Construction method and system for pre-buried connecting node for hoisting high-rise building steel structure
By embedding DCNv2 and PSA mechanisms into the YOLOv8 model and combining them with the Wise-IoU loss function, the problem of insufficient accuracy in node detection during high-rise building hoisting is solved, achieving high-precision node identification and deviation measurement, thus ensuring the stability and safety of construction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA NUCLEAR IND HUAXING CONSTR
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-14
AI Technical Summary
Traditional YOLO models face challenges in high-rise building hoisting, including dynamic environments, complex backgrounds, target rotation, and occlusion, leading to insufficient accuracy and affecting construction stability and safety.
By employing the YOLOv8 model combined with DCNv2 deformable convolution, PSA polarization self-attention mechanism, and Wise-IoU loss function, and through image sequence preprocessing and detection model adjustment, detection results of pre-embedded connection nodes are generated, and position and orientation deviations are calculated to adjust the hoisting equipment.
It improves the accuracy and robustness of node detection, ensuring the effectiveness and continuity of node identification and deviation correction in dynamic construction environments, thereby enhancing hoisting efficiency and safety.
Smart Images
Figure CN122391825A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of building construction technology, specifically relating to a construction method and system for pre-embedded connection nodes used in the hoisting of steel structures in high-rise buildings. Background Technology
[0002] With the acceleration of urbanization, the construction of high-rise buildings is increasing globally. As an indispensable core component of modern architecture, the hoisting process of steel structures is crucial for ensuring construction quality and safety. In steel structure construction, pre-embedded connection nodes are key nodes that connect various parts of the steel structure using bolts or welding. The precise hoisting and docking of these nodes not only affects the construction progress but also directly determines the stability and safety of the structure. With the continuous development of hoisting and automation technologies, hoisting assistance systems based on computer vision and deep learning have become a hot topic in research and application. Modern hoisting systems, through the combination of computer vision, deep learning, and sensing technologies, can now achieve automatic detection, deviation measurement, and automatic docking control of pre-embedded nodes, thereby improving hoisting efficiency and accuracy.
[0003] Although some deep learning-based object detection technologies have been applied in the field of building construction, most of these technologies focus on the recognition of static images. Furthermore, while the traditional YOLO model performs well in static object detection, it is prone to insufficient accuracy in dynamic environments, complex backgrounds, and object rotation and occlusion that are common in high-rise building hoisting, which leads to difficulties in subsequent construction. Summary of the Invention
[0004] This invention addresses the shortcomings of existing technologies by providing a construction method and system for pre-embedded connection nodes in the hoisting of steel structures for high-rise buildings. It solves the problem that traditional YOLO models suffer from insufficient accuracy due to dynamic environments, complex backgrounds, and target rotation and occlusion during hoisting, which leads to difficulties in subsequent construction.
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0006] In a first aspect, the present invention provides a construction method for pre-embedded connection nodes for hoisting steel structures in high-rise buildings, comprising the following steps: S1: Collect image sequences of steel structure hoisting, preprocess them, generate a dataset, divide the dataset to obtain training and validation tensors, and form a tensor set; S2: Construct a detection model, embed deformable convolution and polarization self-attention mechanism, take training and validation tensors as input, output node boxes and quintuples, adjust the node boxes through adjustment mechanism, and generate detection results of pre-embedded connected nodes; S3: Based on the detection results, calculate the positional and orientation deviations of the pre-embedded connection nodes. Based on the positional and orientation deviations, generate control quantities and feed them back to the hoisting equipment for construction adjustments. Store the control quantities in the database.
[0007] Optionally, step S1 includes the following sub-steps: Industrial cameras are installed on the hoisting components to acquire a sequence of images of continuous pre-embedded nodes during the hoisting process of high-rise building steel structures. The continuous pre-embedded node image sequence is cleaned and normalized to obtain a standardized image sequence; The intrinsic parameters of the industrial camera were calibrated using Zhang Zhengyou's calibration method to obtain the camera's intrinsic parameter matrix. A dataset is constructed based on standardized image sequences and camera intrinsic parameter matrices; Based on the dataset, standardized image sequences are extracted and divided into training and validation sets, and the height and width of the images are recorded. Based on the training set and validation set, construct their respective pixel matrices, flatten them into input tensors, obtain the training and validation tensors, and integrate them into a tensor set.
[0008] Optionally, in step S2, the process of outputting the node frame and the quintuple is as follows: The training tensor is input into the detection model built on the YOLOv8 model, and high-order feature modeling operations are performed through deformable convolution to generate the expression tensor; By using a polarization self-attention mechanism to enhance the representation tensor with both channel attention and spatial attention, a dual-enhanced tensor is obtained, including a channel attention-enhanced tensor and a spatial attention-enhanced tensor. Perform element-wise multiplication fusion on the dual-enhancement tensor to obtain the fused feature tensor; The fused feature tensor is input into the detection head of the detection model. A convolutional layer is embedded in the detection head, and the multi-scale fused feature tensor is convolved to output node boxes and quintuples. The quintuple includes the center point coordinates, the width and height of the node frame, the node orientation angle, and the node frame confidence level. Define multiple loss functions including Wise-IoU loss, rotation angle loss, and confidence loss. Weight and fuse the multiple loss functions to generate a total loss function and minimize the value of the total loss function. The AdamW optimizer is used for training iterations. During the iteration process, the iteration stops when the loss value of the total loss function no longer decreases, and the trained detection model is output. The validation tensor is input into the trained detection model, and the output is the node bounding box and the quintuple.
[0009] Optionally, in step S2, the process of adjusting the node frame through the adjustment mechanism is as follows: The intersection-union ratio (CUNR) is calculated using the CUNR formula, and a CUNR threshold is set. If the intersection-union ratio (IU) is greater than or equal to the IU threshold, the current node box is retained and its confidence level is updated; otherwise, the current node box is discarded. Based on the updated confidence level, the node boxes are sorted in descending order of confidence level, and the top ones are selected. Each node bounding box is used as the detection result.
[0010] Optionally, in step S2: The formula for expressing the tensor is: ; In the formula, Indicates the first The layer's representation tensor This represents the training tensor. Represents the learnable weight matrix. This indicates the transpose operation. Indicates the first The layer's representation tensor Indicates the first Layer bias terms; The enhancement of channel attention for the representation tensor through the polarization self-attention mechanism is described by the following formula: ; In the formula, The channel attention enhancement tensor representing the output of the polarization self-attention mechanism. , , These represent the Sigmoid activation function, , , These represent the learnable query, key, and value fusion mapping tensors, respectively. This represents the Softmax activation function; The enhancement of spatial attention to the representation tensor through the polarization self-attention mechanism is described by the following formula: ; In the formula, This represents the spatial attention enhancement tensor output by the polarization self-attention mechanism. Indicates global average pooling; The formula for the fused feature tensor is: ; In the formula, Represents the fused feature tensor. This represents an element-wise multiplication operation; The formula for convolving the multi-scale feature tensor is as follows: ; In the formula, This represents the convolution tensor after convolution. This represents a two-dimensional convolution operation.
[0011] Optionally, in step S2: The formula for the Wise-IoU loss function is: ; In the formula, This represents the Wise-IoU loss function value. This represents the intersection-union ratio (IoU) between the predicted bounding boxes and the labeled bounding boxes. The weights representing the intersection-union ratio; The intersection-union ratio (IUU) of the predicted bounding box and the labeled bounding box is obtained by the following formula: ; In the formula, This represents the area or number of pixels of the intersection region between the predicted bounding box and the ground truth bounding box. This represents the area or number of pixels of the union of the predicted bounding box and the ground truth bounding box. Indicates the prediction box. This represents the intersection operation. Represents the true bounding box. This represents the union operation; The formula for the rotation angle loss is:
[0012] In the formula, This represents the value of the rotation angle loss function. Indicates the prediction angle. Indicates the actual angle; The formula for the confidence loss is: ; In the formula, This represents the confidence loss function value. Indicates the true label, Represents the natural logarithm function. This represents the confidence level of the predicted node bounding box; The formula for the total loss function is:
[0013] In the formula, This represents the total loss function value. The weights represent the values of the Wise-IoU loss function. The weights represent the values of the rotation angle loss function. The weights represent the values of the confidence loss function.
[0014] Optionally, in step S2: The confidence level of the node frame is updated using the following formula: ; In the formula, Indicates the updated number The confidence level of each node frame. Indicates the first The original confidence of each node bounding box. The base of the natural logarithm. This represents the calculated intersection-union ratio. This indicates that the parameters are being adjusted.
[0015] Optionally, step S3 includes the following sub-steps: Based on the validation set, a standardized image sequence is extracted, and the coordinates of the center point of the image are solved using the image geometric center calculation method. Based on the detection results, the corresponding center point coordinates are extracted. Using the center point coordinates as a basis, the coordinate deviation is calculated using subtraction to obtain the positional deviation. , Indicates the deviation in the horizontal direction. Indicates deviation in the vertical direction; Obtain the ideal node orientation angle from the design drawings, and then use subtraction to calculate the deviation between the node orientation angle and the ideal node orientation angle, which is defined as the attitude deviation. ; Based on positional deviation and attitude deviation Calculate the control quantity; The control quantity is used as the controller of the hoisting equipment. After the controller of the hoisting equipment analyzes the control quantity, it instructs the hoisting equipment to make horizontal, vertical and rotational adjustments during construction. The control quantities are stored in the database, and a corresponding hoisting equipment ID is added to the control quantity during storage.
[0016] Optionally, in the sub-steps of step S3: The formula for the control quantity is: ; In the formula, Indicates the control quantity. This represents the proportional gain coefficient for position control. This represents the proportional gain coefficient for attitude control.
[0017] Secondly, the present invention provides a construction system for pre-embedded connection nodes in the hoisting of steel structures of high-rise buildings, comprising: The acquisition and partitioning module is used to acquire image sequences of steel structure hoisting, preprocess them, generate a dataset, partition the dataset to obtain training and validation tensors, and form a tensor set. The detection adjustment module is used to build a detection model, embedding deformable convolution and polarization self-attention mechanisms. It takes training and validation tensors as input and outputs node boxes and quintuples. The adjustment mechanism adjusts the node boxes to generate detection results with pre-embedded connected nodes. The calculation and storage module is used to calculate the position and attitude deviations of the pre-embedded connection nodes based on the detection results. Based on the position and attitude deviations, control quantities are generated and fed back to the hoisting equipment for construction adjustments. The control quantities are stored in the database.
[0018] The beneficial effects of this invention are: This invention significantly improves the accuracy and robustness of node detection by introducing a detection model based on the YOLOv8 architecture, combined with DCNv2 deformable convolution, PSA polarization self-attention mechanism, and Wise-IoU loss function. Furthermore, by utilizing post-processing techniques with adjustment mechanisms, false detections caused by node overlap or image blurring during hoisting are avoided. This enables the invention to not only track node positions stably in real time, but also ensure the effectiveness and continuity of node recognition and deviation correction in dynamic construction environments. Attached Figure Description
[0019] Figure 1 A flowchart of a construction method for pre-embedded connection nodes used in the hoisting of steel structures in high-rise buildings; Figure 2 This is a structural diagram of a construction system for pre-embedded connection nodes used in the hoisting of steel structures in high-rise buildings; Figure 3 This is a flowchart for outputting the final detection results. Detailed Implementation
[0020] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0021] In the first embodiment, the present invention proposes a construction method for pre-embedded connection nodes for hoisting steel structures in high-rise buildings, such as... Figure 1 As shown, it includes the following steps: S1: Collect image sequences of steel structure hoisting, preprocess them, generate a dataset, divide the dataset to obtain training and validation tensors, and form a tensor set; Specifically, image sequences of steel structure hoisting are collected, preprocessed, and used to generate a dataset, including: An industrial camera is installed on the hoisting component to acquire a continuous image sequence of pre-embedded nodes during the hoisting process of high-rise building steel structure. The continuous pre-embedded node image sequence is cleaned and normalized to obtain a standardized image sequence. The intrinsic parameters of the industrial camera were calibrated using Zhang Zhengyou's calibration method to obtain the camera's intrinsic parameter matrix. A dataset is constructed based on standardized image sequences and camera intrinsic parameter matrices.
[0022] By cleaning and normalizing the images, the consistency and stability of data quality were ensured, providing high-quality input for subsequent model training. Secondly, the Zhang Zhengyou calibration method was used to calibrate the intrinsic parameters of the industrial camera, ensuring accurate mapping between camera coordinates and physical space. This allows the position and size in the image to correctly reflect the spatial location of the actual component, thereby improving the accuracy of node detection.
[0023] Furthermore, based on the dataset, training and validation tensors are obtained, forming a tensor set, including: Based on the dataset, standardized image sequences are extracted and divided into training and validation sets. The height and width of the images are then recorded. Based on the training set and validation set, construct their respective pixel matrices, flatten them into input tensors, obtain the training and validation tensors, and integrate them into a tensor set.
[0024] By standardizing image sequences and constructing datasets, the consistency and quality of image data are improved, providing accurate input for deep learning models. Transforming image data into pixel matrices and flattening them into input tensors effectively extracts spatial features, supporting object detection tasks. By appropriately dividing the training and validation sets, the model's generalization ability is ensured, avoiding overfitting.
[0025] S2: Construct a detection model, embed deformable convolution and polarization self-attention mechanism, take training and validation tensors as input, output node boxes and quintuples, adjust the node boxes through the adjustment mechanism, and generate detection results of pre-embedded connected nodes.
[0026] Specifically, a detection model is constructed, embedding deformable convolution and polarization self-attention mechanisms. The training and validation tensors are used as input, and the output includes node bounding boxes and quintuples, including: A detection model is built based on the YOLOv8 model architecture and embedded with deformable convolution (DCNv2) and polarized self-attention mechanism.
[0027] Based on the tensor set, the training tensor is extracted and input into the detection model. High-order feature modeling operations are then performed using deformable convolution to generate the representation tensor, as shown in the formula: ; In the formula, Indicates the first The layer's representation tensor This represents the training tensor. Represents the learnable weight matrix. This indicates the transpose operation. Indicates the first The layer's representation tensor Indicates the first Layer bias terms.
[0028] Based on the tensor set, the representation tensor is extracted, and the representation tensor is enhanced by both channel attention and spatial attention through the polarization self-attention mechanism to obtain the dual-enhanced tensor, including the channel attention and spatial attention enhanced tensor.
[0029] Channel attention is enhanced in the representation tensor through a polarization self-attention mechanism, as shown in the formula: ; In the formula, The channel attention enhancement tensor representing the output of the polarization self-attention mechanism. , , These represent the Sigmoid activation function, , , These represent the learnable query, key, and value fusion mapping tensors, respectively. This represents the Softmax activation function.
[0030] Spatial attention is enhanced for the representation tensor through a polarization self-attention mechanism, as shown in the formula: ; In the formula, This represents the spatial attention enhancement tensor output by the polarization self-attention mechanism. This indicates global average pooling.
[0031] Perform element-wise multiplication fusion on the dual-enhancement tensor to obtain the fused feature tensor, as shown in the formula: ; In the formula, Represents the fused feature tensor. This indicates an element-wise multiplication operation.
[0032] The fused feature tensor is used as the input to the detection head in the detection model. A convolutional layer is embedded in the detection head, and after convolution of the multi-scale feature tensor, the detection head outputs the node box and quintuple, which includes the center point coordinates, the width and height of the node box, the node orientation angle, and the node box confidence.
[0033] The formula for convolving multi-scale feature tensors is: ; In the formula, This represents the convolution tensor after convolution. This represents a two-dimensional convolution operation.
[0034] Define multiple loss functions, including Wise-IoU loss, rotation angle loss, and confidence loss. Weight and fuse these multiple loss functions to generate a total loss function, and minimize the value of the total loss function.
[0035] Wise-IoU loss, the formula is: ; In the formula, This represents the Wise-IoU loss function value. This represents the intersection-union ratio (IoU) between the predicted bounding boxes and the labeled bounding boxes. The weight of the intersection-union ratio (this parameter ranges from 0.5 to 2, and the optimal value can be determined by grid search or 1 can be selected as the default value within this range).
[0036] The intersection-union ratio (IUU) of the predicted bounding box and the labeled bounding box is obtained using the following formula: ; In the formula, This represents the area or number of pixels of the intersection region between the predicted bounding box and the ground truth bounding box. This represents the area or number of pixels of the union of the predicted bounding box and the ground truth bounding box. Indicates the prediction box. This represents the intersection operation. Represents the true bounding box (which can be generated through manual annotation). This indicates the union operation.
[0037] The formula for rotation angle loss is: ; In the formula, This represents the value of the rotation angle loss function. Indicates the prediction angle. The actual angle can be represented by the rotation angle of each node box, which can be generated by annotating the image with tools such as LabelImg and CVAT.
[0038] The confidence loss formula is: ; In the formula, This represents the confidence loss function value. This indicates a real label (generated by manual annotation). This indicates that there is a target at this location, meaning that a certain area in the image contains a pre-embedded node. This indicates that there is no target at that location, meaning that the area does not contain any pre-buried nodes. Represents the natural logarithm function. This represents the confidence level of the predicted node bounding box.
[0039] The total loss function is expressed as follows: ; In the formula, This represents the total loss function value. The weights represent the values of the Wise-IoU loss function. The weights represent the values of the rotation angle loss function. The weights represent the values of the confidence loss function.
[0040] , , The range of values and the basis for their selection are as follows: To ensure the predicted bounding box matches the actual bounding box, a value of 0-5 can be set. For situations with high positional accuracy, such as in building hoisting where nodes must be accurately aligned with pre-embedded positions, a value of 2 is recommended as the default. In defect detection tasks, especially when errors in the location of small targets can significantly affect overall detection accuracy, a value of 3 is recommended as the default. The value range can be set from 0.5 to 2. In general scenarios, the impact of rotation angle is relatively small, and most construction projects do not involve significant node rotation, so 1 can be used as the default value. However, in precise docking scenarios, especially when the node installation direction (angle) affects subsequent equipment operation, 2 can be used as the default value. The value range is set to 0.5~2. Since the existence of pre-embedded nodes must be accurately determined in high-rise hoisting, otherwise the docking quality will be affected, 1.5 or 2 can be set as the default value, and subsequent adjustments can be made through grid search.
[0041] The AdamW optimizer is used for training iterations. During the iteration process, the iteration stops when the loss value of the total loss function no longer decreases significantly, and the trained detection model is output.
[0042] The validation tensor is input into the trained detection model, and the output is the node bounding box and the quintuple.
[0043] It should be noted that during the initialization phase, all learnable parameters can be generated through random initialization.
[0044] Furthermore, the node frame is adjusted through an adjustment mechanism to generate the detection results of the pre-embedded connection nodes. The process is as follows: Figure 3 As shown, it includes: Based on the output node bounding boxes, the intersection-union ratio (IUR) of each node bounding box with other node bounding boxes is calculated using the IUR formula. An IUR threshold is set (the value of this threshold can be set from 0.3 to 0.7, and 0.5 can be used as the default value in order to effectively distinguish different targets and avoid excessive suppression).
[0045] If the intersection-union ratio (IU) is greater than or equal to the IU threshold, the current node bounding box is retained, and the confidence score of that node bounding box is updated using the following formula: ; In the formula, Indicates the updated number The confidence level of each node frame. Indicates the first The original confidence of each node bounding box. The base of the natural logarithm. This represents the calculated intersection-union ratio. This indicates the adjustment parameter (the value of this parameter can be set from 0.05 to 0.5. In order to ensure that the detection box is not excessively suppressed in the case of node overlap and shaking, and at the same time to reduce redundant boxes, 0.1 can be used as the default value).
[0046] Otherwise, remove the current node frame.
[0047] Based on the updated confidence level, the node boxes are sorted in descending order of confidence level, and the top ones are selected. A number of (e.g., 5 to 10) node boxes are used as the detection result.
[0048] By embedding deformable convolution (DCNv2), this embodiment can adaptively handle the diverse geometries and sizes of pre-embedded nodes during hoisting, recognizing not only standard rectangular node boxes but also targets with irregular shapes. This step overcomes the limitations of the traditional YOLOv8 model when facing changes in node shape. Furthermore, through the polarization self-attention mechanism (PSA), this embodiment enhances the extraction capability of target features in complex backgrounds, avoiding the influence of interference factors such as steel structures, hoisting equipment, and changes in ambient lighting, thus improving accuracy and robustness in real construction environments. Traditional target detection models often face accuracy issues with small targets and irregularly shaped targets. The Wise-IoU loss function addresses this problem, effectively improving the localization accuracy of small targets and targets with irregular edges. In steel structure hoisting, pre-embedded nodes are typically small in size and have slender edges; Wise-IoU can significantly optimize the accurate detection of these nodes. Furthermore, Wise-IoU, through weighted IoU loss, adjusts weights during the matching process of different targets, effectively preventing traditional IoU methods from over-penalizing small targets or targets with large positional deviations, thereby improving the reliability of the model in complex construction scenarios. Secondly, in steel structure hoisting, the angle of embedded nodes may deviate due to factors such as wind and vibration, leading to misalignment of hoisted components. In this embodiment, rotation angle loss optimizes the target rotation angle to ensure precise alignment of nodes during hoisting, reducing installation difficulties caused by node angle deviations. Thirdly, in high-rise building construction, environmental changes and hoisting vibrations may cause a decrease in the confidence of some node frames. Confidence loss, by strengthening the optimization of confidence judgment, reduces the occurrence of erroneous judgments, improving system stability and hoisting efficiency.
[0049] Therefore, by embedding DCNv2, PSA, Wise-IoU, etc. into the YOLOv8 model, this embodiment greatly improves the accuracy of node identification and the reliability of deviation measurement, and enables this embodiment to maintain high-precision node identification and adjustment capabilities under different construction conditions.
[0050] In steel structure hoisting, pre-embedded connection nodes often overlap due to factors such as wind, component swaying, or occlusion in the image during hoisting. However, traditional non-maximum suppression (NMS) methods often have problems in such situations, especially when multiple nodes are closely arranged or components overlap. Standard NMS may mistakenly delete some valid boxes, leading to missed detections and affecting subsequent deviation measurements and hoisting control. This embodiment solves this problem by smoothly attenuating the confidence of overlapping boxes instead of directly suppressing them. Specifically, by smoothly updating the confidence of boxes overlapping with the current box based on the intersection-over-union (IoU) ratio and attenuation factor, erroneous deletion is avoided. This processing not only preserves multiple target boxes but also effectively reduces the number of redundant boxes, improving the stability and completeness of the detection results.
[0051] S3: Based on the detection results, calculate the positional and orientation deviations of the pre-embedded connection nodes. Based on the positional and orientation deviations, generate control quantities and feed them back to the hoisting equipment for construction adjustments. Store the control quantities in the database.
[0052] Specifically, based on the test results, the positional and orientation deviations of the pre-embedded connection nodes are calculated. Using these deviations as a basis, control parameters are generated and fed back to the hoisting equipment for construction adjustments, including: Based on the validation set, standardized image sequences are extracted. The geometric center of the image is calculated to determine the coordinates of the image's center points. Then, based on the detection results, the corresponding center point coordinates are extracted. Using the center point coordinates as a basis, subtraction is used to calculate the coordinate deviation, thus obtaining the positional deviation. , Indicates the deviation in the horizontal direction. This indicates the deviation in the vertical direction.
[0053] Obtain the ideal node orientation angle from the design drawings, and then use subtraction to calculate the deviation between the node orientation angle and the ideal node orientation angle, which is defined as the attitude deviation. .
[0054] Based on position deviation and attitude deviation, the control quantity is calculated using the following formula: ; In the formula, Indicates the control quantity. This represents the position control proportional gain coefficient (this position control proportional gain coefficient determines the impact of position deviation on the horizontal movement and vertical lifting of the hoisting equipment, and its value range can be set from 0.5 to 2. In order to meet the requirements of higher hoisting accuracy, 0.5 can be used as the default value). This represents the attitude control proportional gain coefficient (this attitude control proportional gain coefficient determines the influence of attitude deviation on the rotation adjustment of the hoisting equipment, that is, controlling the rotation of the hoisting component around its axis, and its value range can be set to 0.1~1, and in this embodiment, 0.1 can be taken as the default value).
[0055] The control quantity is used as the controller of the hoisting equipment. After the controller of the hoisting equipment analyzes the control quantity, it instructs the hoisting equipment to make horizontal, vertical and rotational adjustments during construction.
[0056] The center point of the target node is determined by image geometric center calculation, and the horizontal and vertical deviations are calculated in real time by comparing it with the standard position. This method can accurately obtain the actual position of the node and quickly provide feedback on deviation information, providing a reliable basis for subsequent adjustments. Secondly, by comparing the angle of the target node with the ideal position in the design drawings, the attitude deviation is calculated, providing data support for the angle adjustment of the hoisting components, thereby ensuring precise alignment between the node and the pre-embedded position. When calculating control quantities, this embodiment combines positional and attitude deviations with corresponding proportional gain coefficients to flexibly adjust the movement and rotation of the hoisting equipment. Through control quantity feedback, the hoisting equipment can perform real-time horizontal, vertical, and rotational adjustments, thereby achieving more precise node alignment. This automated process reduces manual intervention, significantly improves construction efficiency, and reduces human error.
[0057] Furthermore, the control variables are stored in a database, including: The control quantities are stored in the database, and a corresponding hoisting equipment ID is added to the control quantity during storage.
[0058] By precisely storing control parameters in a database, the integrity and accuracy of hoisting data are ensured, providing reliable data support for subsequent operations. Through device ID association, the system can track the operational status of different devices in real time, improving work efficiency.
[0059] In another embodiment, the present invention proposes as follows: Figure 2This diagram illustrates a construction system for pre-embedded connection nodes in the hoisting of steel structures for high-rise buildings. It implements a construction method for pre-embedded connection nodes in the hoisting of steel structures for high-rise buildings as described in the preceding embodiments. The system includes an acquisition and partitioning module, a detection and adjustment module, and a computation and storage module. The acquisition and partitioning module acquires image sequences of the steel structure hoisting, preprocesses them to generate a dataset, and partitions the dataset to obtain training and validation tensors, forming a tensor set. The detection and adjustment module constructs a detection model, embedding deformable convolution and polarization self-attention mechanisms. It takes the training and validation tensors as input and outputs node bounding boxes and quintuples. The adjustment mechanism adjusts the node bounding boxes to generate detection results for the pre-embedded connection nodes. The computation and storage module calculates the positional and orientation deviations of the pre-embedded connection nodes based on the detection results. Based on these deviations, it generates control quantities that are fed back to the hoisting equipment for construction adjustments. The control quantities are stored in a database. The functions of each module in this system correspond to the steps of the method in the preceding embodiments, and therefore will not be repeated here.
[0060] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0061] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A construction method for pre-embedded connection nodes used in the hoisting of steel structures in high-rise buildings, characterized in that, Includes the following steps: S1: Collect image sequences of steel structure hoisting, preprocess them, generate a dataset, divide the dataset to obtain training and validation tensors, and form a tensor set; S2: Construct a detection model, embed deformable convolution and polarization self-attention mechanism, take training and validation tensors as input, output node boxes and quintuples, adjust the node boxes through adjustment mechanism, and generate detection results of pre-embedded connected nodes; S3: Based on the detection results, calculate the positional and orientation deviations of the pre-embedded connection nodes. Based on the positional and orientation deviations, generate control quantities and feed them back to the hoisting equipment for construction adjustments. Store the control quantities in the database.
2. The construction method for pre-embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Industrial cameras are installed on the hoisting components to acquire a sequence of images of continuous pre-embedded nodes during the hoisting process of high-rise building steel structures. The continuous pre-embedded node image sequence is cleaned and normalized to obtain a standardized image sequence; The intrinsic parameters of the industrial camera were calibrated using Zhang Zhengyou's calibration method to obtain the camera's intrinsic parameter matrix. A dataset is constructed based on standardized image sequences and camera intrinsic parameter matrices; Based on the dataset, standardized image sequences are extracted and divided into training and validation sets, and the height and width of the images are recorded. Based on the training set and validation set, construct their respective pixel matrices, flatten them into input tensors, obtain the training and validation tensors, and integrate them into a tensor set.
3. The construction method for pre-embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 1, characterized in that, In step S2, the process of outputting the node frame and the quintuple is as follows: The training tensor is input into the detection model built on the YOLOv8 model, and high-order feature modeling operations are performed through deformable convolution to generate the expression tensor; By using a polarization self-attention mechanism to enhance the representation tensor with both channel attention and spatial attention, a dual-enhanced tensor is obtained, including a channel attention-enhanced tensor and a spatial attention-enhanced tensor. Perform element-wise multiplication fusion on the dual-enhancement tensor to obtain the fused feature tensor; The fused feature tensor is input into the detection head of the detection model. A convolutional layer is embedded in the detection head, and the multi-scale fused feature tensor is convolved to output node boxes and quintuples. The quintuple includes the center point coordinates, the width and height of the node frame, the node orientation angle, and the node frame confidence level. Define multiple loss functions including Wise-IoU loss, rotation angle loss, and confidence loss. Weight and fuse the multiple loss functions to generate a total loss function and minimize the value of the total loss function. The AdamW optimizer is used for training iterations. During the iteration process, the iteration stops when the loss value of the total loss function no longer decreases, and the trained detection model is output. The validation tensor is input into the trained detection model, and the output is the node bounding box and the quintuple.
4. The construction method for embedded connection nodes used in the hoisting of steel structures in high-rise buildings as described in claim 1, characterized in that, In step S2, the process of adjusting the node frame through the adjustment mechanism is as follows: The intersection-union ratio (CUNR) is calculated using the CUNR formula, and a CUNR threshold is set. If the intersection-union ratio (IU) is greater than or equal to the IU threshold, the current node box is retained, and the confidence of the node box is updated. Otherwise, remove the current node frame; Based on the updated confidence level, the node boxes are sorted in descending order of confidence level, and the top ones are selected. Each node bounding box is used as the detection result.
5. The construction method for pre-embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 3, characterized in that, In step S2: The formula for expressing the tensor is: ； In the formula, Indicates the first The layer's representation tensor This represents the training tensor. Represents the learnable weight matrix. This indicates the transpose operation. Indicates the first The layer's representation tensor Indicates the first Layer bias terms; The enhancement of channel attention for the representation tensor through the polarization self-attention mechanism is described by the following formula: ； In the formula, The channel attention enhancement tensor representing the output of the polarization self-attention mechanism. , , These represent the Sigmoid activation function, , , These represent the learnable query, key, and value fusion mapping tensors, respectively. This represents the Softmax activation function; The enhancement of spatial attention to the representation tensor through the polarization self-attention mechanism is described by the following formula: ； In the formula, This represents the spatial attention enhancement tensor output by the polarization self-attention mechanism. Indicates global average pooling; The formula for the fused feature tensor is: ； In the formula, Represents the fused feature tensor. This represents an element-wise multiplication operation; The formula for convolving the multi-scale feature tensor is as follows: ； In the formula, This represents the convolution tensor after convolution. This represents a two-dimensional convolution operation.
6. The construction method for embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 3, characterized in that, In step S2: The formula for the Wise-IoU loss function is: ； In the formula, This represents the Wise-IoU loss function value. This represents the intersection-union ratio (IoU) between the predicted bounding boxes and the labeled bounding boxes. The weights representing the intersection-union ratio; The intersection-union ratio (IUU) of the predicted bounding box and the labeled bounding box is obtained by the following formula: ； In the formula, This represents the area or number of pixels of the intersection region between the predicted bounding box and the ground truth bounding box. This represents the area or number of pixels of the union of the predicted bounding box and the ground truth bounding box. Indicates the prediction box. This represents the intersection operation. Represents the true bounding box. This represents the union operation; The formula for the rotation angle loss is: In the formula, This represents the value of the rotation angle loss function. Indicates the prediction angle. Indicates the actual angle; The formula for the confidence loss is: ； In the formula, This represents the confidence loss function value. Indicates the true label, Represents the natural logarithm function. This represents the confidence level of the predicted node bounding box; The formula for the total loss function is: In the formula, This represents the total loss function value. The weights represent the values of the Wise-IoU loss function. The weights represent the values of the rotation angle loss function. The weights represent the values of the confidence loss function.
7. The construction method for embedded connection nodes used in the hoisting of steel structures in high-rise buildings as described in claim 4, characterized in that, In step S2: The confidence level of the node frame is updated using the following formula: ； In the formula, Indicates the updated number The confidence level of each node frame. Indicates the first The original confidence of each node bounding box. The base of the natural logarithm. This represents the calculated intersection-union ratio. This indicates that the parameters are being adjusted.
8. The construction method for embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 1, characterized in that: Step S3 includes the following sub-steps: Based on the validation set, a standardized image sequence is extracted, and the coordinates of the center point of the image are solved using the image geometric center calculation method. Based on the detection results, the corresponding center point coordinates are extracted. Using the center point coordinates as a basis, the coordinate deviation is calculated using subtraction to obtain the positional deviation. , Indicates the deviation in the horizontal direction. Indicates deviation in the vertical direction; Obtain the ideal node orientation angle from the design drawings, and then use subtraction to calculate the deviation between the node orientation angle and the ideal node orientation angle, which is defined as the attitude deviation. ; Based on positional deviation and attitude deviation Calculate the control quantity; The control quantity is used as the controller of the hoisting equipment. After the controller of the hoisting equipment analyzes the control quantity, it instructs the hoisting equipment to make horizontal, vertical and rotational adjustments during construction. The control quantities are stored in the database, and a corresponding hoisting equipment ID is added to the control quantity during storage.
9. The construction method for embedded connection nodes for hoisting steel structures of high-rise buildings as described in claim 8, characterized in that, In the sub-steps of step S3: The formula for the control quantity is: ； In the formula, Indicates the control quantity. This represents the proportional gain coefficient for position control. This represents the proportional gain coefficient for attitude control.
10. A construction system for pre-embedded connection nodes in the hoisting of steel structures for high-rise buildings, characterized in that, include: The acquisition and partitioning module is used to acquire image sequences of steel structure hoisting, preprocess them, generate a dataset, partition the dataset to obtain training and validation tensors, and form a tensor set. The detection and adjustment module is used to build a detection model. It embeds deformable convolution and polarization self-attention mechanisms, takes training and validation tensors as input, and outputs node boxes and quintuples. The adjustment mechanism adjusts the node boxes to generate detection results with pre-embedded connected nodes. The calculation and storage module is used to calculate the position and attitude deviations of the pre-embedded connection nodes based on the detection results. Based on the position and attitude deviations, control quantities are generated and fed back to the hoisting equipment for construction adjustments. The control quantities are stored in the database.