An assembly sequence optimization method for fabricated construction

Abstract

The application provides an assembled construction assembly sequence optimization method, and belongs to the technical field of computer-aided process optimization; first, radar three-dimensional point cloud data of a construction site is collected and acquired, an RGB-D sensor collects images containing color and depth information, an inertial measurement unit collects pose data of the two sensors, then the point cloud data is converted, the point cloud data coordinates are converted, the point cloud data is time-aligned, the point cloud data is fused, 3D point cloud data modeling of the construction site is performed, assembly sequence optimization based on reinforcement learning is performed, and online verification feedback of the assembly sequence is performed; each step is connected with each other, and dynamic optimization of an optimal assembly strategy is realized. Through multi-source sensor fusion, point cloud drift is effectively reduced, modeling accuracy is improved, reinforcement learning is used to autonomously optimize the assembly sequence, compared with a traditional genetic algorithm or artificial rules, the flexibility and adaptability of the assembly strategy are improved, assembly feasibility is fed back in real time, and construction reliability is improved.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided process optimization technology, and particularly relates to a method for optimizing assembly sequence in prefabricated construction. Background Technology

[0002] Prefabricated construction, with its high efficiency, environmental friendliness, and energy conservation, is becoming an important development direction in the global construction industry. Compared with traditional cast-in-place construction methods, prefabricated construction improves construction efficiency, reduces construction waste, and optimizes construction quality through the factory production and on-site assembly of prefabricated components. However, the complexity and dynamism of the construction process, involving the coordination of numerous prefabricated components, machinery, and construction personnel, present significant challenges to assembly management. In actual construction, the rationality of the assembly sequence directly affects construction efficiency and cost. An unreasonable assembly sequence can lead to construction site congestion, low machinery utilization, increased worker waiting time, and even compromised construction safety due to operational conflicts. Therefore, scientifically and rationally optimizing the assembly sequence to make the construction process smoother and more efficient is one of the core issues in prefabricated building management.

[0003] Currently, the level of intelligence in prefabricated construction still needs improvement. Although BIM (Building Information Modeling) is widely used in construction planning, it is mainly used for design and management, rather than for intelligently optimizing assembly sequences. Construction site environments are complex and ever-changing, and existing methods often rely on manual experience or static planning, making it difficult to adapt to dynamic construction needs. On the other hand, with the rapid development of the Internet of Things, sensor technology, and building digital twins, the information perception capabilities of construction sites have been greatly improved. Acquiring real-time data from construction sites through multi-source sensors and combining it with intelligent decision-making methods can effectively optimize assembly processes and improve construction efficiency. However, how to construct an accurate construction site model using this multi-source data in a dynamic environment and optimize the assembly sequence based on real-time data remains a challenge and difficulty in current research. Furthermore, uncertainties at construction sites, such as weather changes, material supply delays, and equipment failures, can all affect the execution of the assembly sequence. Therefore, how to build an intelligent optimization system that can adapt to environmental changes, making the prefabricated construction process more efficient and safer, is a key issue that urgently needs to be addressed.

[0004] Currently, the optimization methods for prefabricated construction assembly sequences mainly rely on two types of technologies:

[0005] There are two main approaches: rule-based manual optimization and mathematical model-based optimization algorithms. The former relies primarily on the experience of construction experts to formulate assembly rules and make manual adjustments. While this method can ensure the feasibility of construction to a certain extent, it has the following drawbacks: (1) poor adaptability: the rules are usually based on historical experience and are difficult to adapt to the complex and ever-changing construction site environment; (2) difficulty in handling large-scale data: as the scale of construction expands, manual optimization methods cannot efficiently handle the arrangement and combination of a large number of assembly components; (3) limited optimization effect: the optimization capability of manual rules is limited and it is difficult to find the truly optimal assembly sequence.

[0006] Optimization methods based on mathematical models, such as linear programming, integer programming, or genetic algorithms, can improve the optimization effect of assembly sequences to a certain extent, but they still have the following limitations: (1) Staticity: Many mathematical optimization methods are based on pre-set construction conditions, while construction sites are usually highly dynamic, which limits the optimization effect of these methods; (2) High computational complexity: When the assembly sequence involves a large number of components, the computational complexity of traditional optimization methods increases sharply, making it difficult to apply them to the construction site in real time; (3) Lack of real-time feedback mechanism: Most existing methods are based on offline optimization and do not fully consider the unexpected situations that may occur during construction, such as component errors and equipment scheduling conflicts, and lack a flexible adjustment mechanism.

[0007] In summary, existing assembly sequence optimization methods struggle to simultaneously achieve real-time performance, adaptability, and optimization efficiency. Therefore, there is an urgent need for an intelligent method that combines on-site perception capabilities with the ability to dynamically adjust optimization strategies to improve the efficiency and intelligence level of prefabricated construction. Summary of the Invention

[0008] To address the above problems, this invention provides a method for optimizing the assembly sequence in prefabricated construction, comprising the following steps:

[0009] S1, acquire three-dimensional point cloud data of the construction site through radar sensors, collect images containing color and depth information through RGB-D sensors, and collect pose data of the above two sensors through inertial measurement units.

[0010] S2 converts the acquired image containing color and depth information into RGB three-dimensional point cloud data;

[0011] S3 performs time alignment on radar 3D point cloud data, RGB 3D point cloud data, and pose data;

[0012] A transformation matrix is ​​constructed using pose data. Based on the constructed transformation matrix, the coordinates of the two time-aligned 3D point cloud data are transformed and mapped to the same coordinate system.

[0013] S4 performs point cloud registration on the two types of 3D point cloud data processed by S3 to obtain fused point cloud data and construct 3D point cloud data of the construction site.

[0014] S5. Construct a reinforcement learning framework, input the 3D point cloud data of the construction site, the set of components to be assembled and the construction status data into the constructed reinforcement learning framework for iterative optimization to obtain the assembly sequence.

[0015] S6 defines a feedback scoring function. Based on assembly error analysis and collision detection, the assembly sequence is optimized to obtain the maximum scoring function. If the score does not meet the standard, it returns to S6 for optimization and adjustment, and finally outputs the optimal assembly sequence.

[0016] Preferably, in step S3, a linear time interpolation formula is used to synchronize the radar 3D point cloud data to RGB 3D point cloud data in time.

[0017] Simultaneously, the pose data is interpolated over time, and the pose data at time t is calculated. D Transformation matrix T at time step IMU (t D (), and time-aligned with RGB 3D point cloud data.

[0018] Preferably, the specific process of obtaining the fused point cloud data in S4 is as follows:

[0019] First, using the extrinsic transformation matrix T L→D Perform coordinate transformation; based on the installation positions of the radar and RGB-D sensor, measure the coordinate rotation relationship between the two sensors to obtain a 3*3 rotation matrix R, and measure the spatial displacement of the two sensors to obtain a 3*1 translation vector T; therefore, the extrinsic parameter transformation matrix T L→D Includes rotation matrix R and translation vector T; T L→D The specific definition of coordinate transformation is as follows: First, use the rotation matrix R to perform matrix multiplication on the radar point cloud data, and then add the translation vector T to transform the radar point cloud data into the RGB-D coordinate system.

[0020] After converting the radar point cloud data to the RGB-D coordinate system, it is directly aligned with the RGB 3D point cloud data obtained from the RGB-D sensor to obtain a coarse registration result. Then, the Iterative Closest Point (ICP) method is used for precise registration. Specifically, based on the coarse registration result, the nearest corresponding points in the two point cloud datasets are matched using Euclidean distance, and the external parameter transformation matrix T is adjusted according to the error of each pair of corresponding points. L→D The rotation matrix R and translation vector T are optimized and adjusted to obtain the optimized extrinsic transformation matrix T. opt ; and using T opt To achieve more accurate coordinate transformations and data alignment;

[0021] Then, the nearest RGB-D color information is found for the radar point. The nearest point index is found by finding the nearest RGB-D point using the method of finding the nearest RGB-D point. The closest RGB-D sampling point is found by calculating the Euclidean distance. After matching the color information of the nearest neighbor point, the RGB-D color information (R, G, B) is assigned to the radar point cloud to generate a color point cloud.

[0022] Preferably, the specific process of constructing 3D point cloud data of the construction site in S4 is as follows:

[0023] First, the point cloud is voxelized to convert it into a regular mesh structure, thereby reducing the computational complexity of the point cloud.

[0024] Secondly, voxel downsampling is performed. After voxelization, each voxel contains multiple points. A labeled point is selected to reduce the amount of data. Through voxelization and downsampling, the amount of data is reduced, and the final 3D construction site model is obtained from P. V (i) The point cloud data is obtained.

[0025] Preferably, the reinforcement learning framework constructed in S5 enables the agent to learn from the 3D construction environment point cloud data P. V (i) The set of components to be assembled, C = {c1, c2, ..., c...} N Under the 3D point cloud data corresponding to each component, the assembly sequence S is learned. * The next optimal installation component c is determined through a reinforcement learning strategy. t The reinforcement learning framework is defined as follows:

[0026] state s t The reinforcement learning agent needs to perceive the state of the construction scene at time t, including the point cloud data of the current assembly; create a component set C, and record the sequence of currently installed components C. t = (c1, c2, ..., c t If the remaining assemblable components are: C r =(c t+1 c N );

[0027] Action a t The agent selects the next component to install. t+1 ;

[0028] Strategy π θ (a t |s t ): Given the current state s t At that time, take action a t The probability distribution;

[0029] Reward R tEvaluate the merits of the current assembly sequence and select the component based on its construction efficiency, safety, and stability;

[0030] Objective: To learn the optimal policy π θ To maximize cumulative rewards; that is, to solve for the assembly order: This represents the i-th installed component, whose index order is t. i Optimization is needed. N represents the total number of components that need to be assembled. At each step t, the agent adjusts its state based on the current state s. t Select the optimal assembly action a t That is, select the next component to install, c. t The optimization objective is:

[0031]

[0032] In the formula, γ is the discount factor, which controls the impact of future rewards on the current decision, and argmax S This means selecting the sequence with the highest cumulative reward among all assembly sequences.

[0033] Preferably, the reinforcement learning framework built on S5 uses the nearest neighbor field policy optimization (PPO) for policy optimization:

[0034] The agent uses the value function V(s) t Assess the current construction status. t The potential long-term benefits, assuming a reasonable construction sequence, are as follows: The future benefit V(s) t The advantage function A is relatively high; and, using the advantage function A t Evaluate the currently selected action a t Is it better than the average strategy?

[0035] The PPO policy is updated, and PPO uses a pruned objective function to stabilize training.

[0036] L(θ)=E π [min(r t (θ)A t ,clip(r t (θ), 1-∈, 1+∈)A t )]

[0037] In the formula, L(θ) represents the policy optimization objective function. This represents the ratio of the probability between the old and new policies, ∈ represents the PPO clipping range, used to limit the policy update magnitude, and clip(·) clips the policy ratio to ensure that the policy update does not deviate excessively.

[0038] Preferably, geometric information from the point cloud is introduced to optimize the assembly sequence. The agent learns to optimize the assembly order through a reward function R.t include:

[0039] Construction time optimization: Component c t Installation time T(c) t The efficiency of construction equipment (v) w Calculate and set the reward item R. time (t)=-T(c t The goal is to minimize construction time, therefore the reward is negative.

[0040] Structural stability constraints: Point cloud centroid calculation is introduced to determine the support relationship of components. If constraint G(c) is violated... i ) z ≥G(c j ) z If so, a negative reward will be given:

[0041]

[0042] Where, λ s G(c) represents the penalty factor. i ) z For component c i The height of the center of gravity, G(c) j ) z Indicates support member c j The height of the center of gravity;

[0043] Construction safety: Define the minimum safe distance D(c) between components. i c j If D(c) i c j )<d min If so, a punishment will be given:

[0044]

[0045] Where, d min λ represents the minimum permissible safety clearance. d This is the collision penalty factor;

[0046] The final reward function is expressed as follows:

[0047] R t =ω1R time (t)+ω2R stability (t)+ω3R safety (t)

[0048] Where ω1, ω2 and ω3 are reward weighting coefficients.

[0049] Preferably, in step S6, based on the obtained assembly sequence S*, combined with real-time construction point cloud data and assembly constraints as input, a feedback verification process is performed. Each time a component is installed, the construction scene is updated synchronously and used as actual environmental information. This allows for interactive feedback with the next component to be placed through dynamic environment modeling.

[0050] First, error analysis and position adjustment are performed for each component c in the assembly sequence. t Its theoretical installation location is P C (c t The actual installation location is determined by the on-site point cloud P. V (i), the deviation between the two is defined as ΔP(c) t If ||ΔP(c) t If || > δ, then the deviation exceeds the limit. Adjust the component position. If an assembly deviation is detected, then the compensation transformation matrix T needs to be calculated. comp ; Obtain the new position P′ of the assembled component C (c t );

[0051] Next, real-time collision detection is performed; nearest neighbor search is used for collision detection:

[0052]

[0053] In the formula, d min (c t c j ) is component c t With assembled component c j The minimum distance between them. P′ C (c t ) and P′ C (c j ) represent the components to be assembled, c and c respectively. t With assembled component c j Given point cloud data, where p and q represent one of the point clouds, if d min (c t c j )<d safe If a collision occurs, then let II(C) collision ) = 1, d safe Indicates the safe distance threshold;

[0054] Finally, a comprehensive evaluation of assembly feedback is conducted, including integrated error analysis and collision detection. A feedback scoring function is defined, and the final optimization objective is to maximize the feedback scoring function as R.

[0055] The optimization objective is determined by maximizing the feedback scoring function R to minimize assembly errors, avoid collisions, and ensure stability. When R exceeds a set threshold, feedback information is sent to S5, and the subsequent component installation sequence is re-optimized based on the current construction environment. If the threshold is not exceeded, the optimal assembly sequence S obtained in S5 is followed. * Continue assembling the components, and finally output an optimal assembly sequence.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. A Multi-Source Sensor Fusion Method for 3D Modeling of Construction Sites: This invention proposes a 3D modeling method for construction sites based on the fusion of LiDAR, RGB-D cameras, and IMU sensors. Compared with existing single-sensor modeling methods, this method fully utilizes the high precision of laser point clouds, the rich texture features of RGB-D visual information, and the attitude compensation of IMU inertial data. Through time synchronization, spatial registration, and data fusion techniques, it achieves higher-precision 3D reconstruction of construction sites.

[0058] 2. Assembly Sequence Optimization Method Based on Proximal Policy Optimization: This invention proposes an assembly sequence optimization method based on Proximal Policy Optimization (PPO). Unlike traditional genetic algorithms or manual rules, this method can autonomously learn the optimal assembly order and adapt to dynamic environmental changes. Through iterative updates of the reinforcement learning strategy, value function optimization, and exploration and utilization mechanisms, the assembly optimization strategy is continuously improved, enhancing assembly efficiency and stability.

[0059] 3. Assembly Sequence Verification and Feedback Method Based on 3D Modeling: This invention proposes a real-time verification and feedback method for assembly sequences based on 3D modeling of the construction site. Unlike existing offline simulation strategies, this method can detect the feasibility of assembly sequences in real time during construction. It employs collision detection, assembly error calculation, and other techniques, and optimizes the assembly sequence through a feedback scoring mechanism, thereby improving construction accuracy and reliability. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0061] Figure 2 A flowchart for modeling multi-source sensing data acquisition and processing.

[0062] Figure 3 This is a schematic diagram of the construction assembly components after data fusion.

[0063] Figure 4 This is a flowchart for assembly sequence optimization based on PPO reinforcement learning.

[0064] Figure 5This is a flowchart for assembly sequence verification feedback.

[0065] Figure 6 This is a comparison chart of the PPO reinforcement learning reward changes and the assembly error of the greedy algorithm in the embodiment.

[0066] Figure 7 This is an assembly sequence diagram of the 10 components in the embodiment. Detailed Implementation

[0067] This invention proposes a prefabricated construction assembly sequence optimization method based on multi-source perception and reinforcement learning to improve the efficiency and accuracy of prefabricated construction. The method mainly includes four core steps: 3D modeling of the construction site, multi-source data fusion, reinforcement learning-based assembly sequence optimization, and online verification and feedback of the assembly sequence. Each step is interconnected to achieve dynamic optimization of the optimal assembly strategy. The overall technical route is as follows: Figure 1 As shown.

[0068] The implementation process of the present invention will be further described below with reference to specific embodiments.

[0069] I. Multi-source data sensing and acquisition

[0070] Prefabricated construction sites present complex environments involving numerous dynamic interactions among components, equipment, and personnel. To achieve efficient and precise assembly, high-precision modeling of the construction site is essential to provide realistic, real-time environmental information. Currently, single-sensor methods are insufficient to meet the complex and ever-changing demands of construction sites. Therefore, this invention employs a multi-source sensing sensor architecture, including LiDAR (Light Detection and Ranging), an RGB-D camera (color-depth camera), and an IMU (Inertial Measurement Unit), to enhance the comprehensiveness, accuracy, and robustness of the data.

[0071] (1) The point cloud data P obtained by LiDAR used in this invention L =(x i y i , z i This represents the high-precision three-dimensional coordinate information (x, y) of the construction site. i y i , z i ), where i represents the i-th point in the point cloud data, x i This represents the coordinates of the point cloud data along the x-axis, i.e., the left and right positions of the point relative to the LiDAR device, and y-axis coordinates. i The z-coordinate represents the point cloud data in the y-axis direction, i.e., the position of the point relative to the front or back of the LiDAR device. iThis represents the coordinates of the point cloud data along the z-axis, i.e., the height information of the point relative to the LiDAR device. 3D point cloud data is characterized by long-distance and high-precision operation, making it suitable for scanning large-scale environments. The sampling frequency is 20Hz, and the laser line count is 64 lines. It generates approximately 1,300,000 high-density 3D point clouds per second.

[0072] (2) The RGB-D camera simultaneously acquires color I RGB With depth data D RGB It provides color images and depth information, supplementing details that LiDAR struggles to capture, such as surface textures and small objects. Depth resolution is 1280*720, RGB resolution is 1920*1080, and the frame rate is 30fps. Specific data collected includes:

[0073] I RGB ={(R u，v G u，v B u，v )} (u，v)∈Ω

[0074] D RGB (u, v) = {d u，v} (u，v)∈Ω

[0075] Where (u, v) represents a pixel. (R) u，v G u，v B u，v ) represents the RGB pixel value at (u, v). Ω represents the set of all pixels in the image; d u，v This represents the depth information corresponding to pixel (u, v).

[0076] (3) IMU pose data pose =(p x p y p z q x q y q z q w ). (p x p y p z ) represents the spatial location of the sensor, (q) x q y q z q w Let be the rotation angle represented by a quaternion. pose It provides spatial position and attitude information of the sensors, ensures spatial alignment of multiple sensor data, improves the accuracy of data fusion, updates data at a frequency of 400Hz, and uses UART / CAN transmission.

[0077] Three sensors collect initial data on all components N and the construction scenario. The data then needs to be processed to better align with the component assembly task.

[0078] Conversion of 2D and 3D point cloud data

[0079] If the initially collected data contains noise, data processing is performed to improve data quality, including data preprocessing and data fusion. Before data fusion, the data collected by each sensor needs to be denoised, transformed, and aligned. After data processing, the environmental point cloud information of the construction site is obtained. Specifically, the following steps are included:

[0080] (1) First, the LiDAR point cloud data is denoised. Since LiDAR point cloud data may contain noise, false detection points, and outliers, statistical filtering methods are used for denoising:

[0081] P′ L ={p∈P L |||p-μ||<kσ}

[0082] Among them, P′ L Let represent the denoised point cloud dataset, μ and σ be the mean and standard deviation of the local region, k be the filtering parameter, and p represent each point in the point cloud dataset.

[0083] Bilateral filtering is used to denoise the depth map to obtain D′. RGB (u, v):

[0084]

[0085] Where W represents the normalization factor, (i, j) represents the index of the neighboring pixel, and G... s (iu, jv) represents the weights calculated based on the spatial distance between pixels, using a Gaussian function; the closer the pixels are, the greater the weight. G r (D RGB (i,j)-D RGB (u, v) represents the weights used to calculate the similarity between depth values. A Gaussian function is used, and the closer the depth values ​​are, the greater the weight.

[0086] (2) Next, the color images and depth maps acquired by the RGB-D camera are converted into RGB 3D point clouds. The calculation formula is as follows:

[0087]

[0088] Where i is the index of the point cloud data point. as well as f represents the coordinates of the transformed 3D point cloud; x fy c represents the camera's focal length. x c y The coordinates of the principal point are used. Therefore, the processed RGB-D point cloud data P D Represented as:

[0089]

[0090] Where i is the index of the point cloud data point. These are the 3D point coordinates converted from RGB-D camera data. This represents the RGB pixel value corresponding to the i-th point cloud data index.

[0091] III. Temporal Interpolation of 3D Point Cloud Data

[0092] Because LiDAR and RGB-D cameras have different sampling times, time synchronization of the sensor data is necessary. Specifically, LiDAR scans point by point, while RGB-D acquires data frame by frame. Therefore, time interpolation is required for the point cloud acquired by LiDAR to align it to the timestamp t of the RGB-D acquisition frame. D This invention employs a linear time interpolation formula for time synchronization between LiDAR and RGB-D:

[0093]

[0094] in, Indicates time t D LiDAR point cloud coordinates below; Indicates t D The coordinates of the most recent LiDAR sampling point; Indicates t D The coordinates of the LiDAR sampling point at the most recent time point; Point timestamp; For point Timestamp.

[0095] Secondly, IMU-to-RGB-D time synchronization is performed. IMU sampling frequencies are typically higher than RGB-D and LiDAR, therefore time interpolation is performed on the IMU data (pose data) to calculate the IMU's position at time t. D Transformation matrix T at time step IMU (t D The IMU transformation matrix T of this invention IMU (t D The linear interpolation formula is as follows:

[0096]

[0097] in, For tD The IMU transformation matrix at the most recent time point; For t D The IMU transformation matrix at the most recent time point; and This is the corresponding IMU sampling timestamp.

[0098] Next, the pose data acquired by the IMU... pose Used to construct the transformation matrix T IMU .

[0099]

[0100] Among them, R IMU =(q x q y q z q w ) T The rotation matrix calculated for the IMU. t IMU =(p x p y p z ) T The position offset vector provided to the IMU. Using the IMU transformation matrix T. IMU (t D Transform LiDAR point cloud to IMU coordinate system:

[0101]

[0102] The transformation from RGB-D to IMU coordinates also uses the formula P′ D =T IMU ·P D accomplish.

[0103] The above process yields time-synchronized LiDAR data P″. L and RGB-D data P′ D .

[0104] IV. Data Fusion

[0105] Even after time synchronization is completed, LiDAR and RGB-D point clouds still have issues such as coordinate system differences, accuracy differences, and data redundancy. Therefore, data fusion is required to form a complete, high-precision, and colored 3D point cloud model.

[0106] First, coordinate transformation is performed. Because the LiDAR and RGB-D sensors are mounted in different locations, their data reside in different coordinate systems, requiring the use of an extrinsic transformation matrix T. L→D Convert LiDAR point clouds to RGB-D coordinate system.

[0107]

[0108]

[0109] in, This is a LiDAR point cloud transformed to RGB-D coordinate system. L→D It is a rotation matrix, t L→D It is a translation vector.

[0110] Secondly, after coordinate transformation, there may still be slight deviations due to sensor calibration errors, so further point cloud registration is required to make the two sets of point clouds more accurately aligned.

[0111] Optimize the transformation matrix Let be the i-th point of the LiDAR point cloud in the RGB-D coordinate system. T is the coordinate transformation matrix for the current iteration. N is the total number of points in the point cloud. After registration, the optimized T is used... opt Transform the LiDAR point cloud to align it with the RGB-D point cloud:

[0112] Because LiDAR sampling points are sparse while RGB-D sampling points are dense, it is necessary to find the nearest RGB-D color information for each LiDAR point. The nearest point index is found using the method of finding the nearest RGB-D point.

[0113]

[0114] Where, j * It is related to LiDAR points The nearest RGB-D point index. The nearest RGB-D sampling point is found by calculating the Euclidean distance. After matching the color information of the nearest neighbor point, the RGB-D color information (R, G, B) is assigned to the LiDAR point to generate a color point cloud.

[0115]

[0116] Among them, P F (i) represents the fused point cloud data; LiDAR point cloud coordinates; (R j G j B j ) represents the matched RGB-D color value.

[0117] V. 3D Construction Site Modeling with Point Cloud Integration

[0118] After fusing LiDAR and RGB-D point cloud data, a complete point cloud dataset with geometric and color information was obtained. This dataset contains spatial coordinates and color information. Next, it will be used to create a 3D construction site model to build a structured 3D environment, supporting subsequent assembly sequence optimization.

[0119] First, to reduce the computational complexity of point clouds, the point clouds are voxelized, transforming them into a regular mesh structure. Voxel index calculation:

[0120]

[0121] Among them, s v V represents the voxel size, controlling the fineness of the voxel network. V(i) represents point P. F (i) The voxel index to which the point belongs. The value is rounded down to ensure that adjacent points belong to the same voxel.

[0122] Secondly, voxel downsampling. After voxelization, each voxel may contain multiple points, so a labeled point needs to be selected to reduce the amount of data and improve computational efficiency. Voxel center point calculation:

[0123]

[0124] In the formula, P V (i) represents the representative point of volume V(i). By reducing the amount of data through voxelization and downsampling, the computational efficiency is improved. The final 3D construction site model is derived from P. V (i) The point cloud data is obtained and used for subsequent construction planning and optimization.

[0125] The overall process of data acquisition, data preprocessing, data fusion, and construction of a 3D point cloud data model of the construction site is as follows: Figure 2 As shown.

[0126] The point cloud and color data are obtained from sensors based on the original construction environment. This data is then fused to create a final 3D point cloud data illustration used for assembly sequence optimization. Figure 3 As shown.

[0127] VI. Design of Assembly Sequence Optimization Module Based on PPO Reinforcement Learning

[0128] In prefabricated construction, different components need to be installed in a reasonable sequence to ensure the stability, safety, and efficiency of the construction. Due to the complex and dynamically changing construction site environment, traditional methods based on manual rules or genetic algorithms are difficult to adapt. Therefore, this invention uses an assembly sequence optimization method based on PPO reinforcement learning, enabling the agent to optimize the assembly sequence using 3D construction environment point cloud data P... V (i) The set of components to be assembled, C = {c1, c2, ..., c...} N} 3D point cloud data corresponding to each component, learning the assembly sequence S *.

[0129] 1. Design of reinforcement learning framework

[0130] Assembly sequence optimization is modeled as a reinforcement learning framework based on Markov decision processes. The agent determines the next optimal assembly component c through a reinforcement learning policy. t This allows the assembly sequence to minimize construction time. It ensures construction safety and satisfies structural stability constraints. The reinforcement learning framework is defined as follows:

[0131] state s t The reinforcement learning agent needs to perceive the state of the construction scene at time t, including the point cloud information P of the current assembly. V (i). Create a component set C, recording the currently installed component sequence C. t = (c1, c2, ..., c t If the remaining assemblable components are: C r =(c t+1 c N ).

[0132] Action a t The agent selects the next component to install. t+1 .

[0133] Strategy π θ (a t |s t ): Given the current state s t At that time, take action a t The probability distribution.

[0134] Reward R t Evaluate the merits of the current assembly sequence. Select the component based on its construction efficiency, safety, and stability.

[0135] Objective: To learn the optimal policy π θ To maximize cumulative rewards, i.e., to determine the assembly order: This represents the i-th installed component, whose index order is t. i Optimization is needed. N is the total number of components that need to be assembled. At each step t, the agent adjusts its state based on the current state s. t Select the optimal assembly action a t That is, select the next component to install, c. t The optimization objective is:

[0136]

[0137] In the formula, γ is the discount factor, which controls the impact of future rewards on the current decision. argmax SThis means selecting the sequence with the highest cumulative reward from all possible assembly sequences.

[0138] 2. Value Function and Policy Update

[0139] In reinforcement learning, the value function is used to evaluate the quality of the current state, while the policy update determines how to optimize the assembly decision. To adapt to the dynamic environment and complex constraints of prefabricated construction, this invention designs policy optimization within the Nearest Neighbor Field Policy Optimization (PPO) framework.

[0140] The agent uses the value function V(s) t Assess the current construction status. t Potential long-term benefits in the future:

[0141]

[0142] In the formula, γ is the discount factor, controlling the impact of future rewards on the current decision. k represents the index of the future time step. If the construction sequence is reasonable, the future reward V(s) t (Higher)

[0143] Furthermore, using the advantage function A t Evaluate the currently selected action a t Is it better than the average strategy?

[0144] A t =R t +γV(s t+1 )-V(s t )

[0145] Furthermore, the PPO strategy is updated. PPO uses a pruned objective function to stabilize training.

[0146] L(θ)=E π [min(r t (θ)A t ,clip(r t (θ), 1-∈, 1+∈)A t )]

[0147] In the formula, L(θ) represents the policy optimization objective function. This represents the ratio of the old and new policy probabilities. ∈ represents the PPO clipping range, used to limit the magnitude of policy updates. clip(·) clips the policy ratio to ensure that policy updates do not deviate excessively.

[0148] 3. Optimization of assembly strategy based on 3D point cloud data

[0149] In the construction environment, assembly optimization requires point cloud data P V(i) Calculate the spatial relationships of components and influence the agent's policy decisions. This invention introduces geometric information from point clouds to optimize the assembly sequence. The agent learns to optimize the assembly order through a reward function. The reward function R... t It consists of the following parts:

[0150] 1) Construction time optimization. Component c t Installation time T(c) t The efficiency of construction equipment (v) w calculate:

[0151]

[0152] Wherein, V(c) t ) represents component c t The volume. And set the reward item R. time (t)=-T(c t The objective is to minimize construction time, therefore the reward is negative.

[0153] 2) Structural stability constraints. During assembly, some components can only be assembled after certain supporting components have been installed. Therefore, this invention introduces point cloud centroid calculation to determine the support relationship between components:

[0154]

[0155] G(c i ) z ≥G(c j ) z

[0156] In the formula, G(c t ) represents component c t The center of gravity, G(c i ) z For component c i The height of the center of gravity, G(c) j ) z Indicates support member c j The height of the center of gravity. If constraint G(c) is violated. i ) z ≥G(c j ) z If so, a negative reward will be given:

[0157]

[0158] Where, λ s This represents the penalty factor.

[0159] 3) Construction safety. Define the minimum safe distance D(c) between components. i c j ):

[0160]

[0161] Among them, P V (i) and P V (j) represent the centroids of the i-th and j-th components, respectively. If D(c i c j )<d min If so, a punishment will be given:

[0162]

[0163] Where, d min λ represents the minimum permissible safety clearance. d This is the collision penalty factor.

[0164] Therefore, the final reward function is expressed as follows:

[0165] R t =ω1R time (t)+ω2R stability (t)+ω3R safety (t)

[0166] Where ω1, ω2 and ω3 are reward weighting coefficients.

[0167] 4. Assembly sequence S * Detailed solution process

[0168] (1) Data Input. Input the 3D point cloud data of the construction site, the set of components to be assembled, and the construction status data information (the set of components to be assembled refers to the dictionary-type data consisting of the types and quantities of prefabricated building components that the construction unit plans to install. The keys of the dictionary represent the names of the prefabricated building components, and the value corresponding to each key is the quantity of that component. The construction status data is the area of ​​the current building where the components have been installed and the area where the components have not been installed, determined based on the 3D construction environment point cloud data. Specifically, it is dictionary-type data, where the keys are the areas in the building where components should be installed, and the value corresponding to each key is whether the components are installed in that area. Installation is represented by 1, and not installation is represented by 0).

[0169] (2) Through strategy π θ Select the next component. At each time step t, the strategy π θ Select an assembly component P θ (c|s t ) represents the probability distribution of the policy network output, i.e., the probability of selecting a certain component.

[0170] (3) Construction status update. Assembled component c t Then, update the construction status: st+1 =f(s) t a t Update the set of assembled components, update the point cloud data, and check if the construction constraints are met. If not, penalize the strategy.

[0171] (4) Calculate immediate rewards. Calculate immediate rewards and impose penalties if construction constraints are not met.

[0172] (5) Calculate the advantage estimate. The formula for calculating the advantage estimate measures the degree of improvement of the current strategy relative to the old strategy. If A t If it is positive, it indicates that the current strategy selection is the assembly action c. t It is effective. Otherwise, the strategy needs to be adjusted to optimize the assembly sequence.

[0173] (6) Policy Update. The policy is updated using the PPO objective function. The root policy update method calculates the L(θ) optimization objective, making the policy more inclined to select assembly orders with higher rewards. Then, the policy parameters are updated. α represents the learning rate, which controls the step size of policy updates. PPO ensures the stability of the assembly sequence optimization process by limiting the policy update magnitude.

[0174] (7) Output assembly sequence S * The output is the assembly sequence calculated based on the optimization objective in S601. This assembly sequence satisfies the optimal construction time, optimal stability, and optimal safety.

[0175] The assembly sequence optimization process based on PPO reinforcement learning is as follows: Figure 4 As shown, therefore, an assembly sequence S* can be obtained through the nearest-neighbor reinforcement learning module. This sequence is taken as the current optimal assembly sequence result. For the current component c t The assembly process is carried out, and the assembly sequence is adjusted in real time through the verification feedback module based on the environmental conditions generated by the currently assembled components.

[0176] VII. Assembly Sequence Verification Feedback Module Design

[0177] In prefabricated construction, traditional offline simulation feedback methods typically rely on preset rules or limited simulation data for evaluation, failing to adapt to real-time changes at the construction site. This can lead to deviations in the optimized assembly sequence S* during actual construction, impacting efficiency and safety. Therefore, this invention designs an assembly sequence verification feedback method based on 3D modeling. By combining online point cloud data updates for real-time collision detection, dynamic stability analysis, and sequence adjustment strategies, efficient and rapid assembly feedback is achieved. The assembly sequence verification feedback process is as follows: Figure 5 As shown.

[0178] According to the obtained assembly sequence S * Combined with real-time construction point cloud data P V (i) and assembly constraints are used as inputs. A feedback verification process is implemented. Whenever the assembly sequence optimization module places a component, the construction scene is updated synchronously and used as actual environmental information. Through dynamic environment modeling, interactive feedback is conducted with the next component to be placed.

[0179] (1) First, error analysis and position adjustment are performed. For each component c in the assembly sequence... t Its theoretical installation location is P C (c t The actual installation location is determined by the on-site point cloud P. V (i), the deviation between the two is defined as:

[0180] ΔP(c t ) = P C (c t )-P V (i)

[0181] Wherein, ΔP(c t ) represents the deviation vector between the theoretical model and the point cloud at the construction site. If ||ΔP(c t If || > δ, then the deviation exceeds the limit, and the component position needs to be adjusted. Furthermore, if an assembly deviation is detected, the compensation transformation matrix T needs to be calculated. comp :

[0182]

[0183] Among them, P C (c t ,j) represents component c t In the theoretical model, the coordinates of the j-th point cloud are given by P. V (i, j) represents the coordinates of the j-th point in the construction point cloud. T is a rigid transformation matrix, including rotation and translation. The new position of the assembled component is further calculated as follows:

[0184] P′ C (c t ) = T comp P C (c t )

[0185] (2) Subsequently, real-time collision detection is performed. To ensure that no collisions occur between components during assembly, this invention employs nearest neighbor search for collision detection:

[0186]

[0187] In the formula, d min (ct c j ) is component c t With assembled component c j The minimum distance between them. P′ C (c t ) and P′ C (c j ) represent the components to be assembled, c and c respectively. t With assembled component c j Given point cloud data, where p and q represent one point cloud instance. If d min (c t c j )<d safe If a collision occurs, then let II(C) collision ) = 1, d safe This indicates the safe distance threshold.

[0188] (3) Finally, a comprehensive evaluation of assembly feedback is performed. This evaluation considers overall error, collision detection, and stability assessments, and defines a feedback scoring function:

[0189]

[0190] In the formula, η1 and η2 are weighting parameters used to balance the influence of assembly errors, collisions, and stability in the scoring. The final optimization objective is to maximize the feedback scoring function:

[0191]

[0192] The optimization objective is to maximize the feedback scoring function R to minimize assembly errors, avoid collisions, and ensure stability. When R exceeds a set threshold, feedback information is sent to the assembly sequence optimization module, which then re-optimizes the subsequent component installation sequence based on the current construction environment. If the threshold is not exceeded, the optimal assembly sequence S obtained from the assembly sequence optimization module is used. * Continuing with component assembly, this method ultimately outputs an optimal assembly sequence.

[0193] VIII. Deployment and Application

[0194] This method is then practically deployed at construction sites or in intelligent assembly systems. The specific deployment process is as follows:

[0195] Hardware components: LiDAR sensors, RGB-D cameras, and IMU sensors are deployed to collect data at the construction site, and high-performance computing servers are used for data processing and optimization calculations. Industrial robots or automated assembly equipment are used to perform assembly tasks, and PLCs (Programmable Logic Controllers) are used to achieve on-site control.

[0196] Software components: The 3D modeling and point cloud processing module runs sensor fusion, time synchronization, point cloud stitching, and optimization algorithms to generate a high-precision 3D model of the construction site. The reinforcement learning assembly optimization module trains and optimizes based on PPO (Programmable Point Object), utilizing a computing platform to accelerate strategy training and assembly sequence optimization. The real-time feedback and adjustment module monitors the assembly status in real time during assembly and dynamically adjusts the assembly sequence using a feedback scoring function.

[0197] System integration and execution. In automated assembly systems at construction sites, the optimal assembly sequence is used as input, and 3D modeling data is combined to precisely control the assembly process. Edge computing and wireless communication are used to connect construction robots or intelligent assembly systems, enabling remote optimization and execution of the assembly sequence.

[0198] To verify the superiority of the assembly sequence optimization method of this invention, a greedy algorithm was used for comparative experiments, and its optimization effect, computational efficiency, and adaptability were evaluated. The comparison results and the assembly sequence ordering diagram generated with ten components are shown below. Figure 6 and Figure 7 As shown.

[0199] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0200] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for optimizing an assembly sequence of a fabricated construction, characterized by, The method comprises the following steps: S1, acquiring radar three-dimensional point cloud data of the construction site through a radar sensor, collecting images containing color and depth information through an RGB-D sensor, and collecting pose data of the two sensors through an inertial measurement unit; S2, converting the collected images containing color and depth information into RGB three-dimensional point cloud data; S3, time aligning the radar three-dimensional point cloud data, the RGB three-dimensional point cloud data, and the pose data; using the pose data to construct a transformation matrix, and performing coordinate conversion on the two kinds of time-aligned three-dimensional point cloud data based on the constructed transformation matrix to map them to the same coordinate system; S4, performing point cloud registration on the two kinds of three-dimensional point cloud data processed in S3 to obtain fused point cloud data and construct 3D point cloud data of the construction site; S5, constructing a reinforcement learning framework, inputting the 3D point cloud data of the construction site, a set of components to be assembled, and construction state data into the constructed reinforcement learning framework for iterative optimization to obtain an assembly sequence; Constructing a reinforcement learning framework is to make the agent learn the assembly sequence under the condition of 3D construction environment point cloud data , a set of components to be assembled , 3D point cloud data corresponding to each component ; Deciding next optimal installation component by reinforcement learning policy The reinforcement learning framework is defined as follows: State : The reinforcement learning agent needs to perceive the state of the construction scene at time t, including the current assembled point cloud data; create a set of components , record the current installed component sequence , then the remaining assemblable components: ; Actions : the agent selects the next component to install ; policy : probability distribution of actions to take given current state ​​ reward : evaluate the pros and cons of the current assembly sequence, select the construction efficiency, safety and stability of the component; Objective: learn optimal policy to maximize cumulative reward; i.e. to solve the assembly sequence: , denotes the i-th component to be installed, its index order to be optimized, N is the total number of components to be assembled, at each step t, the agent selects the optimal assembly action i.e. the next component to be installed , the optimization goal is: ​ ； wherein is a discount factor that controls the influence of future rewards on current decisions, denotes the selection of the sequence with the maximum cumulative reward among all assembly sequences; based on the constructed reinforcement learning framework, using a near field strategy to optimize PPO for policy optimization: The agent evaluates the value function The current construction state is evaluated The long-term benefits that can be obtained in the future, if the construction sequence is reasonable, the future benefits Higher; and, using the advantage function The current selected action is evaluated Whether it is better than the average strategy; performing PPO policy update, and PPO using a clipped objective function to stabilize training: ； In the formula, denotes the policy optimization objective function, denotes the ratio of the new and old policy probabilities, is the PPO clipping range, used to limit the update range of the policy, is the clipping of the policy ratio, to ensure that the policy update will not deviate too much. S6, defining a feedback score function, optimizing the assembly sequence based on assembly error analysis and collision detection to maximize the score function, and if the score does not meet the standard, returning to S6 for optimization and adjustment, and finally outputting the optimal assembly sequence.

2. The assembly sequence optimization method according to claim 1, characterized in that: in S3, linear time interpolation formula is used to synchronize the radar three-dimensional point cloud data and the RGB three-dimensional point cloud data in time; Simultaneously, the pose data is interpolated over time, and the pose data is calculated in time. Transformation matrix at time step Time alignment with RGB 3D point cloud data.

3. The method of claim 1, wherein: in S4, the specific process of obtaining the fused point cloud data is as follows: First, the extrinsic transformation matrix is used to perform coordinate conversion; according to the installation positions of the radar and the RGB-D sensor, the coordinate rotation relationship of the two sensors is measured to obtain a 3*3 rotation matrix , and the spatial displacement amount of the two sensors is measured to obtain a 3*1 translation vector ; the rotation matrix is used for matrix multiplication on the radar point cloud data, and then the translation vector is added to realize the conversion of the radar point cloud data to the RGB-D coordinate system; after converting the radar point cloud data to the RGB-D coordinate system, aligning it directly with the RGB three-dimensional point cloud data obtained by the RGB-D sensor to obtain a coarse registration result; then, using the iterative closest point (ICP) method for accurate registration; Based on the coarse registration result, the nearest corresponding points in the two point cloud data are matched by the Euclidean distance, and the rotation matrix and the translation vector of the external parameter transformation matrix are adjusted and optimized according to the error of each set of corresponding points, to obtain an optimized external parameter transformation matrix ; and more accurate coordinate transformation and data alignment are realized by using . then find the nearest RGB-D color information for the radar point, use the method of finding the nearest RGB-D point to find the nearest point index, and find the nearest RGB-D sampling point by calculating the Euclidean distance; after matching the color information of the nearest neighbor points, assign the RGB-D color information (R, G, B) to the radar point cloud to generate a color point cloud.

4. The method of claim 3, wherein the method further comprises: in S4, the specific process of constructing the 3D point cloud data of the construction site is as follows: first, voxelize the point cloud to convert it into a regular grid structure to reduce the computational complexity of the point cloud; Secondly, the voxel down-sampling is carried out, and each voxel contains multiple points after voxelization. A point is selected to reduce the data amount. Through voxelization and down-sampling, the data amount is reduced, and finally the 3D construction site model is obtained from the point cloud data composed of ​ 5. The method of claim 1, wherein: Introducing the geometric information of point cloud to optimize the assembly sequence, the agent learns to optimize the assembly sequence through a reward function, the reward function includes: Construction time optimization: component installation time is calculated from the work efficiency of the construction equipment and a reward term is set , the goal being to minimize the construction time, so the reward takes negative values; Structure stability constraint: Introduce the point cloud barycenter calculation to judge the component support relationship, if the constraint is violated then give negative reward: ； wherein represents a penalty factor, is the height of the center of gravity of the component is the height of the center of gravity of the component represents the height of the center of gravity of the support component represents the height of the center of gravity of the support component Construction safety: define minimum safety distance between components , if then give penalty: ； wherein, represents the minimum safety distance allowed, is a collision penalty factor; the final reward function is represented as follows: ； wherein , and are the reward weighting coefficients.

6. The method of claim 1, wherein: In the S6, according to the obtained assembly sequence , combined with real-time construction point cloud data and assembly constraints as input, a feedback verification process is carried out, and the construction scene is updated synchronously and used as actual environment information whenever a component is installed, and through dynamic environment modeling, the next component to be placed is interactively fed back: First, error analysis and position adjustment are performed, for each component in the assembly sequence , its theoretical installation position is , and the actual installation position is from the point cloud data on site , the deviation between the two is defined as , if , the deviation is out of limit, adjust the component position, if the assembly deviation is detected, the compensation transformation matrix needs to be calculated ; Obtaining new positions of assembly members ; then, real-time collision detection is performed; nearest neighbor search is used for collision detection: ； wherein, is the component to be assembled with the already assembled component and denote the point cloud data of the component to be assembled and the already assembled component and denote one of them, respectively, and if , denotes a safety distance threshold value; Finally, an assembly feedback synthesis evaluation is performed, combining error analysis and collision detection, defining a feedback score function, and maximizing the feedback score function as the final optimization goal ; Determining a maximization feedback score function of an optimization objective With the minimization of assembly errors, collision avoidance and stability assurance, when feedback information to S5 when the set threshold is exceeded, re-optimizing the subsequent component installation sequence based on the current existing construction environment, if the threshold is not exceeded, then following the optimal assembly sequence obtained in S5 Continue to assemble components, and finally output an optimal assembly sequence .

Images