A steel structure component positioning method based on laser scanning and deep learning

Abstract

The present application relates to the technical field of construction and computer vision, and discloses a steel structure component positioning method based on laser scanning and deep learning, which comprises a three-dimensional laser scanning module, a feature extraction module, a theoretical modeling module, a correction instruction generation module, a matching verification module, an installation quality evaluation module and a model self-optimization module; the present application changes post-detection into in-process guidance, greatly reduces repeated adjustment and rework time, and improves hoisting efficiency; deep learning and traditional ICP algorithm are combined, advantages are complementary, and overall positioning accuracy is controlled; geometric positioning and mechanical property prediction are combined, more scientific and comprehensive installation quality evaluation is given, and structure safety is ensured from the source; the present application has closed-loop learning ability and can continuously learn and optimize from historical installation data; through AR technology, complex three-dimensional deviation data is intuitively presented to operating personnel, and the technical threshold requirement for operating personnel is reduced.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of building construction and computer vision, and more specifically to a method for locating steel structure components based on laser scanning and deep learning. Background Technology

[0002] In the construction of large steel structures, the installation and positioning accuracy of components is directly related to the safety and stability of the overall structure.

[0003] Currently, the positioning of components in the construction of large steel structures mainly adopts optical measurement technology such as total station. This method involves professional surveyors operating instruments to aim and measure the distance to the pre-set control points of the components, obtain the three-dimensional coordinate data of key points, and determine the spatial position and orientation of the components through manual recording and calculation. This technology has formed a standardized operating procedure in practice and can meet the general construction accuracy requirements. It is a traditional positioning method widely used in the current steel structure installation process.

[0004] However, the following drawbacks still exist:

[0005] Existing technologies are based on a post-installation testing model, separating the measurement and installation processes. This leads to significant time-consuming issues such as repeated positioning, waiting for verification, and rework, resulting in low overall hoisting efficiency.

[0006] Traditional methods rely on manual aiming and discrete point measurement, which are easily affected by operational experience and environmental interference. They are difficult to fully verify the spatial attitude of complex components, have limitations in precision control, and lack reliability.

[0007] Existing acceptance methods only focus on geometric deviations and lack a scientific assessment of the impact of installation deviations on mechanical properties. The quality evaluation has a single dimension and cannot predict structural safety risks from the source.

[0008] Traditional methods lack data accumulation and self-optimization capabilities. Their effectiveness depends entirely on the technical level of operators and cannot adapt to changes in different projects and complex working conditions. Their level of intelligence is low.

[0009] Operators need to rely on abstract numerical reports and drawings to visualize the space. There is a lack of intuitive three-dimensional operation guidance on site. The technical threshold is high, the training cost is large, and human error is easy to occur.

[0010] Therefore, methods that offer real-time performance and high efficiency, high accuracy and reliability, comprehensive quality assessment, self-learning and self-adaptation, and intuitive visualization are needed to solve the above problems. Summary of the Invention

[0011] In order to overcome the above-mentioned defects of the prior art, the present invention provides a steel structure component positioning method based on laser scanning and deep learning to solve the problems existing in the background art.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a method for locating steel structure components based on laser scanning and deep learning, comprising the following steps:

[0013] S1. The target area is scanned from multiple angles before, during and after component hoisting using a 3D laser scanning module to obtain point cloud data and generate uncertain and stable models.

[0014] S2. Use a deep learning network to extract features from the uncertain model and match them with the stable model through the feature extraction module, and output the initial transformation matrix between the component and the design position;

[0015] S3. The initial transformation matrix output by the deep learning network is combined with the original BIM model through the theoretical modeling module to generate the theoretical installation and positioning model of the component in the current state.

[0016] S4. The theoretical installation positioning model and the stable model are registered with high-precision algorithm through the correction instruction generation module, the residual is calculated and the correction instruction is output and presented visually.

[0017] S5. Perform theoretical and practical matching verification between the theoretical installation positioning model and the stable model through the matching verification module;

[0018] S6. The installation quality evaluation module calculates the position deviation score and installation strength score based on the data after hoisting and the stability model data, and performs a weighted calculation to output the comprehensive installation quality score.

[0019] S7. The model self-optimization module uses the initial feature vector, correction instructions, and final score of each installation as training samples to train the model, enabling continuous self-optimization.

[0020] The technical effects and advantages of this invention are as follows:

[0021] 1. This invention is the first to propose a point cloud intelligent coarse registration method based on deep learning, which effectively overcomes the technical bottleneck of the traditional ICP algorithm being sensitive to the initial position, transforms post-event detection into real-time in-process guidance, greatly reduces repeated adjustments and rework, and improves hoisting efficiency;

[0022] 2. This invention constructs a two-stage high-precision positioning technology route of deep learning coarse registration + ICP fine registration, giving full play to the advantages of both, accurately controlling the overall positioning accuracy, and solving the industry problem of ultra-high precision construction of complex steel structure nodes;

[0023] 3. This invention integrates geometric deviation with mechanical performance prediction. By integrating a lightweight FEA engine, it achieves a quantitative assessment leap from installation accuracy to structural safety, providing a scientific and comprehensive comprehensive score for installation quality and ensuring structural safety from the source.

[0024] 4. This invention establishes a closed-loop self-optimization system based on reinforcement learning, which enables the system to continuously learn from historical installation data. Its intelligence level increases with usage time, significantly improving its adaptability and generalization ability under different projects, components and complex working conditions.

[0025] 5. This invention develops a three-dimensional deviation visualization human-computer interaction interface based on AR technology, which converts the abstract six-degree-of-freedom deviation parameters into intuitive graphic arrows and virtual contours, greatly reducing the requirements for operators' technical experience and spatial imagination. Attached Figure Description

[0026] Figure 1 This is a structural block diagram of the present invention.

[0027] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The steel structure component positioning method based on laser scanning and deep learning involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 This invention provides a method for locating steel structure components based on laser scanning and deep learning, including a three-dimensional laser scanning module, a feature extraction module, a theoretical modeling module, a correction instruction generation module, a matching verification module, an installation quality evaluation module, and a self-optimization module.

[0030] Reference Figure 2 The specific implementation steps of the present invention include the following steps:

[0031] S1. The target area is scanned from multiple angles before, during and after component hoisting using a 3D laser scanning module to obtain point cloud data and generate uncertain and stable models.

[0032] It should be specifically noted that the stable model is as follows:

[0033] The high-precision point cloud converted from the original BIM model, as well as the point cloud of the installed components that have passed acceptance and are correctly positioned, serve as a benchmark, ensuring stable and reliable coordinates.

[0034] The uncertainty model is specifically as follows:

[0035] The real-time point cloud data obtained from scanning the components currently being hoisted is characterized by occlusion and noise, and the position is in a state of flux and uncalibration.

[0036] S2. Use a deep learning network to extract features from the uncertain model through the feature extraction module and match the features with the stable model to output the initial transformation matrix between the component and the design position.

[0037] It should be specifically noted that the steps of extracting features from the uncertain model and matching them with the features of the stable model are as follows:

[0038] A1. Input the uncertain model point cloud obtained from the current hoisting component scan;

[0039] A2. Input the uncertain model point cloud into a pre-trained deep learning network FCGF feature extraction network;

[0040] A3. Feature extraction for uncertain texts;

[0041] Generate a compact and discriminative feature descriptor for each point in the point cloud; points with similar geometric structures will have similar feature vectors.

[0042] A4. Perform feature matching with stable models;

[0043] Deep learning networks can understand the geometry of components, such as how many holes a steel node has, the cross-sectional shape of a beam, and the style of a connecting plate. Even if the currently scanned point cloud is incomplete, the network can infer what the whole is and which part of the target model should be matched based on local features.

[0044] A5. Output the initial transformation matrix, specifically:

[0045] ;

[0046] Where T is the initial transformation matrix, which contains all the parameters needed to rotate and translate the currently scanned component from its random initial pose to a position roughly aligned with the original BIM model. Homogeneous coordinates are used to represent rotation and translation in a single matrix.

[0047] R is a 3×3 matrix, which is the rotation matrix that controls the rotation of the object. The three column vectors represent the directions of the X, Y, and Z axes in the original coordinate system after the rotation.

[0048] t is a 3×1 vector, which is a translation vector and represents the distance the object moves in the X, Y, and Z directions;

[0049] The data at the bottom are filler values ​​necessary for the homogeneous coordinate representation introduced to achieve a linear combination of translation and rotation.

[0050] S3. The initial transformation matrix output by the deep learning network is combined with the original BIM model through the theoretical modeling module to generate a theoretical installation and positioning model of the component in its current state.

[0051] It should be specifically explained that the theoretical installation positioning model is as follows: the initial transformation matrix is ​​applied to each vertex of the original BIM model, and the three-dimensional model with updated position formed by all the transformed vertices is the theoretical installation positioning model.

[0052] The transformed vertex coordinates are specifically as follows:

[0053] V = T * Vo;

[0054] Where V is the final three-dimensional coordinate of a point in the theoretical installation positioning model; T is the initial transformation matrix described in S2; and Vo is the initial three-dimensional coordinate of the corresponding point in the original BIM design model.

[0055] S4. The theoretical installation positioning model and the stable model are registered with high precision algorithm through the correction instruction generation module, the residual is calculated and the correction instruction is output and presented visually.

[0056] It should be noted that the input consists of the theoretical installation positioning model point cloud and the stable model high-precision point cloud generated by S3; the ICP iterative nearest point algorithm is selected for high-precision algorithm registration; and finally, an accurate final transformation matrix is ​​output, which represents the final adjustment required from the position of the theoretical installation positioning model to the position of the stable model.

[0057] It should be explained that, when the initial positions are very close, ICP is the optimal and most efficient choice for micrometer-level fine registration. The specific process of high-precision algorithm registration is as follows:

[0058] B1. Nearest point search: For each point in the theoretical installation positioning model, find the nearest point in the stable model and establish a point-to-point correspondence.

[0059] B2. Calculate the transformation: Based on these corresponding points, calculate an optimal transformation matrix that minimizes the sum of the distances between all corresponding points, including rotation and translation;

[0060] B3. Application of Transformation: Apply this transformation matrix to the theoretical installation positioning model to make it closer to the stable model;

[0061] B4. Iteration: Repeat the above steps to find the nearest point → calculate the new transformation → apply the transformation until the transformation value between two iterations is less than a preset threshold, or the error no longer decreases significantly. At this point, the algorithm is considered to have converged and the final transformation matrix is ​​output.

[0062] The specific steps for calculating the residual and outputting the correction instruction are as follows:

[0063] Decomposing the final transformation matrix, a spatial transformation matrix can be decomposed into parameters with six degrees of freedom, specifically including:

[0064] Three translational deviations (ΔX, ΔY, ΔZ): These represent the distances that need to be moved in the X, Y, and Z coordinate axes, respectively.

[0065] For example: ΔX=+3.5mm means shifting 3.5mm to the right, ΔY=-1.2mm means shifting 1.2mm backward, and ΔZ=+5.7mm means shifting 5.7mm upward.

[0066] Three rotational deviations (Δα, Δβ, Δγ): representing the angles of rotation required around the X, Y, and Z axes, respectively;

[0067] For example: Δγ = -0.5° means rotating 0.5 degrees clockwise around the Z-axis;

[0068] It should be explained that these six parameters constitute the final correction instruction to be output.

[0069] The visualization is specifically delivered to the operator's AR glasses in the form of intuitive graphic arrows and numbers, including:

[0070] Three-dimensional arrows, superimposed on real components, display huge colored arrows that intuitively indicate the direction and size of movement required;

[0071] Digital label: Displays the current deviation value in real time. As the operator makes adjustments, the number will continuously change until it reaches zero.

[0072] Virtual outline: A semi-transparent virtual model that shows the final position where the component should be. The operator needs to align the real component with this virtual model.

[0073] S5. Theoretical and practical matching verification of the theoretical installation positioning model and the stable model are performed through the matching verification module.

[0074] It should be specifically noted that an automated check is performed in virtual space before or simultaneously with the output correction command. This involves inputting the theoretical installation positioning model and the stable model, performing virtual assembly and collision detection, assembling the theoretical installation positioning model and the stable model in virtual space, and conducting in-depth spatial relationship analysis. The theoretical matching verification specifically includes:

[0075] Calculate the overlap index: Based on the root mean square error (RMSE) of point cloud registration, calculate the square root of the mean of the sum of squares of the distances between the nearest points in the two point clouds. The smaller the RMSE value, the higher the overlap and the more accurate the localization.

[0076] Check for interference and collisions: Check whether the theoretically installed positioning model penetrates the surrounding stable structure;

[0077] Verify key interfaces: Pay attention to the matching of key interfaces such as bolt hole alignment and connection plate gap.

[0078] Based on the preset thresholds of RMSE < 2mm, no penetration, and interface matching, a theoretical matching verification judgment is made:

[0079] If any condition is not met, the process fails: this means there is a major problem with the digital model itself, triggering an alarm and suspending the sending of corrective instructions to the operator, prompting technical personnel to intervene and investigate.

[0080] If all conditions are met, the system passes: This means that the location of the digital twin model is theoretically correct and feasible, the system can safely send correction instructions to the operator, and the operator performs the final acceptance check after completing the physical world adjustment operation according to AR guidance.

[0081] The actual matching verification specifically refers to:

[0082] The hoisting operator wears AR glasses and operates the crane according to the 3D arrows and digital commands issued by the system, gradually adjusting the component to the required position until the deviation value on the AR display reaches zero. After the operator completes the adjustment, the system will restart the laser scanner to quickly scan the just-adjusted component and obtain its latest actual position point cloud data. This scan only needs to be performed on a single component. The final scanned point cloud is then quickly registered with the stable model using ICP, and the final actual deviation value is calculated. If the actual deviation value is within the tolerance range of ±2mm, the system indicates that the actual installation verification has passed. If the actual deviation still exceeds the range, a new set of correction commands is generated, and the operator continues to make fine adjustments. This process is repeated until the verification is passed.

[0083] S6. The installation quality evaluation module calculates the position deviation score and installation strength score based on the data after hoisting and the stability model data, and performs a weighted calculation to output the comprehensive installation quality score.

[0084] It should be specifically noted that the position deviation score is as follows:

[0085] P = w1*RMSE + w2*M + w3*I;

[0086] Where P is the position deviation score, which quantifies the geometric installation accuracy. The lower the score, the higher the positioning accuracy.

[0087] RMSE is the root mean square error, which reflects the average deviation level of the overall point cloud and measures the overall fit; M is the maximum deviation, which finds the point in the point cloud that deviates the farthest, and this value is used to control local extreme errors; I is the interface gap error, which is the most critical indicator, representing the gap at critical connections, such as the gap between node plates and the alignment deviation of bolt holes. The accuracy of these interfaces is directly related to the transmission of structural forces.

[0088] w1, w2, and w3 are weights, which are pre-set by domain experts based on extensive sample analysis or engineering experience, according to the mechanical importance of the component in the overall structure and the node type factor.

[0089] The installation strength score is a score predicting the mechanical performance guarantee. The higher the score, the closer the mechanical performance is to the ideal state. Geometric data of key interfaces are extracted, including actual bolt hole deviations, actual gaps between connecting plates, and actual angles of contact surfaces. These are then calculated using integrated mechanical simulation. Specifically:

[0090] The system incorporates a lightweight finite element analysis (FEA) engine, which is a simplified and fast mechanical simulation module. The extracted actual geometric deviation data is input into the FEA engine, which quickly simulates the mechanical response of the node under load pressure, tension, and bending moment in the installation state. It outputs the reduction coefficients of key mechanical indicators, summarizes all reduction coefficients, and normalizes them to a percentage score, namely S. If S=100, it means that there is zero loss in mechanical performance.

[0091] The comprehensive installation quality score is as follows:

[0092] A = α*(100-P) + β*S;

[0093] Where A is the overall installation quality score, the higher the score, the better the overall installation quality; 100-P converts the position deviation score into a higher score; S is the installation strength score; α and β are the weights, which are pre-set by domain experts based on a large number of sample analyses or engineering experience, according to the mechanical importance of the component in the overall structure and the node type factors.

[0094] S7. The model self-optimization module uses the initial feature vector, correction instructions, and final score of each installation as training samples to train the model, enabling continuous self-optimization.

[0095] It should be noted that each complete installation constitutes a training sample (L, X, A), where L is the feature vector of the uncertain model obtained from the initial scan, that is, the high-dimensional feature vector extracted by the deep learning network in S2. This feature vector describes the current installation scenario.

[0096] X is the correction instruction for the predicted output, which is an instruction containing 6-DOF transformation parameters. This instruction is a decision on how to adjust the components.

[0097] A represents the overall installation quality score calculated after the correction was performed. A high score indicates that the correction instructions were correct, while a low score indicates that the action was poor.

[0098] After each installation task is completed, the (L, X, A) training samples are timestamped and labeled with scene tags and stored in a dedicated experience replay buffer. The optimization process is triggered once every 100 samples collected, or when low scores occur consecutively n times.

[0099] The specific steps of the model self-optimization are as follows:

[0100] C1. Run the DDPG reinforcement learning algorithm and randomly select a batch of training samples from the experience replay buffer;

[0101] C2. Perform algorithm analysis to find a set of feature vectors L and their corresponding correction instructions X, so that the actions taken in similar situations in the future can maximize the expected comprehensive installation quality score A.

[0102] C3. Fine-tune the point cloud registration deep learning network in S2 using the gradient descent algorithm. The adjustment direction is to enable the output of a prediction that is closer to the successfully corrected sample when a state similar to the successfully corrected sample is found.

[0103] C4. Output an updated point cloud registration model. This new model is deployed to the edge computing unit to replace the old model for the next installation task.

[0104] It is important to note that special attention should be paid to samples with abnormal installation quality scores or conditions. These data will be labeled and given higher sampling weights to ensure that they are learned more during training. By focusing on learning and analyzing failure cases, the system can make up for its shortcomings more quickly, significantly improve its processing capabilities in complex and challenging conditions, and enhance its generalization performance.

[0105] Through the above description of the embodiments, those skilled in the art can clearly understand that the various embodiments of this application can be implemented by means of software or software combined with necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware functions. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The software product is stored in a storage medium and includes several instructions to cause a computer device, such as including but not limited to a personal computer, server, or network device, to execute all or part of the steps of the method described in any embodiment of this application.

[0106] The foregoing has described exemplary embodiments of this application. It should be understood that the above exemplary embodiments are not restrictive but illustrative, and the scope of protection of this application is not limited thereto. It should be understood that those skilled in the art can make modifications and variations to the embodiments of this application without departing from the spirit and scope of this application, and such modifications and variations should be within the scope of protection of this application.

Claims

1. A method for locating steel structure components based on laser scanning and deep learning, characterized in that, Specifically, it includes: S1. The target area is scanned from multiple angles before, during and after component hoisting using a 3D laser scanning module to obtain point cloud data and generate uncertain and stable models. The stable model specifically refers to: a high-precision point cloud converted from the original BIM model, and a point cloud of installed components that have passed acceptance and are in the correct position; The uncertainty model specifically refers to the real-time point cloud obtained by scanning the component currently being hoisted; S2. Use a deep learning network to extract features from the uncertain model and match them with the stable model through the feature extraction module, and output the initial transformation matrix between the component and the design position; S3. The initial transformation matrix output by the deep learning network is combined with the original BIM model through the theoretical modeling module to generate the theoretical installation and positioning model of the component in the current state. S4. The theoretical installation positioning model and the stable model are registered with high-precision algorithm through the correction instruction generation module, the residual is calculated and the correction instruction is output and presented visually. S5. Perform theoretical and practical matching verification between the theoretical installation positioning model and the stable model through the matching verification module; S6. The installation quality evaluation module calculates the position deviation score and installation strength score based on the data after hoisting and the stability model data, and performs a weighted calculation to output the comprehensive installation quality score. S7. The model self-optimization module uses the initial feature vector, correction instructions, and final score of each installation as training samples to train the model, enabling continuous self-optimization.

2. The method for locating steel structure components based on laser scanning and deep learning according to claim 1, characterized in that: The steps for extracting features from the uncertain model and matching features with the stable model are as follows: A1. Input the uncertain model obtained by scanning the current hoisting component; A2. Input the uncertain model into a pre-trained deep learning network FCGF feature extraction network; A3. Extract features from the uncertain model; A4. Perform feature matching with the stable model; A5. Output the initial transformation matrix, which is specifically: ； Where T is the initial transformation matrix, which contains all the parameters required to rotate and translate the currently scanned component from its random initial pose to a position aligned with the original BIM model; R is a 3×3 matrix, which is the rotation matrix, and the three column vectors represent the new directions that the X, Y, and Z axes in the original coordinate system point to after the rotation. t is a 3×1 vector, a translation vector, representing the distance the object moves in the X, Y, and Z directions; the data at the bottom are the fill values ​​required for homogeneous coordinate representation.

3. The method for locating steel structure components based on laser scanning and deep learning according to claim 1, characterized in that: The theoretical installation positioning model is specifically defined as: the updated 3D model formed by applying the initial transformation matrix to each vertex of the original BIM model; the transformed vertex coordinates are specifically: V = T * Vo; Where V is the final three-dimensional coordinate of a point in the theoretical installation positioning model; T is the initial transformation matrix described in S2; and Vo is the initial three-dimensional coordinate of the corresponding point in the original BIM design model.

4. The method for locating steel structure components based on laser scanning and deep learning according to claim 3, characterized in that: The specific process of high-precision algorithm registration is as follows: B1. Nearest point search: For each point in the theoretical installation positioning model, find the nearest point in the stable model and establish a point-to-point correspondence. B2. Calculate the transformation: Based on these corresponding point pairs, calculate an optimal transformation matrix that minimizes the sum of distances between all corresponding points, including rotation and translation; B3. Apply the transformation: Apply this transformation matrix to the theoretical installation positioning model to make it closer to the stable model; B4. Iterate: Repeat the above steps to find the nearest point → calculate the new transformation → apply the transformation until the transformation value between two iterations is less than a preset threshold, or the error no longer decreases significantly. The algorithm then declares convergence and outputs the final transformation matrix.

5. The method for locating steel structure components based on laser scanning and deep learning according to claim 4, characterized in that: The specific steps for calculating the residual and outputting the correction instruction are as follows: The final transformation matrix is ​​decomposed into six degrees of freedom parameters, specifically including: three translational deviations, ΔX, ΔY, and ΔZ, which represent the distances that need to be moved in the X, Y, and Z coordinate axes, respectively; and three rotational deviations, Δα, Δβ, and Δγ, which represent the angles that need to be rotated around the X, Y, and Z axes, respectively. These six parameters constitute the final correction command to be output.

6. The method for locating steel structure components based on laser scanning and deep learning according to claim 1, characterized in that: The theoretical matching verification specifically involves: calculating the overlap index: based on the root mean square error (RMSE) of point cloud registration, calculating the square root of the mean of the sum of squares of the distances between the nearest points between the two point clouds; checking for interference and collision: checking whether the theoretical installation and positioning model penetrates the surrounding stable structure. Verify key interfaces: Pay attention to the alignment of bolt holes and the matching of gaps between connecting plates; if RMSE < 2mm, the structure has no penetration and the interfaces match, the theoretical matching verification is passed; if any one of them does not meet the requirements, the theoretical matching verification fails. The actual matching verification process is as follows: The hoisting operator adjusts the component to the position required by the instructions based on the 3D arrows and digital commands issued by the system; after the operator completes the adjustment, the system will restart the laser scanner to quickly scan the adjusted component and obtain its latest actual position point cloud data. This scan only needs to be performed on a single component; the final scanned point cloud is then quickly registered with the stable model using ICP to calculate the final actual deviation value. If the actual deviation is within the tolerance range of ±2mm, the actual installation verification is passed; if the actual deviation still exceeds the range, the actual installation verification is failed.

7. The method for locating steel structure components based on laser scanning and deep learning according to claim 1, characterized in that: The comprehensive installation quality score is as follows: A = α*(100-P) + β*S; Where A is the overall installation quality score, 100-P is the score converted from the position deviation score to a higher score, S is the installation strength score, and α and β are the weights, which are pre-set by domain experts based on a large number of sample analyses or engineering experience, according to the mechanical importance of the component in the overall structure and the node type factors. The position deviation score is specifically: P=w1*RMSE+w2*M+w3*I; where P is the position deviation score, RMSE is the root mean square error, M is the maximum deviation, which is the deviation of the point farthest from the point cloud, and I is the interface gap error, which is the gap error at the critical connection. w1, w2, and w3 are weights, pre-set by domain experts based on extensive sample analysis or engineering experience, according to the mechanical importance of the component in the overall structure and node type factors. The installation strength score is a score predicting the mechanical performance guarantee. The installation strength score is calculated by extracting the geometric data of key interfaces, including actual bolt hole deviations, actual gaps between connecting plates, and actual angles of contact surfaces. A lightweight finite element analysis (FEA) engine is built into the system. The extracted actual geometric deviation data is input into the FEA engine, which quickly simulates the mechanical response of the node under load pressure, tension, and bending moment in this installation state, outputs the reduction coefficients of key mechanical indicators, summarizes all reduction coefficients, and normalizes them to a percentage score, i.e., S.

8. The method for locating steel structure components based on laser scanning and deep learning according to claim 1, characterized in that: The specific steps of the model self-optimization are as follows: C1. Run the DDPG reinforcement learning algorithm and randomly select a batch of training samples from the experience replay buffer; C2. Perform algorithm analysis to find a set of feature vectors L and their corresponding correction instructions X, so that the actions taken in similar situations in the future can maximize the expected installation quality comprehensive score A. C3. Fine-tune the point cloud registration deep learning network using the gradient descent algorithm. The adjustment is to enable the output of a prediction that is closer to the successfully corrected sample when a state similar to the successfully corrected sample is found. C4. Output an updated point cloud registration model. This new model is deployed to the edge computing unit to replace the old model for the next installation task.

Images