Intelligent building process three-dimensional modeling and quality detection analysis method

Abstract

The invention relates to a three-dimensional modeling and quality detection analysis method for an intelligent construction process. The method comprises the following steps: S1, preprocessing multi-source visual sensing data; s2, extracting multi-scale depth representation; s3, constructing a multi-scale multi-view fusion representation; s4, constructing a 3D decoding module based on the multi-scale multi-view fusion representation, and carrying out 3D CNN decoding to obtain a three-dimensional reconstruction result; s5, constructing a building defect detection network to obtain a building defect detection result; s6, constructing a dynamic deviation prediction network to obtain a dynamic deviation prediction result; and S7, in combination with the building defect detection result and the dynamic deviation prediction result, outputting a feedback result for building defect detection and deviation prediction. The system can monitor the occurrence of deviation in real time in the construction process, and can carry out prediction and feedback output in time. According to the method, advanced early warning can be provided for construction management, so that the construction deviation can be identified and corrected in an early stage.

Description

Technical Field

[0001] The present invention relates to a method for three-dimensional modeling and quality inspection and analysis of an intelligent construction process, and in particular to a method for three-dimensional modeling and quality inspection and analysis of a multi-visual sensing deep fusion perception. Background Art

[0002] As one of the key technologies to promote the digital transformation of the construction industry, intelligent construction is gaining increasing attention. With the increasing complexity and scale of construction projects, traditional construction methods and quality control methods can no longer meet the needs of modern construction. Real-time monitoring, quality inspection and progress tracking of construction sites are important links to ensure the smooth implementation of construction projects. However, existing technologies often rely on manual inspection or a single sensor, which makes it difficult to effectively respond to dynamic changes and quality risks in the construction process. Especially in terms of high-precision three-dimensional modeling, although technologies such as LiDAR can provide higher measurement accuracy, their high equipment cost and cumbersome operation limit their popularization and application in small and medium-sized projects. Therefore, the research on low-cost, high-efficiency intelligent construction technology that can achieve accurate three-dimensional modeling and comprehensive quality monitoring has become an important issue that needs to be solved in the construction industry.

[0003] Existing technologies, such as the Chinese invention patent application with publication number CN117150627A, which proposes a warehouse construction method based on 3D modeling digital twins, and the Chinese invention patent with publication number CN117291086A, which proposes a digital twin intelligent construction method based on extra-long tunnel TBM engineering, provide certain solutions. However, the above technologies still have the following limitations: The balance between 3D modeling accuracy and cost: Most existing architectural 3D modeling technologies rely on LiDAR for measurement. Although it has high accuracy, it faces high equipment costs and complex operating requirements. This makes it difficult for LiDAR technology to be widely used in small or low-budget projects, and due to the limitations of the fixed equipment and scanning angles, it may not be able to fully cover the complex structures and dynamic changes of the construction site. In contrast, 3D modeling methods based on visual sensors (such as cameras and depth cameras) can provide more flexible and low-cost solutions, but they still face challenges in terms of accuracy and robustness, especially in modeling effects under lighting changes or occlusion.

[0004] Insufficient accuracy and adaptability of building defect detection: Traditional building defect detection methods rely on manual inspection or traditional sensor technology, which have problems such as low efficiency, missed detection, and false detection. Although some automated detection methods based on computer vision have been applied in recent years, these methods still have great limitations when dealing with multi-scale and multi-view building defects, especially in complex construction environments, such as occlusion of construction materials, changes in viewing angles, or lighting problems. These factors can significantly affect the accuracy and reliability of defect detection. In addition, when dealing with complex defects on large construction sites, existing methods usually rely on fixed viewing angles or single sensor data, making it difficult to fully obtain all quality information during the construction process.

[0005] Lack of dynamic deviation prediction and early warning mechanism: Current construction process monitoring mainly focuses on static quality inspection and progress tracking, while relatively little attention is paid to dynamic deviation prediction and real-time feedback mechanisms during the construction process. Existing technologies make it difficult to combine real-time 3D modeling data and defect detection results to perform multi-dimensional deviation prediction. Due to the large number of variables in the construction process, traditional methods rely more on historical data or post-review, and lack an immediate early warning system to deal with quality deviations or safety hazards that may occur during construction. Existing technologies are difficult to provide real-time and accurate deviation feedback, especially in complex construction environments, where it is difficult to achieve advance warning and immediate adjustment. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a method for 3D modeling and quality inspection and analysis of intelligent construction process, which comprises the following steps: S1. Preprocessing of multi-source visual sensor data: standardizing and denoising the multi-source visual sensor data, and uniformly adjusting the spatial resolution and brightness; S2, using Swin Transformer to extract multi-scale deep representation of multi-visual sensor data after S1 preprocessing; S3, based on the multi-scale depth representation of multi-vision obtained in S2, is input into the multi-view fusion encoder for deep adaptive fusion encoding, and a multi-scale multi-view fusion representation is constructed at the same time, which includes the following two steps: S31, multi-view fusion coding, through a multi-view fusion encoder composed of two layers of Transformer Encoder, deeply encodes the multi-visual representation of each scale output by S2, captures the semantic interaction and correlation between each view, and generates enhanced multi-view and multi-scale feature representation; S32, multi-scale representation fusion, concatenates the input representation and the encoding representation in the first step, and performs feature fusion through pooling operation, integrating multi-scale and multi-view information to generate the final multi-scale and multi-view fusion representation, providing high-quality feature input for subsequent defect detection and dynamic deviation prediction; S4, based on the multi-scale and multi-view coding representation obtained in S3, construct a 3D decoding module, perform 3D CNN decoding, and input the constructed 3D feature fusion network to obtain a 3D reconstruction result; S5, constructing a building defect detection network and inputting the multi-scale and multi-view fusion deep representation obtained in S3 to obtain the building defect detection result; S6, construct a dynamic deviation prediction network and input the multi-scale and multi-view fusion deep representation obtained in S3 to obtain a dynamic deviation prediction result; S7, combining the building defect detection results obtained in S5 and the dynamic deviation prediction results in S6, outputs feedback results of defect detection and deviation prediction of buildings in the intelligent construction process.

[0007] Based on the above technical solution, step S1 includes the following steps: S10, format standardization, converting the RGB image data collected by N multi-source visual sensors into a unified standard tensor format; S11, pixel value normalization, using a linear normalization method to adjust the pixel value range of the image data to 0, 1, thereby reducing the interference of the numerical scale on the subsequent depth representation extraction; S12, denoising, using classic image denoising algorithms, Gaussian filtering and bilateral filtering, to smooth small-scale noise in the image while retaining edge information, using Gaussian kernel function to perform weighted averaging of image pixel values ​​to smooth random noise, and smoothing in both the spatial domain and the pixel intensity domain to retain edge clarity; S13, unified resolution, for the different resolutions of images collected by multi-source sensors, a bilinear interpolation algorithm is used to scale all images to a unified size of 768×768 pixels; S14, brightness adjustment and contrast enhancement, performing histogram equalization on the pre-processed image data, redistributing pixel values ​​to make the image brightness distribution more uniform, and enhancing the texture details and defect features of the building surface; S15. Data quality improvement and storage. After preprocessing, the multi-source visual data is stored in the form of standardized tensors to form a unified input data set, and the relevant preprocessing parameters are recorded to support the calling and reproduction of subsequent modules.

[0008] Based on the above technical solution, the Swin Transformer block and window attention mechanism in step S2 divides the image into a fixed-size window of 96×96, and calculates local attention inside the window to extract local features; multi-scale feature extraction extracts key depth representation information of buildings at different scales through a hierarchical feature extraction structure; feature integration realizes information interaction between different windows through alternating window movement operations, captures long-range dependencies and complex geometric information of building images, and converts input data into Indicates that Indicates The image data of the visual sensor, the corresponding multi-scale depth representation extracted by Swin Transformer is recorded as ,here Indicates The scale features of the layer are 4 layers, so as to achieve the purpose of multi-scale representation expression; the whole S2 multi-scale feature extraction process is expressed as: ,in represents Swin Transformer, Represents the serial number of the vision sensor.

[0009] Based on the above technical solution, step S31 specifically uses Transformer Encoder to extract deep semantic correlation across perspectives for multi-visual representation of each scale, strengthen the global information of feature representation, and the input representation of each sensor After two layers of Transformer Encoder, the integrity and complementarity of representations from different perspectives are improved. Transformer Encoder includes self-attention mechanism, cross-attention mechanism and multi-layer perceptron module. First, a nonlinear transformation is performed through a multi-layer perceptron to improve the complexity and adaptability of the representation, which can be expressed as:

[0010] After each The self-attention mechanism is used to analyze the relationship between local and global features within the same perspective, and the internal feature expression of each perspective is optimized to obtain :

[0011] Finally, for all Perform cross attention mechanism to obtain corresponding representation :

[0012] After the Transformer Encoder, each scale of the multi-visual representation will generate a corresponding deep encoding representation. This operation maps the multi-view features to a higher-dimensional semantic space through layer-by-layer semantic interactions, and finally obtains a multi-scale multi-view encoding representation. : .

[0013] Based on the above technical solution, the step S32 specifically comprises firstly performing splicing and average pooling operations on multi-view data of different layers to obtain multi-view fusion features, and then performing splicing and average pooling operations on the multi-view fusion features again to obtain multi-scale and multi-view fusion features. For each scale, the original representation and the representation after multi-view fusion encoding are uniformly spliced ​​in the feature dimension. This process is expressed as:

[0014] here Represents the The view sensor data is stored in the multi-view fusion encoder. The feature representation corresponding to the layer, where 1 is the input multi-view fusion encoder as the first layer, the output of the first layer Transformer Encoder in the multi-view fusion encoder as the second layer, and the final output of the multi-view fusion encoder as the third layer, so there are 3 layers in total; Then the final fusion representation is obtained through average pooling:

[0015] in represents the average pooling operation, It is the final multi-scale and multi-view fusion representation.

[0016] Based on the above technical solution, step S4 includes 3D CNN decoding and 3D feature fusion to complete 3D reconstruction. The purpose of 3D CNN decoding is to process the multi-scale multi-view encoding representation through a series of convolution and attention mechanisms, gradually restore and extract the structural features in the three-dimensional space. Finally, through this process, a three-dimensional feature map with spatial semantic information is obtained, and the multi-scale multi-view representation from the S3 stage is input. , which is a deep feature map containing multi-view and multi-scale information. The input is passed to the 3D CNN decoding module for decoding. First, two layers of 3x3x3 three-dimensional convolution operations are used for processing. Each layer of convolution operation further refines the spatial structure, captures local spatial information and structural details, and maintains spatial consistency in the input feature map. The role of this step is to extract local spatial information from multi-scale features and create a clearer three-dimensional feature representation, which is expressed as:

[0017] After two layers of 3D convolution, the self-attention mechanism is connected. The self-attention mechanism allows the model to capture long-distance dependencies between features globally, helping the network to better understand the complex relationships between spatial features. In this way, the network can improve the expression ability of spatial features from a global perspective, which can be expressed as:

[0018] Then the feature map after the self-attention mechanism will be further processed by the cross-attention mechanism. The purpose of the cross-attention mechanism is to process the association between feature maps from different perspectives and scales. Through this mechanism, the model can effectively fuse information from different input sources, so that the contextual dependency between feature maps can be better modeled, which is expressed as:

[0019] Finally, a two-layer 3x3x3 three-dimensional convolution operation is performed on the feature map processed by the cross attention mechanism. This stage further refines and optimizes the detail information in the feature map, strengthens the expression ability of spatial features, and helps to finally restore the three-dimensional spatial structure of the object. It is expressed as:

[0020] 3D feature fusion completes 3D reconstruction as follows: First, all representations are expressed as:

[0021] After the convolution operation, the feature maps of each layer are concatenated to merge the features from different sources, which can be expressed as:

[0022] The concatenated feature map undergoes a 3x3x3 convolution operation to further integrate spatial information and generate more refined three-dimensional spatial features. Finally, each voxel is normalized through the Softmax operation and the final three-dimensional reconstruction result is output. The final result is: .

[0023] Based on the above technical solution, step S5 is specifically as follows: First, input it into the building defect detection network, that is, the two-layer MLP, and get the output defect detection result:

[0024] Indicates whether the building has defects. Since it is a binary classification problem, the network will output a value between 0 and 1, indicating whether the building has defects. If it is greater than a certain threshold, it is predicted that the building has defects, otherwise it is considered that the building has no defects. The prediction process can be expressed as:

[0025] in Represents the prediction result, The activation function compresses the network output into the range of [0, 1] for binary classification tasks.

[0026] Based on the above technical solution, step S6 is to first construct a dynamic deviation prediction network, which receives the multi-scale multi-view fusion depth representation obtained in step S3 and predicts the building The minimum possible dynamic deviation in three dimensions and maximum value The prediction process is implemented through a two-layer MLP network. Each layer of the MLP network maps the input features into a new representation through a fully connected layer and predicts the dynamic deviation range of the building. First, input it into the dynamic deviation prediction network, that is, the two-layer MLP, and get the output dynamic deviation prediction result:

[0027] in Respectively indicate the buildings in The minimum and maximum values ​​of the range of dynamic deviation in three directions.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1) Low-cost and high-efficiency 3D modeling: Compared with the traditional LiDAR measurement method, this invention adopts the deep fusion technology of multiple visual sensors, uses low-cost equipment such as cameras and depth cameras, and combines deep learning and image processing technology to successfully achieve high-precision 3D modeling of buildings. This method can effectively cope with the challenges of lighting changes, obstructions and building structure complexity in complex construction environments, thereby improving the accuracy and robustness of modeling, significantly reducing equipment costs, and providing an efficient and economical solution for intelligent construction.

[0029] 2) Multi-scale, multi-perspective fusion building defect detection: Existing defect detection methods often rely on a single sensor or fixed perspective, which poses the risk of missed detection and false detection. The multi-scale, multi-perspective fusion defect detection method proposed in the present invention can intelligently identify various quality defects in the construction process, such as cracks, deviations or other structural problems, by combining visual data at different angles and scales. Through the fusion processing of multimodal data, this method greatly improves the comprehensiveness and accuracy of defect detection and ensures real-time monitoring of construction quality.

[0030] 3) Dynamic deviation prediction and feedback mechanism: Traditional construction deviation prediction methods often rely on historical data or post-review, lacking real-time and accuracy. By combining real-time 3D modeling data and defect detection results, the present invention provides a dynamic deviation prediction and feedback method that can monitor the occurrence of deviations in real time during the construction process, and make predictions and feedback outputs in a timely manner. This method can provide advanced warning for construction management, so that construction deviations can be identified and corrected at an early stage, thereby reducing quality risks and safety hazards during the construction process and improving the control capability of the construction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings constituting a part of this document are used to provide a further understanding of this document. The illustrative embodiments and descriptions of this document are used to explain this document and do not constitute improper limitations on this document.

[0032] Figure 1 It is the overall structure diagram of 3D modeling and quality inspection analysis of intelligent construction process based on deep fusion of multi-visual sensors; Figure 2 This is the structure diagram of a single-layer Transformer Encoder of a multi-view fusion encoder; Figure 3 It is a structural diagram of the multi-view representation fusion method; Figure 4 This is the 3D CNN Decoder structure diagram; Figure 5 It is the 3D feature fusion network structure diagram; Figure 6 It is a diagram of the defect detection network structure; Figure 7 It is a structural diagram of the dynamic deviation prediction network; DETAILED DESCRIPTION In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other arbitrarily.

[0033] Based on this, Figures 1 to 7 As shown, the exemplary embodiment of the present application provides a method for three-dimensional modeling and quality inspection and analysis of an intelligent construction process, comprising the following steps: S1. Preprocessing of multi-source visual sensor data: standardizing and denoising the multi-source visual sensor data, and uniformly adjusting the spatial resolution and brightness; Multi-source visual sensor data preprocessing is the starting step of this method, which aims to improve the quality and consistency of multi-source data and lay the foundation for subsequent multi-scale depth representation extraction. S2, using Swin Transformer to extract multi-scale deep representation of multi-visual sensor data after S1 preprocessing; S3, based on the multi-scale depth representation of multi-vision obtained in S2, is input into the multi-view fusion encoder for deep adaptive fusion encoding, and a multi-scale multi-view fusion representation is constructed at the same time, which includes the following two steps: S31, multi-view fusion coding, its structure diagram is as follows Figure 2 As shown in the figure, a multi-view fusion encoder composed of two layers of Transformer Encoder is used to deeply encode the multi-visual representation of each scale output by S2, capture the semantic interaction and correlation between each view, and generate an enhanced multi-view and multi-scale feature representation; S32, multi-scale representation fusion, its process is as follows Figure 3 As shown in the figure, the input representation and the encoding representation of the first step are concatenated, and feature fusion is performed through pooling operation to integrate multi-scale and multi-view information to generate the final multi-scale and multi-view fusion representation, providing high-quality feature input for subsequent defect detection and dynamic deviation prediction; S4, based on the multi-scale multi-view coding representation obtained in S3, construct a 3D decoding module, perform 3D CNN decoding, and input the constructed 3D feature fusion network to obtain the 3D reconstruction result; This step is based on the multi-scale multi-view coding representation obtained in S3, and construct a 3D decoding module for 3D reconstruction. Through the 3D CNN decoding module and feature fusion network, the object structure in the 3D space is further restored, and the final 3D reconstruction result is generated.

[0034] S5, constructing a building defect detection network and inputting the multi-scale and multi-view fusion deep representation obtained in S3 to obtain the building defect detection result; S6, construct a dynamic deviation prediction network and input the multi-scale and multi-view fusion deep representation obtained in S3 to obtain a dynamic deviation prediction result; S7, combining the building defect detection results obtained in S5 and the dynamic deviation prediction results in S6, outputs feedback results of defect detection and deviation prediction of buildings in the intelligent construction process.

[0035] Wherein, the step S1 comprises the following steps: S10, format standardization, converting the RGB image data collected by N multi-source visual sensors into a unified standard tensor format (such as N, C, H, W, where C = 3 represents the number of channels, and H and W are unified spatial resolutions) to ensure consistency and compatibility of subsequent processing steps; S11, pixel value normalization, in order to reduce the difference in dynamic range of different sensor data, the linear normalization method is used to adjust the pixel value range of the image data to 0,1, thereby reducing the interference of the numerical scale on the subsequent depth representation extraction; S12, denoising, using classic image denoising algorithms, Gaussian filtering and bilateral filtering, to smooth small-scale noise in the image while retaining edge information, using Gaussian kernel function to perform weighted averaging of image pixel values ​​to smooth random noise, and smoothing in both the spatial domain and the pixel intensity domain to retain edge clarity; S13, unified resolution. In view of the different resolutions of images collected by multi-source sensors, a bilinear interpolation algorithm is used to scale all images to a unified 768×768 pixel size. This resolution takes into account both processing efficiency and feature retention capabilities, and meets the needs of building scenes of different scales. S14, brightness adjustment and contrast enhancement, in order to improve the visual quality and feature contrast of the image, the pre-processed image data is subjected to histogram equalization processing, by redistributing pixel values, the brightness distribution of the image is made more uniform, and the texture details and defect features of the building surface are enhanced; S15. Data quality improvement and storage. After preprocessing, the multi-source visual data is stored in the form of standardized tensors to form a unified input data set, and the relevant preprocessing parameters (such as filter kernel size, scaling ratio, etc.) are recorded to support the calling and reproduction of subsequent modules.

[0036] In step S2, the Swin Transformer's block and window attention mechanism divides the image into fixed-size windows of 96×96, calculates local attention within the window, and extracts local features; multi-scale feature extraction extracts key deep representation information of buildings at different scales through a hierarchical feature extraction structure, taking into account the detailed expression of global and local features; feature integration realizes information interaction between different windows through alternating window movement operations, captures long-range dependencies and complex geometric information of building images, and the extracted multi-scale deep representation will serve as the key input for subsequent modules (such as multi-view fusion encoders and building defect detection networks), providing rich and detailed feature support for the entire intelligent construction process, and using the input data to integrate the information between different windows. Indicates that Indicates The image data of the visual sensor, the corresponding multi-scale depth representation extracted by Swin Transformer is recorded as ,here Indicates The scale features of the layer are 4 layers, so as to achieve the purpose of multi-scale representation expression; the whole S2 multi-scale feature extraction process is expressed as: ,in represents Swin Transformer, Represents the serial number of the vision sensor.

[0037] Preferably, the step S31 specifically comprises: for each scale of multi-visual representation, using TransformerEncoder to extract deep semantic correlation across perspectives, strengthen the global information of feature representation, and the input representation of each sensor After two layers of Transformer Encoder, the integrity and complementarity of representations from different perspectives are improved. The specific structure of Transformer Encoder is as follows: Figure 2 As shown in Figure 1, it includes self-attention mechanism (SA), cross-attention mechanism and multi-layer perceptron (MLP) modules. First, a nonlinear transformation is performed through a multi-layer perceptron to improve the complexity and adaptability of the representation, which can be expressed as:

[0038] After each The self-attention mechanism is used to analyze the relationship between local and global features within the same perspective, and the internal feature expression of each perspective is optimized to obtain :

[0039] Finally, for all The cross-attention mechanism is used to realize interactive encoding of features between different perspectives, enhance the complementarity between perspectives, and enable the information between perspectives to be effectively integrated to obtain corresponding representations. :

[0040] After the Transformer Encoder, each scale of the multi-visual representation will generate a corresponding deep encoding representation. This operation maps the multi-view features to a higher-dimensional semantic space through layer-by-layer semantic interactions, and finally obtains a multi-scale multi-view encoding representation. :

[0041] Furthermore, the multi-scale representation fusion in step S32 integrates the multi-scale and multi-view representations to generate a global feature representation to support comprehensive modeling of subsequent tasks. Specifically, the multi-view data of different layers are first concatenated and averaged pooled to obtain multi-view fusion features, and then the multi-view fusion features are concatenated and averaged pooled again to obtain multi-scale and multi-view fusion features. The operation process is as follows: Figure 3 shown.

[0042] For each scale’s original representation and the representation after multi-view fusion encoding, they are uniformly spliced ​​in the feature dimension. This step not only retains the diversity of the original representation, but also combines the optimized multi-view information to ensure the integrity and richness of the fusion. This process is expressed as:

[0043] here Represents the The view sensor data is stored in the multi-view fusion encoder. The feature representation corresponding to the layer, where 1 is the input multi-view fusion encoder as the first layer, the output of the first layer Transformer Encoder in the multi-view fusion encoder as the second layer, and the final output of the multi-view fusion encoder as the third layer, so there are 3 layers in total; Then the final fusion representation is obtained through average pooling:

[0044] in represents the average pooling operation, For the final multi-scale and multi-view fusion representation, S3 realizes the transformation from multi-visual representation to global fusion features through multi-view fusion encoding and multi-scale representation fusion, integrating information of different scales, perspectives and semantics, and providing high-quality input for subsequent 3D modeling, quality inspection and dynamic deviation prediction tasks.

[0045] Preferably, step S4 includes 3D CNN decoding and 3D feature fusion to complete 3D reconstruction. The purpose of 3D CNN decoding is to process the multi-scale multi-view coding representation through a series of convolution and attention mechanisms, gradually restore and extract the structural features in the 3D space, and finally obtain a 3D feature map with spatial semantic information through this process. The network architecture of the entire decoding module is as follows: Figure 4 As shown, the input comes from the multi-scale and multi-view representation of the S3 stage , which is a deep feature map containing multi-view and multi-scale information. The input is passed to the 3D CNN decoding module for decoding. First, two layers of 3x3x3 three-dimensional convolution operations are used for processing. Each layer of convolution operation further refines the spatial structure, captures local spatial information and structural details, and maintains spatial consistency in the input feature map. The role of this step is to extract local spatial information from multi-scale features and create a clearer three-dimensional feature representation, which is expressed as:

[0046] After two layers of 3D convolution, the self-attention mechanism is connected. The self-attention mechanism allows the model to capture long-distance dependencies between features globally, helping the network to better understand the complex relationships between spatial features. In this way, the network can improve the expression ability of spatial features from a global perspective, which can be expressed as:

[0047] Then the feature map after the self-attention mechanism will be further processed by the cross-attention mechanism. The purpose of the cross-attention mechanism is to process the association between feature maps from different perspectives and scales. Through this mechanism, the model can effectively fuse information from different input sources, so that the contextual dependency between feature maps can be better modeled, which is expressed as:

[0048] Finally, a two-layer 3x3x3 three-dimensional convolution operation is performed on the feature map processed by the cross attention mechanism. This stage further refines and optimizes the detail information in the feature map, strengthens the expression ability of spatial features, and helps to finally restore the three-dimensional spatial structure of the object. It is expressed as:

[0049] Through the feature fusion network, the features of different perspectives and scales are fused to output the final 3D reconstruction result. Through this step, the spatial information of multiple feature sources is effectively integrated to form a more accurate 3D reconstruction image. The structure of the entire fusion network is as follows: Figure 5 As shown, 3D feature fusion completes 3D reconstruction as follows: First, all representations (outputs from the 3D CNN decoding module) will first undergo a 2-layer 3x3x3 3D ​​convolution operation. This operation helps to further refine the features, enhance the expression of spatial information, and make the spatial details between features more prominent. It can be expressed as:

[0050] After the convolution operation, the feature maps of each layer are spliced ​​together to combine features from different sources (such as feature maps of different perspectives and scales) to ensure that all information is fully expressed. It can be expressed as:

[0051] The concatenated feature map undergoes a 3x3x3 convolution operation to further integrate spatial information and generate more refined three-dimensional spatial features. Finally, each voxel is normalized through the Softmax operation and the final three-dimensional reconstruction result is output. The final result is: .

[0052] By combining the 3D CNN decoding module and the 3D feature fusion network, S4 achieves 3D space reconstruction with multi-scale and multi-view encoding representation. The 3D CNN decoding module extracts spatial structure features through convolution and attention mechanisms, while the 3D feature fusion network further optimizes feature fusion through convolution and splicing operations, and finally outputs accurate 3D reconstruction results through Softmax operations. This method effectively restores the spatial structure and provides strong support for 3D modeling in intelligent construction.

[0053] As an implementation method, the defect detection network in step S5 is as follows: Figure 6 As shown, the network receives the multi-scale multi-view fusion deep representation from the S3 stage, and through the two-layer MLP (multi-layer perceptron) network, each layer of MLP realizes nonlinear mapping of features through a fully connected layer, thereby improving the model's ability to detect defects, and finally predicting whether the current building has defects. The step S5 is specifically as follows: First, input it into the building defect detection network, that is, the two-layer MLP, and get the output defect detection result:

[0054] Indicates whether the building has defects. Since it is a binary classification problem, the network will output a value between 0 and 1, indicating whether the building has defects. If it is greater than a certain threshold (set to 0.5 in this invention), it is predicted that the building has defects, otherwise it is considered that the building has no defects. The prediction process can be expressed as:

[0055] in Represents the prediction result, The activation function compresses the network output into the range of [0, 1] for binary classification tasks.

[0056] like Figure 7 As shown, the step S6 specifically includes first constructing a dynamic deviation prediction network, which receives the multi-scale multi-view fusion depth representation obtained in step S3 and predicts the building The minimum possible dynamic deviation in three dimensions and maximum value The prediction process is implemented through a two-layer MLP network. Each layer of the MLP network maps the input features into a new representation through a fully connected layer and predicts the dynamic deviation range of the building. First, input it into the dynamic deviation prediction network, that is, the two-layer MLP, and get the output dynamic deviation prediction result:

[0057] in Respectively indicate the buildings in The dynamic deviation prediction network uses a two-layer MLP network to predict the possible dynamic deviation of a building. Through multi-scale and multi-view fusion representation, the network can predict the dynamic deviation of a building based on its three-dimensional structure. The dynamic deviation range in direction can be determined, thus providing important reference information for subsequent quality inspection and construction process monitoring.

[0058] Step S7 is to combine the building defect detection results obtained in S5 and the dynamic deviation prediction results in S6 to output feedback of defect detection and deviation prediction for the building in the intelligent construction process. The output feedback result is:

[0059] This feedback data provides a reference for the monitoring of the intelligent construction process. By integrating the results of defect detection and deviation prediction, it can provide decision-making basis and data support for early warning, correction or optimization suggestions of the intelligent construction process, thereby realizing real-time tracking and control of the construction process.

[0060] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0061] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the intention of the present application also includes these modifications and variations.

Claims

1. A method for three-dimensional modeling and quality inspection and analysis of intelligent construction process, characterized in that: The steps include: S1. Preprocessing of multi-source visual sensor data: standardizing and denoising the multi-source visual sensor data, and uniformly adjusting the spatial resolution and brightness; S2, using Swin Transformer to extract multi-scale deep representation of multi-visual sensor data after S1 preprocessing; S3, based on the multi-scale depth representation of multi-vision obtained in S2, is input into the multi-view fusion encoder for deep adaptive fusion encoding, and a multi-scale multi-view fusion representation is constructed at the same time, which includes the following two steps: S31, multi-view fusion coding, through a multi-view fusion encoder composed of two layers of Transformer Encoder, deeply encodes the multi-visual representation of each scale output by S2, captures the semantic interaction and correlation between each view, and generates enhanced multi-view and multi-scale feature representation; S32, multi-scale representation fusion, concatenates the input representation and the encoding representation in the first step, and performs feature fusion through pooling operation, integrating multi-scale and multi-view information to generate the final multi-scale and multi-view fusion representation, providing high-quality feature input for subsequent defect detection and dynamic deviation prediction; S4, based on the multi-scale and multi-view coding representation obtained in S3, construct a 3D decoding module, perform 3D CNN decoding, and input the constructed 3D feature fusion network to obtain a 3D reconstruction result; S5, constructing a building defect detection network and inputting the multi-scale and multi-view fusion deep representation obtained in S3 to obtain the building defect detection result; S6, construct a dynamic deviation prediction network and input the multi-scale and multi-view fusion deep representation obtained in S3 to obtain a dynamic deviation prediction result; S7, combining the building defect detection results obtained in S5 and the dynamic deviation prediction results in S6, outputs feedback results of defect detection and deviation prediction of buildings in the intelligent construction process.

2. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 1 is characterized in that: The step S1 comprises the following steps: S10, format standardization, converting the RGB image data collected by N multi-source visual sensors into a unified standard tensor format; S11, pixel value normalization, using a linear normalization method to adjust the pixel value range of the image data to 0, 1, thereby reducing the interference of the numerical scale on the subsequent depth representation extraction; S12, denoising, using classic image denoising algorithms, Gaussian filtering and bilateral filtering, to smooth small-scale noise in the image while retaining edge information, using Gaussian kernel function to perform weighted averaging of image pixel values ​​to smooth random noise, and smoothing in both the spatial domain and the pixel intensity domain to retain edge clarity; S13, unified resolution, for the different resolutions of images collected by multi-source sensors, a bilinear interpolation algorithm is used to scale all images to a unified size of 768×768 pixels; S14, brightness adjustment and contrast enhancement, performing histogram equalization on the pre-processed image data, redistributing pixel values ​​to make the image brightness distribution more uniform, and enhancing the texture details and defect features of the building surface; S15. Data quality improvement and storage. After preprocessing, the multi-source visual data is stored in the form of standardized tensors to form a unified input data set, and the relevant preprocessing parameters are recorded to support the calling and reproduction of subsequent modules.

3. A method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 1 or 2, characterized in that: The Swin Transformer block and window attention mechanism in step S2 divides the image into fixed-size windows of 96×96, and calculates local attention within the window to extract local features; multi-scale feature extraction extracts key depth representation information of buildings at different scales through a hierarchical feature extraction structure; feature integration realizes information interaction between different windows through alternating window movement operations, captures long-range dependencies and complex geometric information of building images, and converts input data into Indicates that Indicates The image data of the visual sensor, the corresponding multi-scale depth representation extracted by Swin Transformer is recorded as ,here Indicates The scale features of the layer are 4 layers, so as to achieve the purpose of multi-scale representation expression; the whole S2 multi-scale feature extraction process is expressed as: ,in represents Swin Transformer, Represents the serial number of the vision sensor.

4. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 3 is characterized in that: The step S31 specifically uses Transformer Encoder to extract deep semantic correlation across perspectives for multi-visual representation of each scale, strengthen the global information of feature representation, and the input representation of each sensor After two layers of Transformer Encoder, the integrity and complementarity of representations from different perspectives are improved. Transformer Encoder includes self-attention mechanism, cross-attention mechanism and multi-layer perceptron module. First, a nonlinear transformation is performed through a multi-layer perceptron to improve the complexity and adaptability of the representation, which can be expressed as: After each The self-attention mechanism is used to analyze the relationship between local and global features within the same perspective, and the internal feature expression of each perspective is optimized to obtain : Finally, for all Perform cross attention mechanism to obtain corresponding representation : After the Transformer Encoder, each scale of the multi-visual representation will generate a corresponding deep encoding representation. This operation maps the multi-view features to a higher-dimensional semantic space through layer-by-layer semantic interactions, and finally obtains a multi-scale multi-view encoding representation. : 。 5. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 4 is characterized in that: The step S32 specifically includes firstly performing splicing and average pooling operations on multi-view data of different layers to obtain multi-view fusion features, and then performing splicing and average pooling operations on the multi-view fusion features again to obtain multi-scale and multi-view fusion features. For each scale, the original representation and the representation after multi-view fusion encoding are uniformly spliced ​​in the feature dimension. This process is expressed as: here Represents the The view sensor data is stored in the multi-view fusion encoder. The feature representation corresponding to the layer, where 1 is the input multi-view fusion encoder as the first layer, the output of the first layer Transformer Encoder in the multi-view fusion encoder as the second layer, and the final output of the multi-view fusion encoder as the third layer, so there are 3 layers in total; Then the final fusion representation is obtained through average pooling: in represents the average pooling operation, It is the final multi-scale and multi-view fusion representation.

6. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 1, characterized in that: The step S4 includes 3D CNN decoding and 3D feature fusion to complete 3D reconstruction. The purpose of the 3D CNN decoding is to process the multi-scale multi-view encoding representation through a series of convolution and attention mechanisms, gradually restore and extract the structural features in the three-dimensional space. Finally, through this process, a three-dimensional feature map with spatial semantic information is obtained, and the multi-scale multi-view representation from the S3 stage is input. , which is a deep feature map containing multi-view and multi-scale information. The input is passed to the 3D CNN decoding module for decoding. First, two layers of 3x3x3 three-dimensional convolution operations are used for processing. Each layer of convolution operation further refines the spatial structure, captures local spatial information and structural details, and maintains spatial consistency in the input feature map. The role of this step is to extract local spatial information from multi-scale features and create a clearer three-dimensional feature representation, which is expressed as: After two layers of 3D convolution, the self-attention mechanism is connected. The self-attention mechanism allows the model to capture long-distance dependencies between features globally, helping the network to better understand the complex relationships between spatial features. In this way, the network can improve the expression ability of spatial features from a global perspective, which can be expressed as: Then the feature map after the self-attention mechanism will be further processed by the cross-attention mechanism. The purpose of the cross-attention mechanism is to process the association between feature maps from different perspectives and scales. Through this mechanism, the model can effectively fuse information from different input sources, so that the contextual dependency between feature maps can be better modeled, which is expressed as: Finally, a two-layer 3x3x3 three-dimensional convolution operation is performed on the feature map processed by the cross attention mechanism. This stage further refines and optimizes the detail information in the feature map, strengthens the expression ability of spatial features, and helps to finally restore the three-dimensional spatial structure of the object. It is expressed as: 3D feature fusion completes 3D reconstruction as follows: First, all representations are expressed as: After the convolution operation, the feature maps of each layer are concatenated to merge the features from different sources, which can be expressed as: The concatenated feature map undergoes a 3x3x3 convolution operation to further integrate spatial information and generate more refined three-dimensional spatial features. Finally, each voxel is normalized through the Softmax operation and the final three-dimensional reconstruction result is output. The final result is: 。 7. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 5, characterized in that: The step S5 is specifically as follows: First, input it into the building defect detection network, that is, the two-layer MLP, and get the output defect detection result: Indicates whether the building has defects. Since it is a binary classification problem, the network will output a value between 0 and 1, indicating whether the building has defects. If it is greater than a certain threshold, it is predicted that the building has defects, otherwise it is considered that the building has no defects. The prediction process can be expressed as: in Represents the prediction result, The activation function compresses the network output into the range of [0, 1] for binary classification tasks.

8. The method for three-dimensional modeling and quality inspection and analysis of intelligent construction process according to claim 7, characterized in that: The step S6 is to first construct a dynamic deviation prediction network, which receives the multi-scale multi-view fusion depth representation obtained in step S3 and predicts the building The minimum possible dynamic deviation in three dimensions and maximum value The prediction process is implemented through a two-layer MLP network. Each layer of the MLP network maps the input features into a new representation through a fully connected layer and predicts the dynamic deviation range of the building. First, input it into the dynamic deviation prediction network, that is, the two-layer MLP, and get the output dynamic deviation prediction result: in Respectively indicate the buildings in The minimum and maximum values ​​of the range of dynamic deviation in three directions.

Images