An automatic detection method and system for building curtain walls

Abstract

This invention discloses an automatic inspection method and system for building curtain walls, relating to the field of visual inspection technology for building curtain walls. It acquires high-resolution images covering the visible to near-infrared bands, combines them with a support vector machine algorithm to classify chemical components, and generates a chemical composition classification map containing the distribution areas of corrosion products and dirt types. Further, it acquires three-dimensional point cloud data using a laser rangefinder, fuses it with multispectral images, inputs it into a convolutional neural network to predict the thickness of the deposits, and generates a thickness distribution map. A graded classification model generates a graded distribution map of aging levels, and combined with aging correlation analysis, determines priority maintenance areas and treatment methods. This invention accurately identifies the types of deposits on the curtain wall surface and accurately measures their thickness and distribution range through automated technical means, achieving precise assessment of the aging state of the curtain wall surface and intelligent maintenance decision-making, significantly improving maintenance efficiency and curtain wall performance.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology for building curtain walls, and particularly to an automatic inspection method and system for building curtain walls. Background Technology

[0002] As a crucial component of modern building facades, building curtain walls not only affect aesthetics but also directly impact safety, energy efficiency, and lifespan. With the rapid growth of high-rise buildings in cities, the need for curtain wall maintenance is increasingly urgent, making automated inspection technology a key means to ensure curtain wall performance and safety. Traditional curtain wall inspection methods primarily rely on manual inspections or simple instrument measurements, resulting in low efficiency and limited coverage. A deeper problem lies in the difficulty of accurately distinguishing different types of deposits on the curtain wall surface, such as dirt, oxide layers, or corrosion products, and the inability to accurately quantify their thickness and distribution. This limitation leads to a lack of scientific basis for maintenance decisions, increasing safety hazards and maintenance costs.

[0003] During the long-term use of curtain walls, the accumulation caused by material aging becomes a core challenge in inspection. Environmental factors gradually form various deposits on the curtain wall surface. These deposits have complex chemical compositions and uneven distribution, making it difficult for traditional single-method testing to accurately identify their type. This directly affects the judgment of the deposit's nature; for example, it cannot distinguish between corrosion products and ordinary dirt, leading to maintenance strategies that may be overly aggressive or insufficient. Furthermore, because the thickness and distribution range of the deposits are difficult to measure precisely, maintenance personnel cannot accurately assess the degree of aging. For example, on the glass curtain wall of a high-rise building, local areas may have reduced light transmittance due to oxide layer accumulation, but current technology cannot quickly locate these areas and quantify their severity, thus delaying targeted maintenance measures.

[0004] Therefore, how to accurately identify the type of deposits on the surface of curtain walls and accurately measure their thickness and distribution range through automated technical means has become a key issue in the field of automatic inspection of building curtain walls. Summary of the Invention

[0005] This invention provides an automatic detection method and system for building curtain walls, which uses automated technology to accurately identify the type of deposits on the curtain wall surface and accurately measure their thickness and distribution range, thereby achieving accurate assessment of the aging state of the curtain wall surface and intelligent maintenance decision-making, significantly improving maintenance efficiency and curtain wall performance.

[0006] The automatic detection method for building curtain walls provided by this invention is executed by a computer and includes:

[0007] Acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data;

[0008] Based on the first multispectral image, a support vector machine algorithm is used to classify the spectral features of the first multispectral image to generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution area of ​​corrosion products and dirt types on the curtain wall surface.

[0009] Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits;

[0010] Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface to obtain the first point cloud data.

[0011] Based on the first point cloud data and the segmented image, a feature fusion mechanism is used to combine the spatial information of the first point cloud data and the segmented image to determine the thickness distribution map corresponding to the thickness of the accumulation.

[0012] Based on the thickness distribution map, spatial pattern features are extracted using a convolutional neural network to generate an aging degree grade distribution map;

[0013] The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

[0014] According to the automatic detection method for building curtain walls provided by the present invention, the step of classifying the spectral features of the first multispectral image based on the first multispectral image using a support vector machine algorithm to generate a chemical composition classification map includes:

[0015] Based on the first multispectral image, preprocessing techniques are used to remove noise and background interference to generate a second multispectral image;

[0016] Based on the spectral features in the second multispectral image, the spectral reflectance of each pixel is calculated using a spectral analysis algorithm to obtain a spectral feature set;

[0017] Based on the spectral feature set, the support vector machine algorithm is used to classify the spectral feature set to obtain a preliminary classification result, wherein the preliminary classification result includes the corrosion product region corresponding to the corrosion product feature and the dirt type region corresponding to the dirt type feature.

[0018] Based on the preliminary classification results, and in conjunction with the chemical composition database, the chemical composition of corrosion product characteristics and fouling type characteristics is matched to determine the chemical composition distribution of each corrosion product region or fouling type region.

[0019] By using image processing technology, the distribution of chemical components is mapped onto the second multispectral image to generate a chemical component classification map.

[0020] According to the automatic detection method for building curtain walls provided by the present invention, the step of generating a segmented image containing the spatial distribution range of the deposits based on the chemical composition classification map and using image segmentation technology includes:

[0021] Based on the chemical composition classification map, noise reduction and enhancement are performed using preprocessing techniques to obtain the first processed image;

[0022] Based on the first processed image, a semantic segmentation algorithm is used to generate the regional boundaries of corrosion products and dirt to obtain a second processed image.

[0023] If the boundary clarity of the second processed image is lower than a preset threshold, the boundary of the second processed image is optimized by an edge detection algorithm to obtain a third processed image.

[0024] Based on the third processed image, the spatial distribution characteristics of corrosion products and dirt types are calculated to determine the range of the deposits and obtain a distribution feature map.

[0025] Based on the area ratio of corrosion products and dirt in the distribution feature map, a clustering algorithm is used to divide the accumulation area and obtain the area classification result.

[0026] Based on the region classification results, a segmented image containing the spatial distribution range of the accumulated objects is generated.

[0027] According to the automatic detection method for building curtain walls provided by the present invention, the step of acquiring three-dimensional point cloud data of the deposits on the curtain wall surface using a laser rangefinder based on the segmented image to obtain first point cloud data includes:

[0028] Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface, and a second point cloud data is generated.

[0029] Based on the second point cloud data, noise reduction is performed using a point cloud filtering algorithm to obtain the third point cloud data;

[0030] If the point density of the third point cloud data is lower than the preset density threshold, then based on the third point cloud data, an interpolation algorithm is used to supplement the missing points and generate the fourth point cloud data.

[0031] Based on the fourth point cloud data, the iterative nearest point algorithm is applied for registration processing to obtain the fifth point cloud data;

[0032] Based on the fifth point cloud data, the geometric features of the curtain wall surface are extracted to generate surface feature data;

[0033] Based on the surface feature data, the distribution area containing the deposits is detected, and the deposit distribution data is generated.

[0034] Based on the distribution data of the deposits, a clustering algorithm is used for classification to determine the first point cloud data, which includes the deposit type and deposit thickness.

[0035] According to the automatic detection method for building curtain walls provided by the present invention, the step of combining the spatial information of the first point cloud data and the segmented image using a feature fusion mechanism to determine the thickness distribution map corresponding to the thickness of the deposit includes:

[0036] Based on the first point cloud data and the segmented image, a point cloud registration algorithm is used to align the first point cloud data and the segmented image to obtain a registered point cloud image dataset.

[0037] Based on the registered point cloud image dataset, the spatial information is fused through a feature fusion mechanism to generate a fused feature set;

[0038] Based on the fused feature set, a convolutional neural network is used to process the fused feature set, extract high-dimensional features, and obtain a feature vector set;

[0039] The feature vector set is input into the thickness regression model to obtain the predicted thickness value corresponding to the thickness of the deposit.

[0040] Based on the predicted thickness value, a three-dimensional interpolation algorithm is used to generate thickness distribution data and a thickness distribution map.

[0041] According to the automatic inspection method for building curtain walls provided by the present invention, the step of extracting spatial pattern features based on the thickness distribution map using a convolutional neural network to generate an aging degree classification distribution map includes:

[0042] Based on the thickness distribution data of the thickness distribution map, a data augmentation method is used to perform rotation and flipping operations to obtain an augmented thickness distribution dataset;

[0043] Based on the enhanced thickness distribution dataset, spatial pattern features are extracted using a convolutional neural network to obtain a feature vector set;

[0044] If the feature vector set meets the preset feature integrity threshold, then the feature vector set is input into the hierarchical classification model to determine the aging degree classification result;

[0045] Based on the aging degree classification results, a visualization algorithm is used to generate graded distribution data;

[0046] The hierarchical distribution data is spatially mapped to generate high-resolution aging degree distribution data, resulting in an optimized distribution map.

[0047] If the optimized distribution map meets the preset clarity threshold, then image smoothing technology is used for processing to obtain the aging degree classification distribution map.

[0048] According to the automatic detection method for building curtain walls provided by the present invention, determining the priority maintenance area and the maintenance treatment method for the priority maintenance area in the building curtain wall includes the following steps:

[0049] Based on the areas of damaged light transmittance in the aging grade distribution map of the curtain wall, a light transmission model is used, combined with the spectral characteristics of the areas of damaged light transmittance, to determine the degree of light transmittance reduction.

[0050] Based on the degree of light transmittance reduction, a convolutional neural network is used to perform spatial pattern analysis and generate a light transmittance distribution map.

[0051] Based on the light transmittance distribution map and the aging degree classification distribution map, a decision tree algorithm is used to determine the priority maintenance areas and maintenance methods.

[0052] The present invention also provides an automatic detection system for building curtain walls, comprising:

[0053] The acquisition module is used to acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data.

[0054] The classification map generation module is used to classify the spectral features of the first multispectral image based on the first multispectral image using a support vector machine algorithm, and generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution area of ​​corrosion products and dirt types on the curtain wall surface.

[0055] Image segmentation and generation: Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits.

[0056] The point cloud data generation module is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface based on the segmented image using a laser rangefinder, and obtain the first point cloud data.

[0057] The distribution map generation module is used to combine the spatial information of the first point cloud data and the segmented image based on the first point cloud data and the segmented image using a feature fusion mechanism to determine the thickness distribution map corresponding to the thickness of the deposit.

[0058] The graded distribution map generation module is used to extract spatial pattern features from the thickness distribution map using a convolutional neural network to generate a graded distribution map of aging degree.

[0059] The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

[0060] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the automatic detection method for building curtain walls as described in any of the preceding claims.

[0061] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the automatic detection method for building curtain walls as described in any of the preceding claims.

[0062] This invention provides an automatic inspection method and system for building curtain walls, addressing the comprehensive detection and maintenance issues of corrosion, dirt distribution, and decreased light transmittance on curtain wall surfaces. It acquires high-resolution images covering the visible to near-infrared bands, and uses a support vector machine algorithm to classify chemical components, generating a chemical composition classification map that includes the distribution areas of corrosion products and dirt types. Further, it acquires three-dimensional point cloud data using a laser rangefinder, denoises and registers it, and fuses it with multispectral images. This data is then input into a convolutional neural network to predict the thickness of the deposits, generating a thickness distribution map. A graded classification model generates a graded distribution map of aging levels, and combined with aging correlation analysis, it determines priority maintenance areas and treatment methods. This invention accurately identifies the types of deposits on curtain wall surfaces and precisely measures their thickness and distribution range through automated techniques, achieving accurate assessment of the aging state of curtain wall surfaces and intelligent maintenance decisions, significantly improving maintenance efficiency and curtain wall performance. Attached Figure Description

[0063] Figure 1 This is one of the flowcharts of the automatic detection method for building curtain walls provided in the embodiments of the present invention;

[0064] Figure 2 This is a second schematic flowchart of the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0065] Figure 3 This is the third flowchart of the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0066] Figure 4 This is the fourth flowchart of the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0067] Figure 5 This is the fifth flowchart illustrating the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0068] Figure 6This is the sixth flowchart illustrating the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0069] Figure 7 This is the seventh flowchart of the automatic detection method for building curtain walls provided in this embodiment of the invention;

[0070] Figure 8 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0071] 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.

[0072] Reference Figure 1 This invention provides an automatic detection method for building curtain walls, executed by a computer, comprising the following steps:

[0073] Step 100: Acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data;

[0074] The first step is to acquire a first multispectral image representing the surface condition of the curtain wall. This first multispectral image not only contains visible light information, but more importantly, it captures the reflection or radiation characteristics of the material in specific non-visible light bands, such as near-infrared and short-wave infrared. The spectral characteristics of these bands are closely related to the molecular structure and chemical composition of the curtain wall surface material. For example, different types of materials have their own unique spectral fingerprints or spectral characteristics. Corrosion products include rust and verdigris, while dirt includes oil, dust, and biofilm.

[0075] Multispectral imaging technology is used to acquire image data covering the visible to near-infrared bands from the curtain wall surface, generating an initial multispectral image. Principal component analysis (PCA) is used to perform dimensionality reduction on the initial multispectral image, extracting key spectral features to obtain a feature dataset. If the spectral features of the feature dataset match a preset threshold in a pre-defined spectral feature library for accumulated materials, the presence of accumulated materials is confirmed, and an accumulated material distribution map is generated. Based on the accumulated material distribution map, K-means clustering is used to segment the accumulated material regions, obtaining region segmentation images. The spectral reflectance of each segmented region is calculated using the region segmentation images, generating a spectral reflectance dataset. If the slope of the reflectance curve in the spectral reflectance dataset exceeds a preset threshold, an anomaly in the accumulated material thickness is identified, generating a thickness anomaly region map. Based on the thickness anomaly region map, image fusion technology is used to overlay the anomaly region with the initial multispectral image to generate the first multispectral image.

[0076] Step 200: Based on the first multispectral image, a support vector machine algorithm is used to classify the spectral features of the first multispectral image to generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution areas of corrosion products and dirt types on the curtain wall surface.

[0077] After obtaining the first multispectral image, the Support Vector Machine (SVM) algorithm is used to intelligently identify and classify the spectral features contained in the first multispectral image. The SVM algorithm is a mature supervised machine learning method, particularly adept at handling high-dimensional data and finding optimal classification boundaries. It automatically analyzes and determines the chemical category to which the spectral features represented by each pixel in the image belong by learning a classification model established from known samples (i.e., spectral data of labeled corrosion products and dirt). It can effectively distinguish subtle differences between spectral features, thereby accurately identifying different types of chemical substances.

[0078] Based on the point-by-point classification results of the spectral features of all pixels using the Support Vector Machine (SVM) algorithm, a chemical composition classification map is obtained. This map is a spatial distribution map containing the chemical composition attributes of different regions on the curtain wall surface. It depicts the specific location, morphology, and distribution range of various corrosion products and different types of dirt on the curtain wall surface. As a spatial distribution analysis map of the chemical substances on the curtain wall surface obtained through intelligent analysis, the chemical composition classification map contains key information on the chemical decay and pollution status of the curtain wall materials, thus providing crucial data support for subsequent curtain wall health assessment and maintenance decisions.

[0079] Support Vector Machine (SVM) is a highly efficient and accurate multispectral data classifier. It utilizes maximum margin classification and kernel function techniques to categorize complex high-dimensional spectral data into predefined chemical categories, resulting in a chemical composition classification map that includes the chemical state of the curtain wall surface. Specifically, the SVM algorithm classifies the spectral features of the input first multispectral image, determining the category to which each spectral vector in the image belongs. For example, it can determine whether the spectral features belong to rust, verdigris, oil stains, dust, or normal curtain wall materials. The SVM algorithm classifies the entire first multispectral image pixel-by-pixel, assigning a category label to each pixel. All these classification results are then integrated to generate a chemical composition classification map. Due to the maximum margin principle in the SVM algorithm, predictions for unknown data have high confidence and robustness, i.e., strong anti-interference ability. This means that even for ambiguous pixels whose spectral features fall between two categories, the SVM algorithm can provide a very firm and stable classification result. This ensures that the final chemical composition classification map has high accuracy and reliability, providing stable and reliable classification results.

[0080] Step 300: Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits;

[0081] The core task of this step is to identify and aggregate all pixels in the image that belong to the target deposit (such as a specific type of corrosion product or severe dirt requiring special attention) based on these pre-defined chemical category labels. Specifically, image segmentation technology is used to analyze image data, separating pixel regions with the same or similar attributes from the background or other regions with different attributes. The attributes used for segmentation are precisely the specific chemical category information corresponding to each pixel, provided by the chemical composition classification map. Based on the chemical composition classification map, the image segmentation algorithm automatically finds and labels all continuous pixel blocks classified as the target deposit, generating a segmented image containing the spatial distribution range of the deposit. The segmented image includes the actual location, specific shape outline, and continuous coverage area of ​​the target deposit on the two-dimensional surface of the curtain wall.

[0082] Image segmentation technology plays a crucial role in converting chemical composition identification into physical spatial localization. It serves as a spatial information extraction and region localization tool, analyzing image data from chemical composition classification maps. It separates pixel regions with similar or identical attributes from the background or other regions with different attributes. Through similarity-based region aggregation or edge detection, it outputs a segmented image representing the physical spatial extent. This acts as a core bridge between chemical composition identification and physical state assessment, transferring data from chemical attributes to spatial attributes, thus transforming abstract category information into concrete, quantifiable geometric regions. For example, image segmentation technology can aggregate all spatially continuous pixels belonging to the same target category in a chemical composition classification map, forming a complete, connected region object. This defines clear and precise physical boundaries for these chemical substances, enabling the identification of the extent, area, and shape of corrosion products or contaminants covering the curtain wall. Image segmentation technology can also distinguish between deposits of interest and normal curtain wall background material, not only separating deposits from the curtain wall substrate but also further separating different types of deposits (such as corrosion products and contaminants).

[0083] Step 400: Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface to obtain the first point cloud data;

[0084] Based on segmented images, a laser rangefinder is used to scan and measure the curtain wall surface. The laser rangefinder emits a laser beam towards the curtain wall surface and receives the reflected signal, accurately calculating the distance from the emission point to the surface reflection point. By systematically moving the rangefinder or utilizing its built-in scanning device, the spatial coordinates of a large number of discrete points on the curtain wall surface can be densely acquired. Through laser scanning, a set of points representing the three-dimensional topography of the curtain wall surface, i.e., the first point cloud data, is collected and generated. The first point cloud data contains the overall three-dimensional geometric information of the curtain wall surface, clearly distinguishing which points belong to the target accumulation area and which belong to the unaffected curtain wall base area. Therefore, the first point cloud data provides the distribution range of the target accumulation in three-dimensional space, including the planar coverage location of the target accumulation.

[0085] Step 500: Based on the first point cloud data and the segmented image, a feature fusion mechanism is used to combine the spatial information of the first point cloud data and the segmented image to determine the thickness distribution map corresponding to the thickness of the accumulation.

[0086] The feature fusion mechanism first relies on strict spatial registration to ensure that each 3D point in the first point cloud data can establish a precise one-to-one correspondence with the corresponding 2D pixel position in the segmented image. This registration is the foundation of all subsequent analyses, guaranteeing seamless integration of the 3D geometric information of the first point cloud data with the segmented image containing 2D chemical category and region information within the same spatial reference frame. Under the premise of precise registration, utilizing the high-precision 3D coordinate information contained in the first point cloud data, the corresponding 3D point cloud set is extracted for each region identified as an accumulation in the segmented image. Then, by calculating the vertical distance or height difference of each point in these point sets relative to the adjacent, confirmed uncorroded or uncontaminated original curtain wall substrate surface, the actual thickness value of the accumulation at that point location can be determined. Finally, the thickness of all relevant points within the target accumulation coverage area is calculated using the feature fusion mechanism to generate a thickness distribution map. A thickness distribution map is a two-dimensional spatial mapping map in which the value of each pixel (or location point) represents the specific thickness of the target deposit at that location. It shows the thickness difference, spatial gradient and overall thickness pattern of the deposit at different locations on the curtain wall surface, such as where it is thicker, where it is thinner, and whether it is uniform or clustered.

[0087] The feature fusion mechanism integrates the feature information from different input data through standardized and unified processing, forming a joint feature representation that more comprehensively and accurately describes the object being tested. Imagine a doctor needing to diagnose a condition; he would look at both X-rays (reflecting bone structure, similar to the three-dimensional geometric information provided by point cloud data) and blood test results (reflecting chemical composition indicators, similar to the spectral features provided by multispectral images). If he only relies on X-rays, he might miss information about inflammation; relying solely on test results, he might not be able to locate the lesion. Feature fusion is like the doctor precisely combining the "abnormal bone location shown on the X-ray" with the "abnormal indicators corresponding to that area in the test results" in his mind, thus making a more accurate diagnosis of "a specific inflammation occurring at a specific location." Therefore, by combining point cloud data with the spatial information of the segmented image, the feature fusion mechanism directly answers the question of the specific thickness of a certain chemical component at a certain location. Ultimately, the feature fusion mechanism generates a physical thickness map, i.e., a thickness distribution map, that has been interpreted and verified by the chemical information. The thickness distribution map identifies regions where thickness is of concern due to the presence of harmful chemicals, while also indicating areas where height variations may be harmless, such as those caused by structural adhesives. Therefore, feature fusion is a core strategy for collaborative analysis of multi-source heterogeneous data. By creating a richer joint feature representation through feature-level fusion, it ultimately achieves precise quantification of the spatial distribution relationship between the chemical composition and physical thickness of deposits on curtain wall surfaces.

[0088] Step 600: Based on the thickness distribution map, spatial pattern features are extracted using a convolutional neural network to generate an aging degree grade distribution map;

[0089] The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

[0090] While thickness information is an important indicator of aging, material aging often manifests as complex spatial patterns, such as thickness gradients, local aggregation patterns, and spatial relationships with structural features (e.g., joints, edges). To capture these complex spatial features containing crucial aging information, convolutional neural networks (CNNs), a deep learning technique, were introduced. CNNs are specifically trained to analyze the two-dimensional spatial data contained in thickness distribution maps. By learning the inherent patterns of a large number of samples with known aging levels (i.e., thickness distribution maps labeled with different aging grades), a mapping model from the spatial distribution characteristics of thickness to the severity of aging is established. When a new thickness distribution map is input, the trained CNN analyzes its thickness value pixel-by-pixel or region-by-region, along with the thickness variation patterns and spatial structural features of its surrounding neighborhood, to comprehensively determine the overall aging state characteristics of that local area.

[0091] Finally, through deep analysis using a convolutional neural network, an aging severity grading distribution map is generated. This map completely covers the entire target curtain wall surface in a spatial dimension, with each point (or area) in the map containing a label indicating its severity of aging. The aging severity grading distribution map not only quantifies the degree of aging but also includes information on the spatial heterogeneity of aging on the curtain wall surface, clearly identifying which areas are most severely aged (priority maintenance areas) and which areas are relatively intact. Furthermore, based on the aging severity grading distribution map, combined with a pre-set maintenance strategy knowledge base or rule engine, it is possible to intelligently derive differentiated maintenance treatment methods recommended for areas with different aging levels. For example, slightly aged areas may only require cleaning, moderately aged areas require local repairs, and severely aged areas require material replacement.

[0092] This invention provides an automatic inspection method for building curtain walls, addressing the comprehensive detection and maintenance issues of corrosion, dirt distribution, and decreased light transmittance on curtain wall surfaces. It acquires high-resolution images covering the visible to near-infrared bands, and uses a support vector machine algorithm to classify chemical components, generating a chemical composition classification map that includes the distribution areas of corrosion products and dirt types. Further, it acquires three-dimensional point cloud data using a laser rangefinder, denoises and registers it, and fuses it with multispectral images. This data is then input into a convolutional neural network to predict the thickness of the deposits, generating a thickness distribution map. A graded classification model generates a graded distribution map of aging levels, and combined with aging correlation analysis, it determines priority maintenance areas and treatment methods. This invention accurately identifies the types of deposits on curtain wall surfaces and precisely measures their thickness and distribution range through automated techniques, achieving accurate assessment of the aging state of curtain wall surfaces and intelligent maintenance decisions, significantly improving maintenance efficiency and curtain wall performance.

[0093] In one embodiment, please refer to Figure 2 The step of classifying the spectral features of the first multispectral image using a support vector machine algorithm to generate a chemical composition classification map based on the first multispectral image includes:

[0094] Step 201: Based on the first multispectral image, preprocessing techniques are used to remove noise and background interference to generate a second multispectral image;

[0095] Step 202: Based on the spectral features in the second multispectral image, calculate the spectral reflectance of each pixel using a spectral analysis algorithm to obtain a spectral feature set;

[0096] Step 203: Based on the spectral feature set, the support vector machine algorithm is used to classify the spectral feature set to obtain a preliminary classification result, wherein the preliminary classification result includes the corrosion product region corresponding to the corrosion product feature and the dirt type region corresponding to the dirt type feature.

[0097] Step 204: Based on the preliminary classification results, and in conjunction with the chemical composition database, match the chemical composition of corrosion product characteristics and fouling type characteristics to determine the chemical composition distribution of each corrosion product region or fouling type region.

[0098] Step 205: Using image processing technology, the distribution of chemical components is mapped onto the second multispectral image to generate a chemical component classification map.

[0099] A first multispectral image is acquired by collecting multispectral data of the curtain wall surface using a spectral imaging device, resulting in raw image data containing multiple bands. Preprocessing techniques are used to remove noise and background interference from the raw image data, generating a second multispectral image. Spectral features are extracted from the second multispectral image, and the spectral reflectance of each pixel is calculated using a spectral analysis algorithm to obtain a spectral feature set. A support vector machine algorithm is used to classify the spectral feature set, obtaining preliminary classification results. If the spectral reflectance matches preset corrosion product characteristics, the preliminary classification result is a corrosion product region; if it matches preset dirt type characteristics, the preliminary classification result is a dirt type region. Based on the preliminary classification results, and combined with a chemical composition database, the chemical components of the corrosion product characteristics and dirt type characteristics are matched to determine the chemical composition distribution of each region. Image processing techniques are used to map the chemical composition distribution onto the second multispectral image, generating a chemical composition classification map. Further post-processing can be performed on the chemical composition classification map, using a smoothing algorithm to optimize the region boundaries, generating the final chemical composition classification map.

[0100] Preprocessing techniques, in particular, involve a series of targeted operations performed before the core algorithm analyzes the data. These operations suppress or eliminate noise, interference, and unnecessary information in the data, while enhancing or highlighting valuable target information, thereby transforming raw, coarse data into high-quality, analysis-ready data. Preprocessing techniques mainly include denoising and normalization operations, as well as enhancement and feature highlighting operations. This improves the signal-to-noise ratio, ensures the reliability of the results, highlights key features, and enhances analytical accuracy. For example, preprocessing techniques for multispectral images can include radiometric calibration and atmospheric correction, image denoising, and image enhancement.

[0101] Spectral analysis algorithms are used to transform multispectral image data into standardized, quantifiable material characteristic data. Because matter can interact with electromagnetic radiation, and different chemical substances, due to their unique molecular structures, chemical bonds, and atomic compositions, exhibit selective absorption, reflection, or emission characteristics at specific wavelengths of light, thus forming unique "spectral fingerprints." This selective absorption and reflection results in light reflected back to the sensor exhibiting reflection peaks at certain wavelengths (strong reflection) and reflection valleys at other wavelengths (strong absorption). This curve of reflectivity versus wavelength is the "spectral characteristic" or "spectral signature" of the substance. Therefore, the task of spectral analysis algorithms is to identify the material composition represented by each pixel by interpreting its spectral curve (i.e., its "fingerprint") and comparing it with a database of known substance spectral fingerprints.

[0102] In this embodiment, the Support Vector Machine (SVM) algorithm is used to find the optimal hyperplane or decision boundary in a high-dimensional feature space composed of multispectral bands. This hyperplane or decision boundary can most clearly and reliably distinguish spectral feature points of different categories (such as different corrosion products and different types of dirt). Based on this boundary, unknown pixels are classified. As an efficient, accurate, and robust classifier, it achieves automated classification of massive pixel spectral data, generating a preliminary classification map that identifies the spatial distribution of corrosion products and dirt types. Therefore, the raw data obtained from the previous spectral analysis is transformed into a preliminary classification result with clear semantic information. Thus, the SVM algorithm can analyze the high-dimensional spectral feature vector pixel by pixel and, based on the trained model, determine which predefined category it is most likely to belong to, such as type A rust, type B verdigris, organic dirt, inorganic dust, etc. After classifying all pixels, the output result is the preliminary classification result. The preliminary classification result is essentially a label map, where the value of each pixel is no longer spectral reflectance, but rather its category number or category attribute.

[0103] The image processing techniques used in this embodiment are employed to achieve spatial alignment and attribute information assignment. This includes ensuring that the distribution of chemical components—specifically, the preliminary results obtained by the support vector machine algorithm, where each pixel corresponds to a chemical category—is spatially aligned with the second multispectral image at the pixel level. Furthermore, the chemical category attribute of each pixel is subsequently converted into an intuitive visual signal using predefined color coding rules, thereby generating a chemical component classification map that can be interpreted by both the human eye and a machine.

[0104] This embodiment proposes a high-precision, automated, and clearly chemically indicative method for spatial analysis of the state of curtain wall surfaces. It achieves deep information mining and standardized identification, utilizes a support vector machine algorithm for pixel-level classification to initially identify material category regions, and introduces a chemical composition database for precise matching. This enables the identification and labeling of specific chemical substances, improving the objectivity and automation level of the detection. The entire process is highly dependent on the algorithm, minimizing the intervention of human subjective judgment and significantly improving the automation level of the detection process and the objectivity and consistency of result interpretation.

[0105] In one embodiment, please refer to Figure 3 The step of generating a segmented image containing the spatial distribution range of the deposits based on the chemical composition classification map and using image segmentation technology includes:

[0106] Step 301: Based on the chemical composition classification map, the image is denoised and enhanced using preprocessing techniques to obtain the first processed image;

[0107] Step 302: Based on the first processed image, a semantic segmentation algorithm is used to generate the regional boundaries of corrosion products and dirt to obtain a second processed image;

[0108] Step 303: If the boundary clarity of the second processed image is lower than a preset threshold, the boundary of the second processed image is optimized by an edge detection algorithm to obtain a third processed image;

[0109] Step 304: Based on the third processed image, calculate the spatial distribution characteristics of corrosion products and dirt types, determine the range of the deposits, and obtain a distribution feature map;

[0110] Step 305: Based on the area ratio of corrosion products and dirt in the distribution feature map, a clustering algorithm is used to divide the accumulation area to obtain the area classification result;

[0111] Step 306: Based on the region classification results, generate a segmented image containing the spatial distribution range of the accumulated objects.

[0112] A chemical composition classification image is obtained, and the image is denoised and enhanced using preprocessing techniques to obtain a first processed image. Based on the chemical classification features in the first processed image, a semantic segmentation algorithm is used to generate the regional boundaries of corrosion products and dirt, resulting in a second processed image. If the boundary clarity of the second processed image is lower than a preset threshold, an edge detection algorithm is used to optimize the boundary of the second processed image, resulting in a third processed image. Based on the third processed image, the spatial distribution features of corrosion products and dirt are calculated to determine the extent of the accumulation, resulting in a distribution feature map. The area ratio of corrosion products and dirt is extracted from the distribution feature map, and a clustering algorithm is used to divide the accumulation regions, resulting in a region classification result. Based on the region classification result, the spatial distribution range of the accumulation is generated, resulting in a segmented image. Furthermore, geometric analysis can be performed on the distribution image to calculate the boundary coordinates of the accumulation, thus determining the spatial distribution range of the segmented image.

[0113] The semantic segmentation algorithm performs intensive classification prediction on each pixel in the image. It not only identifies the object category to which the pixel belongs, but more importantly, utilizes the surrounding context information to ensure consistency in classification results within the same object region, thus accurately delineating the boundaries of different semantic category regions. The input to the semantic segmentation algorithm is the preliminary classification result. Since each pixel already has a chemical category label, but there may be noise and isolated misclassified pixels, the semantic segmentation algorithm uses its context-aware capabilities to eliminate these isolated noise points and aggregate spatially continuous pixels of the same category to form complete, connected region objects. The output, the second-processed image, clearly delineates the precise boundaries, morphology, and spatial extent of different corrosion products and dirt types, no longer just scattered label points, but including regional entities that can be used for subsequent quantitative analysis.

[0114] Edge detection algorithms are enhanced post-processing techniques that refine and optimize the boundaries of preliminary segmentation results. They are used to detect points in an image where the grayscale values ​​(or more broadly, feature values) of local regions change drastically or are discontinuous, thus locating and identifying the boundaries between different regions. Since the grayscale, color, or texture attributes of pixels at the boundaries of different objects and the intersections of different regions in an image often undergo significant and sudden changes, edge detection algorithms act as detectors for these changes. Unlike semantic segmentation algorithms, edge detection algorithms do not perform initial segmentation but rather optimize and refine the preliminary boundaries generated by semantic segmentation algorithms. Specifically, they optimize and sharpen blurred boundaries, compensating for contextual blind spots in semantic segmentation.

[0115] In this embodiment, the clustering algorithm automatically discovers and divides different data points—that is, the natural groups formed by various regions of the curtain wall—within a one-dimensional or multi-dimensional feature space of area proportion, without any pre-defined labels or thresholds. This ensures that regions within the same group have highly similar area proportions, while regions in different groups have significantly different area proportions. The input to the clustering algorithm is area proportion data, where each region of the curtain wall is pre-divided. This region may be a grid cell or a connected region obtained from semantic segmentation. The feature value of each data point is the area proportion of corrosion products and dirt in that region. The similarity between different regions is measured by calculating distance. In this one-dimensional space, the distance is calculated as the absolute difference between the area proportion values ​​of two regions. The smaller the difference, the more similar the pollution coverage of the two regions, and the more likely they are to be grouped into the same group. Based on the similarity of the area proportion feature, iterative optimization automatically finds natural groups in the data. Therefore, the core function of the clustering algorithm is to serve as an objective and automatic data analysis tool, replacing subjective experience-based judgment, and realizing intelligent, data-driven hierarchical classification of the severity of surface deposits on curtain walls.

[0116] This application's embodiments achieve high-precision, adaptive, and intelligent spatial range definition and regional classification of complex deposits on curtain wall surfaces. Compared with traditional threshold segmentation or manual delineation methods, it achieves high-precision, automated, and intelligent extraction and classification of the spatial range of deposits on curtain wall surfaces through adaptive boundary optimization, spatial feature quantification analysis, and intelligent clustering and partitioning.

[0117] In one embodiment, please refer to Figure 4 Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface to obtain the first point cloud data, including:

[0118] Step 401: Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface, and a second point cloud data is generated.

[0119] Step 402: Based on the second point cloud data, noise reduction processing is performed using a point cloud filtering algorithm to obtain the third point cloud data;

[0120] Step 403: If the point density of the third point cloud data is lower than the preset density threshold, then based on the third point cloud data, an interpolation algorithm is used to supplement the missing points and generate the fourth point cloud data.

[0121] Step 404: Based on the fourth point cloud data, the iterative nearest point algorithm is applied for registration processing to obtain the fifth point cloud data;

[0122] Step 405: Based on the fifth point cloud data, extract the geometric features of the curtain wall surface and generate surface feature data;

[0123] Step 406: Based on the surface feature data, detect the distribution area containing the deposits and generate deposit distribution data;

[0124] Step 407: Based on the distribution data of the deposits, a clustering algorithm is used to classify the data and determine the first point cloud data containing the deposit type and deposit thickness.

[0125] The core of this embodiment lies in acquiring and analyzing the three-dimensional geometric morphology of the deposits on the curtain wall surface, thereby identifying their type and degree of accumulation. Based on a segmented image defining the spatial extent of the deposits, a laser rangefinder is used to perform targeted three-dimensional scanning of the curtain wall surface, including the deposited areas identified in the segmented image. This process generates a dense set of three-dimensional spatial points, i.e., the second point cloud data, which directly captures the actual three-dimensional coordinate information of the target deposits and the surrounding curtain wall substrate surface. To ensure the accuracy and usability of the data, the second point cloud data needs to be processed by a point cloud filtering algorithm. The point cloud filtering algorithm is used to identify and remove outliers introduced by environmental interference, measurement errors, or sensor noise, resulting in the third point cloud data.

[0126] Considering that actual scanning conditions may lead to insufficient point density in some areas, a density judgment and point cloud data interpolation process is set up to automatically evaluate the point density of the third point cloud data and interpolate it. The specific steps are as follows: If the third point cloud data is lower than a preset density threshold, it indicates that the 3D information in that area may be sparse or incomplete. The interpolation algorithm is then automatically triggered. Based on the spatial position and geometric relationship of any valid point adjacent to any point cloud data in the third point cloud data, the interpolation algorithm intelligently estimates and supplements the missing 3D points, generating a fourth point cloud data with more uniform coverage and more complete information, effectively filling information gaps. Subsequently, to ensure seamless alignment of point cloud data acquired from multiple scans or different perspectives in a unified spatial coordinate system, especially for large-area curtain walls requiring segmented scanning, an iterative nearest-point algorithm is applied to perform precise registration processing on the fourth point cloud data. This algorithm iteratively calculates the optimal spatial transformation parameters to align fourth point cloud data from different sources or locations to a common reference coordinate system, outputting the fifth point cloud data and laying the foundation for global consistency analysis.

[0127] After obtaining high-quality, registered fifth point cloud data, key surface feature data is extracted. This data includes calculating the normal vector, curvature, and elevation gradient of the deposited areas on the curtain wall surface to describe local geometric features, as well as regional features, such as the flatness or undulation of the deposits. By analyzing the extracted surface feature data, the distribution areas truly containing significant deposits in 3D space are detected and confirmed, generating deposit distribution data. This step verifies and refines the 2D segmentation information in 3D space. Finally, based on the deposit distribution data, a clustering algorithm is used to classify the deposits. The clustering algorithm groups deposits with similar 3D features into different categories, thereby determining the deposit type and thickness of the first point cloud data. Deposits include loose porous dirt and dense layered corrosion products, and deposit thickness includes slight protrusions and significant deposits.

[0128] Point cloud filtering algorithms are used to identify and remove sparse, isolated noise points that do not conform to the main distribution pattern, based on the distribution relationship between each point and its neighbors in three-dimensional space. This retains valid data points representing the actual curtain wall surface. The calculation process is not a simple uniform smoothing, but rather an intelligent discrimination based on neighborhood statistical characteristics. The core principle of point cloud filtering algorithms lies in intelligently identifying and removing statistically significant outlier noise points based on the local spatial statistical characteristics of point cloud data (such as the average neighborhood distance). Its core function is as an essential data preprocessing technique to purify the original point cloud and improve data quality.

[0129] Interpolation algorithms are a key data processing technique that repairs missing data and constructs continuous distribution models based on the principle of spatial correlation. They are used to estimate or predict values ​​at unsampled locations between or outside the known points by establishing appropriate mathematical models or functional relationships, using a set of known discrete data points. Essentially, interpolation algorithms are a type of data repair and continuous surface builder. Their principle is based on the numerical values ​​and spatial relationships of known sampled points, using weighted averaging or geostatistical methods to predict the values ​​of unsampled points. This can solve the problem of data sparsity, transforming discrete point data into a continuous, complete, and analytically usable spatial distribution model.

[0130] The Iterative Closest Point Algorithm (IBPA) is based on iterative spatial transformation optimization using nearest-point search. Through iterative calculation, it continuously optimizes a spatial transformation to minimize the sum of distances between corresponding points in two point clouds, thus achieving precise alignment of the two point clouds in 3D space. Its calculation process is a continuous iterative and progressively refined optimization process, which can solve the problem of spatial inconsistency in point cloud data caused by block scanning or multiple scans, laying the foundation for all subsequent analyses based on a unified coordinate system. For large curtain walls, multiple scans from multiple angles are usually required to cover the entire surface. The point cloud obtained from each scan is located in its independent scanner coordinate system. The IBPA can accurately register these point clouds in different coordinate systems (fourth point cloud data) to the same global coordinate system, forming a complete and seamless 3D model of the curtain wall (fifth point cloud data), as if it were completed in a single scan. Therefore, as a high-precision 3D data registration technique, the IBPA can unify multi-source, multi-view point cloud data into the same coordinate system.

[0131] For example, for the first segmented image, a laser rangefinder was used to acquire 3D point cloud data of the deposits on the curtain wall surface. Specifically, the laser rangefinder was used with a scanning frequency of 976,000 points / second and a point cloud resolution of 0.019 degrees, covering a 10m × 10m area on the curtain wall surface, generating original point cloud data containing approximately 5 million points with a point spacing of approximately 1.5mm. The point cloud data includes x, y, and z coordinates and reflection intensity values, with an accuracy of ±1mm. For noise reduction, a statistical filtering algorithm was used, setting the neighborhood radius to 0.05m and the standard deviation threshold to 2, removing outliers that deviated from the mean by twice the standard deviation, reducing noise points by approximately 2%, and retaining approximately 4.9 million points. Next, registration was performed using an iterative nearest-point algorithm to align the point clouds from multiple frames to a unified coordinate system. The initial alignment was based on feature point matching, selecting the corner points of the curtain wall edge, with an error threshold set to 0.01m. After 20 iterations, the point cloud alignment error decreased to below 0.005m, generating the second point cloud data. The analysis process includes calculating point cloud density, ensuring uniformity of point spacing, and detecting local areas where the point density is below 0.8 points / ... When the time is right, the rescan logic is triggered, and the laser rangefinder is called to rescan the low-density area. The final point cloud data is used for subsequent accumulation thickness analysis. Combined with the curtain wall design model, the accumulation height difference is calculated, with an average thickness of 2.3 mm and a standard deviation of 0.15 mm, ensuring that the data can be used for subsequent cleaning planning.

[0132] In this embodiment, adaptive point cloud optimization, high-precision registration, deep extraction of three-dimensional geometric features, and feature-based intelligent clustering are used to achieve intelligent determination of the type and degree of the deposits on the curtain wall surface from the three-dimensional spatial morphology. This provides accurate spatial location and thickness information, as well as the physical properties and severity level of the deposits.

[0133] In one embodiment, please refer to Figure 5 The step of determining the thickness distribution map corresponding to the thickness of the accumulation material by combining the spatial information of the first point cloud data and the segmented image using a feature fusion mechanism, based on the first point cloud data and the segmented image, includes:

[0134] Step 501: Based on the first point cloud data and the segmented image, a point cloud registration algorithm is used to align the first point cloud data and the segmented image to obtain a registered point cloud image dataset.

[0135] Step 502: Based on the registered point cloud image dataset, the spatial information therein is fused through a feature fusion mechanism to generate a fused feature set;

[0136] Step 503: Based on the fused feature set, a convolutional neural network is used to process the fused feature set to extract high-dimensional features and obtain a feature vector set;

[0137] Step 504: Input the feature vector set into the thickness regression model to obtain the thickness prediction value corresponding to the thickness of the deposit;

[0138] Step 505: Based on the predicted thickness value, a three-dimensional interpolation algorithm is used to generate thickness distribution data and generate a thickness distribution map.

[0139] The core objective of this stage is to accurately quantify the thickness and spatial distribution of the deposits on the curtain wall surface. The key lies in fully utilizing and fusing complementary information from multispectral imagers and laser rangefinders. First, a first point cloud dataset containing information on the deposit type and thickness, along with segmented images defining the two-dimensional spatial extent of the deposits, is used. To achieve collaborative analysis of geometric and spectral information, a point cloud registration algorithm is first employed. This algorithm rigorously aligns the three-dimensional spatial coordinates of the first point cloud dataset with the two-dimensional pixel coordinates and spectral information of the segmented image source. This registration process ensures that each three-dimensional point in the point cloud precisely corresponds to a specific pixel location on the multispectral image, forming a spatially consistent registered point cloud image dataset, laying the geometric foundation for subsequent fusion.

[0140] Based on the registered dataset, a feature fusion mechanism is implemented to deeply fuse the 3D geometric information and spectral features provided by the registered point cloud image dataset in space, generating a fused feature set that simultaneously contains geometric and spectral attributes. This fused feature set more comprehensively characterizes the physical morphology and chemical composition of the deposit. Subsequently, to extract implicit, deeper, and thickness-related complex patterns from the fused feature set, a convolutional neural network (CNN) is employed. CNNs possess powerful spatial feature extraction capabilities, automatically learning the local and global correlation patterns of the fused features on the curtain wall surface. This transforms the fused feature set into a feature vector set containing higher-level information, where the feature vectors contain the nonlinear relationship between the deposit's thickness and the surrounding geometric and spectral environment. The feature vector set is then input into a pre-trained thickness regression model, which can be a deep learning-based regression network. This thickness regression model directly predicts the deposit's thickness at each location based on the learned patterns, achieving a mapping from multimodal features to thickness measurements. Finally, for the smooth thickness values ​​at all locations, a three-dimensional interpolation algorithm is used to generate continuous and smooth thickness distribution data on the two-dimensional surface of the curtain wall, and the final thickness distribution map is drawn accordingly. The resulting thickness distribution map shows the detailed spatial variation of the deposit thickness on the entire curtain wall surface in the form of pseudo-color or contour lines.

[0141] Point cloud registration algorithms are used to solve for the mapping relationship between two-dimensional and three-dimensional spaces. Their core principle is to find an optimal spatial transformation parameter to achieve precise pixel-level alignment between the three-dimensional point cloud data (the first point cloud data) and the two-dimensional segmented image, thus establishing a one-to-one correspondence between each three-dimensional point in the point cloud and each pixel in the image. This includes coordinate system and mapping model establishment, feature-based matching and transformation solving, and reprojection and accuracy evaluation. Specifically, a mathematical mapping relationship is established between the three-dimensional world coordinate system (the space where the point cloud is located) and the two-dimensional image pixel coordinate system (the space where the segmented image is located). Then, more fundamental common features are extracted for precise alignment. The algorithm extracts easily identifiable feature points such as corners or edges from the point cloud data and establishes a correspondence between the found three-dimensional feature points and two-dimensional pixels through feature descriptor matching or geometric constraints. Using these 2D-3D point pair correspondences, the optimal solution is found, i.e., determining the rotation matrix R and translation vector T required to accurately project the three-dimensional point cloud onto the two-dimensional image plane. The calculated transformation parameters are applied to the entire first point cloud data, reprojecting it onto a 2D image plane. The degree of match between the reprojected point cloud contour and the edge contours of the regions in the segmented image is evaluated. Through iterative optimization, the error between the two is minimized. Finally, when the registration accuracy meets the requirements, the registered point cloud image dataset is obtained. This dataset means that for any 3D point in the point cloud, its corresponding pixel position and label (such as the type of erosion) on the segmented image can be uniquely determined; conversely, for any pixel on the segmented image, given its depth information, its corresponding 3D point in the point cloud can also be found.

[0142] The principle behind the thickness regression model is a mathematical function trained through machine learning. It learns from high-dimensional fused features and establishes a complex nonlinear mapping to thickness values. Its role is to serve as the intelligent decision-making endpoint in the entire thickness calculation pipeline, transforming abstract multimodal features into accurate, robust, and physically meaningful thickness predictions. The input to the thickness regression model is a set of feature vectors, where each feature vector is a high-dimensional array extracted from the fused feature set by the convolutional neural network. Its output is a continuous scalar value, representing the predicted thickness of the deposit at that point or in that region. Essentially, the thickness regression model is an intelligent thickness estimator that transforms the abstract fused features extracted in the previous step—features that are difficult for humans to intuitively understand—into accurate thickness data with clear physical meaning and engineering value. This achieves end-to-end accurate thickness estimation, improving the robustness and accuracy of thickness measurement.

[0143] The core principle of 3D interpolation algorithms is to construct a continuous distribution model in 3D space by using the numerical values ​​and 3D coordinates of a series of known scattered points, based on the principles of spatial proximity and correlation, to predict the attribute values ​​of unknown points from discrete sampling points. This allows for the prediction of numerical values ​​at unknown locations, thus forming a continuous 3D data volume or surface. Essentially, 3D interpolation algorithms are builders from discrete samples to continuous distribution models. Their core function is as a key converter from discrete computation to continuous output, transforming sparse thickness predictions into continuous, complete thickness distribution data and visualization maps, generating continuous spatial thickness distribution data. Since the thickness predictions or smoothed thickness values ​​generated by thickness regression models are discrete, they only exist at the locations of point cloud data, and most areas of the curtain wall surface lack directly measurable values. 3D interpolation algorithms fill these gaps, generating continuous thickness distribution data covering the entire region of interest of the curtain wall. This allows us to know the thickness value at any point on the surface, not just at the measurement point.

[0144] This application's embodiments achieve high-precision, robust, and spatially continuous measurement and visualization of the thickness of deposits on curtain wall surfaces through innovative multimodal fusion, deep learning prediction, adaptive post-processing, and spatial interpolation. It transforms abstract sensor data into intuitive, quantitative, and directly guides maintenance actions with a three-dimensional thickness map, providing powerful data-driven support for preventive maintenance, life prediction, and resource optimization of building curtain walls.

[0145] In one embodiment, please refer to Figure 6 The step of extracting spatial pattern features from the thickness distribution map using a convolutional neural network to generate an aging degree grading distribution map includes:

[0146] Step 601: Based on the thickness distribution data of the thickness distribution map, perform rotation and flip operations using a data augmentation method to obtain an augmented thickness distribution dataset;

[0147] Step 602: Based on the enhanced thickness distribution dataset, spatial pattern features are extracted using a convolutional neural network to obtain a feature vector set;

[0148] Step 603: If the feature vector set meets the preset feature integrity threshold, then input the feature vector set into the hierarchical classification model to determine the aging degree classification result.

[0149] Step 604: Based on the aging degree classification results, a visualization algorithm is used to generate graded distribution data;

[0150] Step 605: Perform spatial mapping processing on the hierarchical distribution data to generate high-resolution aging degree distribution data, and obtain an optimized distribution map;

[0151] Step 606: If the optimized distribution map meets the preset clarity threshold, then image smoothing technology is used for processing to obtain the aging degree classification distribution map.

[0152] Thickness distribution data is acquired, and data augmentation methods are used to perform rotation and flipping operations to obtain an augmented thickness distribution dataset. A convolutional neural network is then used to process the augmented thickness distribution dataset to extract spatial pattern features, resulting in a feature vector set. If the feature vector set meets a preset feature integrity threshold, it is input into a hierarchical classification model, and the aging degree grading result is determined based on a pre-established material aging degree standard. Based on the aging degree grading result, a visualization algorithm is used to generate a hierarchical distribution of aging degree, resulting in hierarchical distribution data. Spatial mapping processing is performed on the hierarchical distribution data to generate high-resolution aging degree distribution data, resulting in an optimized distribution map. If the optimized distribution map meets a preset sharpness threshold, image smoothing techniques are used to process the distribution map, resulting in a hierarchical aging degree distribution map. Furthermore, based on the hierarchical aging degree distribution map, a data compression algorithm is used to store the results, resulting in a compressed hierarchical aging degree distribution map.

[0153] The hierarchical classification model is a machine learning-trained discriminative model that learns from high-dimensional feature vectors and establishes a complex nonlinear mapping between ordered aging levels. Its core function is to simultaneously perform classification and ranking. The input to the hierarchical classification model is a set of feature vectors, where each feature vector represents a comprehensive state description of a small area or point on the curtain wall surface. These vectors are extracted from thickness distribution data by the convolutional neural network and contain rich spatial pattern features, such as thickness uniformity, gradient directionality, texture roughness, and local shape. The output of the model is an ordered discrete label, i.e., the aging level classification result, which can be defined as slight, moderate, or severe. The hierarchical classification model is a machine learning model capable of understanding the complex mapping relationship between features and ordered levels. It performs intelligent, context-aware assessment by comprehensively analyzing the spatial pattern features of the thickness distribution. Its core role is as the "decision brain" of the entire inspection pipeline, transforming physical measurement data into aging level information with engineering semantics, providing the final, directly executable judgment basis for achieving precise and automated maintenance decisions.

[0154] The principle of the visualization algorithm is to transform the abstract, digitized aging severity classification results into intuitive, spatial visual symbols through a set of predefined visual encoding rules, thereby generating hierarchical distribution data that retains the original spatial information while conveying the meaning of the classification levels. Its technical principle includes two key operations: visual encoding and spatial mapping. Visual encoding is the core of the algorithm; it assigns a unique and easily distinguishable visual channel value to each discrete aging level, with color being the most commonly used and effective visual channel. After visual encoding, a spatial mapping process is performed, traversing every spatial location in the aging severity classification result data, such as every pixel or grid cell in an image. At each location, the algorithm reads its aging level label, then finds the corresponding color value according to a color lookup table, and assigns this color value to the pixel at the same spatial location in the hierarchical distribution data. The final generated hierarchical distribution data is essentially a pseudo-color image, where the color of each pixel no longer represents the true color but rather the severity of aging at that location. In summary, visualization algorithms map discrete level labels to color values ​​based on visual encoding rules and accurately maintain their spatial relationships. Their core function is to act as a bridge for human-computer interaction, transforming the digital results output by machine learning into an intuitive, spatial "maintenance map," thereby maximizing the release of the insights contained in the data and empowering accurate decision-making and efficient operation.

[0155] For example, based on a thickness distribution map, a data augmentation method is used to process the thickness distribution data. Assume the original data is a 100×100 pixel thickness distribution matrix with values ​​ranging from 0.1 to 5.0 mm. First, augmented data is generated through rotation operations. Specifically, the matrix is ​​rotated clockwise by 90°, 180°, and 270° to generate three new matrices. Simultaneously, horizontal and vertical flipping is performed to generate two flipped matrices, resulting in a total of five augmented datasets. The rotation algorithm uses bilinear interpolation to ensure smooth transitions in pixel values.

[0156] For example, after a 90° rotation, the original coordinates (10, 20) with a value of 4.2 mm are mapped to the new coordinates (20, 90). Subsequently, the original and augmented data are fed into a convolutional neural network. The network structure consists of three convolutional layers (3×3 kernel size, 1 stride, 1 padding), each followed by a ReLU activation function and a 2×2 max-pooling layer to extract spatial features of the thickness distribution, such as edge gradients and region uniformity. After feature extraction, a 256-dimensional feature vector is obtained. This is based on a pre-established material aging grading standard.

[0157] For example, thicknesses less than 1.0 mm are classified as severely aged, 1.0-2.5 mm as moderately aged, and greater than 2.5 mm as mildly aged. A hierarchical classification model is used to map feature vectors to three aging levels. During training, the cross-entropy loss function is used, with a batch size of 32, and 20 training epochs, achieving a validation set accuracy of 92%. Finally, an aging level distribution map is generated by back-mapping the classification results to the original matrix coordinates, generating a 100×100 level matrix. Severely aged areas are marked in red, moderately aged in yellow, and mildly aged in green. Analysis shows that severely aged areas are concentrated in the boundary regions with thicknesses less than 1.0 mm, accounting for approximately 15%, indicating a need to optimize the material boundary treatment process.

[0158] This embodiment achieves the transformation from physical thickness measurement to engineering aging assessment by deeply integrating thickness spatial pattern analysis and intelligent classification, guiding the formulation of differentiated maintenance strategies, such as emergency repair of red areas and regular monitoring of yellow areas, which significantly improves maintenance efficiency and reduces unexpected risks.

[0159] In one embodiment, please refer to Figure 7 The determination of priority maintenance areas and maintenance methods for these priority maintenance areas in a building curtain wall includes the following steps:

[0160] Step 701: Based on the areas of damaged light transmittance of the curtain wall in the aging degree classification distribution map, a light transmission model is used, combined with the spectral characteristics of the areas of damaged light transmittance, to determine the degree of light transmittance reduction.

[0161] Step 702: Based on the degree of decrease in transmittance, a convolutional neural network is used to perform spatial pattern analysis to generate a transmittance distribution map;

[0162] Step 703: Based on the light transmittance distribution map and the aging degree classification distribution map, a decision tree algorithm is used to determine the priority maintenance area and maintenance treatment method.

[0163] Data on areas with impaired light transmittance of the curtain wall were obtained from the aging degree classification distribution map. A region segmentation algorithm was used to extract the boundaries of these impaired areas, resulting in a dataset of impaired areas. Based on this dataset and the thickness distribution map, a light transmittance model was used to calculate the light transmittance of each area, yielding light transmittance distribution data. If the light transmittance distribution data met a preset integrity threshold, a spectral feature extraction algorithm was used to obtain the spectral feature vectors of the impaired areas, indicating the degree of light transmittance reduction. Based on the spectral feature dataset corresponding to the degree of light transmittance reduction, a convolutional neural network was used for spatial pattern analysis to obtain a set of light transmittance reduction feature vectors. If the set of light transmittance reduction feature vectors met a preset feature significance threshold, a spatial mapping algorithm was used to map the set of light transmittance reduction feature vectors to the spatial distribution of the curtain wall material, resulting in a preliminary light transmittance distribution map. An image enhancement algorithm was used to optimize the resolution of the preliminary light transmittance distribution map, resulting in a high-resolution light transmittance distribution map.

[0164] Furthermore, based on the light transmittance distribution map and the aging degree classification distribution map, image processing techniques can be used to extract data features from these maps, resulting in a feature dataset. This dataset includes light transmittance distribution features and aging degree classification distribution features. Based on the data features of light transmittance distribution and aging degree classification in the feature dataset, correlation analysis is used to calculate the aging correlation between the two, resulting in a correlation coefficient matrix. If the correlation coefficient of a certain region in the correlation coefficient matrix is ​​higher than a preset threshold, the light transmittance distribution and aging degree classification are classified according to the decision model to determine region priorities, resulting in a priority maintenance region list. This priority maintenance region list includes the priority maintenance regions and their order. For the performance degradation trend of each priority maintenance region in the priority maintenance region list, a rule matching method is used to determine the processing method, resulting in a maintenance processing method. Alternatively, the processing method for each region in the set of maintenance processing methods can be combined with the priority maintenance region list to generate a maintenance decision basis. Based on the maintenance decision basis, data integration techniques are used to output the priority maintenance regions and processing methods in a structured manner, resulting in a maintenance decision scheme.

[0165] The process begins by acquiring a transmittance distribution map from a multispectral imaging device, calculating the coverage rate, and determining if the coverage rate meets a preset threshold. If the coverage rate does not meet the preset threshold, the scanning path of the multispectral imaging device is optimized based on the missing areas in the transmittance distribution map, resulting in an adjusted scanning path. Using the adjusted scanning path, the multispectral imaging device acquires supplementary multispectral image data, resulting in supplementary image data. This supplementary image data is then preprocessed, employing a mean filtering algorithm to remove noise, resulting in denoised supplementary image data. The denoised supplementary image data is then fused with the transmittance distribution map, and a weighted average algorithm is used to generate an extended transmittance distribution map, resulting in an extended distribution map. Based on this extended distribution map, the distribution features of the deposits are extracted, generating extended deposit distribution information, resulting in deposit distribution information. This deposit distribution information is then verified. If the coverage rate still does not meet the preset threshold, the process returns to the scan path adjustment step, repeating the acquisition and fusion process until deposit distribution information that meets the threshold is obtained.

[0166] Furthermore, chemical composition data can be obtained from the distribution information of the sediment. Spectral analysis is used to process the collected samples to obtain composition classification results. For the composition classification results, image segmentation is performed, dividing the image into regions using a preset threshold to obtain a composition segmentation image. Chemical composition data is extracted from the composition segmentation image, and laser point cloud data is collected. Point cloud registration technology is used to integrate the chemical composition data and laser point cloud data to obtain a three-dimensional point cloud model. Based on the three-dimensional point cloud model, a random forest algorithm is used to predict the thickness distribution, obtaining predicted thickness values. If the predicted thickness distribution value exceeds a preset threshold, it is marked as an abnormal region, resulting in a thickness distribution prediction result. Based on the thickness distribution prediction result, the degree of aging is analyzed. Time series analysis is used to calculate aging characteristic values, obtaining the aging degree distribution. Key parameters are extracted from the aging degree distribution. If the key parameter values ​​exceed a preset threshold, it is determined to be a maintenance-required area, obtaining a maintenance priority area. Based on the maintenance priority area, combined with the chemical composition data and the thickness distribution prediction results, a weighted scoring method is used to generate a maintenance decision basis.

[0167] The principle of the light transmission model is to establish a physical-mathematical model describing the propagation and attenuation process of light in a specific material. By using the spectral characteristics of the material as key input parameters, it simulates and calculates the intensity attenuation of light after passing through the material layer, thereby quantitatively assessing the degree of decline in its light transmittance. The model's input is not merely the presence or absence of deposits, but rather the specific spectral characteristics of the deposits in the areas where light transmittance is impaired, i.e., their reflectivity or absorptivity data across multiple wavelengths. These spectral characteristics are the "fingerprints" of the material, directly reflecting its chemical composition and microstructure, thus determining its absorption and scattering coefficients. The model's output is the degree of light transmittance reduction, quantitatively representing the proportion of light that should have passed through the glass due to the presence of deposits. The light transmission model is essentially a quantitative evaluator from chemical morphology to functional performance. Its role transcends traditional testing methods that only focus on the substance's type and thickness, thereby quantitatively assessing the impact of these deposits on the core physical function of the curtain wall—light transmittance—and its actual effects, achieving a quantitative and non-destructive evaluation of the curtain wall's core physical function, light transmittance.

[0168] The principle of convolutional neural networks (CNNs) is to automatically learn the spatial patterns and contextual features of transmittance degradation data by utilizing their inherent local connections, weight sharing, and hierarchical abstraction mechanisms. Through an encoder-decoder structure, they reconstruct a high-resolution distribution map. Their core function is as an advanced intelligent spatial data analysis tool, transforming sparse transmittance indicators into an accurate, detailed, and reliable spatialized functional performance map. As a powerful spatial feature extractor, CNNs can automatically learn the local and global patterns of transmittance degradation data on a two-dimensional curtain wall surface, mapping these complex spatial features into a more accurate and detailed transmittance performance distribution map.

[0169] The principle of the decision tree algorithm is to simulate the decision-making process of human experts. Through a series of pre-learned or pre-defined conditional rules based on input features, it performs multi-dimensional conditional judgments on each area of ​​the curtain wall, ultimately arriving at a clear classification decision: maintenance priority and treatment method. Its input includes two core features: the area's level in the aging degree distribution map and the degree of light transmittance reduction in the area's light transmittance distribution map. The output is a clear decision conclusion: the determination of the priority maintenance area and the recommended maintenance treatment method. The decision tree algorithm, through a tree-like, rule-based sequence of conditional judgments, performs multi-dimensional fusion analysis of the aging degree and functional impact of each area. Its core role is to act as the decision terminal of the entire automated inspection system, transforming the results of all preceding technical processes into an optimal, interpretable, and executable maintenance action plan.

[0170] This embodiment achieves functional impact assessment and precise maintenance decision-making for curtain wall inspection through quantitative modeling of light transmittance performance, fusion of multi-dimensional spatial information, and rule-based intelligent decision-making.

[0171] The automatic building curtain wall detection system provided by the present invention is described below. The automatic building curtain wall detection system described below can be referred to in correspondence with the automatic building curtain wall detection method described above.

[0172] The present invention also provides an automatic detection system for building curtain walls, comprising:

[0173] The acquisition module is used to acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data.

[0174] The classification map generation module is used to classify the spectral features of the first multispectral image based on the first multispectral image using a support vector machine algorithm, and generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution area of ​​corrosion products and dirt types on the curtain wall surface.

[0175] Image segmentation and generation: Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits.

[0176] The point cloud data generation module is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface based on the segmented image using a laser rangefinder, and obtain the first point cloud data.

[0177] The distribution map generation module is used to combine the spatial information of the first point cloud data and the segmented image based on the first point cloud data and the segmented image using a feature fusion mechanism to determine the thickness distribution map corresponding to the thickness of the deposit.

[0178] The graded distribution map generation module is used to extract spatial pattern features from the thickness distribution map using a convolutional neural network to generate a graded distribution map of aging degree.

[0179] The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

[0180] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute an automatic building curtain wall detection method.

[0181] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the automatic detection method for building curtain walls provided by the above methods.

[0183] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0184] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic detection method for building curtain walls, characterized in that, Executed by a computer, including: Acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data; Based on the first multispectral image, a support vector machine algorithm is used to classify the spectral features of the first multispectral image to generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution area of ​​corrosion products and dirt types on the curtain wall surface. Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits; Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface to obtain the first point cloud data. Based on the first point cloud data and the segmented image, a feature fusion mechanism is used to combine the spatial information of the first point cloud data and the segmented image to determine the thickness distribution map corresponding to the thickness of the accumulation. Based on the thickness distribution map, spatial pattern features are extracted using a convolutional neural network to generate an aging degree grade distribution map; The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

2. The automatic detection method for building curtain walls according to claim 1, characterized in that, The step of classifying the spectral features of the first multispectral image using a support vector machine algorithm to generate a chemical composition classification map based on the first multispectral image includes: Based on the first multispectral image, preprocessing techniques are used to remove noise and background interference to generate a second multispectral image; Based on the spectral features in the second multispectral image, the spectral reflectance of each pixel is calculated using a spectral analysis algorithm to obtain a spectral feature set; Based on the spectral feature set, the support vector machine algorithm is used to classify the spectral feature set to obtain a preliminary classification result, wherein the preliminary classification result includes the corrosion product region corresponding to the corrosion product feature and the dirt type region corresponding to the dirt type feature. Based on the preliminary classification results, and in conjunction with the chemical composition database, the chemical composition of corrosion product characteristics and fouling type characteristics is matched to determine the chemical composition distribution of each corrosion product region or fouling type region. By using image processing technology, the distribution of chemical components is mapped onto the second multispectral image to generate a chemical component classification map.

3. The automatic detection method for building curtain walls according to claim 1, characterized in that, Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits, including: Based on the chemical composition classification map, noise reduction and enhancement are performed using preprocessing techniques to obtain the first processed image; Based on the first processed image, a semantic segmentation algorithm is used to generate the regional boundaries of corrosion products and dirt to obtain a second processed image. If the boundary clarity of the second processed image is lower than a preset threshold, the boundary of the second processed image is optimized by an edge detection algorithm to obtain a third processed image. Based on the third processed image, the spatial distribution characteristics of corrosion products and dirt types are calculated to determine the range of the deposits and obtain a distribution feature map. Based on the area ratio of corrosion products and dirt in the distribution feature map, a clustering algorithm is used to divide the accumulation area and obtain the area classification result. Based on the region classification results, a segmented image containing the spatial distribution range of the accumulated objects is generated.

4. The automatic detection method for building curtain walls according to claim 1, characterized in that, Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface, resulting in the first point cloud data, including: Based on the segmented image, a laser rangefinder is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface, and a second point cloud data is generated. Based on the second point cloud data, noise reduction is performed using a point cloud filtering algorithm to obtain the third point cloud data; If the point density of the third point cloud data is lower than the preset density threshold, then based on the third point cloud data, an interpolation algorithm is used to supplement the missing points and generate the fourth point cloud data. Based on the fourth point cloud data, the iterative nearest point algorithm is applied for registration processing to obtain the fifth point cloud data. Based on the fifth point cloud data, the geometric features of the curtain wall surface are extracted to generate surface feature data; Based on the surface feature data, the distribution area containing the deposits is detected, and the deposit distribution data is generated. Based on the distribution data of the deposits, a clustering algorithm is used for classification to determine the first point cloud data, which includes the deposit type and deposit thickness.

5. The automatic detection method for building curtain walls according to claim 1, characterized in that, The step of combining the spatial information of the first point cloud data and the segmented image using a feature fusion mechanism to determine the thickness distribution map corresponding to the thickness of the accumulation material includes: Based on the first point cloud data and the segmented image, a point cloud registration algorithm is used to align the first point cloud data and the segmented image to obtain a registered point cloud image dataset. Based on the registered point cloud image dataset, the spatial information is fused through a feature fusion mechanism to generate a fused feature set; Based on the fused feature set, a convolutional neural network is used to process the fused feature set, extract high-dimensional features, and obtain a feature vector set; The feature vector set is input into the thickness regression model to obtain the predicted thickness value corresponding to the thickness of the deposit. Based on the predicted thickness values, a three-dimensional interpolation algorithm is used to generate thickness distribution data and a thickness distribution map.

6. The automatic detection method for building curtain walls according to claim 1, characterized in that, The step of extracting spatial pattern features from the thickness distribution map using a convolutional neural network to generate an aging degree grading distribution map includes: Based on the thickness distribution data of the thickness distribution map, a data augmentation method is used to perform rotation and flipping operations to obtain an augmented thickness distribution dataset; Based on the enhanced thickness distribution dataset, spatial pattern features are extracted using a convolutional neural network to obtain a feature vector set; If the feature vector set meets the preset feature integrity threshold, then the feature vector set is input into the hierarchical classification model to determine the aging degree classification result; Based on the aging degree classification results, a visualization algorithm is used to generate graded distribution data; The hierarchical distribution data is spatially mapped to generate high-resolution aging degree distribution data, resulting in an optimized distribution map. If the optimized distribution map meets the preset clarity threshold, then image smoothing technology is used for processing to obtain the aging degree classification distribution map.

7. The automatic detection method for building curtain walls according to claim 1, characterized in that, The process of determining the priority maintenance areas and maintenance methods for the priority maintenance areas in the building curtain wall includes the following steps: Based on the areas of damaged light transmittance in the aging grade distribution map of the curtain wall, a light transmission model is used, combined with the spectral characteristics of the areas of damaged light transmittance, to determine the degree of light transmittance reduction. Based on the degree of light transmittance reduction, a convolutional neural network is used to perform spatial pattern analysis and generate a light transmittance distribution map. Based on the light transmittance distribution map and the aging degree classification distribution map, a decision tree algorithm is used to determine the priority maintenance areas and maintenance methods.

8. An automatic inspection system for building curtain walls, characterized in that, include: The acquisition module is used to acquire high-resolution multispectral image data of the curtain wall surface covering the visible to near-infrared bands, and generate a first multispectral image containing the spectral characteristics of the deposits based on the high-resolution multispectral image data. The classification map generation module is used to classify the spectral features of the first multispectral image based on the first multispectral image using a support vector machine algorithm, and generate a chemical composition classification map, wherein the chemical composition classification map includes the distribution area of ​​corrosion products and dirt types on the curtain wall surface. Image segmentation and generation: Based on the chemical composition classification map, image segmentation technology is used to generate a segmented image containing the spatial distribution range of the deposits. The point cloud data generation module is used to acquire three-dimensional point cloud data of the deposits on the curtain wall surface based on the segmented image using a laser rangefinder, and obtain the first point cloud data. The distribution map generation module is used to combine the spatial information of the first point cloud data and the segmented image based on the first point cloud data and the segmented image using a feature fusion mechanism to determine the thickness distribution map corresponding to the thickness of the deposit. The graded distribution map generation module is used to extract spatial pattern features from the thickness distribution map using a convolutional neural network to generate a graded distribution map of aging degree. The aging degree classification distribution map is used to determine the priority maintenance areas in the building curtain wall and the maintenance treatment methods for the priority maintenance areas.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the automatic detection method for building curtain walls as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the automatic detection method for building curtain walls as described in any one of claims 1 to 7.

Images