A steel net rack construction precision prediction method based on full-component 3D scanning data

Abstract

The application relates to the technical field of construction prediction, and provides a steel net rack construction precision prediction method based on full-component 3D scanning data, which comprises the following steps: in the construction process of a steel net rack, point cloud data of an installed structure is collected by using a three-dimensional laser scanner, and a data sequence synchronous with the construction progress is formed; bolt ball nodes and rod assemblies in the point cloud are identified by using a random sampling consistency algorithm, actual coordinates of each node are calculated, and three-dimensional deviation vectors are obtained by comparing the actual coordinates with design coordinates; a structure state in the construction process is abstracted into a dynamic evolution graph data structure, and the size and features of the graph are continuously updated; a hybrid prediction model is established, wherein a graph convolution network is responsible for learning a spatial propagation mode of the deviation in a structure topology, a long short-term memory network is responsible for capturing an evolution law of the deviation with time, and the two can realize prediction of geometric deviation of a node to be installed in a next construction stage, so that a change from post-inspection to pre-prediction of construction quality control is realized.

Description

Technical Field

[0001] This invention relates to the field of construction prediction technology, and more specifically, to a method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components. Background Technology

[0002] Construction precision control of large-span steel space frame structures is an important research direction in the field of building engineering. Reference 1, "Automated progress tracking using 4Dschedule and 3D sensing technologies." (Yelda Turkan, et al. Automation in Construction, Volume 22, March 2012, Pages 414-421, https: / / doi.org / 10.1016 / j.autcon.2011.10.003), proposes a "Scan-vs-BIM" method. This technology geometrically compares high-density point cloud data obtained from 3D laser scanning with the building information model (BIM) from the design phase to generate a deviation cloud map, intuitively identifying areas exceeding tolerances. This achieves comprehensive inspection of construction quality, significantly improving efficiency and coverage compared to traditional total station sampling methods. However, the Scan-vs-BIM method has a fundamental flaw: it is essentially still a post-verification mode, which can only find problems after the components are installed, and cannot answer "where might go wrong next". Furthermore, due to the idealized representation characteristics of the BIM model, the precise deviation information contained in the real point cloud is oversimplified or lost during the comparison with the BIM model, forming an expressive bottleneck.

[0003] To address the problem of engineering prediction, reference 2, "Attention-based LSTM predictive model for the attitude and position of shield machine in tunneling," Kang Q, et al., UndergroundSpace, Volume 13, December 2023, Pages 335-350; DOI: 10.1016 / j.undsp.2023.05.006, successfully applies Long Short-Term Memory (LSTM) networks to the real-time prediction of shield machine attitude in tunnel engineering. This method utilizes the temporal modeling capability of LSTM networks to predict the continuous attitude changes of the shield machine during the tunneling process based on historical construction data, realizing the transformation from passive monitoring to active prediction, and significantly improving construction safety and efficiency. However, this type of prediction method is mainly aimed at continuous construction processes, while steel space frame construction is a discrete assembly process. Errors propagate in time and space through complex topological connections. Existing time series prediction models cannot effectively handle the spatial dependence and propagation law of geometric deviations in discrete structures, resulting in a significant gap in discrete assembly prediction.

[0004] Therefore, there is an urgent need in this field for a novel technical solution that can simultaneously address the reactive nature and expressive bottlenecks of existing Scan-vs-BIM methods, as well as the limitations of existing prediction methods in applying to discrete assembly processes. This solution would enable dynamic and proactive prediction of geometric deviations during steel space frame construction, thereby shifting quality control from post-confirmation to pre-confirmation. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, this invention provides a method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components. This method expresses the construction process as a dynamic graph characterized by node deviation vectors and uses a model combining graph convolutional networks and long short-term memory networks for prediction: the graph convolutional network analyzes the propagation law of errors in spatial topology, and the long short-term memory network learns its evolution law in time series. The combination of the two solves the prediction gap of discrete assembly and realizes the active prediction of geometric deviations.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components includes the following steps:

[0008] Step S1: After each discrete construction step is completed during the steel space frame construction process, the three-dimensional point cloud data of the currently installed structural part is obtained, thereby obtaining the point cloud dataset time series.

[0009] Step S2 involves processing the point cloud data to extract and calculate the 3D node deviation vector as a structured feature. This step specifically includes:

[0010] Step S21: The point cloud data is segmented using a random sampling consensus algorithm to identify and segment point cloud clusters representing bolt ball nodes and point cloud clusters representing rods.

[0011] Step S22: Calculate the three-dimensional coordinates of the center of the fitted sphere model of each identified node cloud cluster to obtain the measured node coordinates.

[0012] Step S23: Determine the measured node coordinates as the first node coordinates, and query the pre-stored building information model to obtain the design node coordinates of each node and determine them as the second node coordinates;

[0013] Step S24: Calculate the three-dimensional node deviation vector of each node by subtracting the coordinates of the second node from the coordinates of the first node.

[0014] Step S3: Based on the three-dimensional node deviation vector, construct a graph representing the current structural state in each construction step. The graph is defined by a node set, an edge set, and a node feature matrix. Each row of the node feature matrix stores the three-dimensional node deviation vector of the corresponding node, thereby generating a graph time series that dynamically evolves in size and node features over time.

[0015] Step S4: Input the graph time series into a pre-trained hybrid neural network model composed of graph convolutional network modules and long short-term memory network modules to predict the geometric deviation of the new node to be installed in the next construction step and output the predicted deviation vector of the new node to be installed.

[0016] As a further aspect of the present invention, in step S4, the graph convolutional network module in the hybrid neural network model receives the current state graph as input at each time step, aggregates the deviation vector information of neighboring nodes through a message passing mechanism, and aggregates the embedding vectors of all nodes through a graph pooling layer to output a graph embedding vector representing the global geometric deviation state of the entire structure at this time point.

[0017] As a further aspect of the present invention, the Long Short-Term Memory (LSTM) network module in the hybrid neural network model receives the graph embedding vector time series generated by the graph convolutional network module as its input sequence, and uses its gating mechanism to learn the time dependency of the global geometric deviation state in the sequence from one construction step to the next construction step; and outputs the final hidden state of the LSM network module at this time step, inputting it into a fully connected neural network prediction head, which outputs the prediction deviation vector of the new node to be installed.

[0018] At each time step, the graph convolutional network module in the hybrid neural network model receives a state graph representing the current structural state as input, specifically including receiving the node feature matrix and adjacency matrix of the graph. The graph convolutional network module extracts the spatial topological dependencies of node deviations by performing graph convolution operations. The steps are as follows: for each target node in the graph, firstly, all directly connected neighboring nodes are found according to the adjacency matrix; then, the deviation vector of each neighboring node is multiplied by a trainable shared weight matrix, and all multiplication results are aggregated (e.g., summed or averaged); finally, the aggregated result is processed by a non-linear activation function to generate a new generation of embedding vectors for the target node, which incorporates the deviation information of its local neighborhood. After one or more of the above graph convolution operations, a global mean pooling layer is used to calculate the element-wise average of the final embedding vectors of all nodes in the graph, thereby aggregating them into a single, fixed-dimensional graph embedding vector, which numerically represents the global geometric deviation state of the entire structure at this time point.

[0019] Subsequently, the Long Short-Term Memory (LSTM) network module in the hybrid neural network model receives a time series consisting of graph embedding vectors generated by the graph convolutional network module over all past time steps (from the initial state to the current time step) as its input. To learn the dynamic evolution of the global geometric deviation state from one construction step to the next in this sequence, the module performs internal operations including a forget gate, an input gate, and an output gate at each time step. Specifically, the module concatenates the hidden state vector from the previous time step with the graph embedding vector input at the current time step, and inputs the concatenated vector into three independent gating units parameterized by different weight matrices and bias vectors. The forget gate unit calculates a value between 0 and 1 using a sigmoid activation function to determine how much information to forget from the memory of the previous time step; the input gate unit determines how much new information to update the memory; and the output gate unit determines which information from the memory to use as the output at the current time step. By selectively forgetting memories from the previous moment and selectively combining them with new information from the current moment, the module updates its internal cellular state, thereby achieving effective memorization of the accumulated deviations and evolutionary patterns throughout the entire construction process.

[0020] Finally, after the Long Short-Term Memory (LSTM) network module processes the entire graph embedding vector time series, its final hidden state vector, output at the last time step, is input into a fully connected neural network prediction head. This final hidden state vector is a high-dimensional numerical summary of the entire spatiotemporal bias history up to the current time. The fully connected neural network prediction head consists of at least one linear layer, each containing its own trainable weight matrix and bias vector. The prediction head performs an affine transformation on the input final hidden state vector, ultimately outputting a three-dimensional vector containing three values. This three-dimensional vector represents the predicted magnitude of the geometric bias of the next node to be installed along the three coordinate axes.

[0021] As a further aspect of the present invention, step S3, the step of constructing a diagram representing the current structural state at each construction step, specifically includes:

[0022] Step S31: Determine that the node set is the set of all nodes that have been installed and identified in the current construction step;

[0023] Step S32: Determine the edge set as the set of identified physical rods that connect the nodes in the node set, and encode this connection relationship into an adjacency matrix;

[0024] Step S33: Construct the node feature matrix. The dimension of the matrix is ​​a number multiplied by three, where the number is the number of nodes in the node set. Each row of the matrix stores the three-dimensional node deviation vector of its corresponding node.

[0025] Step S34: After a new construction step is completed, the newly installed nodes and edges are incrementally added to the graph to form a new graph, thereby generating the time series of the graph.

[0026] As a further aspect of the present invention, step S34, which incrementally adds newly installed nodes and edges to the graph, specifically includes:

[0027] Step A1: Identify the newly installed nodes and edges in the new construction step;

[0028] Step A2: Perform a set union operation on the newly installed nodes identified and the node set determined in step S31 to form an updated node set, which serves as the node set of the new graph.

[0029] Step A3: Perform a set union operation on the newly installed edges identified and the edge set determined in step S32 to form an updated edge set, which is used as the edge set of the new graph.

[0030] Step A4: Calculate the three-dimensional node deviation vector of each of the newly installed nodes to form a new set of deviation vectors. Then, perform a matrix concatenation operation between the new set of deviation vectors and the node feature matrix constructed in step S33 to form an updated node feature matrix, which serves as the node feature matrix of the new graph.

[0031] As a further aspect of the present invention, in step S1, the step of obtaining the three-dimensional point cloud data of the currently installed structure specifically involves: after each preset discrete construction step is completed, using a ground-based three-dimensional laser scanner to perform a three-dimensional scan on all currently installed structural parts to obtain three-dimensional point cloud data.

[0032] As a further aspect of the present invention, the training steps of the hybrid neural network model include:

[0033] Step B1: Using a completed steel space frame project with complete construction phase scan data as the training set, execute steps S1-S3 to generate a complete graph time series and the corresponding true node deviation vector.

[0034] Step B2 involves using the mean square error between the predicted bias vector and the true node bias vector as the objective function, and performing end-to-end supervised training on all parameters of the graph convolutional network module and the long short-term memory network module through the backpropagation algorithm.

[0035] As a further aspect of the present invention, the random sampling consensus algorithm uses sphere model fitting to identify and segment point cloud clusters representing bolt sphere nodes, and uses cylinder model fitting to identify and segment point cloud clusters representing rod members.

[0036] As a further embodiment of the present invention, the graph pooling layer is a global mean pooling layer, which calculates the average value of all node embedding vectors output by the GCN module to obtain the graph embedding vector of the fixed dimension.

[0037] As a further aspect of the present invention, the message passing mechanism includes: for any node in the graph, weighted aggregation of the deviation vectors of all its neighboring nodes, then combining the aggregation result with the node's own deviation vector, and updating the node's embedding vector through a nonlinear transformation.

[0038] Compared with existing technologies, the beneficial effects of the steel space frame construction accuracy prediction method based on full component 3D scanning data of the present invention are as follows:

[0039] This invention provides a method for proactively predicting the construction accuracy of steel space frames. Compared with the existing "Scan-vs-BIM" method, this method realizes the transformation from post-event verification to pre-event prediction. It can quantitatively predict the possible positional deviations of components before installation, enabling the construction team to provide early warnings and take intervention measures. This transforms quality control from post-event remediation to pre-event prevention, thereby significantly reducing potential rework costs and project delays.

[0040] This invention employs a hybrid architecture combining graph convolutional networks (GCNNs) and long short-term memory (LSTM) networks. The GCNN learns the propagation patterns of deviations across the spatial topology through a message-passing mechanism, while the LTM learns the temporal evolution of deviation states across construction steps. This spatiotemporal decoupling and re-fusion modeling approach effectively addresses the complex characteristics of deviation propagation in steel space frame construction, which exhibit both spatial correlation and temporal dependence. Existing technologies either only consider time-series features while neglecting spatial relationships, or only analyze spatial distribution without considering temporal dynamics. The hybrid architecture of this method can simultaneously process information from both dimensions, improving the accuracy and reliability of predictions. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the construction accuracy prediction method for steel space frame based on 3D scanning data of all components according to the present invention.

[0042] Figure 2 This is a schematic diagram of the dynamic spatiotemporal map construction of a method for predicting the construction accuracy of steel space frame based on 3D scanning data of all components according to the present invention.

[0043] Figure 3 This is a comprehensive performance evaluation diagram of the steel structure construction deviation prediction model of the steel space frame construction accuracy prediction method based on full component 3D scanning data of the present invention.

[0044] Figure 4 This is a diagram illustrating the deviation characteristics of the steel structure construction process in the present invention, which is a method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components.

[0045] Figure 5 This is a scatter plot of the prediction accuracy of a steel space frame construction accuracy prediction method based on full component 3D scanning data according to the present invention. Detailed Implementation

[0046] The technical solutions of this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] Example 1

[0048] A method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components includes the following steps:

[0049] Step S1: After each discrete construction step is completed during the steel space frame construction process, the three-dimensional point cloud data of the currently installed structural part is obtained, thereby obtaining a time series of point cloud dataset that grows synchronously with the construction progress.

[0050] Step S2: Process the point cloud data to extract and calculate the three-dimensional node deviation vector as a structured feature;

[0051] Step S3: Based on the three-dimensional node deviation vector, construct a graph representing the current structural state at each construction step, and generate a graph time series that dynamically evolves over time in terms of size and node features.

[0052] Step S4: Input the graph time series into a pre-trained hybrid neural network model composed of graph convolutional network modules and long short-term memory network modules to predict the geometric deviation of the new node to be installed in the next construction step and output the predicted deviation vector of the new node to be installed.

[0053] In step S1 of this embodiment of the invention, the step of obtaining the three-dimensional point cloud data of the currently installed structure specifically involves: after each preset discrete construction step is completed, a ground-based three-dimensional laser scanner is used to perform a three-dimensional scan of all currently installed structural parts to obtain high-density three-dimensional point cloud data. This process is repeated as construction progresses, forming a time series of a point cloud dataset.

[0054] In this embodiment of the invention, the specific implementation of the method begins with data loading and preprocessing. This process aims to obtain the design coordinates and measured coordinates of the structure. In this embodiment, design geometric information is first obtained from a design blueprint containing 300 nodes and 740 edges. Simultaneously, 198 three-dimensional point cloud files representing different construction stages are loaded as data sources for subsequent analysis.

[0055] In step S2 of this embodiment of the invention, automated processing is performed on each collected point cloud data to extract structured features for subsequent modeling. This step specifically includes:

[0056] First, the point cloud data is segmented using the Random Sample Consensus (RANSAC) algorithm. RANSAC is a classic robust parameter estimation method in the field of computer vision. This embodiment of the invention directly applies its standard implementation, using a sphere model fitting to identify and segment point cloud clusters representing bolt sphere nodes, and simultaneously using a cylinder model fitting to identify and segment point cloud clusters representing rod members.

[0057] Then, for each identified cluster of node points, the measured node coordinates are obtained by calculating the three-dimensional coordinates of the center of its fitted spherical model. These coordinates are defined as the first node coordinates.

[0058] Next, the coordinates of the design node corresponding to each node are retrieved from the pre-stored Building Information Model (BIM) or Computer-Aided Design (CAD) file, and these coordinates are determined as the coordinates of the second node.

[0059] Finally, the 3D node deviation vector for each node is calculated. This is achieved by subtracting the three components of the design node coordinate vector from the three components (x, y, z) of the measured node coordinate vector. This calculation process ensures dimensional consistency. This deviation vector is the core input feature for subsequent modeling.

[0060] like Figure 2 The diagram illustrates the ideal design state of a steel space frame structure compared to its actual completed state. Due to various uncertainties during construction, the coordinates of nodes on the actual completed structure will spatially deviate from their theoretical coordinates in the original design blueprint. This method quantifies this spatial deviation into a three-dimensional node offset vector, obtained by subtracting the design coordinates from the measured coordinates of the node. This offset vector serves as the core digital feature of the corresponding node in the graphical model, providing a foundation for subsequent spatiotemporal deviation propagation modeling and prediction.

[0061] In step S3 of this embodiment of the invention, the structural state during construction is transformed into a graph structure time series that can be processed by a machine learning model. At each construction step, a graph representing the current structural state is constructed based on the features extracted in step S2.

[0062] The specific steps for constructing the graph are as follows:

[0063] First, the node set is defined as the set of all nodes that have been installed and identified in the current construction step;

[0064] Secondly, the edge set is determined to be the set of identified physical poles that connect the nodes in the node set, and this connection relationship is encoded into an adjacency matrix;

[0065] Then, a node feature matrix is ​​constructed, which has a dimension of N×3, where N is the number of nodes in this step. Each row of the matrix stores the three-dimensional node deviation vector of its corresponding node.

[0066] As construction progresses, upon completion of each new construction step, newly installed nodes and edges are incrementally added to the graph, forming a new graph. This incremental addition process involves: identifying newly installed nodes and edges; performing a union operation between the newly identified nodes and edges and the node and edge sets from the previous state, resulting in updated node and edge sets; calculating the deviation vector of the newly installed nodes and concatenating it with the node feature matrix from the previous state, forming an updated node feature matrix. This process generates a time series of the graph, which dynamically evolves in size and node features over time, accurately mapping the physical construction process.

[0067] This invention employs temporal interpolation and process reconstruction techniques to construct a high-fidelity dynamic evolution process that conforms to construction logic. Its construction follows a bottom-up installation sequence and a non-linear construction progress conforming to an S-curve. In generating this temporal diagram, rich features are designed for each node, including not only the core three-dimensional deviation vector but also the node's current three-dimensional spatial coordinates, its current time step number, standardized construction progress, and an environmental impact factor, ultimately forming a 9-dimensional high-dimensional feature vector. By processing 198 point cloud files as described above, 2772 samples suitable for model training are generated.

[0068] In step S4 of this embodiment of the invention, the graph time series generated in step S3 is input into a pre-trained hybrid neural network model.

[0069] The Graph Convolutional Network (GCN) module in this model learns the feature representation of each node in the structural space by aggregating information from neighboring nodes. This includes the following computational process:

[0070] First, the adjacency matrix is ​​standardized. This process involves adding a connection to itself for each node and normalizing it based on the number of connections for each node. The goal is to stably balance the information of the node itself and its neighbors when aggregating information.

[0071] Next, for each node in the graph, the current feature vectors of all its neighboring nodes (including the node itself) are weighted and summed according to the weights defined by the aforementioned standardized adjacency matrix to obtain an aggregated feature vector.

[0072] Finally, the aggregated feature vector is linearly transformed through a trainable weight matrix, and a non-linear activation function (such as the ReLU function) is applied to the transformation result to calculate the feature vector of the node in the new layer of the network.

[0073] In this embodiment of the invention, the GCN module consists of three stacked computational layers. Each layer performs the aforementioned computation, enabling the features of each node to incorporate bias information from its three-hop neighbors. The final output of the GCN module passes through a global mean pooling layer, which sums the final embedding vectors of all nodes bit-by-bit and calculates the average, forming a fixed-dimensional graph embedding vector representing the global geometric bias state of the entire structure at this point in time.

[0074] Subsequently, the Long Short-Term Memory (LSTM) module in the model receives the graph embedding vector time series generated by the GCN module as its input sequence. LSTM is a classic recurrent neural network, its core being the learning of long-term dependencies in the sequence through a gating mechanism consisting of input gates, forget gates, and output gates. These gating units work together to determine which information should be discarded from the memory units, which new information should be stored, and what to output at the current time step. This mechanism enables LSTM to effectively capture the dynamic patterns of deviation states evolving with construction steps.

[0075] Finally, the final hidden state output of the LSTM module at the last time step of the time series is input into a fully connected neural network prediction head, which outputs one or more three-dimensional vectors, each of which is the prediction bias vector for the new node to be installed in the next construction step.

[0076] The training steps of the hybrid neural network model include:

[0077] First, one or more completed steel space frame projects with complete construction phase scan data are used as the training set. Steps S1 to S3 are then performed on this training data to generate a complete graph time series and the corresponding true node deviation vectors for newly installed nodes in each step.

[0078] Then, with the objective function of minimizing the mean squared error between the predicted bias vector of the model output and the true node bias vector, the Adam optimizer is used to perform end-to-end supervised training on all parameters of the graph convolutional network module and the long short-term memory network module through the backpropagation algorithm.

[0079] In step S4 of this embodiment, the generated samples are input into a pre-trained hybrid neural network model. The specific structure of this model is as follows: a spatial feature extraction module consisting of a three-layer Graph Convolutional Network (GCN), a temporal dynamic learning module consisting of a one-layer Long Short-Term Memory Network (LSTM), and a nonlinear prediction head consisting of two fully connected layers. The entire model contains 173,379 trainable parameters. Its data processing flow is as follows: at each time step, the node features of the graph are first processed by the three-layer GCN to aggregate deviation information within the spatial neighborhood; subsequently, the output of the GCN is pooled into a graph embedding vector representing the global state; the time series composed of these vectors is fed into the LSTM to learn the evolution of deviation over time; finally, the final state of the LSTM is processed by the prediction head to output a prediction vector of the geometric deviation of the new node to be installed in the next construction step.

[0080] In this embodiment of the invention, the Adam optimizer is used for training, along with dynamic learning rate scheduling, gradient pruning, and early stopping (Patience=15) strategies to ensure efficient and stable training. In this embodiment, during the processing of construction data, training was automatically terminated early at the 29th epoch, achieving a minimum validation loss of 0.000003. Analysis of the construction data shows that the average construction deviation is 0.2 cm, while the maximum deviation reaches 39.7 cm.

[0081] In the model training phase of this embodiment of the invention, the output log is as follows: Epoch [ 25 / 100] | Train:0.000003 | Val: 0.000003 | LR: 3.50e-04 This indicates that in the 25th training epoch, the loss on both the training set and the validation set has been reduced to an extremely low level, and the learning rate at this time is 0.00035.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components, characterized in that, Includes the following steps: Step S1: After each discrete construction step is completed during the steel space frame construction process, the three-dimensional point cloud data of the currently installed structural part is obtained, thereby obtaining a time series of point cloud dataset that grows synchronously with the construction progress. Step S2 involves processing the point cloud data to extract and calculate the 3D node deviation vector as a structured feature. This step specifically includes: Step S21: The point cloud data is segmented using a random sampling consensus algorithm to identify and segment point cloud clusters representing bolt ball nodes and point cloud clusters representing rods. Step S22: Calculate the three-dimensional coordinates of the center of the fitted sphere model of each identified node cloud cluster to obtain the measured node coordinates. Step S23: Determine the measured node coordinates as the first node coordinates, and query the pre-stored building information model to obtain the design node coordinates of each node and determine them as the second node coordinates; Step S24: Calculate the three-dimensional node deviation vector of each node by subtracting the coordinates of the second node from the coordinates of the first node. Step S3: Based on the three-dimensional node deviation vector, construct a graph representing the current structural state in each construction step. The graph is defined by a node set, an edge set, and a node feature matrix. Each row of the node feature matrix stores the three-dimensional node deviation vector of the corresponding node, thereby generating a graph time series that dynamically evolves in size and node features over time. Step S4: Input the graph time series into a pre-trained hybrid neural network model composed of graph convolutional network modules and long short-term memory network modules to predict the geometric deviation of the new node to be installed in the next construction step and output the predicted deviation vector of the new node to be installed.

2. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 1, characterized in that, In step S4, the graph convolutional network module in the hybrid neural network model receives the current state graph as input at each time step, aggregates the deviation vector information of neighboring nodes through a message passing mechanism, and aggregates the embedding vectors of all nodes through a graph pooling layer to output a graph embedding vector representing the global geometric deviation state of the entire structure at this time point.

3. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 2, characterized in that, The Long Short-Term Memory (LSTM) network module in the hybrid neural network model receives the graph embedding vector time series generated by the graph convolutional network module as its input sequence. It uses its gating mechanism to learn the time dependency of the global geometric deviation state in the sequence from one construction step to the next. The LSM network module outputs the final hidden state at this time step and inputs it into a fully connected neural network prediction head. The fully connected neural network prediction head outputs the prediction deviation vector of the new node to be installed.

4. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 1, characterized in that, Step S3, the step of constructing a diagram representing the current structural state at each construction step, specifically includes: Step S31: Determine that the node set is the set of all nodes that have been installed and identified in the current construction step; Step S32: Determine the edge set as the set of identified physical members that connect nodes in the node set, and encode this connection relationship into an adjacency matrix; Step S33: Construct the node feature matrix. The dimension of the matrix is ​​a numerical value multiplied by three, where the numerical value is the number of nodes in the node set. Each row of the matrix stores the three-dimensional node deviation vector of its corresponding node. Step S34: After a new construction step is completed, the newly installed nodes and edges are incrementally added to the graph to form a new graph, thereby generating the time series of the graph.

5. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 4, characterized in that, The step S34, which incrementally adds newly installed nodes and edges to the graph, specifically includes: Step A1: Identify the newly installed nodes and edges in the new construction step; Step A2: Perform a set union operation on the newly installed nodes identified and the node set determined in step S31 to form an updated node set, which serves as the node set of the new graph. Step A3: Perform a set union operation on the newly installed edges identified and the edge set determined in step S32 to form an updated edge set, which is used as the edge set of the new graph. Step A4: Calculate the three-dimensional node deviation vector of each of the newly installed nodes to form a new set of deviation vectors. Then, perform a matrix concatenation operation between the new set of deviation vectors and the node feature matrix constructed in step S33 to form an updated node feature matrix, which serves as the node feature matrix of the new graph.

6. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 1, characterized in that, In step S1, the step of obtaining the three-dimensional point cloud data of the currently installed structure specifically involves: after each preset discrete construction step is completed, using a ground-based three-dimensional laser scanner to perform a three-dimensional scan of all currently installed structural parts to obtain three-dimensional point cloud data.

7. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 1, characterized in that, The training steps of the hybrid neural network model include: Step B1: Using a completed steel space frame project with complete construction phase scan data as the training set, execute steps S1-S3 to generate a complete graph time series and the corresponding true node deviation vector. Step B2 involves using the mean square error between the predicted bias vector and the true node bias vector as the objective function, and performing end-to-end supervised training on all parameters of the graph convolutional network module and the long short-term memory network module through the backpropagation algorithm.

8. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 1, characterized in that, The random sampling consensus algorithm uses a sphere model fitting to identify and segment point cloud clusters representing bolt sphere nodes, and a cylinder model fitting to identify and segment point cloud clusters representing rod members.

9. The method for predicting the construction accuracy of steel space frames based on 3D scanning data of all components according to claim 2, characterized in that, The graph pooling layer is a global mean pooling layer, which calculates the average value of all node embedding vectors output by the GCN module to obtain a graph embedding vector of fixed dimension.

10. The method for predicting the construction accuracy of a steel space frame based on 3D scanning data of all components according to claim 2, characterized in that, The message passing mechanism includes: for any node in the graph, weighted aggregation of the deviation vectors of all its neighboring nodes, then combining the aggregation result with the node's own deviation vector, and updating the node's embedding vector through a nonlinear transformation.

Images