Intelligent construction control method and system for bridge girder erection machine based on machine vision

Abstract

The invention provides a bridge girder erection machine intelligent construction control method and system based on machine vision, and relates to the technical field of intelligent control. Image information of a construction site is collected through a plurality of high-definition cameras arranged on a bridge girder erection machine; inputting the image information into a pre-trained deep learning model, identifying and positioning a target in the construction scene, and constructing a digital twinborn model of the construction scene in combination with a graph structure of a construction scene topological structure subjected to spatial perception graph convolution operation; the intelligent decision agent generates an initial construction strategy according to the current construction scene state, an improved Monte Carlo tree is adopted to search and optimize the initial construction strategy, and an optimal construction strategy is generated through multi-objective solution; and after the construction is completed, based on the digital twin model and the whole process data, analyzing the construction quality and efficiency by using a graph neural network, and generating a multi-dimensional evaluation report and optimization suggestions.

Description

Technical Field

[0001] The invention relates to intelligent control technology, and in particular to an intelligent construction control method and system for a bridge erecting machine based on machine vision. Background Art

[0002] Traditional bridge erection machine control mainly relies on manual operation and simple mechanical control systems, which have the following limitations:

[0003] First, the accuracy and consistency of manual operation are difficult to guarantee. In a complex construction environment, operators need to pay attention to multiple parameters and targets at the same time, which can easily lead to operational errors or misjudgment. Especially in the case of long-term continuous operation, operator fatigue will further increase the risk of errors.

[0004] Secondly, traditional mechanical control systems lack the ability to perceive the construction environment in real time. The environment that a bridge crane faces during construction is dynamically changing, including weather conditions, terrain changes, obstacles on the construction site, etc. Simple mechanical control systems have difficulty adapting to these changes and cannot make timely adjustments and responses.

[0005] Thirdly, it is difficult to achieve high-precision positioning and attitude control with traditional methods. In large-scale bridge construction, accurate positioning and attitude control are essential to ensure the correct installation of components. However, it is difficult to achieve the high-precision standards required by modern bridge engineering by relying solely on manual measurement and simple mechanical control. Summary of the invention

[0006] The embodiments of the present invention provide a machine vision-based intelligent construction control method and system for a bridge erection machine, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a machine vision-based intelligent construction control method for bridge erection machines, including:

[0009] The image information of the construction site is collected by multiple high-definition cameras installed on the bridge erection machine; the image information is input into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, a digital twin model of the construction scene is constructed;

[0010] Using the reinforcement learning algorithm, the intelligent decision-making agent is trained based on the digital twin model and the predefined reward and punishment function; the intelligent decision-making agent generates an initial construction strategy according to the current state of the construction scene, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates the optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine;

[0011] After the construction is completed, the construction quality and efficiency are analyzed using graph neural networks based on the digital twin model and full-process data to generate multi-dimensional evaluation reports and optimization suggestions.

[0012] In an optional embodiment,

[0013] The steps of collecting image information of the construction site by multiple high-definition cameras installed on the bridge erecting machine; inputting the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various parts of the bridge erecting machine, bridge components and the surrounding environment include:

[0014] The image information of the construction site is collected by multiple high-definition cameras arranged on the bridge erection machine; the image information is preprocessed by adaptive histogram equalization; the preprocessed image information is input into a pre-trained improved Mask R-CNN deep learning model, and a spatial pyramid pooling module is introduced into the feature extraction network of the improved Mask R-CNN deep learning model;

[0015] An attention mechanism module is added after the region of interest alignment layer of the improved Mask R-CNN deep learning model to highlight features and suppress irrelevant background information by generating spatial attention maps and channel attention maps;

[0016] Using the improved Mask R-CNN deep learning model, the targets in the construction scene are identified and located, including the components of the bridge erecting machine, bridge components and the surrounding environment;

[0017] The recognition results of the targets in the construction scene are compared with the pre-built special data set containing bridge erection machine parts and bridge components; the focal loss function is used as the loss function in the training process of the improved Mask R-CNN deep learning model, and the online case mining technology is introduced to dynamically select difficult samples for training;

[0018] The trained improved Mask R-CNN deep learning model is quantized and pruned to convert the 32-bit floating-point numbers in the model into 8-bit integers. The processed improved Mask R-CNN deep learning model is used to perform real-time target detection and recognition in the construction scene, and the target category, location, and contour information are output. The detection and recognition results are transmitted to the control system of the bridge erection machine.

[0019] In an optional embodiment,

[0020] Based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, the steps of constructing a digital twin model of the construction scene include:

[0021] Construct a graph structure representing the topological structure of the construction scene, wherein the nodes of the graph structure represent the identified targets or elements in the preset model, and the edges of the graph structure represent the spatial relationship between the nodes;

[0022] Initialize the node feature matrix based on the constructed graph structure;

[0023] Applying a spatially-aware graph convolution operation to the graph structure, the spatially-aware graph convolution operation comprising:

[0024] Calculate the Euclidean distance between nodes in the graph structure;

[0025] Based on the calculated Euclidean distance, the spatial distance weights between nodes are calculated using the Gaussian kernel function to obtain a spatial distance weight matrix;

[0026] Construct the adjacency matrix of the graph structure and add self-connection to the adjacency matrix;

[0027] Calculate the degree matrix of the adjacency matrix after adding self-connection;

[0028] Multiplying the spatial distance weight matrix with the adjacency matrix to obtain a weighted adjacency matrix;

[0029] Apply a weighted adjacency matrix to the node feature matrix and multiply it with a learnable weight matrix;

[0030] Apply a nonlinear activation function to the product result to obtain the updated node feature matrix;

[0031] Based on the updated node feature matrix, feature extraction is performed on the two-dimensional target recognition results and three-dimensional model data in the construction scene;

[0032] The extracted two-dimensional features and three-dimensional features are input into the multimodal feature fusion module, which uses a fully connected layer to process features and calculates fusion weights through an attention mechanism;

[0033] Perform weighted fusion on the processed features according to the fusion weights to obtain fusion features;

[0034] Based on the fused features and the graph structure after the spatial-aware graph convolution operation, a digital twin model of the construction scene is constructed.

[0035] In an optional embodiment,

[0036] The intelligent decision-making agent generates an initial construction strategy based on the current construction scene status, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates the optimal construction strategy through multi-objective solving. The steps include:

[0037] Using intelligent decision-making agents to generate initial construction strategies based on the current state of the construction scene;

[0038] An improved Monte Carlo tree search is used to optimize the initial construction strategy and generate an optimization result; the improved Monte Carlo tree search includes introducing a dynamic action generator, using a UCT algorithm with adaptive exploration parameters, realizing tree parallelization, and applying heuristic pruning;

[0039] In the UCT algorithm of the adaptive exploration parameter, a dynamically adjusted exploration parameter c is introduced:

[0040] c=c_base*(1+α*exp(-β*t / T));

[0041] Among them, c_base is the basic exploration parameter, α and β are adjustment factors, t is the current iteration number, and T is the total iteration number;

[0042] The optimization results of the initial construction strategy are solved for multiple objectives, including defining the objective functions of construction efficiency and energy consumption. An improved NSGA-III algorithm is used for multi-objective optimization. The improved NSGA-III algorithm includes adaptive crossover and mutation, local search and reference point adaptation. The optimal construction strategy is generated based on the multi-objective optimization solution results.

[0043] In an optional embodiment,

[0044] The improved NSGA-III algorithm includes the steps of adaptive crossover and mutation, local search and reference point adaptation, including:

[0045] Introducing an adaptive crossover and mutation mechanism based on population diversity, dynamically adjusting the crossover probability and mutation probability according to the current population diversity metric, which is calculated using the entropy weight method, and increasing the intensity of genetic operations when population diversity decreases;

[0046] After each iteration of the NSGA-III algorithm, the objective function values ​​of all individuals in the current population are evaluated, a non-dominated solution set is identified and constructed, multiple solutions are selected from the non-dominated solution set for local search, and multiple simulated annealing iterations are performed on each selected solution. The simulated annealing iteration adopts an exponentially decaying temperature schedule, and a neighborhood solution is generated in each iteration and the acceptance probability is calculated based on the energy difference. If the solution obtained by the local search dominates the original solution, the new solution replaces the original solution.

[0047] A reference point adaptive mechanism is adopted, including uniformly generating initial reference points in the standardized target space, periodically evaluating the correlation degree of each reference point during the evolution process, using the density peak clustering algorithm for low-correlation reference points to identify high-density areas in the target space, moving low-correlation reference points to the nearest high-density area while maintaining dispersion, and introducing regularization terms to limit the large-scale migration of reference points.

[0048] In an optional embodiment,

[0049] During the construction process, the data of various sensors of the bridge erecting machine are continuously collected, and the data of various sensors and the feedback data of the actuators are integrated to generate a state estimation of the bridge erecting machine. Based on the state estimation, the steps of closed-loop control of the bridge erecting machine using the nonlinear model predictive control algorithm include:

[0050] Using an inertial measurement unit, a global positioning system, a laser rangefinder, a strain sensor and an actuator encoder to collect state information of the bridge erecting machine, the state information including acceleration, angular velocity, position, relative distance, structural stress, hydraulic cylinder position and motor speed;

[0051] Constructing a state vector including position, acceleration, angular velocity and force / torque, and establishing a discrete-time nonlinear state equation and an observation equation based on the state vector;

[0052] An extended Kalman filter is used to perform state estimation on discrete-time nonlinear state equations and observation equations, including executing a prediction step and an update step, wherein the prediction step includes state prediction and covariance prediction, and the update step includes calculating Kalman gain, state update, and covariance update;

[0053] Based on the state estimation results, a nonlinear model predictive control optimization problem is constructed. The optimization objectives include tracking the predetermined trajectory, minimizing energy consumption, and satisfying safety constraints.

[0054] A real-time iteration strategy is adopted to solve the nonlinear model predictive control optimization problem, and the solution process is divided into a preparation phase and a feedback phase, wherein the preparation phase includes linearizing the system dynamics and constructing a quadratic programming subproblem, and the feedback phase includes updating the initial state, solving the quadratic programming subproblem and applying the control input;

[0055] Based on the solution results of nonlinear model predictive control, the optimal control input is generated, and the optimal control input is converted into the control signal of the bridge-building machine actuator to perform closed-loop control of the bridge-building machine.

[0056] In an optional embodiment,

[0057] After the construction is completed, the steps of analyzing the construction quality and efficiency using graph neural networks based on the digital twin model and the whole process data, and generating a multi-dimensional evaluation report and optimization suggestions include:

[0058] The bridge erection machine and the bridge structure are represented as a heterogeneous graph, where nodes represent components or monitoring points, edges represent physical connections or logical relationships between components, and each node and edge is accompanied by a multi-dimensional feature vector; a special feature transformation matrix is ​​applied to the multi-dimensional feature vectors of different types of nodes and edges to generate initial node and edge representations;

[0059] Based on the initial node and edge representation, an improved graph neural network model is constructed, which includes node-level attention mechanism and edge-level attention mechanism. Node-level attention is used to learn the importance weights between different features, and edge-level attention is used to learn the influence of different neighbor nodes.

[0060] A temporal convolutional network is introduced after the graph convolutional layer of the graph neural network model. Causal convolution and dilated convolution are used to capture temporal features and fuse them with the output of the graph convolutional layer.

[0061] Use historical data from the bridge erection machine construction system to train the graph neural network model, and iterate the training until convergence;

[0062] Use the trained graph neural network model to analyze the quality and efficiency of the new bridge-building machine construction system data and obtain the prediction results of various indicators. Based on the prediction results, generate a multidimensional evaluation report that includes overall score, analysis of various indicators, identification of abnormal events and their root cause analysis. According to the problems and potential risks identified in the multidimensional evaluation report, combined with historical construction experience, generate optimization suggestions and improvement measures.

[0063] According to a second aspect of the embodiments of the present invention,

[0064] Provide intelligent construction control system for bridge erection machine based on machine vision, including:

[0065] The first unit is used to collect image information of the construction site through multiple high-definition cameras installed on the bridge erection machine; input the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, a digital twin model of the construction scene is constructed;

[0066] The second unit is used to train an intelligent decision-making agent using a reinforcement learning algorithm based on a digital twin model and a predefined reward and punishment function; the intelligent decision-making agent generates an initial construction strategy based on the current state of the construction scene, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates an optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine;

[0067] The third unit is used to analyze the construction quality and efficiency after the construction is completed, based on the digital twin model and the data of the whole process, using the graph neural network to generate a multi-dimensional evaluation report and optimization suggestions.

[0068] According to a third aspect of the embodiments of the present invention,

[0069] An electronic device is provided, comprising:

[0070] processor;

[0071] a memory for storing processor-executable instructions;

[0072] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0073] A fourth aspect of the embodiments of the present invention is:

[0074] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0075] The construction strategy generation method based on intelligent decision-making agent, improved Monte Carlo tree search and multi-objective optimization in this paper has many beneficial effects. First, it can automatically generate high-quality construction strategies according to the complex construction scene status, greatly reducing the burden of manual decision-making. Secondly, the improved MCTS significantly improves the efficiency and quality of strategy optimization through technologies such as dynamic action generation, adaptive exploration and parallelization. Furthermore, multi-objective optimization takes into account multiple objectives such as construction efficiency and energy consumption, making the generated strategy more comprehensive and practical. Finally, the adaptability and flexibility of this method enable it to adapt to various complex construction environments and requirements, providing strong technical support for intelligent and refined bridge crane construction management.

[0076] The graph neural network-based bridge-erecting machine construction quality and efficiency analysis method adopted in this paper has many beneficial effects. First, it can comprehensively consider the complex topological relationship between the bridge-erecting machine and the bridge structure, capture the mutual influence between components, and provide more accurate analysis results than traditional methods. Secondly, the introduced attention mechanism and temporal convolutional network enable the model to adaptively focus on important features and long-term dependencies, improving the accuracy and robustness of the analysis. Furthermore, this method can achieve end-to-end learning, directly derive evaluation results from raw data, and reduce the interference of human factors. Finally, the generated multi-dimensional evaluation report and optimization suggestions provide intuitive and comprehensive decision-making support for construction managers, which helps to improve construction quality, efficiency and safety. The application of this method can not only optimize current construction projects, but also promote technological progress in the entire bridge construction industry by accumulating and analyzing large amounts of data, continuously improving construction technology and management strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of an intelligent construction control method of a bridge erecting machine based on machine vision according to an embodiment of the present invention;

[0078] Figure 2 Schematic diagram of the structure of the bridge erecting machine according to an embodiment of the present invention, in which:

[0079] 1. Right end beam assembly; 2. Front support system; 3. Transport vehicle; 4. Crossbeam; 5. Camera; 6. Overhead crane system; 8. Middle support bracket; 9. Auxiliary bracket; 10. Front support system; 11. Left end beam assembly; 12. Main beam;

[0080] Figure 3 It is a structural schematic diagram of an intelligent construction control system for a bridge erecting machine based on machine vision according to an embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0083] Figure 1FIG. 1 is a flow chart of an intelligent construction control method for a bridge erecting machine based on machine vision according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0084] S1. Collect image information of the construction site through multiple high-definition cameras installed on the bridge erection machine; input the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, construct a digital twin model of the construction scene;

[0085] S2. Using the reinforcement learning algorithm, based on the digital twin model and the predefined reward and punishment function, the intelligent decision-making agent is trained; the intelligent decision-making agent generates an initial construction strategy according to the current state of the construction scene, optimizes the initial construction strategy using the improved Monte Carlo tree search, and generates the optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine;

[0086] S3. After the construction is completed, the construction quality and efficiency are analyzed using graph neural networks based on the digital twin model and the data of the entire process, and a multi-dimensional evaluation report and optimization suggestions are generated.

[0087] In an optional embodiment,

[0088] The steps of collecting image information of the construction site by multiple high-definition cameras installed on the bridge erecting machine; inputting the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various parts of the bridge erecting machine, bridge components and the surrounding environment include:

[0089] The image information of the construction site is collected by multiple high-definition cameras arranged on the bridge erection machine; the image information is preprocessed by adaptive histogram equalization; the preprocessed image information is input into a pre-trained improved Mask R-CNN deep learning model, and a spatial pyramid pooling module is introduced into the feature extraction network of the improved Mask R-CNN deep learning model;

[0090] An attention mechanism module is added after the region of interest alignment layer of the improved Mask R-CNN deep learning model to highlight features and suppress irrelevant background information by generating spatial attention maps and channel attention maps;

[0091] Using the improved Mask R-CNN deep learning model, the targets in the construction scene are identified and located, including the components of the bridge erecting machine, bridge components and the surrounding environment;

[0092] The recognition results of the targets in the construction scene are compared with the pre-built special data set containing bridge erection machine parts and bridge components; the focal loss function is used as the loss function in the training process of the improved Mask R-CNN deep learning model, and the online case mining technology is introduced to dynamically select difficult samples for training;

[0093] The trained improved Mask R-CNN deep learning model is quantized and pruned to convert the 32-bit floating-point numbers in the model into 8-bit integers. The processed improved Mask R-CNN deep learning model is used to perform real-time target detection and recognition in the construction scene, and the target category, location, and contour information are output. The detection and recognition results are transmitted to the control system of the bridge erection machine.

[0094] For example, the improved Mask R-CNN-based method for target recognition in bridge erection machine construction scenes is an advanced computer vision technology that aims to improve the safety and efficiency of bridge erection machine construction. The method achieves high-precision target detection and recognition through multiple steps.

[0095] First, multiple high-definition cameras are installed on the bridge erector to ensure full coverage of the construction site. These cameras usually use 4K resolution and a frame rate of no less than 30fps to capture high-quality image information. The camera installation locations include but are not limited to key parts such as the main beam, cross beam, and hoist to obtain images from different angles and fields of view.

[0096] The collected original image is preprocessed by adaptive histogram equalization. This step is crucial for processing images under complex lighting conditions. Specifically, the image is divided into 8x8 blocks, each block is histogram equalized separately, and then the results of adjacent blocks are merged using bilinear interpolation. This method can effectively enhance local contrast, so that targets in shadow areas or bright light areas can also be clearly identified.

[0097] The preprocessed images are input to the improved Mask R-CNN deep learning model. The main network of this model adopts the ResNet-101 architecture, and introduces the spatial pyramid pooling module on its basis. The spatial pyramid pooling module contains four scales of maximum pooling operations: 1x1, 2x2, 4x4, and 8x8. Each pooling is followed by a 1x1 convolution layer, and finally all outputs are concatenated. This design can effectively capture multi-scale features and improve the model's ability to detect objects of different sizes.

[0098] After the ROI alignment layer, an attention mechanism module is added. The spatial attention mechanism highlights important spatial location information by generating a 2D attention map. In the specific implementation, the feature map is average pooled and max pooled, and then the results are concatenated and passed through a 7x7 convolutional layer to obtain the spatial attention map. The channel attention mechanism emphasizes important feature channels by generating 1D vectors. In the implementation, the feature map is globally average pooled and globally max pooled, and then the channel attention vector is obtained through two fully connected layers. The combination of these two attention mechanisms can effectively highlight important features and suppress irrelevant background information.

[0099] The focal loss function is used as the loss function during the model training process. The focal loss function solves the problem of class imbalance by adjusting the weights of difficult and easy samples. In practical applications, the alpha parameter is set to 0.25 and the gamma parameter is set to 2.0. In addition, online case mining technology is introduced to dynamically select difficult samples for training. Specifically, in each mini-batch, the loss of all samples is calculated, and the top 33% of samples with the highest loss are selected as cases, and only these samples are used to update the model parameters.

[0100] After training is completed, the model is quantized and pruned. The quantization process converts 32-bit floating point numbers into 8-bit integers. The specific steps include determining the quantization range, calculating the quantization parameters, quantizing weights and activation values, and recalibrating.

[0101] The optimized model is used to detect and identify targets in real time in construction scenes. The model output includes the target category (such as main beam, cross beam, hanger, etc.), location (expressed in bounding box coordinates) and contour information (expressed in pixel-level mask). This information is transmitted to the control system of the bridge crane via a high-speed network to achieve real-time monitoring and auxiliary decision-making.

[0102] This paper significantly improves the safety of the bridge crane construction process through high-precision target detection and recognition, and can detect potential dangers and issue warnings in a timely manner. Secondly, real-time monitoring and analysis improve construction efficiency and optimize process flow. Thirdly, the optimization of the model enables it to run on embedded systems, realizes edge computing, and reduces data transmission delays.

[0103] In an optional embodiment,

[0104] Based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, the steps of constructing a digital twin model of the construction scene include:

[0105] Construct a graph structure representing the topological structure of the construction scene, wherein the nodes of the graph structure represent the identified targets or elements in the preset model, and the edges of the graph structure represent the spatial relationship between the nodes;

[0106] Initialize the node feature matrix based on the constructed graph structure;

[0107] Applying a spatially-aware graph convolution operation to the graph structure, the spatially-aware graph convolution operation comprising:

[0108] Calculate the Euclidean distance between nodes in the graph structure;

[0109] Based on the calculated Euclidean distance, the spatial distance weights between nodes are calculated using the Gaussian kernel function to obtain a spatial distance weight matrix;

[0110] Construct the adjacency matrix of the graph structure and add self-connection to the adjacency matrix;

[0111] Calculate the degree matrix of the adjacency matrix after adding self-connection;

[0112] Multiplying the spatial distance weight matrix with the adjacency matrix to obtain a weighted adjacency matrix;

[0113] Apply a weighted adjacency matrix to the node feature matrix and multiply it with a learnable weight matrix;

[0114] Apply a nonlinear activation function to the product result to obtain the updated node feature matrix;

[0115] Based on the updated node feature matrix, feature extraction is performed on the two-dimensional target recognition results and three-dimensional model data in the construction scene;

[0116] The extracted two-dimensional features and three-dimensional features are input into the multimodal feature fusion module, which uses a fully connected layer to process features and calculates fusion weights through an attention mechanism;

[0117] Perform weighted fusion on the processed features according to the fusion weights to obtain fusion features;

[0118] Based on the fused features and the graph structure after the spatial-aware graph convolution operation, a digital twin model of the construction scene is constructed.

[0119] For example, the construction scene digital twin model construction method based on target recognition results and spatial perception graph convolution is an innovative technical solution that aims to accurately restore the complex bridge crane construction environment. This method integrates advanced technologies such as computer vision, graph neural network and multimodal feature fusion to achieve efficient conversion from two-dimensional images to three-dimensional digital twin models.

[0120] First, a graph structure is constructed to represent the topological structure of the construction scene. The nodes in the graph structure represent the elements in the identified targets or preset models, such as main beams, cross beams, support frames, etc. The edges of the graph structure represent the spatial relationships between nodes, such as adjacent, connected, or supported. For example, in a typical bridge crane construction scene, there may be 20 nodes, representing the main structural components and surrounding environment elements.

[0121] Next, initialize the node feature matrix. The initial features of each node may include its location coordinates, size, category, etc. For a graph structure containing 20 nodes, the dimension of the initial feature matrix may be 20x64, where 64 is the feature dimension of each node.

[0122] Then, a spatially aware graph convolution operation is applied to the graph structure. The core of this operation is to consider the spatial distance information between nodes, so as to more accurately capture the relationship between the elements in the construction scene. In the specific implementation, the Euclidean distance between the nodes in the graph structure is first calculated. For example, for two nodes, whose three-dimensional coordinates are (1,2,3) and (4,5,6), the Euclidean distance between them is about 5.2.

[0123] Based on the calculated Euclidean distance, the Gaussian kernel function is used to calculate the spatial distance weights between nodes. The Gaussian kernel function is in the form of an exponential function, where the distance is the negative part of the exponent and the sigma parameter controls the decay rate of the function. Usually sigma is set to half of the average node distance. This step results in a spatial distance weight matrix with a matrix size of 20x20 (assuming there are 20 nodes).

[0124] Next, we construct the adjacency matrix of the graph structure and add self-connection to the adjacency matrix. The adjacency matrix is ​​a 20x20 matrix (assuming there are 20 nodes), where an element of 1 indicates that two nodes are connected, and an element of 0 indicates that they are not connected. Adding self-connection means setting the diagonal elements of the adjacency matrix to 1.

[0125] Calculate the degree matrix of the adjacency matrix after adding self-connection. The degree matrix is ​​a diagonal matrix, and the elements on the diagonal represent the degree of the corresponding node (the number of edges connected to it). For example, if a node is connected to 3 other nodes, then the corresponding diagonal element in the degree matrix is ​​4 (including self-connection).

[0126] Multiply the spatial distance weight matrix with the adjacency matrix to obtain the weighted adjacency matrix. This step introduces spatial distance information into the graph structure, making the connection weight between nodes that are closer in space larger.

[0127] A weighted adjacency matrix is ​​applied to the node feature matrix and multiplied by a learnable weight matrix. The size of the learnable weight matrix depends on the input feature dimension and the desired output feature dimension, for example it may be a 64x128 matrix, mapping the node features from 64 dimensions to 128 dimensions.

[0128] Apply a nonlinear activation function, such as the ReLU function, to the product result to obtain the updated node feature matrix. This step introduces nonlinearity and enhances the expressiveness of the model. The dimension of the updated node feature matrix may become 20x128.

[0129] Based on the updated node feature matrix, feature extraction is performed on the 2D target recognition results and 3D model data in the construction scene. 2D features may include the bounding box coordinates and category probability of the target, and 3D features may include point cloud data and volume information.

[0130] The extracted 2D and 3D features are input into the multimodal feature fusion module. This module uses a fully connected layer to process the 2D and 3D features separately, and then calculates the fusion weights through the attention mechanism. The attention mechanism can adaptively adjust the importance of different modal features. For example, in some cases, it may rely more on 2D features, while in other cases, it may rely more on 3D features.

[0131] The processed features are weighted and fused according to the calculated fusion weights to obtain fusion features. The fusion features integrate the two-dimensional image information and the three-dimensional structure information, providing rich input for building an accurate digital twin model.

[0132] Finally, based on the fusion features and the graph structure after the spatial-aware graph convolution operation, a digital twin model of the construction scene is constructed. This digital twin model not only contains the geometric information of each component, but also the topological and spatial relationships between them, which can accurately reflect the status of the actual construction scene.

[0133] This method of constructing a digital twin model based on spatial-aware graph convolution and multimodal feature fusion has many beneficial effects. First, it can more accurately capture the spatial relationship between elements in the construction scene and improve the geometric accuracy of the model. Secondly, by fusing two-dimensional and three-dimensional features, this method makes full use of information from different modalities and enhances the robustness and adaptability of the model. Furthermore, the spatial-aware graph convolution operation enables the model to effectively handle large-scale and complex construction scenes with good scalability. Finally, the constructed digital twin model provides a solid foundation for subsequent construction monitoring, safety warning and intelligent decision-making, which helps to improve the efficiency and safety of bridge crane construction.

[0134] In an optional embodiment,

[0135] The intelligent decision-making agent generates an initial construction strategy based on the current construction scene status, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates the optimal construction strategy through multi-objective solving. The steps include:

[0136] Using intelligent decision-making agents to generate initial construction strategies based on the current state of the construction scene;

[0137] An improved Monte Carlo tree search is used to optimize the initial construction strategy and generate an optimization result; the improved Monte Carlo tree search includes introducing a dynamic action generator, using a UCT algorithm with adaptive exploration parameters, realizing tree parallelization, and applying heuristic pruning;

[0138] In the UCT algorithm of the adaptive exploration parameter, a dynamically adjusted exploration parameter c is introduced:

[0139] c=c_base*(1+α*exp(-β*t / T));

[0140] Among them, c_base is the basic exploration parameter, α and β are adjustment factors, t is the current iteration number, and T is the total iteration number;

[0141] The optimization results of the initial construction strategy are solved for multiple objectives, including defining the objective functions of construction efficiency and energy consumption. An improved NSGA-III algorithm is used for multi-objective optimization. The improved NSGA-III algorithm includes adaptive crossover and mutation, local search and reference point adaptation. The optimal construction strategy is generated based on the multi-objective optimization solution results.

[0142] Exemplarily, the method of generating the optimal construction strategy by combining an intelligent decision-making agent with an improved Monte Carlo tree search and a multi-objective optimization algorithm is an advanced artificial intelligence technology that aims to improve the efficiency and safety of bridge erection machine construction. This method integrates multiple technologies such as reinforcement learning, heuristic search, and evolutionary algorithms to achieve intelligent conversion from the current construction scene state to the optimal construction strategy.

[0143] First, an intelligent decision-making agent is used to generate an initial construction strategy based on the current state of the construction scene. The intelligent decision-making agent is trained using a deep reinforcement learning model, such as the Proximal Policy Optimization algorithm. The agent receives state information of the construction scene, including the position, posture, and construction progress of each component, and outputs a series of initial construction actions. For example, for a typical bridge crane construction scene, the initial strategy may include 20 consecutive construction actions, each of which includes specific parameters such as main beam movement, crossbeam adjustment, and hoist operation.

[0144] Next, the initial construction strategy is optimized using an improved Monte Carlo Tree Search (MCTS). The improved MCTS contains multiple innovations. First, a dynamic action generator is introduced. This generator dynamically generates a set of feasible actions based on the current state, rather than using a fixed action space. For example, in some states, only some actions may be safe or effective, and the dynamic generator will adjust the optional actions accordingly.

[0145] The UCT (game tree search) algorithm with adaptive exploration parameters is used in MCTS. The core of the UCT algorithm is to balance exploration and exploitation, in which the exploration parameter c plays a key role. In this method, the c value is dynamically adjusted as the search progresses:

[0146] c=c_base*(1+α*exp(-β*t / T));

[0147] Among them, c_base is the basic exploration parameter, α and β are adjustment factors, t is the current iteration number, and T is the total iteration number.

[0148] Specifically, the calculation of the c value takes into account the basic exploration parameters, the current number of iterations, and the total number of iterations. For example, assuming the basic exploration parameter is 1.4, the adjustment factor α is 0.5, β is 2, and the total number of iterations is 1000, then at the 100th iteration, the c value is about 1.9, and at the 900th iteration, the c value is about 1.45. This dynamic adjustment makes the algorithm more inclined to exploration in the early stage of the search and more focused on exploitation in the later stage.

[0149] Another improvement of MCTS is to implement tree parallelism. Taking advantage of multi-core processors, multiple search trees are expanded at the same time, each responsible for a different sequence of actions. This greatly improves the search efficiency and enables a larger decision space to be explored within a limited time. For example, on an 8-core processor, 8 search trees can be expanded at the same time, each exploring a different initial action.

[0150] In addition, MCTS also applies heuristic pruning technology. Based on domain knowledge and historical data, some obviously suboptimal or unsafe action sequences are pruned to reduce the search space. For example, if an action sequence causes a component collision, the sequence and its subsequences will be directly pruned and the search will not continue.

[0151] The optimization result of MCTS is a series of improved construction action sequences, which usually better balance efficiency and safety than the initial strategy.

[0152] The results of MCTS optimization are solved by multi-objectives to further optimize the construction strategy. First, multiple objective functions are defined. Typical objectives include construction efficiency and energy consumption. Construction efficiency can be quantified by the time required to complete a specific construction task, while energy consumption can be measured by the accumulated electricity or fuel consumption.

[0153] The multi-objective optimization adopts the improved NSGA-III (Non-dominated Sorting Genetic Algorithm III) algorithm. The improvement of the algorithm includes three aspects: adaptive crossover and mutation, local search and reference point adaptation. Adaptive crossover and mutation dynamically adjust the crossover rate and mutation rate according to the population diversity. For example, when the population diversity is low, the mutation rate is increased to improve the exploration ability; when the diversity is high, the crossover rate is increased to enhance the utilization.

[0154] The local search strategy fine-tunes some excellent individuals after each generation of evolution. Specifically, the simulated annealing algorithm can be used to locally optimize the selected individuals to further improve the quality of the solution. The reference point adaptation dynamically adjusts the position and density of the reference point according to the current population distribution to ensure that the algorithm can evenly explore the entire Pareto frontier.

[0155] Through multi-objective optimization, a set of non-dominated solutions, namely the Pareto optimal solution set, is finally obtained. Each solution represents a construction strategy that achieves a different balance between construction efficiency and energy consumption. Decision makers can choose the most suitable strategy from this set of solutions according to specific needs. For example, under time constraints, a strategy with higher efficiency but slightly higher energy consumption may be selected; while in conventional construction, a strategy with a more balanced efficiency and energy consumption may be selected.

[0156] The construction strategy generation method based on intelligent decision-making agent, improved Monte Carlo tree search and multi-objective optimization in this paper has many beneficial effects. First, it can automatically generate high-quality construction strategies according to the complex construction scene status, greatly reducing the burden of manual decision-making. Secondly, the improved MCTS significantly improves the efficiency and quality of strategy optimization through technologies such as dynamic action generation, adaptive exploration and parallelization. Furthermore, multi-objective optimization takes into account multiple objectives such as construction efficiency and energy consumption, making the generated strategy more comprehensive and practical. Finally, the adaptability and flexibility of this method enable it to adapt to various complex construction environments and requirements, providing strong technical support for intelligent and refined bridge crane construction management.

[0157] In an optional embodiment,

[0158] The improved NSGA-III algorithm includes the steps of adaptive crossover and mutation, local search and reference point adaptation, including:

[0159] Introducing an adaptive crossover and mutation mechanism based on population diversity, dynamically adjusting the crossover probability and mutation probability according to the current population diversity metric, which is calculated using the entropy weight method, and increasing the intensity of genetic operations when population diversity decreases;

[0160] After each iteration of the NSGA-III algorithm, the objective function values ​​of all individuals in the current population are evaluated, a non-dominated solution set is identified and constructed, multiple solutions are selected from the non-dominated solution set for local search, and multiple simulated annealing iterations are performed on each selected solution. The simulated annealing iteration adopts an exponentially decaying temperature schedule, and a neighborhood solution is generated in each iteration and the acceptance probability is calculated based on the energy difference. If the solution obtained by the local search dominates the original solution, the new solution replaces the original solution.

[0161] A reference point adaptive mechanism is adopted, including uniformly generating initial reference points in the standardized target space, periodically evaluating the correlation degree of each reference point during the evolution process, using the density peak clustering algorithm for low-correlation reference points to identify high-density areas in the target space, moving low-correlation reference points to the nearest high-density area while maintaining dispersion, and introducing regularization terms to limit the large-scale migration of reference points.

[0162] For example, the improved NSGA-III algorithm is an efficient multi-objective optimization method that introduces innovative mechanisms such as adaptive crossover and mutation, local search, and reference point adaptation for complex bridge construction strategy optimization problems. These improvements are designed to improve the convergence speed of the algorithm and the quality of the solution while maintaining the diversity and uniform distribution of the population.

[0163] First, an adaptive crossover and mutation mechanism based on population diversity is introduced. This mechanism dynamically adjusts the intensity of genetic operations to adapt to the evolutionary state of the population. Specifically, the entropy weight method is used to calculate the diversity measure of the population. The entropy weight method takes into account the distribution of each decision variable in the population, calculates the information entropy of each variable, and then comprehensively obtains the overall diversity index. For example, suppose there is a population of 100 individuals, each of which has 10 decision variables. For each decision variable, its value range is divided into 10 intervals, the number of individuals falling in each interval is counted, and the information entropy of the variable is calculated. Then the information entropy of all variables is weighted averaged to obtain the diversity measure of the population.

[0164] According to the calculated diversity metric, the crossover probability and mutation probability are dynamically adjusted. When the diversity metric is lower than the preset threshold (such as 0.6), the mutation probability is increased to improve the exploration ability of the population. For example, the basic mutation probability of 0.01 can be increased to 0.05. At the same time, the crossover probability is appropriately reduced, such as from 0.9 to 0.8, to reduce the damage to existing genes. On the contrary, when the diversity metric is higher than another threshold (such as 0.8), the crossover probability is increased and the mutation probability is reduced to enhance the utilization of existing excellent genes.

[0165] After each iteration of the NSGA-III algorithm, a local search mechanism is introduced to further improve the quality of the solution. First, the objective function values ​​of all individuals in the current population are evaluated, and a non-dominated solution set is constructed by non-dominated sorting. From this non-dominated solution set, multiple (e.g., 10) representative solutions are selected for local search. The selection criteria can be based on the distribution of solutions in the target space to ensure that different areas of the Pareto front are covered.

[0166] Multiple simulated annealing iterations are performed for each selected solution. The simulated annealing algorithm uses an exponentially decaying temperature schedule, with the initial temperature set to 100 and the temperature decreasing by 1% for each iteration. In each iteration, a neighborhood solution of the current solution is generated by a small random perturbation. For example, for a solution containing 20 decision variables, 1-3 variables can be randomly selected and adjusted within ±5% of their current values ​​to generate new candidate solutions.

[0167] The energy difference is calculated based on the objective function value of the new solution and the current solution, and the acceptance probability is calculated based on the energy difference and the current temperature. If the new solution is better than the current solution, it is directly accepted; otherwise, the inferior solution is accepted with a certain probability, and this probability decreases as the temperature decreases. Through this mechanism, the algorithm can jump out of the local optimum in the local search. Perform 100 simulated annealing iterations for each selected solution. If the new solution obtained by the local search is better than the original solution in any objective and is not inferior to the original solution in other objectives (that is, the new solution dominates the original solution), the new solution replaces the original solution.

[0168] In order to better guide the population to converge to the Pareto frontier and maintain the diversity of solutions, a reference point adaptive mechanism is adopted. First, the initial reference points are uniformly generated in the standardized target space. For example, for a three-objective optimization problem, 100 uniformly distributed reference points can be generated on the unit hyperplane.

[0169] During the evolution process, the degree of association of each reference point is evaluated periodically (e.g., every 10 generations). The degree of association is measured by calculating the number of individuals closest to each reference point. For reference points with an association below a threshold (e.g., 3), they are considered to be in a low-density area and need to be adjusted.

[0170] Use the density peak clustering algorithm to identify high-density areas in the target space. The algorithm first calculates the local density and distance to high-density points for each solution, and then identifies density peak points, which have both high local density and are far away from other high-density points. These density peak points represent high-density areas in the target space.

[0171] Move low-correlation reference points to the nearest high-density area. When moving, it is necessary to maintain the dispersion between reference points to avoid multiple reference points gathering in the same area. Specifically, the reference points can be moved to the edge of the high-density area instead of the center to maintain the coverage. At the same time, a regularization term is introduced to limit the large-scale migration of reference points. For example, the maximum moving distance of each adjustment can be limited to no more than 20% of the average distance between reference points.

[0172] The improved NSGA-III algorithm adopted in this paper has many beneficial effects in the optimization of bridge-building machine construction strategy. First, the adaptive crossover and mutation mechanism can dynamically adjust the genetic operation intensity according to the population state, effectively balance the exploration and utilization capabilities of the algorithm, and accelerate the convergence speed. Secondly, the local search strategy fine-tunes the excellent solution through the simulated annealing algorithm, significantly improving the quality of the solution, especially in complex decision spaces. The better solution can be found. Furthermore, the reference point adaptive mechanism enables the algorithm to better adapt to the distribution characteristics of the target space, improve the diversity and uniformity of the solution, and provide decision makers with a more comprehensive Pareto optimal solution set. Finally, these improvements enable the algorithm to perform well in dealing with high-dimensional, multi-objective construction strategy optimization problems, and can quickly generate high-quality and diversified construction plans under limited computing resources, providing strong decision support for intelligent and refined bridge-building machine construction management.

[0173] In an optional embodiment,

[0174] During the construction process, the data of various sensors of the bridge erecting machine are continuously collected, and the data of various sensors and the feedback data of the actuators are integrated to generate a state estimation of the bridge erecting machine. Based on the state estimation, the steps of closed-loop control of the bridge erecting machine using the nonlinear model predictive control algorithm include:

[0175] Using an inertial measurement unit, a global positioning system, a laser rangefinder, a strain sensor and an actuator encoder to collect state information of the bridge erecting machine, the state information including acceleration, angular velocity, position, relative distance, structural stress, hydraulic cylinder position and motor speed;

[0176] Constructing a state vector including position, acceleration, angular velocity and force / torque, and establishing a discrete-time nonlinear state equation and an observation equation based on the state vector;

[0177] An extended Kalman filter is used to perform state estimation on discrete-time nonlinear state equations and observation equations, including executing a prediction step and an update step, wherein the prediction step includes state prediction and covariance prediction, and the update step includes calculating Kalman gain, state update, and covariance update;

[0178] Based on the state estimation results, a nonlinear model predictive control optimization problem is constructed. The optimization objectives include tracking the predetermined trajectory, minimizing energy consumption, and satisfying safety constraints.

[0179] A real-time iteration strategy is adopted to solve the nonlinear model predictive control optimization problem, and the solution process is divided into a preparation phase and a feedback phase, wherein the preparation phase includes linearizing the system dynamics and constructing a quadratic programming subproblem, and the feedback phase includes updating the initial state, solving the quadratic programming subproblem and applying the control input;

[0180] Based on the solution results of nonlinear model predictive control, the optimal control input is generated, and the optimal control input is converted into the control signal of the bridge-building machine actuator to perform closed-loop control of the bridge-building machine.

[0181] For example, the intelligent control system of the bridge erection machine is a complex integration of multi-sensor fusion and advanced control algorithms. The system continuously collects and processes multi-source sensor data to accurately estimate the state of the bridge erection machine, and based on these estimation results, uses a nonlinear model predictive control algorithm for closed-loop control to ensure the accuracy, safety and energy efficiency of the construction process.

[0182] First, the system uses a variety of sensors to collect information about the state of the bridge crane. The inertial measurement unit (IMU) is used to measure the acceleration and angular velocity of the bridge crane, with a typical sampling frequency of 100 Hz and an accuracy of up to 0.01 m / s. 2 and 0.01° / s. The global positioning system (GPS) provides the absolute position information of the bridge erector, usually using RTK-GPS technology, which can achieve centimeter-level positioning accuracy with an update frequency of 10Hz. The laser rangefinder is used to measure the relative distance between the bridge erector and the surrounding environment, with a measurement range of up to 100m and an accuracy of ±1mm. Strain sensors are installed in key structural parts of the bridge erector to monitor structural stress, with a sampling frequency of usually 1kHz. The actuator encoder records the hydraulic cylinder position and motor speed with a resolution of up to 0.1mm and 1rpm.

[0183] After preliminary processing, these sensor data are used to construct a state vector containing position, acceleration, angular velocity and force / torque. For example, a typical state vector may contain 18 variables: 3 spatial position coordinates, 3 Euler angles, 3 linear velocity components, 3 angular velocity components and 6 main forces / torques. Based on this state vector, discrete time nonlinear state equations and observation equations are established. The state equation describes the evolution of the system state, taking into account the kinematic and dynamic characteristics of the bridge crane. The observation equation establishes the relationship between the sensor measurements and the system state.

[0184] In order to obtain accurate state estimation from the noise-contaminated sensor data, an extended Kalman filter (EKF) is used for state estimation. The execution of the EKF includes two main steps: prediction and update. In the prediction step, the state equation is used to predict the state and state covariance at the current moment based on the state estimate and control input at the previous moment. For example, if the estimated position of the bridge crane at the previous moment is (10m, 5m, 2m) and the speed is (0.5m / s, 0.3m / s, 0.1m / s), the position after 0.1 seconds can be predicted to be approximately (10.05m, 5.03m, 2.01m).

[0185] The update step first calculates the Kalman gain, which is a key parameter that weighs the prior estimate and the measurement information. Then, based on the difference between the actual measurement and the predicted value, the state estimate and covariance are updated. For example, if the GPS measurement shows that the actual position is (10.06m, 5.04m, 2.02m), the EKF will correct the predicted position according to the Kalman gain to obtain a more accurate state estimate.

[0186] Based on the state estimation results of EKF, a nonlinear model predictive control (NMPC) optimization problem is constructed. The optimization objectives of NMPC usually include three aspects: tracking the predetermined trajectory, minimizing energy consumption, and satisfying safety constraints. Tracking the predetermined trajectory can be defined as the sum of squared deviations between the actual trajectory and the reference trajectory. Minimizing energy consumption can be approximated by minimizing the sum of squares of control inputs. Safety constraints include speed limits, acceleration limits, joint angle limits, etc.

[0187] The solution of the NMPC optimization problem adopts a real-time iteration strategy, which divides the solution process into a preparation phase and a feedback phase. In the preparation phase, the system dynamics are first linearized. For example, a linear approximation can be performed using a Taylor expansion near the current operating point. Then, a quadratic programming subproblem is constructed based on the linearized model. This subproblem is a local approximation of the original nonlinear problem, but can be solved quickly using an efficient quadratic programming solver.

[0188] In the feedback phase, the initial state is first updated using the latest state estimation results. Then the constructed quadratic programming subproblems are solved to obtain a series of optimal control inputs. For example, for an NMPC problem with a prediction horizon of 2 seconds and a control interval of 0.1 seconds, each solution will obtain 20 control inputs. Finally, the first control input obtained from the solution is applied to the system, and the remaining control inputs are used as initial guesses for the next optimization.

[0189] Based on the solution results of NMPC, the optimal control input is generated and converted into specific control signals for the bridge crane actuators. For example, the control input may include the main beam travel speed, the crossbeam pitch angular velocity, the hook speed of the lifting mechanism, etc. These control signals are converted into corresponding voltage or current signals through digital-to-analog converters and sent to the drivers of each actuator, thereby realizing closed-loop control of the bridge crane.

[0190] The intelligent control system of the bridge-building machine based on multi-sensor fusion and nonlinear model predictive control adopted in this paper has many beneficial effects. First, multi-sensor fusion and EKF state estimation provide high-precision and robust estimation of the state of the bridge-building machine, laying a solid foundation for subsequent control. Secondly, the NMPC algorithm can simultaneously consider the nonlinear characteristics of the system, multi-objective optimization and various constraints to generate the optimal control strategy. This not only ensures that the bridge-building machine can accurately track the predetermined trajectory, but also minimizes energy consumption while ensuring safety. Furthermore, the adoption of real-time iterative strategies enables NMPC to obtain effective solutions within a limited computing time, meeting the needs of real-time control. Finally, this closed-loop control method can effectively cope with external disturbances and model uncertainties, significantly improve the accuracy, efficiency and safety of bridge-building machine construction, and provide advanced technical support for large-scale infrastructure construction.

[0191] In an optional embodiment,

[0192] After the construction is completed, the steps of analyzing the construction quality and efficiency using graph neural networks based on the digital twin model and the whole process data, and generating a multi-dimensional evaluation report and optimization suggestions include:

[0193] The bridge erection machine and the bridge structure are represented as a heterogeneous graph, where nodes represent components or monitoring points, edges represent physical connections or logical relationships between components, and each node and edge is accompanied by a multi-dimensional feature vector; a special feature transformation matrix is ​​applied to the multi-dimensional feature vectors of different types of nodes and edges to generate initial node and edge representations;

[0194] Based on the initial node and edge representation, an improved graph neural network model is constructed, which includes node-level attention mechanism and edge-level attention mechanism. Node-level attention is used to learn the importance weights between different features, and edge-level attention is used to learn the influence of different neighbor nodes.

[0195] A temporal convolutional network is introduced after the graph convolutional layer of the graph neural network model. Causal convolution and dilated convolution are used to capture temporal features and fuse them with the output of the graph convolutional layer.

[0196] Use historical data from the bridge erection machine construction system to train the graph neural network model, and iterate the training until convergence;

[0197] Use the trained graph neural network model to analyze the quality and efficiency of the new bridge-building machine construction system data and obtain the prediction results of various indicators. Based on the prediction results, generate a multidimensional evaluation report that includes overall score, analysis of various indicators, identification of abnormal events and their root cause analysis. According to the problems and potential risks identified in the multidimensional evaluation report, combined with historical construction experience, generate optimization suggestions and improvement measures.

[0198] For example, after the construction of the bridge erection machine is completed, it is a complex and important task to use the digital twin model and the whole process data to analyze the quality and efficiency. This technical solution proposes an innovative method based on graph neural network, which can comprehensively analyze the construction quality and efficiency and generate multi-dimensional evaluation reports and optimization suggestions.

[0199] First, the bridge erection machine and the bridge structure are represented as a heterogeneous graph. In this graph, nodes can represent different components or monitoring points. For example, a node may represent a main beam, a cross beam, a leg, or a specific stress monitoring point. The edge represents the physical connection or logical relationship between components, such as the connection between the main beam and the cross beam, or the spatial relationship between adjacent monitoring points. Each node and edge is accompanied by a multidimensional feature vector containing rich information. For nodes, features may include location coordinates, geometric dimensions, material properties, stress and strain values, etc. For example, a node representing a main beam may have the following features: [10.5, 2.3, 1.8, 500, 7850, 210, 150], representing length (m), width (m), height (m), mass (kg), density (kg / m 3 ), elastic modulus (GPa), and maximum stress (MPa). The characteristics of the edge may include connection type, stiffness coefficient, force transmission characteristics, etc.

[0200] In order to handle these heterogeneous features, special feature transformation matrices are applied to different types of nodes and edges. The purpose of this step is to convert the original features into vector representations with the same dimensions to facilitate subsequent graph neural network processing. For example, a 100x7 transformation matrix can be used to convert the above 7-dimensional main beam features into a 100-dimensional vector representation. This process is essentially a feature embedding that can capture the potential relationship between features.

[0201] Based on the initial node and edge representations, an improved graph neural network model is constructed. This model contains two key attention mechanisms: node-level attention and edge-level attention. The node-level attention mechanism is used to learn the importance weights between different features. For example, when evaluating the quality of the main beam, the stress value may be more important than the density, so it will be given a higher weight. The edge-level attention mechanism is used to learn the influence of different neighbor nodes. For example, when analyzing the abnormal situation of a monitoring point, the directly connected components may have a greater influence than the distant components.

[0202] The temporal convolutional network is introduced after the graph convolutional layer of the graph neural network model. This design is intended to capture temporal features. Causal convolution is used to ensure that the model only uses past and current information for prediction to avoid information leakage. Dilated convolution enables the model to effectively capture long-term dependencies by increasing the receptive field. For example, 3 layers of dilated convolution can be used with dilation rates of 1, 2, and 4, respectively, so that a longer time span can be covered without increasing the number of parameters. The output of the temporal convolutional network is fused with the output of the graph convolutional layer to obtain a feature representation that combines spatial and temporal information.

[0203] The graph neural network model is trained using historical data from the bridge erection machine construction system. The training data includes various sensor data collected during the construction process, construction records, and the final quality assessment results. The training process uses a small batch stochastic gradient descent method, with each batch containing 64 samples. The learning rate is initially set to 0.001, and a learning rate decay strategy is used. The model training is iterative until the loss function on the validation set converges or the preset number of iterations (such as 1000 rounds) is reached.

[0204] Use the trained graph neural network model to analyze the quality and efficiency of the new bridge-building machine construction system data. The model can predict various indicators, such as structural stress distribution, deformation, construction efficiency, etc. For example, for a new construction project, the model may predict that the maximum stress of the main beam is 145MPa, the deformation is 5mm, and the construction efficiency is 95%.

[0205] Based on the prediction results, a multi-dimensional evaluation report is generated. This report includes an overall score, analysis of various indicators, identification of abnormal events and their root cause analysis. The overall score can be based on a percentage system, taking into account the performance of multiple aspects. The analysis of various indicators shows in detail the predicted value, historical comparison and industry standard comparison of each key indicator. For example, the report may indicate that the stress of the main beam is within the normal range, but it is 5% higher than the historical average, and further monitoring is recommended. The abnormal event identification section lists all indicators or abnormal patterns that exceed the preset threshold, and uses the interpretability characteristics of the graph neural network to perform root cause analysis. For example, if the stress of a node is detected to be abnormally high, the model can trace back the key factors and propagation paths that affect the node.

[0206] Based on the problems and potential risks identified in the multi-dimensional assessment report and combined with historical construction experience, optimization suggestions and improvement measures are generated. These suggestions may include adjusting the construction process, optimizing resource allocation, strengthening monitoring of specific parts, etc. For example, if the report indicates that a certain type of connector frequently experiences stress concentration, the system may recommend improving the connection design or using higher-strength materials.

[0207] The graph neural network-based bridge-erecting machine construction quality and efficiency analysis method adopted in this paper has many beneficial effects. First, it can comprehensively consider the complex topological relationship between the bridge-erecting machine and the bridge structure, capture the mutual influence between components, and provide more accurate analysis results than traditional methods. Secondly, the introduced attention mechanism and temporal convolutional network enable the model to adaptively focus on important features and long-term dependencies, improving the accuracy and robustness of the analysis. Furthermore, this method can achieve end-to-end learning, directly derive evaluation results from raw data, and reduce the interference of human factors. Finally, the generated multi-dimensional evaluation report and optimization suggestions provide intuitive and comprehensive decision-making support for construction managers, which helps to improve construction quality, efficiency and safety. The application of this method can not only optimize current construction projects, but also promote technological progress in the entire bridge construction industry by accumulating and analyzing large amounts of data, continuously improving construction technology and management strategies.

[0208] In an optional embodiment,

[0209] like Figure 2 As shown, the bridge erecting machine structure of the present invention includes: a crossbeam, a camera, and a transport vehicle.

[0210] Among them, number 1 is the right end beam assembly, which is used to provide the support structure, lateral stability and load-bearing function of the right end of the bridge-building machine; number 2 is the front support system, which is used to support the front end of the main beam, provide stability in the forward direction and bear the front load distribution; number 3 is the transport vehicle, which is used to transport bridge segments, realize the precise positioning of bridge segments and cooperate with the movement requirements of the installation process; number 4 is the crossbeam, which is used to provide lateral structural support, connect various support systems and ensure the overall structural rigidity; number 5 is the camera, which is used to obtain image information of the construction environment where the bridge-building machine is located, and can be transmitted to the controller by wire or wireless means to control the bridge-building machine accordingly; number 6 is the overhead crane system, which is used to realize the lifting and installation of bridge components. It can perform precise position adjustment and complete horizontal and vertical movement control; No. 8 is the middle support bracket, which is used to support the middle part of the main beam, disperse the bearing pressure and provide stable support during construction; No. 9 is the auxiliary bracket, which is used to provide additional support force, increase the overall structural stability and assist in the support requirements under special working conditions; No. 10 is the front support system, which is used in conjunction with No. 2 bracket to enhance the front structural strength and provide a construction operation platform; No. 11 is the left end beam assembly, which is used as the left end support structure of the bridge erecting machine, balance the overall force and cooperate with the right end to form a complete support system; No. 12 is the main beam, which is used as the main load-bearing component of the entire bridge erecting machine, provides longitudinal support and guidance, and at the same time carries the installation basis of various mechanisms and equipment.

[0211] The main parameters are: rated lifting capacity: 800t; suitable span of erected beams: ≤32m; whole machine through-hole speed: 3m / min; lifting speed of hanging beam: no-load: 0~1m / min, heavy load: 0~0.5m / min; longitudinal movement speed of hanging beam: 0~3m / min; longitudinal movement speed of guide beam: 0~3m / min; lateral movement of arm: No. 1 column ±250mm, No. 2 column ±350mm; lateral movement speed of arm: 0.3m / min; suitable slope: ≤30‰; minimum curve radius: 1600m; installed capacity: 230kw; system pressure: 31.5Mpa; working level: A3; maximum overall dimensions: 82.6mx16.8mX12m;

[0212] Its working process includes:

[0213] Preparation stage: The bridge erection machine is in place, spanning across the two bridge piers; system calibration, including initialization of the machine vision system;

[0214] Beam segment transportation: the bridge erection machine moves to the prefabricated beam segment storage area (speed: 0-3m / min); the hoist descends to grab the beam segment (descending speed: 0-1m / min);

[0215] Beam segment installation: The bridge erection machine carries the beam segment to the installation location; uses the machine vision system for precise positioning; and slowly lowers the beam segment to the predetermined location (lowering speed: 0-0.5m / min);

[0216] Precise adjustment: The machine vision system monitors the beam segment position in real time; fine-tuning is performed through lateral and longitudinal sliding (±250mm and ±350mm); ensuring that the beam segment positioning error is within the allowable range;

[0217] Fixing and verification: Fixing of beam segments after they are in place; final inspection and data recording by machine vision system;

[0218] Prepare for the next cycle: the bridge erecting machine moves to the next working position (moving speed: 0.3m / min).

[0219] The technical effect of the bridge erection machine provided by the present invention is mainly reflected in the high-precision installation. The lateral and longitudinal sliding mechanisms cooperate with the machine vision system to achieve accurate positioning of the beam segment, with high installation accuracy, ensuring the high quality and stability of the bridge structure. The bridge erection machine has a large load capacity, with a rated lifting capacity of 800 tons, and can handle large prefabricated beam segments, meeting the needs of modern bridge engineering for large spans and heavy loads.

[0220] In terms of flexibility and adaptability, the maximum span can reach 32 meters, suitable for various bridge types, and the minimum curve radius is 1,600 meters, which can adapt to different bridge lines. Efficient construction is another major feature. The whole machine has a hole-passing speed of 3m / min, a walking speed of 0.3m / min, and a hoisting speed of 0-1m / min, which enables fast construction. The high degree of automation reduces manual operations and improves construction efficiency.

[0221] In terms of safety performance, the system pressure is 31.5MPa, ensuring stability under heavy loads. The machine vision system monitors in real time to reduce human errors and safety risks. Precision control is its important feature. The hoisting speed can be precisely controlled to ensure that the beam section is in place smoothly. The lateral and longitudinal moving speeds can reach 0-3m / min, which is convenient for fine adjustment.

[0222] The bridge erecting machine has large-size adaptability, with a maximum size of 82.6m×16.8m×12m, which is suitable for large bridge projects. It has strong environmental adaptability and working level A3, indicating strong environmental adaptability and stability. Intelligent construction is its significant advantage. Combined with the machine vision system, it realizes intelligent positioning, automatic adjustment and real-time monitoring, improves construction accuracy and reduces human errors.

[0223] With strong data recording and analysis capabilities, the machine vision system can record construction data for quality control and subsequent optimization. The modular design facilitates transportation, assembly and maintenance. Its versatility makes it not only used for beam segment installation, but also for the lifting and positioning of other bridge components.

[0224] This bridge-building machine structure greatly improves the efficiency, accuracy and safety of bridge construction by combining advanced mechanical design and intelligent control systems, and represents the development direction of modern bridge engineering technology.

[0225] Figure 3 FIG. 1 is a schematic diagram of the structure of an intelligent construction control system for a bridge erecting machine based on machine vision according to an embodiment of the present invention. Figure 3 As shown, the system comprises:

[0226] The first unit is used to collect image information of the construction site through multiple high-definition cameras installed on the bridge erection machine; input the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, a digital twin model of the construction scene is constructed;

[0227] The second unit is used to train an intelligent decision-making agent using a reinforcement learning algorithm based on a digital twin model and a predefined reward and punishment function; the intelligent decision-making agent generates an initial construction strategy based on the current state of the construction scene, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates an optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine;

[0228] The third unit is used to analyze the construction quality and efficiency after the construction is completed, based on the digital twin model and the data of the whole process, using the graph neural network to generate a multi-dimensional evaluation report and optimization suggestions.

[0229] According to a third aspect of the embodiments of the present invention,

[0230] An electronic device is provided, comprising:

[0231] processor;

[0232] a memory for storing processor-executable instructions;

[0233] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0234] According to a fourth aspect of the embodiments of the present invention,

[0235] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0236] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The intelligent construction control method of bridge erection machine based on machine vision is characterized in that: include: The image information of the construction site is collected by multiple high-definition cameras installed on the bridge erection machine; the image information is input into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, a digital twin model of the construction scene is constructed; Using reinforcement learning algorithms, intelligent decision-making agents are trained based on digital twin models and predefined reward and punishment functions; The intelligent decision-making agent generates an initial construction strategy according to the current state of the construction scene, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates an optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine; After the construction is completed, the construction quality and efficiency are analyzed using graph neural networks based on the digital twin model and full-process data to generate multi-dimensional evaluation reports and optimization suggestions.

2. The method according to claim 1, characterized in that The steps of collecting image information of the construction site by multiple high-definition cameras installed on the bridge erecting machine; inputting the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various parts of the bridge erecting machine, bridge components and the surrounding environment include: The image information of the construction site is collected by multiple high-definition cameras arranged on the bridge erecting machine; the image information is preprocessed by adaptive histogram equalization; the preprocessed image information is input into a pre-trained improved Mask R-CNN deep learning model, and a spatial pyramid pooling module is introduced into the feature extraction network of the improved Mask R-CNN deep learning model; An attention mechanism module is added after the region of interest alignment layer of the improved Mask R-CNN deep learning model to highlight features and suppress irrelevant background information by generating spatial attention maps and channel attention maps; Using the improved Mask R-CNN deep learning model, the targets in the construction scene are identified and located, including the components of the bridge erecting machine, bridge components and the surrounding environment; The recognition results of the targets in the construction scene are compared with the pre-built special data set containing bridge erection machine parts and bridge components; the focal loss function is used as the loss function in the training process of the improved Mask R-CNN deep learning model, and the online case mining technology is introduced to dynamically select difficult samples for training; The trained improved Mask R-CNN deep learning model is quantized and pruned to convert the 32-bit floating-point numbers in the model into 8-bit integers. The processed improved Mask R-CNN deep learning model is used to perform real-time target detection and recognition in the construction scene, and the target category, location, and contour information are output. The detection and recognition results are transmitted to the control system of the bridge erection machine.

3. The method according to claim 1, characterized in that Based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, the steps of constructing a digital twin model of the construction scene include: Construct a graph structure representing the topological structure of the construction scene, wherein the nodes of the graph structure represent the identified targets or elements in the preset model, and the edges of the graph structure represent the spatial relationship between the nodes; Initialize the node feature matrix based on the constructed graph structure; Applying a spatially-aware graph convolution operation to the graph structure, the spatially-aware graph convolution operation comprising: Calculate the Euclidean distance between nodes in the graph structure; Based on the calculated Euclidean distance, the spatial distance weights between nodes are calculated using the Gaussian kernel function to obtain a spatial distance weight matrix; Construct the adjacency matrix of the graph structure and add self-connection to the adjacency matrix; Calculate the degree matrix of the adjacency matrix after adding self-connection; Multiplying the spatial distance weight matrix with the adjacency matrix to obtain a weighted adjacency matrix; Apply a weighted adjacency matrix to the node feature matrix and multiply it with a learnable weight matrix; Apply a nonlinear activation function to the product result to obtain the updated node feature matrix; Based on the updated node feature matrix, feature extraction is performed on the two-dimensional target recognition results and three-dimensional model data in the construction scene; The extracted two-dimensional features and three-dimensional features are input into the multimodal feature fusion module, which uses a fully connected layer to process features and calculates fusion weights through an attention mechanism; Perform weighted fusion on the processed features according to the fusion weights to obtain fusion features; Based on the fused features and the graph structure after the spatial-aware graph convolution operation, a digital twin model of the construction scene is constructed.

4. The method according to claim 1, characterized in that: The intelligent decision-making agent generates an initial construction strategy based on the current construction scene status, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates the optimal construction strategy through multi-objective solving. The steps include: Using intelligent decision-making agents to generate initial construction strategies based on the current state of the construction scene; An improved Monte Carlo tree search is used to optimize the initial construction strategy and generate an optimization result; the improved Monte Carlo tree search includes introducing a dynamic action generator, using a UCT algorithm with adaptive exploration parameters, realizing tree parallelization, and applying heuristic pruning; In the UCT algorithm of the adaptive exploration parameter, a dynamically adjusted exploration parameter c is introduced: c=c_base*(1+α*exp(-β*t / T)); Among them, c_base is the basic exploration parameter, α and β are adjustment factors, t is the current iteration number, and T is the total iteration number; The optimization results of the initial construction strategy are solved for multiple objectives, including defining the objective functions of construction efficiency and energy consumption, and the improved NSGA-III algorithm is used for multi-objective optimization. The improved NSGA-III algorithm includes adaptive crossover and mutation, local search and reference point adaptation. The optimal construction strategy is generated according to the multi-objective optimization solution results.

5. The method according to claim 4, characterized in that The improved NSGA-III algorithm includes the steps of adaptive crossover and mutation, local search and reference point adaptation, including: Introducing an adaptive crossover and mutation mechanism based on population diversity, dynamically adjusting the crossover probability and mutation probability according to the current population diversity metric, which is calculated using the entropy weight method, and increasing the intensity of genetic operations when population diversity decreases; After each iteration of the NSGA-III algorithm, the objective function values ​​of all individuals in the current population are evaluated, a non-dominated solution set is identified and constructed, multiple solutions are selected from the non-dominated solution set for local search, and multiple simulated annealing iterations are performed on each selected solution. The simulated annealing iteration adopts an exponentially decaying temperature schedule, and a neighborhood solution is generated in each iteration and the acceptance probability is calculated based on the energy difference. If the solution obtained by the local search dominates the original solution, the new solution replaces the original solution. A reference point adaptive mechanism is adopted, including uniformly generating initial reference points in the standardized target space, periodically evaluating the correlation degree of each reference point during the evolution process, using the density peak clustering algorithm for low-correlation reference points to identify high-density areas in the target space, moving low-correlation reference points to the nearest high-density area while maintaining dispersion, and introducing regularization terms to limit the large-scale migration of reference points.

6. The method according to claim 1, characterized in that During the construction process, the data of each sensor of the bridge-erecting machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-erecting machine; Based on state estimation, the steps of closed-loop control of the bridge erecting machine using the nonlinear model predictive control algorithm include: Using an inertial measurement unit, a global positioning system, a laser rangefinder, a strain sensor and an actuator encoder to collect state information of the bridge erecting machine, the state information including acceleration, angular velocity, position, relative distance, structural stress, hydraulic cylinder position and motor speed; Constructing a state vector including position, acceleration, angular velocity and force / torque, and establishing a discrete-time nonlinear state equation and an observation equation based on the state vector; An extended Kalman filter is used to perform state estimation on discrete-time nonlinear state equations and observation equations, including executing a prediction step and an update step, wherein the prediction step includes state prediction and covariance prediction, and the update step includes calculating Kalman gain, state update, and covariance update; Based on the state estimation results, a nonlinear model predictive control optimization problem is constructed. The optimization objectives include tracking the predetermined trajectory, minimizing energy consumption, and satisfying safety constraints. A real-time iteration strategy is adopted to solve the nonlinear model predictive control optimization problem, and the solution process is divided into a preparation phase and a feedback phase, wherein the preparation phase includes linearizing the system dynamics and constructing a quadratic programming subproblem, and the feedback phase includes updating the initial state, solving the quadratic programming subproblem and applying the control input; Based on the solution results of nonlinear model predictive control, the optimal control input is generated, and the optimal control input is converted into the control signal of the bridge-building machine actuator to perform closed-loop control of the bridge-building machine.

7. The method according to claim 1, characterized in that After the construction is completed, the steps of analyzing the construction quality and efficiency using graph neural networks based on the digital twin model and the whole process data, and generating a multi-dimensional evaluation report and optimization suggestions include: The bridge erection machine and the bridge structure are represented as a heterogeneous graph, where nodes represent components or monitoring points, edges represent physical connections or logical relationships between components, and each node and edge is accompanied by a multi-dimensional feature vector; a special feature transformation matrix is ​​applied to the multi-dimensional feature vectors of different types of nodes and edges to generate initial node and edge representations; Based on the initial node and edge representation, an improved graph neural network model is constructed, which includes node-level attention mechanism and edge-level attention mechanism. Node-level attention is used to learn the importance weights between different features, and edge-level attention is used to learn the influence of different neighbor nodes. A temporal convolutional network is introduced after the graph convolutional layer of the graph neural network model. Causal convolution and dilated convolution are used to capture temporal features and fuse them with the output of the graph convolutional layer. Use historical data from the bridge erection machine construction system to train the graph neural network model, and iterate the training until convergence; Use the trained graph neural network model to analyze the quality and efficiency of the new bridge-building machine construction system data and obtain the prediction results of various indicators. Based on the prediction results, generate a multidimensional evaluation report that includes overall score, analysis of various indicators, identification of abnormal events and their root cause analysis. According to the problems and potential risks identified in the multidimensional evaluation report, combined with historical construction experience, generate optimization suggestions and improvement measures.

8. A machine vision-based intelligent construction control system for bridge erection machines, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect image information of the construction site through multiple high-definition cameras installed on the bridge erection machine; input the image information into a pre-trained deep learning model to identify and locate targets in the construction scene, including various components of the bridge erection machine, bridge components and the surrounding environment; based on the recognition results of the targets in the construction scene, combined with the graph structure of the topological structure of the construction scene after the spatial perception graph convolution operation, a digital twin model of the construction scene is constructed; The second unit is used to train intelligent decision-making agents based on digital twin models and predefined reward and punishment functions using reinforcement learning algorithms; The intelligent decision-making agent generates an initial construction strategy according to the current state of the construction scene, optimizes the initial construction strategy using an improved Monte Carlo tree search, and generates an optimal construction strategy through multi-objective solution; the optimal construction strategy is converted into an intelligent control instruction, and the intelligent control instruction is sent to the control system of the bridge-building machine to achieve precise control of each actuator of the bridge-building machine; during the construction process, the data of each sensor of the bridge-building machine is continuously collected, and the data of each sensor and the feedback data of the actuator are integrated to generate a state estimation of the bridge-building machine; based on the state estimation, the nonlinear model predictive control algorithm is used to perform closed-loop control of the bridge-building machine; The third unit is used to analyze the construction quality and efficiency after the construction is completed, based on the digital twin model and the data of the whole process, using the graph neural network to generate a multi-dimensional evaluation report and optimization suggestions.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images