Cooperative control system based on digital twinning in spandex fiber production

Abstract

The invention provides a cooperative control system based on digital twinning in spandex fiber production, and the system comprises a global optimization layer which comprises a global digital twinning model, and the global digital twinning model comprises a hybrid dynamic system model and a multi-target optimizer; the edge collaboration layer is deployed at an edge computing node, the edge collaboration layer comprises a function element decomposition engine, a space-time planner and a terminal execution layer, the terminal execution layer is provided with an adaptive controller, and the adaptive controller is used for generating an optimal control parameter according to a function element with a space-time voucher; and the feedback self-evolution module comprises a product compliance verification unit and a targeted model correction unit. The method and the device are used for solving the technical problems of low dynamic task response efficiency and insufficient global optimization capability.

Description

Technical Field

[0001] This invention relates to the technical field of polyurethane production, and more particularly to a control system for spandex fiber production. Background Technology

[0002] With the development of modern manufacturing, especially process industries based on production lines, small-batch and multi-batch production has become a challenge for capacity adaptation in the face of increasingly refined and diversified market demands. In this context, the traditional manufacturing system architecture based on the ISA-95 standard is ill-suited to the dynamic and complex production tasks, representing a lag in industrial control.

[0003] Traditional control systems have the following shortcomings when dealing with the above-mentioned new requirements:

[0004] In traditional control systems, control systems at different levels, such as ERP and MES, and at different processes, such as polymerization and spinning, are independent and form an island-like structure. As a result, handling emergency order insertions or production change tasks used to be a cumbersome and time-consuming process, requiring manual modification of production plans one by one, which was time-consuming and inefficient.

[0005] The lack of overall coordination leads to poor resource allocation, serious conflicts, low efficiency, and waste of resources when cross-process tasks are executed.

[0006] The system cannot achieve overall optimization; upgrades to the control system are limited to individual modules, resulting in an overly simplistic approach that fails to achieve optimal global performance. Summary of the Invention

[0007] To address the technical problems of inefficient dynamic task response, severe cross-process resource conflicts, and insufficient global optimization capabilities caused by hierarchical and process silos, this invention proposes a collaborative control system based on digital twins for spandex fiber production.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0009] A digital twin-based collaborative control system for spandex fiber production includes:

[0010] A global optimization layer, deployed in the cloud, includes a global digital twin model, which in turn includes a hybrid dynamic system model and a multi-objective optimizer. The hybrid dynamic system model is used to predict future states and optimization objectives based on multi-dimensional real-time data from the entire plant. The multi-objective optimizer is used to call the hybrid dynamic system model to solve for the Pareto optimal production plan set based on the NSGA-II algorithm.

[0011] An edge collaboration layer is deployed on edge computing nodes. The edge collaboration layer includes a functional primitive decomposition engine and a spatiotemporal planner. The functional primitive decomposition engine is used to decompose the Pareto optimal production plan into standardized functional primitives. The spatiotemporal planner is used to allocate corresponding spatiotemporal credentials to the functional primitives.

[0012] The terminal execution layer is deployed on the edge device and is equipped with an adaptive controller. The adaptive controller is used to generate optimal control parameters based on the functional primitives with spatiotemporal credentials. The edge device executes production tasks based on the optimal control parameters and obtains actual product output data.

[0013] The feedback self-evolution module includes a product compliance verification unit and a targeted model correction unit. The product compliance verification unit is used to compare the actual product output data with the predicted data of the hybrid dynamic system model, and the targeted model correction unit is used to correct the parameters of the hybrid dynamic system model based on the comparison results.

[0014] Preferably, the hybrid dynamic system model is described globally using a discrete-time nonlinear state-space model, where the state in the hybrid dynamic system model is a high-dimensional state vector describing the specific physical information in the spandex fiber production process. This includes the state variables of the reactor at time k, the state variables of the AGV at time k, and the state variables of the warehouse at time k; control input vector. The executable macroscopic commands output by the global optimization layer include at least planned output, reactor equipment start / stop signals, and AGV target point coordinates; system output vector The outputs of the three sub-models include final product output, total energy consumption, and risk score; the dynamic evolution of the state variables of the reactor, AGV, and warehouse at time k adopts an additive hybrid form of mechanistic model and data-driven model.

[0015]

[0016] in, This represents a mechanistic model describing the production process of spandex fibers. This is a data-driven model used to compensate for the discrepancy between the predicted output and the actual output of the mechanistic model. Mechanism parameters, These are parameters for the data-driven model.

[0017] Preferably, in the multi-objective optimizer, calling the hybrid dynamic system model to solve for the Pareto optimal production schedule set based on the NSGA-II algorithm includes:

[0018] To control the input vector To optimize the variables, the system output vector is used. The objective function is constructed with the goals of maximizing product output, minimizing total energy consumption, and minimizing risk integral, along with system dynamics constraints, state constraints, input constraints, and output constraints. During simulation, a model reduction technique based on intrinsic orthogonal decomposition is used to reduce the order of sub-models. Distributed parallel simulation is then performed on the reduced sub-models, and the NSGA-II algorithm is used for optimization during simulation to obtain the Pareto optimal production plan set.

[0019] Preferably, the method for optimization using the NSGA-II algorithm is as follows:

[0020] Step 1.1: Based on the optimization variables, perform individual encoding for each control input component: ,in, This represents the planned output of product e; and It is the start and end time window of the plan; Mode is an enumeration variable representing the production mode;

[0021] Step 1.2: Define the population size N and control the upper and lower limits of the input. , Each individual is randomly initialized to obtain the initial population. The fitness of each individual is calculated based on the objective function.

[0022] Step 1.3: For the current population, obtain the non-dominated front set by performing non-dominated sorting based on the fitness of each individual, and calculate the crowding degree based on the non-dominated front set to obtain the crowding degree value of each individual;

[0023] Step 1.4: Randomly select l individuals from the current population. Select individuals with low non-dominance levels in the non-dominance front set. If the non-dominance levels are the same, select individuals with high crowding values. Continue until N parent individuals are selected.

[0024] Step 1.5: Simulate binary crossover on the selected parent individuals to generate offspring. For the offspring individuals, use polynomial mutation to change their gene values ​​with a preset probability to obtain N offspring individuals.

[0025] Step 1.6: Merge the parent individuals and child individuals into a population of size 2N, calculate the fitness of each individual, repeat the non-dominated sorting, calculate the crowding degree, and select the parent individuals to obtain the new generation population. When the algorithm reaches the maximum number of generations or the population converges, the termination condition is met.

[0026] Preferably, in the functional primitive decomposition engine, the Pareto optimal production plan is decomposed into standardized functional primitives, including:

[0027] An ontology-based semantic network technology is used to construct a process knowledge rule base by structurally representing process knowledge.

[0028] Functional primitives are generated by recursively decomposing the selected production plan by searching the process knowledge rule base.

[0029] The data structure representation of the functional primitive is as follows: ;

[0030] in, It is a globally unique identifier; Task type; The execution resource binder is used to specify the specific physical device ID for executing each functional base; The parameterized execution payload is used to provide the information required to perform the task; This is the set of dependencies for the predecessor task.

[0031] Preferably, in the spatiotemporal planner, the corresponding spatiotemporal credentials are assigned to the functional primitives, including:

[0032] Step 2.1: Construct a unified spatiotemporal resource table for storing spatiotemporal credentials based on spatiotemporal discretization and four-dimensional hash mapping;

[0033] When applying for a spatiotemporal certificate for each functional primitive, calculate the key of all discrete spatiotemporal points within the spatiotemporal range required to execute each functional primitive, and query the unified spatiotemporal resource table based on the key;

[0034] Step 2.2: When encountering resource conflicts while querying the unified spatiotemporal resource table, employ diverse resolution rules, including at least:

[0035] Time offset processing: fine-tuning the start time of the task;

[0036] Spatial rerouting: Planning a slightly longer but conflict-free alternative path for the AGV;

[0037] Resource replacement: If multiple similar resources are available, the task will be reassigned to another idle resource.

[0038] Preferably, in the adaptive controller, optimal control parameters are generated based on functional primitives with spatiotemporal credentials, including:

[0039] Step 3.1: The functional primitives adaptively update the gain parameters through the scheduling function fitted offline, and obtain the control parameters for the polymerization reactor through the ADRC controller based on the updated gain parameters;

[0040] Step 3.2: Find the optimal path based on the functional primitives using the adaptive A-path planning algorithm, and calculate the control parameters for AGV execution based on the optimal path.

[0041] Specifically, step 3.1 includes:

[0042] Fitting the scheduling function ;in , and For the fitting parameters, For gain parameters, The target molecular weight;

[0043] The target molecular weight is obtained from the parameterized execution load of the functional primitives, and the gain parameter is calculated based on the fitted scheduling function. ;

[0044] The control parameters executed by the polymerization reactor are obtained based on the calculation formula of the ADRC controller control output.

[0045] Specifically, step 3.2 includes:

[0046] Step 3.21: Define the actual cost, heuristic cost, and total cost that take risk weights into account;

[0047] Define the actual cost of each candidate path from the starting point to the current node n:

[0048]

[0049] in, Risk weights are dynamically provided by the parameterized execution load of the functional primitives. For path length, This is the cumulative risk value;

[0050] Define the heuristic cost from the current node n to the destination E:

[0051]

[0052] in, For standard heuristic functions;

[0053] Define total cost ;

[0054] Step 3.22: Based on the defined total cost, the adaptive A-path planning algorithm is used to solve the problem with the goal of minimizing the actual cost, and the optimal path for AGV execution is obtained.

[0055] Step 3.23: Calculate the execution speed of each sub-path based on the runtime window provided by the parameterized execution load of the functional primitive and the optimal execution path of the AGV, and use the execution speed of each sub-path as the control parameter for AGV execution.

[0056] Preferably, in the product compliance verification unit, the actual product output data is compared with the predicted data of the hybrid dynamic system model, including: monitoring the final product output and comparing it with the predicted output value of the digital twin model in the global optimization stage, calculating the absolute value deviation, and triggering a correction process when the absolute value deviation exceeds a preset deviation threshold in multiple preset batches.

[0057] In the targeted model correction unit, the parameters of the hybrid dynamic system model are corrected based on the comparison results, including:

[0058] Step 4.1: Calculate the gradient of the predicted output value with respect to all parameters of the current hybrid dynamic system model to represent the sensitivity. Based on the sensitivity, select the process sub-model that is most closely related to the output deviation. When calculating the sensitivity, forward automatic differentiation is preferred.

[0059] Step 4.2: For the selected process sub-model most closely related to the production deviation, incremental learning is used to update the targeted parameters. The beneficial effects of this invention are:

[0060] By constructing a three-tiered collaborative architecture consisting of a cloud-based global optimization layer, an edge collaboration layer, and a terminal execution layer, combined with a hybrid dynamic global digital twin model, a functional primitive decomposition engine, and a spatiotemporal planner for task processing, and through the terminal execution layer for specific operations, seamless integration between processes and efficient resource scheduling are achieved, breaking down process barriers. At the same time, it enables rapid response and global planning for dynamic production tasks (such as emergency order insertion and production changeover), effectively improving the response efficiency of dynamic tasks.

[0061] By standardizing the decomposition of functional primitives (FP) and the allocation mechanism of space-time credentials (STC), combined with the unified space-time resource table (USTRT), conflict-free scheduling and visual management of cross-process resources are achieved, which significantly reduces resource conflicts and production waiting time and improves resource utilization.

[0062] The Pareto optimal plan set is generated by the multi-objective global optimization algorithm (NSGA-II), and the digital twin model is continuously optimized by relying on the targeted model self-evolution technology. This breaks through the limitations of local optimization and achieves the global optimization of multi-objectives such as capacity, energy consumption and risk at the plant level, meeting the flexible and refined needs of small-batch and multi-variety production. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0065] Figure 2 This is a schematic diagram of the global optimization layer structure of the present invention.

[0066] Figure 3 This is a schematic diagram of the edge collaboration layer structure of the present invention.

[0067] Figure 4 This is a schematic diagram of the terminal execution layer structure of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] A digital twin-based collaborative control system for spandex fiber production, such as... Figure 1 As shown, it includes:

[0070] A global optimization layer, deployed in the cloud, includes a global digital twin model, which comprises a hybrid dynamic system model and a multi-objective optimizer. The hybrid dynamic system model is used to predict future states and optimization objectives based on multi-dimensional real-time data from the entire plant. The multi-objective optimizer uses the hybrid dynamic system model to solve for the Pareto optimal production plan set based on the NSGA-II algorithm. Figure 2 As shown.

[0071] In this embodiment of the application, the hybrid dynamic system model is a federated model system, specifically composed of multiple process sub-models representing different physical units or business logics. In this embodiment of the application, it includes a polymerization reactor sub-model M_R, an AGV logistics sub-model M_L, and a warehouse sub-model M_E.

[0072] The hybrid dynamic system model describes the entire system's dynamics using a discrete-time nonlinear state-space model, expressed as:

[0073]

[0074] Where is the discrete-time index. Let n be the global state vector at time k, and n be the dimension of the global state vector. Its elements represent specific physical or informational quantities in the manufacturing process. Let be the system control input vector at time k, and m be the dimension of the control input vector, representing the implementable macroscopic instructions output by the global optimization layer; Let p be a measurable disturbance vector, such as ambient temperature or upstream material concentration, and p be the dimension of the measurable disturbance vector. For model parameter vectors, The process noise at time k, The measurement noise at time k is... Let q be the system output vector, i.e., the optimization objective, and q be the dimension of the system output vector. A nonlinear function vector describing the evolution of the system state. This is a vector of nonlinear functions that describe the mapping relationship between the system outputs.

[0075] In this embodiment of the application, the global state vector A high-dimensional state vector includes: the state variables of the reactor at time k, the state variables of the AGV at time k, and the state variables of the warehouse at time k.

[0076] Specifically, the state variables of the reactor at time k:

[0077] in, Let K be the temperature of the reactor at time k. Let K be the pressure of the reactor at time k. The concentration of reactants in the reactor at time k, The number-average molecular weight at time k in the reactor is a core indicator for measuring the molecular chain length and performance of a product (such as the elasticity and strength of spandex fibers).

[0078] The state variables of the AGV at time k:

[0079] in, Let x be the two-dimensional plane x-coordinate of the AGV at time k. Let be the y-coordinate of the AGV in the two-dimensional plane at time k. Let be the heading angle of the AGV at time k. Let be the battery level of the AGV at time k.

[0080] The warehouse's state variables at time k:

[0081] in, Let be the inventory of reactants in the warehouse at time k.

[0082] In this embodiment of the application, the system control input vector This includes: planned output, target molecular weight, equipment start / stop signals, target point coordinates, and planned time window.

[0083] In this embodiment of the application, the system output vector Includes: final product output Total energy consumption and risk points .

[0084] In this embodiment of the application, a hybrid form of mechanism model and data-driven model is adopted for each sub-model in specific implementation.

[0085] Specifically, firstly, the state of the entire system is represented by a high-dimensional state vector S(t):

[0086]

[0087] Each sub-vector represents the state of a sub-model. For the state variables of the polymerization reactor sub-model, These are the state variables of the AGV logistics sub-model. These are the state variables of the warehouse sub-model.

[0088] Furthermore, the dynamic evolution of the system follows a system of differential equations of the general form:

[0089]

[0090] in, Let be the global state vector at time t. This is the external control input vector. For modeling dynamic or external disturbance vectors, such as changes in ambient temperature, fluctuations in raw material batches, etc. For the model's parameter vector, The vector of nonlinear functions representing the dynamic evolution of the system, including the dynamic evolution function of the polymerization reactor sub-model. Dynamic evolution function of AGV logistics sub-model Dynamic evolution function of warehouse sub-model .

[0091] Furthermore, for each sub-model, its dynamic evolution function is constructed as an additive hybrid of the mechanistic model and the data-driven model:

[0092]

[0093] in, This refers to a mechanistic model established based on first principles such as the conservation of material and energy. This is a data-driven model used to compensate for discrepancies between the mechanistic model and the real world. Mechanism parameters, These are the parameters for the data-driven model. The data-driven model uses an LSTM network or an MLP model to train the output compensation quantity of the data-driven model based on the state of the mechanistic model, and uses the deviation between the predicted output and the actual output of the mechanistic model as the label, with the mean squared error as the loss function.

[0094] In this embodiment, the multi-objective optimizer's processing includes determining optimization variables and constructing optimization objective functions and constraints. Given the computational complexity of the plant-wide digital twin model, the multi-objective optimizer employs model order reduction and distributed parallel simulation strategies to ensure its simulation efficiency, thereby meeting the real-time requirements of cloud-based optimization. During the simulation process, the NSGA-II algorithm is used to call the global digital twin model for simulation calculation.

[0095] In this embodiment of the application, constructing the optimization objective function and constraints includes:

[0096] First, determine the optimization variables:

[0097]

[0098] in, This represents the control input sequence within a finite prediction time domain, encompassing time from the current k to the future. All control commands within a given time step, each control input following the semantic rules of the prediction time | decision time. This represents the control input at time k, i.e., the immediate control action. , representing the state at the current time k, with respect to future times. The control input, i.e., predictive control action. To predict the length of the time domain.

[0099] Furthermore, construct the optimization objective function:

[0100]

[0101] in, Let be the output objective function, which measures the output performance of the production system. The negative sign in the formula transforms "maximizing output" into "minimizing negative output," facilitating a unified solution for multi-objective optimization. Let be the energy consumption objective function, measuring the energy consumption of the production process, with the goal of minimizing energy consumption. The risk objective function measures various risks in the production process, with the goal of minimizing these risks.

[0102] Further, determine the constraints, including:

[0103] System dynamic constraints:

[0104]

[0105] in, This represents the system's global state prediction vector at time k+i for time k. This is the system's dynamic evolution function, describing how the state evolves from time k+i to time k+i+1. The inputs include: the state x at time k+i, the control input u, the measurable disturbance d, and the model parameters Θ. Let be the predictable perturbation vector at time k+i for time k.

[0106] State constraints:

[0107]

[0108] in, , This represents the upper and lower bound constraint vectors for the global state of the system. It ensures that the state remains within the physically feasible range, such as non-negative concentration in the reactor, warehouse inventory not falling below the lower safety limit, and equipment temperature not exceeding the upper safety limit.

[0109] Input constraints:

[0110]

[0111] in, , To control the upper and lower limit constraint vectors of the input, ensuring that the control commands are within the device's capabilities.

[0112] Output constraints:

[0113] in, For the risk-related output predictor variables at time k+i, This serves as a risk upper limit threshold to ensure that the risks in the production process are within a safe range.

[0114] The model reduction and distributed parallel simulation strategies include:

[0115] First, by employing Model Rank Reduction (ROM) technology, the sub-models of the process are reduced in rank using methods such as Orthogonal Decomposition (POD). This simplifies the sub-models from a complex form with millions of degrees of freedom to a real-time model with only tens of degrees of freedom, thereby improving the computation speed by several orders of magnitude while ensuring accuracy.

[0116] Furthermore, by employing distributed parallel simulation, the various process sub-models (M_R, M_L, M_E) of the hybrid dynamic system model can be processed in parallel using multiple computing nodes on the cloud platform; by adopting the NSGA-II algorithm, a fast non-dominated sorting genetic algorithm, the efficiency of evaluating individual populations is significantly improved by distributing a large number of simulation tasks in the computing cluster for parallel computation; this parallel processing method reduces the simulation time, which may have previously taken several days, to hours or even minutes.

[0117] Furthermore, to balance simulation accuracy and speed, the system can dynamically adjust the simulation accuracy according to the optimization stage. In the early stage of optimization, a low-precision fast model is used for exploration, and in the later stage, a high-precision model is used for accurate evaluation of the optimal solution set.

[0118] In this embodiment, the optimization problem is a complex, computationally intensive, constrained multi-objective optimization problem. This invention employs a non-dominated sorting genetic algorithm with an elitist strategy (NSGA-II) to solve it. The implementation method of the NSGA-II algorithm is as follows:

[0119] First, based on the aforementioned optimization variables, chromosome (individual) encoding is performed for each control input component. ,in, and These represent the planned output of products A and B, respectively. and It is the start and end time window of the plan. , The target molecular weights for products A and B are given; Mode is an enumerated variable representing the production mode, such as high-speed mode and energy-saving mode.

[0120] Furthermore, we define the population size N and control the upper and lower limits of the input. , Each individual is randomly initialized to obtain the initial population. The fitness of each individual is calculated based on the objective function.

[0121] Furthermore, for the current population (initially...) (Subsequently, the population is merged). Based on the fitness of each individual, a set of non-dominated fronts (i.e., Pareto fronts) is obtained by performing non-dominated sorting. The crowding degree is calculated based on the set of non-dominated fronts to obtain the crowding degree value of each individual.

[0122] Furthermore, randomly select l individuals from the current population, select individuals with low non-dominance levels in the non-dominance front set, and if the non-dominance levels are the same, select individuals with high crowding values, until N parent individuals are selected.

[0123] Furthermore, simulated binary crossover (SBX) is applied to the selected parent individuals to generate offspring. For the offspring individuals, polynomial mutation is applied to change their gene values ​​with a preset probability to obtain N child individuals.

[0124] Furthermore, the parent and child individuals are merged into a population of size 2N, and the fitness of each individual is calculated. The non-dominated sorting, crowding calculation, and parent selection are repeated to obtain the next generation of the population. The termination condition is met when the algorithm reaches the maximum number of generations or the population converges. At this point, the algorithm outputs all individuals on the first Pareto front, forming a Pareto optimal solution set. These solution sets represent the optimal state where no individual benefit can be increased without reducing any individual benefit, providing decision-makers with a basis for balancing efficiency and fairness.

[0125] An edge collaboration layer, deployed on edge computing nodes, includes a functional primitive decomposition engine and a spatiotemporal planner. The functional primitive decomposition engine breaks down the Pareto optimal production plan into standardized functional primitives, and the spatiotemporal planner allocates corresponding spatiotemporal credentials to these primitives. These credentials include execution time windows and spatial resource allocation. The edge collaboration layer is used for action decomposition and real-time coordination, and its core lies in the FP-STC mechanism, such as... Figure 3 As shown.

[0126] In this embodiment of the application, the functional primitive decomposition engine uses ontology-based semantic network technology to construct a process knowledge rule base by structurally representing process knowledge. As a rule-based reasoning tool, the functional primitive decomposition engine generates functional primitives (FPs) by recursively decomposing and instantiating selected production plans by retrieving the process knowledge rule base.

[0127] An ontology-based semantic network technology is used to construct a process knowledge rule base by structurally representing process knowledge.

[0128] Functional primitives are generated by recursively decomposing the selected production plan by searching the process knowledge rule base.

[0129] The data structure representation of the functional primitive is as follows:

[0130] in, It serves as a globally unique identifier for full lifecycle tracking, such as FP-001; Based on task type, such as material handling or heating; To execute the resource binder, specify the specific physical device ID that executes the functional primitive (FP), such as AGV-01 or polymerization reactor 1; To parameterize the execution load, a self-contained set of parameters provides all the information needed to execute the task. For example, in the material handling functional primitive (FP), its load includes the starting location, destination location, material type, material quantity, priority, risk weighting factor, and runtime window; This is a set of predecessor task dependencies, a list of IDs of one or more predecessor FPs, such as FP2 depending on FP1 to complete.

[0131] In this embodiment of the application, the spatiotemporal planner constructs a unified spatiotemporal resource table (USTRT) based on spatiotemporal discretization and four-dimensional hash mapping, and allocates a spatiotemporal credential (STC) to each functional primitive by querying the unified spatiotemporal resource table (USTRT).

[0132] The construction of a unified spatiotemporal resource table (USTRT) based on spatiotemporal discretization and four-dimensional hash mapping includes:

[0133] First, define the physical boundaries of the factory: determine the three-dimensional coordinate range of the production area, covering the entire production space including equipment layout, logistics channels, and warehousing areas; set time dimension parameters: define a fixed duration for time slices, such as 1 second, and define the total time window. For example, in a 24-hour period, calculate the total number of time slices. .

[0134] Furthermore, spatiotemporal discretization is performed: the three-dimensional space of the factory is divided into voxels using a multi-level granularity strategy, with voxel indices as follows: Corresponding to the spatial coordinates x, y, and z axes, for fixed equipment (such as reactors), coarse-grained voxels (e.g., 1 m³ / unit) are used to define its fixed operating space to ensure that the overall space of the equipment is completely covered by a single or continuous voxel. For mobile equipment (AGV), fine-grained voxels (e.g., 0.1 m³) are used to define its path corridor to match its path flexibility requirements. The total time window is divided into continuous time slices according to a preset fixed duration, and each time slice is assigned a unique index t.

[0135] Furthermore, three-dimensional spatial coordinate encoding: The three-dimensional index of each voxel is generated using a Hilbert space-filling curve. Mapped to a unique one-dimensional integer Spatial encoding ensures that adjacent voxels are mapped to a continuous one-dimensional index, improving query locality. Four-dimensional spatiotemporal unique key generation: The spatial encoding h is combined with the time slice index t to generate a four-dimensional hash key. Ensure that each "voxel-time slice" combination corresponds to a unique key, supporting O(1) complexity queries.

[0136] The Unified Spatiotemporal Resource Table (USTRT) is essentially a distributed hash table deployed locally on edge computing nodes. It employs a sharded storage strategy, supports parallel read and write operations across multiple nodes, and is adapted to the real-time scheduling requirements of the edge collaboration layer. The data structure is as follows (appropriate):

[0137] Key: The generated unique four-dimensional hash key (Integer type);

[0138] Value: Designed as a structured object, containing 3 core fields:

[0139] status: Resource occupancy status, Free = Idle, Reserved = Reserved, Occupied = Occupied;

[0140] stc_id: Spacetime credential ID;

[0141] resource_info: Details of resource usage.

[0142] When requesting an STC for each FP, the spatiotemporal planner calculates the key of all discrete spatiotemporal points within its required spatiotemporal range (i.e., a cuboid space and a time window) and queries the USTRT accordingly. Only when all points are available are these points marked as Reserved and an STC generated.

[0143] Furthermore, to further optimize the scheduling process, the time-space planner, upon detecting resource conflicts, is not limited to a postponement strategy. Its built-in conflict resolution engine is equipped with diverse resolution rules, including but not limited to:

[0144] Time offset processing allows for fine-tuning of task start times, such as delaying FP1 to 8:05-8:35.

[0145] Spatial rerouting plans a slightly longer but conflict-free alternative path for the AGV;

[0146] Resource replacement involves reassigning a task to another idle resource when multiple similar resources are available. For example, assigning AGV-02 to replace AGV-01 for execution.

[0147] The final output is a set of standardized functional primitives with a unique spatiotemporal credential (STC), such as FP1+STC1, FP2+STC2, and FP3+STC3.

[0148] A terminal execution layer, deployed on an edge device, includes an adaptive controller. This controller generates optimal control parameters based on functional primitives with spatiotemporal credentials. The edge device executes production tasks based on these optimal control parameters to obtain actual product output data. Figure 4 As shown.

[0149] In this embodiment, the adaptive controller includes: adaptively updating the gain parameters based on the scheduling function fitted by the functional primitives through offline optimization; obtaining the control parameters executed by the polymerization reactor through the ADRC controller based on the updated gain parameters; finding the optimal path based on the functional primitives through the adaptive A-path planning algorithm; and calculating the control parameters executed by the AGV based on the optimal path.

[0150] Specifically, for obtaining the control parameters executed by the polymerization reactor, an adaptive ADRC controller is set at the reactor terminal in practical implementation. The core idea of ​​ADRC control is to estimate and compensate for the total system disturbance. The formula for calculating its control output u(k) is:

[0151]

[0152] in, These are the control parameters executed by the polymerization reactor. It is the ADRC controller's estimate of the total disturbance. It is a key parameter related to the gain of the controlled object.

[0153] Gain parameters The gain is closely related to the polymerization reaction, and this gain varies with the target molecular weight. The change. The optimization problem and adaptive mechanism lie in the gain parameter. Optimal gain parameter This can be determined through an offline optimization problem:

[0154]

[0155] in, This is a curve showing the change in molecular weight of the product over time.

[0156] Its solution can be expressed as a scheduling function: , The scheduling function obtained through offline optimization can be fitted using offline experimental data:

[0157]

[0158] in, , and These are the fitting parameters.

[0159] When the controller receives a message containing When FP is reached, it is immediately dispatched via the scheduling function. Calculate the new The t-value is then updated in the ADRC controller to calculate the control parameters executed by the polymerization reactor. This approach successfully and seamlessly translates macroscopic process objectives (FP) into optimal parameter configurations for low-level control, thereby ensuring optimal control performance under various operating conditions.

[0160] Specifically, for calculating the control parameters executed by the AGV, the adaptive A-path planning algorithm finds the optimal path and introduces risk weights from the FP.

[0161] First, we define the actual cost, heuristic cost, and total cost that take risk weights into account.

[0162] Define the actual cost of each candidate path from the starting point to the current node n:

[0163]

[0164] in, The risk weight is dynamically provided by the parameterized execution load of FP. For path length, This represents the cumulative risk value.

[0165] Define the prediction cost, or heuristic cost, from the current node n to the destination E:

[0166]

[0167] in, For standard heuristic functions, such as the Euclidean distance from node n to the endpoint, As a risk heuristic function, the AGV terminal periodically downloads a two-dimensional risk map from the cloud digital twin model. The coordinate value RiskMap[nx][ny] in the risk map is the risk value of the location of node n.

[0168] Define total cost .

[0169] Furthermore, based on the defined total cost, an adaptive A-path planning algorithm is used to solve the problem with the goal of minimizing the actual cost, thereby obtaining the optimal path for AGV execution.

[0170] Furthermore, based on the runtime window provided by FP and the optimal path for AGV execution, the execution speed of each sub-path is calculated, and the execution speed of each sub-path is used as the control parameter for AGV execution.

[0171] Ultimately, the polymerization reactor performs production based on the control parameters executed by the polymerization reactor, obtaining actual product output data, while the AGV transports materials based on the control parameters executed by the AGV.

[0172] The feedback self-evolution module includes a product compliance verification unit and a targeted model correction unit. The product compliance verification unit is used to compare the actual product output data with the predicted data of the hybrid dynamic system model, and the targeted model correction unit is used to correct the parameters of the hybrid dynamic system model based on the comparison results.

[0173] In this embodiment, the product compliance verification unit processes the following steps: monitoring the final product output and comparing it with the predicted output value generated during the planning phase using a digital twin. The comparison is performed, and the absolute value deviation is calculated. When the absolute value deviation of multiple batches exceeds the preset deviation threshold, the correction process is triggered.

[0174] When the correction process is triggered, the targeted model correction unit performs the following procedure:

[0175] First, calculate the predicted output value. For the gradients of all model parameters in the hybrid dynamic system model, the sensitivity is represented by the gradient. Based on the sensitivity, the process sub-model most closely related to the production deviation is selected. When calculating the sensitivity, forward automatic differentiation is preferentially used. Forward automatic differentiation can efficiently and accurately calculate the gradient of the output with respect to all parameters while performing digital twin model simulation. Its computational cost is only a few times that of a single simulation and far lower than the finite difference method. The sensitivity calculation formula is:

[0176]

[0177] in, This represents the i-th parameter. The larger the value, the more significant the parameter. The greater the impact on the final output, the more likely the system will identify the parameter with the highest sensitivity value and its corresponding process sub-model, determining it as the primary cause of the deviation.

[0178] Furthermore, for the selected process sub-model that is most closely related to the production deviation, incremental learning is used to update the targeted parameters, such as the gradient descent method that drives the output in the embodiment of this application:

[0179]

[0180] in, It's the learning rate. It is the momentum coefficient. It's momentum. These are the model parameters for the sub-model with the highest sensitivity. This targeted correction avoids retraining the entire complex model, has minimal computational overhead, and enables online and real-time model self-evolution.

[0181] To ensure the stability of the model correction and avoid overfitting to a single batch of data, the target model correction unit does not update the parameter θ_k immediately. Instead, it adopts a sliding window update mechanism: it collects the production deviation data and gradient information of N batches (e.g., N=10) with similar time periods and performs mini-batch gradient descent updates. This can effectively smooth noise and further enhance the robustness of the correction results.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for collaborative control in the production of spandex fibers based on digital twinning, characterized by, Comprise: A global optimization layer deployed in the cloud, comprising a global digital twin model, the global digital twin model comprising a hybrid dynamic system model and a multi-objective optimizer; the hybrid dynamic system model is used for future state prediction and optimization target prediction according to multi-dimensional real-time data of the whole plant; The multi-objective optimizer is used to call the hybrid dynamic system model to solve the Pareto optimal production plan set based on the NSGA-II algorithm; An edge collaboration layer deployed in an edge computing node, comprising a functional primitive decomposition engine and a space-time planner, the functional primitive decomposition engine is used to decompose the Pareto optimal production plan into standardized functional primitives, and the space-time planner is used to assign corresponding space-time credentials to the functional primitives; A terminal execution layer deployed on an edge device, and the terminal execution layer is provided with an adaptive controller, the adaptive controller is used to generate optimal control parameters according to the functional primitives with space-time credentials, and the edge device executes production tasks based on the optimal control parameters to obtain actual product yield data; A feedback self-evolution module, comprising a product compliance verification unit and a targeted model correction unit, the product compliance verification unit is used to compare the actual product yield data with the prediction data of the hybrid dynamic system model, and the targeted model correction unit is used to correct the parameters of the hybrid dynamic system model according to the comparison result.

2. The collaborative control system based on digital twin in spandex fiber production according to claim 1, wherein, The mixed dynamic system model is integrally described by a discrete-time nonlinear state space model, and a high-dimensional state vector in the mixed dynamic system model is used to describe specific physical information in the spandex fiber production process , including state variables of the reaction kettle at time k, state variables of the AGV at time k, and state variables of the warehouse at time k; a control input vector is an implementable macro instruction output by the global optimization layer, and at least includes a planned yield, a reaction kettle equipment start-stop signal, and AGV target point coordinates; and a system output vector is an output of the three sub-models, including a final product yield, total energy consumption, and a risk integral; and dynamic evolutions of the state variables of the reaction kettle at time k, the state variables of the AGV at time k, and the state variables of the warehouse at time k adopt an additive mixed form of a mechanism model and a data-driven model. wherein, represents a mechanistic model describing the production of spandex fibers, is a data-driven model for compensating for deviations of the mechanistic model's prediction output from the real output, is a mechanistic parameter, is a data-driven model parameter.

3. The system for collaborative control based on digital twin in spandex fiber production according to claim 2, wherein In the multi-objective optimizer, the hybrid dynamic system model is called to solve the Pareto optimal production plan set based on the NSGA-II algorithm, comprising: with the control input vector An optimization objective function is constructed with the product output vector maximizing product output, minimizing total energy consumption and minimizing risk integral as the target, system dynamics constraints, state constraints, input constraints and output constraints; in simulation solving, a model order reduction technology based on intrinsic orthogonal decomposition is used for sub-model order reduction, distributed parallel simulation is performed on the reduced sub-model, NSGA-II algorithm is used for optimization solving in simulation, and a Pareto optimal production plan set is obtained.

4. The system for collaborative control based on digital twin in spandex fiber production according to claim 3, wherein, The method for optimization and solution by using the NSGA-II algorithm is: Step 1.

1. Individual coding for each control input component based on the optimization variables: where, represents the planned production of product e; and is the planned start and end time window; Mode is an enumeration variable representing the production mode; Step 1.2, define population size N, control input upper and lower limits , , randomly initialize each individual to get initial population , calculate fitness of each individual based on the objective function; Step 1.3, for the current population, the non-dominated front set is obtained by non-dominated sorting based on the fitness of each individual, and the crowding degree value of each individual is obtained by calculating the crowding degree based on the non-dominated front set; Step 1.4, randomly select l individuals from the current population, select individuals with low non-dominated level in the non-dominated front set, if the non-dominated level is the same, select individuals with large crowding degree value, until N parent individuals are selected; Step 1.5, simulate binary crossover on the selected parent individuals to generate offspring, and use polynomial mutation on the offspring individuals to change their gene values with a preset probability to obtain N offspring individuals; Step 1.6, combine the parent individuals and offspring individuals into a population with a size of 2N, calculate the fitness of each individual, and repeat the non-dominated sorting, crowding degree calculation and parent individual selection to obtain a new generation population, when the algorithm reaches the maximum evolution generation or the population converges, the termination condition is satisfied.

5. The system for collaborative control based on digital twin in spandex fiber production according to any one of claims 1-4, characterized in that, In the functional primitive decomposition engine, the Pareto optimal production plan is decomposed into standardized functional primitives, comprising: Using ontology-based semantic network technology to structure the process knowledge and construct a process knowledge rule base; Recursively decompose the selected production plan to generate functional primitives by searching the process knowledge rule base; The data structure representation of the functional motif is: ; wherein, is a globally unique identifier; is a task type; is an execution resource binder, used to specify the specific physical device ID to execute each function base; is a parameterized execution payload, used to provide information required to execute the task; is a predecessor task dependency set.

6. The system for collaborative control in the production of spandex fibers based on digital twinning according to any one of claims 1-4, characterized in that, In the space-time planner, the functional primitives are assigned corresponding space-time credentials, comprising: Step 2.1, based on space-time discretization and four-dimensional hash mapping, a unified space-time resource table for storing space-time credentials is constructed; When applying the space-time credentials for each functional element, the Key of all discrete space-time points in the space-time range required for the execution of each functional element is calculated, and the unified space-time resource table is queried according to the Key; Step 2.2, when querying the unified space-time resource table, in the event of resource conflicts, a variety of solving rules are adopted, at least including: Time offset processing: fine-tuning the start time of the task; Space rerouting: planning an alternative path for the AGV that is slightly longer but conflict-free; Resource replacement: if multiple similar resources are available, the task is re-assigned to another idle resource.

7. The system for collaborative control based on digital twin in spandex fiber production according to claim 6, wherein, In the adaptive controller, the optimal control parameters are generated based on the functional elements with space-time credentials, including: Step 3.1, the scheduling function fitted by offline optimization is used to update the gain parameters of the functional elements, and the control parameters for the execution of the AGV are obtained based on the updated gain parameters through the ADRC controller; Step 3.2, the optimal path is found based on the adaptive A path planning algorithm, and the control parameters for the execution of the AGV are calculated based on the optimal path.

8. The system for collaborative control based on digital twin in spandex fiber production according to claim 7, wherein, Step 3.1 includes: Fitting a scheduling function ; wherein , and are fitting parameters, is a gain parameter, is a target molecular weight; The calculated molecular weight of the target molecule is obtained from the parameterization of the functional motifs performing the load, based on the fitted scheduling function, the gain parameters are calculated ; The control parameters for the execution of the AGV are obtained based on the calculation formula of the ADRC controller control output.

9. The digital twin based collaborative control system in spandex fiber production according to claim 7 or 8, characterized in that, Step 3.2 includes: Step 3.21, define the actual cost, heuristic cost and total cost considering the risk weight; Define the actual cost of each candidate path from the starting point to the current node n: wherein, is a risk weight, provided dynamically by the parameterization of the functional primitive of the load, is a path length, is an accumulated risk value; Define the heuristic cost from the current node n to the end point E: wherein is the standard heuristic function; define total cost ; Step 3.22, based on the defined total cost, use the adaptive A path planning algorithm to solve the problem with the goal of minimizing the actual cost, and obtain the optimal path for the execution of the AGV; Step 3.23, based on the execution speed of each sub-path calculated based on the running time window provided by the parameterized execution load of the functional element and the optimal path for the execution of the AGV, the execution speed of each sub-path is used as the control parameter for the execution of the AGV.

10. The digital twin based collaborative control system in spandex fiber production according to claim 9, wherein, In the product compliance verification unit, the actual product yield data and the prediction data of the hybrid dynamic system model are compared, including: monitoring the final product yield and comparing it with the predicted yield value of the digital twin model in the global optimization stage, calculating the absolute value deviation, and when the absolute value deviation exceeds the preset deviation threshold in multiple batches, triggering the correction process; In the targeted model correction unit, the hybrid dynamic system model parameters are corrected according to the comparison results, including: Step 4.1, calculate the gradient of the predicted yield value to all parameters of the current hybrid dynamic system model to represent the sensitivity, select the process sub-model closest to the yield deviation based on the sensitivity, and preferentially use forward automatic differentiation for sensitivity calculation; Step 4.2, for the selected process sub-model closest to the yield deviation, use incremental learning for targeted parameter update.

Images