Intelligent size precision regulation and control system and method based on rolled finished product

Abstract

The invention relates to the technical field of intelligent regulation and control of fusion data, in particular to a rolled finished product-based size precision intelligent regulation and control system and method, which comprises a data acquisition unit, a data fusion unit, a deep learning prediction unit and a compensation execution unit, multi-source data such as temperature gradient, roll gap fluctuation and tension spectrum are collected in real time through a sensor array, thermal expansion compensation correction is carried out through a data fusion unit based on a rolling elasticoplastic deformation equation, an anti-interference coupling feature vector is generated by using adaptive Kalman filtering, and a deep learning prediction unit is used for predicting the rolling elasticoplastic deformation through a dual-channel space-time convolutional network. The thickness error and the width error are predicted by combining the constraint of a rolling force balance equation, the compensation execution unit dynamically generates the roll gap adjusting amount and the rolling force compensation amount according to a three-level response mechanism, phase lag compensation current is injected to suppress resonance, the high-end steel yield and the rolling precision are improved, and the rolling quality is improved. And an intelligent closed-loop control scheme is provided for high-precision rolling under complex working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent regulation of fusion data, in particular to a size precision intelligent regulation system and method based on rolled products. BACKGROUND

[0002] Intelligent regulation of fusion data is an important technology. In modern rolling production, size precision directly determines product quality and added value. An intelligent regulation system corrects rolling parameters dynamically by real-time sensing of multi-source data such as temperature, roll gap, and tension in the rolling process, and is a core technology for achieving high-precision and high-stability rolling.

[0003] However, the traditional size precision regulation technology has core problems of multi-physical field data fragmentation and insufficient dynamic coupling modeling. Existing solutions rely on single temperature field models or mechanical vibration models for independent regulation, lack collaborative analysis of multiple factors such as temperature gradient, roll gap fluctuation, and tension oscillation, and when correcting the roll gap by temperature compensation, the mechanical resonance caused by tension fluctuation is not considered simultaneously, which may cause interference between thermal crown compensation and vibration suppression measures. A single prediction model based on physical equations cannot capture nonlinear deformation characteristics. In high-speed rolling, the thickness error prediction deviation can reach ±5 microns due to sensor noise and clock drift problems. In addition, the traditional actuator uses fixed gain compensation, which is prone to system oscillation due to regulation lag or excessive compensation when the size error exceeds the critical value, resulting in an increase in out-of-tolerance rate of finished product size, which seriously affects the yield and production efficiency. To solve this problem, we provide a size precision intelligent regulation system and method based on rolled products. SUMMARY

[0004] The present application aims to provide a size precision intelligent regulation system and method based on rolled products to solve the problems raised in the background.

[0005] Because the traditional scheme has fragmented multi-physical field data and lacks collaborative analysis, resulting in interference between regulation measures and large errors, this case uses a sensor array to collect multi-source data, a data fusion unit to perform thermal expansion compensation correction and adaptive filtering fusion, which can achieve collaborative analysis of multi-source data and improve error prediction accuracy.

[0006] Because the traditional actuator uses fixed gain compensation, which is prone to oscillation due to regulation lag or excessive compensation, this case divides the compensation execution unit into three response zones, dynamically generates roll gap adjustment and rolling force compensation, and injects compensation current, which can accurately and dynamically compensate, reduce the risk of system oscillation, and improve the stability of size precision control.

[0007] One of the objectives of the present application is to provide a size precision intelligent regulation system based on rolled products, comprising: The data acquisition unit collects data streams in the rolling process through a sensor array, the data streams including temperature gradient distribution values in the width direction of the rolled piece, roll gap dynamic fluctuation values, tension fluctuation frequency spectrum characteristic values between rolling sections, and speed harmonic distortion rates of the main transmission of the rolling mill; The data fusion unit performs thermal expansion compensation correction on the temperature gradient distribution values based on an elastic-plastic deformation equation for rolling, generates a mechanism corrected temperature field, and fuses the roll gap dynamic fluctuation values and the tension fluctuation frequency spectrum characteristic values by using an adaptive Kalman filter to generate an anti-interference coupling characteristic vector; The deep learning prediction unit includes a double-channel spatio-temporal convolution network, a first channel of which inputs the mechanism corrected temperature field to extract the mapping relationship between the spatial temperature distribution and the thickness error, and a second channel of which inputs the anti-interference coupling characteristic vector to extract the roll gap-tension coupling oscillation mode, and the rolling force balance equation is embedded into the long short-term memory network weight updating process as a regularization term to output the thickness error prediction value and the width error confidence interval; The compensation execution unit generates roll gap adjustment amounts and rolling force compensation amounts according to the thickness error prediction value and the width error confidence interval, and injects compensation currents in the reverse direction according to the speed harmonic distortion rates.

[0008] As a further improvement of the technical solution, the sensor array configuration method comprises the following steps: An infrared sensor matrix is used in the width direction of the rolled piece, and a multi-spectral fusion technology is used to locate the hot spot area through the inlet / outlet double-waveband compensator; A sensor array and a piezoelectric force sensor are installed on the bearing seat of the work roll, a coupling phase difference between the rolling force and the roll gap vibration is captured based on a phase synchronization technology, and a mechanical resonance characteristic fingerprint library is established; A fiber grating sensor group is arranged at the end of the tension roll shaft of the coiler, a stress gradient distribution cloud map in the width direction of the strip steel is solved through the Bragg wavelength shift amount, and a stress concentration point of the edge wave defect is identified.

[0009] As a further improvement of the technical solution, the method of thermal expansion compensation correction in the data fusion unit comprises: Based on the phase change expansion characteristics of different steel grades, a nonlinear mapping relationship library of temperature gradient to thermal strain is established, and the austenite transformation critical point in the continuous cooling transformation curve is dynamically matched; The thermal expansion equivalent is converted into a dynamic control instruction of the segmented cooling system of the roll, and the thermal crown distortion of the work roll is offset in real time through closed-loop adjustment of the cooling liquid flow; A transmission delay estimation model is used to perform spatio-temporal coordinate transformation on the temperature measurement data, and the thermal conduction lag effect is compensated in the reverse direction with the center of the deformation zone as the reference point.

[0010] As a further improvement of the technical solution, the method of generating the anti-interference coupling characteristic vector in the deep learning prediction unit is: Gaussian noise model is adopted in the low-speed rolling stage, and the high-speed stage is enabled A stable distribution model is adopted, and the model is seamlessly switched through a speed threshold. The low speed is less than 3 m / s, and the high speed is greater than 8 m / s. The amplitude attenuation rate of the fundamental frequency in the tension spectrum is extracted as the mechanical resonance energy index, a resonance danger level evaluation matrix is constructed, and the timestamp information of the high-frequency laser range finder is used to perform sub-millisecond alignment on the roll gap fluctuation sequence and the tension fluctuation sequence, so as to eliminate the clock drift of the sensor.

[0011] As a further improvement of the technical solution, the double-channel spatio-temporal convolution network operation method in the deep learning prediction unit comprises: A three-dimensional hollow convolutional neural network is used to extract the spatio-temporal conduction path features of the temperature field, an implicit hot spot area is located through gradient weight visualization technology, a roll gap-tension frequency domain energy spectrum is input into a graph convolutional network, a propagation topology graph of mechanical vibration in the rolling mill housing is generated, a key conduction path of vibration is identified, and finally a temperature-vibration feature saliency weight distributor is designed to dynamically adjust the double-channel feature fusion weight according to the rolling schedule parameters.

[0012] As a further improvement of the technical solution, the method for embedding the rolling force balance equation in the deep learning prediction unit is: In the back propagation of the long short-term memory network, the predicted rolling force value is forced to be within the ±15% interval of the theoretical value calculated by the Hill formula. When the predicted value deviates from the theoretical value by more than the threshold, the physical regularization weight is automatically increased to 3 times until convergence. The constraint strength is adaptively adjusted based on the material deformation resistance parameter library, realizing zero-sample transfer from high-strength low-alloy steel to silicon steel.

[0013] As a further improvement of the technical solution, the thickness error compensation triggering mechanism in the compensation execution unit comprises: The three-level response zones of emergency, early warning and monitoring are divided with the predicted deviation value, confidence probability and historical gradient as the coordinate axes. When the absolute value of the deviation gradient is greater than 1 micrometer per second for 5 consecutive prediction periods, the preventive compensation is started 200 milliseconds in advance. A regulation amount decay function is established based on the compensation effect historical database to dynamically shrink the output amplitude, so as to avoid system oscillation.

[0014] As a further improvement of the technical solution, the generation method of the compensation current injected in reverse according to the speed harmonic distortion rate is: The inherent resonance source of the rolling mill main motor or roll gear box is located through the odd harmonic proportion and phase characteristics of the main drive current spectrum. The harmonic phase drift amount is dynamically captured by using a digital phase-locked loop technology to compensate for the phase lag caused by the switch delay of the inverter. The harmonic suppression gain is adjusted in sections according to the rolling load rate. The gain is automatically attenuated by 40% at heavy load to prevent overmodulation.

[0015] As a further improvement of the technical solution, the method for generating the roll gap adjustment amount and the rolling force compensation amount is: The following collaborative logic is executed based on the upper and lower boundary values of the width error confidence interval, that is, when the width error confidence interval is less than or equal to 1.5 microns, the rolling force compensation amount is adjusted, and when the width error confidence interval is greater than 1.5 microns, the roll gap adjustment amount proportion is linearly increased from 60% to 95%; The generated roll gap adjustment amount is subjected to a change rate constraint, a first-order inertia link is output, and the current generated rolling force compensation amount is subjected to decay correction according to a historical compensation effect database.

[0016] The second object of the present application is to provide a method for implementing a size precision intelligent control system based on rolling products, comprising the following steps: S1, position the hot spot area through the infrared sensor matrix and the double-band compensator, synchronously capture the roll gap dynamic fluctuation value and the tension fluctuation frequency spectrum characteristic value, use the fiber grating sensor group to solve the strip stress gradient distribution cloud picture, combine the high-frequency laser range finder time stamp to align the multi-source data, and generate the time and space synchronous rolling process data stream; S2, establish a temperature-thermal strain nonlinear mapping library based on the phase change characteristics of the steel grade, perform thermal expansion compensation correction through a transmission delay estimation model, generate a mechanism corrected temperature field, and use an adaptive noise model to fuse the roll gap and tension data, construct an anti-interference coupling feature vector, and extract the fundamental frequency amplitude decay rate to evaluate the resonance danger level; S3, the double-channel space-time convolution network respectively processes the temperature field space-time conduction characteristics and the vibration propagation topology, fuses the double-channel outputs through a dynamic weight distributor, takes the rolling force balance equation as a long short-term memory network regularization term, constrains the prediction value within the interval of the theoretical value of the Hill formula ±15%, and outputs the thickness error prediction value and the width error confidence interval; S4, according to the prediction deviation value, the confidence probability and the historical gradient, three response zones are divided, the compensation is triggered 200ms in advance, the roll gap adjustment amount and the rolling force compensation amount are generated based on the boundary value of the width error confidence interval, and the change rate constraint and the historical decay correction are applied, the main drive harmonic source is positioned synchronously, and the phase lag compensation current is injected to suppress resonance.

[0017] Compared with the prior art, the present application has the following advantages: The application synchronously collects multi-source data such as temperature gradient, roll gap fluctuation, tension spectrum, etc. through a sensor array, combines the thermal expansion compensation correction and adaptive Kalman filtering of the data fusion unit, achieves the function of time and space synchronous multi-physical field data, has the effect of eliminating data fragmentation and improving feature coupling analysis accuracy, solves the problem of large traditional single model regulation error, through the double-channel time and space convolution network of the deep learning prediction unit, fuses the mechanism correction temperature field and anti-disturbance coupling feature vector, embeds the rolling force balance equation constraint prediction value, achieves the function of capturing nonlinear deformation characteristics and suppressing mechanical resonance, solves the problem of insufficient modeling accuracy under complex working conditions, through the three-level response mechanism and dynamic compensation algorithm of the compensation execution unit, combines the speed harmonic distortion rate injection phase lag compensation current, solves the problem of traditional fixed gain compensation lag or overcompensation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 is the overall block diagram of the application; Fig. 2 is the overall flowchart of the application.

[0019] The meanings of various numbers in the figure are as follows: 1, data acquisition unit; 2, data fusion unit; 3, deep learning prediction unit; 4, compensation execution unit. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0021] One of the purposes of the application is to provide a size precision intelligent regulation and control system based on rolled products, please refer to Figs. 1-2 as shown, comprising: The data acquisition unit 1 collects data streams in the rolling process through a sensor array, and the data streams include temperature gradient distribution values in the width direction of the rolled piece, roll gap dynamic fluctuation values, tension fluctuation spectrum characteristic values between rolling sections, and rolling mill main drive speed harmonic distortion rate; When constructing the size precision intelligent regulation and control system based on rolled products, the accurate configuration of the sensor array is the basis for realizing multi-source data acquisition, and the specific implementation manner is as follows: The sensor array configuration method comprises the following steps: To monitor the temperature distribution of the rolled piece in real time, a matrix of infrared sensors is uniformly arranged in the width direction of the rolled piece (e.g., an infrared temperature measurement probe is installed every 50 mm), covering the entire rolling area. Through double-band compensators (using visible light and near-infrared light double-band detection) at the inlet and outlet, the infrared temperature measurement data is compensated for environmental radiation. A multi-spectral fusion algorithm is used to eliminate the influence of iron oxide scale thickness on temperature measurement accuracy, and to locate the hot spot area of temperature anomalies (areas with a temperature deviation of more than ± 15°C from the average value). For example, when a sudden temperature rise is detected at the edge of the rolled piece, the system can quickly identify the location of the hot spot and trigger the cooling system to adjust, reducing temperature measurement errors and improving the accuracy of thermal expansion compensation. A piezoelectric force sensor and an acceleration sensor array are integrated into the work roll chock. The former is used to collect rolling force fluctuation signals, and the latter is used to monitor roll gap vibration acceleration. Based on phase synchronization technology, the phase difference between rolling force and roll gap vibration signals is calculated through Hilbert transform, and the coupling phase difference of the two is captured in real time. When the phase difference exceeds ± 30°, it is determined that there is a risk of mechanical resonance. By continuously collecting phase difference data under different rolling conditions, a mechanical resonance characteristic fingerprint library is established. The characteristic fingerprint library contains information such as steel grade, rolling speed, and phase difference threshold. For example, a dedicated fingerprint model is established for silicon steel rolling to improve the accuracy of resonance warning. Compared with traditional single sensor monitoring, it is improved. Optical fiber grating sensors are deployed at the ends of the tension roll of the coiler (4-6 sensors are installed at each end). Using the Bragg wavelength shift principle (wavelength shift is proportional to stress), the stress gradient distribution cloud map in the width direction of the strip is calculated in real time. When the wavelength shift at the edge area exceeds 2 pm / µm, it is determined that there is a stress concentration point of edge wave defect. The defect position is located through the stress-strain inverse calculation model. This detection method can identify potential defects before the strip is completely cooled, reducing the edge wave defect rate. Through the above sensor array configuration, the system realizes multi-dimensional accurate monitoring of temperature, force-vibration coupling, and stress distribution during rolling. The data of each sensor is time and space aligned through a synchronous clock, providing high-reliability raw data for subsequent data fusion and precision control, and enhancing the state perception ability of the rolling process, laying a solid foundation for intelligent control of size precision.

[0022] After completing the multi-source data collection during rolling, the data fusion unit 2 needs to perform thermal expansion compensation correction on the temperature gradient data to eliminate the influence of temperature changes on the size precision of the rolled piece. The specific implementation is as follows: The data fusion unit 2 performs thermal expansion compensation correction on the temperature gradient distribution value based on the elastic-plastic deformation equation of rolling, generates a mechanism-corrected temperature field, and uses adaptive Kalman filtering to fuse the roll gap dynamic fluctuation value and the tension fluctuation spectral characteristic value, generating an anti-disturbance coupling feature vector. The method for thermal expansion compensation correction in the data fusion unit 2 comprises: According to the phase transition expansion characteristics of different steel grades, the corresponding data of temperature gradient and thermal strain are obtained through thermal simulation experiments, a nonlinear mapping relationship library is established by using a polynomial fitting algorithm, and meanwhile, the critical points of austenite transformation of each steel grade are determined based on continuous cooling transformation curves. When the measured temperature exceeds the critical point, the high-temperature expansion coefficient model is automatically switched to, for example, for Q345B steel, when the temperature is higher than 727 DEG C, the austenite zone expansion coefficient 0.012 mm / (m o C) is enabled, and when the temperature is lower than this temperature, the ferrite zone coefficient 0.010 mm / (m o C) is used, thereby reducing the thermal strain calculation error and improving the adaptability of the model to the differences between steel grades; The calculated thermal expansion equivalent (unit: μm) is converted into control instructions of the segmented cooling system of the roll, and the cooling liquid flow of each cooling section (for example, divided into 10 cooling sections, and each section is independently controlled) is adjusted through a proportional-integral-derivative (PID) controller. When the thermal expansion equivalent of a certain region exceeds 50 μm, the cooling flow of the corresponding cooling section is increased by 20%, and the thermal crown distortion of the work roll is real-timely offset through closed-loop feedback control (for example, the thermal crown is reduced from 0.08 mm to 0.03 mm). For example, when rolling silicon steel, the temperature difference between the middle and the edge of the roll is reduced from 12 DEG C to 5 DEG C by dynamically adjusting the cooling flow, which significantly reduces the thickness deviation caused by the thermal deformation of the roll, and the thickness precision is improved to ±5 μm; Since there is a delay in the transmission of temperature signals from the surface of the rolled piece to the sensor, a transmission delay estimation model (based on the Fourier heat conduction equation) is used to perform space-time coordinate transformation on the temperature measurement data. Taking the center of the deformation zone as the reference point, the delay time is calculated according to the heat conduction distance and the material thermal conductivity coefficient (for example, when the thickness of the rolled piece is 20 mm, the delay time is about 0.8 seconds), and the measured temperature data is compensated to the coordinates at the time of deformation, thereby eliminating the influence of the lag effect. For example, the inlet temperature data is mapped to the deformation zone 0.8 seconds in advance, so that the time synchronization error of the temperature field and the deformation of the rolled piece is reduced from ±1.5 seconds to ±0.2 seconds, thereby improving the real-time performance of the thermal expansion compensation, and cooperating with the cooling system adjustment to reduce the size error caused by thermal deformation; Through the above-mentioned thermal expansion compensation correction method, the system realizes dynamic compensation of the whole process from temperature data acquisition, model calculation to cooling control, effectively solves the problem of size precision fluctuation caused by temperature change in the rolling process, and in actual application, the mechanism improves the accuracy of thermal expansion compensation for different steel grades, significantly improves the stability of rolling size precision, and especially in multi-steel grade switching production, manual adjustment of compensation parameters is not required, thereby improving the production efficiency.

[0023] After the data fusion unit 2 completes the preliminary processing of the anti-interference features, the deep learning prediction unit 3 needs to generate an anti-interference coupling feature vector to improve the robustness of the prediction model, and the specific implementation manner is as follows: The deep learning prediction unit 3 comprises a dual-channel space-time convolution network, a first channel inputs a mechanism corrected temperature field, extracts a mapping relationship between a spatial temperature distribution and a thickness error, a second channel inputs an anti-interference coupling feature vector, extracts a roll gap-tension coupling oscillation mode, and embeds a rolling force balance equation into a long short-term memory network weight updating process as a regularization term, and outputs a thickness error prediction value and a width error confidence interval; The anti-interference coupling feature vector generation method in the deep learning prediction unit 3 is as follows: Considering the noise characteristic difference between low-speed and high-speed rolling stages, the system presets 3 m / s and 8 m / s as speed thresholds, when the rolling mill main drive speed is less than 3 m / s, it is determined that it is in a low-speed stage, a Gaussian noise model is used to describe the random interference of tension and roll gap fluctuation, the model constructs a normal distribution curve by calculating the mean and variance of historical data, effectively inhibiting the influence of low-frequency random noise on feature extraction, when the speed is greater than 8 m / s, a stable distribution model is enabled, the feature parameters are fitted by the maximum likelihood estimation method, and the pulse noise interference under high speed is coped with, the rolling speed is monitored in real time by a speed sensor, the model is triggered to realize seamless switching logic, and the noise suppression efficiency under different working conditions is improved; The collected tension fluctuation signal is subjected to fast Fourier transform, the fundamental frequency component (such as a typical fundamental frequency of 20 Hz) in the frequency spectrum is extracted, the fundamental frequency amplitude decay rate (unit: dB / s) is calculated as a mechanical resonance energy index, when the decay rate is greater than 5 dB / s, it is determined that the resonance energy is low, and the system is safe, when the decay rate is less than 2 dB / s, a three-dimensional resonance danger level evaluation matrix containing amplitude, frequency and steel grade is constructed, and potential resonance risks are warned in advance, meanwhile, the nanosecond level timestamp information of the high-frequency laser range finder is used to perform sub-millisecond level time alignment on the roll gap fluctuation sequence and the tension fluctuation sequence: The time points of the two sequences are matched through a dynamic time warping algorithm, the phase deviation caused by sensor clock drift is eliminated, the space-time consistency of the coupling feature is ensured, the phase error of the feature vector is reduced, and the accuracy of subsequent vibration mode analysis is improved; Through the above anti-interference coupling feature vector generation method, the system realizes adaptive suppression of rolling process noise and precise alignment of multi-source signals, effectively improves the reliability of the feature vector, provides high-quality input data for accurate prediction of the dual-channel space-time convolution network, reduces the root mean square error of the thickness error prediction, and improves the coverage of the width error confidence interval.

[0024] When the rolling force prediction model is constructed in the deep learning prediction unit 3, in order to ensure that the prediction result conforms to the physical law, the rolling force balance equation needs to be embedded into the training process of the long short-term memory network as a regularization term, and the specific implementation is as follows: The method for embedding the rolling force balance equation in the deep learning prediction unit 3 is as follows: Based on the Hill formula (describing the relationship between rolling force and material deformation, roll gap, etc.), the theoretical rolling force value is calculated in advance , and the constraint interval of the predicted value is set as [0.85 , 1.15 ]. During the back propagation process of the long short-term memory network, the deviation between the predicted value and the theoretical value is included in the loss function calculation, and the formula is ; wherein, is the regularization weight coefficient, and the initial value is set as 1. When the predicted value exceeds the constraint interval, the physical regularization mechanism is activated to force the adjustment of the network weight to make the predicted value converge to a reasonable range, avoiding abnormal prediction that violates the principles of rolling force, such as preventing the predicted rolling force from being lower than the empty load value or exceeding the rated load of the equipment, and improving the physical interpretability of the model; The degree of deviation of the predicted value from the theoretical value is monitored in real time, and when the deviation exceeds the threshold, the regularization weight coefficient is increased to 3 times, enhancing the influence of physical constraints on model training. For example, if the prediction deviation reaches 20% at a certain time, the weight is increased from 1 to 3, increasing the proportion of the physical constraint term in the loss function, forcing the network to adjust the weight to reduce the deviation, until the predicted value returns to the ±15% interval. At the same time, the constraint strength is adaptively adjusted based on the material deformation resistance parameter library: for high-strength low-alloy steel, due to its high deformation resistance and large rolling force fluctuation range, the constraint interval is dynamically relaxed to ±20%, and for soft magnetic materials such as silicon steel, the constraint interval is tightened to ±10%, ensuring the prediction accuracy of different steel grades; Using the prior knowledge in the material deformation resistance parameter library, when switching to a new steel grade, no additional data labeling is required, and the yield strength and rolling temperature in the parameter library are directly used to recalculate the theoretical rolling force based on the Hill formula, wherein, ; is the width of the rolled piece, is the average rolling thickness, and the input feature weight of the long short-term memory network is automatically adjusted to realize zero-shot transfer, for example, when rolling silicon steel, the model automatically reduces the rolling force prediction value based on the yield strength of silicon steel in the parameter library, significantly improving the adaptability of the model to new steel grades, reducing the cost of manual parameter adjustment, and shortening the model training time when switching steel grades; By embedding the rolling force balance equation, the system deeply integrates physical constraints and data-driven models, making the prediction results conform to the rolling force laws and adapt to the process differences of different steel grades. When switching between high-strength low-alloy steel and silicon steel with large differences, the prediction accuracy remains stable, and the generalization ability of the model is significantly enhanced. This provides a solid prediction foundation for the compensation execution unit 4 to generate reliable roll gap adjustment and rolling force compensation, and effectively improves the overall size accuracy control.

[0025] After the deep learning prediction unit 3 outputs the thickness error prediction value, the compensation execution unit 4 needs to dynamically trigger the compensation mechanism according to the error trend to achieve accurate control of the size accuracy. The specific implementation is as follows: The compensation execution unit 4 generates roll gap adjustment and rolling force compensation based on the thickness error prediction value and width error confidence interval, and injects compensation current in reverse according to the speed harmonic distortion rate.

[0026] The thickness error compensation triggering mechanism in the compensation execution unit 4 includes: To achieve hierarchical control of thickness error, the system uses prediction deviation (unit: microns), confidence probability (model confidence in prediction results), and historical gradient (unit: microns / second) as three-dimensional coordinate axes to construct a three-level response space. The monitoring area is the area where the prediction deviation is less than or equal to 5 microns, the confidence probability is less than 0.8, and the absolute value of the historical gradient is less than 1 micron / second. Only data monitoring is performed. The warning area is the area where the prediction deviation is 5-10 microns or the confidence probability is greater than or equal to 0.8, and the absolute value of the historical gradient is 1-2 microns / second. Trigger a warning prompt. The emergency area is the area where the prediction deviation is greater than 10 microns and the confidence probability is greater than or equal to 0.9, and the absolute value of the historical gradient is greater than or equal to 2 microns / second. Immediately start compensation, for example, when the prediction deviation is 8 microns, the confidence probability is 0.85, and the historical gradient is 1.5 microns / second, it is determined that it enters the warning area, and the system response speed is improved by preparing compensation instructions in advance. To avoid continuous error expansion, when the absolute value of the deviation gradient of the last 5 prediction periods (each period is 200 milliseconds) is greater than 1 micron / second, it is determined that the error is accelerating and deteriorating, and the preventive compensation is triggered 200 milliseconds in advance (i.e. adjust the execution earlier than the theoretical trigger time point), for example, if the current period deviation is 4 microns, the next period increases to 5.5 microns, and the gradient is 1.5 microns / second, after 5 times, the system triggers compensation in advance, so that the thickness error starts to converge before reaching the emergency threshold. Compared with traditional lag compensation, the error peak is reduced. Based on the compensation effect historical database (storing adjustment amount and error correction data under different working conditions), an adjustment amount decay function is established ; wherein, is the number of compensations, is the attenuation coefficient, taking a value of 0.8-0.95, the output amplitude is automatically contracted after each compensation, for example, the first compensation amount is 20 microns, the second compensation amount is attenuated to 20x0.9=18 microns, to avoid system oscillation caused by repeated compensation, by real-time monitoring of error change rate, dynamic adjustment of attenuation coefficient: if the error correction rate is greater than 1.2 microns / sec, the attenuation coefficient is 0.8 to quickly converge, if the correction rate is less than 0.5 microns / sec, the attenuation coefficient is 0.95 to prevent over-regulation, the thickness error fluctuation after compensation is reduced by 55%, the system stability is significantly improved; Through the above thickness error compensation trigger mechanism, the system realizes intelligent control of the whole process from risk warning, early intervention to dynamic adjustment, effectively solving the problems of hysteresis and over-regulation of traditional fixed threshold compensation, and cooperating with the synergistic effect of rolling force compensation and roll gap adjustment, the thickness precision is controlled within ±5 microns, the width precision is controlled within ±8 microns, and the size precision stability and production efficiency of the rolled product are significantly improved.

[0027] While the compensation execution unit 4 generates the roll gap adjustment amount and the rolling force compensation amount, it needs to suppress the harmonic interference of the main drive system of the rolling mill. The vibration suppression is realized by reverse injection of compensation current. The specific implementation is as follows: The generation method of reverse injection of compensation current according to the speed harmonic distortion rate is as follows: The main drive current signal is collected by the current sensor, and the odd harmonic components in the frequency spectrum are analyzed by fast Fourier transform. When the proportion of a certain frequency harmonic exceeds 15% of the fundamental wave, it is determined as a natural harmonic source. Combined with phase characteristic analysis, if the phase difference between the odd harmonic and the fundamental wave exceeds ±60°, the harmonic source position is further confirmed: The main motor resonance usually shows that the 3rd harmonic proportion is high and the phase lags, and the gear box resonance shows that the 5th harmonic proportion is high and the phase leads. For example, when the 5th harmonic proportion is detected to be 20% and the phase leads by 45°, the gear box of the roll system is located as the harmonic source, providing a basis for accurate compensation; The phase drift of the odd harmonic is dynamically captured by using the digital phase-locked loop technology (by tracking the harmonic phase through closed-loop feedback), the phase lag time caused by the inverter switching delay is calculated, and the compensation signal opposite to the lagging phase is injected in advance in the current command through the feedforward compensation algorithm, for example, when the measured lagging phase is 30°, a compensation phase of-30° is generated, the harmonic phase error is reduced, the influence of switching delay on harmonic suppression effect is effectively offset, and the real-time performance of the compensation current is improved; According to the rolling load rate (the ratio of the actual rolling force to the rated rolling force), the compensation process is divided into three sections: light load (less than 50%), medium load (50%-80%), and heavy load (more than 80%), and the harmonic suppression gain coefficient is set respectively. The gain is set to 1.0 when the load is light, it is automatically attenuated to 0.7 when the load is medium, and it is further attenuated by 40% to 60% when the load is heavy. For example, when the load rate reaches 90%, the gain decreases from 1.0 to 0.6, which limits the compensation current amplitude within 10% of the rated current, preventing the risk of motor overheating caused by over-modulation under heavy load. By monitoring the load sensor data in real time, the gain level is dynamically switched, and the switching delay is less than 100 milliseconds, ensuring the balance between harmonic suppression effect and system safety under different working conditions. Through the above-mentioned generation method of reverse injection compensation current, the system realizes accurate positioning and dynamic suppression of the inherent resonance of the main drive system, significantly improves the operation stability of the main drive system of the rolling mill, and cooperates with the roll gap and rolling force compensation to improve the comprehensive control efficiency of the finished product size precision, providing reliable power system protection for high-precision rolling production.

[0028] When the compensation execution unit 4 generates the control command according to the width error confidence interval output by the deep learning prediction unit 3, the size precision is controlled in detail through cooperative logic and dynamic correction. The specific implementation is as follows: The method for generating the roll gap adjustment amount and the rolling force compensation amount is: To balance the stability of the rolling force adjustment and the sensitivity of the roll gap adjustment, the system executes differentiated adjustment strategies according to the upper and lower boundary values of the width error confidence interval (reflecting the fluctuation range of the predicted error). When the confidence interval is less than or equal to 1.5 microns, it indicates that the width error prediction value is highly reliable and has small fluctuations, and the error is corrected by the rolling force compensation amount adjustment first, because the rolling force adjustment has more linear characteristics on the width error, which can avoid mechanical wear caused by frequent roll gap actions. For example, when the confidence interval is 1.2 microns, the width error convergence speed is improved by rolling force adjustment. When the confidence interval is greater than 1.5 microns, it is determined that the error fluctuation is large, and the proportion of roll gap adjustment needs to be increased until the interval shrinks to the target range. For example, when the confidence interval is 2.0 microns, the proportion of roll gap adjustment is increased, and the rolling force compensation is used to shorten the convergence time of the width error, effectively meeting the rapid correction demand in the large error scenario. To prevent the roll gap from adjusting too fast and causing abnormal plate shape, the generated adjustment amount is subjected to a change rate constraint, and the output command is smoothed through a first-order inertia link, so that the roll gap opening change rate is limited to within 0.5 mm / s. For example, a 200-micron adjustment amount is executed gradually in 5 cycles (40 milliseconds per cycle), reducing mechanical impact. At the same time, based on the historical compensation effect database (storing adjustment amount and error correction data for different steel types and rolling speeds), the rolling force compensation amount is subjected to attenuation correction. If the error correction rate after the previous compensation is less than the expected value, the current compensation amount is multiplied by the decay coefficient 0.9 to avoid over-regulation caused by accumulated errors, and if the correction rate exceeds the expected value, the original compensation amount is maintained, for example, when rolling silicon steel, historical data shows that the decay coefficient of rolling force compensation is usually 0.95, through dynamic correction, the effectiveness of the compensation amount is improved, and the average correction accuracy of the width error is improved; Through the above-mentioned synergistic regulation and dynamic correction mechanism, the system realizes multi-mode accurate control of the width error, which not only ensures the smoothness of the regulation in the small error scene, but also improves the response speed in the large error scene, reduces the energy loss of the rolling force compensation, cooperates with the temperature compensation and vibration suppression module, improves the comprehensive efficiency of the overall size precision regulation, and significantly improves the width precision stability and production efficiency of the rolling product.

[0029] The present application collects temperature gradient, roll gap fluctuation, tension spectrum and other multi-source data in real time through the sensor array of the data acquisition unit 1, performs thermal expansion compensation correction based on the rolling elastic-plastic deformation equation through the data fusion unit 2, and generates an anti-interference coupling feature vector using adaptive Kalman filtering, the deep learning prediction unit 3 realizes the prediction of thickness error and width error through a double-channel spatio-temporal convolution network combined with rolling force balance equation constraint, and the compensation execution unit 4 dynamically generates roll gap adjustment and rolling force compensation according to the three-level response mechanism, and injects phase lag compensation current to suppress resonance, improves the finished product rate and rolling precision of high-end steel, and provides an intelligent closed-loop control scheme for high-precision rolling under complex working conditions.

[0030] The second object of the present application is to provide a method for realizing a size precision intelligent regulation system based on rolling products, comprising the following steps: S1, positioning the hot spot area through the infrared sensor matrix and the double-waveband compensator, synchronously capturing the roll gap dynamic fluctuation value and the tension fluctuation frequency spectrum characteristic value, using the fiber grating sensor group to solve the strip stress gradient distribution cloud picture, combining the time stamp alignment multi-source data of the high-frequency laser range finder, and generating a spatio-temporal synchronous rolling process data stream; S2, establishing a temperature-thermal strain nonlinear mapping library based on the phase change characteristics of the steel grade, performing thermal expansion compensation correction through a transmission delay estimation model, generating a mechanism correction temperature field, and using an adaptive noise model to fuse the roll gap and tension data, constructing an anti-interference coupling feature vector, and extracting the fundamental frequency amplitude decay rate to evaluate the resonance danger level; S3, the double-channel spatio-temporal convolution network respectively processes the temperature field spatio-temporal conduction characteristics and the vibration propagation topology, fuses the double-channel outputs through a dynamic weight distributor, takes the rolling force balance equation as a long short-term memory network regularization term, constrains the prediction value within the interval of ±15% of the theoretical value of the Hill formula, and outputs the thickness error prediction value and the width error confidence interval; S4, according to the predicted bias value, confidence probability and historical gradient, three response zones are divided, 200ms in advance trigger compensation, based on the width error confidence interval boundary value, the roll gap adjustment amount and the rolling force compensation amount are generated, and the change rate constraint and the historical decay correction are applied, the main drive harmonic source is positioned synchronously, and the phase lag compensation current is injected to suppress the resonance.

[0031] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A size accuracy intelligent control system based on rolled products, characterized by, The application relates to a rolling process thickness and width error prediction method based on a sensor array and a data fusion unit. The data acquisition unit (1) collects data streams in a rolling process through a sensor array, the data streams comprising temperature gradient distribution values in a width direction of a rolled piece, roll gap dynamic fluctuation values, tension fluctuation frequency spectrum characteristic values between rolling sections and a speed harmonic distortion rate of a rolling mill main drive; The data fusion unit (2) performs thermal expansion compensation correction on the temperature gradient distribution values based on an elastic-plastic deformation equation of rolling, generates a mechanism correction temperature field, and fuses the roll gap dynamic fluctuation values and the tension fluctuation frequency spectrum characteristic values by adopting an adaptive Kalman filter to generate an anti-interference coupling characteristic vector; The deep learning prediction unit (3) comprises a double-channel space-time convolution network, a first channel inputs the mechanism correction temperature field, extracts a mapping relationship between a spatial temperature distribution and a thickness error, a second channel inputs the anti-interference coupling characteristic vector, extracts a roll gap-tension coupling oscillation mode, and embeds a rolling force balance equation into a long short-term memory network weight updating process as a regularization term, and outputs a thickness error prediction value and a width error confidence interval; The compensation execution unit (4) generates roll gap adjustment amounts and rolling force compensation amounts according to the thickness error prediction value and the width error confidence interval, and injects compensation currents in a reverse direction according to the speed harmonic distortion rate.

2. The intelligent control system for dimensional accuracy of rolled products according to claim 1, wherein, The sensor array configuration method comprises the following steps: An infrared sensor matrix is arranged in the width direction of the rolled piece, a multi-spectrum fusion technology is used to position a hot spot area in combination with an inlet / outlet double-waveband compensator; A sensor array and a piezoelectric force sensor are arranged on a work roll bearing seat, a coupling phase difference between rolling force and roll gap vibration is captured based on a phase synchronization technology, and a mechanical resonance characteristic fingerprint library is established; A group of fiber grating sensors are arranged on the shaft end of a coiler tension roll, a stress gradient distribution cloud picture in the width direction of the strip steel is solved through a Bragg wavelength shift amount, and a stress concentration point of an edge wave defect is identified.

3. The system according to claim 1, wherein, The method for thermal expansion compensation correction in the data fusion unit (2) comprises the following steps: A nonlinear mapping relationship library of temperature gradient to thermal strain is established based on phase change expansion characteristics of different steel grades, and austenite transformation critical points in a continuous cooling transformation curve are dynamically matched; A thermal expansion equivalent is converted into a dynamic control instruction of a segmented cooling system of a roll, and a work roll thermal crown distortion is offset in real time through cooling liquid flow closed loop adjustment; A space-time coordinate transformation is performed on temperature measurement data by using a transmission delay estimation model, and a thermal conduction lag effect is reversely compensated with a deformation area center as a reference point.

4. The intelligent control system for dimensional accuracy of rolled products according to claim 1, wherein, The method for generating the anti-interference coupling characteristic vector in the deep learning prediction unit (3) comprises the following steps: Gaussian noise model is used in low speed rolling stage, and high speed stage is enabled Stable distribution model, seamlessly switch the model through the speed threshold trigger, the low speed is less than 3 m / s, high speed is greater than 8 m / s; A mechanical resonance energy index is constructed by extracting a fundamental frequency amplitude attenuation rate in a tension spectrum, a resonance danger level evaluation matrix is constructed, and a roll gap fluctuation sequence and a tension fluctuation sequence are aligned at a sub-millisecond level by using timestamp information of a high-frequency laser range finder, so that sensor clock drift is eliminated.

5. The intelligent control system for dimensional accuracy of rolled products according to claim 1, wherein, The operation method of the double-channel space-time convolution network in the deep learning prediction unit (3) comprises the following steps: The three-dimensional hollow convolutional neural network is used to extract the space-time conduction path features of the temperature field, the implicit hot spot area is located through the gradient weight visualization technology, the roll gap-tension frequency energy spectrum is input into the graph convolutional network to generate the propagation topology graph of mechanical vibration in the rolling mill housing, the vibration key conduction path is identified, and finally, a temperature-vibration feature saliency weight distributor is designed to dynamically adjust the double-channel feature fusion weight according to the rolling schedule parameters.

6. The system for intelligent control of dimensional accuracy of rolled products according to claim 1, wherein The method for embedding the rolling force balance equation in the deep learning prediction unit (3) is: In the long short-term memory network back propagation, the predicted rolling force value is forced to be within the ±15% interval of the theoretical value calculated by the Hill formula. When the predicted value deviates from the theoretical value by more than a threshold value, the physical regularization weight is automatically increased to 3 times until convergence. The constraint strength is adaptively adjusted based on the material deformation resistance parameter library, realizing zero-sample transfer from high-strength low-alloy steel to silicon steel.

7. The system for intelligent control of dimensional accuracy of rolled products according to claim 1, wherein The thickness error compensation trigger mechanism in the compensation execution unit (4) includes: The three-level response zones of emergency, early warning and monitoring are divided by taking the predicted deviation value, confidence probability and historical gradient as the coordinate axes. When the absolute value of the deviation gradient is greater than 1 micrometer per second for 5 consecutive prediction periods, the preventive compensation is started 200 milliseconds in advance. The adjustment amount decay function is established based on the compensation effect historical database to dynamically shrink the output amplitude, which is used to avoid system oscillation.

8. The system for intelligent control of dimensional accuracy of rolled products according to claim 1, wherein The generation method of the compensation current injected in reverse according to the speed harmonic distortion rate is: The inherent resonance source of the rolling mill main motor or roll gear box is located through the odd harmonic proportion and phase characteristics of the main drive current spectrum. The digital phase-locked loop technology is used to dynamically capture the harmonic phase drift. The phase lag caused by the inverter switching delay is compensated. The harmonic suppression gain is adjusted in sections according to the rolling load rate. The gain is automatically attenuated by 40% at heavy load to prevent over-modulation.

9. The system for intelligent control of dimensional accuracy of rolled products according to claim 1, wherein The method for generating the roll gap adjustment amount and the rolling force compensation amount is: The following collaborative logic is executed based on the upper and lower boundary values of the width error confidence interval, that is, when the width error confidence interval is less than or equal to 1.5 micrometers, the rolling force compensation amount is adjusted, and when the width error confidence interval is greater than 1.5 micrometers, the roll gap adjustment amount proportion is linearly increased from 60% to 95%; The generated roll gap adjustment amount is subjected to a change rate constraint, the instruction is output through a first-order inertia link, and the current generated rolling force compensation amount is subjected to decay correction based on the historical compensation effect database.

10. A method for implementing a system for intelligent control of dimensional accuracy of rolled products according to any one of claims 1-9, characterized in that, The method comprises the following steps: S1, the hot spot area is located through the infrared sensor matrix and the double-band compensator, the roll gap dynamic fluctuation value and the tension fluctuation frequency spectrum characteristic value are synchronously captured, the strip stress gradient distribution cloud picture is solved by using the fiber grating sensor group, the time stamp alignment multi-source data of the high-frequency laser range finder is combined, and the space-time synchronous rolling process data stream is generated; S2, a temperature-thermal strain nonlinear mapping library is established based on the phase change characteristics of the steel grade, a thermal expansion compensation correction is performed through a transmission delay estimation model, a mechanism corrected temperature field is generated, an adaptive noise model is used to fuse the roll gap and tension data, an anti-interference coupled feature vector is constructed, and the resonance danger level is evaluated by using the fundamental frequency amplitude decay rate. S3, the dual-channel space-time convolution network processes the temperature field space-time conduction characteristics and vibration propagation topology respectively, fuses the dual-channel outputs through a dynamic weight distributor, takes the rolling force balance equation as a long short-term memory network regularization term, and constrains the prediction value within a ±15% interval of the theoretical value of the Hill formula, to output the thickness error prediction value and the width error confidence interval; S4, according to the prediction deviation value, the confidence probability and the historical gradient, three response zones are divided, the compensation is triggered 200ms in advance, based on the boundary value of the width error confidence interval, the roll gap adjustment amount and the rolling force compensation amount are generated, and the change rate constraint and the historical decay correction are applied, the main drive harmonic source is positioned synchronously, and the phase lag compensation current is injected to suppress the resonance.

Images