An assembly workshop production line construction system based on digital twinning

Abstract

The application discloses an assembly workshop production line construction system based on digital twinning, and belongs to the technical field of intelligent manufacturing. The system comprises a multi-source data sensing and fusion module, which is used for collecting production order data, physical environment data, equipment state data and material attribute data, and performing fusion processing on the collected data to generate a physical world state vector; and an assembly strategy deep reinforcement learning optimization module, which is used for generating a candidate assembly strategy, and performing virtual pre-performance on the candidate assembly strategy by using a predictive digital twinning modeling and pre-assembly module to optimize a comprehensive reward function, and finally generating a candidate optimal assembly strategy and corresponding model prediction uncertainty. The application prevents safety accidents caused by heat accumulation out of control, and realizes the unification of rapid flexibility and safety reliability.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to an assembly workshop production line construction system based on digital twins. Background Art

[0002] New energy vehicle power battery packs are complex integrated products with high energy density and high safety standards. Their assembly process, especially key processes such as cell stacking, module fixing, and welding, has extremely high requirements for assembly accuracy and process thermal management. To meet the diverse market demands, modern battery pack production lines are developing towards flexible assembly, which means they can quickly switch between orders for battery packs of different models and specifications.

[0003] However, existing technologies face a difficult technical contradiction in achieving flexible assembly. On the one hand, the high frequency of order switching requires physical equipment such as assembly robots and conveyors to have the ability to respond quickly and adapt adaptively. On the other hand, the thermodynamic properties of different battery packs, coupled with fluctuations in workshop ambient temperature, make the diffusion and cumulative effects of heat generated during the assembly process highly uncertain. This uncertainty poses a serious threat to the ultimate safety and performance consistency of the battery pack.

[0004] Current solutions are often isolated, such as using a posteriori quality testing, static model digital twins, or separate offline simulation and online control. These methods cannot effectively address the fundamental technical difficulty of achieving both fast flexibility and safety and reliability due to the dynamic coupling effect of assembly precision adaptability and real-time thermal diffusion prediction under high-frequency switching and environmental fluctuations.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide an assembly workshop production line construction system based on digital twins to solve the problems raised in the above background technology.

[0007] The technical solution of the present invention comprises: a multi-source data perception and fusion module for collecting production order data, physical environment data, equipment status data and material attribute data, and fusing the collected data to generate a physical world state vector;

[0008] The assembly strategy deep reinforcement learning optimization module is used to generate candidate assembly strategies and virtually preview the candidate assembly strategies using predictive digital twin modeling and pre-assembly modules to optimize the comprehensive reward function, ultimately generating candidate optimal assembly strategies and corresponding model prediction uncertainties.

[0009] The predictive digital twin modeling and pre-assembly module is used to receive the physical world state vector and candidate assembly strategies during the virtual preview process to solve the corresponding predicted temperature field and predicted stress field for use by the assembly strategy deep reinforcement learning optimization module;

[0010] The adaptive safety boundary and active intervention module receives candidate optimal assembly strategies and model prediction uncertainty, obtains environmental disturbance uncertainty and absolute safety redline temperature, makes safety boundary determinations, and generates final safety instructions.

[0011] The physical production line execution and closed-loop feedback module is used to execute the final safety instructions and collect real physical data to update the physical world state vector, thereby forming a data closed loop.

[0012] Preferably, the predictive digital twin modeling and pre-assembly module predicts the evolution of the predicted temperature field and the predicted stress field by solving the temperature change rate and the stress change rate.

[0013] Preferably, the temperature change rate calculation process includes:

[0014] Determine the volumetric heat source generation rate; combine the volumetric heat source generation rate with the material's inherent thermophysical properties derived from material attribute data to calculate the temperature change rate; the volumetric heat source generation rate is used to characterize the heat generated during the assembly process, and its value is determined by the welding heat source and the contact heat transfer effect; the size of the welding heat source is related to the welding current derived from the equipment status data; the size of the contact heat transfer effect is related to the assembly pressure instructions output by the deep reinforcement learning optimization module of the assembly strategy.

[0015] Preferably, the stress change rate calculation process includes:

[0016] Determine the total strain; solve the time rate of change of the total strain to obtain the rate of change of stress based on the material Young's modulus derived from the material property data; the total strain is composed of mechanical strain and thermal strain; the mechanical strain is determined by the assembly pressure command; the thermal strain is determined by the predicted temperature field and the material thermal expansion coefficient derived from the material property data.

[0017] Preferably, the optimization process of the comprehensive reward function includes:

[0018] Determine efficiency rewards, safety penalties, and quality penalties; perform a weighted summation of the efficiency rewards, safety penalties, and quality penalties to generate a comprehensive reward function; the efficiency reward is generated based on the predicted cycle time; the safety penalty is generated when the global maximum temperature of the predicted temperature field exceeds the preset safety soft threshold; the quality penalty is generated based on the predicted final assembly dimension deviation.

[0019] Preferably, the efficiency importance weight, safety importance weight and quality importance weight used for weighted summation are dynamically adjusted by the system based on parsing the order type and priority information in the production order data.

[0020] Preferably, the process of the adaptive safety boundary and active intervention module performing safety boundary determination includes:

[0021] Determine a dynamic safety buffer margin; subtract the dynamic safety buffer margin from the absolute safety red line temperature derived from production order data to generate a dynamic safety margin; and compare the predicted maximum temperature corresponding to the candidate optimal assembly strategy with the dynamic safety margin.

[0022] Preferably, the process of determining the dynamic safety buffer margin includes:

[0023] Determine the model risk term and the environmental risk term; sum the model risk term and the environmental risk term to generate a dynamic safety buffer margin; the model risk term is generated based on the model prediction uncertainty and the preset risk aversion coefficient; the environmental risk term is generated based on the environmental disturbance uncertainty and the preset risk aversion coefficient.

[0024] Preferably, the subsequent processing logic of the comparison includes:

[0025] If the predicted maximum temperature is less than the dynamic safety margin, the candidate optimal assembly strategy is determined to be safe and is output as the final safety instruction;

[0026] If the predicted maximum temperature is greater than or equal to the dynamic safety margin, the candidate optimal assembly strategy is determined to be unsafe. After the candidate optimal assembly strategy is revised, the safety margin is re-determined until the revised strategy is determined to be safe.

[0027] The present invention provides an assembly workshop production line construction system based on digital twins through improvements. Compared with the existing technology, it has the following improvements and advantages:

[0028] 1. This technology enables accurate prediction of the thermal-mechanical coupling effects of the assembly process, improving the fidelity of the digital twin model. The predictive digital twin modeling and pre-assembly module in this invention achieves a breakthrough by constructing specific temperature and stress rate-of-change models. Bidirectional coupling modeling enables the system to preview the actual physical state evolution within the battery pack due to the combined effects of force and heat under specific assembly strategies, achieving prediction accuracy far exceeding existing offline simulations.

[0029] 2. The production line is empowered to adaptively optimize process parameters based on business needs, achieving truly flexible production. This is achieved by introducing a deep reinforcement learning optimization module for assembly strategies and designing a sophisticated integrated reward function. When the system receives an expedited order for a conventional battery pack, it analyzes the production order data and automatically increases the efficiency importance weight, driving the DRL algorithm to find a strategy that minimizes assembly cycle time. Conversely, when producing high-precision sample battery packs, the safety and quality importance weights are automatically increased, allowing the algorithm to prioritize robust strategies that minimize deviations between predicted maximum temperatures and final assembly dimensions. This mechanism extends the flexibility of the production line beyond the adaptability of hardware to include the intelligence and adaptability of the control strategy.

[0030] 3. An active safety defense system based on uncertainty quantification has been established, elevating safety control from post-detection to pre-emptive prevention. The adaptive safety margin and active intervention module have achieved revolutionary progress by defining a dynamic safety buffer margin. The safety margin is no longer a fixed value, but is dynamically adjusted based on the uncertainty of model predictions and environmental disturbances. If the system has a low confidence level in predicting an uncommon assembly strategy, or if sensors detect drastic temperature fluctuations in the workshop environment, the value will automatically increase, tightening the safety margin and forcing the system to make more conservative adjustments to the strategy before execution. This active intervention logic ensures that any instructions issued to the physical production line fully account for real-world uncertainties, preventing safety accidents caused by uncontrolled heat accumulation, and achieving a unified combination of speed, flexibility, safety, and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be further explained below in conjunction with the accompanying drawings and Examples:

[0032] Figure 1 It is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0034] Example 1:

[0035] See also Figure 1 , the present invention provides a technical solution for building an assembly workshop production line system based on digital twins, including: a multi-source data perception and fusion module for collecting production order data, physical environment data, equipment status data and material attribute data, and fusing the collected data to generate a physical world state vector;

[0036] The assembly strategy deep reinforcement learning optimization module is used to generate candidate assembly strategies and virtually preview the candidate assembly strategies using predictive digital twin modeling and pre-assembly modules to optimize the comprehensive reward function, ultimately generating candidate optimal assembly strategies and corresponding model prediction uncertainties.

[0037] The predictive digital twin modeling and pre-assembly module is used to receive the physical world state vector and candidate assembly strategies during the virtual preview process to solve the corresponding predicted temperature field and predicted stress field for use by the assembly strategy deep reinforcement learning optimization module;

[0038] The adaptive safety boundary and active intervention module receives candidate optimal assembly strategies and model prediction uncertainty, obtains environmental disturbance uncertainty and absolute safety redline temperature, makes safety boundary determinations, and generates final safety instructions.

[0039] The physical production line execution and closed-loop feedback module is used to execute the final safety instructions and collect real physical data to update the physical world state vector, thereby forming a data closed loop.

[0040] In this embodiment, the system aims to overcome the technical challenges of balancing speed, flexibility, safety, and reliability in the field of new energy vehicle power battery packaging, caused by frequent order switching and dynamic environmental fluctuations. The system achieves accurate prediction and proactive control of the production process by building a closed-loop system consisting of data perception, strategy optimization, virtual rehearsal, safety intervention, and physical execution.

[0041] The technical core of the system's overall architecture lies in its five logically tightly coupled functional modules. The workflow begins with the multi-source data perception and fusion module, which serves as the system's perception basis and provides a high-fidelity initial state vector of the physical world for all subsequent calculations in the digital space. The assembly strategy deep reinforcement learning optimization module, as the system's decision-making optimization unit, interacts frequently with the predictive digital twin modeling and pre-assembly modules. The former is responsible for generating innovative assembly strategies, while the latter is responsible for quickly and accurately rehearsing the physical consequences that may result from these strategies. The product of this interactive optimization, the candidate optimal assembly strategy, must undergo final review and potential correction by the adaptive safety boundary and active intervention module before being released to the physical world to ensure absolute safety of operation. The safety-audited instructions are accurately executed in the physical world by the physical production line execution and closed-loop feedback module, and new real data is collected in the process, forming a data closed loop that drives continuous iterative optimization of the entire system.

[0042] Example 2

[0043] The predictive digital twin modeling and pre-assembly module predicts the evolution of the predicted temperature field and the predicted stress field by solving the temperature change rate and stress change rate.

[0044] The core function of the predictive digital twin modeling and pre-assembly module in this embodiment is to build a multi-physics model that can perform forward-looking calculations on the thermal-mechanical coupling effects during the battery pack assembly process. This module does not use static mapping, but rather calculates the thermal-mechanical coupling effects at any position of the assembly. Temperature and stress Perform temporal evolution prediction to dynamically reveal the future impact of assembly strategies on product status;

[0045] To achieve this kind of evolution prediction, this embodiment adopts the first-order Euler integration method to convert the current moment The physical world state vector , moving forward a very short time step , to calculate the next moment The predicted state vector ; This process is specifically implemented through the following two core evolution equations:

[0046] ;

[0047] ;

[0048] in, : Next moment The predicted temperature state; : The temperature state of the physical world at the current time t; : an extremely short time step;

[0049] : Next moment The predicted stress state; : The stress state of the physical world at the current time t; : an extremely short time step; : stress change rate;

[0050] The key here is to understand the temperature change rate and stress change rate Accurate modeling and solution of the evolution prediction mechanism; this module can receive candidate assembly strategies and, based on the current physical state, quickly generate a complete set of prediction data on future temperature and stress fields, providing key judgment basis for subsequent strategy evaluation and optimization; this evolution prediction mechanism based on the rate of change gives the digital twin model the ability to make forward-looking predictions, which is the fundamental prerequisite for achieving active control and risk avoidance.

[0051] Example 3

[0052] The temperature change rate calculation process includes:

[0053] Determine the volumetric heat source generation rate; combine the volumetric heat source generation rate with the material's inherent thermophysical properties derived from material attribute data to calculate the temperature change rate. The volumetric heat source generation rate is used to characterize the heat generated during the assembly process, and its value is determined by both the welding heat source and the contact heat transfer effect. The magnitude of the welding heat source is related to the welding current derived from equipment status data; the magnitude of the contact heat transfer effect is related to the assembly pressure instructions output by the assembly strategy deep reinforcement learning optimization module.

[0054] The stress rate of change solution process includes:

[0055] Determine the total strain; solve the time rate of change of the total strain to obtain the rate of change of stress based on the material Young's modulus derived from the material property data; the total strain is composed of mechanical strain and thermal strain; the mechanical strain is determined by the assembly pressure command; the thermal strain is determined by the predicted temperature field and the material thermal expansion coefficient derived from the material property data.

[0056] To achieve the above-mentioned high-fidelity evolution prediction, this embodiment performs customized coupling modeling on the solution process of temperature change rate and stress change rate;

[0057] For temperature change rate , its solution is based on the improvement of the classical Fourier heat transfer law, and its core lies in the accurate description of the heat source term; the mathematical expression is:

[0058] Formula (1): ;

[0059] This formula aims to establish a mathematical model that can accurately reflect the physical process of heat generation and transfer during battery pack assembly. The motivation is that existing models often ignore the significant effect of assembly pressure on the heat conduction efficiency of the contact surface. This model uses a custom heat source term to , the key process parameter of mechanical pressure is internalized into the thermodynamic calculation, thereby achieving deep coupling of thermal-mechanical effects and ensuring that both ends of the formula are the temperature change rate. consistency;

[0060] in, represents the material density, represents the specific heat capacity of the material, Indicates the thermal conductivity of the material. These three are the inherent thermophysical properties of the material. The values ​​are based on the BOM list in the production order data and are obtained from the material attribute data. Direct query to obtain;

[0061] represents the heat conduction term; is defined in this embodiment and is located at position The volumetric heat source generation rate, in units of , used to characterize the total heat generated during the assembly process;

[0062] : Temperature change rate; : Temperature at position i;

[0063] The application of this formula lies in the ability to accurately calculate the rate of change of temperature over time at any point within the battery pack by inputting material properties and a custom heat source term. The technical effect is that the temperature field prediction no longer relies solely on conventional heat conduction and heat radiation, but can accurately reflect the heat generation and exchange process driven by core process parameters such as welding current and assembly pressure, greatly improving the fidelity of temperature prediction.

[0064] Furthermore, the volumetric heat source generation rate Innovatively defined as:

[0065] Formula (2): ;

[0066] This formulation aims to decompose complex assembly heat sources into two main components: active heat sources directly introduced by processes such as welding, and equivalent heat sources generated by mechanical pressure changing the thermal resistance of the contact interface. This is motivated by the desire to clearly separate the different physical mechanisms, electrothermal conversion and pressure-heat conduction coupling, to facilitate modeling and parameter calibration.

[0067] in, Represents the welding heat source, whose size is determined by the device status data The welding current obtained from Strong correlation can be calculated using the well-known Goldak double ellipsoid heat source model;

[0068] :Contact surface heat transfer coefficient, is the assembly pressure function;

[0069] : The temperature of the contact point on another component adjacent to position i;

[0070] : Temperature at position i; : Characteristic heat transfer distance of the control volume;

[0071] Taking the typical 2mm thick aluminum alloy shell laser welding as an example, the heat source parameters of the Goldak model can be set as: the length of the front half ellipsoid , the length of the rear half ellipsoid , heat source width , heat source depth The energy distribution coefficients of the front and rear hemi-ellipsoids are and Total heat input power The welding current Functions such as ,in For welding efficiency, is the welding voltage;

[0072] The second term is the contact heat transfer effect term, which is only effective at the contact interface between different components; It is the key parameter for achieving thermal-mechanical coupling, namely the heat transfer coefficient of the contact surface. It is not a fixed value, but the assembly pressure Function ,This functional relationship is obtained by experimental calibration of a specific material combination;

[0073] For example, for the contact interface between a common aluminum alloy housing and a polymer diaphragm, this functional relationship can be modeled as a power law empirical formula: ;in, and are experimentally determined coefficients, e.g., , , The pressure index is usually between 0.5 and 0.8. In the system's material property database, function coefficients such as C1, C2, and n reference tables obtained through experimental calibration are pre-stored for key material combinations;

[0074] It is the assembly pressure instruction output by the assembly strategy deep reinforcement learning optimization module; Is with location The temperature of the contact point on another adjacent component; is the characteristic heat transfer distance of the control volume;

[0075] This distance is more specifically defined as half the thickness of the component at the contact interface, e.g., for two components with thicknesses of and The plate is in position Contact occurs, and the characteristic distance used to calculate heat transfer can be taken as , which is perpendicular to the contact surface; this represents the average path length for heat to diffuse from the interface into the material;

[0076] During virtual pre-assembly, the system generates candidate assembly pressures based on the output of the DRL module. , query or calculate the corresponding The value is substituted into formulas (2) and (1) to solve the temperature change rate. This application method enables the force in the mechanical domain to directly modulate the heat exchange in the thermodynamic domain. The technical effect is that a direct and quantitative causal relationship between assembly pressure and temperature evolution is established in the digital twin model, enabling the model to preview the subtle effects of different press-fitting strategies on the internal temperature rise of the battery pack.

[0077] For stress change rate from solid mechanics , which is solved by extending Hooke's law to a time-differential form that includes thermal strains:

[0078] Formula (3): ;

[0079] Formula (4): ;

[0080] Formula (3) is the basic differential form of the stress-strain relationship. The motivation for formula (4) is to accurately distinguish and couple the two physical sources of total strain: the mechanical strain caused by the external mechanical load and the thermal strain caused by the material's own expansion and contraction due to temperature changes, thus ensuring the physical authenticity of the model.

[0081] in, Represents the Young's modulus of the material, whose value is obtained from the material property data Obtained from

[0082] is the dimensionless total strain; is the mechanical strain, the size of which is determined by the mechanical stress Decision, and According to the assembly pressure instruction Make estimates;

[0083] : stress change rate; : Young's modulus of the material;

[0084] :Total strain; : Mechanical strain, caused by mechanical stress Decide; : Mechanical stress, according to assembly pressure instructions Make estimates;

[0085] : Material thermal expansion coefficient; :Current temperature; :Assembly reference temperature; : partial differential with respect to time;

[0086] For simple geometries, this estimation is done based on Hertzian contact theory; for more complex geometries, it is done using a pre-computed surrogate model based on finite element analysis: ;

[0087] in is a response surface model, such as polynomial regression or a trained neural network, developed through a series of offline finite element simulations to map assembly pressures to positions The stress response at Refers to a specific location point i on the assembly;

[0088] is the thermal strain, the magnitude of which is determined by the current temperature , Assembly reference temperature and the material's thermal expansion coefficient Jointly determined, the latter two are also based on material attribute data and obtained from process documents;

[0089] In the virtual preview, the system will use the predicted temperature field calculated by formula (1) and (2) Substitute into formula (4), and the assembly pressure The total strain rate is calculated together, and the stress change rate is solved by formula (3). This interlocking calculation process enables the stress field evolution prediction to respond to both the mechanical assembly action and the heat accumulation effect. The technical effect is that it can proactively identify local stress concentrations that may be caused by improper thermal-mechanical coupling management, thereby providing a reliable prediction method for preventing quality defects such as deformation and sealing failure of the battery pack due to excessive internal stress.

[0090] Example 4

[0091] The optimization process of the comprehensive reward function includes:

[0092] Determine efficiency rewards, safety penalties, and quality penalties; perform a weighted summation of these rewards to generate a comprehensive reward function; the efficiency reward is generated based on the predicted cycle time; the safety penalty is generated when the global maximum temperature of the predicted temperature field exceeds a preset safety soft threshold; and the quality penalty is generated based on the predicted final assembly dimension deviation.

[0093] The efficiency importance weight, safety importance weight, and quality importance weight used for weighted summation are dynamically adjusted by the system based on the analysis of order type and priority information in the production order data.

[0094] The core mechanism of the deep reinforcement learning optimization module for assembly strategies in this embodiment lies in designing and optimizing a comprehensive reward function that quantifies the core contradiction between speed and flexibility and safety and reliability. This function provides a clear optimization goal for the deep reinforcement learning agent, namely, finding an assembly strategy that maximizes long-term cumulative rewards.

[0095] In a preferred embodiment, the DRL agent uses a soft actor-critic algorithm; its state space contains the physical world state vector and current order demand information; its action space is continuous and is used to define assembly pressure , welding current The actor network and the critic network both use multi-layer perceptrons, which contain three hidden layers of 256 neurons each, and use ReLU as the activation function.

[0096] The comprehensive reward function The construction of reflects the comprehensive consideration of multiple dimensions of production objectives, which can be expressed mathematically as follows:

[0097] Formula (5): ;

[0098] The motivation for designing this formula is to transform the abstract production goals of high efficiency, high safety, and high quality into a specific, computable, dimensionless scalar reward value. By combining positive rewards representing efficiency with negative penalties representing safety and quality risks, it provides a clear guide for the exploration and learning process of the DRL algorithm, enabling it to automatically balance the pros and cons of different goals.

[0099] This function consists of three parts; the first is the efficiency reward, where is the assembly cycle time required for the current strategy predicted by the predictive digital twin modeling and pre-assembly module, As a reference to the benchmark time, this incentive algorithm seeks a faster production cycle; the second is a safety penalty, in which is the predicted global maximum temperature; The function here is to implement a judgment condition: the safety penalty term will be activated and calculated only when the predicted maximum temperature exceeds the preset safety soft threshold;

[0100] It is a safety soft threshold that is pre-set based on historical production data and safety engineering practices. The penalty is activated only when the predicted temperature exceeds the limit.

[0101] is a predetermined normalization coefficient with temperature dimension set to ensure the dimensionless nature of the penalty term; the third term is the quality penalty,

[0102] in is the predicted final assembly dimensional deviation, is a normalization coefficient with length dimension, which is also set for dimensionless purpose;

[0103] : comprehensive reward function; :Efficiency importance weight; : The predicted assembly cycle time required for the current strategy;

[0104] : Safety importance weight; : predicted global maximum temperature; :Safety soft threshold; : Normalization coefficient for dimensionless transformation, with temperature dimension; :Quality importance weight; : Predicted final assembly dimension deviation;

[0105] : Normalization coefficient for dimensionless transformation, with length dimension;

[0106] These coefficients are predetermined based on historical data and engineering specifications; for example, Can be set to absolute safety red line temperature With safety soft threshold The difference is:

[0107] ;

[0108] Similarly, The maximum allowable design tolerance that can be set for the assembly dimension;

[0109] During the optimization process, the DRL agent continuously generates candidate assembly strategies and obtains 、 and , substitute into formula (5) to calculate the reward value of the current strategy, and adjust its strategy network based on this value; the technical effect is to create a decision optimization unit that can learn and optimize autonomously, which can not only find a combination of process parameters that exceeds manual experience, but also find the optimal balance between the three mutually constrained goals of efficiency, safety, and quality based on quantitative reward feedback;

[0110] A core technical feature is that it determines the efficiency importance weight of the above three trade-offs , safety importance weight and quality importance weights It is not fixed; on the contrary, the system will actively analyze the production order data The system automatically increases the order type and priority of the battery when receiving an urgent order. The value of makes the optimization algorithm more inclined to find a more time-efficient assembly solution; when processing high-specification sample orders, the value of and , guiding the algorithm to generate more robust and accurate assembly strategies; this dynamic weight adjustment mechanism enables the optimization of assembly strategies to closely fit specific business needs, giving the production line real flexibility.

[0111] Example 5

[0112] The process of adaptive security boundary and active intervention module to determine the security boundary includes:

[0113] Determine a dynamic safety buffer margin; subtract the dynamic safety buffer margin from the absolute safety redline temperature derived from production order data to generate a dynamic safety margin; compare the predicted maximum temperature corresponding to the candidate optimal assembly strategy with the dynamic safety margin;

[0114] The process of determining the dynamic safety buffer margin includes:

[0115] Determine the model risk term and the environmental risk term; sum the model risk term and the environmental risk term to generate a dynamic safety buffer margin; the model risk term is generated based on the model prediction uncertainty and the preset risk aversion coefficient; the environmental risk term is generated based on the environmental disturbance uncertainty and the preset risk aversion coefficient;

[0116] The subsequent processing logic of the comparison includes:

[0117] If the predicted maximum temperature is less than the dynamic safety margin, the candidate optimal assembly strategy is determined to be safe and is output as the final safety instruction;

[0118] If the predicted maximum temperature is greater than or equal to the dynamic safety margin, the candidate optimal assembly strategy is determined to be unsafe. After the candidate optimal assembly strategy is revised, the safety margin is re-determined until the revised strategy is determined to be safe.

[0119] The rules for modifying the candidate optimal assembly strategy include: first modifying the process parameters that have the greatest impact on the predicted maximum temperature; for example, based on the physical model of the predicted temperature such as formula (1) and (2), sensitivity analysis is performed, if the welding heat source is the main influencing factor, then the welding current in the strategy The parameter is reduced by a preset step size such as 2%, and virtual preview and safety judgment are repeated; if the contact heat transfer effect is the main factor, the assembly pressure is reduced. Reduce a preset step size; the correction process is iterated until the predicted maximum temperature of the new strategy is less than the dynamic safety margin;

[0120] The adaptive safety boundary and active intervention module in this embodiment serves as the final active safety intervention mechanism before physical execution. This module's decisive innovation lies in abandoning the traditional fixed safety threshold approach and instead building an adaptive safety boundary that can dynamically shrink or expand based on the system's current uncertainty level.

[0121] This dynamic security boundary The judgment process begins with a core calculation model:

[0122] Formula (6): ;

[0123] This formula aims to define a maximum temperature limit that is valid in real time at time t and allows for operation. Its core motivation is to recognize that a fixed safety threshold cannot cope with the dual uncertainties of model prediction errors and environmental fluctuations. Therefore, a dynamic safety buffer must be introduced to ensure that even in the worst case scenario, the actual temperature does not exceed the absolute safety red line.

[0124] in, Represents the absolute safety red line temperature clearly specified in the battery design specification. This is an insurmountable rigid constraint, and its value is determined by the production order data. Read directly from the associated technical specification file;

[0125] This is the core of this embodiment - the dynamic safety buffer margin, which is a positive value that changes with time and represents the safety space reserved by the system to deal with uncertainty.

[0126] :Dynamic security boundary; : Absolute safety red line temperature; : Dynamic safety buffer margin;

[0127] After receiving the optimal strategy output by the DRL module and its predicted maximum temperature After that, this module first calculates the current dynamic safety boundary through formula (6); its technical effect is to establish an intelligent and variable dynamic safety boundary mechanism, which automatically adjusts its threshold according to the system's own prediction confidence and the stability of the external environment. Compared with a fixed threshold, it can provide a larger operating space for the pursuit of extreme efficiency while ensuring absolute safety.

[0128] The dynamic safety buffer margin The quantization of is further precisely defined by the following model:

[0129] Formula (7): ;

[0130] The motivation for this formula is to decompose abstract uncertainty into two specific and quantifiable risk sources: a model risk term arising from the limitations of the digital twin model’s own predictive capabilities, and an environmental risk term arising from temperature fluctuations in the physical workshop environment;

[0131] Among them, the first item is the model risk item, That is, the model prediction uncertainty, which is the maximum temperature predicted by the internal probability model when the DRL module performs strategy evaluation. The standard deviation of quantifies the confidence level of the model in this prediction;

[0132] This uncertainty quantification is achieved by using the Monte Carlo dropout technique in the inference phase of the DRL agent’s value network or Q network, where for the same input state, times, for example, The forward propagation with activated Dropout can obtain a prediction of the maximum temperature The standard deviation of this distribution is used as the model prediction uncertainty: ;

[0133] The second item is the environmental risk item. That is, environmental disturbance uncertainty, which is the environmental temperature recently collected by the multi-source data perception and fusion module The data is obtained through statistical analysis, which reflects the real-time fluctuation intensity of the environment; and are two dimensionless risk aversion coefficients, which are adjustable parameters that characterize risk preferences and are configured by system managers according to specific production stages or safety requirements; : Number of forward propagation executions with Dropout activation; : The predicted maximum temperature obtained by the i-th forward propagation; : The average value of the predicted maximum temperature obtained by N forward propagations;

[0134] As a guideline, typical values ​​for these coefficients are between 1.0 and 3.0, e.g. Corresponding to a safety margin of one standard deviation, the value is , similar to the 3-sigma principle, provides a higher confidence level, ensuring that the actual temperature will not exceed the dynamic boundary, for first production or high-value sample orders, and It can be set to 3.0; for mature, large-scale production, it can be appropriately reduced to 1.5 to improve production efficiency while ensuring safety;

[0135] : Dynamic safety buffer margin; : Risk aversion coefficient; : Model prediction uncertainty is a quantification of the model's confidence in this prediction;

[0136] : Risk aversion coefficient; : Environmental disturbance uncertainty, reflecting the severity of real-time environmental fluctuations;

[0137] The system uses formula (7) to sum the model uncertainty with the environmental uncertainty to generate the final safety margin; its technical effect is to create a sophisticated mechanism that can quantify and manage risks; when the model has a low confidence in predicting a novel assembly strategy, Increase or workshop temperature fluctuates violently, When the value increases, the safety margin will automatically increase, and the safety margin will be tightened, forcing the system to take more conservative actions. Conversely, under stable and certain working conditions, the margin will be appropriately relaxed to release production efficiency.

[0138] Based on the above calculations, the module predicts the maximum temperature corresponding to the optimal assembly strategy and the calculated dynamic security boundary Compare; if , then the strategy is determined to be safe and is directly sent to the physical production line execution and closed-loop feedback module as the final safety instruction; on the contrary, if , then the strategy is judged to be risky and the active intervention mechanism is triggered. The system will automatically correct the strategy according to the preset rules and drive the corrected strategy back to the predictive digital twin modeling and pre-assembly module for re-virtual rehearsal and safety boundary judgment until it is finally judged to be safe. This closed-loop correction logic fundamentally eliminates the possibility of issuing any operation instructions with potential risks to the physical world, thereby achieving the ultimate guarantee of production safety.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A digital twin-based assembly workshop production line construction system, characterized by: include: Multi-source data perception and fusion module, used to collect production order data, physical environment data, equipment status data and material attribute data, and fuse the collected data to generate a physical world state vector; The assembly strategy deep reinforcement learning optimization module is used to generate candidate assembly strategies and virtually preview the candidate assembly strategies using predictive digital twin modeling and pre-assembly modules to optimize the comprehensive reward function, ultimately generating candidate optimal assembly strategies and corresponding model prediction uncertainties. The predictive digital twin modeling and pre-assembly module is used to receive the physical world state vector and candidate assembly strategies during the virtual preview process to solve the corresponding predicted temperature field and predicted stress field for use by the assembly strategy deep reinforcement learning optimization module; The adaptive safety boundary and active intervention module receives candidate optimal assembly strategies and model prediction uncertainty, obtains environmental disturbance uncertainty and absolute safety redline temperature, makes safety boundary determinations, and generates final safety instructions. The physical production line execution and closed-loop feedback module is used to execute the final safety instructions and collect real physical data to update the physical world state vector, thus forming a data closed loop; The optimization process of the comprehensive reward function includes: Determine efficiency rewards, safety penalties, and quality penalties; perform a weighted summation of these rewards to generate a comprehensive reward function; the efficiency reward is generated based on the predicted cycle time; the safety penalty is generated when the global maximum temperature of the predicted temperature field exceeds a preset safety soft threshold; and the quality penalty is generated based on the predicted final assembly dimension deviation. The efficiency importance weight, safety importance weight, and quality importance weight used for weighted summation are dynamically adjusted by the system based on the analysis of order type and priority information in the production order data; The process of adaptive security boundary and active intervention module to determine the security boundary includes: Determine a dynamic safety buffer margin; subtract the dynamic safety buffer margin from the absolute safety redline temperature derived from production order data to generate a dynamic safety margin; compare the predicted maximum temperature corresponding to the candidate optimal assembly strategy with the dynamic safety margin; The process of determining the dynamic safety buffer margin includes: Determine the model risk term and the environmental risk term; sum the model risk term and the environmental risk term to generate a dynamic safety buffer margin; the model risk term is generated based on the model prediction uncertainty and the preset risk aversion coefficient; the environmental risk term is generated based on the environmental disturbance uncertainty and the preset risk aversion coefficient.

2. The assembly workshop production line construction system based on digital twin according to claim 1 is characterized in that: The predictive digital twin modeling and pre-assembly module predicts the evolution of the predicted temperature field and the predicted stress field by solving the temperature change rate and stress change rate.

3. The assembly workshop production line construction system based on digital twin according to claim 2 is characterized in that: The temperature change rate calculation process includes: Determine the volumetric heat source generation rate; combine the volumetric heat source generation rate with the material's inherent thermophysical properties derived from material attribute data to calculate the temperature change rate; the volumetric heat source generation rate is used to characterize the heat generated during the assembly process, and its value is determined by the welding heat source and the contact heat transfer effect; the size of the welding heat source is related to the welding current derived from the equipment status data; the size of the contact heat transfer effect is related to the assembly pressure instructions output by the deep reinforcement learning optimization module of the assembly strategy.

4. The assembly workshop production line construction system based on digital twin according to claim 2 is characterized in that: The stress rate of change solution process includes: Determine the total strain; solve the time rate of change of the total strain to obtain the rate of change of stress based on the material Young's modulus derived from the material property data; the total strain is composed of mechanical strain and thermal strain; the mechanical strain is determined by the assembly pressure command; the thermal strain is determined by the predicted temperature field and the material thermal expansion coefficient derived from the material property data.

5. The assembly workshop production line construction system based on digital twin according to claim 4 is characterized in that: The subsequent processing logic of the comparison includes: If the predicted maximum temperature is less than the dynamic safety margin, the candidate optimal assembly strategy is determined to be safe and is output as the final safety instruction; If the predicted maximum temperature is greater than or equal to the dynamic safety margin, the candidate optimal assembly strategy is determined to be unsafe. After the candidate optimal assembly strategy is revised, the safety margin is re-determined until the revised strategy is determined to be safe.

Images