A method and system for automatic assembly of a passenger car body

Abstract

This application provides a method and system for automatic assembly of bus bodies, relating to the field of automatic control. The method includes: acquiring CAD models of bus body components and a workshop assembly scene; constructing a digital assembly scene driven by a digital twin of the bus body for pre-assembly simulation; using a genetic algorithm to optimize assembly process parameters based on the simulation results of the pre-assembly simulation; completing the welding of the body frame under 3D laser vision guidance and outputting the dimensional data of the welded frame; obtaining the body structure based on the frame dimensional data and optimized assembly process parameters, and adjusting the adsorption pressure during the skin bonding process using a BP neural network; performing quality inspection on the completed body structure to obtain traceability data; and using a deep Q-network based on reinforcement learning to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control commands. This application, used in the automatic assembly process of bus bodies, solves the technical problems of high rework rate and low equipment utilization in existing technologies.

Description

Technical Field

[0001] This application relates to the field of automatic control, and in particular to a method and system for automatic assembly of bus bodies. Background Technology

[0002] Existing automated bus body assembly systems typically employ offline process planning and fixed parameter control. Before production, simulation software is used to determine the welding path, assembly sequence, and process parameters, which are then executed according to the fixed parameters during actual production.

[0003] However, in real production environments, factors such as material batch performance fluctuations, changes in equipment operating status, fixture wear, and the accumulation of assembly errors across multiple processes are difficult to accurately model in the offline simulation stage, leading to deviations between actual production conditions and preset process parameters. These deviations may manifest as welding deformation and dimensional drift in the skeleton welding stage, uneven bonding and localized stress concentration in the skin bonding stage, and inconsistent connection strength or appearance defects in the riveting stage. These problems not only increase rework rates but also cause fluctuations in production cycle time and a decrease in equipment utilization. More critically, in existing systems, process parameter optimization, quality inspection analysis, and equipment scheduling strategies are typically independent: welding parameter adjustments rely mainly on experience or single-dimensional inspection results; quality inspection results are only used to determine pass / fail status without providing feedback for subsequent processes; and the scheduling system focuses on capacity allocation without considering real-time quality status and dimensional deviation risks. This "fragmented control" makes it difficult to form a closed-loop adjustment mechanism across processes, preventing dimensional deviations from being suppressed in early stages. Problems are often only discovered in the final inspection stage, leading to widespread rework. Therefore, the lack of data linkage and feedback adjustment mechanisms between various processes in the current body assembly technology leads to the inability to dynamically optimize process parameters based on real-time quality status, the gradual accumulation of dimensional errors, and the disconnect between equipment scheduling and quality control. This results in high rework rates and low equipment utilization, which are problems that urgently need to be solved. Summary of the Invention

[0004] This application provides a method and system for automatic assembly of bus bodies, which solves the technical problems of high rework rate and low equipment utilization rate in the prior art.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for automatically assembling a bus body is provided, comprising: acquiring CAD models of bus body components and a workshop assembly scenario; constructing a digital assembly scenario driven by a digital twin of the bus body components and the workshop assembly scenario for pre-assembly simulation, and using a genetic algorithm to optimize the assembly process parameters of the pre-assembly simulation results to obtain optimized assembly process parameters; based on the optimized assembly process parameters, completing the welding of the body frame under the guidance of three-dimensional laser vision, and outputting the size data of the welded frame; according to the frame size data and the optimized assembly process parameters, adjusting the adsorption pressure during the skin bonding process through a BP neural network, performing skin bonding and riveting connection with the frame, and outputting the body structure after skin bonding and riveting; performing quality inspection on the body structure after skin bonding and riveting, and outputting the quality inspection results through a support vector machine; recording the quality inspection results and corresponding process information to obtain traceability data; based on the traceability data, using a deep Q-network of reinforcement learning to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control instructions for task scheduling control of the next body assembly process.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves constructing a digital twin-driven digital assembly scenario for the bus body based on CAD models of bus body components and a workshop assembly scene. This scenario is then used for pre-assembly simulation, and a genetic algorithm is employed to optimize the assembly process parameters based on the simulation results. The optimized assembly process parameters include: constructing a digital twin assembly scene that corresponds one-to-one with the actual production line, based on the CAD models of the bus body components and the spatial layout and motion constraints of welding robots, assembly robots, tooling fixtures, and AGVs in the workshop assembly scene; and within the digital twin assembly scene, optimizing the assembly process parameters for the bus body frame... The assembly process of frame welding, skin bonding, and component installation is pre-assembled and simulated to generate simulation result data containing component interference information, tooling collision information, and equipment motion conflict information. The simulation result data is used as input to construct a genetic algorithm optimization model, and the welding path nodes, riveting sequence, and equipment operation sequence are encoded as individual genes of the genetic algorithm. Based on the genetic algorithm optimization model, the fitness function is set to minimize welding path length, minimize equipment motion conflict, and balance workstation load. The individual genes are iteratively evolved to obtain the optimal process solution that meets the optimization conditions. The optimal process solution is then analyzed into optimized assembly process parameters.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, based on optimized assembly process parameters, the welding of the vehicle body frame is completed under the guidance of 3D laser vision, and the size data of the welded frame is output. This includes: controlling the 3D laser vision system to scan the area to be welded on the vehicle body frame according to the optimized assembly process parameters, obtaining the weld position and bevel geometry, and guiding the welding robot to complete the positioning of the welding start position and welding path; during the welding process, welding timing data is collected in real time and input into the CNN-LSTM fusion model to obtain adjustment parameters for welding speed, wire feed, and number of welding layers, and dynamically controlling the welding; the welding timing data includes: welding current, voltage, molten pool temperature, and bevel size; after the frame welding is completed, the 3D laser vision system is controlled to perform an overall scan of the welded vehicle body frame to obtain the actual size data of the frame; the actual size data of the frame is compared with the standard size in the digital twin model, and when the size deviation exceeds a preset threshold, a repair welding or welding path adjustment operation is triggered, and the corresponding welding parameters are updated until the size deviation meets the preset threshold condition and the frame size data is output; when the size deviation does not exceed the preset threshold, the size data of the welded frame is output.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, based on the frame size data and optimized assembly process parameters, the adsorption pressure during the skin bonding process is adjusted using a BP neural network to perform skin bonding and frame riveting connection, outputting the completed skin bonding and riveting body structure. This includes: determining the curvature distribution, dimensional deviation, and local assembly tolerance characteristics of the skin bonding area based on the welded frame size data and optimized assembly process parameters; determining the characteristic parameters of the skin bonding area; and inputting the characteristic parameters of the skin bonding area into the BP neural network to determine the skin bonding... During the skin bonding process, the target adsorption pressure value of each adsorption unit is determined. Based on the target adsorption pressure value, the skin is adsorbed in sections to complete the flexible bonding operation. After the skin bonding operation is completed, the machine vision system is used to identify the rivet holes of the skin and the frame, and obtain the spatial position and orientation information of the rivet holes. Based on the spatial position and orientation information of the rivet holes and the optimized assembly process parameters, the rivet pressure, rivet speed and rivet sequence are determined, and the riveting connection between the skin and the frame is completed. After the skin bonding and riveting connection are completed, the body structure with the skin bonding and riveting completed is output.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, quality inspection is performed on the body structure after skin bonding and riveting, and the quality inspection results are output through a support vector machine. This includes: inspecting the body structure after skin bonding and riveting, collecting quality inspection data, including appearance images, riveting point morphology, key dimensional deviations, and assembly gaps; extracting features from the quality inspection data to obtain quality feature vectors; and inputting the quality feature vectors into the support vector machine model to determine the corresponding quality inspection results for the body structure.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, after determining the quality inspection result corresponding to the vehicle body structure, the method further includes: when the quality inspection result is unqualified, according to the defect type corresponding to the quality inspection result, calling the preset defect handling rules to perform rework operation or parameter adjustment operation; when the quality inspection result is qualified, allowing the vehicle body structure to enter the next production process or complete the off-line operation.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves using a deep Q-network based on reinforcement learning to dynamically allocate and adjust equipment tasks based on traceability data, thereby obtaining assembly scheduling control instructions for subsequent body assembly process task scheduling control. This includes: extracting equipment load status, process quality status, work-in-process quantity, and production cycle time information from the traceability data to construct a scheduling state vector; inputting the scheduling state vector into the deep Q-network model to determine the corresponding task allocation actions, including equipment selection and process sequencing; generating assembly scheduling control instructions based on the task allocation actions to control the corresponding equipment to perform the task allocation operation and obtain updated production status data; the production status data includes production efficiency, quality pass rate, and equipment utilization rate.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, after obtaining the updated production status data, the method further includes: constructing a reward function based on the updated production status data, calculating a reward value, and updating the parameters of the deep Q-network model based on the reward value.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the reward function satisfies the following formula:

[0014]

[0015] Where k takes values ​​from 1 to 5, This represents the qualified output rate per unit time. Indicates the average equipment utilization rate; Indicates the coefficient of variation of cycle time; Indicates the return rate; This indicates the percentage of abnormal line outages; These are dynamic weighting coefficients; This is a penalty term for the coupling of beat fluctuation and quality fluctuation. The coefficient of the coupling penalty term; This is a set of bottleneck workstations. To improve the utilization rate of bottleneck workstations, This is the coefficient for the bottleneck workstation utilization rate item; , This is the deviation function of production indicators from target values.

[0016] In a second aspect, an automatic assembly system for bus bodies is provided, characterized in that the system includes: a data acquisition module, a simulation module, a welding control module, a bonding control module, a quality inspection module, and a scheduling optimization module; The system comprises the following modules: a data acquisition module for acquiring CAD models of bus body components and workshop assembly scenarios; a simulation module for constructing a digital twin-driven digital assembly scenario based on the CAD models and workshop assembly scenarios of the bus body components, performing pre-assembly simulation, and using a genetic algorithm to optimize the assembly process parameters based on the simulation results of the pre-assembly simulation; a welding control module for completing the welding of the body frame under the guidance of 3D laser vision based on the optimized assembly process parameters, and outputting the size data of the welded frame; a bonding control module for adjusting the adsorption pressure during the skin bonding process using a BP neural network based on the frame size data and the optimized assembly process parameters, performing skin bonding and frame riveting connection, and outputting the body structure after skin bonding and riveting; a quality inspection module for performing quality inspection on the body structure after skin bonding and riveting, and outputting the quality inspection results through a support vector machine; and a scheduling optimization module for recording the quality inspection results and corresponding process information to obtain traceability data. Based on the traceability data, a deep Q-network using reinforcement learning is used to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control instructions and perform body assembly.

[0017] This application provides an automated assembly method and system for bus bodies. By constructing a closed-loop collaborative control mechanism encompassing through-frame welding, skin bonding and riveting, quality inspection, and equipment scheduling, process parameters can be dynamically optimized based on real-time dimensional data and quality inspection results. Furthermore, the system adaptively schedules equipment tasks in conjunction with production status, thereby suppressing the cumulative transmission of dimensional errors during assembly, improving welding and bonding consistency, and reducing rework rates. Simultaneously, by linking quality results with scheduling strategies, the system optimizes production cycle time under quality constraints, improving equipment utilization and overall production efficiency. It also enhances the system's adaptability to material fluctuations and equipment status changes, resulting in significant stability improvements and optimized overall benefits. This addresses the technical problems of high rework rates and low equipment utilization in existing bus body assembly technologies. Attached Figure Description

[0018] Figure 1 A system architecture diagram of an automated bus body assembly system provided in this application embodiment; Figure 2 A flowchart illustrating an automatic assembly method for a bus body provided in this application embodiment; Figure 3 A flowchart illustrating another automatic bus body assembly method provided in this application embodiment; Figure 4 A flowchart illustrating another automatic bus body assembly method provided in this application embodiment; Figure 5 This is a flowchart illustrating another automatic bus body assembly method provided in an embodiment of this application. Detailed Implementation

[0019] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0020] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0021] The automatic assembly method for bus bodies provided in this application embodiment can be applied to, for example... Figure 1 In the automated bus body assembly system shown, such as Figure 1 As shown, the system includes: a data acquisition module 101, a simulation module 102, a welding control module 103, a bonding control module 104, a quality inspection module 105, and a scheduling optimization module 106. The system includes: a data acquisition module 101 for acquiring CAD models of bus body components and workshop assembly scenarios; a simulation module 102 for constructing a digital assembly scenario driven by a digital twin of the bus body based on the CAD models of the bus body components and the workshop assembly scenarios, performing pre-assembly simulation, and using a genetic algorithm to optimize the assembly process parameters based on the simulation results of the pre-assembly simulation to obtain optimized assembly process parameters; a welding control module 103 for completing the welding of the body frame under the guidance of three-dimensional laser vision based on the optimized assembly process parameters, and outputting the size data of the welded frame; and a bonding control module 104 for... The dimensional data and optimized assembly process parameters are used to adjust the adsorption pressure during the skin bonding process through a BP neural network, and the skin bonding and frame riveting connection is performed, outputting the body structure with completed skin bonding and riveting. The quality inspection module 105 is used to perform quality inspection on the body structure with completed skin bonding and riveting, and outputs the quality inspection results through a support vector machine. The scheduling optimization module 106 is used to record the quality inspection results and corresponding process information to obtain traceability data. Based on the traceability data, a deep Q-network of reinforcement learning is used to dynamically allocate and adjust the equipment tasks, obtain assembly scheduling control instructions, and perform body assembly.

[0022] To address the technical problems of high rework rate and low equipment utilization in existing vehicle body assembly, this application provides an automatic assembly method for bus bodies.

[0023] Figure 2 This is a flowchart illustrating the automatic assembly method for bus bodies provided in an embodiment of this application, as shown below. Figure 2 As shown, the method includes: S201. Obtain CAD models of bus body components and workshop assembly scenes.

[0024] Among them, the CAD model of bus body components refers to the three-dimensional parametric model data of structures such as frame beams, columns, crossbeams, and skin panels; the workshop assembly scene refers to the three-dimensional workshop environment model and its equipment operation parameter data, including robots, welding stations, fixtures, conveyor lines and spatial layout information.

[0025] In one possible implementation, standardized CAD 3D models of all body components, including the frame, skin, hinges, and crossbeams, are retrieved from the bus body design end. The models contain complete design parameters such as the dimensions, materials, assembly relationships, and hole coordinates of each component. Simultaneously, a 3D scanning device is used to perform a full-area scan of the assembly workshop, collecting spatial information such as the installation positions, movement ranges, and equipment parameters of industrial robots, AGVs, CNC bending machines, and tooling fixtures, as well as workstation layouts, production line directions, and logistics channels. This forms a structured workshop assembly scene dataset. The CAD models and the workshop assembly scene dataset are then imported into the model library of the digital twin system after being formatted in a unified way to complete the data acquisition.

[0026] It should be noted that during the coordinate unification process, a global coordinate system should be established with the workshop's main reference point as a reference to avoid spatial offset between subsequent simulation and actual execution; at the same time, structural layering should be performed on large assemblies to ensure the efficiency of subsequent simulation calculations.

[0027] As an example, for a 12-meter bus, CAD models of more than 200 parts, including the body frame, side panels, and front components, were retrieved from the design stage. The bus assembly workshop was scanned using a laser 3D scanner to collect the spatial positions and technical parameters of 6 welding robots, 8 AGVs, 2 sets of CNC bending machines, and the layout information of 10 assembly stations. All data was converted into STL format and then imported into the digital twin system.

[0028] This step achieves unified modeling and spatial alignment between vehicle body structure data and actual workshop environment data, providing accurate input data for subsequent digital twin pre-assembly simulation and reducing the error amplification problem caused by virtual-real deviation.

[0029] S202. Based on the CAD model of the bus body parts and the workshop assembly scenario, a digital assembly scenario driven by the digital twin of the bus body is constructed for pre-assembly simulation. The simulation results of the pre-assembly simulation are then used to optimize the assembly process parameters to obtain the optimized assembly process parameters.

[0030] Among them, the digital twin-driven digital assembly scenario refers to a simulation system that synchronously maps the motion characteristics and process constraints of real equipment in a virtual environment; the assembly process parameters include parameters such as welding sequence, weld spacing, welding current, voltage and assembly sequence.

[0031] In one possible implementation, the vehicle body model and workshop scene are imported into the simulation platform to establish a welding path planning model and an assembly sequence constraint model. The dimensional deformation, assembly interference, and cycle time are set as evaluation indicators, and the pre-assembly simulation is run to obtain the initial simulation results. The process parameters are used as encoding genes to construct a population and perform selection, crossover, and mutation operations. The deformation and cycle time under each parameter combination are iteratively calculated, and the optimal assembly process parameters are output.

[0032] It should be noted that during the iteration process of the genetic algorithm, a convergence threshold and a maximum number of iterations need to be set to prevent overcomputation; at the same time, boundary limits should be imposed on constraints (such as the maximum allowable deformation).

[0033] As an example, the welding sequence and current value are used as optimization variables. An initial population size of 50 is set. After 30 iterations, the parameter combination with the smallest deformation and the required cycle time is selected as the output.

[0034] Based on the above steps, this step achieves global search optimization of multiple parameter combinations before production, reducing the risk of welding deformation while taking into account the production cycle, and providing a better process solution for subsequent actual welding.

[0035] S203. Based on the optimized assembly process parameters, the vehicle body frame welding is completed under the guidance of three-dimensional laser vision, and the size data of the welded frame is output.

[0036] Among them, 3D laser vision guidance refers to a control method that uses laser scanning sensors to acquire spatial point cloud data of the welding area in real time and correct the path; the skeleton size data refers to the key dimensional deviation values ​​after welding. The body skeleton welding refers to the process of connecting the steel structure components such as longitudinal beams, cross beams, and pillars of the bus body into an integral body skeleton by welding, which is the basic process of bus body assembly.

[0037] In one possible implementation, optimized assembly process parameters are sent to the welding robot and the 3D laser vision system. The 3D laser vision system performs a full-area laser scan of the body frame to be welded according to the welding area positioning requirements in the process parameters, collecting feature data such as the spatial coordinates of the weld, bevel dimensions, and angles. Using a point cloud processing algorithm, the scanned data is converted into motion path instructions recognizable by the robot, guiding it to the designated welding position with positioning accuracy controlled within a preset range. The welding robot starts the welding operation according to the welding current, voltage, welding speed, and other parameters in the process parameters. During the operation, it collects data such as the molten pool temperature and wire feed rate in real time, and automatically increases the number of welding layers in high-stress areas according to the process parameters. After the welding operation is completed, the 3D laser vision system performs another full-area scan of the entire body frame, collecting data such as the actual dimensions and geometric tolerances of each part of the frame. After data standardization processing, structured frame dimension data is generated, synchronized to the central control system, and output.

[0038] Based on the above steps, this step achieves real-time error suppression during the welding process, reduces the accumulation of structural deformation, obtains quantifiable skeleton size data, and provides a basis for subsequent bonding control.

[0039] S204. Based on the frame size data and optimized assembly process parameters, the adsorption pressure during the skin bonding process is adjusted through a BP neural network to perform skin bonding and frame riveting connection, and output the body structure with completed skin bonding and riveting.

[0040] Among them, the BP neural network is an error backpropagation neural network, which is a multi-layer feedforward neural network. It calculates the output results through forward propagation and corrects the network weights through backpropagation to achieve nonlinear fitting of the input data. In this step, it is used to adjust the adsorption pressure of the skin bonding according to the skeleton size deviation. The adsorption pressure refers to the pressure value of each suction cup when the vacuum suction cup robot adsorbs the skin. It is a key parameter affecting the tightness of the skin bonding with the skeleton. The riveting connection refers to connecting the bonded skin to the body skeleton with rivets to fix the skin to the skeleton. It is the core connection method of bus body skin assembly.

[0041] In one possible implementation, based on the dimensions of the welded frame and optimized assembly process parameters, the curvature distribution, dimensional deviations, and local assembly tolerance characteristics of the skin bonding area are determined, thus defining the characteristic parameters of the skin bonding area. These characteristic parameters are then input into a BP neural network to determine the target adsorption pressure value for each adsorption unit during the skin bonding process. Based on the target adsorption pressure value, the skin is partitioned for adsorption, completing the flexible skin bonding operation. After the skin bonding operation is completed, a machine vision system is used to identify the rivet holes between the skin and the frame, acquiring their spatial position and orientation information. Based on the spatial position and orientation information of the rivet holes and the optimized assembly process parameters, the riveting pressure, riveting speed, and riveting sequence are determined, and the riveting connection between the skin and the frame is completed. Finally, after completing the skin bonding and riveting connection, the completed body structure with skin bonding and riveting is output.

[0042] As an example, the frame dimensions of a 12-meter bus (with a dimensional deviation of ±0.4mm) and process parameters were input into a BP neural network model. The model identified a 0.4mm dimensional deviation in the right side of the frame and calculated that the suction pressure in that area should be adjusted from 0.5MPa to 0.7MPa, while keeping the pressure constant at 0.5MPa in the remaining areas. A vacuum suction robot then applied this pressure to the pre-formed skin. After bonding, the maximum gap between the skin and the frame was measured to be 0.25mm. A riveting robot applied a riveting pressure of 8MPa to complete the riveting of 200 points, ultimately outputting the bus body structure with the skin bonded and riveted.

[0043] Based on the above steps, this step can achieve adaptive and precise adjustment of the skin bonding and adsorption pressure through a BP neural network, effectively compensating for the dimensional deviation of the body frame, ensuring the tightness of the skin and frame bonding, and solving the problem of uneven gaps in traditional skin bonding; at the same time, the riveting operation is completed according to standardized process parameters, improving the firmness of the connection between the skin and the frame, and the output complete body structure can directly enter the subsequent quality inspection stage, improving the connection efficiency of the assembly process.

[0044] S205. Conduct quality inspection on the body structure after skin bonding and riveting, and output the quality inspection results through support vector machine.

[0045] Support Vector Machines (SVMs) are machine learning algorithms based on statistical learning theory. They map a low-dimensional feature space to a high-dimensional feature space using kernel functions, searching for the optimal classification hyperplane to achieve accurate sample classification. In this step, SVMs are used to determine the quality level of the vehicle body assembly. The quality inspection results are determined based on the bus body assembly quality standards, after inspecting the completed skin bonding and riveting of the body structure. For example, these results may include three levels: qualified, requiring rework, and scrapped.

[0046] In one possible implementation, the completed body structure with skin bonding and riveting is inspected, and quality inspection data is collected, including appearance images, riveting point morphology, key dimensional deviations, and assembly gaps. Feature extraction is performed on the quality inspection data to obtain a quality feature vector. This quality feature vector is then input into a support vector machine model to determine the corresponding quality inspection result for the body structure. When the quality inspection result is unqualified, based on the defect type corresponding to the quality inspection result, preset defect handling rules are invoked to perform rework or parameter adjustment operations. When the quality inspection result is qualified, the body structure is allowed to proceed to the next production process or be completed and rolled off the production line.

[0047] As an example, a quality inspection was conducted on the body structure of a 12-meter bus after the skin was fitted and riveted. The visual inspection equipment did not identify welding defects, riveting defects, or skin surface problems. The dimensional inspection instrument detected that the maximum fitting gap between the skin and the frame was 0.25mm, and the overall body shape and position tolerance deviation was ±0.4mm. All inspection data met industry standards. The inspection features were input into a support vector machine model, and after classification and calculation, the model output a "qualified" quality inspection result.

[0048] Based on the above steps, this step can achieve comprehensive control over the assembly quality of the vehicle body through multi-dimensional quality inspection. Combined with the support vector machine model, it can achieve accurate and rapid determination of quality level. Compared with manual judgment, it greatly improves inspection efficiency and judgment accuracy. At the same time, clear quality inspection results can promptly report assembly quality problems, which facilitates the rapid initiation of rework process, prevents unqualified products from flowing into subsequent processes, and improves the overall product quality.

[0049] S206. Record the quality inspection results and corresponding process information to obtain traceability data.

[0050] In one possible implementation, unique electronic tags are affixed to the bus body frame, skin, and key components. The system reads the tag's identification information, binds it to the corresponding quality inspection results, and simultaneously retrieves all process information for the bus body structure during assembly, including the equipment number, operator information, operation time, process parameter execution status, and equipment operating status for each process. The identification information, quality inspection results, and all process information are structured and integrated according to the principle of "one file per vehicle" to form a complete traceability dataset containing data collection time, data source, and data content. The traceability data is uploaded to the quality traceability platform for storage and simultaneously synchronized to the tags, completing the recording of traceability data.

[0051] As an example, a unique RFID tag is affixed to the body frame and side skin of a 12-meter bus, with the tag identification being "GJ12001". The system binds this tag to the "qualified" quality inspection result, and at the same time retrieves the job numbers of 6 welding robots and 2 riveting robots during the assembly process of the bus body, the information of 3 operators, the operation time and process parameter execution data of each process, integrates all the information to form a traceability dataset, uploads it to the quality traceability platform and writes it into the RFID tag.

[0052] Based on the above steps, this step can achieve integrated recording of bus body assembly quality inspection results and information of the entire process, forming a traceable and complete data archive, providing accurate and comprehensive data support for subsequent quality problem investigation, significantly shortening the source tracing time of quality problems, and at the same time, the traceability data can be used to analyze the operation quality of each process, providing data basis for process optimization.

[0053] S207. Based on traceability data, a deep Q-network using reinforcement learning is used to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control instructions, which are then used for task scheduling control in the next vehicle body assembly process.

[0054] Among them, the Deep Q-Network (DQN) reinforcement learning algorithm combines deep neural networks with Q-learning. It uses a deep neural network to fit the Q-value function, enabling action decisions in complex state spaces. In this step, it is used to achieve intelligent dynamic allocation of workshop equipment tasks based on production data. Dynamic allocation and adjustment of equipment tasks refers to adjusting the operational tasks of welding robots, AGVs, riveting robots, and other equipment in real time based on factors such as the operating status of workshop equipment, workstation load, and changes in production plans, achieving load balance between each piece of equipment and workstation. Assembly scheduling control instructions are standardized instructions output by the Deep Q-Network model, used to guide the execution of tasks and workstation coordination among workshop equipment. These instructions include equipment task allocation, AGV transfer scheduling, and assembly cycle adjustment.

[0055] In one possible implementation, based on traceability data, the load status of the equipment, the quality status of the process, the quantity of work-in-process, and the production cycle time information are extracted to construct a scheduling state vector. The scheduling state vector is then input into a deep Q-network model of reinforcement learning to determine the corresponding task allocation actions, which include equipment selection and process sequencing. Based on the task allocation actions, assembly scheduling control instructions are generated to control the corresponding equipment to perform the task allocation operations and obtain updated production status data. The production status data includes production efficiency, quality pass rate, and equipment utilization rate.

[0056] Based on the above steps, this step uses a deep Q-network model to perform in-depth analysis of traceability data, enabling intelligent dynamic allocation and adjustment of equipment tasks. This effectively solves problems such as unbalanced equipment load in the workshop and low AGV transfer efficiency, achieving collaborative and efficient operation of each workstation and equipment. At the same time, the output standardized assembly scheduling and control instructions can provide precise scheduling guidance for the next body assembly, continuously optimize the production process, improve overall assembly efficiency and production stability, and achieve closed-loop optimization of the production process.

[0057] This application's embodiment establishes an automated bus body assembly method that achieves a closed-loop control mechanism covering the entire process from digital modeling, simulation optimization, welding execution, bonding control, quality judgment, data traceability, to intelligent scheduling. Before production, the method optimizes process parameters using digital twins and genetic algorithms. During production, it combines 3D laser vision and neural networks to achieve real-time correction of dimensional deviations and adaptive adjustment of bonding pressure. After production, it uses support vector machines for quality judgment and generates traceable data. Furthermore, a deep Q-network is used to inversely influence the quality results on the equipment scheduling strategy, enabling data linkage and dynamic collaboration between process control, quality control, and production scheduling. Compared to existing methods that control each process independently, this solution effectively suppresses the cumulative transmission of dimensional errors, reduces rework rates and quality fluctuation risks, while improving equipment utilization and production cycle stability. It also enhances the system's adaptability to material fluctuations and equipment status changes, achieving synergistic optimization of quality and efficiency, and solving the technical problems of high rework rates and low equipment utilization in existing bus body assembly technologies.

[0058] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S202 can be specifically implemented through the following S301 to S305, which are explained in detail below: S301. Based on the CAD model of the bus body parts, and combined with the spatial layout and motion constraints of welding robots, assembly robots, tooling fixtures and AGVs in the workshop assembly scene, a digital twin assembly scene corresponding to the actual production line is constructed.

[0059] Motion constraints refer to the physical movement limitations of welding robots, AGVs, and other equipment, including boundary parameters such as movement range, working radius, and moving speed, which can be set by the user.

[0060] In one possible implementation, based on the CAD model data of the vehicle body parts from the simulation platform, the spatial coordinates and motion constraint parameters of the robots, tooling, and AGVs in the workshop are extracted. The 3D model is completed according to the actual production line layout, the virtual and physical workshop coordinate systems are unified, the equipment motion and component assembly constraint rules are set, and a digital twin assembly scene that is completely matched with the actual production line is constructed.

[0061] It should be noted that the core dimension error of the CAD model should be ≤ ±0.05mm, and the motion constraint parameters of the equipment should be consistent with the actual technical manual to ensure that the virtual scene corresponds one-to-one with the real space position.

[0062] As an example, the CAD model of more than 200 parts of a 12-meter bus was converted into STL format and imported into the platform. The spatial positions of 6 welding robots and 8 AGVs were restored according to the actual layout of the workshop. The ±180° rotation constraints of the robots and the fixed running channels of the AGVs were set to complete the construction of the digital twin assembly scene.

[0063] Based on the above steps, this step can accurately replicate the physical production line, providing a realistic virtual environment for subsequent pre-assembly simulation, avoiding the disconnect between simulation results and actual production, and ensuring the accuracy of equipment motion simulation.

[0064] S302. In the digital twin assembly scenario, pre-assembly simulation is performed on the assembly process of vehicle body frame welding, skin bonding and component installation, generating simulation result data containing component interference information, tooling collision information and equipment motion conflict information.

[0065] The simulation result data consists of structured problem data collected during pre-assembly simulation, including the location, type, and involved object information of component interference, tooling collision, and equipment motion conflict.

[0066] In one possible implementation, basic assembly process parameters are imported, and the entire assembly process, including skeleton welding and skin bonding, is simulated according to the actual production rhythm. The spatial status of components, tooling, and equipment is detected in real time, and interference, collision, and conflict problems are identified and recorded. After the simulation is completed, the problems are standardized and coded to generate simulation result data containing problem codes, position coordinates, and descriptions.

[0067] It should be noted that the simulation cycle time must be consistent with the actual production, the detection must be free of omissions, instantaneous conflicts or slight interference, and the result data encoding must be easy for subsequent algorithm recognition and processing.

[0068] As an example, the assembly of a 12-meter bus was simulated at a cycle time of 40 minutes per unit. Three parts interferences, two tooling collisions, and one equipment movement conflict were detected. The objects and locations involved were recorded after being coded by "type-number" to generate simulation result data.

[0069] Based on the above steps, this step can identify potential production problems in advance in a virtual environment, obtain accurate problem data without physical trial production, avoid on-site process problems from the source, reduce material and labor losses, and provide clear targets for process optimization.

[0070] S303. Using the simulation results data as input, construct a genetic algorithm optimization model, and encode the welding path nodes, riveting sequence, and equipment operation sequence into individual genes of the genetic algorithm.

[0071] Among them, the individual gene is the digital code of the welding path node, riveting sequence, and equipment operation sequence, and is the basic unit of the genetic algorithm for iterative evolution; the welding path node is the key coordinate point that determines the motion trajectory of the welding robot, including the welding start point, inflection point, and end point.

[0072] In one possible implementation, the simulation results data are used as input constraints to clarify that the optimization scheme must eliminate all the problems found; the welding path nodes are converted into decimal number strings, the riveting sequence is encoded into a number sequence, and the equipment operation sequence is encoded into a number combination, which are then spliced ​​into a unique number string as the individual gene; the population size, crossover or mutation probability are set, and a genetic algorithm optimization model is constructed.

[0073] It should be noted that individual genes must be unique and decodable, the encoding length must be adapted to the complexity of the process, and the constraints must be strictly set based on the simulation results.

[0074] As an example, with the constraint of eliminating the assembly problem of a 12-meter bus, the 10 weld path nodes, the sequence of 200 riveting holes, and the operation sequence of 6 robots are encoded into a 580-bit digital string as individual genes. The population is set to 50, the crossover probability is 0.8, and the mutation probability is 0.05 to complete the model construction.

[0075] Based on the above steps, this step transforms the optimization of process parameters into a digital iterative problem, realizes the digital expression of process parameters, ensures that the direction of algorithm optimization is consistent with production needs, and provides a standardized and feasible model foundation for subsequent iterative evolution.

[0076] S304. Based on the genetic algorithm optimization model, the fitness function is to use the minimum welding path length, the minimum equipment movement conflict, and the balanced workstation load to iteratively evolve individual genes and obtain the optimal process solution that meets the optimization conditions.

[0077] The fitness function is the core function for evaluating the quality of individual genes, with the optimization objectives being the shortest welding path, the least equipment conflict, and a balanced workstation load. The optimal process solution is the individual gene with the highest fitness value after the genetic algorithm iterations, corresponding to the optimal assembly process parameter scheme.

[0078] In one possible implementation, the genes of individuals in the initial population are decoded and their fitness values ​​are calculated. A roulette wheel method is used to select high-quality parents. Offspring are generated through crossover and mutation. This process is repeated iteratively until a preset number of iterations is reached or the fitness value stabilizes. The individual gene with the highest fitness value is selected as the optimal process solution.

[0079] It should be noted that the fitness function needs to set reasonable weights for each optimization objective, retain the best gene in each generation during iteration, and pre-set the number of iterations and the lower limit of the fitness value.

[0080] Based on the above steps, this step uses multi-objective iterative optimization to obtain the optimal process solution, which can completely solve the problems found in the simulation, take into account the core production indicators, and is more efficient and scientific than manual optimization, thus improving the rationality and optimality of process parameters.

[0081] S305. Decompose the optimal process into optimized assembly process parameters.

[0082] The analysis process involves converting the digitally optimal process solution into standardized assembly process parameters that can be executed by the production equipment.

[0083] In one possible implementation, the optimal process is decomposed into welding path, riveting sequence, and equipment operation sequence sub-codes according to the coding rules. These are then converted into actual parameters such as weld coordinates, riveting hole number sequence, and equipment station allocation. The parameters are then categorized and organized by equipment and process to generate a standardized process parameter file containing parameter names, values, and executing equipment. This file is the optimized assembly process parameter.

[0084] It should be noted that the parsing must strictly follow the coding rules, the parameters must match the equipment control protocol, and the process parameter files must be stored in a standardized and categorized manner.

[0085] Based on the above steps, this step realizes the transformation from algorithm optimization to production application. The parsed parameters can be directly sent to the equipment to ensure the rapid execution of the optimization plan. The categorized and organized files facilitate subsequent parameter management, retrieval and modification, and improve application efficiency.

[0086] This application's embodiments achieve precise replication of the physical production line through end-to-end collaboration, proactively avoiding process issues such as component interference and equipment movement conflicts, thereby reducing on-site trial production rework and material waste from the source. It utilizes genetic algorithms to achieve multi-objective scientific optimization of parameters such as welding paths and equipment operations, balancing production efficiency, equipment load, and operational smoothness, replacing manual experience-based process planning and significantly improving the efficiency and scientific rigor of parameter optimization. Simultaneously, the optimal solution from the algorithm is transformed into standardized process parameters that can be directly executed by the equipment, achieving seamless implementation from virtual simulation to actual production. The output unified process parameters provide accurate and standardized execution guidelines for subsequent body assembly, ensuring consistency and standardization of operations at each stage and reducing overall process trial-and-error costs.

[0087] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S203 can be specifically implemented through the following S401 to S405, which are explained in detail below: S401. Based on the optimized assembly process parameters, control the three-dimensional laser vision system to scan the area to be welded on the vehicle body frame, obtain the weld position and bevel geometry, and guide the welding robot to complete the positioning of the welding start position and welding path.

[0088] Among them, the geometric characteristics of the weld groove are the core parameters of the weld groove, including the groove angle, depth, width, and root gap, which are the key basis for determining the welding parameters.

[0089] In one possible implementation, the optimized process parameters are sent to the laser vision system and the welding robot. The system scans the area to be welded along a preset path, extracts point cloud data, and identifies the weld position and bevel geometry. After matching it with the standard welding path, it sends a positioning command to the robot. The robot moves to the welding start position and adjusts the welding torch posture to complete the precise positioning.

[0090] As an example, the process parameters of a 12-meter bus were sent to the equipment, and the laser system scanned 30 welds, extracting features such as a 60° bevel and an 8mm depth, to guide the positioning of 6 robots.

[0091] Based on the above steps, this step achieves automated acquisition of weld features without the need for manual measurement, improves positioning efficiency, ensures a high degree of consistency between welding operations and optimized parameters, and provides a fundamental guarantee for the stability of welding quality.

[0092] S402. During the welding process, welding timing data is collected in real time and input into the CNN-LSTM fusion model to obtain adjustment parameters for welding speed, wire feed amount and number of welding layers, and to dynamically control the welding.

[0093] The welding timing data includes welding current, voltage, molten pool temperature, and bevel size; the CNN-LSTM fusion model is a deep learning model that combines CNN local feature extraction with LSTM temporal pattern capture.

[0094] In one possible implementation, welding timing data is collected at a preset frequency, and after normalization and denoising preprocessing, it is input into a baseline CNN-LSTM fusion model. The model extracts features and captures temporal changes, outputs welding speed, wire feed amount, and welding layer number adjustment parameters, and sends the parameters to the robot to achieve dynamic welding control.

[0095] As an example, when the molten pool temperature is collected at 1500℃, the model outputs a welding speed of 8mm / s and 1 layer of welding parameters; when the temperature in the high-stress area drops to 1450℃, the model is adjusted to a speed of 6mm / s and 2 layers of welding, and the robot responds and executes in real time.

[0096] Based on the above steps, this step achieves adaptive dynamic control of the welding process, avoids defects caused by changes in molten pool temperature and groove size, improves the consistency of welding quality, and automatically adds layers in high-stress areas to enhance the strength of the skeleton structure.

[0097] S403. After the frame welding is completed, control the three-dimensional laser vision system to perform an overall scan of the welded car body frame to obtain the actual size data of the frame.

[0098] The actual dimensions of the skeleton are the geometric dimensions and form and position tolerances of the skeleton after welding, including length, width, height, component installation dimensions, straightness of welding nodes, and flatness.

[0099] In one possible implementation, after welding is completed, the skeleton is scanned without blind spots according to a preset global path to collect global view cloud data. After registration and segmentation, the actual size and geometric tolerance of the skeleton are extracted. After standardization, structured skeleton actual size data containing parameter names, measurement values ​​and positions are generated.

[0100] Based on the above steps, this step achieves automated and high-precision acquisition of skeleton dimensions, replacing manual measurement, improving detection efficiency and accuracy, and standardizing data for easy comparison with standard dimensions, providing accurate basis for deviation judgment and calibration.

[0101] S404. Compare the actual size data of the skeleton with the standard size in the digital twin model. When the size deviation exceeds the preset threshold, trigger the repair welding or welding path adjustment operation and update the corresponding welding parameters until the size deviation meets the preset threshold condition and then output the skeleton size data.

[0102] The preset threshold is the allowable range of dimensional deviations set according to the bus body assembly quality standard, which is ±0.5mm in this application.

[0103] In one possible implementation, the actual size data of the skeleton is compared with the standard size of the digital twin model parameter by parameter, the deviation value is calculated, and if it exceeds the threshold, the cause of the deviation is identified. If the welding amount is insufficient, the welding is triggered to repair; if the path deviation is triggered, the path adjustment is triggered. The corresponding welding parameters are updated and calibration is performed. After calibration, the scan is re-inspected until all deviations meet the threshold conditions, and then the skeleton size data is output.

[0104] It should be noted that the size comparison must be carried out parameter by parameter without omission, the amount of additional welding and the amount of path adjustment must be calculated accurately, and the calibration parameters must be matched with the original parameters.

[0105] Based on the above steps, this step achieves automated judgment and precise calibration of dimensional deviations, addresses deviation issues in a targeted manner, eliminates the need for manual intervention, improves correction efficiency, ensures that all output skeleton dimensions meet the standards, and avoids problems in subsequent skin bonding processes.

[0106] S405. When the dimensional deviation does not exceed the preset threshold, output the dimensional data of the welded skeleton.

[0107] In one possible implementation, if all dimensional deviations do not exceed a preset threshold, the skeleton welding is deemed qualified. The actual dimensional data is then categorized and organized to generate a standardized inspection report containing the inspection time, skeleton number, and judgment result. The report and dimensional data are stored in a central database and sent to the equipment control system for the subsequent skin bonding process to complete the data output.

[0108] This application embodiment relies on a CNN-LSTM fusion model to perform real-time analysis of welding time-series data, enabling dynamic adaptive adjustment of welding parameters. This effectively avoids welding defects caused by changes in molten pool temperature and bevel size. Simultaneously, it automatically adds layers to high-stress areas, enhancing the welding strength and quality consistency of the skeleton. After welding, the actual size data of the skeleton is obtained through laser full-domain scanning and compared with the standard dimensions of the digital twin to complete automated deviation calibration, ensuring that all skeleton size parameters meet the preset threshold requirements. This fundamentally avoids problems in subsequent processes such as skin bonding caused by skeleton size deviations.

[0109] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S207 can be specifically implemented through the following S501 to S503, which are explained in detail below: S501. Based on the traceability data, extract the equipment load status, process quality status, work-in-process quantity and production cycle information, and construct a scheduling status vector.

[0110] The scheduling state vector is a multi-dimensional numerical vector composed of quantified values ​​of production indicators such as equipment load status and process quality status, and serves as the input data for the DQN model. Equipment load status is quantified by the ratio of actual equipment operating time to rated operating time, with a value ranging from 0 to 1; the closer the value is to 1, the higher the load.

[0111] In one possible implementation, relevant data such as equipment operation records, quality inspection results, and production records are extracted from traceability data. The equipment load status and process quality status are quantified respectively, the quantity of work-in-process is counted, and the production cycle is calculated. These indicators are arranged in a preset order, normalized, and mapped to the 0-1 interval to generate a standardized scheduling state vector.

[0112] It should be noted that deep learning or machine learning methods can be used to encode the scheduling state vector, but this application does not limit this.

[0113] Based on the above steps, this step transforms massive traceability data into standardized vectors that the model can process, comprehensively and accurately reflecting the real-time status of the production line, providing standardized input data for intelligent scheduling decisions, and improving the model's processing efficiency.

[0114] S502. Input the scheduling state vector into the deep Q-network model of reinforcement learning to determine the corresponding task allocation action.

[0115] The task allocation process includes equipment selection and process sequencing.

[0116] In one possible implementation, the scheduling state vector is input into the baseline DQN model. The model extracts and fuses production state features through a hidden layer, fits the Q-values ​​of all possible actions, and uses a greedy algorithm to select the action with the highest Q-value to determine the equipment task allocation and process execution sequence.

[0117] Based on the above steps, this step uses a model to achieve precise matching between production status and scheduling actions, specifically addressing issues such as load imbalance and work-in-process accumulation. Compared to manual scheduling, it offers faster response, more scientific decision-making, and improves scheduling accuracy and rationality.

[0118] S503: Generate assembly scheduling control instructions based on task allocation actions, control the corresponding equipment to perform task allocation operations, and obtain updated production status data.

[0119] The production status data includes production efficiency, quality pass rate, and equipment utilization rate. Assembly scheduling control instructions are standardized equipment control instructions transformed from task allocation actions, containing equipment number, task content, and execution sequence. Updated production status data is the production data collected after the equipment executes the scheduling instructions, including production efficiency, quality pass rate, and equipment utilization rate.

[0120] In one possible implementation, the task allocation action is parsed, broken down into specific task requirements according to equipment type, and transformed into scheduling control instructions that match the equipment control protocol. These instructions are then sent point-to-point to the corresponding equipment, which executes the task adjustment according to the instructions. Data such as equipment operation and product quality are collected in real time to calculate production efficiency, pass rate, and equipment utilization rate, and then compiled into updated production status data.

[0121] It should be noted that after obtaining the updated production status data, a reward function is constructed based on the updated production status data, the reward value is calculated, and the parameters of the deep Q-network model are updated based on the reward value.

[0122] Preferably, the reward function satisfies the following formula:

[0123]

[0124] Where k takes values ​​from 1 to 5, This represents the qualified output rate per unit time. Indicates the average equipment utilization rate; Indicates the coefficient of variation of cycle time; Indicates the return rate; This indicates the percentage of abnormal line outages; These are dynamic weighting coefficients; This is a penalty term for the coupling of beat fluctuation and quality fluctuation. The coefficient of the coupling penalty term; This is a set of bottleneck workstations. To improve the utilization rate of bottleneck workstations, This is the coefficient for the bottleneck workstation utilization rate item; , This is the deviation function of production indicators from target values.

[0125] Based on the above steps, this step enables the rapid implementation of scheduling decisions, ensuring that the optimized solutions are applied to production in a timely manner. At the same time, the collected status data intuitively reflects the scheduling effect, providing practical data support for subsequent scheduling optimization, realizing closed-loop optimization of production scheduling, and continuously improving production efficiency and stability.

[0126] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method of automatically assembling a body of a passenger vehicle, characterized by, include: Obtain CAD models of bus body components and workshop assembly scenes; Based on the CAD model of the bus body components and the workshop assembly scenario, a digital assembly scenario driven by the digital twin of the bus body is constructed for pre-assembly simulation. A genetic algorithm is then used to optimize the assembly process parameters based on the simulation results of the pre-assembly simulation, resulting in optimized assembly process parameters. Based on the optimized assembly process parameters, the vehicle body frame welding is completed under the guidance of three-dimensional laser vision, and the size data of the welded frame is output. Based on the skeleton size data and the optimized assembly process parameters, the adsorption pressure during the skin bonding process is adjusted through a BP neural network to perform skin bonding and riveting connection with the skeleton, and output the body structure with completed skin bonding and riveting. The quality of the completed skin bonding and riveting body structure is inspected, and the quality inspection results are output through a support vector machine. The quality inspection results and corresponding process information are recorded to obtain traceability data; Based on the traceability data, a deep Q-network using reinforcement learning is used to dynamically allocate and adjust equipment tasks, thereby obtaining assembly scheduling control instructions for task scheduling control in the next vehicle body assembly process.

2. The method of claim 1, wherein, Based on the CAD model of the bus body components and the workshop assembly scenario, a digital assembly scenario driven by a digital twin of the bus body is constructed for pre-assembly simulation. A genetic algorithm is then used to optimize the assembly process parameters based on the simulation results of the pre-assembly simulation, resulting in optimized assembly process parameters, including: Based on the CAD model of the bus body parts, and combined with the spatial layout and motion constraints of welding robots, assembly robots, tooling fixtures and AGVs in the workshop assembly scene, a digital twin assembly scene corresponding to the actual production line is constructed. In the digital twin assembly scenario, the assembly process of welding the car body frame, attaching the skin, and installing parts is pre-assembled and simulated to generate simulation result data containing part interference information, tooling collision information, and equipment motion conflict information. Using the simulation results as input, a genetic algorithm optimization model is constructed, and the welding path nodes, riveting sequence, and equipment operation sequence are encoded as individual genes of the genetic algorithm. Based on the genetic algorithm optimization model, the fitness function is to use the minimum welding path length, the minimum equipment movement conflict, and the balanced workstation load to iteratively evolve the individual genes and obtain the optimal process solution that satisfies the optimization conditions. The optimal process is analyzed into optimized assembly process parameters.

3. The method of claim 1, wherein, The process of welding the vehicle body frame under three-dimensional laser vision guidance, based on the optimized assembly process parameters, and outputting the dimensional data of the welded frame, includes: Based on the optimized assembly process parameters, the three-dimensional laser vision system is controlled to scan the area to be welded on the car body frame to obtain the weld position and bevel geometry features, and guide the welding robot to complete the positioning of the welding start position and welding path. During the welding process, welding timing data is collected in real time and input into the CNN-LSTM fusion model to obtain adjustment parameters for welding speed, wire feed, and number of welding layers, thereby dynamically controlling the welding. The welding timing data includes welding current, voltage, molten pool temperature, and bevel size. After the frame welding is completed, the three-dimensional laser vision system is controlled to perform an overall scan of the welded car body frame to obtain the actual size data of the frame. The actual size data of the skeleton is compared with the standard size in the digital twin model. When the size deviation exceeds the preset threshold, the repair welding or welding path adjustment operation is triggered, and the corresponding welding parameters are updated until the size deviation meets the preset threshold condition and the skeleton size data is output. When the dimensional deviation does not exceed the preset threshold, the dimensional data of the welded skeleton is output.

4. The method of claim 1, wherein, The process involves adjusting the adsorption pressure during skin bonding using a BP neural network based on the frame size data and optimized assembly process parameters, performing skin bonding and frame riveting connection, and outputting a complete vehicle body structure with skin bonding and riveting, including: Based on the dimensions of the welded skeleton and the optimized assembly process parameters, the curvature distribution, dimensional deviation and local assembly tolerance characteristics of the skin bonding area are determined, and the characteristic parameters of the skin bonding area are determined. The feature parameters of the skin bonding area are input into a BP neural network to determine the target adsorption pressure value of each adsorption unit during the skin bonding process. Based on the target adsorption pressure value, the skin is adsorbed in sections to complete the flexible bonding operation of the skin. After the skin bonding operation is completed, the machine vision system is used to identify the rivet holes between the skin and the skeleton to obtain the spatial position and orientation information of the rivet holes. Based on the spatial position and orientation information of the rivet holes and the optimized assembly process parameters, the rivet pressure, rivet speed and rivet sequence are determined, and the rivet connection between the skin and the skeleton is completed. After completing the skin bonding and riveting connection, the vehicle body structure with the skin bonding and riveting completed is output.

5. The method of claim 1, wherein, The quality inspection of the completed skin bonding and riveting body structure, and the output of the quality inspection results through a support vector machine, includes: The body structure after skin bonding and riveting is inspected, and quality inspection data is collected, including appearance images, riveting point morphology, key dimensional deviations and assembly gaps. Feature extraction is performed on the quality inspection data to obtain a quality feature vector; The quality feature vector is input into the support vector machine model to determine the quality inspection result corresponding to the vehicle body structure.

6. The method of claim 5, wherein, After determining the quality inspection results corresponding to the vehicle body structure, the method further includes: When the quality inspection result is unqualified, the preset defect handling rules are invoked according to the defect type corresponding to the quality inspection result to perform rework operation or parameter adjustment operation; when the quality inspection result is qualified, the body structure is allowed to enter the next production process or complete the roll-off.

7. The method according to claim 1, characterized in that, Based on the traceability data, a deep Q-network using reinforcement learning is used to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control instructions, which are used for task scheduling control in the subsequent body assembly process, including: Based on the traceability data, extract the equipment load status, process quality status, work-in-process quantity and production cycle information, and construct a scheduling status vector; The scheduling state vector is input into a deep Q-network model of reinforcement learning to determine the corresponding task allocation action, which includes equipment selection and process sequencing. Based on the task allocation action, an assembly scheduling control instruction is generated to control the corresponding equipment to perform the task allocation operation and obtain updated production status data; the production status data includes production efficiency, quality pass rate and equipment utilization rate.

8. The method according to claim 7, characterized in that, After obtaining the updated production status data, the method further includes: constructing a reward function based on the updated production status data, calculating a reward value, and updating the parameters of the deep Q-network model based on the reward value.

9. The method according to claim 8, characterized in that, The reward function satisfies the following formula: Where k takes values ​​from 1 to 5, This represents the qualified output rate per unit time. Indicates the average equipment utilization rate; Indicates the coefficient of variation of cycle time; Indicates the return rate; This indicates the percentage of abnormal line outages; These are dynamic weighting coefficients; This is a penalty term for the coupling of beat fluctuation and quality fluctuation. The coefficient of the coupling penalty term; This is a set of bottleneck workstations; To improve the utilization rate of bottleneck workstations, This is the coefficient for the bottleneck workstation utilization rate item; , This is the deviation function of production indicators from target values.

10. An automated assembly system for a bus body, characterized in that, The system includes: a data acquisition module, a simulation module, a welding control module, a bonding control module, a quality inspection module, and a scheduling optimization module; The data acquisition module is used to acquire CAD models of bus body components and workshop assembly scenes; The simulation module is used to construct a digital assembly scenario driven by a digital twin of the bus body based on the CAD model of the bus body parts and the workshop assembly scenario, to perform pre-assembly simulation, and to use a genetic algorithm to optimize the assembly process parameters of the simulation results of the pre-assembly simulation to obtain the optimized assembly process parameters. The welding control module is used to complete the welding of the car body frame under the guidance of three-dimensional laser vision based on the optimized assembly process parameters, and output the size data of the welded frame. The bonding control module is used to adjust the adsorption pressure during the skin bonding process through a BP neural network based on the skeleton size data and optimized assembly process parameters, to perform skin bonding and riveting connection with the skeleton, and output the body structure with skin bonding and riveting completed. The quality inspection module is used to perform quality inspection on the body structure after the skin bonding and riveting are completed, and outputs the quality inspection results through a support vector machine. The scheduling optimization module is used to record the quality inspection results and corresponding process information to obtain traceability data; based on the traceability data, a deep Q-network using reinforcement learning is used to dynamically allocate and adjust equipment tasks to obtain assembly scheduling control instructions, and then perform vehicle body assembly.

Images