Finishing multi-machine cluster simulation control method and system based on attention mechanism and mechanism constraint

Abstract

The application provides a kind of based on attention mechanism and mechanism constraint's finish rolling multi-machine cluster simulation control method and system, belongs to hot rolling process multi-machine cluster control field.The method constructs finish rolling multi-machine cluster simulation model and runs, generates local state and global state, constructs global state space;Every rack constructs intelligent agent actor network, extracts rack features based on attention mechanism, obtains PID control parameter;To the intelligent agent actor network of all racks, construct shared critic network, extract quality indicators based on attention mechanism, complete network screening;Based on process mechanism constraint, construct multi-objective optimization function;Pre-set finish rolling unit control index, establish optimization direction, train global intelligent agent actor network, shared critic network;Finally, the trained global intelligent agent actor network is deployed to rack workstation, and the PID control parameter is calculated.The application realizes the collaborative control of finish rolling process and improves the control precision of finish rolling process.

Description

Technical Field

[0001] This invention belongs to the field of multi-machine cluster control in hot rolling processes, specifically involving a simulation control method and system for multi-machine cluster control in finishing rolling based on attention mechanisms and mechanistic constraints. Background Technology

[0002] Hot rolling is a process that involves handling thick slabs through heating furnaces, roughing mills, hot output roller conveyors and flying shears, finishing mills, laminar flow cooling, and coiling mills to produce thin-gauge steel coils that meet the requirements of downstream industries. Figure 1 As shown. Among them, the finishing mill is the core link of the entire hot rolling production line, directly determining the thickness accuracy, shape quality, and microstructure of the final product. Typically, the finishing mill consists of seven stands arranged in sequence, F1 to F7. The slab passes through each stand in sequence at a high temperature, undergoing deformation with decreasing thickness under the squeezing action of the rolls in each stand, thereby producing steel coil products that meet the set thickness and other key quality requirements.

[0003] To achieve high-precision target thickness, finishing mills are generally equipped with Automatic Gauge Control (AGC) systems. These systems consist of three coupled dynamic subsystems: ① Automatic Position Control (APC) system: This system rapidly adjusts the roll gap position via hydraulic cylinders, directly altering the strip exit thickness; ② Automatic Speed ​​Regulator (ASR) system: This system precisely controls the rotational speed of each stand's rolls to regulate the strip flow rate per second, maintaining speed matching between stands; ③ Hydraulic looper system: This system controls the angle and tension of the looper supports between adjacent stands, buffering strip stacking caused by speed differences and providing closed-loop control of strip tension. These three systems exhibit complex coupling characteristics: APC adjusting the roll gap alters the rolling force, affecting the exit product thickness and the motor load and speed, prompting a response from the ASR; changes in ASR speed between upstream and downstream stands further alter strip tension, triggering the intervention of the hydraulic looper system. This high-dimensional, strongly coupled, and nonlinear dynamic mechanical characteristic makes the finishing rolling process a typical complex industrial control object.

[0004] Currently, researchers have proposed various control algorithms for such complex and high-dimensional process models, mainly including: traditional proportional-integral-derivative (PID) control algorithms, neural network-based adaptive PID control algorithms, and multi-machine cluster control algorithms. Among these, traditional PID control algorithms are mostly used in traditional industrial control systems. The control of thickness, pressure, tension, and looper angle in finishing mills still primarily relies on distributed PID control based on expert experience and mechanistic knowledge. Each stand's APC, ASR, and hydraulic looper system uses an independent single-input single-output PID controller, exchanging limited information through their respective preset fixed control parameters, and is managed independently by different operators. While this framework is simple and easy to deploy, it essentially decomposes the multi-stand coupled system into multiple independent loops, lacking collaborative optimization capabilities between controllers. To address the nonlinear and time-varying characteristics caused by physical mechanisms, some research combines neural networks with PID control, resulting in neural network-based adaptive PID control algorithms. This approach utilizes neural networks to identify the dynamic characteristics of controlled objects in each rack online, and adjusts the proportional, integral, and derivative gains of the PID controller in real time based on the errors of the controlled objects. While this framework automatically extracts the nonlinear characteristics of the loop and adjusts control parameters in real time, enhancing adaptability under non-ideal conditions and avoiding the resource waste of traditional parameter tuning, it still belongs to a decentralized structure, making it difficult to establish unified collaborative decision-making. In contrast, multi-machine cluster control algorithms emphasize solving overall control tasks that a single agent cannot handle through collaborative decision-making and parameter optimization among multiple agents. By utilizing the coupling relationships between multiple agents and the final control effect, algorithm parameters are continuously optimized, resulting in a mapping from a high-dimensional state space to a multi-rack control action space. Future development holds promise for integration with industrial process mechanism models, sharing feature representations and knowledge parameters.

[0005] In existing technologies, researchers have proposed various algorithms for the cluster control of multi-stand finishing mills, mainly including: Particle Swarm Optimization (PSO), Model Predictive Control (MPC), and Reinforcement Learning (RL). Among them, the PSO algorithm can search for a set of optimal multi-loop PID control parameters offline by simulating population optimization, competition and cooperation mechanisms to balance the performance indicators of each stand and achieve overall collaborative optimization. However, in finishing mill units, changes in stand entry thickness disturbances and tension speed compensation can cause rolling characteristics to drift. When the actual state deviates from the offline standard state, the fixed parameters after tuning cannot be adjusted in real time, resulting in a decrease in control performance. Furthermore, offline tuning requires a large number of simulations, which cannot adapt to the frequent specification changes required by the production line. Distributed MPC (DMPC) strategies solve local optimization problems in the rolling time domain, compensating for MPC through local information exchange and achieving global coordination. While this effectively reduces the computational burden of MPC, it is highly dependent on the modeling accuracy of the simulation control system. Finishing mills involve multiple physical mechanisms, such as metal plastic deformation, making it difficult to establish accurate simulation control systems. Model mismatch leads to the accumulation of prediction biases in DMPC, resulting in decreased control performance. Furthermore, it still requires solving complex quadratic programming problems online, resulting in a high computational load. RL algorithms iteratively optimize the action space through interaction with the environment using Deep Deterministic Policy Gradient (DDPG). However, RL algorithms still treat the physical models of the APC system, ASR system, and hydraulic looper system in each stand as the entire high-dimensional space, leading to dimensional explosion, training instability, and an overestimation bias in the shared reviewer network evaluation metrics, resulting in unstable updates and potentially causing serious failures. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a simulation control method and system for a multi-mill cluster in finishing mills based on attention mechanisms and mechanistic constraints. By combining attention mechanisms and process mechanism constraints, a multi-mill cluster collaborative control model is constructed, which includes the coupling characteristics of multiple stands and multiple control loops in the finishing mill. Through simulation of the finishing mill process, collaborative control of the finishing mill process is achieved, thereby improving the control accuracy of the finishing mill process.

[0007] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows: In a first aspect, embodiments of the present invention provide a simulation control method for a multi-mill cluster in finishing mills based on attention mechanisms and mechanistic constraints, the method comprising the following steps: Construct a multi-mill cluster simulation model of the finishing mill unit that includes all finishing mill stands and AGC-related components; Run the simulation model to generate a corresponding local state for each rack, and construct the global state from the local states of all racks; the local states include the states of all critical masses; An independent agent network is constructed for each rack, and all racks form a global agent network. The structure of each agent network includes an input layer, a normalization layer, a rack feature extraction layer, and a fully connected layer. The input of the agent network is the local state of the multi-machine cluster simulation model, and the output is PID control parameters. A global state space is constructed based on the global states of all rack simulation models, and a global action space is constructed based on the PID control parameters output by the agent actor network. For all racks of intelligent agent actor networks, a shared critic network is constructed; the structure of the shared critic network includes a quality feature extraction layer and a fully connected layer, the input is the global state space and the global action space, and the output is the quality index of the global state and the global action; A multi-objective optimization function is constructed based on the process mechanism constraints of a multi-machine cluster simulation model of a finishing mill; the simulation model is run and sampled data is recorded to construct a training set in the global state space and global action space; Preset control parameters for the finishing mill, establish the optimization direction of the multi-objective optimization function, and construct the target actor network and the target critic network; based on the training set, the target actor network, and the target critic network, train the global agent actor network and the shared critic network, and optimize the hyperparameters of the global agent actor network and the shared critic network; The trained global agent network is deployed to the workstations in the corresponding racks. The local state of each rack is collected in real time and input into the corresponding agent network to calculate the corresponding PID control parameters.

[0008] As a preferred embodiment of the present invention, the formula for describing the local state of each rack is as follows: (9) In equation (9), For rack serial number; This represents the current partial state of the rack. This is a critical quality status. This includes the displacement and pressure of the APC system physical model, the rotational speed and torque of the ASR system physical model, and the setpoint deviations of the tension and angle of the hydraulic looper system physical model. Deviation integral , deviation differential and setting value , The number of critical masses in the current rack, and the last rack. The number of racks is 4, and the remaining racks are 6.

[0009] In a preferred embodiment of the present invention, the rack feature extraction layer employs an attention mechanism to extract rack features from the local state of each rack; and the extracted rack features include: The input layer of the intelligent agent actor network receives the local state of the current sampling step and normalizes it; In the feature extraction layer, the attention mechanism is used to extract features from the normalized local state in order to adaptively mine the weights of different state variables in the coupling feature and mechanism feature space within the rack, thereby obtaining the adaptively weighted local state. The weighted local state is then input into the multilayer perceptron (MLP) of the rack feature extraction layer to extract the current rack features layer by layer.

[0010] As a preferred embodiment of the present invention, the normalization formula for the local state is as follows: (10) In equation (10), This represents the current partial state of the rack. This represents the normalized local state. and Preset mean and standard deviation for different key quality parameters; This represents the number of critical mass components in the current rack. The feature extraction layer uses an attention mechanism to extract the weighted local state space, as shown in the following formula: (11) In equation (11), , , These are all learnable weight matrices used in attention mechanisms. For activation function, for The dimension; Indicates query vector features; Represents key vector features; Represents the features of a value vector; This represents the extracted attention weights; Represents the weighted local state space; MLP extracts current rack features layer by layer. The formula is as follows: (12) In equation (12), For activation function, and For the first The weight and bias matrix of the layer perceptron, Indicates the number of layers in the MLP; The current rack features extracted by the last layer of the sensor are: .

[0011] In a preferred embodiment of the present invention, in the fully connected layer of the agent actor network, Mapped into the output space and transformed into the distribution of PID control parameters, the formula is as follows: (13) In equation (13), For activation function, and To output the weights and biases of the mapping layer, and This defines the range of PID control parameter values ​​for each controller in the current rack. These are the PID control parameters before inverse normalization. For the final All controllers in the rack are in PID control parameters for sampling time step.

[0012] In a preferred embodiment of the present invention, a global state space is constructed based on the global states of all rack simulation models, as shown in the following formula: (14) In equation (14), These are global operating parameters. This represents the number of critical mass components in the current rack. For the dimensions of global operating condition parameters, Represents the global state space. Indicates the first The local state of the rack, where N represents the number of racks; The complete PID control parameters of the finishing mill are obtained through a network of N intelligent agents, and the global action space is constructed as follows: (15) In equation (15), Represents the global action space; The first feature extracted by the rack feature extraction layer All controllers in the rack are in PID control parameters for sampling time step.

[0013] In a preferred embodiment of the present invention, the quality feature extraction layer employs an attention mechanism to extract quality features from the global state space and the global action space; and the extracted quality features include: Step S51: Adaptively generate the weighted global state space and the weighted global action space using an attention mechanism, as shown in the following formula: (16) In equation (16), and For a normalized global state space and global action space, This represents the number of critical mass components in the current rack. The dimension for global operating condition parameters; , , All are learnable weight matrices from the shared commentator network attention mechanism. , and These represent the query vector features, key vector features, and value vector features in the shared commentator network attention mechanism, respectively. for Dimensions For activation function, This represents the weighted global state space and the weighted global action space; Step S52, will The input is fed into a multilayer perceptron (MLP) for quality feature extraction, where it is extracted and evaluated layer by layer. The quality features of the global state-global action at the sampling time step are extracted using the following formula: (17) In equation (17), For activation function, and For the first The weight and bias matrix of the layer perceptron, Indicates the number of MLP layers in the quality feature extraction layer; For the first The quality features extracted by the layer perceptron, among which The quality features of the global state-global action extracted by the last layer of the perceptron are: .

[0014] As a preferred embodiment of the present invention, the multi-objective optimization function A reward function is used; for a 7-stand finishing mill, the formula is as follows:

[0015] (19) In equation (19), , Assigning weights to different reward items, For rack serial number, This represents the number of critical mass components in the current rack. Indicates the first rack number Deviation of the set value of a key quality; For the first Roll gap setting value for each frame For the first Simulated value of slab thickness at the t-th sampling point at the entrance of each rack. For the first The simulated displacement value of the t-th sampling point obtained from the physical model simulation of the rack hydraulic servo APC system. For the first The simulated pressure value at the t-th sampling point obtained from the physical model simulation of the rack-mounted hydraulic servo APC system. For the first Mill rigidity of individual stands , For the first The linear velocity setting value of the slab at the exit of each stand. For the first Simulated value of the slab linear velocity at the t-th sampling point at the entrance of each rack. For the first and The initial tension setting value of the looper between each frame. For the first The simulated tension increment at the t-th sampling point in the physical model of the hydraulic looper system of a single frame. This is a constant weight between the tension and velocity difference.

[0016] In a preferred embodiment of the present invention, the training process for the global agent actor network and the shared critic network is as follows: Step S71: Record the sampling data at each sampling time step and store it in the dataset. ;from The batch size is randomly selected from the middle. The training set is used; samples from the training set are input into the target actor network and the target critic network, and the PID control parameters of the actor network corresponding to each frame in the target actor network and the quality index of the target critic network are calculated: (20) In equation (20), and These are the actor network and the target critic network corresponding to the nth rack in the target actor network, respectively. and For the internal parameters of the corresponding target network, Indicates the serial number of the shared critic network; This represents the global state space of the (t+1)th sampling point. Let N represent the local state space of the (t+1)th sampling point of the actor network input corresponding to the nth rack in the target actor network, where N represents the number of racks. This represents the global operating condition parameters at the (t+1)th sampling point. This represents the global action space for the (t+1)th sampling point. This represents the PID control parameters at the (t+1)th sampling point of the actor network output corresponding to the nth frame in the target actor network. This represents the weights of immediate rewards and future global action space quality metrics in the corresponding target network. Represents a multi-objective optimization function; The target quality indicator; Step S72: Construct a loss function based on the target quality index and calculate the error of the quality index output by the shared reviewer network. Backpropagation updates the current shared critic network and the target critic network; The loss function formula is as follows: ,(twenty one) In equation (21), The error in the quality metrics output by the current shared critic network; pass Update the internal parameters of the shared critics network. ; The target commentator network update formula is as follows: ,(twenty two) In equation (22), The learning rate is updated softly; Step S73: Set a threshold for the number of updates to the shared critic network; when the number of updates reaches the threshold, update the global agent actor network and the target actor network using the quality index of the target critic network; for each agent actor network, the update formula is as follows: ,(twenty three) In equation (23), For the current number A network of intelligent agents, For the present Network internal parameters; The number of samples in the training set; The actor network update formula for each rack in the target actor network is as follows: ,(twenty four); Step S74: Determine whether the iteration threshold has been reached; if it has, output the global agent actor network; if not, return to step S71.

[0017] Secondly, embodiments of the present invention also provide a simulation control system for a multi-machine cluster of finishing mills based on attention mechanisms and mechanistic constraints. The system includes a simulation model construction module, a local and global state generation module, an agent actor network construction module, a state and action space construction module, a shared critic network construction module, a multi-objective optimization function construction module, a network hyperparameter optimization module, and a distributed execution control module. The system executes the above-described simulation control method for a multi-machine cluster of finishing mills based on attention mechanisms and mechanistic constraints.

[0018] The solutions of the embodiments of the present invention have the following beneficial effects: The present invention provides a simulation control method and system for a multi-mill cluster in hot rolling mills based on attention mechanisms and mechanistic constraints. Based on the Simulink simulation platform of MATLAB, a simulation model of the finishing mill unit related to AGC is built. MATLAB-compiled code is used to construct a multi-mill cluster simulation control system for the hot rolling process based on attention mechanisms and mechanistic constraints, together with the simulation model. By real-time monitoring and closed-loop control of core physical quantities of the finishing mill unit's AGC in the hot rolling process, such as the exit slab thickness of the hydraulic servo APC system physical model of each stand, the roll speed of the asynchronous motor ASR system physical model, and the tension of the hydraulic looper system physical model, coordinated control of the finishing rolling process is achieved, improving the control accuracy of the finishing rolling process.

[0019] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0021] Figure 1 This is a flow chart of the hot rolling process in the existing technology; Figure 2 This is a schematic diagram of the simulation control method for multi-machine cluster of finishing mills based on attention mechanism and mechanistic constraints as described in the embodiments of the present invention. Figure 3 This is a flowchart of the simulation control method for multi-machine cluster of finishing mills based on attention mechanism and mechanistic constraints as described in the embodiments of the present invention; Figure 4This is a schematic diagram of the physical model structure of the asynchronous motor ASR system based on vector control in an embodiment of the present invention; Figure 5 This is a structural diagram of the transfer function of the hydraulic pressing element in the physical model of the hydraulic servo APC system in this embodiment of the invention; Figure 6 This is a structural diagram of the transfer function of the hydraulic pressure element in the physical model of the hydraulic servo APC system in this embodiment of the invention; Figure 7 This is a schematic diagram of the physical model structure of the hydraulic looper system in an embodiment of the present invention; Figure 8 This is a schematic diagram of the multi-mill cluster simulation model structure of the finishing mill in this embodiment of the invention; Figure 9 This is an example diagram of the slab thickness response curve at the outlet of each frame in an embodiment of the present invention; Figure 10 This is an example diagram of the linear velocity response curves of the slab at the entrance of each rack in an embodiment of the present invention; Figure 11 This is an example diagram of the tension response curves between the frames in an embodiment of the present invention. Detailed Implementation

[0022] After discovering the aforementioned problems, the inventors of this application conducted a detailed study on existing multi-mill cluster control methods for finishing mills. The study found that some researchers have already used Deep RL (DRL) to synchronously optimize multi-PID controllers. This is achieved by designing a DRL agent containing an actor-critic network, using the error sequence of each loop and the setpoint as state inputs, directly outputting multi-PID control parameters, and maximizing cumulative rewards through repeated interaction with the environment. This allows the agent to learn the mapping from state to the global action space. After deployment, the agent can adjust all PID control parameters in real time according to the current production environment, achieving collaborative optimization of high-dimensional process models. Currently, most methods rely on data-driven algorithms to construct network structures, which cannot explicitly fit real physical scenarios. They are more suitable for low-dimensional systems such as general chemical processes, and are prone to training instability and overestimation bias in high-dimensional scenarios. Therefore, they have not yet been applied to high-dimensional, strongly coupled, and time-varying scenarios such as finishing mills.

[0023] Meanwhile, the inventors also discovered that most existing simulation control systems use simplified models such as one-dimensional transfer function models and linearized state-space models to simulate their working states. These models cannot simulate the dynamic coupling of multiple stands and the nonlinear coupling of thickness-speed-tension, nor can they embed intelligent multi-machine cluster control algorithms to simulate closed-loop iterative training and on-site parameter setting. Therefore, the thickness-tension-speed control between multiple stands in the finishing mill is mutually coupled. Traditional simulation control systems struggle to simulate the strong nonlinearity, high dimensionality, and strong coupling characteristics of the hot rolling process. Furthermore, traditional single-agent control methods cannot simultaneously meet the requirements of high-precision positioning, rapid speed response, and stable tension response, affecting the coordinated optimization of quality indicators such as finished product thickness and speed. Traditional reinforcement learning algorithms face challenges such as the explosion of state-action space dimensions, difficulty in handling inter-stand coupling, and the inability of independent training of each controller to achieve global coordinated optimization. Purely data-driven learning strategies suffer from unstable training, poor interpretability, difficulty in incorporating prior industrial knowledge, and heavy computational burden during online execution, failing to meet the real-time requirements of the finishing mill. Currently, there is no simulation control system that possesses the following multiple characteristics: ① It includes a complete 7-stand finishing mill model, and the model parameters can simulate different steel grades, specifications, and incoming material conditions; ② It simulates the nonlinear coupling between the physical models of the APC system, ASR system, and hydraulic looper system; ③ It introduces an intelligent multi-stand cluster control method and connects it to the simulation control system for closed-loop training and verification. Therefore, existing methods struggle to handle the strong coupling characteristics between multiple stands, and are unable to efficiently coordinate and optimize multiple control loops. They cannot meet the high-precision, high-real-time, and robust multi-stand cluster control requirements of hot rolling finishing mills, severely restricting the efficiency of transforming multi-stand cluster control methods from theoretical research and simulation to industrial engineering applications.

[0024] It should be noted that the defects in the above-mentioned prior art solutions are all the result of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of the present invention in the following text should be the inventors' contributions to the present invention.

[0025] 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 a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. It should be noted that, without conflict, the embodiments and features in the embodiments of the present invention can also be combined with each other.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the embodiments of the present invention, the terms "first," "second," "third," "fourth," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, sometimes a subscript such as W1 may be written in a non-subscript form such as W1, and their meanings are consistent unless the distinction is emphasized.

[0027] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0028] After the above in-depth analysis, this invention provides a simulation control method and system for a multi-mill cluster in finishing mills based on attention mechanisms and mechanistic constraints. A multi-mill cluster control method for finishing mills is constructed based on reinforcement learning, integrating attention mechanisms and process mechanism constraints. A simulation model of the finishing mill is built based on the physical mechanism and nonlinear coupling characteristics of the rolling process. Control of the finishing mill in the hot rolling process is achieved through centralized training and distributed execution of the control method, as well as simulation of the model. The multi-mill cluster simulation control method treats each stand as an agent network based on an attention mechanism, from which the coupling characteristics of the control loops within the stand can be extracted. All agent networks share two commentator networks to extract the coupling characteristics of the multi-stands, realistically reflecting the dynamic coupling of the multi-stands. Furthermore, an evaluation system and indicators conforming to the process mechanism of the finishing mill are constructed by integrating the finishing mill physical model to guide the parameter updates of each agent network. An APC system physical model is constructed based on dual-input dual-output PID control to ensure accurate mill settings and determine the thickness and quality of the hot-rolled finished product. An ASR system physical model is constructed based on vector control to stabilize the rolling mill. The process involves controlling the metal flow rate per second to ensure precision in thickness and speed control. A physical model of the hydraulic looper system is constructed based on feedforward compensation decoupling control to guarantee constant tension rolling and assist in regulating the thickness and speed of hot-rolled products. Based on the bounce equation, flow rate conservation per second, tension-speed compensation, and spatiotemporal lag, the dynamic coupling of multiple stands and control loops is simulated. A forward displacement-pressure control strategy and a reverse speed-tension coordination strategy for the stands are constructed. Through a finishing mill simulation model, the actual incoming material thickness, rolling requirements, and spatiotemporal lag characteristics are simulated under closed-loop conditions, providing reliable technical support for the simulation control of multiple finishing mill clusters in the hot rolling process and improving the accuracy and precision of finishing mill control. During the centralized training phase, a shared commentator network, a global state space, and a reward function based on mechanistic constraints are combined to interactively train with the finishing mill simulation model, optimizing the collaborative strategies among all agent networks. In the distributed execution phase, each agent network relies only on local observation states to output the global action space, achieving adaptive collaborative control of multiple stands and control loops.

[0029] like Figure 2 and Figure 3 As shown, the simulation control method for multi-machine clusters of finishing mills based on attention mechanisms and mechanistic constraints includes the following steps: Step S1: Construct a multi-mill cluster simulation model of the finishing mill units that includes all finishing mill stands and AGC-related components.

[0030] In this step, the multi-machine cluster simulation model of the finishing mill unit includes simulation models of all finishing mill stands coupled by the physical models of the asynchronous motor automatic speed regulation (ASR) system, the hydraulic servo automatic position control (APC) system, and the hydraulic looper system. Each stand simulation model is constructed according to the thickness-speed-tension coupling mechanism between stands, including: bounce equation, metal flow rate conservation equation, tension-speed compensation equation, and time-space lag equation.

[0031] Specifically, this step includes: Step S11: Construct the physical model of the asynchronous motor ASR system.

[0032] The physical model of the ASR system ensures that the rotational speed of each frame can quickly, stably, and error-free follow the set value, thereby maintaining constant strip tension between frames. During model construction, vector control uses coordinate transformation to equate the three-phase asynchronous motor to a DC motor, thus achieving decoupled control of flux linkage and torque. Through analysis of the voltage equation, flux linkage equation, etc., in the synchronous rotating coordinate system, the state equations of the asynchronous motors in the simulation models of each frame are constructed as follows: (1) In equation (1), The rotor's electric angular velocity, This represents the number of pole pairs of the motor. The mutual inductance between the coaxial equivalent windings of the stator and rotor. For rotational inertia, The self-inductance of the rotor's equivalent two-phase winding is... and The stator current components are in a synchronously rotating coordinate system. , The rotor electromagnetic time constant is For load torque, The leakage coefficient is... For the self-inductance of the equivalent two-phase stator winding, The resistance per phase of the stator, The resistance per phase of the rotor, To synchronize rotational angular velocity, and For stator voltage components, This is the slip angular frequency, which enables the physical modeling of the asynchronous motor ASR system in each rack.

[0033] like Figure 4 As shown, the physical model of the asynchronous motor ASR system receives a given speed. Given rotor flux and its simulation values and The given value of the stator current torque component is output through the speed and torque controller. The given value of the stator current excitation component is output through the flux linkage controller. The given vector can be converted into the instantaneous given value of the three-phase current using the Park and Clark inverse transforms. and compared with simulation values The input is fed into the current tracking pulse width modulation module to generate three-phase voltage. And drive the motor to run. Meanwhile, Automatic regulation can be achieved through a three-phase current sensor. Quality indicators such as speed and torque can be controlled through corresponding feedback loops. The ASR system physical model provides the control principle for the metal flow rate conservation and tension-velocity compensation equations.

[0034] Step S12: Construct the physical model of the hydraulic servo APC system.

[0035] In this step, the physical model of the hydraulic servo APC system adopts a master-slave cooperative dual-input dual-output PID control. The position loop directly determines the opening degree of the mill roll gap; the pressure loop senses the load changes during rolling and assists the position loop in control.

[0036] The position loop can be mapped to a transfer function model in the complex frequency domain, such as... Figure 5 As shown. Among them, Set the displacement value. and The transfer functions of the servo amplifier and the displacement sensor are respectively, and both can be regarded as proportional amplification components. For the transfer function of the servo valve, Let be the transfer function of the hydraulic cylinder. The second-order models for both are constructed as follows: (2) In equation (2), This refers to the flow gain coefficient of the servo valve. It is a second-order oscillatory element. The damping coefficient is... For traffic, This is the input current. For the piston control chamber area, , For the rigidity of the rolling mill, For the plastic stiffness of steel plate, , Main sliding radial clearance, For the area gradient of the valve core, This refers to the viscosity of the oil. The leakage coefficient of the hydraulic cylinder. The angular frequency of the inertial element. The angular frequency of the second-order oscillating element. This is the damping coefficient of the looper cylinder.

[0037] The transfer function of the pressure ring is similar to that of the hydraulic pressing element, and its structure is as follows: Figure 6 As shown. Among them Set the pressure value. This is the transfer function of the pressure sensor. Its response frequency is much higher than the natural frequency of the cylinder load, and can be regarded as a proportional amplification element. The physical model of the APC system provides the control principle for the bounce equation through the coupling effect of the position loop and the pressure loop, thereby realizing roll gap compensation and obtaining the true exit thickness of each frame.

[0038] Step S13: Construct a physical model of the hydraulic looper system.

[0039] The physical model of the hydraulic looper system is input to the current reference value deviation and the upstream frame speed reference value deviation. By adjusting the hydraulic cylinder thrust and the upstream frame speed, the looper angle and tension are changed, thereby absorbing the speed difference between the frames, avoiding strip stacking, and ensuring thickness-speed stability. The tension increment is related to the... and The speed difference between the individual frames is related, and the looper angle is affected by the combined effects of tension torque, gravitational torque, and dynamic torque. Therefore, the equation for tension versus angle increment is: (3) In equation (3), For the tension between the frames, For elastic modulus, This refers to the distance between racks. and The linear velocity of the rolls. and For rack exit speed and inlet speed, The forward slip coefficient, This is the backslip coefficient. To create a flexible upward angle, For the output torque of the hydraulic cylinder, For tension torque, It is the sum of the rotational inertia of the looper and the hydraulic cylinder.

[0040] The physical model of a hydraulic looper system contains numerous and deeply coupled control parameters, necessitating the introduction of a decoupling controller to aid in model construction. This is based on the deviation of the looper current reference value. Deviation from upstream rack speed reference value As input, in loop angle increments and tension increment To obtain the model output, through analysis and order reduction of the coefficients at each order, we can obtain the following: Figure 7 The physical model shown. Wherein, , , and Both are second-order transfer functions, and the two feedforward decoupling elements are: (4) In equation (4), and This is the transfer function corresponding to the PID controller. The physical model of the hydraulic looper system can provide the control principle for tension-speed compensation.

[0041] Step S14: Using the physical models of the asynchronous motor ASR system, the hydraulic servo APC system, and the hydraulic looper system, a simulation model of a single stand in the finishing mill is jointly constructed. Then, based on the bounce equation, the metal flow rate conservation equation, the tension-velocity compensation equation, and the spatiotemporal lag, the thickness-velocity-tension coupling between the seven stands is achieved, thereby constructing a multi-machine cluster simulation model of the finishing mill.

[0042] The bounce equation is the core equation relating the pressure loop and position loop in the physical model of each frame's hydraulic servo APC system. Individual rack outlet thickness for: (5) In equation (5), , and The first The initial roll gap, roll gap setting, and mill stiffness of each stand. , and For the first The displacement simulation values, displacement setpoint values, and pressure simulation values ​​are obtained from the physical model simulation of the individual frame hydraulic servo APC system. When fluctuations in incoming material thickness, etc., cause... When changes occur, the bounce equation can calculate the increment of the exit thickness, providing a basis for the physical model of the subsequent rack APC system.

[0043] The metal flow rate per second conservation equation is the physical basis of thickness-velocity coupling between adjacent stands, and is the core equation relating the physical models of the asynchronous motor ASR system and the hydraulic servo APC system in adjacent stands. During steady-state rolling, the metal flow rate per second of the strip through each stand remains constant, i.e.: (6) In equation (6), For the first The linear velocity setting value of the slab at the exit of each stand. For the first Simulated values ​​of slab thickness at the entrance of each rack. For the first Simulated values ​​of the linear velocity of the slab at the entrance of each rack. For the first The radius of the rolls in each stand, and For the first and The speed setpoint and simulation values ​​obtained from the physical model simulation of the ASR system of the rack asynchronous motor are shown. The influence of the forward slip and backward slip coefficients will be eliminated by the tension speed compensation circuit.

[0044] Tension-speed compensation is the core mechanism for maintaining constant strip tension and speed-tension coupling between adjacent stands. It is also the core equation relating the physical models of the asynchronous motor ASR system and the hydraulic looper system in adjacent stands. During finishing rolling, the tension between adjacent stands is determined by the difference in inlet velocities between the upstream and downstream stands. (7) In equation (7), For the first and The initial tension setting value of the looper between each frame. The constant weight between tension and velocity difference. For the first The simulated tension increment values ​​are obtained from the physical model of the hydraulic looper system of each stand. Based on this, a speed-tension coupled physical model can be constructed. Taking the deviation between the simulated and set tension as input, the controller calculates the required speed compensation and dynamically adjusts the speed of the upstream stand. The models are coupled based on the distance between stands and the exit linear velocity of the upstream stand, and feedforward links of different stand control loops are embedded to simulate the spatiotemporal lag between multiple stands in an actual finishing mill. (8) In equation (8), For the first The hysteresis factor of each rack. For ease of modeling, the distance between racks is the same.

[0045] Using the physical models of the asynchronous motor ASR system, hydraulic servo APC system, and hydraulic looper system in the above-mentioned stand simulation models, the coupling mechanism of thickness-speed-tension between stands, and the spatiotemporal lag model, a multi-machine cluster simulation model of the finishing mill related to AGC can be constructed, such as... Figure 8 As shown.

[0046] The simulation model of the finishing mill focuses on quality indicators such as the exit thickness, exit linear speed, and inter-stand tension of each stand. Therefore, it is necessary to set the reduction amount for each stand to ensure a uniform distribution of workload and resources, as well as precise control of the final product exit thickness. The exit linear speed of the last stand also needs to be set to ensure it meets the requirements of subsequent laminar cooling and coiling processes. The exit linear speeds of other stands need to be adaptively set based on the simulation speed of the downstream stands and the simulated inter-stand tension. Inter-stand tension also needs to be set, typically constant and equal, to avoid strip stacking between stands. The hydraulic control loop of the simulation model sequentially controls stands 1-7, ensuring that each stand can accurately complete its task and that the finished product thickness at the last stand meets production requirements. In the speed-tension coupled physical model, to ensure that the speed of the last stand accurately follows the set value, the speed of the upstream stand needs to be adjusted, i.e., constructing a reverse control mechanism for stands 7-1. Simultaneously, hysteresis modules are inserted between the simulation models of each stand to simulate the spatiotemporal lag characteristics between stands.

[0047] When the APC system physical model receives the simulated inlet thickness and the setpoint outlet thickness, it first compensates for the bounce equation to obtain the displacement and pressure setpoints. In the coupled control of the ASR system physical model and the hydraulic looper system physical model, the ASR system physical model receives the simulated inlet linear velocity of the downstream stand and the simulated inlet and outlet thicknesses of the current stand. Through flow rate conservation and tension-velocity compensation, it obtains the setpoint inlet linear velocity of the current stand's ASR system physical model. The simulated inlet linear velocity can be used to analyze whether the simulated velocity compensation meets the setpoint to ensure constant tension rolling.

[0048] In the constructed multi-machine cluster simulation model of the finishing mill, each stand includes displacement and pressure controllers from the APC system physical model, speed and torque controllers from the ASR system physical model, and tension and angle controllers from the hydraulic looper system physical model. Traditional algorithms such as PID control are difficult to handle the aforementioned high-dimensional, strongly coupled dynamic interactions. Each controller corresponds to a specific critical quality.

[0049] Step S2: Run the simulation model to generate a corresponding local state for each rack. The local states of all racks are used to construct the global state. The local states include the states of all critical masses.

[0050] In this step, the formulas describing the local state of each rack of the simulation model generator are as follows: (9) In equation (9), For rack serial number; This represents the current partial state of the rack. This is a critical quality status. This includes the displacement and pressure of the APC system physical model, the rotational speed and torque of the ASR system physical model, and the setpoint deviations of the tension and angle of the hydraulic looper system physical model. Deviation integral , deviation differential and setting value , The number of critical masses in the current rack, and the last rack. The number of racks is 4, and the remaining racks are 6.

[0051] Step S3: Construct an independent agent network for each rack, and all racks form a global agent network. The structure of each agent network includes an input layer, a normalization layer, a rack feature extraction layer, and a fully connected layer. The input of the agent network is the local state of the multi-machine cluster simulation model, and the output is PID control parameters. The rack feature extraction layer uses an attention mechanism to extract rack features from the local state of each rack.

[0052] In this step, each agent network corresponds to a different mill stand and has a different number of PID controllers. For example, for a 7-stand finishing mill, the agent network corresponding to the first 6 stands includes 6 PID controllers, outputting 6 sets of PID control parameters; while the 7th stand only includes 4 PID controllers, outputting 4 sets of PID control parameters. The PID control parameters are the PID gain parameters.

[0053] The input layer of the intelligent agent actor network receives the local state of the current sampling step and normalizes it: (10) In equation (10), This represents the normalized local state. and The preset mean and standard deviation for different key quality parameters.

[0054] In the feature extraction layer, an attention mechanism is used to extract features from the normalized local states, adaptively mining the weights of different state variables in the coupled feature and mechanistic feature space within the rack, thus obtaining an adaptively weighted local state space. The feature extraction formula for the local states of each rack using the attention mechanism is as follows: (11) In equation (11), , , These are all learnable weight matrices used in attention mechanisms. For activation function, for The dimension; Indicates query vector features; Represents key vector features; Represents the features of a value vector; This represents the extracted attention weights; This represents the weighted local state space.

[0055] Then weight the local state The input is fed into a multilayer perceptron (MLP) for rack feature extraction, which extracts the current rack features layer by layer. The frame features include thickness-speed-tension coupling characteristics and the mechanistic characteristics of each layer. The extraction formula is as follows: (12) In equation (12), For activation function, and For the first The weight and bias matrix of the layer perceptron, Indicates the number of layers in the MLP; preferably, the number of layers in the MLP . For the first The features extracted by the layer perceptron, among which The features of the current rack extracted by the last layer of the perceptron are: .

[0056] Will The output space is mapped through a fully connected layer and transformed into the distribution of PID control parameters: (13) In equation (13), For activation function, and To output the weights and biases of the mapping layer, and This defines the range of PID control parameter values ​​for each controller in the current rack. These are the PID control parameters before inverse normalization. For the final All controllers in the rack are in The PID control parameters are sampled at the time step. To ensure safe and reliable control operation, parameter mapping has been used to ensure that the gain parameter of each controller is limited to a feasible range. Inside.

[0057] Step S4: Construct a global state space based on the global states of all rack simulation models, and construct a global action space based on the PID control parameters output by the agent actor network.

[0058] In this step, a global state space is constructed based on the global states of all rack simulation models, as shown in the following formula: (14) In equation (14), These are global operating parameters. For the dimensions of global operating condition parameters, Let N represent the global state space, and let N represent the number of racks. Preferably, N=7.

[0059] The complete PID control parameters of the finishing mill are obtained through a network of N intelligent agents, thereby constructing the global action space: (15) In equation (15), Represents the global action space.

[0060] Step S5: Construct a shared critic network for all the agent actor networks in all racks; the shared critic network includes a quality feature extraction layer and a fully connected layer, with the input being the global state space and the global action space, and the output being the quality indicators of the global state and global action, thereby evaluating the control effect of the agent actor network; wherein, the quality feature extraction layer uses an attention mechanism to extract quality features from the global state space and the global action space.

[0061] In one executable embodiment, all racks in this step share two shared critic networks with different initial parameters to evaluate the performance of all agent actor states and actions. During the centralized training phase, the shared critic network uses the global state space and global action space to calculate quality indicators, guiding each agent actor network to update its parameters; during the distributed execution phase, each agent actor network independently outputs PID control parameters based solely on the current rack state. This architecture combines the advantages of clear division of labor among multiple agents, strong scalability, and global coordination, making it suitable for strongly coupled, high-dimensional, and nonlinear hot rolling mill models.

[0062] The quality feature extraction layer includes an attention mechanism and a multilayer perceptron (MLP). In the quality feature extraction layer, the attention mechanism is used to extract features from the global state space and the global action space, including: Step S51: Adaptively generate the weighted global state space and the weighted global action space using an attention mechanism, as shown in the following formula: (16) In equation (16), and For a normalized global state space and global action space; , , All are learnable weight matrices from the shared commentator network attention mechanism. , and These represent the query vector features, key vector features, and value vector features in the shared commentator network attention mechanism, respectively. for Dimensions For activation function, This represents the weighted global state space and the weighted global action space.

[0063] Step S52: Input the weighted global state space and the weighted global action space into the MLP of the quality feature extraction layer, and extract and evaluate the first layer layer by layer. The quality features of the global state-global action at each sampling time step are mapped to quality indicators through a fully connected layer. The extraction formula for the quality features is as follows: (17) In equation (17), and For the first The weight and bias matrix of the layer perceptron, This indicates the number of MLP layers in the quality feature extraction layer; preferably, ; For the first The quality features extracted by the layer perceptron, among which The quality features of the global state-global action extracted by the last layer of the perceptron are: .

[0064] Will Mapping to the quality space yields quality indicators. The mapping formula is as follows: (18) In equation (18), and This is used to output the weights and bias matrices of the mapping layer.

[0065] In the evaluation algorithm based on dual-delay deep deterministic policy gradient, two shared reviewer networks with identical structures, inputs, and outputs but different initializations are constructed. Two different quality metrics are obtained at each sampling time step. and And a smaller value needs to be selected to suppress over-evaluation.

[0066] Through the forward propagation process of the intelligent agent actor network and the shared critic network based on the simulation model described above, it can be clarified how the state space is mapped to the action space, and how the global state-action is evaluated.

[0067] Step S6: Construct a multi-objective optimization function based on the process mechanism constraints of the multi-machine cluster simulation model of the finishing mill; run the simulation model and record the sampled data to construct a training set in the global state space and global action space.

[0068] In this step, the multi-objective optimization function A reward function is employed to enable the shared critic network to correctly evaluate the quality metrics of an action space and to optimize the global agent network towards improving these metrics. The reward function formula is as follows: (19) In equation (19), , Assigning weights to different reward items, For rack serial number; Indicates the first rack number Deviation of the set value of a key quality; For the first Roll gap setting value for each frame For the first Simulated value of slab thickness at the t-th sampling point at the entrance of each rack. For the first The simulated displacement value of the t-th sampling point obtained from the physical model simulation of the rack hydraulic servo APC system. For the first The simulated pressure value at the t-th sampling point obtained from the physical model simulation of the rack-mounted hydraulic servo APC system. For the first Mill rigidity of individual stands , For the first The linear velocity setting value of the slab at the exit of each stand. For the first Simulated value of the slab linear velocity at the t-th sampling point at the entrance of each rack. For the first and The initial tension setting value of the looper between each frame. For the first The simulated tension increment at the t-th sampling point in the physical model of the hydraulic looper system of a single frame. The constant weight between tension and speed difference is defined as follows: The first term of the reward function is the weighted squared error of the key mass of the physical models of the ASR system, APC system, and hydraulic looper system in each stand of the finishing mill; the second term is the constraint of the bouncing equation, metal flow rate conservation, and tension-speed compensation mechanism that the key mass of the entire finishing mill must satisfy. This enables the global intelligent agent network to extract relevant features based on the simulation model and the thickness-speed-tension coupling mechanism, thereby improving the convergence efficiency of the shared critic network in evaluating the quality indicators of the global action space and the PID control parameters.

[0069] Step S7: Preset the control indicators of the finishing mill, establish the optimization direction of the multi-objective optimization function, and construct the target actor network and the target critic network; based on the training set, the target actor network, and the target critic network, train the global agent actor network and the shared critic network, and optimize the hyperparameters of the global agent actor network and the shared critic network.

[0070] In this step, when training the global agent actor network and the shared critic network based on the training set, the sampling data at each sampling time step is recorded. and store in the dataset In the middle. Whenever the simulation model runs After one sampling time step, the global agent actor network and the shared critic network enter the training phase. The training process is as follows: Step S71, from A batch of size is randomly selected from the batch. The training set is used; samples from the training set are input into the target actor network and the target critic network, and the PID control parameters of the actor network corresponding to each frame in the target actor network and the target quality index of the target critic network are calculated: (20) In equation (20), and These are the actor network and the target critic network corresponding to the nth frame in the target actor network, respectively, and the target network structure is completely consistent with the current network structure; and For the internal parameters of the corresponding target network, Indicates the serial number of the shared critic network; via To set the weights for immediate rewards and future global action space quality metrics in the target network. The target quality indicator.

[0071] Step S72: Construct a loss function based on the target quality index and calculate the error of the quality index output by the shared reviewer network. Backpropagation updates the current shared critic network and the target critic network.

[0072] In this step, the loss function formula is as follows: ,(twenty one) In equation (21), The error in the quality metrics output by the current shared critic network; pass Update the internal parameters of the shared critics network. ; The target commentator network update formula is as follows: ,(twenty two) In equation (22), The learning rate is updated softly.

[0073] Step S73: Set a threshold for the number of updates to the shared critic network; when the number of updates reaches the threshold, update the global agent actor network and the target actor network using the quality index of the target critic network.

[0074] In this step, for each agent actor network, the update formula is as follows: ,(twenty three) In equation (23), For the current number A network of intelligent agents, For the present Network internal parameters; The number of samples in the training set; The actor network update formula for each rack in the target actor network is as follows: ,(twenty four) Step S74: Determine whether the iteration threshold has been reached; if it has, output the global agent actor network; if not, return to step S71.

[0075] The dataset is obtained by repeatedly sampling the simulation model controlled by AMC-MACC. The current shared critic network and the global agent actor network can be iteratively trained. Under the constraint of the reward function, the shared critic network can accurately evaluate the quality index of the global state-action, thereby enabling the agent actor network to extract the thickness-velocity-tension coupling features and their respective mechanistic features in the current frame, and map the features to the output space to obtain the optimal PID control parameters.

[0076] Step S8: Deploy the trained global agent network to the workstations in the corresponding racks, and collect the local state of each rack during real-time operation. The corresponding intelligent agent network is input to calculate the most suitable PID control parameters for the current state. It is then directly applied to multiple PID controllers in the current rack.

[0077] Based on the same approach, this invention also provides a simulation control system for a multi-machine cluster of finishing mills based on attention mechanisms and mechanistic constraints. The system includes a simulation model construction module, a local and global state generation module, an agent actor network construction module, a state and action space construction module, a shared critic network construction module, a multi-objective optimization function construction module, a network hyperparameter optimization module, and a distributed execution control module. The simulation model building module is used to build a multi-machine cluster simulation model of the finishing mill unit that includes all finishing mill stands and AGC-related components. The local and global state generation modules are used to run the simulation model, generate corresponding local states for each rack, and construct the global state from the local states of all racks. The intelligent agent actor network construction module is used to build an independent intelligent agent actor network for each rack, and all racks form a global intelligent agent actor network; the structure of each intelligent agent actor network includes an input layer, a normalization layer, a rack feature extraction layer and a fully connected layer; the input of the intelligent agent actor network is the local state of the multi-machine cluster simulation model, and the output is PID control parameters; The state and action space construction module is used to construct a global state space based on the global state of all rack simulation models, and to construct a global action space based on the PID control parameters output by the intelligent agent actor network. The shared critic network construction module is used to construct a shared critic network for all the intelligent agent actor networks in all racks; the structure of the shared critic network includes a quality feature extraction layer and a fully connected layer, the input is the global state space and the global action space, and the output is the quality index of the global state and the global action; The multi-objective optimization function construction module is used to construct multi-objective optimization functions based on the process mechanism constraints of the multi-machine cluster simulation model of the finishing mill. The network hyperparameter optimization module is used to run the simulation model and record sampled data, and construct a training set in the global state space and global action space; preset the control indicators of the finishing mill, establish the optimization direction of the multi-objective optimization function, and construct the target actor network and the target critic network; based on the training set, the target actor network, and the target critic network, train the global agent actor network and the shared critic network, and optimize the hyperparameters of the global agent actor network and the shared critic network. The distributed execution control module is used to deploy the trained global agent network to the workstations in the corresponding racks, collect the local state of each rack during real-time operation, and input the corresponding agent network to calculate the corresponding PID control parameters.

[0078] The system or device for executing the method in this embodiment of the invention can be a terminal or a server. The system includes a processor, a memory, and / or a transceiver, etc., and is connected via a communication bus. Each module can be implemented using a processor, memory, and / or transceiver, etc. The processor can be, but is not limited to, one or more microprocessors, central processing units, network processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) and other programmable logic devices, discrete gates, transistor logic devices, discrete hardware components, etc. The memory includes random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EPR), optical disc drive (ODM), and / or non-volatile memory, etc. The transceiver is used to communicate with network devices or with terminal devices, and includes a receiver and a transmitter. The memory and transceiver can be integrated with the processor or exist independently. It should also be noted that the system and method described in this embodiment correspond to each other; the description and limitations of the method also apply to the system, and will not be repeated here.

[0079] The simulation control method and system for multi-mill cluster of finishing mills based on attention mechanism and mechanistic constraints described in this invention are applied to the control of the finishing mill unit of a hot rolling production line.

[0080] The current finishing mill has an overall inlet plate thickness of 10mm, an outlet finished product thickness requirement of 3mm, and an outlet linear velocity requirement of 25m / s. To evenly distribute the workload and resources of each stand, the reduction displacement of each stand is set to 1mm. Based on the above actual operating conditions, a multi-mill cluster simulation model of the finishing mill with AGC related stands 1-7 is constructed. For each stand, the aforementioned agent actor network and shared critic network are built. After network optimization based on the collected raw dataset, the current actual operating conditions are input, and the finishing mill is controlled. The results are plotted as shown below. Figures 9 to 11 The following shows the slab thickness response curve at the exit of each stand, the slab linear velocity response curve at the entrance of each stand, and the tension response curve between each stand. (The following is combined with...) Figures 9 to 11 The simulation response curves are used to conduct a detailed analysis of the dynamic behavior of the simulation control system and the optimization results of the PID control parameters.

[0081] Figure 9 The time-domain response processes for the sheet metal thicknesses at the seven rack exits are presented. For example... Figure 9As shown, since the stands in the finishing mill are arranged sequentially along the rolling line, fluctuations in the thickness at the exit of the upstream stand are transmitted downstream via the strip conveying speed, resulting in a significant time lag effect. Through a trained AMC-MACC, the thickness at the exit of each stand has been instantaneously adjusted under the PID control parameters of the corresponding agent network. The mechanism-data fusion driving strategy effectively suppresses the problem of progressively amplified incoming material thickness deviations.

[0082] like Figure 10 As shown, the roll speed is a key actuator for flow rate conservation and tension-speed compensation between stands. Each stand's speed exhibits high-frequency, small-amplitude corrections while maintaining the setpoint speed of the final stand. The timing of these adjustments matches the fluctuations in thickness and tension, achieving coordinated control of thickness and speed. Thanks to the accurate simulation model and AMC-MACC, the speed response maintains reasonable overshoot and remains synchronized with thickness control. When tension deviates from the target, the speed is quickly pulled back to the tension setpoint through cascaded adjustments. Although phase differences due to time lag are visible in the curves, the envelopes of the speed curves for each stand are stable, and no risks of steel piling or pulling due to speed mismatch occur, verifying the robustness of AMC-MACC.

[0083] like Figure 11 As shown, inter-stand tension is a crucial indicator of the stability of the finishing mill. The closer an adjacent stand is to the mill exit, the faster its rotational speed and the more pronounced its speed fluctuations, making coordinated speed-tension control between stands increasingly difficult. However, through the finishing mill mechanism constraints and adaptive optimization using AMC-MACC, tension fluctuations between adjacent stands are consistently constrained within a preset process window, preventing deviation and wrinkling caused by tension relaxation. This control effect benefits from the combined effect of the thickness-tension-speed coupled dynamic equations in the simulation control system and the centralized training and distributed execution of AMC-MACC, ultimately achieving coordinated control of thickness, speed, and tension, ensuring constant tension rolling, and improving the stability of the thickness and speed of the finished product at the exit.

[0084] In summary, spatiotemporal lag exists between each stand, and all curves exhibit lag characteristics. Plate thickness and rotational speed are essentially controlled instantaneously, and the exit thickness is guaranteed to meet the set value through a bounce equation. The maximum tension between each stand is within a reasonable range, ensuring the forming quality of the finished product exiting the finishing mill. The simulation control system can significantly reduce the difference between simulation results and actual operating results, laying the foundation for online real-time hardware control. The simulated data can effectively support the tuning of PID control parameters by the proposed algorithm. Simultaneously, the AMC-MACC in this system adaptively optimizes the PID control parameters of each loop, ultimately enabling the system to achieve the desired results. Figures 9 to 11 The high-efficiency control effect shown.

[0085] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0086] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed, and is not intended to limit the scope of the claimed invention, but only to illustrate preferred embodiments of the invention. Those skilled in the art should understand that the scope of the invention is not limited to the specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A simulation control method for a multi-machine cluster of finishing mills based on attention mechanisms and mechanistic constraints, characterized in that, The method includes the following steps: Construct a multi-mill cluster simulation model of the finishing mill unit that includes all finishing mill stands and AGC-related components; Run the simulation model to generate a corresponding local state for each rack, and construct the global state from the local states of all racks; the local states include the states of all critical masses; An independent agent network is constructed for each rack, and all racks form a global agent network. The structure of each agent network includes an input layer, a normalization layer, a rack feature extraction layer, and a fully connected layer. The input of the agent network is the local state of the multi-machine cluster simulation model, and the output is PID control parameters. A global state space is constructed based on the global states of all rack simulation models, and a global action space is constructed based on the PID control parameters output by the agent actor network. For all racks of intelligent agent actor networks, a shared critic network is constructed; the structure of the shared critic network includes a quality feature extraction layer and a fully connected layer, the input is the global state space and the global action space, and the output is the quality index of the global state and the global action; A multi-objective optimization function is constructed based on the process mechanism constraints of a multi-machine cluster simulation model of a finishing mill; the simulation model is run and sampled data is recorded to construct a training set in the global state space and global action space; Preset control parameters for the finishing mill, establish the optimization direction of the multi-objective optimization function, and construct the target actor network and the target critic network; based on the training set, the target actor network, and the target critic network, train the global agent actor network and the shared critic network, and optimize the hyperparameters of the global agent actor network and the shared critic network; The trained global agent network is deployed to the workstations in the corresponding racks. The local state of each rack is collected in real time and input into the corresponding agent network to calculate the corresponding PID control parameters.

2. The method according to claim 1, characterized in that, The formulas describing the local state of each rack are as follows: ，（9） In equation (9), For rack serial number; This represents the current partial state of the rack. This is a critical quality status. This includes the displacement and pressure of the APC system physical model, the rotational speed and torque of the ASR system physical model, and the setpoint deviations of the tension and angle of the hydraulic looper system physical model. Deviation integral , deviation differential and setting value , The number of critical masses in the current rack, and the last rack. The number of racks is 4, and the remaining racks are 6.

3. The method according to claim 1, characterized in that, The rack feature extraction layer employs an attention mechanism to extract rack features from the local state of each rack; and the extracted rack features include: The input layer of the intelligent agent actor network receives the local state of the current sampling step and normalizes it; In the feature extraction layer, the attention mechanism is used to extract features from the normalized local state in order to adaptively mine the weights of different state variables in the coupling feature and mechanism feature space within the rack, thereby obtaining the adaptively weighted local state. The weighted local state is then input into the multilayer perceptron (MLP) of the rack feature extraction layer to extract the current rack features layer by layer.

4. The method according to claim 3, characterized in that, The normalization formula for the local state is as follows: ，（10） In equation (10), This represents the current partial state of the rack. This represents the normalized local state. and Preset mean and standard deviation for different key quality parameters; This represents the number of critical mass components in the current rack. The feature extraction layer uses an attention mechanism to extract the weighted local state space, as shown in the following formula: ，（11） In equation (11), , , These are all learnable weight matrices used in attention mechanisms. For activation function, for The dimension; Indicates query vector features; Represents key vector features; Represents the features of a value vector; This represents the extracted attention weights; Represents the weighted local state space; MLP extracts current rack features layer by layer. The formula is as follows: ，（12） In equation (12), For activation function, and For the first The weight and bias matrix of the layer perceptron, Indicates the number of layers in the MLP; The current rack features extracted by the last layer of the sensor are: .

5. The method according to claim 4, characterized in that, In the fully connected layer of the agent actor network, Mapped into the output space and transformed into the distribution of PID control parameters, the formula is as follows: ，（13） In equation (13), For activation function, and To output the weights and biases of the mapping layer, and This defines the range of PID control parameter values ​​for each controller in the current rack. These are the PID control parameters before inverse normalization. For the final All controllers in the rack are in PID control parameters for sampling time step.

6. The method according to claim 1, characterized in that, The global state space is constructed based on the global states of all rack simulation models, as shown in the following formula: ，（14） In equation (14), These are global operating parameters. This represents the number of critical mass components in the current rack. For the dimensions of global operating condition parameters, Represents the global state space. Indicates the first The local state of the rack, where N represents the number of racks; The complete PID control parameters of the finishing mill are obtained through a network of N intelligent agents, and the global action space is constructed as follows: ，（15） In equation (15), Represents the global action space; The first feature extracted by the rack feature extraction layer All controllers in the rack are in PID control parameters for sampling time step.

7. The method according to claim 1, characterized in that, The quality feature extraction layer uses an attention mechanism to extract quality features from the global state space and the global action space. And the extracted quality features include: Step S51: Adaptively generate the weighted global state space and the weighted global action space using an attention mechanism, as shown in the following formula: ，（16） In equation (16), and For a normalized global state space and global action space, This represents the number of critical mass components in the current rack. The dimension for global operating condition parameters; , , All are learnable weight matrices from the shared commentator network attention mechanism. , and These represent the query vector features, key vector features, and value vector features in the shared commentator network attention mechanism, respectively. for Dimensions For activation function, This represents the weighted global state space and the weighted global action space; Step S52, will The input is fed into a multilayer perceptron (MLP) for quality feature extraction, where it is extracted and evaluated layer by layer. The quality features of the global state-global action at the sampling time step are extracted using the following formula: ，（17） In equation (17), For activation function, and For the first The weight and bias matrix of the layer perceptron, Indicates the number of MLP layers in the quality feature extraction layer; For the first The quality features extracted by the layer perceptron, among which The quality features of the global state-global action extracted by the last layer of the perceptron are: .

8. The method according to claim 1, characterized in that, The multi-objective optimization function A reward function is used; for a 7-stand finishing mill, the formula is as follows: ，（19） In equation (19), , Assigning weights to different reward items, For rack serial number, This represents the number of critical mass components in the current rack. Indicates the first rack number Deviation of the set value of a key quality; For the first Roll gap setting value for each frame For the first Simulated value of slab thickness at the t-th sampling point at the entrance of each rack. For the first The simulated displacement value of the t-th sampling point obtained from the physical model simulation of the rack hydraulic servo APC system. For the first The simulated pressure value at the t-th sampling point obtained from the physical model simulation of the rack-mounted hydraulic servo APC system. For the first Mill rigidity of individual stands , For the first The linear velocity setting value of the slab at the exit of each stand. For the first Simulated value of the slab linear velocity at the t-th sampling point at the entrance of each rack. For the first and The initial tension setting value of the looper between each frame. For the first The simulated tension increment at the t-th sampling point in the physical model of the hydraulic looper system of a single frame. This is a constant weight between the tension and velocity difference.

9. The method according to claim 1, characterized in that, The process of training the global agent actor network and the shared critic network is as follows: Step S71: Record the sampling data at each sampling time step and store it in the dataset. ;from The batch size is randomly selected from the middle. The training set is used; samples from the training set are input into the target actor network and the target critic network, and the PID control parameters of the actor network corresponding to each frame in the target actor network and the quality index of the target critic network are calculated: ，（20） In equation (20), and These are the actor network and the target critic network corresponding to the nth rack in the target actor network, respectively. and For the internal parameters of the corresponding target network, Indicates the serial number of the shared critic network; This represents the global state space of the (t+1)th sampling point. Let N represent the local state space of the (t+1)th sampling point of the actor network input corresponding to the nth rack in the target actor network, where N represents the number of racks. This represents the global operating condition parameters at the (t+1)th sampling point. This represents the global action space for the (t+1)th sampling point. This represents the PID control parameters at the (t+1)th sampling point of the actor network output corresponding to the nth frame in the target actor network. This represents the weights of immediate rewards and future global action space quality metrics in the corresponding target network. Represents a multi-objective optimization function; The target quality indicator; Step S72: Construct a loss function based on the target quality index and calculate the error of the quality index output by the shared reviewer network. Backpropagation updates the current shared critic network and the target critic network; The loss function formula is as follows: ，（21） In equation (21), The error in the quality metrics output by the current shared critic network; pass Update the internal parameters of the shared critics network. ; The target commentator network update formula is as follows: ，（22） In equation (22), The learning rate is updated softly; Step S73: Set a threshold for the number of updates to the shared critic network; when the number of updates reaches the threshold, update the global agent actor network and the target actor network using the quality index of the target critic network; for each agent actor network, the update formula is as follows: ，（23） In equation (23), For the current number A network of intelligent agents, For the present Network internal parameters; The number of samples in the training set; The actor network update formula for each rack in the target actor network is as follows: ，（24）； Step S74: Determine whether the iteration threshold has been reached; if it has, output the global agent actor network; if not, return to step S71.

10. A simulation control system for a multi-mill cluster in finishing mills based on attention mechanisms and mechanistic constraints, characterized in that, The system includes a simulation model construction module, a local and global state generation module, an agent actor network construction module, a state and action space construction module, a shared critic network construction module, a multi-objective optimization function construction module, a network hyperparameter optimization module, and a distributed execution control module; among which... The simulation model building module is used to build a multi-machine cluster simulation model of the finishing mill unit that includes all finishing mill stands and AGC-related components. The local and global state generation modules are used to run the simulation model, generate corresponding local states for each rack, and construct the global state from the local states of all racks. The intelligent agent actor network construction module is used to build an independent intelligent agent actor network for each rack, and all racks form a global intelligent agent actor network; the structure of each intelligent agent actor network includes an input layer, a normalization layer, a rack feature extraction layer and a fully connected layer; the input of the intelligent agent actor network is the local state of the multi-machine cluster simulation model, and the output is PID control parameters; The state and action space construction module is used to construct a global state space based on the global state of all rack simulation models, and to construct a global action space based on the PID control parameters output by the intelligent agent actor network. The shared critic network construction module is used to construct a shared critic network for all the intelligent agent actor networks in all racks; the structure of the shared critic network includes a quality feature extraction layer and a fully connected layer, the input is the global state space and the global action space, and the output is the quality index of the global state and the global action; The multi-objective optimization function construction module is used to construct multi-objective optimization functions based on the process mechanism constraints of the multi-machine cluster simulation model of the finishing mill. The network hyperparameter optimization module is used to run the simulation model and record sampled data, and construct a training set in the global state space and global action space; preset the control indicators of the finishing mill, establish the optimization direction of the multi-objective optimization function, and construct the target actor network and the target critic network; based on the training set, the target actor network, and the target critic network, train the global agent actor network and the shared critic network, and optimize the hyperparameters of the global agent actor network and the shared critic network. The distributed execution control module is used to deploy the trained global agent network to the workstations in the corresponding racks, collect the local state of each rack during real-time operation, and input the corresponding agent network to calculate the corresponding PID control parameters.

Images