Bidirectional tension control withdrawal and straightening method for loose edge defect of cold-rolled 2B plate

Abstract

The invention provides a bidirectional tension control withdrawal and straightening method for loose edge defects of a cold-rolled 2B plate, which comprises the following steps of: S1, acquiring transverse strain distribution data of the plate in real time through a distributed optical fiber strain sensor array at the sampling frequency of more than or equal to 10kHz; s2, inputting the strain data and the process parameters into a physical-data hybrid drive prediction model, and generating an optimized combination matrix of longitudinal tension sigma long and transverse tension sigma trans by the model through a neural network constrained by an embedded Hollomon elastic-plastic equation; s3, the magnetic field intensity of the magnetorheological fluid clamping roller is dynamically adjusted according to the optimization matrix, so that the transverse tension is continuously adjustable within the range of 0.1-50 N / mm < 2 >; s4, a multi-target reinforcement learning algorithm is adopted, model parameters are updated on line based on real-time defect detection data, and the loose edge elimination rate and the energy consumption index are balanced. The method has the advantages that longitudinal and transverse tension coordinated regulation and control are realized, the loose edge defect elimination precision is improved, and the real-time adaptability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of cold-rolled sheet processing, and particularly to a two-way tension control straightening method for the loose edge defect of cold-rolled 2B sheets. Background Art

[0002] During the production process of cold-rolled sheets, the straightening process is a key link to ensure the flatness and dimensional accuracy of the sheets. However, there are many technical bottlenecks in the traditional straightening process: First, the existing technologies mainly rely on longitudinal tension control and cannot effectively adjust the transverse stress distribution of the sheets, resulting in stress concentration or relaxation at the edges, seriously affecting the sheet quality; Second, most of the existing intelligent control methods use offline training models and are difficult to adapt to the real-time fluctuations of material properties, with insufficient control accuracy and stability; Third, the adjustment accuracy of traditional mechanical segmented rolls is relatively low, usually with an error range of ±5 MPa, which easily causes over-tensioning or under-tensioning at the edges and cannot meet the production requirements of high-precision cold-rolled thin sheets. In addition, the existing technologies lack the ability to monitor the dynamic strain of the sheets in real time and are difficult to achieve the coordinated optimization control of longitudinal and transverse tensions. In view of the above problems, the existing technologies urgently need to be improved. Summary of the Invention

[0003] To solve the above technical problems, the present application provides a two-way tension control straightening method for the loose edge defect of cold-rolled 2B sheets, which has the advantages of realizing the coordinated regulation of longitudinal and transverse tensions, improving the elimination accuracy of loose edge defects, and enhancing the real-time adaptability of the model.

[0004] To solve the above technical problems, the present invention adopts the following technical solutions:

[0005] The present application provides a two-way tension control straightening method for the loose edge defect of cold-rolled 2B sheets, and the technical solution is as follows:

[0006] It includes the following steps: S1. Real-time collect the transverse strain distribution data of the sheet through a distributed fiber optic strain sensor array, and the sampling frequency ≥ 10 kHz; S2. Input the strain data and process parameters into a physical-data hybrid-driven prediction model, and the model generates an optimization combination matrix of longitudinal tension σ_long and transverse tension σ_trans through a neural network embedded with the Hollomon elastoplastic equation constraint; S3. Dynamically adjust the magnetic field strength of the magnetorheological fluid clamping roll according to the optimization matrix, so that the transverse tension is continuously adjustable within the range of 0.1 - 50 N / mm 2 ; S4. Adopt a multi-objective reinforcement learning algorithm to online update the model parameters based on the real-time defect detection data and balance the loose edge elimination rate and the energy consumption index.

[0007] Furthermore, the present application also proposes that in step S2, the physical-data hybrid-driven prediction model includes: a bidirectional LSTM network layer for processing time-series process parameters; a graph convolutional network layer for modeling the contact relationship between the roll system and the sheet; and a physical constraint layer for embedding the material constitutive equation through a differentiable solver.

[0008] Furthermore, the present application also proposes that in step S2, the Hollomon elastoplastic equation is σ = K·ε^n where: σ: true stress, unit is MPa; ε: true plastic strain, dimensionless; K: strength coefficient, reflecting the hardening ability after the initial yield of the material, unit is MPa; n: strain hardening index, characterizing the work hardening rate of the material, dimensionless (0 < n < 1).

[0009] Furthermore, the present application also proposes that in step S2, the optimization combination matrix of the longitudinal tension σ_long and the transverse tension σ_trans is a 10×10 optimized tension matrix, with the longitudinal tension σ_long (50 - 150 MPa), the transverse tension σ_trans (0.1 - 50 MPa), step size 0.5 MPa, and satisfying

[0010] Furthermore, the present application also proposes that step S3 includes generating a magnetic field intensity distribution map according to the tension matrix, and a PID controller adjusts the current. The formula for the transverse tension is σ trans = μ·B 2 , where μ is the viscosity coefficient of the magnetorheological fluid.

[0011] Furthermore, the present application also proposes that step S1 includes arranging a set of sensor arrays at the inlet and outlet of the tension leveling machine, with a spacing of 50 cm, for capturing the dynamic strain propagation, and using a C-band demodulator to measure and obtain the strain data, with a sampling frequency of 12.5 kHz.

[0012] Furthermore, the present application also proposes that step S1 further includes monitoring the surface temperature field of the sheet using a multi-spectral thermal imager and measuring the running speed of the sheet in real time using a laser velocimeter.

[0013] Furthermore, the present application also proposes that step S1 further includes achieving multi-sensor clock synchronization based on the Precision Time Protocol, with a time deviation ≤ 10 μs.

[0014] Furthermore, the present application also proposes that step S4 includes, when the right loose edge is detected, triggering the reinforcement learning agent to generate a compensation strategy; the magnetorheological clamping roll increases the right transverse tension to 20 MPa within 0.2 s, while reducing the left tension to 16 MPa; the online learning module records the adjustment effect and updates the GCN edge weights.

[0015] Furthermore, the present application also proposes that the judgment criterion for detecting the loose right edge is that the residual stress deviation > 12%

[0016] As can be seen from the above, a two-way tension control and straightening method for the loose edge defect of cold-rolled 2B plates provided by the present application collects high-precision strain data in real time through a distributed optical fiber strain sensor array, generates a longitudinal and transverse tension optimization matrix by combining a hybrid drive model with physical constraints, uses a magnetorheological fluid clamping roll to achieve continuous regulation of the transverse tension, and cooperates with an online reinforcement learning algorithm to dynamically optimize control parameters, effectively solving the technical problems of insufficient transverse stress adjustment ability and poor model adaptability of the traditional straightening process, and having the advantages of high success rate in eliminating loose edge defects and precise energy consumption control. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of the steps of a two-way tension control and straightening method for the loose edge defect of cold-rolled 2B plates according to the present invention;

[0018] Figure 2 is a structural block diagram of the physical-data hybrid drive prediction model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The technical solutions in the present application will be clearly and completely described below with reference to the accompanying drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the present application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application to be protected, but only represents the selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts belong to the scope of protection of the present application. It should be noted that: similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present application, terms such as "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0020] In the prior art, the cold-rolled thin plate straightening process has long faced the problem of uncontrollable transverse stress distribution. The traditional method relies on a single longitudinal tension adjustment mechanism and cannot effectively eliminate the relaxation defect at the edge of the plate. The mechanical segmented roll is limited by the mechanical structure and has defects such as adjustment hysteresis and insufficient accuracy. The off-line trained control model is difficult to adapt to the real-time fluctuations of the material properties during the production process, resulting in a mismatch between the control strategy and the actual working conditions.

[0021] To solve the above problems, the R & D team found that precise regulation of the transverse stress distribution is the key breakthrough for solving the loose edge defect. Based on the theoretical analysis of the elastic-plastic deformation of materials, a collaborative regulation mechanism combining longitudinal and transverse tension control was proposed. To address the problem of insufficient model generalization ability, a technical path of embedding the material constitutive equation into the data-driven model was explored. To solve the bottleneck of the response speed of the actuator, magnetorheological fluid was selected as the core regulation medium.

[0022] Referring to the attached Figure 1-2 As shown, therefore, this application proposes a technical solution including the following steps: real-time collecting the transverse strain distribution data of the sheet by a distributed fiber optic strain sensor array, with the sampling frequency reaching or exceeding 10 kHz; inputting the strain data and process parameters into a physical-data hybrid-driven prediction model, which generates an optimization combination matrix of longitudinal and transverse tensions through a neural network constrained by the embedded Hollomon elastic-plastic equation; dynamically adjusting the magnetic field intensity of the magnetorheological fluid clamping roll according to the optimization matrix, so that the transverse tension is continuously adjustable within a set range; adopting a multi-objective reinforcement learning algorithm to online update the model parameters based on real-time defect detection data and balance the loose edge elimination rate and energy consumption index.

[0023] Among them, the distributed fiber optic strain sensor array refers to a fiber optic sensor network evenly arranged along the width direction of the sheet, and a Bragg grating array can be used to realize strain measurement, and the local strain value is deduced by detecting the wavelength shift of the optical wave. The physical-data hybrid-driven prediction model refers to a computational model that integrates material mechanics equations and machine learning algorithms. A bidirectional long short-term memory network can be used to process the time-series process parameters, and a graph convolutional network is combined to establish the topological structure of the roll contact relationship. The magnetorheological fluid clamping roll refers to an intelligent roll device filled with magnetorheological fluid, and the viscosity of the fluid is adjusted by changing the current intensity of the electromagnetic coil, thereby controlling the friction coefficient between the roll surface and the sheet. The multi-objective reinforcement learning algorithm refers to an online learning mechanism that simultaneously optimizes multiple control objectives, and a deep deterministic policy gradient algorithm can be used to achieve dynamic parameter adjustment.

[0024] Specifically, the high-frequency strain data acquisition system monitors the stress distribution state of the sheet in real time and provides dynamic input for the control model. While processing the time-series characteristics of the process parameters, the hybrid prediction model enforces the material constitutive law through a differentiable solver to ensure that the generated tension combination conforms to physical constraints. The magnetic field regulation system converts the optimization matrix into a current command to precisely control the viscosity of the magnetorheological fluid and achieve continuous adjustment of the transverse tension. The online learning module continuously updates the network weights according to the actual control effect, so that the control strategy automatically adapts to the change of material properties.

[0025] Compared with the prior art, the traditional method can only adjust the stress of the sheet by longitudinal stretching, while this solution actively intervenes in the transverse stress distribution through the coordinated control of bidirectional tension. The existing control model relies on offline training data, while this solution realizes the dynamic optimization of model parameters through online reinforcement learning. The traditional mechanical rollers are limited by the discrete adjustment mode, while this solution adopts magnetorheological intelligent rollers to achieve stepless continuous control.

[0026] Through the above technical solutions, this application realizes the coordinated optimization control of longitudinal and transverse tensions, effectively balancing the stress distribution in the width direction of the sheet. The dynamic learning mechanism ensures the adaptability of the control strategy to the fluctuations of material properties and solves the problem of offline model mismatch. The intelligent clamping device breaks through the accuracy limitation of mechanical adjustment and realizes precise tension control with a sub-second response speed.

[0027] This application further proposes a physical-data hybrid-driven prediction model, which includes a bidirectional LSTM network layer for processing time-series process parameters, a graph convolutional network layer for modeling the contact relationship between the roll system and the sheet, and a physical constraint layer for embedding the material constitutive equation through a differentiable solver.

[0028] Among them, the bidirectional LSTM network layer refers to a recurrent neural network structure with the ability to process bidirectional time series, which can be specifically implemented by using hidden units containing the time backpropagation mechanism and is used to capture the temporal dependencies of process parameters before and after. The graph convolutional network layer refers to a convolutional neural network for feature extraction based on graph-structured data, which can be specifically implemented by using an adjacency matrix to describe the contact relationship between roll system nodes and sheet nodes and is used to establish a spatial mechanical interaction model between the roll system and the sheet. The physical constraint layer refers to a module that transforms the material constitutive equation into a differentiable form and embeds it into neural network training, which can be specifically implemented by using automatic differentiation technology to construct the coupling relationship between the Hollomon equation and the loss function and is used to ensure that the prediction results conform to the elastic-plastic deformation law of materials.

[0029] Specifically, in the physical-data hybrid-driven prediction model, the bidirectional LSTM network layer receives inputs of time-series process parameters such as temperature, speed, and tension set values, and extracts feature vectors through forward and backward time series analysis. The graph convolutional network layer receives the geometric parameters of the roll system and the operating state data of the sheet, constructs a graph node feature matrix containing attributes such as roll diameter, roll spacing, and contact pressure, and updates the node state using multi-layer graph convolutional operations to characterize the stress transfer characteristics of the contact area. The physical constraint layer substitutes the predicted stress output by the neural network into the differentiable Hollomon equation, generates a physical loss function by calculating the residual between the predicted strain and the measured strain, and simultaneously optimizes the data fitting error and the physical law constraint conditions during the backpropagation process.

[0030] Compared with the prior art, traditional prediction models only adopt a single data-driven method and lack physical constraints on the constitutive relationship of materials. When the material properties fluctuate, prediction results that do not conform to mechanical principles are likely to occur. Existing graph neural network models do not consider the dynamic change characteristics of the contact relationship between the roll system and the sheet, and cannot accurately describe the spatial coupling effect during the tension transfer process.

[0031] Through the above technical solutions, this application can dynamically adapt to the performance changes of different batches of materials, ensure the rationality of the predicted tension through physical constraints, and at the same time accurately model the complex contact relationship between the roll system and the sheet, providing an accurate mechanical state prediction for real-time adjustment of the lateral tension.

[0032] This application further proposes that in step S2, the Hollomon elastoplastic equation is σ = K·ε^n, where: σ is the true stress, with the unit of MPa; ε is the true plastic strain, dimensionless; K is the strength coefficient, reflecting the hardening ability of the material after initial yield, with the unit of MPa; n is the strain hardening index, characterizing the work hardening rate of the material, dimensionless and within the range of 0 to 1.

[0033] Among them, the true stress refers to the stress value actually borne by the material during the plastic deformation stage. Specifically, the stress-strain curve can be obtained through a tensile test, and then calculated by area correction. This parameter is used to accurately characterize the load state of the sheet during the tension leveling process.

[0034] Among them, the true plastic strain refers to the strain value when the material undergoes irreversible deformation. Specifically, it can be calculated through the logarithmic strain formula. This parameter is used to quantify the degree of plastic deformation of the sheet.

[0035] Among them, the strength coefficient refers to the resistance index when the material continues to harden after yielding. Specifically, it can be determined by fitting experimental data. This parameter reflects the ability of the material to resist subsequent plastic deformation.

[0036] Among them, the strain hardening index refers to the quantitative index of the work hardening rate of the material during the plastic deformation stage. Specifically, it can be obtained through the power-law fitting method. This parameter determines the rate of stress increase with strain.

[0037] Specifically, by embedding the Hollomon elastoplastic equation as a physical constraint into the neural network training process, a non-linear mapping relationship between the true stress and plastic strain of the material is established. The combination of the strength coefficient and the strain hardening index in this equation can accurately characterize the hardening characteristics of different materials. During model training, backpropagation calculation is performed through a differentiable solver, so that the prediction results always follow the material constitutive relationship. When the material properties fluctuate, the stress prediction value is adjusted in real time through the online updated strength coefficient and strain hardening index, thereby eliminating the prediction deviation caused by the curing of material parameters in the offline model.

[0038] Compared with the prior art, the traditional method only relies on a data-driven model for stress prediction and lacks the physical constraints of the material constitutive relationship. When the material batch changes or the performance fluctuates due to environmental factors, the prediction model needs to be retrained offline. In this solution, by embedding the differentiable Hollomon equation in the neural network, the prediction result automatically satisfies the material elastoplastic deformation law. Even in the case of material performance fluctuations, only the strength coefficient and hardening index in the equation need to be adjusted to complete the online update of the model.

[0039] Through the above technical solution, the present application can dynamically adjust the stress prediction model according to the real-time detected material performance parameters, effectively eliminating the model prediction deviation caused by material batch differences, temperature changes or changes in the work-hardening state, enabling the tension control system to adapt to the material property changes under different working conditions, and significantly improving the stability of the loose edge defect elimination process.

[0040] The present application further proposes that in step S2, the optimization combination matrix of the longitudinal tension and the transverse tension is an optimization tension matrix of 10×10. The longitudinal tension is set in the range of 50 - 150 MPa, the transverse tension is set in the range of 0.1 - 50 MPa, the step size is set to 0.5 MPa, and the stress divergence constraint condition is satisfied.

[0041] Among them, the optimization tension matrix refers to a two-dimensional discrete parameter combination set composed of the longitudinal tension and the transverse tension. Specifically, it can be realized by using a grid search algorithm to generate a matrix structure covering the process parameter space, providing an optimization space for the system through the discrete tension combination. The longitudinal tension range of 50 - 150 MPa refers to the tensile stress interval applied along the rolling direction of the sheet, which can be specifically realized by the coordinated control of the main drive motor and the tension roller, and this range covers the typical load requirements in the cold rolling thin sheet processing process. The transverse tension range of 0.1 - 50 MPa refers to the stress adjustment interval perpendicular to the rolling direction, which can be specifically realized by regulating the magnetic field strength of the magnetorheological fluid clamping roller to adapt to the weak stress adjustment requirements in the edge area of the sheet. The step size of 0.5 MPa refers to the minimum adjustment increment of adjacent tension parameters, which can be specifically realized by the high-precision servo drive system and the real-time feedback control of the magnetorheological fluid viscosity, breaking through the precision limitation of the traditional mechanical adjustment device. The stress divergence constraint condition refers to a mathematical expression that the sum of the longitudinal and transverse stress gradients does not exceed the material yield strength, which can be specifically realized by online calculating the divergence value of the stress tensor and comparing it with the material yield strength database in real time to ensure that the sheet is in the elastic deformation state.

[0042] Specifically, this technical solution combines the wide-range adjustment of longitudinal tension with the fine adjustment ability of transverse tension by constructing a 10×10 two-dimensional tension parameter matrix. Each node in the matrix represents a specific combination of longitudinal and transverse tensions, and the stress combination is optimized by traversing the matrix nodes. The wide range setting of longitudinal tension can adapt to the processing requirements of plates with different thickness specifications, and the fine-span design of transverse tension focuses on solving the problem of weak stress adjustment in the edge area. The step size of 0.5 MPa enables the tension adjustment accuracy to reach the sub-megapascal level, significantly superior to the adjustment ability of traditional mechanical devices. The stress divergence constraint condition mathematically enforces the composite stress level under any tension combination, ensuring that the equivalent stress in each area of the plate is always lower than the material yield limit, maintaining the elastic deformation state and avoiding local overload. This combination of discretized parameter space and continuous physical constraints realizes the global optimization search of tension parameters while ensuring material safety.

[0043] Compared with the prior art, the traditional stretching process only uses longitudinal tension control with low adjustment accuracy, unable to actively control the transverse stress distribution, and prone to over-stretching or under-stretching in the edge area. The existing mechanical segmented roller device is limited by the adjustment accuracy of ±5 MPa and is difficult to achieve fine tension control at the sub-megapascal level. This technical solution realizes the coordinated control of stress distribution in both longitudinal and transverse dimensions by constructing a two-dimensional tension optimization matrix, combined with high-precision step size setting and stress constraint conditions, not only solving the technical defect of uncontrollable transverse stress in the traditional process but also overcoming the limitation of insufficient adjustment accuracy of mechanical devices.

[0044] Through the above technical solution, this application realizes the precise coordinated control of longitudinal and transverse tensions, effectively avoiding the plastic deformation problem caused by local stress overrun. The discretized tension matrix provides a global optimization search space for the system, combined with the stress divergence constraint condition to ensure that the material is within the safe stress range, enabling the plate to eliminate the loose edge defect and maintain the elastic deformation state during the stretching process. The fine adjustment ability of transverse tension is particularly suitable for the precise control of the weak stress area at the edge, solving the technical problems of over-stretching or under-stretching at the edge in the traditional process.

[0045] This application further proposes to generate a magnetic field intensity distribution map according to the tension matrix, adjust the current through a PID controller, and the transverse tension calculation formula is σtrans = μ·B 2 , where μ is the viscosity coefficient of the magnetorheological fluid.

[0046] Among them, the magnetic field intensity distribution map refers to a two-dimensional magnetic field control instruction map generated according to the transverse tension target value in the optimized tension matrix, which can be specifically realized by calculating the coil current distribution of the magnetorheological fluid clamping roller using finite element electromagnetic field simulation software. Through the mapping relationship between the spatial magnetic field intensity distribution and the target tension distribution, continuous gradient control of transverse tension is achieved.

[0047] Among them, the PID controller adjusts the current, which means that the proportional-integral-derivative control algorithm performs closed-loop regulation on the drive current of the magnetorheological fluid clamping roller. Specifically, an incremental PID algorithm with feedforward compensation can be used to achieve this. By comparing the deviation between the set magnetic field strength and the actual detected value in real time, the duty cycle of the coil drive current is dynamically adjusted.

[0048] Among them, the formula for the transverse tension σ trans = μ·B 2 refers to a physical model established based on the quadratic relationship between the viscosity coefficient of the magnetorheological fluid calibrated through experiments and the magnetic field strength. Specifically, it can be achieved by measuring the shear stress characteristic curve of the magnetorheological fluid at different magnetic field strengths with a rheometer, and establishing a quantitative conversion relationship between the magnetic field strength and the equivalent transverse tension.

[0049] Specifically, during the tension control process, first, the transverse tension target value in the optimization matrix is input into the magnetic-force conversion module, and the required magnetic field strength distribution is calculated according to the inverse function of σ trans = μ·B 2 . Subsequently, the initial values of the coil currents of each clamping roller unit are generated through finite element electromagnetic field calculation, and the PID controller performs feedback regulation according to the magnetic field strength actually measured by the Hall sensor. When it is detected that the actual magnetic field strength deviates from the set value, the PID algorithm automatically corrects the pulse width of the drive current, so that the shear stress of the magnetorheological fluid quickly converges to the target tension value. During this process, the characteristic of the square relationship of the magnetic field strength enables the transverse tension to produce a smooth non-linear response with the change of the current, avoiding the step change of the mechanical adjustment mechanism.

[0050] Compared with the prior art, the traditional mechanical segmented roller realizes discrete transverse tension adjustment by mechanically adjusting the roll gap height. Limited by the mechanical transmission clearance and stiffness characteristics, it is difficult to achieve continuous gradient control. However, this solution can achieve stepless continuous adjustment of the transverse tension through the rheological characteristics of the magnetorheological fluid and the continuously adjustable electromagnetic field, combined with closed-loop current control.

[0051] Through the above technical solutions, this application can eliminate the inherent response hysteresis and position deviation of the mechanical adjustment mechanism, and achieve high-precision dynamic control of the transverse tension. During the cold rolling of thin sheets, this solution can avoid the edge over-tension or under-tension phenomena caused by sudden changes in the transverse tension, ensure uniform stress distribution across the sheet cross-section, and effectively eliminate the loose edge defect.

[0052] The present application further proposes a technical solution for strain dynamic monitoring during the control of the loose edge defect of cold-rolled sheets by arranging two sets of sensor arrays and adopting high-frequency demodulation technology. During specific implementation, a set of distributed fiber optic strain sensor arrays are respectively installed at the inlet and outlet positions of the tension leveling machine. The distance between the two arrays is set to 50 cm. A C-band demodulator is used for data measurement, and the sampling frequency is set to 12.5 kHz.

[0053] Among them, the distributed fiber optic strain sensor array refers to a set of sensing units continuously arranged along the transverse direction of the sheet. Specifically, it can be realized by using a fiber Bragg grating array sensor. Each sensing unit is arranged at an interval of 10 mm, covering the full width of the sheet. The C-band demodulator refers to an optical demodulation device with a working wavelength in the range of 1525 - 1565 nm. Specifically, it can be realized by using a wavelength scanning system based on a tunable laser. The strain value is calculated by detecting the wavelength shift of the fiber Bragg grating reflection. The 12.5 kHz sampling frequency means that 12,500 strain data are collected per second. Specifically, it can be realized by equipping a high-speed data acquisition card. This frequency setting can meet the strain change tracking requirements when the running speed of the cold-rolled sheet exceeds 300 m / min.

[0054] Specifically, the dual-station sensor arrays at the inlet and outlet form a spatially continuous monitoring network. When the sheet passes through the tension leveling machine at high speed, the inlet array detects the initial strain distribution, and the outlet array captures the strain change after tension adjustment. The setting of the 50 cm distance ensures that the two arrays can not only completely cover the deformation area of the sheet during the tension leveling process but also avoid data redundancy caused by the overlap of the signal acquisition areas. The C-band demodulator converts the wavelength shift of the fiber Bragg grating into a micro-strain value through phase-sensitive optical time domain reflectometry technology. Its wavelength resolution can reach the level of 1 picometer, corresponding to a strain measurement accuracy of ±2 micro-strain. The 12.5 kHz sampling frequency is realized through a high-speed analog-to-digital converter, which can complete parallel data acquisition of 1024 sensing points in each sampling period, ensuring the complete capture of transient strain fluctuations.

[0055] Compared with the prior art, traditional methods mostly adopt single-point strain sensors or low-frequency sampling systems, which cannot effectively capture the dynamic strain propagation process of high-speed running sheets. In the prior art, the sampling frequency of typical contact strain gauges is usually lower than 1 kHz, and there are defects such as fixed installation positions and insufficient spatial resolution. However, this solution realizes full-width, high spatio-temporal resolution monitoring of the surface strain field of the sheet by combining a distributed array layout and high-frequency demodulation technology, and can accurately track the propagation trajectory of the strain wave in the advancing direction of the sheet.

[0056] Through the above technical solution, the present application effectively solves the technical problem of difficult real-time capture of dynamic strain propagation in the detection of loose edge defects of cold-rolled sheets. The arrangement of the dual-station sensor array forms a complete monitoring chain for the strain propagation path, the high-frequency sampling ability ensures the complete capture of the strain fluctuation characteristics, and the C-band demodulation technology provides high-precision strain quantification data. This solution provides strain field data support with high spatio-temporal resolution for subsequent tension control decisions, enabling the control system to accurately identify the formation process of loose edge defects and intervene and adjust in a timely manner.

[0057] The present application further proposes a technical means of arranging a set of sensor arrays at the inlet and outlet of the temper mill, and also using a multi-spectral thermal imager to monitor the surface temperature field of the sheet and a laser velocimeter to measure the running speed of the sheet in real time.

[0058] Among them, the multi-spectral thermal imager is a measuring instrument that obtains the surface temperature distribution of an object through thermal radiation detection devices with different wavelengths. Specifically, it can be realized by using an infrared focal plane array detector with a multi-band filter. Its function is to capture the heat distribution gradient on the surface of the sheet caused by plastic deformation and provide temperature field boundary conditions for elastoplastic deformation analysis. The laser velocimeter is a linear velocity measuring device based on the Doppler effect principle. Specifically, it can be realized by using a dual-beam interferometric laser velocimeter system. Its function is to quantify the longitudinal movement speed of the sheet in real time and eliminate the transmission chain error in traditional encoder measurements.

[0059] Specifically, during the process of the distributed fiber optic strain sensor array collecting transverse strain distribution data, the multi-spectral thermal imager obtains the surface temperature field data of the sheet in a non-contact manner. By establishing the mapping relationship between the temperature gradient and the change in material yield strength, the material constitutive parameters in the physical driving model are corrected; the laser velocimeter calculates the instantaneous value of the running speed in real time by detecting the frequency shift of the scattered light on the surface of the sheet, and inputs the speed data, strain data, and temperature data into the prediction model synchronously. The temperature field data is used to compensate for the interference of the material thermal expansion effect on strain measurement, and the running speed data is used to calculate the phase difference between the tension application and the material displacement, thereby realizing the spatio-temporal matching of the tension control command and the material motion state.

[0060] Compared with the prior art, the traditional tempering process only relies on strain sensors and encoders to achieve single-point speed measurement, cannot eliminate the influence of temperature gradient on the rheological properties of materials, and there is a mechanical transmission lag in speed measurement. Through the collaborative application of the multi-spectral thermal imager and the laser velocimeter, this solution adds new temperature field boundary conditions and real-time dynamic displacement parameters on the basis of maintaining the original strain monitoring accuracy, enabling the hybrid driving model to accurately reflect the thermo-mechanical coupling effect of materials and eliminating the control command delay caused by speed measurement lag.

[0061] Through the above technical solution, the present application effectively solves the problem of inaccurate control of the transverse stress distribution caused by the lack of real-time temperature field and accurate velocity data. By establishing a multi-dimensional physical quantity collaborative monitoring mechanism, it provides a complete temperature compensation benchmark and dynamic displacement parameters for subsequent tension optimization calculation, significantly improving the control synchronization and accuracy of the transverse tension distribution.

[0062] The present application further proposes to achieve multi-sensor clock synchronization based on the Precision Time Protocol in step S1, with the time deviation controlled within 10 microseconds.

[0063] Among them, the Precision Time Protocol refers to the networked clock synchronization standard protocol, which can specifically be implemented using the IEEE 1588 protocol. It achieves microsecond-level time synchronization through the master-slave clock architecture, and its role is to eliminate delay jitter in long-distance transmission.

[0064] Among them, the time deviation refers to the maximum allowable difference in the sampling moments of multi-sensors, which can specifically be achieved through hardware timestamp marking and software compensation algorithms. The limitation of controlling the time deviation within 10 microseconds can ensure the timing consistency of strain, temperature, and velocity data.

[0065] Specifically, in a multi-physical field measurement system composed of a distributed fiber optic strain sensor array, a multi-spectral thermal imager, and a laser velocimeter, a hierarchical synchronization network of the master clock and each sensor slave clock is established by deploying a synchronization module that supports the IEEE 1588 protocol. The master clock periodically broadcasts synchronization messages, and the slave clocks calculate the clock offset based on the message transmission delay and adjust the local clock frequency through a phase-locked loop circuit. During the data acquisition stage, each sensor triggers the sampling action based on the synchronized clock, records the actual sampling moment using a hardware timestamp, and then compensates for the residual deviation within 10 microseconds through a time-domain interpolation algorithm. Finally, the multi-source data is aligned at the sub-millisecond level in the timing dimension. Thus, the correlation analysis of dynamic strain propagation, temperature field changes, and the running speed of the sheet metal can be carried out based on a strictly synchronized time benchmark, avoiding calculation errors of transverse stress caused by clock deviation.

[0066] Compared with the prior art, traditional methods usually use a hardware trigger line or a simple network time protocol to achieve synchronization, which has problems such as limited transmission distance or millisecond-level synchronization errors. This solution breaks through the physical distance limitation of hardware triggers through the hierarchical synchronization mechanism of the Precision Time Protocol, and at the same time improves the synchronization accuracy to the microsecond level, meeting the timing accuracy requirements of multi-source data fusion in high-speed dynamic measurement scenarios.

[0067] Through the above technical solution, the present application effectively solves the problem of dynamic strain measurement deviation caused by clock asynchronization during multi-sensor data acquisition, ensuring that the subsequent physical-data hybrid-driven prediction model can obtain multi-physical field input data with strict time-domain alignment, so as to accurately calculate the transverse stress distribution and optimize the tension control strategy.

[0068] The present application further proposes a technical solution in which when the loose side on the right is detected, a reinforcement learning agent is triggered to generate a compensation strategy. The magnetorheological pinch roll increases the transverse tension on the right to the target value within a set time, and at the same time reduces the tension on the left to the target value. The online learning module records the adjustment effect and updates the edge weights of the graph convolutional network.

[0069] Among them, the reinforcement learning agent refers to an intelligent agent that can autonomously optimize the decision-making strategy by interacting with the environment. Specifically, it can be implemented by the deep deterministic policy gradient algorithm, and a tension compensation strategy is generated by establishing a state-action value function mapping relationship.

[0070] Among them, the magnetorheological pinch roll refers to an intelligent actuator based on the change of the viscosity of the magnetorheological fluid with the magnetic field strength. Specifically, it can be implemented by the structure of a toroidal electromagnet and a ferromagnetic particle suspension, and the clamping force distribution is adjusted in real time by changing the current intensity.

[0071] Among them, the transverse tension adjustment refers to the stress control operation applied along the width direction of the sheet. Specifically, it can be achieved by adjusting the magnetic field strength in different regions of the pinch roll, forming a transverse tension gradient to eliminate local stress relaxation.

[0072] Among them, the online learning module refers to a computing unit with real-time data acquisition and model update functions. Specifically, it can be implemented by a sliding window mechanism, and the topological connection weights of the graph convolutional network are updated by storing the latest operation records.

[0073] Among them, the update of the edge weights of the graph convolutional network refers to dynamically adjusting the modeling parameters of the contact relationship between the roll system and the sheet. Specifically, it can be achieved by the backpropagation algorithm, and the actual compensation effect is quantified as the loss function gradient for network optimization.

[0074] Specifically, after the detection module identifies that the residual stress deviation on the right side exceeds the threshold, the reinforcement learning agent is triggered to read the current process parameters and sensor data, and generate a compensation strategy including right-side tension increase and left-side tension decrease based on the Markov decision process. After receiving the control instruction, the magnetorheological clamping roller adjusts the current intensity of the electromagnetic coil to increase the magnetic field intensity in the right-side area and decrease the magnetic field intensity in the left-side area within 0.2 seconds respectively, forming a lateral tension difference to balance the stress distribution of the sheet. During the execution process, the temperature sensor and strain measurement data are synchronously collected and input into the online learning module. By comparing the theoretical predicted value with the actual compensation effect, the parameter correction amount is calculated, and the connection weight coefficient of the roller system nodes in the graph convolutional network is updated using the gradient descent method, enabling the physics-data hybrid-driven model to continuously adapt to the changes in material properties.

[0075] In some specific embodiments, when it is detected that wavy deformation appears in the right-edge area, the local temperature distribution data can be obtained through an infrared thermal imager, and the degree of loose edge can be judged in combination with the strain measurement value. The magnetorheological clamping roller can adopt a zone control mode, dividing the roller surface into several independent electromagnetic control units. For example, eight control sections are arranged along the axial direction, and each section is equipped with a separate excitation coil and current controller. The reinforcement learning agent can construct a state space including elements such as the sheet speed, temperature field, and historical tension value, and the output action space includes the current adjustment amplitude and response timing of each electromagnetic section.

[0076] Compared with the prior art, the traditional method relies on a preset empirical formula for tension compensation, suffering from problems such as response lag and insufficient adjustment accuracy. The present invention can generate a dynamic compensation strategy in real time through the reinforcement learning agent, which can autonomously optimize the action sequence according to the online monitoring data, and realize the rapid reconstruction of the tension field in combination with the millisecond-level response characteristics of the magnetorheological clamping roller. Compared with the discrete adjustment method of the mechanical segmented roller, the present invention realizes the stepless adjustment of the tension gradient by continuously adjusting the magnetic field intensity, and at the same time effectively overcomes the influence of material property fluctuations on the model prediction accuracy by using the online update mechanism of the graph convolutional network.

[0077] Through the above technical solutions, the present application realizes the real-time closed-loop control of the loose-edge defect of cold-rolled sheets. When the stress anomaly on the right side is detected, the optimal compensation strategy is quickly generated and executed, and the local stress relaxation is eliminated by dynamically adjusting the lateral tension gradient. This solution effectively solves the problem of adjustment failure caused by model mismatch in the traditional method, and continuously optimizes the control model parameters through the online learning mechanism to ensure the control stability of the system under the condition of changing material properties. At the same time, the rapid response characteristics of the magnetorheological actuator and the synergistic effect of the reinforcement learning decision-making module significantly improve the timeliness and accuracy of the lateral tension adjustment, avoiding production interruption caused by manual intervention.

[0078] The present application further proposes that the judgment criterion for detecting the loose edge on the right side is that the residual stress deviation exceeds 12%.

[0079] Among them, the residual stress deviation refers to the degree of deviation between the actually measured stress in the edge area of the sheet and the reference stress value. Specifically, a distributed fiber optic strain sensor array can be used to collect transverse stress distribution data in real time, and the deviation percentage can be calculated by comparing with a preset reference stress field. This feature establishes a defect determination standard by setting a quantization threshold to ensure the objectivity and consistency of defect identification.

[0080] Among them, the judgment criterion refers to the critical condition for triggering the tension compensation strategy. Specifically, it can be dynamically calibrated through multi-sensor data fusion technology combined with the analysis of material yield characteristics. This feature is set based on the mechanical properties of materials and the law of dynamic strain propagation, and can accurately distinguish the critical state between normal process fluctuations and loose edge defects.

[0081] Specifically, when the distributed fiber optic sensor detects the transverse stress distribution data in the right area of the sheet, by calculating the deviation amplitude between the current stress and the reference stress field, if the deviation exceeds the set threshold of 12%, the system immediately determines that a loose edge defect has occurred. At this time, the control module immediately activates the tension compensation program of the magnetorheological fluid clamping roller and adjusts the transverse tension distribution according to the strategy generated by the reinforcement learning algorithm. This judgment criterion is calibrated by combining the material yield strength and the dynamic strain propagation speed to ensure the accuracy of defect detection even when the running speed of the sheet changes or the material properties fluctuate.

[0082] Compared with the prior art, the traditional method relies on visual inspection by operators or simple comparison based on a fixed stress difference, and there are risks of judgment lag and misjudgment. However, this solution realizes the automatic and accurate identification of loose edge defects by setting a residual stress deviation threshold based on the mechanical properties of materials and combining real-time sensor data, avoiding the response delay caused by manual intervention.

[0083] Through the above technical solution, this application can effectively eliminate the phenomena of over-tensioning or under-tensioning at the edges caused by inaccurate defect identification during the mechanical segmented roller adjustment process, ensuring that the tension adjustment action is triggered only when there is actually a loose edge defect. This solution significantly improves the defect handling accuracy through a quantization judgment criterion and an automatic detection mechanism, avoiding energy waste and equipment loss caused by ineffective adjustment.

[0084] The above are only embodiments of this application and are not used to limit the protection scope of this application. For those skilled in the art, this application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included in the protection scope of this application.

Claims

1. A two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B plates, characterized in that, It includes the following steps: S1. Real-time collect the transverse strain distribution data of the sheet through a distributed fiber optic strain sensor array, with a sampling frequency ≥ 10 kHz; S2. Input the strain data and process parameters into a physical-data hybrid-driven prediction model, which generates an optimized combination matrix of longitudinal tension σ_long and transverse tension σ_trans through a neural network embedded with the Hollomon elastoplastic equation constraint; S3. Dynamically adjust the magnetic field intensity of the magnetorheological fluid clamping roller according to the optimization matrix to make the transverse tension continuously adjustable within the range of 0.1 - 50 N / mm 2 ; S4. Adopt a multi-objective reinforcement learning algorithm to online update the model parameters based on real-time defect detection data, and balance the loose edge elimination rate and the energy consumption index.

2. The two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B sheet according to claim 1, characterized in that, In the step S2, the physical-data hybrid-driven prediction model includes: A bidirectional LSTM network layer to process the time-series process parameters; A graph convolutional network layer to model the contact relationship between the roll system and the sheet; A physical constraint layer to embed the material constitutive equation through a differentiable solver.

3. The two-way tension control straightening method for the loose edge defect of cold-rolled 2B sheet according to claim 1, characterized in that, In the step S2, the Hollomon elastoplastic equation is σ = K·ε^n Where: σ: True stress, unit is MPa; ε: True plastic strain, dimensionless; K: Strength coefficient, reflecting the hardening ability after the initial yield of the material, unit is MPa; n: Strain hardening index, characterizing the work hardening rate of the material, dimensionless (0 < n < 1).

4. The two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B plates according to claim 1, wherein, In the step S2, the optimization combination matrix of the longitudinal tension σ_long and the transverse tension σ_trans is a 10×10 optimized tension matrix, where the longitudinal tension σ_long is (50 - 150 MPa), the transverse tension σ_trans is (0.1 - 50 MPa), the step size is 0.5 MPa, and it satisfies 5. The two-way tension control stretcher leveling method for the loose edge defect of cold-rolled 2B sheet according to claim 1, characterized in that, The step S3 includes generating a magnetic field intensity distribution map according to the tension matrix, and a PID controller adjusts the current. The formula for the lateral tension is σ trans = μ·B 2 , where μ is the viscosity coefficient of the magnetorheological fluid.

6. The two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B sheet according to claim 1, characterized in that, The step S1 includes arranging a set of sensor arrays at the inlet and outlet of the tension leveller respectively, with a spacing of 50 cm, for capturing dynamic strain propagation, and measuring the strain data using a C-band demodulator, with a sampling frequency of 12.5 kHz.

7. The two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B sheet according to claim 6, characterized in that, The step S1 also includes monitoring the surface temperature field of the sheet using a multi-spectral thermal imager and measuring the running speed of the sheet in real time using a laser velocimeter.

8. The two-way tension control straightening method for the edge-loosening defect of cold-rolled 2B sheet according to claim 7, characterized in that, The step S1 also includes achieving multi-sensor clock synchronization based on the precise time protocol, with a time deviation ≤ 10 μs.

9. The two-way tension control and straightening method for the edge-loosening defect of cold-rolled 2B sheet according to claim 1, characterized in that The step S4 includes, when the right loose edge is detected, triggering the reinforcement learning agent to generate a compensation strategy; the magnetorheological clamping roll increases the right transverse tension to 20 MPa within 0.2 s, and at the same time reduces the left tension to 16 MPa; The online learning module records the adjustment effect and updates the GCN edge weights.

10. The two-way tension control straightening method for the loose edge defect of cold-rolled 2B plate according to claim 9, characterized in that, The judgment criterion for detecting the right loose edge is that the residual stress deviation > 12%.

Images