Building quality risk prediction method and system

Abstract

The invention discloses a building quality risk prediction method and system, and relates to the technical field of image data processing. Acquiring an enhanced image of a target area in the target building, and performing feature extraction on the enhanced image to obtain deformation data and crack data; calculating an aging score according to the deformation data and the crack data, and performing aging judgment on the target building according to the aging score; if the aging score is greater than an aging threshold value, obtaining multi-modal data, calculating a quality score according to the multi-modal data, and carrying out building quality risk early warning according to the quality score; building aging is judged through image data, then multi-modal quality analysis is carried out, layered screening can be carried out, aging features are rapidly recognized through images, and only multi-modal detection is started for aged buildings so as to reduce cost and improve efficiency; the image is combined with multi-modal data, a multi-level evaluation system is constructed, and the accuracy is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image data processing, and in particular relates to a method and system for predicting building quality risks. Background Art

[0002] With the development of urbanization and infrastructure, building quality assessment has become crucial for ensuring safety and sustainable development. Traditional assessments rely on manual inspections, which are inefficient and susceptible to subjective factors. In recent years, the development of emerging technologies such as computer technology, image processing, machine learning, and the Internet of Things has provided new solutions for building quality assessment, enabling automated, efficient, and accurate evaluations. Research on the application of these technologies in building quality assessment has important theoretical and practical implications, with far-reaching implications for improving building quality, ensuring safety, and promoting the sustainable development of the industry.

[0003] Patent number: CN119443975A, discloses a construction quality prediction method and system based on data analysis; by collecting data from the construction site in real time, using the Gaussian process model and adjusting the kernel function parameters to provide deep insights into the complex data relationships on the construction site, thereby optimizing the accuracy and reliability of the model. The real-time adjustment of the model achieved by the incremental learning method ensures that the model and construction site data are updated synchronously, improving the real-time and adaptability of the prediction. In addition, through the use of confidence intervals, it provides a more comprehensive risk assessment for the construction team, realizes early warning of potential quality problems, and reduces structural safety risks and maintenance costs.

[0004] In the existing technology, building quality assessment is carried out by obtaining a large amount of building data. Although high-precision quality detection is achieved, it is easy to waste resources and increase the amount of data processing. Moreover, a single modality cannot take into account both macroscopic appearance and microscopic mechanism, resulting in a lack of accuracy in quality assessment. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of quality assessment of buildings by obtaining a large amount of building data. Although high-precision quality detection is achieved, it is easy to waste resources and increase the amount of data processing. Moreover, a single modality is difficult to take into account both macroscopic appearance and microscopic mechanism, resulting in a lack of accuracy in quality assessment. A building quality risk prediction method and system are proposed.

[0006] In a first aspect of the present invention, a method for predicting building quality risks is first proposed, the method comprising:

[0007] Acquire a reinforced image of a target area in a target building, and perform feature extraction on the reinforced image to obtain deformation data and crack data;

[0008] Calculating an aging score based on the deformation data and the crack data, and making an aging judgment on the target building based on the aging score;

[0009] If the aging score is greater than the aging threshold, multimodal data is obtained, a quality score is calculated based on the multimodal data, and a building quality risk warning is issued based on the quality score; the multimodal data includes: stress data, inclination data, vibration data and sound wave data.

[0010] Optionally, a positioning tag is deployed in the target area; before extracting features from the enhanced image to obtain deformation data and crack data, the method further includes:

[0011] Obtaining a reference image and image data of a target area, and performing similarity calculation on positioning tags in the image data based on the reference image to obtain target similarity;

[0012] If the target similarity of the positioning tag is less than a similarity threshold, feature extraction is performed on the positioning tag in the image data to obtain a binary mask and a region image, the region image and the binary mask are encoded into a feature vector, and the feature vector is input into an image restoration model to perform structure extraction to obtain a structure sketch;

[0013] Texture filling is performed on the structural sketch to obtain an initial repair image, the initial repair image is iteratively optimized to obtain a target repair image, and the target repair image is fused with the image data to obtain an enhanced image.

[0014] Optionally, feature extraction is performed on the image data to obtain deformation data and crack data, and extracting deformation data on the enhanced image includes:

[0015] Extracting coordinates of the enhanced image and the reference image to obtain enhanced coordinates and reference coordinates, respectively, and calculating a displacement difference based on the enhanced coordinates and the reference coordinates;

[0016] Acquire a scale image and a temperature change, calculate a scale factor based on the scale image, and calculate a temperature difference deformation based on the temperature change and the scale factor;

[0017] Deformation data is calculated based on the displacement difference and the temperature difference deformation amount.

[0018] Optionally, performing feature extraction on the image data to obtain deformation data and crack data, and extracting crack data from the reinforced image using a target YOLOv8 model includes:

[0019] Improvements to the target YOLOv8 model include:

[0020] The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model;

[0021] The working principle of the improved backbone network includes:

[0022] Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module is composed of a context enhancement module and an HG module, the second hierarchical module is composed of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module is composed of a DWConv module and an HG module.

[0023] Optionally, the improvement of the target Conv module includes:

[0024] Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module;

[0025] The target CIoU loss function includes:

[0026]

[0027] Among them, L MBCIoU Represents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

[0028] In a second aspect of the present invention, a building quality risk prediction system is proposed, comprising: a feature extraction module, an aging judgment module, and a risk assessment module:

[0029] The feature extraction module is used to obtain a reinforced image of a target area in a target building, and perform feature extraction on the reinforced image to obtain deformation data and crack data;

[0030] The aging judgment module is used to calculate an aging score based on the deformation data and the crack data, and perform aging judgment on the target building based on the aging score;

[0031] The risk assessment module is used to obtain multimodal data if the aging score is greater than the aging threshold, calculate a quality score based on the multimodal data, and issue a building quality risk warning based on the quality score; the multimodal data includes: stress data, inclination data, vibration data and sound wave data.

[0032] Optionally, the system further includes: a similarity judgment module, a structure extraction module and a fusion enhancement module:

[0033] The similarity judgment module is used to obtain a reference image and image data of a target area, and perform similarity calculation on the positioning tags in the image data according to the reference image to obtain target similarity;

[0034] The structure extraction module is configured to, if the target similarity of the positioning tags is less than a similarity threshold, perform feature extraction on the positioning tags in the image data to obtain a binary mask and a region image, encode the region image and the binary mask into a feature vector, and input the feature vector into an image restoration model to perform structure extraction to obtain a structure sketch;

[0035] The fusion and enhancement module is used to perform texture filling on the structural sketch to obtain an initial repair image, iteratively optimize the initial repair image to obtain a target repair image, and fuse the target repair image with the image data to obtain an enhanced image.

[0036] Optionally, the fusion enhancement module includes: a displacement calculation module, a deformation calculation and deformation data acquisition module:

[0037] The displacement calculation module is configured to extract coordinates of the enhanced image and the reference image to obtain enhanced coordinates and reference coordinates, respectively, and calculate a displacement difference based on the enhanced coordinates and the reference coordinates;

[0038] The deformation variable calculation module is used to obtain a scale image and a temperature change, calculate a scale factor based on the scale image, and calculate a temperature difference deformation variable based on the temperature change and the scale factor;

[0039] The deformation data acquisition module is used to calculate the deformation data according to the displacement difference and the temperature difference deformation amount.

[0040] Optionally, the target YOLOv8 model is improved by:

[0041] The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model;

[0042] The working principle of the improved backbone network includes:

[0043] Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module is composed of a context enhancement module and an HG module, the second hierarchical module is composed of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module is composed of a DWConv module and an HG module.

[0044] Optionally, the improvement of the target Conv module includes:

[0045] Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module;

[0046] The target CIoU loss function includes:

[0047]

[0048] Among them, L MBCIoU Represents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

[0049] Beneficial effects of the present invention:

[0050] This paper proposes a building quality risk prediction method. This method acquires image data of a target area within a target building, extracts features from the image data to obtain deformation and crack data, then determines the aging of the target building based on the deformation and crack data. If the target building is determined to be aged, a quality inspection is performed to obtain multimodal data. A quality score is calculated based on the multimodal data and input into a risk prediction model to obtain a risk assessment result. Building quality risk warnings are then issued based on the risk assessment results. Image data is used to determine building aging, followed by multimodal quality analysis, enabling hierarchical screening. Aging features are quickly identified based on images, and multimodal inspection is initiated only for aged buildings to reduce costs and improve efficiency. Combining images with multimodal data creates a multi-level assessment system to enhance accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described below with reference to the accompanying drawings.

[0052] Figure 1 A flowchart of a building quality risk prediction method provided by an embodiment of the present invention;

[0053] Figure 2 A model structure diagram of a building quality risk prediction method provided by an embodiment of the present invention;

[0054] Figure 3 A framework diagram of another building quality risk prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments represent only a portion of the embodiments of the present invention, not all of them. The term "and / or" herein simply describes an association relationship between associated objects, indicating that three possible relationships exist. For example, "A" and "B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, references to "first," "second," and so on in the present invention are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one of these features. Furthermore, the technical solutions of the various embodiments may be combined, but only if they are achievable by a person of ordinary skill in the art. If a combination of technical solutions contradicts or is unachievable, such combination shall be deemed non-existent and outside the scope of protection claimed by the present invention.

[0056] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0057] The embodiment of the present invention provides a method for predicting building quality risks. Figure 1 , Figure 1 A flowchart of a method for predicting building quality risks provided by an embodiment of the present invention. The method includes the following steps:

[0058] S101, obtaining a reinforced image of a target area in a target building, and performing feature extraction on the reinforced image to obtain deformation data and crack data;

[0059] S102, calculating an aging score based on the deformation data and the crack data, and performing an aging judgment on the target building based on the aging score;

[0060] S103: If the aging score is greater than the aging threshold, multimodal data is obtained, a quality score is calculated based on the multimodal data, and a building quality risk warning is issued based on the quality score;

[0061] Multimodal data includes stress data, inclination data, vibration data and acoustic wave data.

[0062] A building quality risk prediction method provided by an embodiment of the present invention determines building aging through image data, and then performs multimodal quality analysis, which can be layered for screening: aging characteristics are quickly identified based on images, and multimodal detection is only initiated for aging buildings to reduce costs and improve efficiency; images are combined with multimodal data to construct a multi-level assessment system to enhance accuracy.

[0063] In one implementation, when image data comes from inside the building (for example, load-bearing beams and column surfaces; cameras are installed at the above-mentioned locations), a phased assessment is performed (first, images are initially screened for signs of aging, and then in-depth verification is combined with multimodal data) to achieve efficient data acquisition and building assessment; internal images can capture early deformation and cracks in key load-bearing components and quickly identify high-risk areas; multimodal sensors are then deployed in these areas, and cross-validation is performed through material performance degradation and vibration frequency attenuation data to avoid misjudgment of single image data due to viewing angle limitations or light interference.

[0064] In one implementation, multimodal data, such as fiber Bragg grating (FBG) sensors, are installed at the mid-span and column bases of load-bearing beams to monitor concrete strain (με). Inclinometers are deployed at high-rise corner columns or settlement-sensitive areas to measure tilt angles (°) and settlement amounts (mm). Acoustic emission sensors are deployed at nodes or cantilever roots to capture the energy (dB) and frequency of cracking / fracture events. Accelerometers are used to detect vibration acceleration.

[0065] In one implementation, an aging score is calculated based on deformation data and crack data. Aging score = α * deformation data + β * crack data, where deformation data is the mean of the deformation amount of each monitoring point, and crack data is the mean of the crack area ratio of each monitoring point.

[0066] In one implementation, the aging of the target building is judged based on deformation data and crack data. When the deformation value (deformation data) exceeds the threshold allowed by the building design code, it indicates that the structural stability has decreased, and if the area ratio of the crack data exceeds the threshold, the building is judged to have entered an aging state and the quality inspection process is triggered. The quality score is calculated based on the multimodal data. The quality score obtained by weighted summation of the change in the multimodal data per unit time (normalized and dimensionless) is used to measure the quality of the building.

[0067] In one implementation method, building quality risk warnings are carried out based on quality scores. For example, for normal (90-100 points): system response: maintain regular monitoring frequency (such as once a week), generate health reports for archiving; medium risk (60-89 points): system response: increase monitoring frequency to once a day, automatically generate maintenance work orders and push them to the cloud; high risk (0-59 points): system response: real-time monitoring (updated every minute), trigger sound and light alarms and contact emergency management departments; dynamic strategy: enable drone inspections in high-risk areas 24 hours a day, mark risk components in red through the BIM model, and simultaneously push them to the operation and maintenance APP.

[0068] In one implementation method, aging judgment is made through image data and then multimodal quality analysis is carried out, which can achieve the technical effects of layered screening and precise detection: image data is used to quickly identify intuitive aging characteristics such as deformation and cracks, and multimodal detection is only initiated for buildings judged to be aged, avoiding redundant detection to reduce costs; the scope of quality inspection is narrowed through image analysis, making multimodal data collection more targeted and improving overall process efficiency; image data provides evidence of macroscopic structural aging, and multimodal data supplements microscopic mechanism information. The two are combined to construct a multi-level assessment system from appearance to essence, enhancing the accuracy and comprehensiveness of risk assessment; the aging area is first determined at the macro level through image data, and then quality analysis is performed by obtaining multimodal data of the aging area, thereby reducing the amount of calculation and improving resource utilization efficiency.

[0069] In one embodiment, a positioning tag is deployed in the target area; before extracting features from the enhanced image to obtain deformation data and crack data, the method further includes:

[0070] Obtain a reference image and image data of the target area, and calculate the similarity of the positioning tags in the image data based on the reference image to obtain the target similarity;

[0071] If the target similarity of the positioning label is less than the similarity threshold, the positioning label in the image data is feature extracted to obtain a binary mask and a region image. The region image and the binary mask are encoded into a feature vector, and the feature vector is input into the image restoration model for structure extraction to obtain a structure sketch.

[0072] Texture filling is performed on the structural sketch to obtain an initial repair image, which is then iteratively optimized to obtain a target repair image, which is then fused with the image data to obtain an enhanced image.

[0073] In one implementation, a target area (e.g., a heavy column) is equipped with a positioning tag (a blueprint used for positioning, such as ⊕). Pixel-level comparison with a reference image (e.g., hash value matching or feature point matching) quantifies the integrity of the positioning tag, preventing misjudgments due to lighting changes or device jitter. For example, when a pipeline partially obscures a positioning tag, similarity calculations can quickly identify missing pixels in the marked area (e.g., similarity drops from 100% to 60%). Similarity calculations, such as cosine similarity, convert images to vector form and use cosine similarity to calculate their similarity, can also be used for Euclidean distance, Manhattan distance, and so on). Using similarity thresholds can further reduce false detection rates (e.g., eliminating fluctuations in area ratios caused by dust cover).

[0074] In one implementation, a binary mask (e.g., the 0-1 matrix output by U²-net, where 1 = labeled area and 0 = background) can precisely separate the location label from the background, defining a clear region for subsequent repair. Only the marked portion with a mask value of 1 is repaired, preventing the repair algorithm from accidentally touching background areas (e.g., cracks on the concrete surface). Regional image extraction reduces computational effort and improves repair efficiency.

[0075] In one implementation, the Transformer encoder in the image restoration model extracts global structural features, and based on the mask position encoding, focuses on restoring the geometric structure of the occluded area (such as the position of the blocked circle center), and finally outputs a structural sketch; the image restoration model, for example: an incremental image restoration model based on Transformer structure enhancement.

[0076] In one implementation, iterative optimization can gradually correct the detail errors in the initial restoration (gradually synthesizing details that are consistent with the surrounding area, generating a residual map at each iteration and superimposing it on the previous version of the result). Through the backpropagation of the loss function, the restoration result is made closer to the reference image in terms of pixel value, structure, and semantics.

[0077] In one embodiment, feature extraction is performed based on image data to obtain deformation data and crack data. Extracting deformation data from the enhanced image includes:

[0078] Coordinates of the enhanced image and the reference image are extracted to obtain enhanced coordinates and reference coordinates respectively, and displacement difference is calculated based on the enhanced coordinates and the reference coordinates;

[0079] Obtain a scale image and a temperature change, calculate a scale factor based on the scale image, and calculate a temperature difference deformation based on the temperature change and the scale factor;

[0080] The deformation data is calculated based on the displacement difference and temperature difference deformation.

[0081] In one implementation, the displacement difference (pixel displacement) is calculated based on the enhanced coordinates and the reference coordinates. , ,in, Represents the horizontal pixel displacement difference of the target at time ti, represents the vertical pixel displacement difference of the target at time ti, m represents the total number of detection points, p represents the target number, Represents the horizontal coordinate of the center of mass of the pth target at time ti, represents the center of mass coordinates of the pth target at the initial time t0, Represents the vertical coordinate of the center of mass of the pth target at time ti, Represents the centroid coordinates of the pth target at the initial time t0. By calculating the average of the centroid coordinates of multiple targets, the deviation caused by occlusion or detection error of a single target is reduced, ensuring the robustness of the displacement calculation. p represents the target number, which is also the detection point number.

[0082] In one implementation, the scale image is an image taken of a scale ruler mounted next to the positioning tag (a ruler of known length), and there is no obstruction between the two. The scale factor is calculated based on the scale image. Since the length of the scale ruler is known, the pixel length of the scale image (the pixel coordinates at both ends of the scale ruler) is obtained through the scale image. The scale factor = actual length of the scale ruler / pixel length of the scale ruler. For example: if the actual length of the ruler is 1000 mm and it occupies 500 pixels in the image, the scale factor = 1000 / 500 = 2 mm / px.

[0083] In one implementation, the temperature difference deformation is calculated based on the temperature change and the proportional factor, WC=BL*ΔT, where WC represents the temperature difference deformation, BL represents the proportional factor, and ΔT represents the temperature change; the deformation data is calculated based on the displacement difference and the temperature difference deformation. , where XB represents the deformation data (the deformation of the building in the vertical direction).

[0084] In one implementation, building materials (such as concrete and steel) age. For example, concrete shrinks and deforms, while steel rusts due to long-term exposure to the environment. This reduces the strength and elastic modulus of the materials, making aging buildings more susceptible to deformation when exposed to temperature fluctuations.

[0085] In one implementation, by extracting the target center of mass coordinates in the enhanced image and the reference image and calculating their average displacement difference, the deviation caused by occlusion or detection error is reduced, the accuracy of displacement calculation and the robustness of the system are improved, and the scale factor and temperature difference deformation are calculated using the scale image and temperature change. The deformation caused by temperature change is compensated, further improving the accuracy of deformation monitoring. By combining the displacement difference and temperature difference deformation, the deformation data of the building is calculated, reflecting the deformation of the building.

[0086] In one embodiment, feature extraction is performed based on image data to obtain deformation data and crack data, and crack data extraction is performed on the reinforced image using a target YOLOv8 model, including:

[0087] Improvements to the target YOLOv8 model include:

[0088] The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model;

[0089] The working principles of the improved backbone network include:

[0090] Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module consists of a context enhancement module and an HG module, the second hierarchical module consists of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module consists of a DWConv module and an HG module.

[0091] In one implementation, see Figure 2 , Figure 2A model structure diagram of a building quality risk prediction method provided by an embodiment of the present invention; initial feature extraction modules such as: HGStem, DWConv, ResNet module, etc.; the first hierarchical module consists of a context enhancement module and an HG module, and the context enhancement module and the HG module are connected in sequence; the second hierarchical module consists of a DWConv module, an HG module, an HG module, and an HG module, and the DWConv module, HG module, HG module, and HG module are connected in sequence; the third hierarchical module consists of a DWConv module and an HG module, and the DWConv module and the HG module are connected in sequence.

[0092] In one implementation, the backbone network in the YOLOv8 model is replaced with an improved backbone network. The improved backbone network can more effectively extract feature information in the image through a combination of multi-level feature extraction and hierarchical modules. The working principle of the improved backbone network includes multiple stages of feature extraction and enhancement: initial feature extraction module (such as HGStem): can quickly extract the initial features of the image, providing a basis for subsequent feature enhancement; multi-level hierarchical module: through the combination of context enhancement module and multiple HG modules, the feature expression ability is gradually enhanced. The hierarchical design enables the model to capture feature information at different levels, thereby improving the detection accuracy of the target; output of the SPFF module: the final output feature map has stronger feature expression ability and can better support subsequent detection tasks; the improved backbone network design can significantly improve the model's detection ability for complex scenes and small targets, and improve detection accuracy and robustness.

[0093] In one implementation, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module. Traditional convolution operations have two problems: 1. Traditional convolution is limited to the local receptive field, cannot effectively obtain long-range information, and has a fixed sampling architecture; 2. The square convolution kernel and fixed sampling configuration have poor adaptability to dynamically changing targets. Therefore, the Conv module needs to be improved to achieve high model accuracy while improving operational efficiency.

[0094] In one embodiment, the improvement of the target Conv module includes:

[0095] Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module;

[0096] The target CIoU loss function includes:

[0097]

[0098] Among them, L MBCIoURepresents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

[0099] In one implementation, the Mish function: Mish(x)=x⋅tanh(ln(1+e x Compared to SiLU, Mish has smoother curvature and non-monotonicity, providing a more stable gradient flow during backpropagation. This replacement accelerates model convergence, reduces training loss, and avoids the vanishing gradient problem. Improved microcrack feature extraction: Mish has a higher saturation threshold, retaining more negative feature information (such as dark pixels at crack edges), enhancing sensitivity to low-contrast cracks.

[0100] In one implementation, SENet dynamically recalibrates channel weights through a squeeze-excitation mechanism: Squeeze: Global average pooling compresses spatial dimensions (H×W×C→1×1×C) and aggregates global crack features; Excitation: The fully connected layer learns the nonlinear relationship between channels (W∈R C×C ), strengthen key crack channels (such as the main direction of the crack), and suppress background noise (such as wall texture); Scale: weighted output feature map to improve the response of the crack area.

[0101] In one implementation, the accuracy of bounding box regression in the target CIoU loss function is improved by introducing a penalty for box size differences in the loss function. This improvement can better handle errors between boxes of different sizes, overcoming the limitations of traditional IoU in this context. By optimizing the calculation process, the target CIoU loss function can accelerate model convergence while improving regression accuracy. In terms of gradient calculation, the target CIoU loss function introduces a balance parameter to make the relationship between IoU and perimeter difference adjustable, further optimizing the loss function calculation process and improving the model's stability and performance in various scenarios.

[0102] In one implementation, Represents the IoU value, which ranges from [0,1]. The larger the value, the higher the overlap between the two boxes. + Represents the sum of the squares of the upper left corner coordinate differences, measuring the distance between the predicted box and the upper left corner of the real box; + Represents the sum of the squares of the coordinate differences of the lower right corner, which measures the distance between the predicted box and the lower right corner of the true box; Represents the normalization factor, which normalizes the boundary distance penalty term to eliminate the influence of the true box size on the loss, so that the model maintains consistent sensitivity to cracks of different sizes (such as microcracks and large cracks).

[0103] Based on the same inventive concept, the present invention also provides a building quality risk prediction system. Figure 3 , Figure 3 A framework diagram of a building quality risk prediction system provided by an embodiment of the present invention includes: a feature extraction module, an aging judgment module, and a risk assessment module:

[0104] A feature extraction module is used to obtain a reinforced image of a target area in a target building and perform feature extraction on the reinforced image to obtain deformation data and crack data;

[0105] An aging judgment module is used to calculate an aging score based on deformation data and crack data, and to make an aging judgment on the target building based on the aging score;

[0106] The risk assessment module is used to obtain multimodal data if the aging score is greater than the aging threshold, calculate the quality score based on the multimodal data, and issue a building quality risk warning based on the quality score; the multimodal data includes: stress data, inclination data, vibration data and sound wave data.

[0107] A building quality risk prediction system provided by an embodiment of the present invention determines building aging through image data and then performs multimodal quality analysis, which can be screened in layers: aging characteristics are quickly identified based on images, and multimodal detection is only initiated for aging buildings to reduce costs and improve efficiency; images are combined with multimodal data to construct a multi-level assessment system to enhance accuracy.

[0108] In one embodiment, the system further includes: a similarity judgment module, a structure extraction module, and a fusion enhancement module:

[0109] A similarity judgment module is used to obtain a reference image and image data of a target area, and perform similarity calculation on the positioning tags in the image data based on the reference image to obtain target similarity;

[0110] A structure extraction module is used to extract features from the positioning tags in the image data to obtain a binary mask and a region image if the target similarity of the positioning tags is less than a similarity threshold, encode the region image and the binary mask into a feature vector, and input the feature vector into the image restoration model to perform structure extraction to obtain a structure sketch;

[0111] The fusion and enhancement module is used to perform texture filling on the structural sketch to obtain an initial repair image, iteratively optimize the initial repair image to obtain a target repair image, and fuse the target repair image with the image data to obtain an enhanced image.

[0112] In one embodiment, the fusion enhancement module includes: a displacement calculation module, a deformation calculation and deformation data acquisition module:

[0113] A displacement calculation module is used to extract coordinates of the enhanced image and the reference image to obtain enhanced coordinates and reference coordinates respectively, and calculate the displacement difference based on the enhanced coordinates and the reference coordinates;

[0114] The deformation variable calculation module is used to obtain the scale image and the temperature change, calculate the scale factor according to the scale image, and calculate the temperature difference deformation variable according to the temperature change and the scale factor;

[0115] The deformation data acquisition module is used to calculate the deformation data based on the displacement difference and the temperature difference deformation.

[0116] In one embodiment, improvements to the target YOLOv8 model include:

[0117] The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model;

[0118] The working principles of the improved backbone network include:

[0119] Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module consists of a context enhancement module and an HG module, the second hierarchical module consists of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module consists of a DWConv module and an HG module.

[0120] In one embodiment, the improvement of the target Conv module includes:

[0121] Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module;

[0122] The target CIoU loss function includes:

[0123]

[0124] Among them, L MBCIoU Represents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

[0125] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for predicting building quality risks, characterized in that: The method comprises: Acquire a reinforced image of a target area in a target building, and perform feature extraction on the reinforced image to obtain deformation data and crack data; Calculating an aging score based on the deformation data and the crack data, and making an aging judgment on the target building based on the aging score; If the aging score is greater than the aging threshold, multimodal data is obtained, a quality score is calculated based on the multimodal data, and a building quality risk warning is issued based on the quality score; the multimodal data includes: stress data, inclination data, vibration data and sound wave data.

2. A building quality risk prediction method according to claim 1, characterized in that: The target area is deployed with a positioning tag; Before extracting features from the enhanced image to obtain deformation data and crack data, the method further includes: Obtaining a reference image and image data of a target area, and performing similarity calculation on positioning tags in the image data based on the reference image to obtain target similarity; If the target similarity of the positioning tag is less than a similarity threshold, feature extraction is performed on the positioning tag in the image data to obtain a binary mask and a region image, the region image and the binary mask are encoded into a feature vector, and the feature vector is input into an image restoration model to perform structure extraction to obtain a structure sketch; Texture filling is performed on the structural sketch to obtain an initial repair image, the initial repair image is iteratively optimized to obtain a target repair image, and the target repair image is fused with the image data to obtain an enhanced image.

3. A building quality risk prediction method according to claim 2, characterized in that: Extracting features from the image data to obtain deformation data and crack data, extracting deformation data from the enhanced image includes: Extracting coordinates of the enhanced image and the reference image to obtain enhanced coordinates and reference coordinates, respectively, and calculating a displacement difference based on the enhanced coordinates and the reference coordinates; Acquire a scale image and a temperature change, calculate a scale factor based on the scale image, and calculate a temperature difference deformation based on the temperature change and the scale factor; Deformation data is calculated based on the displacement difference and the temperature difference deformation amount.

4. A building quality risk prediction method according to claim 2, characterized in that: Extracting features from the image data to obtain deformation data and crack data, and extracting crack data from the enhanced image using a target YOLOv8 model includes: Improvements to the target YOLOv8 model include: The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model; The working principle of the improved backbone network includes: Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module is composed of a context enhancement module and an HG module, the second hierarchical module is composed of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module is composed of a DWConv module and an HG module.

5. A building quality risk prediction method according to claim 4, characterized in that: The improvements of the target Conv module include: Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module; The target CIoU loss function includes: ； Among them, L MBCIoU Represents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

6. A building quality risk prediction system, characterized in that: The system includes: a feature extraction module, an aging judgment module and a risk assessment module: The feature extraction module is used to obtain a reinforced image of a target area in a target building, and perform feature extraction on the reinforced image to obtain deformation data and crack data; The aging judgment module is used to calculate an aging score based on the deformation data and the crack data, and perform aging judgment on the target building based on the aging score; The risk assessment module is used to obtain multimodal data if the aging score is greater than the aging threshold, calculate a quality score based on the multimodal data, and issue a building quality risk warning based on the quality score; the multimodal data includes: stress data, inclination data, vibration data and sound wave data.

7. A building quality risk prediction system according to claim 6, characterized in that: The system also includes: a similarity judgment module, a structure extraction module and a fusion enhancement module: The similarity judgment module is used to obtain a reference image and image data of a target area, and perform similarity calculation on the positioning tags in the image data according to the reference image to obtain target similarity; The structure extraction module is configured to, if the target similarity of the positioning tags is less than a similarity threshold, perform feature extraction on the positioning tags in the image data to obtain a binary mask and a region image, encode the region image and the binary mask into a feature vector, and input the feature vector into an image restoration model to perform structure extraction to obtain a structure sketch; The fusion and enhancement module is used to perform texture filling on the structural sketch to obtain an initial repair image, iteratively optimize the initial repair image to obtain a target repair image, and fuse the target repair image with the image data to obtain an enhanced image.

8. A building quality risk prediction system according to claim 7, characterized in that: The fusion enhancement module includes: a displacement calculation module, a deformation variable calculation and deformation data acquisition module: The displacement calculation module is configured to extract coordinates of the enhanced image and the reference image to obtain enhanced coordinates and reference coordinates, respectively, and calculate a displacement difference based on the enhanced coordinates and the reference coordinates; The deformation variable calculation module is used to obtain a scale image and a temperature change, calculate a scale factor based on the scale image, and calculate a temperature difference deformation variable based on the temperature change and the scale factor; The deformation data acquisition module is used to calculate the deformation data according to the displacement difference and the temperature difference deformation amount.

9. A building quality risk prediction system according to claim 7, characterized in that: Improvements to the target YOLOv8 model include: The backbone network in the YOLOv8 model is replaced with the improved backbone network, the Conv module in the neck structure of the YOLOv8 model is replaced with the target Conv module, and the CIoU loss function in the detection head of the YOLOv8 model is replaced with the target CIoU loss function to obtain the improved YOLOv8 model; The working principle of the improved backbone network includes: Obtain an initial feature map, input the initial feature map into the initial feature extraction module to obtain a low-level feature map, input the low-level feature map into the HG module to obtain a first feature map, input the first feature map into the first hierarchical module to obtain a second feature map, input the second feature map into the second hierarchical module to obtain a third feature map, input the third feature map into the third hierarchical module to obtain a fourth feature map, and use the fourth feature map as the output of the SPFF module; the first hierarchical module is composed of a context enhancement module and an HG module, the second hierarchical module is composed of a DWConv module, an HG module, an HG module and an HG module, and the third hierarchical module is composed of a DWConv module and an HG module.

10. A building quality risk prediction system according to claim 9, characterized in that: The improvements of the target Conv module include: Replace the SiLU function in the AKConv module with the Mish function, and add the SENet module after the Mish function in the AKConv module to obtain the target Conv module; The target CIoU loss function includes: ； Among them, L MBCIoU Represents the target CIoU loss function, A represents the predicted bounding box, B represents the real bounding box, A∩B represents the intersection area of ​​the predicted box and the real box, and A∪B represents the union area of ​​the predicted box and the real box. represents the IoU value, + represents the sum of the squares of the upper left corner coordinate differences, + represents the sum of the squares of the coordinate differences of the lower right corner, represents the normalization factor.

Images