image, which is bound to the image's capture time and GPS location parameters. Defect localization: The AI ​​recognition model is used to extract features from the valid image, locate the defect area and mark the bounding box, with the bounding box coordinates accurate to the pixel level; Defect classification: Based on the texture and shape features of the defects, the located defect areas are classified, and their defect types are identified and output. The defect types include at least one of rust, coating peeling, cracks, deformation, and missing bolts.

6. The method according to claim 1, characterized in that, The digital twin model is generated using oblique photography or 3D laser scanning technology, and the model accuracy meets the requirement that the defect location labeling error is ≤0.5m. The association is achieved by converting the GPS spatial location data of the defect into coordinates in the model coordinate system through a reserved data interface, importing it into the digital twin model, and visually labeling different types of defects in the model with different color labels.

7. The method according to claim 1, characterized in that, The inspection report is automatically generated based on a preset industry standard template. The report content integrates at least one of the following: defect image, type, location, size information, inspection time, and inspection personnel information.

8. An intelligent inspection system for the exterior appearance of steel structure buildings, characterized in that, It includes a drone detection unit, a cloud processing unit, a digital management unit, and a digital twin connection unit; among which: The UAV detection unit includes the UAV body, image acquisition module, positioning module and data transmission module, which are used to acquire images of the steel structure appearance and related parameters according to the planned path and upload them to the cloud processing unit. The cloud processing unit includes a cloud server and an AI recognition model, used to receive images, preprocess images, call the AI ​​recognition model to complete defect identification, and output defect data; The digital management unit includes a digital management platform and a database. The digital management platform is used to receive defect data, automatically generate inspection reports, and provide data retrieval and access management functions. The database is used to archive images, defect identification results, and inspection reports and establish correlation relationships. The digital twin connection unit includes a modeling module and a data import module. The modeling module generates a digital twin model through oblique photography or 3D laser scanning technology, and the data import module imports defect data into the model through a reserved interface to realize defect annotation and change comparison.

9. The system according to claim 8, characterized in that, The AI ​​recognition model is deployed on a cloud server and has a built-in trained and optimized YOLO series model, which supports batch processing of valid images and output of defect data.

10. The system according to claim 8, characterized in that, The digital twin connection unit also has a defect change analysis function, which automatically calculates the area change rate and position offset of the same defect in different inspection batches, generates defect change trend charts, and intuitively displays the development of defects.

Images