Intelligent control strategy and framework for continuous charging production line

Abstract

The invention discloses a continuous charging production line intelligent management and control strategy and framework, and the framework further comprises a multi-physics field coupling modeling module which carries out the physical mechanism modeling and state prediction of a production process through a heat-fluid-solid three-field coupling model on the basis of a conventional data collection and control module; the intelligent twin module adopts a mode of dynamically fusing a mechanism model and a data-driven residual model to improve the precision and robustness of system state estimation; the scheduling optimization module is used for realizing dynamic adaptive scheduling of AGV paths and production tasks based on a multi-target deep reinforcement learning strategy; and the multi-layer safety causal reasoning module realizes advanced prediction and active prevention and control of safety risks through a causal map, fuzzy logic and a linkage response mechanism. According to the method, the problem that safety, quality and efficiency are difficult to collaboratively optimize in the prior art is solved, and the predictability, safety, scheduling efficiency and adaptive capacity of the production line are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of drug loading control technology, specifically an intelligent control strategy and framework for a continuous drug loading production line. Background Technology

[0002] Continuous solid propellant manufacturing lines for continuous charging exhibit several challenges during the mixing, feeding, transfer, and casting processes. These challenges include strong coupling of multiple factors, severe operational disturbances, high safety control requirements, complex scheduling tasks, and difficulties in coordinating safety, quality, and efficiency. Traditional SCADA / MES-based production management systems, primarily relying on static process control and limit alarm mechanisms, are ill-suited to the following needs: (1) Real-time collaborative modeling and state perception requirements for multi-physics fields (thermal, structural, and material flow); (2) The need for highly robust scheduling and precise control under sudden disturbances; (3) Edge computing deployment and lightweight operation requirements under real-time feedback control; (4) Feedforward modeling and interpretable early warning requirements for equipment-level fault prediction and safety response.

[0003] Existing technologies are mostly focused on single technical dimensions such as digital twin modeling, AGV scheduling optimization, and intelligent MES, and have not yet formed a systematic closed-loop intelligent management and control system that can be deployed, integrated, and predicted. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide an intelligent control strategy and framework for a continuous drug loading production line. Through thermal-fluid-solid coupling modeling, mechanism-data dual-mode twin modeling, deep reinforcement learning scheduling optimization, and causal graph risk reasoning, the predictability, safety, and response efficiency of the system under complex working conditions are improved.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A smart control strategy and framework for a continuous drug loading production line includes a data acquisition system for collecting sensor data from the production line and a control module for executing control commands. The smart control system further includes: The multiphysics coupling modeling module is configured to establish a system state model describing the thermo-solid-fluid coupling mechanism of the slurry during the mixing process by coupling the control equations of the thermal field, structural field and fluid field, and output the predicted values ​​of key physical states including temperature distribution, structural stress and fluid viscosity. The intelligent twin module is communicatively connected to the multiphysics coupling modeling module and is configured to fuse the mechanistic model prediction results from the multiphysics coupling modeling module with a data-driven residual model prediction result, and output high-precision system state reconstruction values ​​and future state prediction values ​​through a dynamic weighted fusion mechanism. The scheduling optimization module, which is communicatively connected to the intelligent twin module, is configured to receive the system state reconstruction value and the future state prediction value, and generate and output path planning instructions and production task scheduling instructions based on a multi-objective deep reinforcement learning policy network. The multi-layer safety causal reasoning module is communicatively connected to the intelligent twin module and the scheduling optimization module, respectively. It is configured to predict the risks of excessively rapid temperature rise, uneven slurry distribution, or equipment failure based on causal graph reasoning, fuzzy logic judgment, and linkage response mechanism, and output corresponding early warning signals, and / or directly output control commands to the scheduling optimization module or the control module.

[0006] Furthermore, the multiphysics coupling modeling module includes: The thermal field modeling unit is configured to establish a temperature field model of the slurry within the twin-screw extruder based on the partial differential equation of heat conduction. in: Density of the medicated paste; The specific heat capacity of the slurry; For the medicine in position and time Temperature; Thermal conductivity; The volumetric heat source term is modeled as a stirring rate function: , For mixed shaft torque, For rotational speed, The energy conversion efficiency coefficient of stirring; The structural field modeling unit is configured to establish a total stress model for a twin-screw hybrid shaft based on thermo-structural coupled elasticity: in: The total stress; The stress is caused by thermal bending moment, and , For elastic modulus, For thermal expansion curvature; The torque applied to the hybrid shaft; The radius of the axial section; Let be the moment of inertia of the axial cross section; It is the polar moment of inertia; The fluid field modeling unit is configured to establish a slurry flow model based on the generalized Navier-Stokes equations for non-Newtonian fluids: in: It is a velocity vector; For pressure; Density; Effective viscosity depends on shear rate. and temperature ; It is a volume force; The effective viscosity is fitted using either the Carreau model or the Cross model. in: Zero shear viscosity; Viscosity at infinite shear rate; It is a time constant; The liquidity index; The state equation of the system state model is: in: This indicates the internal temperature field, stress field, and heat flow state of the system. This represents the system input control vector, and also represents an external controllable or measurable signal. This represents a system noise or uncertain disturbance vector, used to characterize dynamics not modeled by the model, sensor noise, or external interference. It represents the dynamic matrix or operator of the system, reflecting the coupling relationship and dynamic characteristics of the system, and can be a linear matrix or a nonlinear operator (e.g., thermo-mechanical coupling operator). Indicates the system temperature status; This indicates the thermal stress or thermal strain state of the system. This indicates the heat flux density or heat source intensity within the system. Indicates rotational speed; This indicates the control commands or speed input for the automated guided vehicle (AGV). This indicates the ambient temperature or external thermal boundary conditions.

[0007] Furthermore, the multiphysics coupling modeling module employs an adaptive mesh refinement strategy based on error indices. For regions with drastic changes in thermal gradient, local mesh refinement is performed, including: For state variables; To refine the threshold; , and These represent the temperature field, shear stress, and structural stress in the system state model, respectively. The Sobol global sensitivity analysis method is used to evaluate the impact of input parameters on output state variables. The first-order Sobol sensitivity exponent is defined as: in: For the first The input variable for the first... The percentage of variance explained by each response; For the first One output target state vector; For the first A vector of input physical parameters.

[0008] Furthermore, the intelligent twin module includes a mechanism model, a data model, and a dynamic fusion mechanism; The mechanism model It is configured to run a simplified physical model derived from the multiphysics coupling modeling module; The data model It is configured to run a trained LSTM or Transformer neural network to learn and predict the residuals between the mechanistic model predictions and the actual system state; The dynamic fusion mechanism is configured to output based on the mechanism model. and the output of the data model The fused prediction value is calculated using a Bayesian dynamic weighting mechanism: in: For the dynamic fusion output of the intelligent twin module; This is the output of the mechanistic model; The output of the data model; The fusion weights are calculated based on the mean square error of the sliding residuals: in: The error variance of the mechanistic model; This represents the error variance of the data model.

[0009] Furthermore, the fusion prediction controller is further configured as follows: Define a working condition adaptability factor : in: The disturbance intensity; The disturbance handover threshold; Sensitivity factor; Using the aforementioned working condition adaptability factor Dynamically adjust the output of the fusion prediction so that when When the value is close to 1, the data model dominates the fusion prediction. When the value is close to 0, the mechanistic model dominates the fusion prediction.

[0010] Furthermore, the scheduling optimization module employs a reinforcement learning policy network based on the DDPG or TD3 algorithm, wherein: The state space S is defined as follows: in: Indicates the first The operating status vector of each workstation or equipment unit, including features such as temperature, torque, energy consumption, and process progress; Action space A is defined as: The reward function is defined as: in: This indicates the task completion time or total production cycle time, used to measure overall production efficiency. Indicates the system in time Energy consumption indicators, including energy consumption of motors, heaters, AGVs, etc.; Indicates system operational risk indicators, where local stress Exceeding the stress safety threshold The penalty item is triggered at any time, reflecting the risk of exceeding the limit; This indicates a penalty for frequent changes in process parameters; Indicates localized stress; Indicates the stress safety threshold; , , and These represent the production efficiency weighting coefficient, energy consumption weighting coefficient, safety and risk weighting coefficient, and switching frequency penalty weighting coefficient, respectively.

[0011] Furthermore, the scheduling optimization module adopts a hierarchical reinforcement learning structure, including: High-level task strategy: Input is the task pool Equipment pool and order priority Output task - device matching action : in: Indicates the task layer state, indicating the task pool. Equipment pool and order priority A comprehensive state vector containing information such as [list of information] is used to describe the current task allocation environment of the system; This represents a high-level policy network, which considers the input state. Make decisions and output the matching and allocation actions between tasks and AGVs; that is, determine which AGV will perform which task. Indicates the first task in the task pool Tasks include material transportation, mixing tank feeding, and finished product transfer. Indicates the first An automated guided vehicle (AGV) is responsible for moving and performing the assigned tasks. Low-level path strategy: Inputs include the AGV's current position, environmental obstacles, heat map, and energy consumption estimate; output is the path decision. : in: It represents the path layer status, including the AGV's current position, environmental obstacle map, hot zone distribution map, and energy consumption estimation information, which is used for path planning and obstacle avoidance decisions. This represents a low-level policy network that, given assigned tasks, outputs the optimal path or control action to enable autonomous navigation of the AGV. Indicates the first path in the candidate path set The features or scores of each path are used to select reinforcement learning strategies; The Critic supervisor is used to jointly evaluate the comprehensive return function. in: Incentives for efficient dispatching; Energy consumption penalty per unit task; Penalty for the number of times a player crosses a hot zone; , and These are adaptively adjustable target preference weights.

[0012] The scheduling optimization module integrates a fast transfer meta-learning mechanism and is configured as follows: In multiple source tasks Learning initial policy parameters : When a new task arrives, you can adapt quickly: in: This indicates that the process involves a meta-task (i.e., multiple source tasks). The optimal initial policy parameters obtained during training represent the model's general prior knowledge across multiple tasks, enabling it to converge quickly in new tasks. Indicates the first Individual source task The loss function, which typically corresponds to negative reward or prediction error in reinforcement learning, is used to measure the performance of the policy on the task. This represents the learning rate or step size coefficient, which controls the magnitude of gradient updates per step. Indicates the parameter The gradient operator represents the direction and magnitude of the gradient of the model parameters with respect to the loss function.

[0013] Furthermore, the multi-layered secure causal reasoning module includes a causal graph reasoning layer, a fuzzy logic judgment layer, and a linkage response layer; The causal graph inference layer constructs a Bayesian network based on a variable causal graph to model the causal relationships between process variables. The variable causal graph is represented as follows: in: Indicates ambient temperature; Indicates the temperature of the equipment wall surface; Indicates the viscosity of the slurry; Indicates mixed torque; Indicates the stress on the stirring shaft; In a variable causal graph, each edge is assigned a causal strength. This forms a Bayesian network: The fuzzy logic inference layer stores several semantic rules in the form of "IF-THEN", and uses Gaussian or trapezoidal membership functions to fuzzify input variables into levels such as "high", "medium", and "low", and then calculates the security risk level based on the fuzzy inference rules; the semantic rules in the form of "IF-THEN" are expressed as follows: If temperature rise rate > High AND twin-screw shaft torque > Large, then the hazard level is High. Mathematically, it is expressed as: in: The membership function representing the high-risk level indicates the confidence or credibility of the system being in a high-risk state; A fuzzy membership function representing the rate of temperature rise, used to reflect whether the rate of temperature rise exceeds a safe threshold; A fuzzy membership function representing the stirring torque is used to reflect the mechanical load level during the stirring process; The linkage response layer is configured to map the safety risk level to a predetermined response strategy, wherein a high risk level triggers a stop to mixing and AGV path reconstruction, a medium risk level triggers a reduction in mixing speed and an increase in cooling, and a low risk level triggers logging and an increase in monitoring frequency.

[0014] Furthermore, the fuzzy logic inference layer embeds linear temporal logic rules for modeling time-coupled anomaly patterns. The rule form is as follows: in: Let be the membership function; It represents the rate of temperature rise, i.e., the derivative of temperature with time, and is used to characterize the intensity of the thermal dynamic changes of the system. This indicates the change in torque, reflecting the degree of load fluctuation on the stirring shaft within a short period of time. It is usually related to sudden changes in material viscosity or equipment jamming. The membership function represents the risk level, and the fuzzy confidence level represents the current high-risk state of the system. Indicates the future The linear-time logic (LTL) operator is "finally satisfied" within seconds; The variable causal graph is automatically constructed by Granger causality test and PCMCI+ algorithm to automatically mine the potential causal path structure between sensor variables, and a time series causal discovery algorithm is introduced to construct a data-driven dynamic graph structure. The Granger causality test determines causal relationships based on the predictive power of time lags. The principle is as follows: If added Significant improvement The predictive ability is considered to be... ; in: It represents the value of the explained variable (target variable) at the current time, that is, the current response or observation of the system; Represents the explained variable itself The lagged term represents the past. Historical values ​​within a time step are used to capture its autoregressive characteristics; This indicates that exogenous variables (potential influencing factors) in The value of the first lag time reflects the time influence of external variables on the target variable; Represents random disturbance terms or white noise, used to represent unmodeled noise or random effects; The lag order represents the number of historical steps considered, and is usually determined by information criteria (such as AIC, BIC). and These represent the autoregressive coefficient and the cross-regression coefficient, respectively, describing... Self-history and external variables For the current value The intensity of the impact; The PCMCI+ algorithm is suitable for high-dimensional, multi-lag data, supports pruning and residual analysis to form a dynamically updated directed acyclic graph, where each edge represents a potential causal chain. (Time lag at the edge of causality) in: and They represent the first The and the first A time series variable, representing the target quantity that is affected or predicted; Indicates the time lag, representing the amount of time delay in the causal effect, reflecting... right The lag effect.

[0015] The beneficial effects of this invention are as follows: The intelligent control strategy and framework for the continuous drug loading production line of the present invention, through the construction of a four-dimensional integrated technical architecture of "multi-physics field coupling modeling - intelligent twin fusion - reinforcement learning scheduling - multi-layer safety reasoning", has achieved significant technical progress in state perception, decision optimization and safety control compared with the traditional SCADA / MES system, and has the following technical effects.

[0016] (1) In terms of state perception and prediction accuracy, through thermal-fluid-solid multi-physics coupling modeling, the system has for the first time achieved coordinated, real-time and mechanistic prediction of key physical quantities such as slurry temperature, equipment stress and flow state, overcoming the limitation of traditional systems that rely solely on static limit alarms. Furthermore, the intelligent twin module adopts dynamic fusion of mechanism and data dual models, which can adapt to working condition disturbances and compensate for model deviations in real time, effectively reducing the overall prediction error and significantly improving the system's state perception and prediction capabilities under unsteady working conditions.

[0017] (2) In terms of scheduling efficiency and adaptability, the deep reinforcement learning strategy can integrate high-dimensional information such as AGV position, equipment thermal status, and material status to make joint optimization decisions for multiple objectives (such as task delay, congestion, and thermal load). It breaks through the traditional scheduling logic based on fixed rules, enabling the scheduling strategy to evolve dynamically with the environment, effectively improving scheduling efficiency, reducing task completion time, and demonstrating strong environmental adaptability and optimization potential.

[0018] (3) In terms of initiative and foresight in safety risk prevention and control, through the multi-layered safety architecture of causal graph, fuzzy reasoning and linkage response, a fundamental shift from "post-event alarm" to "pre-event prediction and avoidance" has been achieved. It can explore the deep causal relationship between variables, and quantitatively assess and warn of composite risks in advance. It can effectively extend the warning time of high-risk situations and automatically trigger scheduling or control response, thereby greatly improving the inherent safety level and system robustness of the production process. Attached Figure Description

[0019] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a framework diagram of the intelligent control strategy and framework for the continuous drug loading production line of the present invention. Figure 2 A flowchart for the deployment and implementation of a thermal-fluid-solid three-field modeling and state control system; Figure 3 A block diagram illustrating the integration principle of reinforcement learning schedulers and digital twins; Figure 4 This is a schematic diagram of a safety reasoning system and early warning response. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0021] I. Intelligent Control Strategy and Framework like Figure 1 As shown, the intelligent control strategy and framework of the continuous loading production line in this embodiment constructs a four-dimensional integrated structure for edge deployment, which consists of "multi-physics modeling + intelligent twin fusion + enhanced scheduling + multi-layer security reasoning". Through thermal-fluid-solid coupling modeling, mechanism-data dual-mode twin modeling, deep reinforcement learning scheduling optimization and causal graph risk reasoning, the predictability, safety and response efficiency of the system under complex working conditions are improved.

[0022] Specifically, the intelligent control strategy and framework for the continuous drug loading production line in this embodiment includes a data acquisition system, a multi-physics coupling modeling module, an intelligent twin module, a scheduling optimization module, a multi-layer safety causal reasoning module, and a control module.

[0023] 1.1 Multiphysics Coupled Modeling Module The multiphysics coupling modeling module is configured to establish a system state model describing the thermo-solid-fluid coupling mechanism of the slurry during the mixing process by coupling the governing equations of the thermal field, structural field, and fluid field, and output predicted values ​​of key physical states, including temperature distribution, structural stress, and fluid viscosity. In this embodiment, the multiphysics coupling modeling module includes a thermal field modeling unit, a structural field modeling unit, and a fluid field modeling unit.

[0024] Specifically, the multiphysics coupling modeling module establishes a joint modeling mechanism for the thermal field (based on the PDE conduction model), the structural field (based on the thermally induced stress model), and the fluid field (flow and viscosity change model), forming state-space control equations. It analyzes the thermo-solid-fluid coupling mechanism of the slurry in the twin-screw mixing process, providing a foundation for the coordinated optimization of safety, efficiency, and quality. The module predicts the thermal distribution of the twin-screw through temperature gradients and heat source terms, calculates the thermally induced bending and torsional coupling stress of the twin-screw through the structural field response, and characterizes the slurry flow state and viscosity evolution. This module supports numerical solutions based on the finite element method or finite difference method. The model output can serve as key physical state inputs for subsequent scheduling optimization, safety assessment, and edge control, achieving organic integration from physical mechanism modeling to multi-field predictive control, enhancing the system's ability to perceive and predict state changes under unsteady and complex environments.

[0025] Specifically, the heat conduction partial differential equation modeling results are verified using 1D / 2D sensor array data; structural field modeling employs thermo-structural coupled elasticity, and the structural field modeling results are verified through strain gauge / stress sensor inversion; fluid field modeling is based on non-Newtonian fluid theory, considering the shear rate distribution and viscosity evolution characteristics of the slurry within the twin-screw mixing cylinder, establishing a velocity-pressure coupled Navier-Stokes equation set, and introducing a viscosity model based on temperature and shear rate (such as the Carreau or Cross model) to describe rheological properties. Steady-state or dynamic flow pattern prediction is achieved through speed control coupled with the thermal field boundary, and key modeling results are used for model inversion and error correction using data from flow sensors, pressure sensors, and online viscometers. These three field models constitute a complete thermo-structural-fluid three-dimensional coupled system, providing a high-precision state prediction foundation and controllable parameter space for subsequent residual modeling, twin fusion, scheduling optimization, and safety response modules.

[0026] 1.1.1 Thermal Field Modeling Unit The thermal field modeling unit is configured to establish a temperature field model of the slurry within the twin-screw extruder based on the partial differential equation of heat conduction. Assuming the mixing cylinder is a one-dimensional cylindrical model and ignoring radial non-uniformity, the temperature field model constructed using the partial differential equation of heat conduction is expressed as follows: in: Density of the medicated paste; The specific heat capacity of the slurry; For the medicine in position and time Temperature; Thermal conductivity; This is the volumetric heat source term.

[0027] Specifically, the left side of the formula represents the rate of energy change per unit volume of slurry; the first term is the heat conduction term; the second term is the internal heat source term (such as heat generated by stirring); an external PID controller can be connected to adjust the outer wall temperature of the cylinder (boundary condition) to control the temperature. .

[0028] In practical modeling applications, temperature sensors are installed to obtain boundary and internal temperatures; volumetric heat source term It can be modeled as a stirring rate function: in: For mixed shaft torque, For rotational speed, This is the energy conversion efficiency coefficient for stirring.

[0029] 1.1.2 Structural Field Modeling Unit The structural field modeling unit is configured to establish a total stress model for the twin-screw hybrid shaft based on thermo-structural coupled elasticity. Specifically, considering the impact of thermo-structural coupling on the safety of the hybrid equipment, the stress and deformation of the twin-screw hybrid shaft are modeled. The total stress model (thermal bending + torsional linkage) is expressed as follows: in: The total stress; The stress is caused by thermal bending moment, and , For elastic modulus, For thermal expansion curvature; The torque applied to the hybrid shaft; The radius of the axial section; Let be the moment of inertia of the axial cross section; It is the polar moment of inertia.

[0030] The thermal expansion curvature is caused by the temperature gradient and is expressed as: in: This represents the coefficient of linear expansion. Thus, the thermally induced stress field of the twin-screw mixer can be obtained.

[0031] 1.1.3 Fluid Field Modeling Unit The fluid field modeling unit is configured to establish a slurry flow model based on the generalized Navier-Stokes equations for non-Newtonian fluids. Specifically, to accurately describe the flow characteristics of slurry in a twin-screw mixer in a continuous charging line, a fluid field modeling system based on non-Newtonian fluid dynamics is constructed to establish the coupling relationship between the velocity field, pressure field, and temperature field, which is used to predict the flow behavior, distribution uniformity, and viscosity variation trend of the slurry under different process parameters.

[0032] The fluid field model aims to achieve the following functions: (1) accurately characterize the velocity distribution and shear stress field of the slurry in the twin-screw mixer; (2) characterize the residence time distribution in different regions and evaluate the mixing uniformity; (3) establish the mapping between flow parameters and process adjustments (such as screw speed and cooling temperature); (4) provide basic flow state parameters for equipment thermal load, AGV logistics scheduling and quality control.

[0033] As a shear-thinning non-Newtonian fluid, the flow behavior of the slurry satisfies the generalized Navier-Stokes equations, which can be expressed as follows under incompressible steady-state conditions: in: It is a velocity vector; For pressure; Density; Effective viscosity depends on shear rate. and temperature ; It is a volume force.

[0034] The viscosity of the slurry exhibits shear-thinning characteristics and can be fitted using either the Carreau or Cross model. Specifically, the effective viscosity is fitted using either the Carreau or Cross model. in: Zero shear viscosity; Viscosity at infinite shear rate; It is a time constant; This is the liquidity index.

[0035] The model parameters were obtained by fitting experimental data of actual samples using a rotational rheometer or an online viscometer. The boundary conditions included: (1) Inlet boundary: setting the inlet velocity distribution or screw rotation speed; (2) Wall boundary: setting no-slip condition and thermal boundary (temperature / heat flux); (3) Outlet boundary: setting atmospheric pressure or zero normalized pressure; (4) Coupling condition: with thermal field temperature. Linkage, influence change.

[0036] Numerical discretization was performed using the finite volume method (FVM) or the finite element method (FEM), and modeling and solving were performed using COMSOL Multiphysics or OpenFOAM platforms. For complex boundaries (such as screw sections), the sliding mesh method or boundary mapping could be used. The modeling results were verified and calibrated by the following means: (1) using online pressure sensors to compare the flow pressure at key locations; (2) using online flow meters to monitor the consistency between the outlet flow velocity and the modeling results; (3) using an embedded online viscometer to obtain the viscosity changes under shear conditions in real time and comparing them with the model. Comparison; (4) The Residence Time Distribution method was used to verify the model’s ability to predict residence time and mixing uniformity.

[0037] 1.1.4 State Equations of the System State Model The three fields are unified into a state equation, which is expressed as: in: This indicates the internal temperature field, stress field, and heat flow state of the system. This represents the system input control vector, and also represents an external controllable or measurable signal. This represents a system noise or uncertain disturbance vector, used to characterize dynamics not modeled by the model, sensor noise, or external interference. It represents the dynamic matrix or operator of the system, reflecting the coupling relationship and dynamic characteristics of the system, and can be a linear matrix or a nonlinear operator (e.g., thermo-mechanical coupling operator). Indicates the system temperature status; This indicates the thermal stress or thermal strain state of the system. This indicates the heat flux density or heat source intensity within the system. Indicates rotational speed; This indicates the control commands or speed input for the automated guided vehicle (AGV). This indicates the ambient temperature or external thermal boundary conditions.

[0038] The modeling process is as follows: (1) Establish the state equations for the thermal field, structural field, and material field respectively; (2) Discretize the state equations using finite difference (FDM) or finite element method (FEM); (3) Embed the solver in the SCADA system (such as COMSOLServer / self-built C++ module); (4) Output key state indicators (such as thermal risk index, critical value of axial stress, and AGV conflict probability); (5) Trigger control commands through linkage scheduling strategy or alarm system.

[0039] 1.1.5 Adaptive Multi-Field Mesh Modeling and Sensitivity Analysis of Parameter Uncertainty To address the regions of drastic thermal gradient changes and boundary layer shear zones within the twin-screw hybrid cavity, an error-driven adaptive mesh refinement strategy is introduced. This mechanism adaptively refines the mesh based on transient heat flux changes and structural stress gradients, improving local solution accuracy and suppressing simulation errors caused by thermal mismatch. The Sobol global sensitivity analysis method is employed to systematically evaluate the influence coefficients of core parameters such as boundary heat flux, stirring heat source term, and non-Newtonian fluid exponent on state variables (e.g., peak temperature, shear stress, bending stress), establishing a priority ranking of physical parameters and the boundary of the model convergence domain.

[0040] Specifically, regarding the temperature field in the thermo-fluid-structure interaction model Shear stress Structural stress When the variables exhibit strong nonlinear changes near the boundary of the mixing cavity, the following error index-based approach is introduced. Criteria for local mesh refinement in regions with drastic thermal gradient changes: in: For state variables; To refine the threshold.

[0041] Once refinement is triggered, the local mesh count increases by one level until the maximum depth or the error is controllable. This mechanism enables finite element models to achieve local high resolution with limited resources.

[0042] Define the input physical parameter vector: in: For boundary heat flux; For stirring heat source items; For non-Newtonian fluids, the index is used.

[0043] Output target state vector: The Sobol global sensitivity analysis method is used to evaluate the impact of input parameters on output state variables. The first-order Sobol sensitivity exponent is defined as: in: For the first The input variable for the first... The percentage of variance explained by each response; For the first One output target state vector; For the first A vector of input physical parameters.

[0044] 1.2 Intelligent Twin Module The intelligent twin module is communicatively connected to the multiphysics coupled modeling module and is configured to fuse the mechanistic model prediction results from the multiphysics coupled modeling module with the prediction results of a data-driven residual model. Through a dynamic weighted fusion mechanism, it outputs high-precision system state reconstruction values ​​and future state prediction values.

[0045] The intelligent twin module constructs a residual mapping model driven by a hybrid approach of mechanism (PDE-FEM) and data (LSTM / Transformer). It achieves dual-model fusion prediction through a Bayesian dynamic weighting mechanism, aiming to improve the accuracy and robustness of system state estimation under complex operating conditions and disturbances. The mechanism model uses physical laws to describe the steady-state process, possessing interpretability and generalizability, while the data model compensates for unmodeled errors and dynamic disturbances through residual learning. The fusion weights are estimated using Bayesian methods based on real-time error covariance, enabling the model to automatically switch trust mechanisms according to changes in operating conditions, thereby achieving adaptive suppression of prediction errors and high-precision reconstruction of the system state.

[0046] 1.2.1 Modeling Objectives The intelligent twin module aims to address the problems of insufficient prediction accuracy, model rigidity, and poor adaptability of traditional digital twin systems in environments with multiple perturbations and high nonlinearity. By constructing an intelligent twin system integrating a "mechanism model + data model + dynamic fusion mechanism," it balances physical consistency and data sensitivity, forming an adaptive, high-precision, and deployable prediction core.

[0047] 1.2.2 Modeling Process Traditional digital twins focus on mapping physical processes, but when faced with complex coupling, multiple disturbances, and highly nonlinear systems in continuous charging processes, it is necessary to introduce: (1) constructing a thermal-structure-fluid coupling model based on physical laws and constructing a mechanism model (PDE-FEM). (2) Data model (LSTM / Transformer): Based on historical operating data, learn non-modeling disturbance terms and operating condition drift trends to achieve residual compensation; (3) Dynamic fusion mechanism: With Bayesian error inversion or Kalman filtering as the core, realize the fusion of multi-model prediction results with dynamic weighting and error adaptation.

[0048] In this embodiment, the intelligent twin module includes a mechanism model, a data model, and a dynamic fusion mechanism. Mechanism Model Configured to run a simplified physical model exported by the multiphysics coupling modeling module to solve the thermal-structural-fluid multiphysics model, construct a steady-state prediction backbone, and output prediction data. Data Model It is configured to run a trained LSTM or Transformer neural network to learn and predict the residuals between the mechanistic model's predictions and the actual system state; that is, to learn historical perturbation errors and modeling biases, perform residual predictions, and output predicted data. The dynamic fusion mechanism is configured to output based on a mechanistic model. and the output of the data model The fused predicted value is calculated using a Bayesian dynamic weighting mechanism. By calculating the Bayesian dynamic weights, the fused predicted value and confidence interval are output. The output data is as follows. and .

[0049] (1) Model output In this embodiment, the unified output of the intelligent twin module is represented as follows: in: For the dynamic fusion output of the intelligent twin module; This is the output of the mechanistic model; The output of the data model; To integrate the weights, an adaptive weighting is applied between the prediction results of the mechanistic model and the data model. Time-dependent mechanism, when Time-dependent data.

[0050] In this embodiment, it is assumed that the error of the mechanistic model follows a certain order. The data model error follows The fusion weights are calculated based on the mean square error of the sliding residuals: in: The error variance of the mechanistic model; This represents the error variance of the data model.

[0051] That is, when the physical model is more accurate, the weights are closer to the mechanism; when the data model is more accurate (e.g., with strong perturbations), the focus shifts to data-driven approaches. In engineering terms, this can be approximated as: in: This represents the actual observed value obtained by the sensor or experiment. This represents the sensitivity coefficient.

[0052] When the system operates under stable conditions (small disturbances), the mechanistic model takes precedence to ensure physical interpretability; when disturbances are severe or sensor feedback changes abruptly, the data model takes precedence to improve local adaptability. In this way, the system can automatically switch the main model to improve the prediction accuracy for edge cases such as cooling failure and load fluctuations.

[0053] The mechanistic model is built in COMSOL / Ansys and then trimmed into a lightweight model. The FMU / Python / C++ modules are exported and deployed to Jetson / industrial PCs. The data model is trained in PyTorch using LSTM or Transformer networks, transformed by ONNX, and deployed to edge devices, accelerated by TensorRT. The fusion predictor embeds error sliding window calculation logic within a C++ or Python module, supporting GPU / CPU adaptive inference. The output format is fused predictions (such as thermal risk index, stress prediction value, and flow field stability) sent to the MES / scheduling system via REST API or OPC-UA interface.

[0054] (2) Residual modeling Real system state The mechanistic model predicts the following results: Define residual : Training data model Fitting mapping: in: This represents an LSTM or Transformer network model; Indicates the system status within the time window The historical sequence within; This represents the historical sequence of external inputs (such as control variables, loads, and environmental disturbances) within the same time window. In a twin, there is mechanism prediction. Data prediction with residual correction At this point, the system determines the confidence level (by...). (Dynamic calculation) then execute: The actual system output should be: Therefore, we get: This formula illustrates how the error in the mechanism is predicted by the data model, and how it can be directly corrected in the mechanism output.

[0055] (3) Deployment method The content is a physical field model, and the engineering implementation principle is based on a simplified thermodynamic logistics model of finite difference / FEM, which is deployed at the edge. The core components are residual LSTM / Transformer, implemented using Python + PyTorch, and deployed using an ONNX model. The fusion unit handles fusion weight calculation and inference, implemented using a lightweight C++ module integration that supports GPU / CPU switching. The scheduling system interface provides state inference output, implemented using a REST interface to push control suggestions or anomaly alarms to the MES / PLC. All modules interact via shared memory / Kafka / OPC-UA / REST interfaces.

[0056] The data flow is as follows: (1) The MES / SCADA acquisition system provides the state vector; (2) The state is sent to the PDE solver to generate physical predictions; (3) The residual network is synchronously input to generate residual predictions; (4) The fusion prediction module completes the final prediction and uncertainty assessment; (5) The output is sent to the scheduler, safety assessment module, or MES controller. In scenarios with severe disturbances or large modeling deviations (such as tank surface cooling failure), the data model plays a dominant role; under normal and stable operating conditions, the mechanism model is the main one, improving interpretability and traceability; compared with the single model prediction deviation, the prediction deviation is reduced by more than 40% (simulation verification).

[0057] (4) Module interface definition and extension capability The input interface includes MES (State Variables). It includes SCADA (temperature, pressure, stress), model configuration (network structure, weight path, residual window parameters), and output interfaces such as OPC-UA / REST (thermal risk indicators, structural early warning indicators), internal buffer (for use by the scheduler optimizer), and log recording structure (fusion weights, confidence bands, runtime performance). It supports incorporating extended variables (such as AGV position, stirring viscosity, power fluctuations) into the predictor factor combination.

[0058] (5) Dynamic Adaptability and Mode Switching Strategies Introducing fusion control factors The definition is as follows: in: Indicates the physical model at time 10:00 The estimated value of the prediction error; This indicates that the data-driven model is at time [time]. The estimated value of the prediction error; Represent it as a very small positive number (e.g., 10). -6 Or 10 -8 (), used to prevent the denominator from being zero and to ensure numerical stability.

[0059] when When the data model dominates (due to severe disturbances or large modeling biases); when At this time, the mechanism model dominates (steady-state condition); in actual engineering, threshold switching or S-shaped function can be introduced to smooth the transition and prevent oscillation.

[0060] This mechanism supports abnormal state perception, mode switching management (physical priority / data priority), stable operation under low-frequency conditions, and sensitive response under high-frequency disturbances.

[0061] The effects are as follows: (1) Improved model prediction accuracy: thermal field prediction error RMSE < ±0.7 ℃ , stress prediction error <7%, under typical disturbance conditions (such as screw thermal asymmetry / mixing unevenness / material flow conflict), the fusion model reduces the error by about 40% compared with the single model; (2) enhance system resilience: can maintain prediction stability when equipment state changes suddenly or sensing data fluctuates; (3) support edge deployment: single-step fusion inference time <40ms, total control response delay <150ms, meets the real-time prediction update requirements of 200ms level, working condition switching fusion response time <1s, steady state / disturbance switching residual suppression rate is improved by 30%-40%; (4) strong scalability: has the ability to incorporate other process variables (such as bubble content, pressure pulsation) into dual-mode modeling.

[0062] (5) Scheduling optimizer / alarm PDE solutions and physical modeling can be performed on cloud-based modeling platforms such as COMSOL / Ansys, and simplified models can be exported; lightweight numerical models can be deployed on edge nodes (Jetson Xavier / industrial control computers); and data models can be deployed online for learning within TensorRT or ONNX frameworks.

[0063] The fused prediction results are fed into the system control center, triggering the following modules: Scheduling optimizer: Using predicted states (equipment temperature rise, AGV congestion, material uniformity) as state space input, it drives a reinforcement learning-based multi-objective scheduler (such as DDPG / TD3) to dynamically optimize path planning, transfer priority and task allocation; Alarm module: Embeds the predicted output into a fuzzy inference / causal graph network to achieve early identification of high-risk events (such as thermal runaway, mixed anomalies) and link response mechanisms (automatic parking, path reconstruction, voice alarm, etc.). All control strategies are integrated with MES, RCS, or PLC systems via REST API or OPC-UA interface to achieve unified hardware and software linkage.

[0064] This process, through an architecture of "physical model + data model + dynamic fusion + edge-cloud collaboration," effectively solves the problems of poor adaptability, high latency, and lack of interpretability in traditional single-model control systems under unsteady conditions. It possesses high-precision prediction, highly robust decision-making, and low-latency response capabilities. It supports deployment on integrated production lines for solid propellant mixing, transfer, and loading, and also has portability and universality for expansion to other coupled manufacturing processes.

[0065] 1.3 Scheduling Optimization Module The scheduling optimization module communicates with the intelligent twin module and is configured to receive system state reconstruction values ​​and future state prediction values, and generate and output path planning instructions and production task scheduling instructions based on a multi-objective deep reinforcement learning policy network.

[0066] The scheduling optimization module employs a multi-objective reinforcement learning policy network based on DDPG / TD3, combining high-dimensional inputs such as AGV status, equipment thermal status, and material in-process status to learn the optimal policy. With minimum task delay, optimal path selection, and equipment thermal load balancing as multi-objective optimization directions, it constructs a state-action-reward triplet model to achieve adaptive policy evolution in dynamic and complex environments. This module improves policy convergence and stability by constructing an Actor-Critic structure and using experience replay and a dual-network update mechanism. During the training phase, tens of thousands of policy steps are learned based on simulated workstation diagrams. During the deployment phase, the trained policy network is converted to ONNX format, accelerated by TensorRT, and integrated into industrial PCs or edge computing devices. Scheduling results are sent to the AGV and MES systems in real time via a REST API interface, achieving seamless linkage with the actual production line. This significantly improves transfer efficiency, reduces AGV congestion, lowers equipment thermal load, and achieves dynamic scheduling optimization of the entire process.

[0067] 1.3.1 Modeling Objectives The scheduling task of a continuous loading production line involves multiple key sub-objectives (such as minimizing task completion time, energy consumption control, and equipment thermal load balance), and presents the following challenges: (1) Complex coupling of objective functions: task efficiency and safety indicators are mutually constrained; (2) High-dimensional and dynamic state space: containing multi-source information such as AGV position, path congestion, equipment thermal status, and material status; (3) Frequent operating condition disturbances: frequent occurrences of task batches, temperature fluctuations, cooling failures, and equipment switching; (4) Traditional rule-based scheduling is difficult to generalize: it cannot adapt to multi-objective strategies. Therefore, this embodiment introduces deep reinforcement learning methods based on policy optimization (such as DDPG, TD3, etc.) to construct a trainable, deployable, and updatable autonomous scheduler to achieve intelligent production scheduling.

[0068] 1.3.2 Reinforcement Learning Strategy Network In this embodiment, the scheduling optimization module employs a reinforcement learning policy network based on the DDPG or TD3 algorithm, wherein: The state space S is defined as follows: in: Indicates the first The operating status vector of each workstation or equipment unit, including features such as temperature, torque, energy consumption, and process progress; Action space A is defined as: The reward function is defined as: in: This indicates the task completion time or total production cycle time, used to measure overall production efficiency. Indicates the system in time Energy consumption indicators, including energy consumption of motors, heaters, AGVs, etc.; Indicates system operational risk indicators, where local stress Exceeding the stress safety threshold The penalty item is triggered at any time, reflecting the risk of exceeding the limit; This indicates a penalty for frequent changes in process parameters; Indicates localized stress; Indicates the stress safety threshold; , , and These represent the production efficiency weighting coefficient, energy consumption weighting coefficient, safety and risk weighting coefficient, and switching frequency penalty weighting coefficient, respectively.

[0069] (1) Logistics field modeling: AGV material transfer status and scheduling modeling The logistics problem is modeled as a directed graph path selection problem within a discrete system. Model specification: Nodes represent workstations, edges represent feasible AGV paths, and state variables are defined. For the first Each AGV in time The location. Define the AGV queue scheduling function: in: Thermal state of the target workstation equipment; Current loading status; The transfer speed (including acceleration model).

[0070] The cost function can be defined as: in: Indicates the number of AGVs; Indicates the total transit time; , and These represent the weighting coefficients for path length, dwell time, and overheat zone penalty, respectively.

[0071] Based on this model, heat-sensitive path avoidance scheduling can be designed. All path calculations can be transformed into graph theory shortest path problems or reinforcement learning path strategies, which can be coupled with temperature and stress fields to form a feedback control loop.

[0072] (2) Policy learning algorithm (taking DDPG as an example) Policy Network (Actor): Value Network (Critic): Update target: in: The action value function; State; For action; As a reward; Discount factor; For policy gradient; For mathematical expectation; For the action gradient; For policy network gradients; The state is...

[0073] The convergence guarantee conditions are: (1) using experience replay to prevent correlation; (2) using a dual network (target network) to prevent oscillation; and (3) using Prioritized Replay to improve learning efficiency.

[0074] The system integration includes: (1) a status reading module, which implements the strategy of real-time variable extraction from MES / SCADA / twin platforms; (2) a strategy engine module, which implements the strategy of PyTorch+ONNX model and TensorRT inference acceleration; (3) an action delivery module, which uses REST API or PLC instruction interface; and (4) a multi-objective adjustment module, where weights can be automatically adjusted according to the scenario.

[0075] 1.3.3 Generalization and Self-Evolution Mechanism of Multi-Objective Reinforcement Scheduling Strategies (1) Hierarchical reinforcement learning (HRL) This embodiment divides the scheduling strategy into: 1) a high-level task-level actor to determine task priorities and AGV device assignment; 2) a low-level route-level actor to dynamically avoid congestion and hot spots, achieving the shortest path and minimum energy consumption; and 3) a Critic supervisor to jointly evaluate the comprehensive benefit function of multiple objectives (timeliness, energy consumption, and equipment thermal load). This hierarchical structure supports modular training of the strategy and has strong scenario transfer capabilities.

[0076] The HRL framework is adopted to split the global scheduling task into two-layer policy modules, which have pluggable training and cross-scenario transfer capabilities. Specifically, in this embodiment, the scheduling optimization module adopts a hierarchical reinforcement learning structure, including the following:

[0077] 1) The input to the high-level task-level actor is the task pool. Equipment pool and order priority Output task - device matching action : in: Indicates the task layer state, indicating the task pool. Equipment pool and order priority A comprehensive state vector containing information such as [list of information] is used to describe the current task allocation environment of the system; This represents a high-level policy network, which considers the input state. Make decisions and output the matching and allocation actions between tasks and AGVs; that is, determine which AGV will perform which task. Indicates the first task in the task pool Tasks include material transportation, mixing tank feeding, and finished product transfer. Indicates the first An automated guided vehicle (AGV) is responsible for moving and performing the assigned tasks. The goal of high-level task strategy is to maximize task completion rate and priority completion.

[0078] The low-level route actor takes the AGV's current position, environmental obstacles, heat map, and energy consumption estimate as input, and outputs a route decision. : in: It represents the path layer status, including the AGV's current position, environmental obstacle map, hot zone distribution map, and energy consumption estimation information, which is used for path planning and obstacle avoidance decisions. This represents a low-level policy network that, given assigned tasks, outputs the optimal path or control action to enable autonomous navigation of the AGV. Indicates the first path in the candidate path set The features or scores of each path are used to select reinforcement learning strategies.

[0079] The goal of the low-level path strategy is to achieve the shortest path, the minimum energy consumption, and avoid thermal coupling and overlap.

[0080] The Critic supervisor (multi-objective integrated evaluation) is used to jointly evaluate the comprehensive benefit function: in: Incentives for efficient dispatching; Energy consumption penalty per unit task; Penalty for the number of times a player crosses a hot zone; , and These are adaptively adjustable target preference weights.

[0081] (2) Fast transfer meta-learning mechanism (Model-agnostic meta-learning, MAML) When batch formulations, equipment configurations, or topologies change, the scheduling strategy can be transferred using a few samples based on the MAML training framework. It only requires 50 rounds of fine-tuning to adapt to new scenarios, significantly reducing retraining time.

[0082] To address the rapid changes in production line structure, process batches, and equipment thermal load patterns, a meta-learning algorithm (MAML) is introduced to achieve "rapid adaptation with few samples" for the scheduling strategy. In this embodiment, the scheduling optimization module integrates a fast transfer meta-learning mechanism and is configured as follows: In multiple source tasks Learning initial policy parameters : When a new task arrives, only a small number of samples (e.g., less than 50 rounds of fine-tuning) are needed for rapid adaptation: in: This indicates that the process involves a meta-task (i.e., multiple source tasks). The optimal initial policy parameters obtained during training represent the model's general prior knowledge across multiple tasks, enabling it to converge quickly in new tasks. Indicates the first Individual source task The loss function, which typically corresponds to negative reward or prediction error in reinforcement learning, is used to measure the performance of the policy on the task. This represents the learning rate or step size coefficient, which controls the magnitude of gradient updates per step. Indicates the parameter The gradient operator represents the direction and magnitude of the gradient of the model parameters with respect to the loss function.

[0083] 1.4 Multi-layered secure causal reasoning module The multi-layer safety causal reasoning module is connected to the intelligent twin module and the scheduling optimization module respectively. It is configured to predict the risks of excessive temperature rise, uneven slurry or equipment failure based on causal graph reasoning, fuzzy logic judgment and linkage response mechanism, and output corresponding early warning signals, and / or directly output control commands to the scheduling optimization module or the control module.

[0084] The multi-layered safety causal reasoning module proposes a three-layer structure (causal graph → fuzzy reasoning → linkage response) to predict and respond to risks such as rapid temperature rise, uneven slurry distribution, and equipment failure. The first layer, based on Bayesian networks and dynamic causal graphs, models the causal relationships and triggering links between process variables to capture potential risk sources. The second layer introduces fuzzy membership functions and semantic rules to determine the risk level under combined conditions of multiple source variables (such as temperature gradient, torque fluctuation, start-up / shutdown frequency, etc.). The third layer maps the risk level to response strategies in the scheduling or control system, automatically triggering path reconstruction, cooling start-up, stirring rate adjustment, or alarm linkage mechanisms. This module can run independently on edge nodes, supporting an inference frequency of ≥1Hz to ensure real-time response to highly dynamic anomalies. By embedding safety logic into the scheduling reward function, a closed-loop control of "prediction + avoidance + response" is achieved, effectively improving the system's fault warning capability and operational robustness.

[0085] 1.4.1 Modeling Objectives Continuous charging systems are inherently dangerous, posing risks such as material deterioration due to rapid temperature rise and AGV collisions. Existing systems are mostly based on limit alarms, lacking predictability and interpretability. A multi-layered safety causal reasoning structure is proposed: (1) The causal graph reasoning layer is mainly used for physical causal graph reasoning; (2) The fuzzy logic judgment layer is mainly used for fuzzy logic judgment; (3) The linkage response layer is mainly used for security policy linkage response.

[0086] In this embodiment, the multi-layered secure causal reasoning module includes a causal graph reasoning layer, a fuzzy logic judgment layer, and a linkage response layer.

[0087] 1.4.2 Hierarchical Modeling and Derivation (1) Causal Graph Reasoning Layer The causal graph inference layer constructs a Bayesian network based on variable causal graphs to model the causal relationships between process variables. The variable causal graph is represented as follows: in: Indicates ambient temperature; Indicates the temperature of the equipment wall surface; Indicates the viscosity of the slurry; Indicates mixed torque; This indicates the stress on the stirring shaft.

[0088] In a variable causal graph, each edge is assigned a causal strength. This forms a Bayesian network: Reasoning methods include (1) Bayesian reasoning (static); (2) Dynamic causal graph (DAG-TCN) learning (if the amount of data is sufficient).

[0089] (2) Fuzzy logic reasoning layer The fuzzy logic inference layer stores several semantic rules in the form of "IF-THEN". It then uses Gaussian or trapezoidal membership functions to fuzzify the input variables into levels such as "high", "medium", and "low", and calculates the security risk level based on these fuzzy inference rules. The semantic rules in the form of "IF-THEN" are expressed as follows: If temperature rise rate > High AND twin-screw shaft torque > Large, then the hazard level is High. Mathematically, it is expressed as: in: The membership function representing the high-risk level indicates the confidence or credibility of the system being in a high-risk state; A fuzzy membership function representing the rate of temperature rise, used to reflect whether the rate of temperature rise exceeds a safe threshold; A fuzzy membership function representing the stirring torque is used to reflect the mechanical load level during the stirring process.

[0090] Fuzzy membership functions are defined using Gaussian or trapezoidal functions, allowing for flexible adjustment.

[0091] (3) Linkage Response Layer The linkage response layer is configured to map safety risk levels to predetermined response strategies. High risk levels trigger the cessation of mixing and AGV path reconstruction, medium risk levels trigger the reduction of mixing speed and the increase of cooling, and low risk levels trigger the logging and the increase of monitoring frequency.

[0092] 1.4.3 System Implantation Strategy This embodiment features excellent system integration and real-time application capabilities: the causal graph is stored on an edge server or local PLC, enabling low-latency access and local decision support; fuzzy inference rules are stored in the form of a structured knowledge rule set, facilitating subsequent maintenance and expansion; the inference results are integrated into a digital twin dashboard in a visual manner, allowing maintenance personnel to view system status and risk levels in real time; simultaneously, the risk scoring results are input as constraints to the reinforcement learning scheduler, influencing its task allocation and path planning strategies, thereby achieving closed-loop linkage control between risk perception and task scheduling.

[0093] 4.4.4 Temporal and Causal Reinforcement Modeling of Secure Reasoning Mechanisms (1) Temporal logic fusion (LTL-embedded fuzzy rule system) By embedding LTL into a fuzzy rule system (LTL-FRS) and introducing fuzzy temporal rules, the following type of anomaly chain is modeled: Linear temporal logic rules are embedded in the fuzzy logic inference layer to model time-coupled anomaly patterns. The rule form is as follows: in: Let be the membership function; It represents the rate of temperature rise, i.e., the derivative of temperature with time, and is used to characterize the intensity of the thermal dynamic changes of the system. This indicates the change in torque, reflecting the degree of load fluctuation on the stirring shaft within a short period of time. It is usually related to sudden changes in material viscosity or equipment jamming. The membership function represents the risk level, and the fuzzy confidence level represents the current high-risk state of the system. Indicates the future The linear-time logic (LTL) operator is "finally satisfied" within seconds.

[0094] (2) Automatic generation of dynamic causal graph (Causal discovery augmented graph) The variable causal graph is automatically constructed by mining the potential causal path structure between sensor variables through Granger causality testing and the PCMCI (PCMCI+) algorithm, reducing reliance on manual intervention and adapting to dynamic changes in the system. Simultaneously, a time-series causal discovery algorithm is introduced to construct a data-driven dynamic graph structure (Causal DAG). The core algorithm principle is as follows: If added Significant improvement The predictive ability is considered to be... .

[0095] in: This indicates the current state of the explained variable (target variable). The value of is the current response or observation of the system; Represents the explained variable itself The lagged term represents the past. Historical values ​​within a time step are used to capture its autoregressive characteristics; This indicates that exogenous variables (potential influencing factors) in The value of the first lag time reflects the time influence of external variables on the target variable; Represents random disturbance terms or white noise, used to represent unmodeled noise or random effects; The lag order represents the number of historical steps considered, and is usually determined by information criteria (such as AIC, BIC). and These represent the autoregressive coefficient and the cross-regression coefficient, respectively, describing... Self-history and external variables For the current value The intensity of the impact; The PCMCI+ (Peter-Clark Momentary Conditional Independence) algorithm is suitable for high-dimensional, multi-lag data. It supports pruning (PC stage) + residual analysis (MCI stage) and can effectively reduce the number of spurious causal edges to form a dynamically updated directed acyclic graph, where each edge represents a potential causal chain. (Time lag at the edge of causality) in: and They represent the first The and the first A time series variable, representing the target quantity that is affected or predicted; Indicates the time lag, representing the amount of time delay in the causal effect, reflecting... right The lag effect.

[0096] II. Experimental Verification To verify the effectiveness of the intelligent control strategy and framework for the continuous loading production line proposed in this invention, a hybrid assembly / testing production line simulation platform with real production environment characteristics was constructed, and a comparison group of "traditional scheduling system" and "optimized but not self-evolving scheme A" was introduced as a benchmark.

[0097] 2.1 Experimental Platform Configuration This embodiment constructs a verification platform integrating simulation-driven, data-driven, and physical modeling, employing multiple rounds of experiments to compare the impact of different strategies on system performance. All software is built within a unified virtual simulation platform to ensure experimental consistency. Module communication and closed-loop control between physical modeling, scheduling strategies, state awareness, and safety inference are achieved through a Python scheduling hub.

[0098] This embodiment employs a multi-agent scheduling system, integrating a dual-mode confidence fusion prediction module, a hierarchical reinforcement learning (HRL) scheduling strategy, and a causal enhanced security reasoning mechanism to form an intelligent control system with dynamic perception, intelligent decision-making, and adaptive linkage capabilities. The experiments covered various typical operating conditions, including stable equipment operation, frequent abnormal disturbances, and complex environments such as production line structural topology changes, comprehensively verifying the system's robustness and adaptability. Evaluation metrics covered key performance dimensions such as order delivery timeliness, unit energy consumption reduction rate, task conflict rate, abnormal response time, and system stable runtime, ensuring the effectiveness and universality of the proposed method in real-world complex manufacturing environments.

[0099] To ensure reproducibility and engineering feasibility, this embodiment establishes a complete hardware and software experimental and simulation environment. At the simulation level, production process modeling is completed using AnyLogic 8.8.3 Professional Edition and MATLAB Simulink 2022a in collaboration; multiphysics thermal-fluid-structure interaction simulation is implemented using COMSOL Multiphysics 6.0; the data analysis and model training modules are built using Python 3.10, integrating commonly used machine learning and deep learning tools such as Pandas, Scikit-learn, and PyTorch 2.1; system interaction uses an OPC-UA interface simulator and MQTT Broker to build the device communication protocol; the scheduling controller uses a self-developed reinforcement learning engine, implemented based on OpenAI Gym and Stable-Baselines3, supporting HRL policies and meta-learning transfer mechanisms. The deployment platform is an Ubuntu 20.04 system, equipped with a 16-core Intel Xeon Silver 4310 processor and 128GB of memory on an edge server. The experimental design covers three types of order recipes, three combinations of production line topologies, and two typical equipment fault disturbance injection scenarios to comprehensively verify the robustness, adaptability, and generalization ability of the system.

[0100] 2.2 Implementation Case 1: Thermal-Fluid-Solid Three-Field Modeling and State Control System Deployment like Figure 2As shown, the system in this embodiment is deployed on a continuous charge mixing and casting production line, using 12 temperature sensors, 4 strain gauge sensors, and AGV position encoders to collect thermal field, structural field, and material flow status data. A PDE model is constructed using the COMSOL platform and exported to an edge industrial control computer for deployment, where a simplified finite difference method is used for real-time solution. An LSTM residual model is deployed using Jetson Xavier, with ONNX used to accelerate inference; the system sampling period is 200 ms. After state fusion, the thermal risk index and the upper limit of axial stress are synchronously transmitted to the SCADA system via the OPC-UA interface, achieving dual-channel control between the edge and the center.

[0101] 2.2.1 Implementation Objectives Verify the accuracy and responsiveness of thermal-fluid-solid modeling and state control systems.

[0102] 2.2.2 Implementation Process 1) A hybrid cavity geometry model was created using COMSOL (dimensions: L×W×H = 650mm × 320mm × 200mm); the flow field adopted a non-Newtonian shear thinning model (Carreau-Yasuda); the heat source term adopted screw shear power + cavity wall heat source; the solid domain material adopted Alloy 6061, and the fluid domain material: polymer composite fluid; 2) The mesh type is a hexahedral master mesh with boundary layer body refinement; the initial 150,000 elements are adaptively refined to 450,000; the time step is 0.1s, and the solution time is 300s; the convergence criterion is residual <1×10⁻⁶. -6 ; 3) The residual LSTM module (3 layers × 128 nodes) is used to predict the thermal field and the status is fed back to the Simulink controller to realize the fan power adjustment. The control response delay is controlled within 0.4s, and the spindle shutdown logic is automatically triggered if the temperature rises too fast.

[0103] 4) Twelve K-type thermocouple sensors are axially arranged on the outer wall of the twin-screw mixing cylinder; four resistance strain gauges are arranged on the key sections (heat / stress concentration) of the twin-screw mixing shaft; the AGV system integrates a laser SLAM module or encoder to output position information; all sensors are connected to the SCADA acquisition system through a Modbus to OPC data gateway.

[0104] 5) The heat conduction equation is established based on the one-dimensional unsteady PDE form: the bending-torsional coupled stress model considers the curvature change caused by the temperature gradient; directed graph material flow modeling. The AGV status is mapped to path status variables.

[0105] 6) Output a simplified version of the model through the COMSOL solver and export the FMU (Functional Model Unit) or Python module; deploy the thermo-coupling solver module + LSTM residual model on Jetson Xavier; use ONNX Runtime to accelerate inference, triggering state calculations every 200ms; feed the prediction results back to the SCADA platform through the OPC-UA interface to generate thermal risk index and structural stress early warning indicators.

[0106] 7) The merged state is used to trigger cooling control commands; update the scheduler to evaluate equipment priority; and determine whether the AGV path needs to avoid high-heat equipment.

[0107] 2.2.3 Deployment Parameters To achieve high-precision sensing and modeling control of the continuous loading production line's operational status, this system deploys various types of sensors and a high-efficiency edge computing platform. Twelve temperature sensors are deployed on the outer wall of the mixing cylinder and at multiple key axial measuring points inside to capture dynamic heat distribution; four strain gauge sensors are installed at key cross-sections of the mixing axis to monitor structural load changes; and a high-precision laser-coded positioning module is used for AGV positioning to ensure accurate scheduling paths and task allocation. For physical modeling, the COMSOL Multiphysics platform is used to complete the thermo-structural coupling modeling, and the residual data model is built using PyTorch and exported in ONNX format for efficient cross-platform deployment. Edge inference is deployed on the NVIDIA Jetson Xavier NX platform (16GB RAM), processing state data in real time with a sampling period of 200 ms. Numerical solution employs the simplified one-dimensional finite difference method (FDM) to achieve fast and stable temperature field prediction. Meanwhile, the system achieves interoperability with the SCADA platform based on the OPC-UA protocol, with a data synchronization cycle of 500 ms, ensuring the timeliness and reliability of the system's closed-loop control.

[0108] 2.2.4 Technical Specifications and Effects: The thermal prediction error (RMSE) was reduced to ±0.72°C; the maximum prediction error RMSE = 0.72℃; the match rate between the measured temperature control curve and the predictive control curve reached 96.3%; and the risk of overheating was reduced by 25.6% compared with the traditional lag-type control strategy. The prediction error of the maximum thermally induced bending stress of the hybrid shaft was less than 7%; the system state perception delay (from edge node to SCADA) was < 300 ms; the thermal risk index response time was < 2.5 s (including preprocessing); and the peak bandwidth requirement for three-field state synchronization was < 3 Mbps.

[0109] 2.3 Implementation Case 2: Integration of Reinforcement Learning Scheduler and Digital Twin like Figure 3 As shown, in AGV logistics path planning, a workstation graph model with 12 nodes and 38 edges is established. Task delay, AGV congestion, and equipment thermal state are used as state-space variables to train a scheduling strategy network based on TD3. The enhanced scheduler deployed on the industrial PC runs in TensorRT format, and control commands are sent to the RCS control system via REST API.

[0110] In this embodiment, the implementation process of the reinforcement learning scheduling system includes the following four steps: First, based on the AGV station layout of the continuous loading production line, a directed graph topology model containing 12 nodes and 38 edges is constructed, and a state space is established. The state variables cover the current position information of each AGV, the task queue and load status, the path congestion coefficient, and the thermal state of the equipment it serves (this thermal state is predicted by the thermal-fluid-solid model in Embodiment 1). The action space is defined as the selection strategy for AGV path nodes and task allocation. Second, the system uses the TD3 (Twin Delayed Deep Deterministic Policy Gradient) reinforcement learning algorithm for policy training, constructing a dual neural network model with an Actor-Critic structure. The system continuously runs for 100,000 steps in the constructed virtual simulation environment to train the task queue and path dynamics. The reward function is designed as follows: The system incorporates the equipment's thermal risk level as a penalty factor into the reward function. Third, the trained policy network is exported to ONNX format and deployed to an industrial PC terminal, where it is accelerated for inference using the TensorRT framework. During actual operation, the system schedules the task queue every 100 milliseconds, dynamically generating scheduling actions based on the current AGV status and environmental predictions. Control commands are sent to the RCS scheduling system via a REST API interface, enabling real-time path updates and task distribution. Finally, the system continuously collects AGV operation feedback data for policy update fine-tuning, path occupancy feedback to the thermally sensitive area avoidance mechanism, and records key scheduling status and operation logs in the SCADA system, forming a closed-loop optimized scheduling and thermal management linkage control system.

[0111] To verify the performance advantages of reinforcement learning scheduling system in multi-objective (energy consumption, congestion, efficiency) scheduling, an AnyLogic 8 model was built with five device nodes, three AGVs, and three workstations. The model training settings are as follows: (1) High-level policy network: Actor (256×128×64) + Critic; (2) Low-level policy network: Actor (128×64) + path constraint graph; (3) Reward function: R = αT efficiency + βE energy consumption + γC conflict rate; (4) Sampling frequency: 10 steps / episode; (5) Learning rate: 2×10 -4 Experience recovery size: 10 5 .

[0112] The scheduling algorithm is trained for 200 rounds, with a target of 600 orders and 3 work condition switching points; MAML training is used to migrate to the new order recipe structure, requiring only 50 rounds; the twin system implements the scheduling strategy through MQTT → synchronous execution of the digital twin simulation; the system status is fed back to the reinforcement learning environment as a state variable; an automatic readjustment strategy is implemented when the task failure rate is >5%.

[0113] In this embodiment, the AGV scheduling system is based on a constructed production line workstation diagram structure, containing 12 nodes and 38 directed edges, forming a complete path network topology. The state space has over 25 dimensions, covering the real-time location of the AGVs, task queuing status, path occupancy, and the thermal status information of the equipment they serve. The system uses the TD3 (Twin Delayed DeepDeterministic Policy Gradient) reinforcement learning algorithm for scheduling policy training, with a single round of simulation training taking approximately 15 minutes and policy convergence achieved within 100,000 steps. During the deployment phase, the trained policy network is exported in ONNX format and deployed on an industrial PC equipped with a quad-core CPU. It runs rapidly using the TensorRT inference engine, with each round of scheduling control having a 100-millisecond cycle. Scheduling commands communicate with the RCS control system via a REST API interface, ensuring the real-time performance and reliability of the scheduling process.

[0114] Comparative experiments verified that the scheduling system in this embodiment exhibits significant improvements in several key performance indicators: the average task completion time is reduced by 16.7% compared to traditional methods; the peak waiting time of AGVs is reduced by 27.4%; and under high-load and complex scenarios, the system's task completion rate within 60 seconds is significantly improved from 83.2% of the traditional method to 96.5%. Furthermore, the system's single-step inference control time is less than 20 milliseconds, and it can complete the rescheduling response within 300 milliseconds after an AGV failure, demonstrating excellent real-time performance and robustness. These results indicate that the reinforcement learning scheduling system proposed in this embodiment possesses good scalability and practical application value under complex production and high-concurrency conditions.

[0115] The table shows the significant advantages of the HRL scheduling system and the self-evolving strategy system after integrating the MAML fast migration mechanism proposed in this embodiment compared with the traditional baseline scheme under four key performance indicators: (1) In terms of average energy consumption, the scheme of this embodiment is significantly better than the traditional scheduling strategy. The HRL scheduler reduces the energy consumption per unit task by about 12.5%, while the self-evolving system after integrating MAML further reduces it to 987.2 kJ, with an overall energy consumption reduction of 20.4%; (2) In terms of congestion rate, it significantly decreases from 7.2% to 2.3%, effectively avoiding AGV path conflicts and task accumulation; (3) In terms of average completion time, the self-evolving scheduling system compresses the average task processing time from 29.4 seconds to 19.3 seconds, improving efficiency by more than 34%; (4) Most importantly, the MAML mechanism proposed in this embodiment only requires 50 rounds of fine-tuning training to complete the strategy migration under the new working conditions, which is far lower than the more than 200 rounds of retraining required by conventional reinforcement learning, greatly improving the system's adaptability and deployment efficiency. In summary, the data fully validates the comprehensive superiority of this embodiment in terms of energy consumption optimization, scheduling efficiency improvement, path congestion control, and system migration capabilities, demonstrating its strong industrial applicability.

[0116] 2.4 Implementation Case 3: Security Reasoning System and Early Warning Response like Figure 4 As shown, a dynamic Bayesian graph is constructed with temperature gradient, mixer torque, and equipment start-up frequency as causal variables, and a safety level judgment module is established in conjunction with fuzzy membership functions. The safety risk level is embedded into the scheduler reward function to achieve risk-driven scheduling optimization.

[0117] In a simulated AGV cooling failure test, the system issued a medium-risk warning in advance and triggered automatic transfer path reconstruction, thus avoiding the risk of thermal runaway.

[0118] In the implementation of the safety inference system in this embodiment, multi-source data acquisition and causal modeling are first performed. The core variables collected by the system include: the rate of temperature rise of the equipment (calculated by dividing the sensor difference by the sampling period), the instantaneous fluctuation value of the hybrid shaft torque, and the frequency of equipment start-up and shutdown per unit time (e.g., the number of starts per hour). Based on the above multi-dimensional data, a Bayesian causal network is constructed to complete causal structure learning and parameter training. In this network, nodes represent state variables, event nodes, and risk levels, while edges quantify the causal strength between variables in the form of a conditional probability table (CPT), clarifying the risk mode triggering mechanism, such as the combined path of "temperature rise + severe torque fluctuation → high risk".

[0119] To enhance model interpretability and human-machine collaboration efficiency, the system introduces 18 fuzzy rules to construct a safety inference module, and uses Gaussian membership functions to continuously represent the "high / medium / low" fuzzy intervals. For example: "If the temperature rise rate is high and the torque oscillation amplitude is large, then the risk level is high". The final risk level output is divided into three response levels: a high-risk state will trigger immediate停车 and AGV path reconstruction, a medium-risk triggers a mixed rate reduction operation, while a low-risk only records the log and appropriately increases the monitoring frequency.

[0120] The safety inference module runs at a frequency of 1 Hz, and the inference results can be sent to the MES system in real time through the REST API interface. At the same time, the risk level signal is also input as a constraint condition into the reinforcement learning scheduling system in Embodiment 2 to achieve multi-module linkage control among the hot state, risk, and task scheduling. The system is deployed in industrial edge computing devices and can share hardware resources with other scheduling and control modules to achieve cost optimization and system integration.

[0121] To evaluate the safety event prediction ability of the causal enhanced inference system under dynamic working conditions. Use Simulink to generate 15-dimensional time series signals such as temperature, current, and torque, with a period of 20 ms; inject abnormal temperature rise (2℃ / min) + current jitter; the fault events are abnormal spindle跳动 and speed fluctuation; the inference algorithm is fuzzy rules + LTL embedding, and 50 rule trees are constructed; PCMCI + dynamic causal graph generates causal edges, and the maximum lag step is 10; Granger test significance p<0.01 is used to screen channels; It should be noted that the Chinese character "停车" in the original text seems to be an incomplete expression. I translated it as "停车" for now, but it may need to be adjusted according to the actual situation. Also, the character "跳动" might be more accurately translated as "vibration" or other more appropriate terms depending on the context.The alarm logic triggers a red light and AGV shutdown once the "time-sequence combination trigger" rule is met; the delay time is recorded in real time. This table compares the performance of the safety inference system in this embodiment with that of the non-causal inference reference group under key early warning performance indicators, verifying the significant advantages of this embodiment in time-sequence causal enhanced inference: In terms of average alarm lead time, the system in this embodiment can achieve an early warning of 14.3 seconds, far exceeding the 4.7 seconds of the non-causal inference group, indicating that it can identify potential risks earlier, providing more time for system response and manual intervention; in terms of false positive rate, this embodiment significantly reduces false alarms to only 2.3% by integrating LTL time-sequence rules and automated causal graph mining, a 60.3% decrease compared to the non-causal system (5.8%), improving the reliability of alarms; in terms of false negative rate, after adopting causal enhancement mechanisms such as Granger-PCMCI, the alarm false negative rate is only 1.1%, significantly better than the 8.6% of the traditional system, greatly improving the coverage of high-risk events and the level of system security. In summary, the safety reasoning system in this embodiment is superior to existing technologies in terms of being "earlier, more accurate, and more comprehensive," verifying the important value of introducing a causal temporal reasoning mechanism for risk prediction and response control in complex production lines.

[0122] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. An intelligent control strategy and framework for a continuous drug loading production line, comprising a data acquisition system for collecting sensor data from the production line and a control module for executing control commands, characterized in that: The intelligent control system also includes: The multiphysics coupling modeling module is configured to establish a system state model describing the thermo-solid-fluid coupling mechanism of the slurry during the mixing process by coupling the control equations of the thermal field, structural field and fluid field, and output the predicted values ​​of key physical states including temperature distribution, structural stress and fluid viscosity. The intelligent twin module is communicatively connected to the multiphysics coupling modeling module and is configured to fuse the mechanistic model prediction results from the multiphysics coupling modeling module with a data-driven residual model prediction result, and output high-precision system state reconstruction values ​​and future state prediction values ​​through a dynamic weighted fusion mechanism. The scheduling optimization module, which is communicatively connected to the intelligent twin module, is configured to receive the system state reconstruction value and the future state prediction value, and generate and output path planning instructions and production task scheduling instructions based on a multi-objective deep reinforcement learning policy network. The multi-layer safety causal reasoning module is communicatively connected to the intelligent twin module and the scheduling optimization module, respectively. It is configured to predict the risks of excessively rapid temperature rise, uneven slurry distribution, or equipment failure based on causal graph reasoning, fuzzy logic judgment, and linkage response mechanism, and output corresponding early warning signals, and / or directly output control commands to the scheduling optimization module or the control module.

2. The intelligent control strategy and framework for the continuous loading production line according to claim 1, characterized in that: The multiphysics coupling modeling module includes: The thermal field modeling unit is configured to establish a temperature field model of the slurry within the twin-screw extruder based on the partial differential equation of heat conduction. in: Density of the medicated paste; The specific heat capacity of the slurry; For the medicine in position and time Temperature; Thermal conductivity; The volumetric heat source term is modeled as a stirring rate function: , For mixed shaft torque, For rotational speed, The energy conversion efficiency coefficient of stirring; The structural field modeling unit is configured to establish a total stress model for a twin-screw hybrid shaft based on thermo-structural coupled elasticity: in: The total stress; The stress is caused by thermal bending moment, and , For elastic modulus, For thermal expansion curvature; The torque applied to the hybrid shaft; The radius of the axial section; Let be the moment of inertia of the axial cross section; It is the polar moment of inertia; The fluid field modeling unit is configured to establish a slurry flow model based on the generalized Navier-Stokes equations for non-Newtonian fluids: in: It is a velocity vector; For pressure; Density; Effective viscosity depends on shear rate. and temperature ; It is a volume force; The effective viscosity is fitted using either the Carreau model or the Cross model. in: Zero shear viscosity; Viscosity at infinite shear rate; It is a time constant; The liquidity index; The state equation of the system state model is: in: This indicates the internal temperature field, stress field, and heat flow state of the system. This represents the system input control vector, and also represents an external controllable or measurable signal. This represents a system noise or uncertain disturbance vector, used to characterize dynamics not modeled by the model, sensor noise, or external interference. It represents the dynamic matrix or operator of the system, reflecting the coupling relationship and dynamic characteristics of the system, and can be a linear matrix or a nonlinear operator; Indicates the system temperature status; This indicates the thermal stress or thermal strain state of the system. This indicates the heat flux density or heat source intensity within the system. Indicates rotational speed; This indicates the control commands or speed input for the self-guided vehicle. This indicates the ambient temperature or external thermal boundary conditions.

3. The intelligent control strategy and framework for the continuous loading production line according to claim 1, characterized in that: The multiphysics coupling modeling module adopts an adaptive mesh refinement strategy based on error indicators. For regions with drastic changes in thermal gradient, local mesh refinement is performed, including: For state variables; To refine the threshold; , and These represent the temperature field, shear stress, and structural stress in the system state model, respectively. The Sobol global sensitivity analysis method is used to evaluate the impact of input parameters on output state variables. The first-order Sobol sensitivity exponent is defined as: in: For the first The input variable for the first... The percentage of variance explained by each response; For the first One output target state vector; For the first A vector of input physical parameters.

4. The intelligent control strategy and framework for the continuous loading production line according to claim 1, characterized in that: The intelligent twin module includes a mechanism model, a data model, and a dynamic fusion mechanism; The mechanism model It is configured to run a simplified physical model derived from the multiphysics coupling modeling module; The data model It is configured to run a trained LSTM or Transformer neural network to learn and predict the residuals between the mechanistic model predictions and the actual system state; The dynamic fusion mechanism is configured to output based on the mechanism model. and the output of the data model The fused prediction value is calculated using a Bayesian dynamic weighting mechanism: in: For the dynamic fusion output of the intelligent twin module; This is the output of the mechanistic model; The output of the data model; The fusion weights are calculated based on the mean square error of the sliding residuals: in: The error variance of the mechanistic model; This represents the error variance of the data model.

5. The intelligent control strategy and framework for the continuous loading production line according to claim 4, characterized in that: The fusion prediction controller is further configured to: Define a working condition adaptability factor : in: The disturbance intensity; The disturbance handover threshold; Sensitivity factor; Using the aforementioned working condition adaptability factor Dynamically adjust the output of the fusion prediction so that when When the value is close to 1, the data model dominates the fusion prediction. When the value is close to 0, the mechanistic model dominates the fusion prediction.

6. The intelligent control strategy and framework for the continuous loading production line according to claim 1, characterized in that: The scheduling optimization module employs a reinforcement learning policy network based on the DDPG or TD3 algorithm, wherein: The state space S is defined as follows: in: Indicates the first The operating status vector of each workstation or equipment unit; Action space A is defined as: The reward function is defined as: in: This indicates the task completion time or total production cycle time, used to measure overall production efficiency. Indicates the system in time Energy consumption indicators, including energy consumption of motors, heaters, AGVs, etc.; Indicates system operational risk indicators, where local stress Exceeding the stress safety threshold The penalty item is triggered at any time, reflecting the risk of exceeding the limit; This indicates a penalty for frequent changes in process parameters; Indicates localized stress; Indicates the stress safety threshold; , , and These represent the production efficiency weighting coefficient, energy consumption weighting coefficient, safety and risk weighting coefficient, and switching frequency penalty weighting coefficient, respectively.

7. The intelligent control strategy and framework for the continuous loading production line according to claim 6, characterized in that: The scheduling optimization module adopts a hierarchical reinforcement learning structure, including: High-level task strategy: Input is the task pool Equipment pool and order priority Output task - device matching action : in: Indicates the task layer state, indicating the task pool. Equipment pool and order priority A comprehensive state vector containing information such as [list of information] is used to describe the current task allocation environment of the system; This represents a high-level policy network, which considers the input state. Make decisions and output the matching and allocation actions between tasks and AGVs; that is, determine which AGV will perform which task. Indicates the first task in the task pool Tasks include material transportation, mixing tank feeding, and finished product transfer. Indicates the first An automated guided vehicle (AGV) is responsible for moving and performing the assigned tasks. Low-level path strategy: Inputs include the AGV's current position, environmental obstacles, heat map, and energy consumption estimate; output is the path decision. : in: It represents the path layer status, including the AGV's current position, environmental obstacle map, hot zone distribution map, and energy consumption estimation information, which is used for path planning and obstacle avoidance decisions. This represents a low-level policy network that, given assigned tasks, outputs the optimal path or control action to enable autonomous navigation of the AGV. Indicates the first path in the candidate path set The features or scores of each path are used to select reinforcement learning strategies; The Critic supervisor is used to jointly evaluate the comprehensive return function. in: Incentives for efficient dispatching; Energy consumption penalty per unit task; Penalty for the number of times a player crosses a hot zone; , and These are adaptively adjustable target preference weights.

8. The intelligent control strategy and framework for the continuous loading production line according to claim 6 or 7, characterized in that: The scheduling optimization module integrates a fast transfer meta-learning mechanism and is configured as follows: In multiple source tasks Learning initial policy parameters : When a new task arrives, you can adapt quickly: in: This indicates that the process involves a meta-task (i.e., multiple source tasks). The optimal initial policy parameters obtained during training represent the model's general prior knowledge across multiple tasks, enabling it to converge quickly in new tasks. Indicates the first Individual source task The loss function, which typically corresponds to negative reward or prediction error in reinforcement learning, is used to measure the performance of the policy on the task. This represents the learning rate or step size coefficient, which controls the magnitude of gradient updates per step. Indicates the parameter The gradient operator represents the direction and magnitude of the gradient of the model parameters with respect to the loss function.

9. The intelligent control strategy and framework for the continuous loading production line according to claim 1, characterized in that: The multi-layered secure causal reasoning module includes a causal graph reasoning layer, a fuzzy logic judgment layer, and a linkage response layer. The causal graph inference layer constructs a Bayesian network based on a variable causal graph to model the causal relationships between process variables. The variable causal graph is represented as follows: in: Indicates ambient temperature; Indicates the temperature of the equipment wall surface; Indicates the viscosity of the slurry; Indicates mixed torque; Indicates the stress on the stirring shaft; In a variable causal graph, each edge is assigned a causal strength. This forms a Bayesian network: The fuzzy logic inference layer stores several semantic rules in the form of "IF-THEN", and uses Gaussian or trapezoidal membership functions to fuzzify input variables into levels such as "high", "medium", and "low", and then calculates the security risk level based on the fuzzy inference rules; the semantic rules in the form of "IF-THEN" are expressed as follows: If temperature rise rate > High AND twin-screw shaft torque > Large, then the hazard level is High. Mathematically, it is expressed as: in: The membership function representing the high-risk level indicates the confidence or credibility of the system being in a high-risk state; A fuzzy membership function representing the rate of temperature rise, used to reflect whether the rate of temperature rise exceeds a safe threshold; A fuzzy membership function representing the stirring torque is used to reflect the mechanical load level during the stirring process; The linkage response layer is configured to map the safety risk level to a predetermined response strategy, wherein a high risk level triggers a stop to mixing and AGV path reconstruction, a medium risk level triggers a reduction in mixing speed and an increase in cooling, and a low risk level triggers logging and an increase in monitoring frequency.

10. The intelligent control strategy and framework for the continuous loading production line according to claim 9, characterized in that: The fuzzy logic inference layer embeds linear temporal logic rules for modeling time-coupled anomaly patterns. The rule form is as follows: in: Let be the membership function; It represents the rate of temperature rise, i.e., the derivative of temperature with time, and is used to characterize the intensity of the thermal dynamic changes of the system. This indicates the change in torque, reflecting the degree of load fluctuation on the stirring shaft within a short period of time. It is usually related to sudden changes in material viscosity or equipment jamming. The membership function represents the risk level, and the fuzzy confidence level represents the current high-risk state of the system. Indicates the future Linear sequential logic operators that are "finally satisfied" within seconds; The variable causal graph is automatically constructed by Granger causality test and PCMCI+ algorithm to automatically mine the potential causal path structure between sensor variables, and a time series causal discovery algorithm is introduced to construct a data-driven dynamic graph structure. The Granger causality test determines causal relationships based on the predictive power of time lags. The principle is as follows: If added Significant improvement The predictive ability is considered to be... ; in: This indicates the current state of the explained variable (target variable). The value of is the current response or observation of the system; Represents the explained variable itself The lagged term represents the past. Historical values ​​within a time step are used to capture its autoregressive characteristics; This indicates that exogenous variables (potential influencing factors) in The value of the first lag time reflects the time influence of external variables on the target variable; Represents random disturbance terms or white noise, used to represent unmodeled noise or random effects; The lag order, i.e., the number of historical steps considered, is determined by the information criterion. and These represent the autoregressive coefficient and the cross-regression coefficient, respectively, describing... Self-history and external variables For the current value The intensity of the impact; The PCMCI+ algorithm is suitable for high-dimensional, multi-lag data, supports pruning and residual analysis to form a dynamically updated directed acyclic graph, where each edge represents a potential causal chain. (Time lag at the edge of causality) in: and They represent the first The and the first A time series variable, representing the target quantity that is affected or predicted; Indicates the length of time lag, representing the amount of time delay in the causal effect, reflecting... right The lag effect.

Images