Intelligent management and control method suitable for modular assembly of aluminum alloy body frame

Abstract

This invention discloses an intelligent control method for modular assembly of aluminum alloy car body frames, relating to the field of automotive manufacturing technology. The invention divides the car body frame into multiple functional modules and assigns corresponding assembly stations. It collects internal and external pose data, station parameters, and environmental data through a multi-type sensor network. After preprocessing, it constructs digital twins covering modules, stations, and the environment, achieving precise mapping between virtual and real data. Based on multi-source data, separate process databases are established. An improved Double-DQN deep reinforcement learning algorithm is used to train the first and second models. The first model dynamically generates the optimal assembly path for parts based on internal data and completes the assembly of parts within the module. The second model generates the optimal docking path between modules based on external pose data, achieving overall docking assembly. After assembly and docking are completed, the entire process data is uploaded, and the two models are periodically updated in parallel using new data as samples. This invention achieves high precision, efficiency, and stability in assembly and docking, and has broad application value.

Description

Technical Field

[0001] This invention relates to the field of automotive manufacturing technology, and more specifically to an intelligent control method applicable to the modular assembly of aluminum alloy body frames. Background Technology

[0002] With the rapid development of the new energy vehicle industry, lightweighting has become a core technological path for the automotive industry to achieve energy conservation, emission reduction, and improved driving range. Aluminum alloy body frames, with their significant advantages of low density, high specific strength, corrosion resistance, and recyclability, are gradually replacing traditional steel bodies and becoming the mainstream choice for manufacturing bodies of new energy passenger vehicles and commercial vehicles. Currently, aluminum alloy body frame manufacturing generally adopts a modular assembly mode. This involves dividing the vehicle into independent functional units such as the front compartment module, floor module, side panel module, roof module, and rear compartment module, according to the vehicle's functional structure. Each module is assigned a dedicated assembly station. After the internal parts are assembled, precise docking and integration between modules are performed. This mode not only effectively shortens the assembly cycle and reduces the manufacturing cost per unit but also improves assembly flexibility, adapting to the industry's needs for multi-model co-production. However, aluminum alloys are inherently prone to deformation, difficult to weld, and require high assembly precision. In addition, modular assembly involves complex scenarios such as multi-station collaboration, docking of various types of parts, and multi-source data interaction, which places extremely high demands on intelligent control technology for the assembly process. It is necessary to achieve high-precision assembly of parts within the module, ensure the gap surface difference and connection stability between modules, and cope with assembly deviations caused by uncertain factors such as environmental fluctuations, equipment wear, and material batch differences during the production process. Existing control technologies are no longer sufficient to meet the needs of high-quality industrial development.

[0003] At the same time, the automotive manufacturing industry is accelerating its transformation towards intelligence and digitalization. Emerging technologies such as digital twins, multi-source sensing, and reinforcement learning are gradually being applied to the assembly field, providing technical support for the intelligent management and control of modular assembly of aluminum alloy body frames. For example:

[0004] Chinese patent (publication number CN119494422A) discloses a reinforcement learning hybrid method for assembly sequence planning. This patent proposes a Greedy-QS hybrid algorithm that combines Q-Learning reinforcement learning algorithm and Sarsa algorithm. It abstracts assembly information through Boolean matrix to achieve optimized planning of assembly sequence, focusing on solving the convergence speed and optimality problems of assembly sequence planning. However, this method only relies on the algorithm for control and cannot achieve real-time perception and virtual simulation of the assembly process. Moreover, it does not design a dedicated control model for the dual scenarios of internal parts assembly and inter-module docking in the modular assembly of aluminum alloy car body, and cannot take into account the accuracy and stability requirements of the two scenarios.

[0005] Chinese patent (publication number CN118449948B) discloses a blockchain task allocation method, apparatus, device, and medium based on DQN. This patent constructs a target blockchain task allocation model through a DQN model and performs intelligent task allocation through the target blockchain task allocation model after receiving a blockchain task allocation request. However, it only designs a duration reward function for the predicted execution time and actual execution time of blockchain task allocation, without optimizing core indicators such as accuracy, gap difference, and connection stability for aluminum alloy modular assembly, and cannot achieve accurate optimization and deviation correction of the assembly path. Moreover, the DQN model selects the action that maximizes the Q value every time during training, which often leads to the target value in the actual strategy being higher than the true value.

[0006] Furthermore, most existing assembly control methods based on digital twins only focus on defect monitoring and prediction during the parts production process, without connecting them to core modular assembly aspects such as assembly path planning and real-time accuracy adjustment. Without incorporating artificial intelligence algorithms, they cannot achieve intelligent optimization of assembly paths, nor can they cope with assembly deviations caused by equipment wear and environmental fluctuations during the assembly process, thus failing to meet the full-process control requirements of modular assembly.

[0007] In summary, existing technologies still have many defects and shortcomings. Therefore, there is an urgent need for an intelligent control method that can meet the high-precision, high-coordination, and iterative requirements of actual production to fill the gaps in existing technologies. Summary of the Invention

[0008] Based on the aforementioned technical problems, this application discloses an intelligent control method applicable to the modular assembly of aluminum alloy car body frames, specifically including:

[0009] The aluminum alloy body frame is divided into front cabin module, floor module, side panel module, roof module, and rear cabin module according to its functional structure, and assembly stations are assigned according to the modules.

[0010] Sensor networks are deployed at each assembly station to collect external pose data, internal pose data, station parameter data, and environmental data of each module, and to construct a digital twin. The stations include assembly stations and docking stations.

[0011] Based on internal pose data, assembly station parameter data, and assembly environment data, an internal assembly process database is constructed. Combined with a digital twin, the first model is trained to dynamically generate the optimal assembly path for parts and assemble the parts within each module.

[0012] After the internal assembly of each module is completed, an external docking process database is constructed based on the external part pose data, docking station parameter data, and docking environment data. Combined with the digital twin, a second model is trained to dynamically generate the optimal docking and assembly path between modules and to dock and assemble each module. Both the first and second models are constructed using the improved Double-DQN deep reinforcement learning algorithm.

[0013] After the assembly and docking are completed, the process data will be uploaded, and the first and second models will be updated and optimized regularly based on the newly uploaded data.

[0014] Preferably, the sensor network specifically comprises: deploying a lidar and a binocular structured light camera above the workstation and next to the mounting fixture inside the module, respectively, to collect the external pose of the module and the pose of the internal parts; reading the parameter data of the workstation equipment to obtain the workstation parameter data; and deploying temperature and humidity sensors and vibration sensors around the workstation to collect environmental data.

[0015] Preferably, the construction of the digital twin specifically involves: based on the collected data, building an overall framework for a digital twin model of the modular assembly of the aluminum alloy car body frame, including twins of each module, assembly station twins of each module, and assembly environment twins; wherein, the twins of each module record the external posture of each module, the geometric features of internal parts, material properties, and assembly constraints, and associate the internal posture data of the module to achieve real-time mapping between the external posture and internal state of the module; the assembly station twin corresponds to each dedicated assembly station, and records the station layout and equipment parameters; the environment twin records environmental parameters.

[0016] Preferably, the dynamic generation of the optimal assembly path for the part specifically involves: synchronously inputting the real-time collected internal module pose data, assembly station parameter data, and assembly environment data into the pre-trained first model; combining the real-time data with the relevant parameters in the internal module assembly process database, and dynamically generating the optimal assembly path for the part through the action selection strategy of the improved Double-DQN deep reinforcement learning algorithm; importing the generated optimal assembly path into a digital twin for virtual simulation verification to confirm that the path is free from interference and that the pose deviation meets the requirements; if there are problems, adjusting the path parameters until the assembly requirements are met.

[0017] The optimal assembly path that has passed verification is sent to the control terminal of the corresponding assembly station. Based on the sent assembly path, combined with the real-time collected internal pose data and assembly station parameter data, the entire assembly process, including picking up, docking, and connecting the parts inside the module, is completed. During the assembly process, the first model receives data from the sensors in real time and compares the actual assembly data with the qualified standards in the database. If pose deviation or abnormal process parameters occur, the assembly path and process parameters are adjusted to ensure assembly accuracy. After all parts inside a single module are assembled, the pose data and assembly process parameter data after assembly are collected and compared with the qualified standards to confirm that the assembly is qualified.

[0018] Preferably, the first model is specifically designed to meet the high-precision, multi-source data collaboration, and virtual simulation training requirements of the internal assembly of the aluminum alloy body frame module. It is pre-trained based on the improved Double-DQN deep reinforcement learning algorithm. The pre-trained improved Double-DQN deep reinforcement learning model is input with the module's internal pose data, assembly station parameter data, and assembly environment data, and outputs the optimal assembly path for the parts, including the part grabbing path, docking path, connection path, and the coordinates, motion parameters, and time nodes of each node on the path.

[0019] Preferably, the targeted pre-training based on the improved Double-DQN deep reinforcement learning algorithm specifically involves: the pre-training of the improved Double-DQN deep reinforcement learning algorithm aims to achieve the highest assembly accuracy, optimal assembly efficiency, and minimum error. This is combined with the simulation of actual operation using a digital twin to perform reinforcement learning training. The Q-value update formula during training is:

[0020]

[0021] in, Current action parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Let be the state transition probability. To control the impact of future gear-based action rewards on the current... Discount factor affecting the value The next gear state determined by the action selection network. The optimal equipment setup action, The action evaluation network evaluates the value of this optimal geared action. The instantaneous reward value for the internal assembly of the module is calculated based on the assembly accuracy, efficiency, and error, using the following formula:

[0022]

[0023] in, The deviation between the actual assembly position and the standard position of the internal parts of the module. It is a very small constant. Standard assembly time for internal components of the module. The actual assembly time of the internal parts of the module. To account for assembly errors of internal components of the module , , These are the weighting coefficients for accuracy, efficiency, and error, respectively.

[0024] Preferably, the optimal docking and assembly path between the dynamically generated modules is specifically as follows: accurately transfer each internally assembled qualified module to the docking station, call the twin in the completed digital twin model, import the three types of core data from the external docking process database into the twin, update the status of the module twin, perform accurate mapping between the data and the virtual docking scene between modules, and simulate the real scene of docking and assembly between modules.

[0025] The external pose data of each module, docking station parameter data, and docking environment data collected in real time are used to output the optimal docking and assembly path between modules through a pre-trained second model. The generated optimal docking and assembly path is then imported into a digital twin for virtual simulation verification. The entire docking process between modules is simulated to confirm that there is no interference in the path, the module interface is accurately aligned, and the gap and surface difference meet the requirements. If there are any problems, the path parameters are adjusted until the docking and assembly requirements between modules are met.

[0026] The optimal docking and assembly path between the verified modules is sent to the control terminal of the inter-module docking station. Combined with the real-time collected external pose data of the modules and the parameter data of the docking station, the entire process of precise positioning, interface docking, and connection of each module is completed. During the docking and assembly process, the second model receives data from the sensors in real time and compares the actual docking data with the qualified standards in the external docking process database. If there is a deviation in module pose, out-of-tolerance docking gap, or abnormal process parameters, the docking and assembly path and process parameters are adjusted. After all modules are docked and assembled, the overall pose data, interface gap / surface difference data, and process parameter data are collected and compared with the qualified docking standards to confirm that the docking and assembly is qualified.

[0027] Preferably, the second model is as follows: The second model adopts the same improved Double-DQN deep reinforcement learning algorithm framework as the first model. For the docking and assembly scenario between modules, the variable definition and parameter value are optimized. The input modules are external pose data, docking station parameter data, and docking environment data. The output is the optimal docking and assembly path between modules, including the transfer path, docking positioning path, interface connection path of each module, as well as the coordinates, docking order, action parameters and time nodes of each node in the path.

[0028] Preferably, the pre-trained second model is specifically designed to achieve the highest docking accuracy, most stable connection strength, smallest gap difference, and optimal docking efficiency as training objectives. It is trained using reinforcement learning in conjunction with an external docking process database and a digital twin, and its update formula is as follows:

[0029]

[0030] in, Current state parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Status during assembly The state transition probability, This is a discount factor, with the same value range as the first model. The next gear state determined by the action selection network. The optimal docking action is as follows. The motion evaluation network assesses the value of this optimal docking motion. The instant bonus value for inter-module docking is calculated based on docking accuracy, gap difference, and connection strength, using the following formula:

[0031]

[0032] in, This refers to the deviation between the actual docking position and the standard docking position between modules. It is a very small constant. For the inter-module mating gap / surface difference, The actual time required for docking and assembly between modules. For standard docking and assembly time between modules, To account for inter-module connection errors, , , , These are the weighting coefficients.

[0033] Preferably, the update optimization specifically involves: calling the latest parameters of the currently used first and second models from the model server as the update benchmark; calling the digital twin, importing the filtered sample data, updating the virtual assembly scenario, and performing virtual simulation verification of the model update; the update goal of the first model is set to improve internal assembly accuracy, reduce errors, and improve efficiency; the update goal of the second model is set to improve docking accuracy, reduce gap / surface difference, and enhance connection stability; performing parallel update training of the two models, and verifying the fit of the assembly and docking paths through twin simulation after each iteration; repeating the iteration until the model converges.

[0034] Compared with the prior art, the technical solution of this application has the following technical effects:

[0035] This invention constructs an integrated digital twin model of module twins, workstation twins, and environment twins, specifically adapting to the dual scenarios of internal component assembly and inter-module docking. It achieves accurate virtual simulation, interference prediction, and real-time synchronization of the assembly process, which can avoid assembly interference and deviation accumulation in advance and effectively guide actual assembly operations.

[0036] This invention employs an improved DoubleDQN reinforcement learning algorithm that incorporates experience replay, a target network, and a deep neural network to fit the Q-value function. It constructs dedicated control models adapted to two scenarios, optimizes algorithm variables and reward functions for the two scenarios, and balances the high precision of internal module assembly with the stability of inter-module docking and gap / surface difference control requirements. At the same time, through the dual-network design of DoubleDQN, it effectively alleviates the overestimation problem of the traditional DQN algorithm, further improving the algorithm's convergence speed and control accuracy.

[0037] This invention integrates multi-source sensor data such as part pose, equipment parameters, and ambient temperature and humidity during the assembly process to construct a dedicated assembly process database adapted to dual scenarios. It also designs an online model iteration mechanism that can dynamically optimize model parameters based on changes in actual assembly data, thereby improving the practicality and industrial adaptability of the technology.

[0038] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0039] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0041] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0042] Figure 1 A flowchart of the intelligent control method applicable to modular assembly of aluminum alloy car body frames;

[0043] Figure 2 Architecture diagram for an intelligent control method applicable to modular assembly of aluminum alloy car body frames;

[0044] Figure 3 The architecture diagrams for the first and second models are shown below.

[0045] Figure 4 This is a structural diagram of the improved Double-DQN deep reinforcement learning algorithm;

[0046] Figure 5 This is a data diagram illustrating the training process of the first and second models in the embodiments of this application;

[0047] Figure 6 This is a comparison diagram of the internal configuration and control data of each method in the embodiments of this application;

[0048] Figure 7 This is a comparison diagram of the inter-module interface control data of the various methods in the embodiments of this application;

[0049] Figure 8 This is a comparison diagram of the digital twin stitching effects of various methods in the embodiments of this application;

[0050] Figure 9 This is a comparison chart of the overall performance of the methods in the embodiments of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0052] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0053] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0054] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0055] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0056] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0057] Example 1 mainly describes an intelligent control method applicable to the modular assembly of aluminum alloy car body frames, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0058] The aluminum alloy body frame is divided into front cabin module, floor module, side panel module, roof module, and rear cabin module according to its functional structure, and assembly stations are assigned according to the modules.

[0059] Sensor networks are deployed at each assembly station to collect external pose data, internal pose data, station parameter data, and environmental data of each module, and to construct a digital twin. The stations include assembly stations and docking stations.

[0060] Based on internal pose data, assembly station parameter data, and assembly environment data, an internal assembly process database is constructed. Combined with a digital twin, the first model is trained to dynamically generate the optimal assembly path for parts and assemble the parts within each module.

[0061] After the internal assembly of each module is completed, an external docking process database is constructed based on the external part pose data, docking station parameter data, and docking environment data. Combined with the digital twin, a second model is trained to dynamically generate the optimal docking and assembly path between modules and to dock and assemble each module. Both the first and second models are constructed using the improved Double-DQN deep reinforcement learning algorithm.

[0062] After the assembly and docking are completed, the process data will be uploaded, and the first and second models will be updated and optimized regularly based on the newly uploaded data.

[0063] Furthermore, the sensor network specifically involves: deploying LiDAR and binocular structured light cameras above the workstation and next to the mounting fixtures inside the module, respectively, to collect the external pose of the module and the pose of the internal parts; reading the parameter data of the workstation equipment to obtain the workstation parameter data; and deploying temperature and humidity sensors and vibration sensors around the workstation to collect environmental data.

[0064] Furthermore, the module's internal pose data includes the three-dimensional coordinates and attitude angles (pitch, yaw, roll) of each component within the module; the module's external pose data includes the overall three-dimensional coordinates, attitude angles, and interface feature point coordinates of each complete module; and the environmental data includes workstation temperature, humidity, and equipment vibration amplitude.

[0065] Furthermore, filtering, denoising, and normalization processes are employed to eliminate outlier data, ensuring data accuracy and providing reliable input for subsequent digital twin construction and model training.

[0066] Furthermore, a digital twin is constructed, specifically: based on the collected data, an overall framework for a digital twin model of the modular assembly of the aluminum alloy car body frame is built, including twins of each module, assembly station twins of each module, and assembly environment twins. Among them, the twins of each module record the external posture of each module, the geometric features of internal parts, material properties, and assembly constraints, and associate the internal posture data of the module to achieve real-time mapping between the external posture and internal state of the module; the assembly station twin corresponds to each dedicated assembly station, and the station layout and equipment parameters are recorded; the environment twin records environmental parameters.

[0067] Furthermore, the real-time collected actual data (position, workstation, and environmental data) is compared one by one with the virtual data output by the digital twin model. The deviation between the two is calculated, and the model parameters (geometric parameters, material parameters, assembly constraint parameters, etc.) are corrected until the deviation between the virtual data and the actual data is within the preset accuracy range, ensuring that the virtual model and the actual assembly process are accurately synchronized.

[0068] Furthermore, the optimal assembly path for the parts is dynamically generated. Specifically, the real-time collected pose data inside the module, assembly station parameter data, and assembly environment data are synchronously input into the pre-trained first model. The first model combines the real-time data, calls relevant parameters from the assembly process database inside the module, and dynamically generates the optimal assembly path for the parts through the action selection strategy of the improved Double-DQN deep reinforcement learning algorithm. The generated optimal assembly path is imported into the digital twin for virtual simulation verification to confirm that the path is free from interference and that the pose deviation meets the requirements. If there are any problems, the path parameters are adjusted until the assembly requirements are met.

[0069] The optimal assembly path that has passed verification is sent to the control terminal of the corresponding assembly station. Based on the sent assembly path, combined with the real-time collected internal pose data and assembly station parameter data, the entire assembly process, including picking up, docking, and connecting the parts inside the module, is completed. During the assembly process, the first model receives data from the sensors in real time and compares the actual assembly data with the qualified standards in the database. If pose deviation or abnormal process parameters occur, the assembly path and process parameters are adjusted to ensure assembly accuracy. After all parts inside a single module are assembled, the pose data and assembly process parameter data after assembly are collected and compared with the qualified standards to confirm that the assembly is qualified.

[0070] Furthermore, such as Figure 3(a) shows the architecture diagram of the first model. Specifically, the first model is designed for high-precision, multi-source data collaboration, and virtual simulation training requirements of the internal assembly of the aluminum alloy body frame module. Based on the improved Double-DQN deep reinforcement learning algorithm, targeted pre-training is performed. The improved Double-DQN deep reinforcement learning model is input with the internal pose data of the module, the assembly station parameter data, and the assembly environment data. It outputs the optimal assembly path of the parts, including the part grabbing path, docking path, connection path, and the coordinates, motion parameters, and time nodes of each node of the path.

[0071] Furthermore, targeted pre-training is performed based on the improved Double-DQN deep reinforcement learning algorithm. Specifically, the pre-training of the improved Double-DQN deep reinforcement learning algorithm aims to achieve the highest assembly accuracy, optimal assembly efficiency, and minimum error. This is combined with digital twin simulation of actual operation to conduct reinforcement learning training. The Q-value update formula during training is as follows:

[0072]

[0073] in, Current action parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Let be the state transition probability. To control the impact of future gear-based action rewards on the current... Discount factor affecting the value The next gear state determined by the action selection network. The optimal equipment setup action, The action evaluation network evaluates the value of this optimal geared action. The instantaneous reward value for the internal assembly of the module is calculated based on the assembly accuracy, efficiency, and error, using the following formula:

[0074]

[0075] in, The deviation between the actual assembly position and the standard position of the internal parts of the module. It is a very small constant. Standard assembly time for internal components of the module. The actual assembly time of the internal parts of the module. To account for assembly errors of internal components of the module , , These are the weighting coefficients for accuracy, efficiency, and error, respectively.

[0076] Furthermore, the generated assembly path is imported into the digital twin to verify its compatibility with the virtual scene. The loss formula is as follows:

[0077]

[0078] in, To account for the discrepancy between the actual coordinates of the fitting path and the virtual coordinates of the digital twin, , , These are the three-dimensional coordinates of each node in the assembly path generated from the first model. , , These are the three-dimensional coordinates of the corresponding assembly nodes in the digital twin;

[0079] when If the accuracy is less than the preset accuracy threshold (usually 0.05~0.1mm), the path is acceptable; otherwise, adjust the path parameters.

[0080] Furthermore, after all internal assembly of the modules is completed, three types of core data (after preprocessing) are extracted and used as the basic input for constructing the external docking process database and training the second model, ensuring that the data accuracy meets the docking and assembly requirements between modules:

[0081] External pose data of the modules: The overall three-dimensional coordinates, attitude angles and interface feature point coordinates of each complete module (front compartment, floor, side panel, roof and rear compartment modules) after Gaussian filtering preprocessing. The pose data of the module docking surface and positioning hole are collected in particular, and abnormal data is removed.

[0082] Docking station parameter data: Read parameters such as tightening torque, pressing force, docking stroke, and adhesive application amount from the docking station robot, docking fixture, and SPR / FDS connection equipment between modules, and supplement them with redundant sensor data acquisition to ensure the real-time performance and accuracy of docking process parameters;

[0083] Inter-module environmental data: Temperature, humidity, and equipment vibration amplitude data of the inter-module docking station after being collected and preprocessed by sensors.

[0084] Furthermore, the optimal docking and assembly path between modules is dynamically generated. Specifically, each internally assembled and qualified module is accurately transferred to the docking station, the twin in the completed digital twin model is called, the three types of core data in the external docking process database are imported into the twin, the status of the module twin is updated, and the data is accurately mapped to the virtual docking scenario between modules to simulate the real docking and assembly scenario between modules.

[0085] The external pose data of each module, docking station parameter data, and docking environment data collected in real time are used to output the optimal docking and assembly path between modules through a pre-trained second model. The generated optimal docking and assembly path is then imported into a digital twin for virtual simulation verification. The entire docking process between modules is simulated to confirm that there is no interference in the path, the module interface is accurately aligned, and the gap and surface difference meet the requirements. If there are any problems, the path parameters are adjusted until the docking and assembly requirements between modules are met.

[0086] The optimal docking and assembly path between the verified modules is sent to the control terminal of the inter-module docking station. Combined with the real-time collected external pose data of the modules and the parameter data of the docking station, the entire process of precise positioning, interface docking, and connection of each module is completed. During the docking and assembly process, the second model receives data from the sensors in real time and compares the actual docking data with the qualified standards in the external docking process database. If there is a deviation in module pose, out-of-tolerance docking gap, or abnormal process parameters, the docking and assembly path and process parameters are adjusted. After all modules are docked and assembled, the overall pose data, interface gap / surface difference data, and process parameter data are collected and compared with the qualified docking standards to confirm that the docking and assembly is qualified.

[0087] Furthermore, such as Figure 3 (b) shows the architecture diagram of the second model. Specifically, the second model adopts the same improved Double-DQN deep reinforcement learning algorithm framework as the first model. For the docking and assembly scenario between modules, the variable definition and parameter values ​​are optimized. The input is the external pose data of the module, the docking station parameter data, and the docking environment data. The output is the optimal docking and assembly path between modules, including the transfer path, docking positioning path, interface connection path of each module, as well as the coordinates, docking order, action parameters and time nodes of each node in the path.

[0088] Furthermore, the pre-trained second model, specifically, aims to achieve the highest docking accuracy, most stable connection strength, smallest gap difference, and optimal docking efficiency. It is trained using reinforcement learning, combining an external docking process database and a digital twin. Its update formula is as follows:

[0089]

[0090] in, Current state parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Status during assembly The state transition probability, This is a discount factor, with the same value range as the first model. The next gear state determined by the action selection network. The optimal docking action is as follows. The motion evaluation network assesses the value of this optimal docking motion. The instant bonus value for inter-module docking is calculated based on docking accuracy, gap difference, and connection strength, using the following formula:

[0091]

[0092] in, This refers to the deviation between the actual docking position and the standard docking position between modules. It is a very small constant. For the inter-module mating gap / surface difference, The actual time required for docking and assembly between modules. For standard docking and assembly time between modules, To account for inter-module connection errors, , , , These are the weighting coefficients.

[0093] Furthermore, the loss function for the docking process is:

[0094]

[0095] in, For losses during the docking process, , , These are the three-dimensional coordinates of each node in the inter-module docking and assembly path generated from the second model. , , These are the three-dimensional coordinates of the docking nodes between corresponding modules in the digital twin.

[0096] Furthermore, such as Figure 4 The diagram shows the architecture of the improved DoubleDQN reinforcement learning algorithm. Specifically, the improved DoubleDQN algorithm solves the problem of overestimation of action value in the traditional DQN algorithm through a dual-network collaborative architecture, a targeted reward mechanism, and digital twin closed-loop training. This enables dynamic optimization and intelligent control of the assembly path. Its core algorithm employs an Online Network (ONN) with [parameters not specified]. ) and Action Evaluation Network (TargetNetwork, parameter The algorithm employs a dual-network parallel hierarchical structure, with both networks sharing input and feature extraction links, while decoupling their functions to ensure training stability. The input layer receives multi-source state data from the assembly scenario, including module pose data, workstation parameter data, and environmental data. The feature extraction layer fuses and reduces the dimensionality of the multi-source data, and then extracts deep state features through fully connected layers and hidden layers (containing nonlinear activation functions), which are then input into two parallel networks. The action selection network is responsible for calculating the Q-values ​​of all candidate assembly actions in the current state and selecting the optimal assembly action through a max strategy, providing a basis for real-time assembly path decision-making. The action evaluation network independently calculates the target Q-value and is only used to evaluate the value of the optimal action output by the action selection network, fundamentally avoiding overestimation of action value caused by a single network and improving the accuracy of Q-value updates. The algorithm follows the core logic of DoubleDQN for Q-value updates, with the action selection network determining the optimal action for the next state and the action evaluation network completing the value evaluation.

[0097] To address the differentiated needs of assembly scenarios, the algorithm is designed with a scenario-based real-time reward function: for assembly scenarios within modules, the goal is to achieve the highest accuracy, optimal efficiency, and minimum error; for assembly scenarios between modules, the goal is to achieve the optimal docking accuracy, minimum gap difference, and stable connection. By balancing core indicators such as accuracy, efficiency, and error through multi-dimensional weight coefficients, the algorithm is guided to iteratively optimize towards the preset assembly goals. Simultaneously, the algorithm combines a digital twin to construct a virtual simulation training environment, importing the assembly path output by the model into the twin for interference verification and accuracy validation. Closed-loop training is formed through assembly data fed back from sensors in real time, and the model parameters are updated periodically based on newly collected assembly data. This enables the algorithm to continuously adapt and improve its performance in actual assembly scenarios, ultimately outputting the optimal assembly path including coordinates, motion parameters, and time nodes, supporting the intelligent management and control of the entire process of modular assembly of aluminum alloy body frames.

[0098] Further updates and optimizations are implemented as follows: the latest parameters of the first and second models currently in use are retrieved from the model server as the update benchmark; the digital twin is invoked, the selected sample data is imported, the virtual assembly scenario is updated, and virtual simulation verification of the model update is performed; the update goals of the first model are set to improve internal assembly accuracy, reduce errors, and improve efficiency; the update goals of the second model are set to improve docking accuracy, reduce gap / surface difference, and enhance connection stability; parallel update training of the two models is performed, and after a certain number of iterations, the suitability of the assembly and docking paths is compared through twin simulation verification; the iteration is repeated until the model converges.

[0099] This embodiment details an intelligent control method applicable to modular assembly of aluminum alloy car body frames. The method involves dividing the car body frame into multiple functional modules and assigning corresponding assembly stations. Multiple types of sensor networks are used to collect internal and external poses, station parameters, and environmental data. After preprocessing, a digital twin covering the modules, stations, and environment is constructed to achieve precise mapping between virtual and real data. A process database is established based on multi-source data. An improved Double-DQN deep reinforcement learning algorithm is used to train the first and second models. The first model dynamically generates the optimal assembly path for parts based on internal data and completes the assembly of parts within the module. The second model generates the optimal docking path between modules based on external pose and other data, achieving overall docking assembly. After assembly docking is completed, the entire process data is uploaded, and the two models are periodically updated in parallel using new data as samples.

[0100] Example 2, based on Example 1, describes in detail the intelligent control process of modular assembly of aluminum alloy body frames using the method of the present invention, taking a passenger vehicle aluminum alloy body frame modular assembly production line as an application scenario. The details are as follows:

[0101] The aluminum alloy body frame used is made of 6061 aluminum alloy. The aluminum alloy body frame is functionally divided into a front cabin module, floor module, side panel module, roof module, and rear cabin module, with 5 dedicated assembly stations and 1 inter-module docking station assigned. A sensor network is deployed at each station, specifically as follows: LiDAR and binocular structured light cameras are deployed above each assembly station, docking station, and next to the assembly fixtures inside the modules, to collect the external pose of the modules and the pose of internal parts; parameter data from each station's equipment is read via an industrial bus; temperature and humidity sensors and vibration sensors are deployed around each station to collect environmental data.

[0102] Through the aforementioned sensor network, real-time data collection is achieved of internal pose data, external pose data, workstation parameter data, and environmental data for each module. The internal pose data of each module includes the three-dimensional coordinates (X: 0-2000mm, Y: 0-1500mm, Z: 0-1000mm) and attitude angles (pitch angle: -5°~5°, yaw angle: -5°~5°, roll angle: -5°~5°).

[0103] The module's external pose data includes the overall three-dimensional coordinates, attitude angles, and interface feature point coordinates of each complete module. The focus is on collecting pose data of the module's docking surface and positioning holes, as well as the interface feature point coordinate errors.

[0104] Environmental data include workstation temperature (18-32℃), humidity (30%-70%RH), and equipment vibration amplitude (0.01-0.05mm).

[0105] Workstation parameter data includes assembly robot movement speed (50-150mm / s), fixture clamping force (500-800N), SPR tightening torque, and FDS adhesive application amount;

[0106] The collected data were processed using Gaussian filtering, noise reduction, and normalization to remove outlier data.

[0107] Based on the preprocessed collected data, a digital twin model framework for the modular assembly of the aluminum alloy car body frame was constructed using Unity3D software. This framework includes twins for each module, assembly station twins for each module, and an assembly environment twin. The module twins record the external pose, geometric features of internal parts, material properties, and assembly constraints of each module, and associate them with the module's internal pose data to achieve real-time mapping between the module's external pose and internal state. The assembly station twins correspond to six dedicated assembly stations, recording the station layout (assembly station dimensions 3m×4m, docking station dimensions 5m×6m) and equipment parameters, enabling real-time linkage with the actual station equipment. The environment twins record environmental parameters (temperature, humidity, vibration amplitude) to achieve real-time mapping and monitoring of the environmental state.

[0108] The actual data collected in real time is compared one by one with the virtual data output by the digital twin model. The deviation between the two is calculated, and the model parameters (geometric parameters, material parameters, assembly constraint parameters, etc.) are corrected until the deviation between the virtual data and the actual data is within the preset accuracy range (≤0.05mm). This ensures that the virtual model and the actual assembly process are accurately synchronized. After calibration, the simulation accuracy of the twin reaches 99.5%.

[0109] Based on the preprocessed internal pose data, assembly station parameter data, and assembly environment data, an internal assembly process database was constructed, containing part assembly standard parameters, qualification thresholds, and historical best assembly data, with a total of over 10,000 data entries. Based on the external pose data, docking station parameter data, and docking environment data, an external docking process database was constructed, containing docking standards, clearance and surface difference thresholds, and connection strength requirements for 12 docking interfaces, with a total of over 8,000 data entries.

[0110] Both the first and second models were constructed using the improved Double-DQN deep reinforcement learning algorithm and trained using the TensorFlow framework. Specific training details are as follows:

[0111] The training objective of the first model is to achieve the highest assembly accuracy, optimal assembly efficiency, and minimum error; the training duration is 10,000 epochs, the batch size is 64, the learning rate is 0.001, the discount factor γ = 0.95, the minimum constant δ = 1e-6, and the weight coefficients are... Standard assembly time for internal components of the module The preset accuracy threshold is 0.08mm. When Δ When the thickness is <0.08mm, the path is acceptable; according to Figure 5 As shown in (a) of the data graph of the first model pre-training process, the first model converged after 8521 training rounds, and the prediction accuracy of the model after convergence reached 98.8%.

[0112] The training objective of the second model is to achieve the highest docking accuracy, the most stable connection strength, the smallest gap surface difference, and the optimal docking efficiency. It undergoes 10,000 training epochs with a batch size of 64, a learning rate of 0.001, a discount factor γ = 0.95 (consistent with the first model), a minimum constant δ = 1e-6, and weight coefficients... Standard assembly and docking time between modules The preset accuracy threshold is set to 0.08mm, based on... Figure 5 As shown in the training process data graph of the second model in (b), the model converged after 8245 training rounds, and the prediction accuracy of the model reached 98.5% after convergence.

[0113] Real-time collected pose data within the module, assembly station parameter data, and assembly environment data are synchronously input into the pre-trained first model. The first model, combining the real-time data, calls relevant parameters from the module's internal assembly process database and dynamically generates the optimal assembly path for the part using an improved Double-DQN algorithm action selection strategy. This path includes part grabbing paths, docking paths, and connection paths. The generated optimal assembly path is then imported into the digital twin for virtual simulation verification to confirm that the path is interference-free and that pose deviations meet requirements. The optimal assembly path that has passed verification is sent to the control terminal of the corresponding assembly station. Based on the sent assembly path, combined with the real-time collected internal pose data and assembly station parameter data, the entire assembly process, including picking up, docking, and connecting parts inside the module, is completed, resulting in the following table 1, which shows the statistical table of core assembly data for each module:

[0114] Table 1. Statistics of Core Components for Each Module

[0115] Module type Number of internal parts (pieces) Average assembly time Average deviation of assembly Assembly pass rate Front cabin module 18 22.5min 0.057mm 99.68% Floor modules 22 24.3min 0.063mm 99.57% Side panel module 20 24.8min 0.076mm 99.56% Roof Module 12 22.1min 0.052mm 99.89% Rear cabin module 14 23.3min 0.068mm 99.67% average value 17.2 23.8min 0.063mm 99.87%

[0116] As shown in Table 1, the assembly data of each module are better than the preset standards (assembly deviation ≤ 0.08 mm, pass rate ≥ 99.5%), and the overall assembly stability is strong, which verifies the accuracy and reliability of the first model.

[0117] After the internal assembly of each module is completed, the qualified modules are precisely transported to the docking station. Real-time collected external pose data of each module, docking station parameter data, and docking environment data are used to output the optimal docking and assembly path between modules through a pre-trained second model. This path includes the transport path, docking positioning path, interface connection path of each module, as well as the coordinates, docking sequence, action parameters, and time points of each node. The generated optimal docking and assembly path is imported into a digital twin for virtual simulation verification, simulating the entire docking process between modules to confirm that there is no interference in the path, the module interfaces are accurately aligned, and the gap surface difference meets the requirements. The verified path is then sent to the docking station control terminal. Combined with the real-time collected data, the entire process of precise positioning, interface docking, and connection of each module is completed, resulting in the core docking data statistics table for each module shown in Table 2 below.

[0118] Table 2. Statistics of core data on inter-module connections

[0119] Interface docking module combination Average docking time Average gap between docks Average surface difference of docking docking pass rate Path simulation verification pass rate Interface 1 Front Cabin Module - Floor Module (Front End) 6.2min 0.072mm 0.063mm 99.55% 98.88% Interface 2 Front cabin module - floor module (both sides) 5.8min 0.075mm 0.068mm 99.82% 98.67% Interface 3 Floor module - side panel module (front left) 5.9min 0.078mm 0.065mm 99.73% 98.59% Interface 4 Floor module - side panel module (rear left) 5.7min 0.071mm 0.069mm 99.84% 99.60% Interface 5 Floor module - side panel module (front right) 5.8min 0.067mm 0.061mm 99.91% 98.86% Interface 6 Floor module - side panel module (rear right) 5.6min 0.074mm 0.066mm 99.53% 98.78% Interface 7 Side panel module (left) - Roof module (left side) 5.5min 0.069mm 0.062mm 99.54% 99.71% Interface 8 Side panel module (right) - Roof module (right side) 5.4min 0.065mm 0.058mm 99.75% 99.92% Interface 9 Floor module - Rear cabin module (front end) 4.9min 0.077mm 0.071mm 99.61% 98.56% Interface 10 Floor module - Rear cabin module (both sides) 5.0min 0.073mm 0.067mm 99.52% 98.57% Interface 11 Side panel module (left) - Rear cabin module (left side) 5.2min 0.076mm 0.064mm 99.62% 98.68% Interface 12 Side panel module (right) - Rear compartment module (right side) 5.1min 0.068mm 0.060mm 99.79% 98.59% Overall integration Full module integration 28.5min 0.072mm 0.064mm 99.61% 98.88%

[0120] As shown in Table 2, the docking data between each module is better than the preset standards (dock gap ≤ 0.08 mm, surface difference ≤ 0.08 mm, pass rate ≥ 99.5%). The docking accuracy and stability meet the modular assembly requirements of the aluminum alloy body frame, verifying the effectiveness of the second model and digital twin simulation.

[0121] After the assembly of each vehicle body frame is completed, all data during the assembly process (position data, workstation parameters, environmental data, assembly errors, gap differences, etc.) are uploaded to the model server. Every 7 days is an update cycle, and the first and second models are regularly updated and optimized based on the newly uploaded data.

[0122] This method produced a total of 300 aluminum alloy car body frames in 30 days. Testing showed that the average assembly cycle per unit was 76.2 minutes, with a 99.7% qualification rate for internal module assembly, a 99.3% qualification rate for inter-module docking, and an overall assembly qualification rate of 99.0%. The manual intervention rate during assembly was only 1.2%, and the equipment failure rate was reduced to 0.8%. After four model updates, the assembly cycle per unit was shortened to 73.5 minutes, the average assembly deviation was reduced to 0.054 mm, and the average docking gap was reduced to 0.068 mm, fully meeting the high-precision, high-efficiency, and low-intervention requirements for modular assembly of aluminum alloy car body frames for new energy vehicles.

[0123] Existing methods such as TRL (Reinforcement Learning Assembly Control), BDQN (DQN Assembly Control), DT (Digital Twin Simulation Control), and DQN-BC (DQN-based Blockchain Assembly Control) were selected to generate control schemes based on datasets. Comparative experiments were conducted in a digital twin, and the experimental data of each method were compared. The results are shown in Table 3 below.

[0124] Table 3 Comparative Experimental Data

[0125] Control methods This method TRL BDQN DT DQN-BC Average assembly cycle per unit 76.2min 95.8min 88.6min 89.3min 91.2min Average deviation of internal assembly of module 0.063mm 0.147mm 0.104mm 0.086mm 0.168mm Average gap between modules 0.072mm 0.153mm 0.129mm 0.95mm 0.144mm Overall assembly pass rate 99.75% 92.43% 98.71% 99.24% 93.35% Human intervention rate 1.2% 8.5% 4.8% 3.4% 7.1% Assembly cycle reduction rate after model update 3.5% No (no model update function) 1.8% No (no model update function) 1.5%

[0126] According to Table 1-3 and Figure 6 The comparison chart of module internal assembly control data for each method shows that the average deviation of module internal assembly in this method is 0.063mm, which is less than the model's preset accuracy threshold of 0.08mm, and is significantly better than other methods.

[0127] According to Table 2-3 and Figure 7 The comparison chart of inter-module interface control data for each method shows that the average gap of this method is 0.072mm, which is also less than the accuracy threshold of 0.08mm, and is superior to other methods.

[0128] During the docking experiment of various methods in the digital twin, the docking status at interface 7 was tracked and recorded at 4 typical time points, resulting in... Figure 8 The comparison charts of digital twin stitching effects of various methods are shown below, based on Table 2-3 and... Figure 8 It can be seen that the control scheme generated by this method has a better docking effect in the digital twin and its accuracy is significantly better than other methods;

[0129] According to Table 1-3 and Figure 9 As shown in the comprehensive performance comparison chart, this method has a higher assembly qualification rate, a lower rate of manual intervention, and stronger adaptability and stability, resulting in a significant improvement in assembly efficiency.

[0130] This embodiment describes in detail the intelligent control process of modular assembly of aluminum alloy body frames using the method of the present invention, taking a passenger vehicle aluminum alloy body frame modular assembly production line as an application scenario. During the process, existing commonly used methods are used to generate control schemes, which are verified in a digital twin. The results show that the technical solution of the present invention has significant improvements in assembly efficiency, assembly accuracy, pass rate, and intelligence level compared with the prior art. It can effectively meet the high-precision, high-efficiency, and iterative control requirements of modular assembly of aluminum alloy body frames and has extremely high industrial application value.

[0131] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. An intelligent control method applicable to modular assembly of aluminum alloy car body frames, characterized in that, include: The aluminum alloy body frame is divided into front cabin module, floor module, side panel module, roof module, and rear cabin module according to its functional structure, and assembly stations are assigned according to the modules. Sensor networks are deployed at each assembly station to collect external pose data, internal pose data, station parameter data, and environmental data of each module, and to construct a digital twin. The stations include assembly stations and docking stations. Based on internal pose data, assembly station parameter data, and assembly environment data, an internal assembly process database is constructed. Combined with a digital twin, the first model is trained to dynamically generate the optimal assembly path for parts and assemble the parts within each module. After the internal assembly of each module is completed, an external docking process database is constructed based on the external part pose data, docking station parameter data, and docking environment data. Combined with the digital twin, a second model is trained to dynamically generate the optimal docking and assembly path between modules and to dock and assemble each module. Both the first and second models are constructed using the improved Double-DQN deep reinforcement learning algorithm. After the assembly and docking are completed, the process data will be uploaded, and the first and second models will be updated and optimized regularly based on the newly uploaded data.

2. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 1, characterized in that, The sensor network specifically comprises: deploying LiDAR and binocular structured light cameras above the workstation and next to the mounting fixture inside the module, respectively, to collect the external pose of the module and the pose of the internal parts; reading the parameter data of the workstation equipment to obtain the workstation parameter data; and deploying temperature and humidity sensors and vibration sensors around the workstation to collect environmental data.

3. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 1, characterized in that, The construction of the digital twin specifically involves: based on the collected data, building an overall framework for a digital twin model of the modular assembly of the aluminum alloy car body frame, including twins for each module, assembly station twins for each module, and assembly environment twins; wherein, the twins of each module record the external posture of each module, the geometric features of internal parts, material properties, and assembly constraints, and associate the internal posture data of the module to achieve real-time mapping between the external posture and internal state of the module; the assembly station twins correspond to each dedicated assembly station, recording the station layout and equipment parameters; and the environment twins record environmental parameters.

4. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 1, characterized in that, The dynamic generation of the optimal assembly path for the part specifically involves: synchronously inputting real-time collected pose data within the module, assembly station parameter data, and assembly environment data into the pre-trained first model; combining the real-time data with relevant parameters from the module's internal assembly process database, and dynamically generating the optimal assembly path for the part using the action selection strategy of the improved Double-DQN deep reinforcement learning algorithm; importing the generated optimal assembly path into a digital twin for virtual simulation verification to confirm that the path is free from interference and that pose deviations meet requirements; if problems exist, adjusting the path parameters until the assembly requirements are met. The optimal assembly path that has passed verification is sent to the control terminal of the corresponding assembly station. Based on the sent assembly path, combined with the real-time collected internal pose data and assembly station parameter data, the entire assembly process, including picking up, docking, and connecting the parts inside the module, is completed. During the assembly process, the first model receives data from the sensors in real time and compares the actual assembly data with the qualified standards in the database. If pose deviation or abnormal process parameters occur, the assembly path and process parameters are adjusted to ensure assembly accuracy. After all parts inside a single module are assembled, the pose data and assembly process parameter data after assembly are collected and compared with the qualified standards to confirm that the assembly is qualified.

5. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 4, characterized in that, The first model is specifically designed for high-precision, multi-source data collaboration, and virtual simulation training requirements of the internal assembly of the aluminum alloy body frame module. It is pre-trained based on the improved Double-DQN deep reinforcement learning algorithm. The pre-trained improved Double-DQN deep reinforcement learning model is input with the module's internal pose data, assembly station parameter data, and assembly environment data. It outputs the optimal assembly path for the parts, including the part grabbing path, docking path, connection path, and the coordinates, motion parameters, and time nodes of each node on the path.

6. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 5, characterized in that, The targeted pre-training based on the improved Double-DQN deep reinforcement learning algorithm specifically involves: The pre-training of the improved Double-DQN deep reinforcement learning algorithm aims to achieve the highest assembly accuracy, optimal assembly efficiency, and minimum error. This is combined with digital twin simulation of actual operational effects for reinforcement learning training. The Q-value update formula during training is as follows: in, Current action parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Let be the state transition probability. To control the impact of future gear-based action rewards on the current... Discount factor affecting the value The next gear state determined by the action selection network The optimal equipment setup action, The action evaluation network is used to assess the value of this optimal gear-fitting action. The instantaneous reward value for the module's internal assembly is calculated based on assembly accuracy, efficiency, and error, using the following formula: in, The deviation between the actual assembly position and the standard position of the internal parts of the module. It is a very small constant. Standard assembly time for internal components of the module. The actual assembly time of the internal parts of the module. To account for assembly errors of internal components of the module , , These are the weighting coefficients for accuracy, efficiency, and error, respectively.

7. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 1, characterized in that, The optimal docking and assembly path between the dynamically generated modules is specifically generated as follows: each internally assembled and qualified module is accurately transferred to the docking station, the twin in the completed digital twin model is called, the three types of core data in the external docking process database are imported into the twin, the status of the module twin is updated, and the data is accurately mapped to the virtual docking scene between modules to simulate the real docking and assembly scene between modules. The external pose data of each module, docking station parameter data, and docking environment data collected in real time are used to output the optimal docking and assembly path between modules through a pre-trained second model. The generated optimal docking and assembly path is then imported into a digital twin for virtual simulation verification. The entire docking process between modules is simulated to confirm that there is no interference in the path, the module interface is accurately aligned, and the gap and surface difference meet the requirements. If there are any problems, the path parameters are adjusted until the docking and assembly requirements between modules are met. The optimal docking and assembly path between the verified modules is sent to the control terminal of the inter-module docking station. Combined with the real-time collected external pose data of the modules and the parameter data of the docking station, the entire process of precise positioning, interface docking, and connection of each module is completed. During the docking and assembly process, the second model receives data from the sensors in real time and compares the actual docking data with the qualified standards in the external docking process database. If there is a deviation in module pose, out-of-tolerance docking gap, or abnormal process parameters, the docking and assembly path and process parameters are adjusted. After all modules are docked and assembled, the overall pose data, interface gap / surface difference data, and process parameter data are collected and compared with the qualified docking standards to confirm that the docking and assembly is qualified.

8. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 7, characterized in that, The second model is specifically as follows: The second model adopts the same improved Double-DQN deep reinforcement learning algorithm framework as the first model. For the docking and assembly scenario between modules, the variable definition and parameter value are optimized. The input is the external pose data of the module, the docking station parameter data, and the docking environment data. The output is the optimal docking and assembly path between modules, including the transfer path, docking positioning path, interface connection path of each module, as well as the coordinates, docking order, action parameters and time nodes of each node in the path.

9. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 8, characterized in that, The second pre-trained model, specifically, aims to achieve the highest docking accuracy, most stable connection strength, smallest gap difference, and optimal docking efficiency. It is trained using reinforcement learning, incorporating an external docking process database and a digital twin. Its update formula is as follows: in, Current state parameters In state Next, perform the assembly action. The value of the action, Choose network parameters for the action. To evaluate the network parameters for action assessment, Status during assembly The state transition probability, This is a discount factor, with the same value range as the first model. The next gear state determined by the action selection network The optimal docking action is as follows. The motion evaluation network assesses the value of this optimal docking motion. The instant bonus value for inter-module docking is calculated based on docking accuracy, gap difference, and connection strength, using the following formula: in, This refers to the deviation between the actual docking position and the standard docking position between modules. It is a very small constant. For the inter-module mating gap / surface difference, The actual time required for docking and assembly between modules. For standard docking and assembly time between modules, To account for inter-module connection errors, , , , These are the weighting coefficients.

10. The intelligent control method for modular assembly of aluminum alloy car body frames according to claim 1, characterized in that, The update and optimization specifically involves: calling the latest parameters of the currently used first and second models from the model server as the update benchmark; calling the digital twin, importing the filtered sample data, updating the virtual assembly scene, and performing virtual simulation verification of the model update; the update goal of the first model is set to improve the internal assembly accuracy, reduce errors, and improve efficiency; the update goal of the second model is set to improve docking accuracy, reduce gap / surface difference, and enhance connection stability. The dual-model parallel update training is performed. After a certain number of iterations, the model is verified by twin simulation to compare the fit of the assembly and docking paths. The iteration is repeated until the model converges.

Images