High-precision adaptive assembly system for automobile door cover based on physical field digital twinning

Abstract

The application discloses a high-precision adaptive installation and adjustment system for a car door cover based on physical field digital twinning, comprising: a multi-modal flexible perception system for real-time collection of geometric and mechanical data of the door cover and the door frame; a digital twin and real-time simulation prediction module for constructing and synchronizing a high-fidelity virtual model and predicting the gap and surface difference after assembly based on a physical information neural network; a reinforcement learning intelligent decision model for outputting optimal adjustment instructions according to the prediction deviation; and an automatic guidance and execution module for executing the adjustment instructions and completing the assembly and tightening of the door cover. The high-precision adaptive installation and adjustment system for the car door cover based on the physical field digital twinning realizes high automation of the door cover assembly through intelligent perception, real-time optimization and adaptive control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automobile intelligent manufacturing, and particularly relates to a high-precision self-adaptive installation and adjustment system for automobile door covers based on physical field digital twinning. BACKGROUND

[0002] In the field of automobile manufacturing, door cover (including four doors and two covers) installation and adjustment is an important link in the vehicle process flow. The installation and adjustment precision of the door cover is not only related to the consistency of the vehicle appearance, but also directly affects the sealing performance, safety performance and service life. At present, there are mainly two operation modes for door cover installation in the industry: one is manual installation assisted by a power arm, and the other is automatic installation by using an industrial robot. In the manual installation mode, door cover assembly usually needs to be completed by multiple people, which has high labor intensity, and the installation and adjustment quality is largely dependent on the experience and skill level of the operator, and the consistency problem is prone to occur. In the industrial robot installation mode, although partial automation is realized, due to the problems of insufficient visual perception accuracy, single rigid positioning mode and poor dynamic working condition adaptability, the assembly one-time qualification rate is low, and it is seriously dependent on subsequent manual adjustment and finishing links to meet the installation and adjustment requirements.

[0003] With the transformation of the automobile industry to multi-model, small batch and flexible production, the traditional door cover installation and adjustment mode relying on manual experience and rigid robot teaching has been difficult to meet the comprehensive requirements of high precision, high efficiency and consistency. Therefore, the industry urgently needs a revolutionary technical solution that can break through the existing limitations. It must be able to build a real-time digital twin and intelligent agent that closely approximates the physical reality, and give the automated system the ability to perceive and think, so that it can adapt to the individual differences of each workpiece in real time like an experienced expert, and realize intelligent installation and adjustment with high precision, high efficiency and high consistency. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a high-precision self-adaptive installation and adjustment system for automobile door covers based on physical field digital twinning, which realizes high automation of door cover assembly through intelligent perception, real-time optimization and adaptive control.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A high-precision self-adaptive installation and adjustment system for automobile door covers based on physical field digital twinning, comprising:

[0007] A multi-modal flexible perception system for real-time acquisition of geometric and mechanical data of the door cover and the door frame;

[0008] A digital twin and real-time simulation prediction module for building and synchronizing a high-fidelity virtual model, and predicting the gap and surface difference after assembly based on a physical information neural network;

[0009] The reinforcement learning intelligent decision-making model is used to output optimal adjustment instructions according to the deviation of the predicted gap and surface difference;

[0010] The automatic guiding and executing module is used to execute the adjustment instructions and complete the assembly and tightening of the door cover.

[0011] Further, the multi-modal flexible perception system comprises:

[0012] The three-dimensional optical scanner is used to collect geometric data, including high-density three-dimensional point cloud data of the door cover, door frame and surrounding area, and gap and surface difference data of hundreds of key measurement points distributed along the edge of the door cover.

[0013] The six-dimensional force / torque sensor is used to collect mechanical data, including contact force and torque during assembly.

[0014] Further, the digital twin and real-time simulation prediction module comprises:

[0015] The high-fidelity virtual model contains geometric and physical property models of the door cover, door frame, hinge, sealing strip and bolt.

[0016] The physical information neural network model is used to predict the final gap and surface difference distribution of the door cover after elastic deformation under the action of gravity, assembly stress and sealing strip extrusion force.

[0017] The real-time state synchronization engine is used to synchronize the real-time data collected by the multi-modal flexible perception system with the high-fidelity virtual model at the millisecond level.

[0018] Further, the loss function of the physical information neural network model comprises a data-driven term and a physical law constraint term, expressed as:

[0019]

[0020] Wherein: is the hyperparameter for balancing the two; is the mean square error between the model prediction value and the true measurement value; is the L2 norm integral of the residual based on the Navier-Stokes equation of elasticity mechanics in the solution domain; for displacement field , the residual of which is is defined as:

[0021]

[0022] Wherein: and are the Lame parameters; is the gravity.

[0023] Further, the reinforcement learning intelligent decision model comprises an agent based on reinforcement learning, the agent taking the high-fidelity virtual model as a simulation environment for interactive training thereof, and

[0024] The state of the agent is defined as a vector composed of the gap measurement values of all the key measurement points and the deviations between the surface difference measurement values and ideal target values, denoted as:

[0025]

[0026] wherein: is a state vector; and are a gap deviation term and a surface difference deviation term, respectively, , is the number of key measurement points, and:

[0027]

[0028]

[0029] wherein: and are the gap value and the surface difference value measured for the i-th measurement point at the t-th moment; and are ideal target values of the gap and the surface difference, respectively; The action of the agent is defined as a six-degree-of-freedom fine adjustment instruction vector of three-dimensional space translation and rotation issued by the door cover or hinge adjustment actuator, denoted as:

[0030]

[0031] wherein:

[0032] is an action vector; , and are three-dimensional space translation fine adjustment instructions; , and are three-dimensional space rotation fine adjustment instructions; The reward function of the agent is used to guide the system to quickly converge to the optimal assembly quality, the reward value thereof being inversely proportional to the root mean square error of the adjusted gap and surface difference, and a negative punishment being applied to excessive adjustment actions to ensure process stability, denoted as:

[0033]

[0034]

[0035] ​​

[0036] wherein: is the root mean square error of all gap deviations and face deviation in the new state; is the reward coefficient, is a small constant to prevent the denominator from being zero; is the penalty coefficient.

[0037] Further, the automatic guiding and executing module comprises:

[0038] a door cover grabbing and positioning robot for grabbing a door cover to be assembled and coarsely positioning according to initial instructions;

[0039] a high-precision adjustment robot as an adjustment executing unit for performing micron-level accurate pose adjustment according to optimal adjustment instructions output by the reinforcement learning intelligent decision-making model;

[0040] an automatic tightening robot equipped with a high-precision servo tightening shaft and a bolt automatic feeding system for automatically completing bolt tightening work after the door cover is adjusted in place.

[0041] Further, the high-precision adjustment robot is provided with a sensor suite of the multi-modal flexible perception system at the end thereof.

[0042] Further, the method steps of the high-precision adjustment robot for performing micron-level accurate pose adjustment are:

[0043] S1: converting an action vector into a homogeneous transformation matrix :

[0044]

[0045] wherein: is a 3x3 rotation matrix generated from rotation instructions in the action vector ; and:

[0046]

[0047] wherein: , and are three-dimensional space translation fine-tuning instructions; , and are three-dimensional space rotation fine-tuning instructions;

[0048] S2: solving a target pose of a robot end effector :

[0049]

[0050] ​Wherein: is the current pose and is a 4x4 homogeneous transformation matrix describing the current position and pose of the robot end effector relative to the robot base;

[0051] Using the D-H parameter method to obtain the current pose, for a robot with n joints:

[0052] The D-H parameters of each joint are: ;

[0053] The joint angle reading is: ;

[0054] The transformation matrix for the first joint is established as: ;

[0055]

[0056] Through chain multiplication, we get:

[0057]

[0058] Wherein: represents the total transformation from the base coordinate system to the end effector coordinate system, which is equal to the end-to-end multiplication of all individual joint transformation matrices; represents the link length; represents the link torsion angle; represents the link spacing; represents the joint angle; is the number of joints;

[0059] S3: Through robot inverse kinematics, the target angles of each joint required to achieve are calculated, and a motion trajectory is generated to drive the robot to perform an adjustment action;

[0060] S4: Steps S1 to S3 are executed in a loop until is less than the preset quality threshold.

[0061] Further, it further includes a cloud-edge collaborative computing platform, including a cloud computing server and an edge computing server;

[0062] The cloud computing server is deployed with the digital twin and real-time simulation prediction module and the reinforcement learning intelligent decision model, and is used to store historical production data and high-fidelity virtual models, to perform offline training and iterative optimization of physical information neural networks and reinforcement learning intelligent decision models, and to deploy the optimized model to the edge server;

[0063] The edge computing server is deployed beside the production line, used for receiving a real-time data stream collected by the multi-modal flexible perception system, running a physical information neural network model and a reinforcement learning intelligent decision model which have been trained, performing real-time state synchronization, deviation prediction and optimal adjustment strategy calculation, and issuing optimal adjustment instructions to an automatic guiding and executing module.

[0064] The present application has the following advantages:

[0065] The present application is a high-precision adaptive installation and adjustment system for automobile door covers based on physical field digital twinning, which collects high-precision geometric and mechanical data in real time through a multi-modal flexible perception system, realizes accurate forward-looking prediction of the assembly result relying on a digital twinning and real-time simulation prediction module, and outputs an optimal adjustment strategy using a reinforcement learning intelligent decision module, finally completes micron-level adjustment and automatic tightening by a high-precision robot executing mechanism. The present application overcomes the problems of relying on manual experience, insufficient visual perception accuracy, poor rigidity positioning adaptability and the like in the prior art, realizes stable control of assembly precision at the level of 0.1mm, and improves consistency and quality level. At the same time, the system has strong flexibility and self-learning ability, can quickly adapt to production requirements of multiple vehicle models and multiple batches, completely replaces the traditional manual adjustment link, significantly reduces manufacturing cost, and provides key technical support for realizing intelligent manufacturing and "black light factory". BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to make the purpose, technical scheme and beneficial effects of the present application clearer, the present application provides the following drawings for illustration:

[0067] Figure 1 It is a principle block diagram of the embodiment of the present application, a high-precision adaptive installation and adjustment system for automobile door covers based on physical field digital twinning;

[0068] Figure 2 It is a principle diagram of a physical information neural network;

[0069] Figure 3 It is a principle diagram of a reinforcement learning intelligent decision model;

[0070] Figure 4 It is a flowchart of adaptive installation and adjustment;

[0071] Figure 5 It is a principle diagram of real-time simulation prediction and intelligent decision. DETAILED DESCRIPTION

[0072] The present application will be further described below in combination with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0073] As Figure 1As shown, the high-precision adaptive installation and adjustment system for automobile door covers based on physical field digital twinning in this embodiment includes a multi-modal flexible perception system, a digital twinning and real-time simulation prediction module, a reinforcement learning intelligent decision model, and an automatic guidance and execution module. Specifically, the multi-modal flexible perception system is used to collect geometric and mechanical data of the door cover and the door frame in real time; the digital twinning and real-time simulation prediction module is used to construct and synchronize a high-fidelity virtual model, and predict the gap and surface difference after assembly based on a physical information neural network; the reinforcement learning intelligent decision model is used to output optimal adjustment instructions according to the deviation of the predicted gap and surface difference; and the automatic guidance and execution module is used to execute the adjustment instructions and complete the assembly and tightening of the door cover.

[0074] (1) Multi-modal flexible perception system.

[0075] The multi-modal flexible perception system includes a three-dimensional optical scanner and a six-dimensional force / torque sensor. Specifically, the three-dimensional optical scanner is used to collect geometric data, including high-density three-dimensional point cloud data of the door cover, the door frame and the surrounding area, as well as gap and surface difference data of hundreds of key measurement points distributed along the edge of the door cover; the six-dimensional force / torque sensor is used to collect mechanical data, including contact force and torque during assembly. The multi-modal flexible perception system is responsible for providing high-dimensional photometric and dynamic input data streams for the real-time evolution of digital twinning.

[0076] (2) Digital twinning and real-time simulation prediction module.

[0077] The digital twinning and real-time simulation prediction module is a bridge connecting the physical entity and the virtual model, and runs on a cloud or edge computing platform. In this embodiment, the digital twinning and real-time simulation prediction module includes a high-fidelity virtual model, a physical information neural network model, and a real-time state synchronization engine.

[0078] The high-fidelity virtual model includes geometric and physical property models of the door cover, the door frame, the hinge, the sealing strip, and the bolt. Specifically, the geometric model includes accurate CAD geometric models of the door cover, the door frame, the hinge, the sealing strip, and the bolt; the physical property model defines material properties such as elastic modulus, Poisson's ratio, and density.

[0079] The physical information neural network model is used to predict the final gap and surface difference distribution of the door cover after elastic deformation under the action of gravity, assembly stress, and sealing strip extrusion force. For example, Figure 2 As shown, in this embodiment, the physical information neural network model takes the six-degree-of-freedom pose of the hinge and the pre-tightening force of the bolt as input, and can predict the final gap and surface difference distribution of the door cover after elastic deformation under the combined action of gravity, assembly stress, and sealing strip extrusion force in real time and with high precision. Specifically, in this embodiment, the loss function of the physical information neural network model includes a data-driven term and a physical law constraint term, and is expressed as:

[0080]

[0081] wherein: is a hyperparameter balancing the two; is the mean squared error between the model predicted value and the true measured value; is the L2 norm integral of the residual of the Navier-Stokes equation over the solution domain.

[0082] For the displacement field its residual is defined as:

[0083]

[0084] wherein: and are the Lame parameters; is the gravity.

[0085] The real-time state synchronization engine is used to synchronize the real-time data collected by the multi-modal flexible perception system with the high-fidelity virtual model at the millisecond level, and to map the data and the state, so as to ensure that the state of the model in the virtual space is completely consistent with the state of the physical entity.

[0086] By minimizing the total loss function, the trained physical information neural network model can accurately predict the final deformation under the action of and the tightening force , and further obtain the predicted next state .

[0087] (3) Reinforcement learning intelligent decision-making model.

[0088] As shown in Figure 3 , the reinforcement learning intelligent decision-making model includes an agent based on reinforcement learning, and the agent takes the high-fidelity virtual model as its simulation environment for interactive training.

[0089] The state of the agent is defined as a vector composed of the deviations between the gap measurement values and the ideal target values of all key measurement points and the surface difference measurement values, and is represented as:

[0090]

[0091] wherein: is the state vector; and are the gap deviation term and the surface difference deviation term, respectively, , is the number of key measurement points, and:

[0092]

[0093]

[0094] in: and They are respectively in For the first moment The gap value and surface difference value obtained from the measurement at each measuring point; and These are the ideal target values ​​for gap and surface difference, respectively.

[0095] The action of an intelligent agent is defined as: a six-degree-of-freedom fine-tuning command vector of three-dimensional spatial translation and rotation issued to the door cover or hinge adjustment actuator, expressed as:

[0096]

[0097] in: For action vectors; , and This is a three-dimensional spatial translation fine-tuning command; , and This is a three-dimensional rotation fine-tuning command.

[0098] The goal of this strategy is to maximize the expected cumulative reward. The policy network is trained by interacting with a digital twin environment at each time step. :

[0099]

[0100] in: These are the parameters of the policy network.

[0101] In performing the action Afterwards, the system transitions to a new state. The intelligent agent receives a reward. The agent's reward function guides the system to converge quickly to the optimal assembly quality. Its reward value is inversely proportional to the root mean square error of the adjusted gap and surface difference, and a negative penalty is applied to excessively large adjustment actions to ensure process smoothness. This is expressed as:

[0102]

[0103]

[0104] in: It is the root mean square error of all gap deviations and surface differences under the new condition; It is the reward coefficient. The first term is a small constant to prevent the denominator from being zero; the second term is a penalty term for adjusting the magnitude of the action. It is a penalty coefficient, designed to encourage agents to achieve their goals with smaller adjustments.

[0105] Through millions of training sessions in a digital twin environment, the agent is able to learn a complex nonlinear mapping from arbitrary biased states to optimal adjustment actions, i.e., the optimal adjustment strategy.

[0106] (4) Automated boot and execution module.

[0107] In this embodiment, the automated guidance and execution module includes a door cover gripping and positioning robot, a high-precision adjustment robot, and an automatic tightening robot.

[0108] The door cover gripping and positioning robot is used to accurately grip the door cover to be assembled from the material rack and perform coarse positioning according to the initial instructions.

[0109] The high-precision adjustment robot is used to grip door covers or hinges. As an adjustment execution unit, it performs micron-level precise pose adjustments based on the optimal adjustment instructions output by a reinforcement learning intelligent decision-making model. In this embodiment, the high-precision adjustment robot's end effector is equipped with a sensor suite of a multimodal flexible sensing system.

[0110] In this embodiment, the method steps for high-precision adjustment robot to perform micron-level precise pose adjustment are as follows.

[0111] S1: Action vector output by the decision module The motion vectors need to be converted into high-precision joint movement commands for adjusting the robot. Transform into a homogeneous transformation matrix :

[0112]

[0113] in: It is a 3x3 rotation matrix, composed of motion vectors. Rotation command in These values ​​are generated, and they are usually very small, close to zero.

[0114] and:

[0115]

[0116] in: , and This is a three-dimensional spatial translation fine-tuning command; , and This is a three-dimensional rotation fine-tuning command.

[0117] S2: Solve the target pose of the robot's end effector Target pose of robot end effector The current pose is multiplied by the homogeneous transformation matrix to obtain:

[0118]

[0119] wherein: is the current pose and is a 4x4 homogeneous transformation matrix describing the current position and attitude of the robot end effector relative to the robot base.

[0120] The current pose is obtained using the D-H parameter method, for a robot having joints:

[0121] The D-H parameters for each joint are: ;

[0122] The joint angle readings are: ;

[0123] A transformation matrix is established for the th joint:

[0124]

[0125] Through chain multiplication, we obtain:

[0126]

[0127] wherein: represents the total transformation from the base coordinate system to the end effector coordinate system, which is equal to the end-to-end multiplication of all individual joint transformation matrices; represents the link length; represents the link torsion angle; represents the link spacing; represents the joint angle; is the number of joints.

[0128] S3: Through robot inverse kinematics solving, the target angles of each joint required to achieve are calculated, and a motion trajectory is generated to drive the robot to perform the adjustment action.

[0129] S4: Steps S1 to S3 are executed in a loop until is less than the preset quality threshold.

[0130] The automatic tightening robot is equipped with a high-precision servo tightening shaft and a bolt automatic feeding system, and is responsible for automatically tightening all connecting bolts according to the torque and angle required by the process after the door cover is adjusted in place. The end of the automatic tightening robot of the embodiment is provided with a servo tightening gun with clutch control and a bolt automatic feeding mechanism, which is used to automatically complete the bolt tightening operation after the door cover is adjusted in place.

[0131] (5) A cloud-edge collaborative computing platform.

[0132] The cloud-edge collaborative computing platform of the embodiment includes a cloud computing server and an edge computing server.

[0133] Specifically, the cloud computing server is deployed with a digital twin and real-time simulation prediction module and a reinforcement learning intelligent decision model, and is used to store historical production data and high-fidelity virtual models, perform offline training and iterative optimization of the physical information neural network and the reinforcement learning intelligent decision model, and deploy the optimized model to the edge server.

[0134] The edge computing server is deployed beside the production line and is used to receive real-time data streams collected by the multi-modal flexible perception system, run the trained physical information neural network model and the reinforcement learning intelligent decision model, perform real-time state synchronization, deviation prediction and optimal adjustment strategy calculation, and issue optimal adjustment instructions to the automated guidance and execution module.

[0135] Next, taking the left front door assembly station of an automobile assembly shop as an example, the specific implementation of the automobile door cover high-precision self-adaptive assembly and adjustment system based on physical field digital twinning of the embodiment is further described in detail.

[0136] Specifically, the left front door assembly station is mainly composed of a multi-joint robot cooperation system, and is equipped with an edge computing server beside the production line.

[0137] 1. System hardware deployment.

[0138] (1) The physical entity of the automated guidance and execution module includes a door cover grabbing and positioning robot, a high-precision adjustment robot and an automatic tightening robot.

[0139] Door cover grabbing and positioning robot: a heavy-load six-axis industrial robot, the end of which is provided with a specially designed flexible door cover grabber with a pneumatic suction cup and a positioning pin. The robot is responsible for grabbing the left front door from a dedicated door cover trolley and transporting it to the predetermined position beside the body-in-white.

[0140] High-precision adjustment robot: a high-precision and high-rigidity six-axis industrial robot, whose end flange is directly connected with the door cover gripper or directly clamps the hinge body through a transition device. The robot is the main executor of fine adjustment actions, and its repeat positioning accuracy reaches the micron level.

[0141] Automatic tightening robot: a light-load six-axis robot, whose end is equipped with a high-precision servo tightening gun with clutch control and a bolt automatic feeding mechanism. It is responsible for automatically completing the tightening operation of the four connecting bolts of the upper and lower hinges after the door cover is adjusted in place.

[0142] (2) The physical entity of the multi-modal flexible sensing system is an integrated sensor suite mounted on the end effector of the high-precision adjustment robot, which specifically includes a three-dimensional optical scanner and a six-dimensional force / torque sensor.

[0143] Three-dimensional optical scanner: a lightweight blue light structured light scanner installed on the robot wrist, which can quickly obtain a million-level high-density three-dimensional point cloud of the door flange edge, vehicle body pillar and other areas.

[0144] Six-dimensional force / torque sensor: installed between the robot wrist and the end gripper, used to monitor the three-directional force and three-directional torque changes in the process of the door cover connecting with the vehicle body hinge and being pressed with the sealing strip.

[0145] 2. Cloud-edge collaborative computing platform.

[0146] The software algorithms of digital twin and real-time simulation prediction module and reinforcement learning intelligent decision module are deployed on the cloud-edge collaborative platform, which includes edge computing server and cloud computing platform.

[0147] Edge computing server: deployed beside the production line, responsible for receiving real-time data streams collected by the sensing system, running the trained physical information neural network model and reinforcement learning intelligent decision model, performing real-time state synchronization, deviation prediction and optimal adjustment strategy calculation, and issuing control instructions to the robot controller. This deployment method ensures low delay of decision-making and meets the production rhythm requirements.

[0148] Cloud computing platform: responsible for handling non-real-time and computationally intensive tasks. Including: storing massive amounts of historical production data and twin models; performing offline training and iterative optimization of physical information neural networks and reinforcement learning agents. When the cloud model is updated, the optimized model is deployed to the edge server.

[0149] 3. Self-adaptive assembly and adjustment implementation process.

[0150] As shown in Figure 4 , when a white body passes through the automatic conveying line and stops at the left front door assembly station, the assembly and adjustment process is as follows:

[0151] (1) Preparation and recognition: The labels on the body-in-white are read, and the system retrieves the standard CAD model, process parameters, and initial assembly program for this vehicle model from the database. Meanwhile, the door cover grabbing and positioning robot grabs the corresponding left front door from the trolley.

[0152] (2) Initial scanning and twin instantiation: The high-precision adjustment robot drives the three-dimensional optical scanner at its end to quickly scan the door hole edge of the body-in-white and the four surrounding edges of the left front door to be installed, respectively. The real-time point cloud data collected is sent to the edge server, fused with the standard CAD model, and a high-fidelity digital twin instance reflecting the unique tolerance of the current pair of doors and the body is generated.

[0153] (3) Rough positioning and contact sensing: The door cover grabbing robot moves the left front door to the installation preparation position of the body hinge and performs preliminary hanging. In this process, the six-dimensional force / torque sensor monitors the changes in the force and torque of the left front door when it contacts the hinge and positioning pin, and the data is updated to the digital twin model in real time.

[0154] (4) Real-time simulation prediction and intelligent decision-making: The digital twin and simulation prediction module in the edge server starts. The physical information neural network model predicts the gap and surface difference distribution state of all key measurement points after the door cover is completely fixed within milliseconds according to the initial pose and contact force of the left front door, considering the virtual extrusion force of the door self-weight and the sealing strip. The reinforcement learning intelligent decision-making module receives this predicted deviation state vector, and the agent inside it immediately outputs an optimal six-degree-of-freedom fine-tuning instruction vector. This instruction is the optimal strategy learned by the agent through millions of virtual trial-and-error training with the twin model in the cloud, aiming to reduce all deviations to a minimum within one or a few steps, as shown in the principle diagram Figure 5 .

[0155] (5) Precise adjustment and execution: After receiving this fine-tuning instruction, the high-precision adjustment robot performs complex fine-tuning of translation and rotation of the left front door or hinge with high-precision interpolation motion. At the same time, according to the actual door cover pose state, the cycle is iterated (generally iterated for 3-5 steps) until the accuracy requirement is met.

[0156] (6) Tightening and fixing: Once the adjustment is in place, the system sends a signal, and the automatic tightening robot immediately starts according to the preset tightening strategy to tighten all hinge bolts in turn.

[0157] (7) Quality verification and data back: after the tightening is completed, the three-dimensional optical scanner performs the last comprehensive gap and surface difference scanning on the left front door after the assembly is completed, and generates a digital quality report of this assembly. The measured results are used for final quality confirmation on the one hand, and compared with the predicted value in step 4 on the other hand, and the difference will be uploaded to the cloud as learning data for continuous optimization and iteration of the twin model and decision model, so that the system has self-learning and self-adaptation ability.

[0158] In the above manner, the system uses a new mode of "one-time prediction-intelligent decision-precise execution" to realize the automation, high precision and intelligentization of the automobile door cover assembly.

[0159] The above-described embodiments are only preferred embodiments for fully illustrating the present application, and the protection scope of the present application is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present application are within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. A high-precision adaptive adjustment system for a vehicle door cover based on physical field digital twinning, characterized in that: Comprise: A multi-modal flexible perception system for real-time acquisition of geometric and mechanical data of the door cover and door frame; A digital twin and real-time simulation prediction module for constructing and synchronizing a high-fidelity virtual model, and predicting the gap and face difference after assembly based on a physical information neural network; A reinforcement learning intelligent decision-making model for outputting optimal adjustment instructions according to the deviation of the predicted gap and face difference; An automated guidance and execution module for executing the adjustment instructions and completing the assembly and tightening of the door cover; The digital twin and real-time simulation prediction module comprises: A high-fidelity virtual model containing geometric and physical property models of the door cover, door frame, hinge, sealing strip and bolt; A physical information neural network model for predicting the final gap and face difference distribution of the door cover after elastic deformation under the action of gravity, assembly stress and sealing strip extrusion force; A real-time state synchronization engine for millisecond-level data synchronization and state mapping between the real-time data collected by the multi-modal flexible perception system and the high-fidelity virtual model; The loss function of the physical information neural network model includes a data-driven term and a physical law constraint term, expressed as: where: is a hyperparameter balancing the two; is the mean squared error between the model predictions and the true measurements; is the L2 norm integral over the solution domain of the residual based on the Navier- Cauchy equation of elasticity; for displacement field with residual defined as: wherein: and are Lame parameters; is the gravity.

2. The high-precision adaptive assembly and adjustment system for the automobile door cover based on the digital twinning of physical fields according to claim 1, characterized in that: The multi-modal flexible perception system comprises: A three-dimensional optical scanner for acquiring geometric data, including high-density three-dimensional point cloud data of the door cover, door frame and surrounding area, as well as gap and face difference data of hundreds of key measurement points distributed along the edge of the door cover; A six-dimensional force / torque sensor for acquiring mechanical data, including contact force and torque during assembly.

3. The high-precision adaptive assembly and adjustment system for automobile door covers based on physical field digital twinning according to claim 1, characterized in that: The reinforcement learning intelligent decision-making model includes an agent based on reinforcement learning, which uses the high-fidelity virtual model as its interactive training simulation environment, and: The state of the agent is defined as a vector composed of the deviations between the gap and face difference measurement values of all key measurement points and the ideal target values, expressed as: wherein: is a state vector; and are a gap bias term and a face bias term, respectively, , is the number of key measurement points, and: wherein: and are respectively the gap value and the surface difference value measured at the time instant t for the i-th measuring point; and are respectively the gap value and the surface difference value measured at the time instant t for the i-th measuring point; and are respectively the ideal target value of the gap and of the surface difference. The action of the agent is defined as a six-degree-of-freedom fine adjustment instruction vector for the three-dimensional space translation and rotation of the door cover or hinge adjustment execution mechanism, expressed as: wherein: is a motion vector; , and are three-dimensional spatial translation fine-tuning instructions; , and are three-dimensional spatial rotation fine-tuning instructions; The reward function of the agent is used to guide the system to quickly converge to the best assembly quality, with the reward value being inversely proportional to the root mean square error of the adjusted gap and face difference, and a negative penalty is applied to excessive adjustment actions to ensure process stability, expressed as: wherein: is the root mean square error of all gap deviations and facet deviations in the new state; is a reward coefficient, is a small constant to prevent the denominator from being zero; is a penalty coefficient.

4. The high-precision adaptive assembly and adjustment system for automobile door covers based on physical field digital twinning according to claim 1, characterized in that: The automated guidance and execution module comprises: A door cover grabbing and positioning robot for grabbing the door cover to be assembled and performing coarse positioning according to the initial instructions; A high-precision adjustment robot as an adjustment execution unit, which performs micron-level accurate pose adjustment according to the optimal adjustment instructions output by the reinforcement learning intelligent decision-making model; An automatic tightening robot equipped with a high-precision servo tightening shaft and a bolt automatic feeding system for automatically completing the bolt tightening operation after the door cover is adjusted into position.

5. The high-precision adaptive assembly system for automobile door cover based on physical field digital twinning according to claim 4, characterized in that: The end of the high-precision adjustment robot is equipped with a sensor suite of the multi-modal flexible perception system.

6. The high-precision adaptive assembly and adjustment system for automobile door covers based on physical field digital twinning according to claim 4, characterized in that: The method steps of the high-precision adjustment robot for micron-level accurate pose adjustment are: S1: convert the motion vector to a homogeneous transformation matrix : wherein: is a 3x3 rotation matrix generated from the rotation instructions in the motion vector ; and:​ wherein: , and are three-dimensional spatial translation fine-tuning instructions; , and are three-dimensional spatial rotation fine-tuning instructions; S2: Solving the target pose of the robot end effector : wherein: is the current pose and is a 4x4 homogeneous transformation matrix describing the current position and orientation of the robot end-effector with respect to the robot base; The current pose is obtained using the D-H parameter method for a robot with one joint: The D-H parameters of each joint are: ; Joint angle readings are: ; A transformation matrix is established for the first joint is: Obtained by chain multiplication: wherein: represents the total transformation from the base coordinate frame to the end effector coordinate frame, equal to the concatenated product of all individual joint transformation matrices from head to tail; represents the link length; represents the link twist angle; represents the link separation; represents the joint angle; is the number of joints; S3: Through the inverse kinematics of the robot, the target angles of each joint required to achieve are calculated, and a motion trajectory is generated to drive the robot to perform the adjustment action; S4: cyclically performing steps S1 to S3 until less than a preset quality threshold.

7. The high-precision adaptive assembly system for automobile door covers based on physical field digital twinning according to claim 1, characterized in that: Further comprising a cloud-edge collaborative computing platform including a cloud computing server and an edge computing server; The cloud computing server is deployed with the digital twin and real-time simulation prediction module and the reinforcement learning intelligent decision model, and is used for storing historical production data and high-fidelity virtual models, performing offline training and iterative optimization of the physical information neural network and the reinforcement learning intelligent decision model, and deploying the optimized model to an edge server; The edge computing server is deployed beside the production line, is used for receiving real-time data streams collected by the multi-modal flexible perception system, running the trained physical information neural network model and the reinforcement learning intelligent decision model, performing real-time state synchronization, deviation prediction and optimal adjustment strategy calculation, and issuing optimal adjustment instructions to the automation guiding and executing module.

Images