Multi-process collaborative control method for keel production line based on digital twin

Abstract

This invention relates to the field of process control technology, specifically to a multi-process collaborative control method for a keel production line based on digital twins. The method includes collecting physical operation data such as local tension between frames, real-time current of servo motors, and strip cross-sectional springback deviation during punching and rolling processes; inputting the physical operation data into a graph-temporal attention network to extract physical feature vectors; and using a digital twin inference model to calculate the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing; inputting the cumulative longitudinal tensile deformation as a feedforward compensation parameter into the flying shear process for timing compensation, and updating the phase compensation parameters to reconstruct the timing sequence of the flying shear cut. This invention solves the problem of severe phase misalignment between the end-process execution action commands and the actual dynamic displacement of the strip in large-scale multivariable coupled scenarios by globally collaboratively extracting implicit system disturbances across processes and establishing a dynamic feedforward compensation mechanism.

Description

Technical Field

[0001] This invention relates to the field of process control technology, specifically to a multi-process collaborative control method for a keel production line based on digital twins. Background Technology

[0002] The continuous production of light steel keel is a typical large-scale, multi-variable coupled industrial control scenario, encompassing multiple continuous execution processes such as punching, rolling, and end-effector flying shearing. In existing industrial automation control systems, such multi-process production lines typically employ a distributed or decoupled control architecture, where the servo motion control subsystems of each process operate relatively independently, primarily relying on local closed-loop control loops (such as single encoder position feedback or speed feedback) to track setpoints.

[0003] However, this distributed control system neglects the strong coupling characteristics of multiple processes in continuous production. During the continuous drive of multi-stage servo motors, complex dynamic physical disturbances (such as tension fluctuations and cumulative deformation across processes) are generated within the system. These system disturbances have significant time delays and spatial cumulative effects, and belong to implicit state variables in the control system. Existing control architectures lack a collaborative interaction mechanism for state data between multiple processes, and cannot effectively extract and synchronously transmit these cross-process dynamic disturbances within the global control domain.

[0004] Because the control loop lacks a feedforward compensation mechanism for upstream cumulative state disturbances, the motion controller of the end-effector can only issue motion trigger commands based on the kinematic model under ideal steady state or a single reference signal. This system-level information silo causes a significant phase misalignment between the timing of the motion commands from the underlying servo drive and the actual dynamic displacement state of the controlled object. The control system is unable to adaptively reconstruct the motion timing based on upstream coupled disturbances, ultimately leading to deviations in the dynamic response of the end-effector control system and a severe decrease in execution accuracy.

[0005] To address this, a multi-process collaborative control method for keel production lines based on digital twins is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-process collaborative control method for a keel production line based on digital twins, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A multi-process collaborative control method for keel production lines based on digital twins includes: Collect physical operation data in the keel production line process, which includes punching and rolling processes. The physical operation data includes local tension between frames, real-time current of servo motors, and springback deviation of strip cross-section. The physical operation data is input into the graph spatiotemporal attention network, the punching and rolling processes are set as network topology nodes, and the normalized local tension between the frames is used as the initial edge weight between the nodes. The physical feature vector is extracted by combining the real-time current of the servo motor and the springback deviation of the strip cross section. The physical feature vector is input into the digital twin inference model to calculate the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing. The cumulative longitudinal tensile deformation is input as a feedforward compensation parameter into the flying shear process. The feedforward compensation parameter is used to perform time-series compensation for the cumulative deformation error of the strip and update the phase compensation parameter of the flying shear cutting action. Based on the updated phase compensation parameter, a flying shear trigger command is sent to the motion controller of the flying shear process to reconstruct the timing sequence of the flying shear cutting action.

[0008] Preferably, the physical operation data acquisition process involves: collecting local tension between the frames by using tension sensors deployed between adjacent frames in the punching and rolling processes; collecting the real-time current of the servo motor driving the punching and rolling equipment by reading the external current transformer of the servo driver; acquiring the three-dimensional cross-sectional profile data of the strip by using a laser profile scanner installed at the discharge end of the rolling process; and comparing the three-dimensional cross-sectional profile data with a preset standard digital three-dimensional model to extract the springback deviation of the strip cross-section.

[0009] Preferably, the graph spatiotemporal attention network includes a physical topology encoding module, a spatiotemporal feature joint extraction module, and a deductive feature alignment module connected in sequence. The physical topology encoding module maps the punching and each level of the roll forming machine stand to network topology nodes, obtains the real-time transmission line speed of the strip, calculates the quotient of the physical spatial distance between nodes and the real-time transmission line speed, and defines the quotient as a time delay parameter. The local tension between the stands is normalized to the time delay parameter, and a physical attenuation factor is constructed to nonlinearly reduce the initial edge weights to generate a dynamic adjacency matrix. The spatiotemporal feature joint extraction module generates spatial feature vectors and temporal feature vectors; the inference feature alignment module concatenates the spatial feature vectors and the temporal feature vectors into a tensor, performs a fully connected layer linear mapping on the concatenated tensor, and outputs a physical feature vector.

[0010] Preferably, the spatiotemporal feature joint extraction module includes a spatial attention unit and a temporal attention unit; the spatial attention unit calculates the steady-state fluctuation phase difference of the real-time current of the servo motors of the upstream and downstream racks and defines it as a penalty term, updates the dynamic adjacency matrix using the penalty term, calculates the spatial attention distribution weight based on the updated dynamic adjacency matrix, and outputs a spatial feature vector; the temporal attention unit constructs a variable-step long-term sliding window to sample historical operating data, extracts the temporal change rate of the strip cross-section springback deviation, maps the temporal change rate to a temporal attention score, and outputs a temporal feature vector based on the temporal attention score.

[0011] Preferably, the digital twin inference model maps the physical feature vector into a state-space equation containing the physical operation data, iteratively calculates the strain increment of a single process based on the discrete processing time of each level of the roll press stand, corrects the strain increment of the single process using a nonlinear slip function based on the principle of volume invariance, performs spatial integration in the direction of strip movement, and outputs the initial deformation amount; inputs the observation residual of the state-space equation into the Kalman filter algorithm to output the state correction value, uses the state correction value to perform time-domain compensation on the initial deformation amount, and outputs the cumulative longitudinal tensile deformation amount.

[0012] Preferably, the timing compensation involves adding the cumulative longitudinal tensile deformation to the reference cutting length of the flying shear to obtain the actual expected displacement of the strip; inputting the actual expected displacement of the strip and the real-time feed line speed into the flying shear electronic cam model; using the time offset corresponding to the cumulative longitudinal tensile deformation as a reference, shifting the cutting phase of the synchronous shearing zone in the flying shear electronic cam model to generate a compensated asymmetric electronic cam curve; and based on the compensated asymmetric electronic cam curve, calculating the mechanical starting point offset in reverse when the flying shear servo cutter axis reaches the cutting stage, and updating the mechanical starting point offset to the phase compensation parameters of the flying shear cutting action.

[0013] Preferably, the flying shear electronic cam model is constructed as a piecewise mapping function with the actual expected displacement of the strip as the main axis and the rotational phase angle of the flying shear shaft as the secondary axis, and is divided into an acceleration zone, a synchronous shearing zone, a deceleration zone, and a return zone in sequence. In the synchronous shearing zone, the horizontal component of the tangential linear velocity of the flying shear shaft is constrained to be consistent with the real-time feed linear velocity. The cumulative longitudinal tensile deformation is used as the position compensation value. Under the condition of keeping the total periodic displacement of the main shaft constant and the acceleration continuous, the position compensation value is subtracted from the main shaft displacement span in the acceleration zone, and the position compensation value is increased by an equal amount in the main shaft displacement span in the deceleration zone. The motion curve is asymmetrically reconstructed using a polynomial, and a compensated asymmetrical electronic cam curve is output.

[0014] Preferably, the reconstructed flying shear cutting action timing process involves encapsulating the updated phase compensation parameters with the real-time feed line speed of the strip to generate a flying shear trigger command; sending the flying shear trigger command to the motion controller of the flying shear process; establishing a fixed interpolation calculation cycle with the inherent communication frequency of the underlying hardware of the motion controller; and constructing a mapping space with the strip travel distance as the virtual main axis. Within each interpolation operation cycle, the theoretical displacement increment is calculated based on the real-time feed line speed, and the theoretical displacement increment, the virtual spindle position of the previous cycle, and the phase compensation parameter as the position offset are dynamically summed to generate a new spindle position. The new spindle position is updated in the underlying instruction register, and after interpolation and differential operations, it is converted into a feedforward drive instruction to control the flying shear servo motor to perform shearing according to the compensated action phase offset.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention breaks down the information silos of traditional distributed or decoupled control architectures by inputting the physical operation data of multiple processes in a keel production line into a graph spatiotemporal attention network, setting the punching and rolling processes as network topology nodes, and constructing a dynamic adjacency matrix using time delay parameters. This design globally and collaboratively extracts implicit system disturbances with time-delay and spatiotemporal coupling characteristics, such as local tension between frames, servo motor current, and strip springback deviation, achieving collaborative interaction of multi-process state data and effectively overcoming the shortcomings of existing independent closed-loop control systems that cannot perceive and extract cross-process dynamic physical disturbances.

[0016] 2. This invention constructs a digital twin inference model, maps the extracted physical feature vectors into state-space equations, and combines a nonlinear slip function based on the principle of volume invariance with a Kalman filter algorithm for spatial integration and system state observation correction. This establishes a solution mechanism for continuous stress deformation of strip. This mechanism can calculate the cumulative longitudinal tensile deformation of strip caused by continuous multi-stage cross-process processing, fundamentally filling the technical gap in the existing control loop that lacks a feedforward compensation link for upstream cumulative state disturbances, and providing reliable dynamic compensation parameters for the timing synchronization of subsequent processes.

[0017] 3. This invention introduces the calculated cumulative longitudinal tensile deformation as a feedforward position compensation value into the flying shear electronic cam model to generate a compensated asymmetric electronic cam curve. In the master-slave axis synchronization control module, the phase compensation parameters are dynamically superimposed onto the virtual master axis position of the flying shear servo, adaptively reconstructing the action timing of the flying shear cutting. This strategy completely solves the problem of severe phase misalignment between the timing of the end-process execution action instructions and the actual dynamic displacement state of the strip in large-scale multivariable coupling scenarios, realizing dynamic action timing synchronization based on feedforward compensation. Attached Figure Description

[0018] Figure 1 A schematic diagram of a multi-process collaborative control method for a keel production line based on digital twins; Figure 2 This is a schematic diagram of the physical operation data acquisition process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the digital twin inference model solution process provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a multi-process collaborative control method for a keel production line based on digital twins, the technical solution of which is as follows: A multi-process collaborative control method for keel production lines based on digital twins includes: Collect physical operation data in the keel production line process, which includes punching and rolling processes. The physical operation data includes local tension between frames, real-time current of servo motors, and springback deviation of strip cross-section. The physical operation data is input into the graph spatiotemporal attention network, the punching and rolling processes are set as network topology nodes, and the normalized local tension between the frames is used as the initial edge weight between the nodes. The physical feature vector is extracted by combining the real-time current of the servo motor and the springback deviation of the strip cross section. The physical feature vector is input into the digital twin inference model to calculate the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing. The cumulative longitudinal tensile deformation is input as a feedforward compensation parameter into the flying shear process. The feedforward compensation parameter is used to perform time-series compensation for the cumulative deformation error of the strip and update the phase compensation parameter of the flying shear cutting action. Based on the updated phase compensation parameter, a flying shear trigger command is sent to the motion controller of the flying shear process to reconstruct the timing sequence of the flying shear cutting action.

[0021] Example 1: Collect physical operation data in the keel production line process, which includes punching and rolling processes. The physical operation data includes local tension between frames, real-time current of servo motors, and springback deviation of strip cross-section. See Figure 2The physical operation data acquisition process involves collecting local tension between the frames by using tension sensors deployed between adjacent frames in the punching and rolling processes; collecting the real-time current of the servo motor driving the punching and rolling equipment by reading the external current transformer of the servo driver; acquiring the three-dimensional cross-sectional profile data of the strip by using a laser profile scanner installed at the discharge end of the rolling process; and comparing the three-dimensional cross-sectional profile data with a preset standard digital three-dimensional model to extract the springback deviation of the strip cross-section.

[0022] Specifically, the process of collecting the local tension between the frames includes the adjacent first and second frames in the punching and rolling processes. rack and the A wire-passing mechanism with guide rollers is installed between the frames, and the tension sensor is deployed below the bearing housing of the guide roller. When the strip is continuously conveyed between adjacent frames, the longitudinal tension of the strip applies downward radial pressure to the guide roller. The tension sensor senses this radial pressure through an internal resistance strain gauge and outputs an analog voltage signal. The analog voltage signal is converted from analog to digital by a high-frequency data acquisition card to obtain a digital voltage value. The static calibration parameters are then used to calculate the actual tension value of the strip between the current adjacent frames, which is then used as the local tension between the frames. The formula for calculating the local tension between the frames is as follows:

[0023] in, The zero-point voltage value of the sensor when there is no strip between the racks, the static calibration parameters The method of obtaining the data is to suspend a standard weight of known mass onto the strip between the frames while the equipment is stopped, record the corresponding sensor output voltage increment, and calculate the static calibration parameters by fitting the linear ratio between the voltage increment and the gravity of the standard weight. The process of acquiring the real-time current of the servo motors driving the punching and rolling equipment involves installing external current transformers on the three-phase power supply lines of all servo motors driving the punching and rolling equipment. During the operation of the servo motors driving the frame rolls, the AC current flowing through the three-phase stator windings induces a proportional secondary current signal on the secondary side of the external current transformers. This secondary current signal is then amplified and low-pass filtered by a signal conditioning circuit to remove high-frequency electromagnetic interference noise before being input to the analog input module of the programmable logic controller (PLC). The PLC discretizes the filtered signal and performs amplitude reconstruction calculations based on the fixed turns ratio of the external current transformers to obtain the instantaneous amplitude sequence of the stator current of each stage of the servo motors in the punching and rolling equipment as the real-time current of the servo motors. Three sets of laser profile scanners are installed in a ring array above and on both sides of the strip's running trajectory at the output end of the rolling process. When the strip passes through the scanning area at a constant speed, the three sets of laser profile scanners synchronously emit linear laser beams to irradiate the upper surface and both sides of the strip. The diffuse reflected light is received by the photoelectric sensing element inside the scanner, and the spatial three-dimensional coordinates of each profile reflection point on the strip surface in the coordinate system of the independent scanner are calculated using the triangulation principle. Using the spatial coordinate system transformation matrix, the local coordinate data collected by the three sets of scanners are stitched and merged into a unified global coordinate system to generate the complete actual three-dimensional cross-sectional point cloud data of the currently output strip as the three-dimensional cross-sectional profile data. Read the computer-aided design drawings from the strip processing technology design stage, extract the theoretical cross-sectional geometric feature parameters of the target keel product, including theoretical web width, theoretical flange height, and theoretical bending angle; generate a standard cross-sectional point cloud composed of uniformly distributed discrete spatial points in the three-dimensional parametric software based on the above theoretical cross-sectional geometric feature parameters, and use it as the preset standard digital three-dimensional model; use the nearest point iterative spatial registration algorithm to spatially align the three-dimensional cross-sectional contour data (i.e., the actual three-dimensional cross-sectional point cloud data) with the preset standard digital three-dimensional model in the global coordinate system through translation and rotation operations, and obtain the aligned actual point cloud and standard point cloud; For any point in the aligned actual point cloud located within the keel fracture bend feature region. Traverse the standard point cloud to find the corresponding point that is closest to it in Euclidean distance. ,calculate and The spatial Euclidean distance between them; extract the spatial Euclidean distances corresponding to each point within the characteristic area of ​​all bending angles of the keel section, and calculate the arithmetic mean of these spatial Euclidean distances, using the arithmetic mean as the springback deviation of the strip section; The physical operation data is input into the graph spatiotemporal attention network, the punching and rolling processes are set as network topology nodes, and the normalized local tension between the frames is used as the initial edge weight between the nodes. The physical feature vector is extracted by combining the real-time current of the servo motor and the springback deviation of the strip cross section. The graph-temporal attention network includes a physical topology encoding module, a spatiotemporal feature joint extraction module, and a deductive feature alignment module connected in sequence. The physical topology encoding module maps the punching and rolling mill stands at each level as network topology nodes, obtains the real-time transmission line speed of the strip, calculates the quotient between the physical spatial distance between nodes and the real-time transmission line speed, and defines the quotient as a time delay parameter. The local tension between the stands is normalized and the time delay parameter is normalized. The normalized local tension between the stands is used as the initial edge weight. The time delay parameter is used to construct a physical attenuation factor, and the initial edge weight is nonlinearly reduced to generate a dynamic adjacency matrix. The spatiotemporal feature joint extraction module generates spatial feature vectors and temporal feature vectors; the inference feature alignment module concatenates the spatial feature vectors and the temporal feature vectors into a tensor, performs a fully connected layer linear mapping on the concatenated tensor, and outputs a physical feature vector. The spatiotemporal feature joint extraction module includes a spatial attention unit and a temporal attention unit. The spatial attention unit calculates the steady-state fluctuation phase difference of the real-time current of the servo motors of the upstream and downstream racks as a penalty term, updates the dynamic adjacency matrix using the penalty term, calculates the spatial attention distribution weight based on the updated dynamic adjacency matrix, and outputs a spatial feature vector. The temporal attention unit constructs a variable-step long-term sliding window to sample historical operating data, extracts the temporal change rate of the strip cross-section springback deviation, maps the temporal change rate to a temporal attention score, and outputs a temporal feature vector based on the temporal attention score. Specifically, according to the actual processing flow of the strip, the punching process and the subsequent multi-stage independent roll forming frames are numbered sequentially to construct a network topology node set with a sequential order. ,in The total number of frames; the rotational angular velocity of the main shaft is collected in real time by an incremental photoelectric encoder installed on the main drive shaft of the roll forming equipment, and combined with the actual outer diameter of the main drive roller, the real-time linear velocity of the strip passing through the frame is calculated. ; Extracting adjacent 3D models of equipment manufacturing Node and the The fixed physical space distance between the central axes of the racks corresponding to the nodes, wherein the quotient of the fixed physical space distance and the real-time transmission line speed is the distance from the first node to the second node. rack transfer to the first The physical travel time required for the rack is defined as the time delay parameter. ; The maximum-minimum normalization method was used to analyze the local tension between the racks in the collected continuous time series. and the time delay parameter Perform linear scaling to obtain normalized inter-rack local tension. and normalized time delay parameter The normalized inter-rack local tension is used as the initial edge weight of the connectivity link between corresponding node pairs in the network topology; with the natural base... Using the normalized time delay parameter as the base, A physical attenuation factor is constructed to nonlinearly reduce the initial edge weights, as shown in the formula: in, These are elements in the dynamic adjacency matrix. For the first OK, For the first The column calculates the corresponding element by traversing all adjacent node pairs in the network topology. Then, a matrix array is arranged to generate a dynamic adjacency matrix.

[0024] The spatial attention unit collects data on adjacent upstream and downstream units. rack and the The real-time current sequence of the servo motor of the rack is analyzed in the frequency domain using Fast Fourier Transform. The phase angles of the two current sequences at the equipment's dominant operating frequency are extracted, and the absolute difference between the two phase angles is calculated and defined as the penalty term. The penalty term is transformed into an inhibition factor using an exponential decay function. The dynamic adjacency matrix is ​​multiplied by the inhibition factor to obtain the updated dynamic adjacency matrix elements. The input feature vectors of each rack node and the updated dynamic adjacency matrix are input into the spatial attention calculation function of the graph convolutional network layer. The spatial attention distribution weight is calculated by weighted summation of the features of topological neighbor nodes, and the spatial neighborhood information of each node is aggregated to output a spatial feature vector that integrates the physical coupling interference features of multiple processes. A window step size adjustment function is established based on the real-time transmission linear velocity. When the linear velocity is large, the sampling time step of the sliding window is reduced proportionally; when the linear velocity is small, the sampling time step of the sliding window is increased proportionally. A continuous historical data sequence, including the springback deviation of the strip cross-section, is extracted according to this variable step size rule. The springback deviation values ​​of the strip cross-section at adjacent sampling times within the sliding window are differentiated using a first-order backward difference algorithm to obtain the temporal change rate, which characterizes the severity of the springback deviation evolution. The temporal change rate is input into the Softmax activation function for normalized probability distribution calculation to obtain a time attention score. The time attention score is then used to weight and sum the historical time step feature sequences within the sliding window to output a time feature vector. The spatial feature vector output by the spatial attention unit and the temporal feature vector output by the temporal attention unit are concatenated and spliced ​​to generate a joint tensor that fuses the spatiotemporal dual-domain dimensions; a fully connected network composed of multilayer perceptrons is constructed, and the learnable weight matrix and bias vector inside the fully connected layer network are obtained and loaded; the joint tensor is input into the fully connected network layer, and linear mapping calculation is performed to generate physical feature vectors; By constructing a graph-temporal attention network, the local tension and time delay between racks are deeply integrated into a dynamic adjacency matrix, which accurately maps the real physical coupling and delay characteristics of continuous strip transmission across racks. Furthermore, in the spatial dimension, the phase difference of upstream and downstream motor currents is used as a penalty term to update the adjacency matrix, which effectively blocks the spread of local abnormal fluctuations and enhances the anti-interference ability of feature extraction. At the same time, in the temporal dimension, a variable step-size sliding window mechanism based on real-time linear velocity is introduced, and the time attention weight is dynamically allocated in combination with the temporal change rate of cross-sectional springback deviation, so that the model can adaptively focus on the transient distortion of material internal stress release.

[0025] The physical feature vector is input into the digital twin simulation model to calculate the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing. See Figure 3 The digital twin inference model maps the physical feature vector into a state-space equation containing the physical operation data. Iteratively calculates the strain increment of a single process based on the discrete processing time of each level of the roll press stand. The strain increment of the single process is corrected using a nonlinear slip function based on the principle of volume invariance, and spatial integration is performed in the direction of strip movement to output the initial deformation. The observation residual of the state-space equation is input into the Kalman filter algorithm to output the state correction value. The state correction value is used to perform time-domain compensation on the initial deformation to output the cumulative longitudinal tensile deformation.

[0026] Specifically, the physical feature vector is mapped and transformed into initial prior values ​​containing the physical operation data, and a three-dimensional real observation vector including local tension between frames, real-time current of the servo motor, and springback deviation of the strip cross-section is established. Based on the elastoplastic mechanical transmission relationship between adjacent roller press frames, the physical mapping logic inside the state space equation is constructed: in the state space equation, a self-holding constant characterizing strain and an exponential mapping term characterizing the decay of tension and stress with physical travel time are configured, and the reciprocal of the material stiffness (the product of the material's elastic modulus and real-time cross-sectional area) is used as a coupling parameter to establish the physical mapping from tension to strain; simultaneously, a torque equivalent coupling coefficient for the mapping of strain increment and internal tension to external servo motor current is configured in the state space equation, completing the mapping construction of the state space equation. By acquiring the rotary encoder pulse signals of the spindle servo motors of each level of the roll press, the precise timestamps of the strip as it passes through the bite point of each level of the roll press are extracted, and the precise timestamps are defined as discrete processing moments. In the state space equation, taking the derived state variables of the previous discrete processing moment as the starting point, the local tension between the current level of the roll press and the previous level of the press and the real-time current of the servo motor at the corresponding moment are extracted, and the force change value of the strip at the current discrete processing moment is calculated. The force change value is input into the deformation conversion logic based on elastoplastic mechanics to solve the longitudinal elongation change value caused by the passive stretching of the strip between the current two adjacent levels of the press. The longitudinal elongation change value is used as the strain increment of a single process, and iterative calculation is performed level by level as the strip moves forward between each level of the press. The deformation conversion logic based on elastoplastic mechanics is as follows: To calculate the single-process strain increment of the strip caused by passive stretching between adjacent frames, it is necessary to comprehensively consider both the elastic deformation of the material and the plastic hardening deformation beyond the yield point. The specific expression is as follows: in, For strip material in the current first The single-process strain increment generated between the stage and the previous stage frame, The tension fluctuation is calculated by using the local tension between the racks and the servo current. For the strip to enter the first Real-time cross-sectional area in front of the rack. To measure the elastic modulus of the metal strip material, To test the yield strength of the obtained metal strip material, The plasticity strengthening coefficient of the material. The strain hardening index of the material was measured. The maximum value function is used as a nonlinear activation function in the digital twin simulation model to characterize the yield determination condition of the material and perform feature truncation. The initial cross-sectional area of ​​the strip before entering the current rolling process and the real-time cross-sectional area after passing through the current rolling mill frame are collected. Based on the physical principle that the volume of the material remains constant before and after deformation during metal plastic forming (i.e., the principle of constant volume), the theoretical elongation coefficient is calculated to compensate for the elongation in the length direction due to the reduction in the cross-sectional area of ​​the strip. At the same time, the rotational linear velocity of the roll surface and the actual transmission linear velocity of the strip surface are extracted, and the velocity difference between the two is calculated to construct a nonlinear slip function characterizing the degree of relative sliding between the strip and the roll surface. The theoretical elongation coefficient and the nonlinear slip function are used as correction factors and multiplied with the single-process strain increment calculated iteratively to obtain the true process strain increment after eliminating the slip error. A one-dimensional spatial coordinate system is established along the movement trajectory of the strip with the position of the strip's first end entering the production line as the origin. The true process strain increment is discretely accumulated and summed in the intervals of each frame in the one-dimensional spatial coordinate system to complete the spatial integration operation. The total elongation of the strip obtained by the accumulation and summation is output as the initial deformation amount. The expressions for the theoretical extension coefficient and the nonlinear slip function are as follows:

[0027]

[0028] The expression for the actual process strain increment after incorporating the correction factor is:

[0029] in, This represents the actual process strain increment after both theoretical extension and surface slip correction. This is the theoretical extension coefficient. It is a nonlinear slip function. This refers to the initial cross-sectional area of ​​the strip before it enters the current rolling process. This represents the real-time cross-sectional area of ​​the strip after it passes through the current roll press stand. The linear velocity of the current stand roll surface. This represents the actual linear speed of transmission on the surface of the strip in the current rack. This is the interfacial friction slip coefficient calibrated experimentally.

[0030] At each discrete processing stage, the prior predicted value of the physical state of the strip is calculated using the state space equation, and the actual measured value of the physical operation data corresponding to that time is read synchronously. The difference between the actual measured value and the prior predicted value is calculated and defined as the observation residual. The observation residuals of the state-space equations are input into the Kalman filter algorithm, and the covariance update logic inside the algorithm is used to perform closed-loop recursive updates to calculate the state correction value after eliminating time-varying noise interference. The total running time of the strip in this processing is extracted and divided into continuous discrete time steps. The state correction value is dynamically superimposed on the initial deformation according to the corresponding time step span to complete the time-domain compensation operation. The longitudinal overall tensile length of the strip after time-space dual correction is output as the cumulative longitudinal tensile deformation. The adaptively established dynamic process noise covariance matrix, the measurement noise covariance matrix, and the observation residual are simultaneously fed into the Kalman filter algorithm to calculate the Kalman gain and perform closed-loop recursive updates, thereby solving for the state correction value after eliminating time-varying noise interference. The total running time of the strip in this processing process is extracted and divided into continuous discrete time steps. The state correction value is dynamically superimposed on the initial deformation amount according to the corresponding time step span to complete the time-domain compensation operation. The longitudinal overall tensile length of the strip after spatiotemporal double correction is output as the cumulative longitudinal tensile deformation amount.

[0031] By establishing state-space equations and combining them with iterative solutions at discrete processing times, a deep fusion of physical mechanisms and dynamic data was achieved, characterizing the transient stress-deformation process of the strip. At the same time, a theoretical elongation coefficient and a nonlinear slip function based on the principle of volume invariance were introduced to eliminate non-ideal interferences caused by strip slippage and cross-sectional reduction, ensuring the physical accuracy of the spatial dimension strain increment calculation. Kalman filtering was used to perform closed-loop feedback and time-domain compensation of the observation residuals, overcoming the temperature drift and cumulative errors caused by long-term model operation.

[0032] The cumulative longitudinal tensile deformation is input into the flying shear process as a feedforward compensation parameter. The feedforward compensation parameter is used to perform time-series compensation for the cumulative deformation error of the strip and update the phase compensation parameter of the flying shear cutting action. The timing compensation involves adding the cumulative longitudinal tensile deformation to the reference cutting length of the flying shear to obtain the actual expected displacement of the strip; inputting the actual expected displacement of the strip and the real-time feed line speed into the flying shear electronic cam model; using the time offset corresponding to the cumulative longitudinal tensile deformation as a reference, shifting the cutting phase of the synchronous shearing zone in the flying shear electronic cam model to generate a compensated asymmetric electronic cam curve; and based on the compensated asymmetric electronic cam curve, calculating the mechanical starting point offset in reverse when the flying shear servo cutter axis reaches the cutting stage, and updating the mechanical starting point offset to the phase compensation parameters of the flying shear cutting action.

[0033] The flying shear electronic cam model is constructed using a piecewise mapping function with the actual expected displacement of the strip as the master axis and the rotational phase angle of the flying shear shaft as the slave axis, sequentially divided into an acceleration zone, a synchronous shearing zone, a deceleration zone, and a return zone. Within the synchronous shearing zone, the horizontal component of the tangential linear velocity of the flying shear shaft is constrained to maintain consistency with the real-time feed linear velocity. The cumulative longitudinal tensile deformation is used as a position compensation value. While maintaining a constant total periodic displacement of the master shaft and continuous acceleration, the position compensation value is subtracted from the master shaft displacement span in the acceleration zone, and the position compensation value is increased by an equal amount in the master shaft displacement span in the deceleration zone. A polynomial is used to asymmetrically reconstruct the motion curve, outputting a compensated asymmetric electronic cam curve.

[0034] Specifically, by reading the current batch production order data issued by the production line manufacturing execution system, the reference cutting length of the flying shear specified in the order is extracted; the cumulative longitudinal tensile deformation output by the digital twin model is extracted simultaneously; and an addition operation is performed in the logic operation unit of the motion controller to sum and merge the values ​​of the reference cutting length of the flying shear and the cumulative longitudinal tensile deformation to calculate the actual expected displacement of the strip that includes the influence of the material's actual tensile deformation. A high-precision incremental encoder deployed on the measuring spindle in front of the flying shear inlet is used to collect pulse signals generated by the material transmission driving the measuring spindle to rotate. The frequency of the pulse signals is calculated and combined with the outer diameter circumference of the measuring roller to obtain the real-time feed line speed. The calculated actual expected displacement of the strip and the real-time feed line speed are used as basic motion addressing variables and written into the data interface corresponding to the flying shear electronic cam model. A two-dimensional data table is established in the internal storage space of the flying shear controller. The actual expected displacement of the strip is configured as a horizontal virtual master axis variable, and the rotation phase angle, which represents the physical position of the flying shear servo cutter axis and is fed back by the absolute encoder of the servo motor, is configured as a vertical slave axis variable. According to a complete 360° mechanical rotation cycle of a single cutting operation of the flying shear, based on the physical process of the cutter cutting the strip, the total displacement range of the master axis is divided into four continuous action segments. The first segment is set as the acceleration zone, which controls the cutter axis to accelerate and follow the material from the initial stationary position. The second segment is set as the synchronous shearing zone, which controls the blade to accurately bite, penetrate, and cut the strip. The third segment is set as the deceleration zone, which controls the cutter to gradually reduce its speed after leaving the strip. The fourth segment is set as the return zone, which controls the cutter axis to smoothly return to the initial standby position for the next cycle. For each of these four segments, a corresponding set of position points between the master and slave axes is planned to form a piecewise mapping function with conditional branching. The method for determining different sections in the flying shear electronic cam model is as follows: in the underlying logic operation unit of the motion controller, the total running displacement span value of the main shaft within a complete cutting cycle is extracted. The actual expected displacement of the strip calculated in real time is then moduloed by this total running displacement span value to obtain the real-time cumulative displacement of the main shaft within the current cutting cycle. Simultaneously, the boundary node data of each section divided when the segment mapping function was established are read, and the coordinates of the end point of the acceleration zone, the end point of the synchronous shearing zone, and the end point of the deceleration zone are extracted in sequence. The specific implementation process of determining the segment state based on the real-time periodic cumulative displacement is as follows: In each cyclic scanning cycle of the motion controller, the real-time periodic cumulative displacement is continuously read and the numerical range is compared; when it is determined that the real-time periodic cumulative displacement is greater than or equal to zero and less than the coordinate of the end point of the acceleration zone, it is confirmed that the flying shear axis is in the acceleration zone, and the following acceleration control logic is triggered; when it is determined that the real-time periodic cumulative displacement is greater than or equal to the coordinate of the end point of the acceleration zone and less than the coordinate of the end point of the synchronous shearing zone, it is confirmed that the flying shear axis has entered the synchronous shearing zone, and the internal feedforward speed tracking closed loop is activated to lock the horizontal component of the tangential linear velocity; when it is determined that the real-time periodic cumulative displacement is greater than or equal to the coordinate of the end point of the synchronous shearing zone and less than the coordinate of the end point of the deceleration zone, it is confirmed that the flying shear axis has detached from the strip and entered the deceleration zone, and a deceleration command is issued to dissipate the kinetic energy of the cutter axis; when it is determined that the real-time periodic cumulative displacement is greater than or equal to the coordinate of the end point of the deceleration zone and less than the total running displacement span value, it is confirmed that the flying shear axis is in the return zone, and the cutter axis is driven to rotate smoothly to the angle position indicated by the mechanical starting point offset, waiting for the next cutting cycle to arrive and the value is cleared to zero. When determining the action range corresponding to the synchronous shearing zone, the angular velocity of the cutter shaft rotation is extracted and multiplied with the physical blade rotation radius to obtain the tangential linear velocity of the cutter shaft edge; simultaneously, the current real-time cutter shaft rotation phase angle is obtained, and the projection value of this tangential linear velocity in the horizontal conveying direction of the strip is calculated using cosine trigonometric function logic (because the horizontal projection is usually multiplied by the cosine of the included angle). The horizontal component of the tangential linear velocity is used as the basis for constructing a feedforward speed tracking control closed loop. The horizontal component of the tangential linear velocity is compared with the real-time feed linear velocity collected in real time. The output current of the flying shear servo driver is dynamically adjusted to force the speed values ​​of the two to be equal at all times in the entire synchronous shearing zone, so as to prevent the blade from tearing the strip at the moment of cutting or causing mechanical interference due to speed difference. Extract the data value of the cumulative longitudinal tensile deformation calculated in the previous step and define it as the position compensation value for dynamic shaping; extract the initial running length occupied by each of the four action areas on the main axis, and dynamically adjust the motion boundary without changing the total cumulative length of the main axis in a single cutting cycle; for the acceleration zone, perform a subtraction operation to deduct the position compensation value from its originally allocated main axis displacement span; for the deceleration zone, perform an addition operation to add an equal amount of the position compensation value to its originally allocated main axis displacement span; extract the initial position, termination position, running speed, and instantaneous acceleration value at the boundary node of each area as boundary constraint conditions, forcibly requiring that the speed and acceleration values ​​of two adjacent areas at the junction point are completely equal, ensuring that no torque change occurs during servo motor operation; Using fifth-order or seventh-order polynomial fitting logic, the principal axis displacement spans of each segment after the addition and subtraction span adjustment, along with all the previously extracted boundary constraints, are substituted into the polynomial equation system to solve for the corresponding coefficients of each order. Based on the solved coefficients, a discrete corresponding point matrix of principal axis displacement and slave axis angle is recalculated and generated. By applying reverse displacement span adjustments to the acceleration and deceleration zones, the symmetrical shape of the original electronic cam curve is broken, and the asymmetric reconstruction of the motion trajectory is completed. The reconstructed point matrix data table is output as the compensated asymmetric electronic cam curve. After performing asymmetric reconstruction, the starting reference point on the compensated asymmetric electronic cam curve where the cutter axis just enters the synchronous shearing zone and the horizontal component velocity of the cutter edge just matches and coincides with the feed line velocity is extracted and defined as the state node that has reached the cutting stage. Following the trajectory mapping table of the reconstructed compensated asymmetric electronic cam curve, starting from this state node, a reverse interpolation backtracking operation is performed along the direction of decreasing spindle displacement to solve for the initial rotation angle position that the slave axis must be in when the spindle displacement is zero. The default starting position angle in the initial state of the flying shear system is extracted, and the initial rotation angle position solved in reverse is subtracted from this default starting position angle to obtain the physical angle difference between the two. This physical difference is extracted as the mechanical starting offset. The mechanical starting offset is written into the underlying register of the motion controller, overwriting the original initial phase variable, and used as the phase compensation parameter for the flying shear cutting action. By converting the tensile deformation of the strip into an asymmetrical displacement span adjustment between the acceleration and deceleration zones of the electronic cam, dynamic feedforward compensation for the cumulative deformation error of the material is achieved while maintaining a constant total cutting cycle. Simultaneously, by forcibly constraining the horizontal tangential velocity component of the tool within the synchronous shearing zone to match the feed line speed, the mechanical interference and strip tearing problems caused by the speed mismatch between the two at the moment of cutting are resolved.

[0035] Based on the updated phase compensation parameters, a flying shear trigger command is sent to the motion controller of the flying shear process to reconstruct the timing sequence of the flying shear cut. The reconstructed flying shear cutting action timing process involves encapsulating the updated phase compensation parameters with the real-time feed line speed of the strip to generate a flying shear trigger command; sending the flying shear trigger command to the motion controller of the flying shear process; establishing a fixed interpolation calculation cycle with the inherent communication frequency of the underlying hardware of the motion controller; and constructing a mapping space with the strip travel distance as the virtual main axis. Within each interpolation operation cycle, the theoretical displacement increment is calculated based on the real-time feed line speed, and the theoretical displacement increment, the virtual spindle position of the previous cycle, and the phase compensation parameter as the position offset are dynamically summed to generate a new spindle position. The new spindle position is updated in the underlying instruction register, and after interpolation and differential operations, it is converted into a feedforward drive instruction to control the flying shear servo motor to perform shearing according to the compensated action phase offset.

[0036] Specifically, by reading the data register of the main control unit of the keel production line, the updated phase compensation parameters and the real-time feed line speed of the current strip are extracted; according to the standard format of the industrial Ethernet fieldbus communication protocol, an instruction data frame is constructed; the instruction data frame is divided into a frame header, a data segment, and a checksum; the updated phase compensation parameters are converted into position offset range data and written into the low byte area of ​​the data segment, the real-time feed line speed is converted into frequency range data and written into the high byte area of ​​the data segment, and the characteristic value of the data segment is calculated using a cyclic redundancy check algorithm and filled into the checksum. In the process, the flying shear trigger instruction is encapsulated and generated; the data segment is set as a continuous storage space with a preset fixed bit width (e.g., a 32-bit data register), and the total bit width of the data is divided into upper and lower parts, starting from the least significant bit of the storage space; the low bit width interval (e.g., bits 0 to 15) including the least significant bit is allocated and defined as the low byte region, and its corresponding memory starting address offset is small; the high bit width interval (e.g., bits 16 to 31) including the most significant bit is allocated and defined as the high byte region, and its corresponding memory starting address offset is large. The flying shear trigger command generated by the encapsulation is sent to the communication interface of the motion controller corresponding to the flying shear process through the industrial Ethernet fieldbus network in the form of isochronous synchronous communication, and the position offset range data and frequency range data are parsed out. The frequency of the clock pulse signal output by the motherboard crystal oscillator chip is read by calling the internal system function of the motion controller, and the frequency of the clock pulse signal is used as the inherent communication frequency of the underlying hardware. The interpolation operation period is set as the reciprocal of the inherent communication frequency of the underlying hardware. An independent single-dimensional coordinate system data block is opened in the memory of the motion controller, the initial feeding position of the strip is set as the absolute zero point of the single-dimensional coordinate system, the frequency range data parsed from the flying shear trigger command is multiplied by the real-time running time to convert it into the cumulative running distance, and the cumulative running distance is defined as the position coordinate value in the single-dimensional coordinate system, thus constructing a mapping space with the strip travel distance as the virtual main axis. Within the current interpolation operation cycle, according to the formula Calculate the theoretical displacement increment, where This is the theoretical displacement increment. The real-time feed line speed is... The fixed interpolation operation period is defined as follows: Under the premise that all data involved in the operation are uniformly converted into displacement scalar units, the operation is performed according to the formula... Perform dynamic summation calculation, where For the new spindle position, This is the virtual spindle position of the previous cycle. The phase compensation parameter, which is the position offset, is parsed from the flying shear trigger command. The new spindle position for the current cycle is generated by numerical superposition. The process for determining the virtual spindle position of the previous cycle is as follows: It is determined whether the current interpolation operation cycle is the first operation cycle after the flying shear trigger command is parsed. If it is the first operation cycle, the absolute zero coordinate value of the initial feed position of the strip set in the single-dimensional coordinate system data block is read, and the absolute zero coordinate value is used as the virtual spindle position of the previous cycle for the first operation cycle. If it is not the first operation cycle, the historical state cache register configured inside the motion controller is accessed, and the previous generation of new spindle position data, which has been dynamically summed and overwritten in the strictly adjacent previous interpolation operation cycle, is extracted and used as the virtual spindle position of the previous cycle for the current operation cycle. After all displacement calculations for the current interpolation operation cycle are completed, a data shift command is triggered to overwrite the newly generated spindle position of the current cycle into the historical state cache register.

[0037] The calculated new spindle position is overwritten into the underlying instruction register of the motion controller; the actual rotor position data fed back by the absolute encoder at the tail of the flying shear servo motor is read, and the new spindle position in the underlying instruction register is subtracted from the actual rotor position data to obtain the position tracking deviation; since the interpolation calculation cycle is greater than the internal current loop control cycle of the servo motor, a first-order linear interpolation algorithm is used to proportionally and discretely distribute the position tracking deviation to each internal current loop control cycle, and smoothly convert it into a continuous analog control voltage as a feedforward drive command to drive the rotor of the flying shear servo motor to accelerate or decelerate, so that the falling phase of the final cutting tool strictly follows the new spindle position after the phase compensation parameters are superimposed to perform the cutting action.

[0038] This application constructs a graph-temporal attention network, sets the punching and rolling processes as network topology nodes, and uses the local tension between frames as the initial edge weights. It extracts physical feature vectors by combining the real-time current of the servo motor and the springback deviation of the strip cross section. This breaks the information silos of the traditional distributed control architecture and realizes the global collaborative extraction of dynamic physical disturbances across processes. By inputting the physical feature vectors into a digital twin inference model, the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing is calculated, filling the technical gap of upstream cumulative state disturbance feedforward compensation in the control loop. The cumulative longitudinal tensile deformation is used as a feedforward compensation parameter input to the flying shear process. The timing of the flying shear cutting action is dynamically reconstructed using timing compensation, solving the problem of misalignment between the end action command and the actual dynamic displacement of the strip in large-scale multivariable coupled scenarios.

[0039] Example 2: This embodiment applies a multi-process collaborative control method for keel production lines based on digital twins to the production process of large-scale continuous automated U-shaped light steel keels with a thickness of 1.5mm and a target cut length of 4000mm. The baseline operating speed of the production line is set at 80m / min. The production line hardware sequentially includes a punching process, a continuously 14-stage independently driven roll forming frame, and a high-speed flying shear process at the end. The specific process is as follows: Tension sensors with guide rollers are deployed between the punching process and the first-stage roll forming stand, as well as between the subsequent 14 adjacent roll forming stands, to sense the radial pressure applied to the strip in real time and convert it into local tension between the stands. The real-time current of the servo motors of the punching and roll forming equipment is continuously sampled and calculated by reading the external current transformers that are respectively connected to the three-phase power supply lines of the servo motors driving the punching and 14-stage roll forming equipment. Three sets of laser profile scanners arranged in a ring array at the output end of the 14-stage roll forming process are used to synchronously scan and obtain the three-dimensional cross-sectional profile data of the output strip. The nearest point iteration algorithm is used to perform spatial alignment and distance calculation between the three-dimensional cross-sectional profile data and the pre-set U-shaped keel standard digital three-dimensional model in the global coordinate system, and the springback deviation of the strip cross-section with spatial attributes is extracted. The acquired inter-rack local tension, servo motor real-time current, and strip cross-sectional springback deviation are continuously input into a graph-temporal attention network for deep feature extraction. In the front-end physical topology encoding module of the network, the punching and 14-level rolling processes are mapped into 15 consecutive network topology nodes according to their physical sequence. The real-time strip transmission linear velocity of 80 meters per minute is obtained, and the quotient of the fixed physical distance between adjacent nodes and the real-time transmission linear velocity is calculated to obtain the time delay parameter of transmission between nodes. The real-time measured inter-rack local tension and time delay parameter are normalized to their maximum and minimum values. The normalized inter-rack local tension is used as the initial edge weight, and a negative exponential decay factor is constructed using the time delay parameter and the calibrated spatiotemporal attenuation coefficient. The initial edge weights are then physically reduced to generate a dynamic adjacency matrix. In the spatiotemporal feature joint extraction module, the steady-state fluctuation phase difference of the real-time current of the upstream and downstream rack servo motors under the dominant operating frequency is calculated as a penalty term to update the dynamic adjacency matrix. Based on this, the spatial attention distribution weight is calculated to output the spatial feature vector. At the same time, a variable step-size sliding window based on linear velocity adjustment is constructed to sample historical operating data, and the derivative is calculated to extract the temporal change rate of the strip cross-section springback deviation and mapped into a fractional output time feature vector. Finally, in the deduction feature alignment module, the spatial feature vector and the time feature vector are tensor-concatenated, and a fully connected layer linear mapping is performed to output a physical feature vector that integrates the physical coupling features of multiple processes. The output physical feature vector is used as the initial state parameter input into the digital twin inference model, and the physical feature vector is mapped into a state space equation containing physical operation data. Based on 15 discrete processing moments of the strip passing through punching and 14-stage roll forming stands, the single-process strain increment of the strip due to passive stretching between adjacent stands is iteratively calculated using elastoplastic mechanical deformation conversion logic. By collecting the real-time cross-sectional area of ​​the strip entering and exiting each stand and calculating the speed difference between the actual transmission linear velocity and the rotational linear velocity of the rolls, a nonlinear slip function based on the principle of constant volume and speed difference is constructed. This function is used to correct the single-process strain increment to eliminate slip error. Discrete summation (i.e., spatial integration) is performed on the trajectory coordinate system of the strip moving along the 80 m / min direction, and the initial deformation is output. Furthermore, the observation residual between the prior prediction value and the actual measurement value of the state-space equation at each discrete processing moment is calculated and input into the Kalman filter algorithm to calculate the state correction value. This state correction value is dynamically superimposed on the initial deformation value according to discrete time steps for time-domain closed-loop compensation, and the cumulative longitudinal tensile deformation of the 1.5 mm thick strip due to continuous processing across 15 strongly coupled processes is calculated. The calculated cumulative longitudinal tensile deformation is added to the 4000 mm flying shear reference cutting length in the target order to obtain the actual expected displacement of the strip, which includes the influence of real tensile deformation. The actual expected displacement of the strip and the real-time feed linear velocity obtained by pulse conversion from the encoder pulse of the speed measuring roller in front of the flying shear inlet are used as addressing variables and written into the data interface of the flying shear electronic cam model, with strip displacement as the master axis and cutter shaft phase as the slave axis. Within the synchronous shearing zone of this model, the driving current is dynamically adjusted to force the horizontal component of the tangential linear velocity of the flying shear shaft to remain absolutely consistent with the real-time feed linear velocity. Using the time offset corresponding to the cumulative longitudinal tensile deformation as a reference, it is used as the position compensation value. Under the constraints of maintaining a constant total displacement in a single cut and continuous acceleration at each intersection point, the position compensation value is subtracted from the master shaft displacement span in the acceleration zone, and the position compensation value is increased by an equal amount in the master shaft displacement span in the deceleration zone. Polynomial fitting is used to asymmetrically reconstruct the motion curve, outputting a compensated asymmetric electronic cam curve. Along the compensation asymmetric electronic cam curve, find the starting reference point where the flying shear servo axis just reaches the cutting stage. Perform reverse interpolation backtracking calculation along the direction of decreasing spindle displacement to calculate the mechanical starting offset of the axis when the spindle displacement is cleared. Write the mechanical starting offset into the underlying register to update the phase compensation parameters of the flying shear cutting action. The updated phase compensation parameter is combined with the real-time feed line speed of the strip for data encapsulation to generate a flying shear trigger command, which is then sent to the motion controller of the high-speed flying shear process at the end. A fixed interpolation calculation cycle is established based on the inherent communication frequency of the underlying hardware of the motion controller, and a mapping space with the strip travel distance as the virtual master axis is constructed in the control system. In each interpolation calculation cycle, the theoretical displacement increment of the strip is calculated strictly according to the currently collected real-time feed line speed, and the theoretical displacement increment, the virtual master axis position of the previous calculation cycle, and the phase compensation parameter as the position offset are dynamically summed to generate a new master axis position. The calculated new master axis position is updated in real time to the underlying instruction register of the motion controller, and converted into a feedforward drive command through internal interpolation and differential operations, which directly controls the flying shear servo motor to perform the shearing action according to the compensated action phase offset.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-process collaborative control method for a keel production line based on digital twins, characterized in that, include: Collect physical operation data in the keel production line process, which includes punching and rolling processes. The physical operation data includes local tension between frames, real-time current of servo motors, and springback deviation of strip cross-section. The physical operation data is input into the graph spatiotemporal attention network, the punching and rolling processes are set as network topology nodes, and the normalized local tension between the frames is used as the initial edge weight between the nodes. The physical feature vector is extracted by combining the real-time current of the servo motor and the springback deviation of the strip cross section. The physical feature vector is input into the digital twin inference model to calculate the cumulative longitudinal tensile deformation of the strip caused by continuous cross-process processing. The cumulative longitudinal tensile deformation is input as a feedforward compensation parameter into the flying shear process. The feedforward compensation parameter is used to perform time-series compensation for the cumulative deformation error of the strip and update the phase compensation parameter of the flying shear cutting action. Based on the updated phase compensation parameter, a flying shear trigger command is sent to the motion controller of the flying shear process to reconstruct the timing sequence of the flying shear cutting action.

2. The multi-process collaborative control method for a keel production line based on digital twins according to claim 1, characterized in that, The physical operation data acquisition process involves collecting local tension between the frames using tension sensors deployed between adjacent frames in the punching and rolling processes; collecting the real-time current of the servo motor driving the punching and rolling equipment by reading the external current transformer of the servo driver; acquiring the three-dimensional cross-sectional profile data of the strip using a laser profile scanner installed at the discharge end of the rolling process; and comparing the three-dimensional cross-sectional profile data with a preset standard digital three-dimensional model to extract the springback deviation of the strip cross-section.

3. The multi-process collaborative control method for a keel production line based on digital twins according to claim 1, characterized in that, The graph-temporal attention network includes a physical topology encoding module, a spatiotemporal feature joint extraction module, and a deductive feature alignment module connected in sequence. The physical topology encoding module maps the punching and each level of the roller press stand to a network topology node, obtains the real-time transmission line speed of the strip, calculates the quotient of the physical spatial distance between nodes and the real-time transmission line speed, and defines the quotient as a time delay parameter. The local tension between the stands is normalized to the time delay parameter, and a physical attenuation factor is constructed to nonlinearly reduce the initial edge weights to generate a dynamic adjacency matrix. The spatiotemporal feature joint extraction module generates spatial feature vectors and temporal feature vectors; the inference feature alignment module concatenates the spatial feature vectors and the temporal feature vectors into a tensor, performs a fully connected layer linear mapping on the concatenated tensor, and outputs a physical feature vector.

4. The multi-process collaborative control method for a keel production line based on digital twins according to claim 3, characterized in that, The spatiotemporal feature joint extraction module includes a spatial attention unit and a temporal attention unit. The spatial attention unit calculates the steady-state fluctuation phase difference of the real-time current of the servo motors of the upstream and downstream racks and defines it as a penalty term. It updates the dynamic adjacency matrix using the penalty term, calculates the spatial attention distribution weight based on the updated dynamic adjacency matrix, and outputs a spatial feature vector. The temporal attention unit constructs a variable-step long-term sliding window to sample historical operating data, extracts the temporal change rate of the strip cross-section springback deviation, maps the temporal change rate to a temporal attention score, and outputs a temporal feature vector based on the temporal attention score.

5. The multi-process collaborative control method for a keel production line based on digital twins according to claim 1, characterized in that, The digital twin inference model maps the physical feature vector into a state-space equation containing the physical operation data, iteratively calculates the strain increment of a single process based on the discrete processing time of each level of the roller press stand, corrects the strain increment of a single process using a nonlinear slip function based on the principle of volume invariance, and performs spatial integration in the direction of strip movement to output the initial deformation. The observation residual of the state-space equation is input into the Kalman filter algorithm to output the state correction value. The state correction value is used to perform time-domain compensation on the initial deformation, and the cumulative longitudinal tensile deformation is output.

6. The multi-process collaborative control method for a keel production line based on digital twins according to claim 1, characterized in that, The timing compensation involves adding the cumulative longitudinal tensile deformation to the reference cutting length of the flying shear to obtain the actual expected displacement of the strip; inputting the actual expected displacement of the strip and the real-time feed line speed into the flying shear electronic cam model; using the time offset corresponding to the cumulative longitudinal tensile deformation as a reference, shifting the cutting phase of the synchronous shearing zone in the flying shear electronic cam model to generate a compensated asymmetric electronic cam curve; and based on the compensated asymmetric electronic cam curve, calculating the mechanical starting point offset in reverse when the flying shear servo cutter axis reaches the cutting stage, and updating the mechanical starting point offset to the phase compensation parameters of the flying shear cutting action.

7. The multi-process collaborative control method for a keel production line based on digital twins according to claim 6, characterized in that, The flying shear electronic cam model is constructed using a piecewise mapping function with the actual expected displacement of the strip as the master axis and the rotational phase angle of the flying shear shaft as the slave axis, sequentially divided into an acceleration zone, a synchronous shearing zone, a deceleration zone, and a return zone. Within the synchronous shearing zone, the horizontal component of the tangential linear velocity of the flying shear shaft is constrained to maintain consistency with the real-time feed linear velocity. The cumulative longitudinal tensile deformation is used as the position compensation value. While maintaining a constant total periodic displacement of the master shaft and continuous acceleration, the position compensation value is subtracted from the master shaft displacement span in the acceleration zone, and an equal amount of position compensation value is added to the master shaft displacement span in the deceleration zone. A polynomial is used to asymmetrically reconstruct the motion curve, outputting a compensated asymmetric electronic cam curve.

8. The multi-process collaborative control method for a keel production line based on digital twins according to claim 1, characterized in that, The reconstructed flying shear cutting action timing process involves encapsulating the updated phase compensation parameters with the real-time feed line speed of the strip to generate a flying shear trigger command; sending the flying shear trigger command to the motion controller of the flying shear process; establishing a fixed interpolation calculation cycle with the inherent communication frequency of the underlying hardware of the motion controller; and constructing a mapping space with the strip travel distance as the virtual main axis. Within each interpolation operation cycle, the theoretical displacement increment is calculated based on the real-time feed line speed, and the theoretical displacement increment, the virtual spindle position of the previous cycle, and the phase compensation parameter as the position offset are dynamically summed to generate a new spindle position. The new spindle position is updated in the underlying instruction register, and after interpolation and differential operations, it is converted into a feedforward drive instruction to control the flying shear servo motor to perform shearing according to the compensated action phase offset.

Images