Convex letter steel rolling plate shape control method and system based on deep learning

Abstract

The present application relates to the technical field of deep learning machine vision control convex letter steel rolling, and relates to a convex letter steel rolling plate shape control method and system based on deep learning, which comprises the following steps: heating and descaling a blank, setting a first visual detection unit at the outlet of a heating furnace, triggering hydraulic correction when the centering deviation is greater than a first threshold value; setting a second visual detection unit at a rough rolling mill, extracting the residual area of the rolling piece surface oxide scale through an image segmentation algorithm, and detecting whether to trigger secondary descaling; setting a third visual detection unit at a medium rolling mill, adjusting the flow distribution of the cooling water of a finishing mill if the temperature distribution is uneven, and correcting the side pressure of the finishing mill vertical roll if the corner radius error exceeds a fourth threshold value; setting a fourth visual detection unit at the finishing mill, calculating the roll gap compensation amount and correcting the finishing mill reduction amount in real time when the filling rate is less than a fifth threshold value. The present application combines convex letter steel with deep learning and machine vision, and can make the quality of the manufactured convex letter steel better and the shape more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning machine vision control convex letter steel rolling, in particular to a convex letter steel rolling plate shape control method and system based on deep learning. BACKGROUND

[0002] In the convex letter steel rolling production process, plate shape control is a key link to ensure product quality and production efficiency. The traditional convex letter steel rolling plate shape control method mainly relies on manual experience and some relatively simple sensor monitoring methods, which has many limitations.

[0003] The traditional descaling method often fails to comprehensively and accurately remove the oxide layer. The residual oxide layer not only affects the friction force distribution in the rolling process, but also may cause defects on the surface of the rolled piece, affecting the quality of the finished product. And the precision of the centering method is low, which cannot correct the deviation in time and accurately, leading to inaccurate hole positioning in the subsequent rolling process, affecting the shape and size accuracy of the rolled piece.

[0004] In the traditional rough rolling stage, there is a lack of real-time monitoring and accurate adjustment means, which makes it difficult to ensure that the spread of the rolled piece meets the design requirements, and easily causes size deviation of the rolled piece; in the intermediate rolling stage, the traditional detection method has insufficient accuracy for the fillet radius of the transition area between the flange and the waist of the rolled piece, and cannot timely discover and correct the fillet radius deviation, leading to the shape of the rolled piece not meeting the design requirements; in the finishing rolling stage, there is a lack of effective real-time monitoring and accurate adjustment mechanism, and the detection and processing of size tolerance are also relatively lagging, when the deviation is found, it has often produced unqualified products, causing resource waste and reducing production efficiency. SUMMARY

[0005] The present application provides a convex letter steel rolling plate shape control method and system based on deep learning.

[0006] The technical scheme of the present application is as follows:

[0007] The convex letter steel rolling plate shape control method based on deep learning comprises the following steps:

[0008] S1, heating the billet to reach the preset roughing temperature range, and removing the surface oxide layer by high-pressure water descaling;

[0009] S2, a first visual detection unit is arranged above the conveying roller way between the heating furnace outlet and the blooming mill to identify the edge position of the billet, and the center line offset of the billet is calculated through three-dimensional contour reconstruction, when the centering deviation is greater than a first threshold value, the hydraulic correction is triggered, and the centering data is transmitted to the roughing mill in real time as the reference for the hole positioning of the roughing mill;

[0010] S3, a second visual detection unit is arranged at the outlet side of the rough rolling mill, and the residual area of the oxide scale on the surface of the rolled piece is extracted in real time through an image segmentation algorithm. When the residual area is detected to be greater than a second threshold value, secondary descaling is triggered; the spread amount is calculated synchronously, and if the deviation between the measured spread amount and the theoretical value is greater than a third threshold value, the side pressure amount of the edger roller of the intermediate rolling mill is adjusted;

[0011] S4, a third visual detection unit is arranged at the outlet side of the intermediate rolling mill, and the temperature distribution of the rolled piece is monitored in real time. When the temperature distribution is detected to be uneven, the flow distribution of the cooling water of the finishing rolling mill is adjusted; the transition area between the flange and the waist of the rolled piece is scanned, and if the measured value of the fillet radius deviates from the design value by more than a fourth threshold value, the side pressure amount of the edger roller of the finishing rolling mill is corrected;

[0012] S5, a fourth visual detection unit is arranged at the outlet side of the finishing rolling mill, and the flange cross-sectional profile is generated through three-dimensional point cloud reconstruction. The filling rate is calculated by comparing the theoretical model, and when the filling rate is less than a fifth threshold value, the roll gap compensation amount is calculated, and the reduction amount of the finishing rolling mill is corrected in real time. At the same time, the cross-sectional size tolerance of the rolled piece is detected, and if the flange deviation is greater than a sixth threshold value, the rolled piece is sent back to the finishing rolling mill for re-rolling until the output rolled piece with the flange deviation within the qualified range.

[0013] The specific operation of S3 for adjusting the side pressure amount of the edger roller of the intermediate rolling mill is to calculate the side pressure adjustment amount of the edger roller:

[0014] ΔS=αΔW+C,

[0015] Where ΔS is the side pressure adjustment amount of the edger roller, α is the spread compensation coefficient, ΔW is the deviation between the measured spread amount and the theoretical value, and C is the system error compensation term.

[0016]

[0017] Where σ s is the yield strength of the rolled piece material, and β is the equalization coefficient, usually 300.

[0018] The third visual detection unit is also provided with an infrared camera to detect the temperature difference of the rolled piece. If the temperature difference between the lowest temperature and the highest temperature of the rolled piece is greater than the maximum temperature difference threshold, the flow of the cooling water to the high-temperature area corresponding to the highest temperature is increased to Q new :

[0019] Q new =Q0×[1+0.02(ΔT-50)],

[0020] Where Q new is the adjusted cooling water flow, Q0 is the original set cooling water flow, and ΔT is the temperature difference of the rolled piece.

[0021] At the same time, the rolling speed of the rough rolling mill is increased to v new :

[0022] vnew = v0 x [1 + 0.01 (50 - ΔT)],

[0023] wherein, v new is the adjusted rolling speed, v0 is the original set rolling speed.

[0024] In S2, the three-dimensional profile reconstruction can also calculate the cross-sectional asymmetry of the rolled piece. If the cross-sectional asymmetry exceeds the sixth threshold value, the rolled piece is sent back to the heating furnace for reheating.

[0025] The specific operation of correcting the finishing mill reduction in S5 is to adjust the finishing mill reduction to Δh:

[0026]

[0027] wherein, Δh is the compensated total reduction, h0 is the original set reduction, η is the flange filling rate, and K is the material deformation compensation coefficient.

[0028] After adjusting the finishing mill reduction, the rolling speed is reduced by 10%-15% of the original rolling speed.

[0029] The fourth visual detection unit identifies the scale indentation defects on the rolled piece based on the spectral characteristics of the scale. If the scale indentation defect rate increases, the second threshold value and the pressure of the high-pressure water descaling are dynamically adjusted.

[0030] The fourth visual detection unit can also detect surface defects of the rolled piece with flange deviation within the qualified range in real time. If there are surface defects, the rolled piece is automatically sprayed with a code for marking and rejection.

[0031] The deep learning-based convex steel rolling shape control system comprises:

[0032] The heating furnace heats the billet to reach the preset rolling temperature range and removes the surface oxide layer by high-pressure water descaling.

[0033] The blooming mill preliminarily rolls the heated billet into an intermediate billet.

[0034] The rough rolling mill rolls the intermediate billet to preliminarily form the basic shape profile of the rolled piece.

[0035] The intermediate rolling mill continues to roll the rolled piece on the basis of rough rolling.

[0036] The finishing mill finally rolls the rolled piece.

[0037] The first visual detection unit identifies the edge position of the billet, calculates the billet center line offset through three-dimensional profile reconstruction, and triggers hydraulic correction when the centering deviation is greater than the first threshold value. The centering data is transmitted to the rough rolling mill in real time as the reference for the rough rolling mill pass positioning.

[0038] The second visual detection unit extracts the residual area of the oxide scale on the surface of the rolled piece in real time through an image segmentation algorithm, and triggers secondary descaling when the detected residual area is greater than a second threshold value; the spread amount is calculated synchronously, and if the actual spread amount deviates from the theoretical value by more than a third threshold value, the side pressure amount of the intermediate rolling mill stand is adjusted;

[0039] The third visual detection unit monitors the temperature distribution of the rolled piece in real time, adjusts the flow distribution of the cooling water of the finishing mill when the detected temperature distribution is uneven, and scans the transition area between the flange and the waist of the rolled piece; if the measured value of the corner radius deviates from the designed value by more than a fourth threshold value, the side pressure amount of the finishing mill stand is corrected;

[0040] The fourth visual detection unit generates a flange cross-sectional profile through three-dimensional point cloud reconstruction, compares the filling rate calculated based on a theoretical model, and calculates the roll gap compensation amount when the filling rate is less than a fifth threshold value, and adjusts the roll gap of the finishing mill in real time; at the same time, the cross-sectional size tolerance of the rolled piece is detected, and if the flange deviation is greater than a sixth threshold value, the rolled piece is sent back to the finishing mill for re-rolling until the rolled piece with a flange deviation within the qualified range is output.

[0041] The beneficial effects of the present application are:

[0042] Improve the quality of the billet processing: the first visual detection unit identifies the edge position of the billet in real time, calculates the center line deviation, and automatically triggers hydraulic correction when the centering deviation is greater than a first threshold value, and transmits the centering data to the rough rolling mill as the hole type alignment reference, greatly improving the accuracy of the billet centering, ensuring the accuracy of the hole type alignment in the subsequent rolling process, and laying a foundation for high-quality rolling.

[0043] Optimize the control of the rough rolling process: the second visual detection unit is arranged at the outlet side of the rough rolling mill, and the image segmentation algorithm is used to extract the residual area of the oxide scale on the surface of the rolled piece in real time, effectively solving the problem of residual oxide scale and ensuring the surface quality of the rolled piece. Real-time calculation of the spread amount realizes accurate control of the spread amount and improves the size accuracy of the rolled piece.

[0044] Precise control of the intermediate rolling stage: the third visual detection unit monitors the temperature distribution of the rolled piece in real time, ensures that the temperature of each part of the rolled piece is uniform, and makes the deformation of the rolled piece more uniform, which is beneficial to improving the shape quality.

[0045] Accurate control of the finishing quality: the flange cross-sectional profile is generated through three-dimensional point cloud reconstruction, and the filling rate is calculated by comparing the theoretical model, to ensure that the shape of the flange of the rolled piece meets the design requirements. At the same time, the rolled piece with a large error is sent back to the finishing mill for re-rolling until the rolled piece with a flange deviation within the qualified range is output, effectively avoiding the production of unqualified products and improving the production efficiency and product quality.

[0046] The method and system are based on deep learning technology, realize real-time monitoring, intelligent analysis and automatic adjustment of the whole process of convex character steel rolling, reduce manual intervention, improve the automation and intelligent level of production, reduce labor intensity, and improve production efficiency and product quality stability. Compared with the original process, the surface defect occurrence rate can be reduced by 40% after using the method of the application. DETAILED DESCRIPTION

[0047] Embodiment 1

[0048] The technical scheme of the application is as follows:

[0049] S1, heating the blank to reach the preset rolling temperature range, and removing the surface oxide layer by high-pressure water descaling.

[0050] The blank is provided with the required heat to reach the preset rolling temperature range, so that the blank has good plastic deformation ability in the subsequent rolling process, reduces the rolling force, and improves the rolling efficiency and product quality. If the temperature is insufficient, the blank is prone to cracking, rolling and other problems; if the temperature is too high, it may cause grain coarsening and serious oxidation.

[0051] High-temperature oxidation of metal blanks in the heating furnace will form a dense oxide layer. If these oxide layers are not removed, they will directly embed into the surface of the subsequent rolling or forging metal, causing products to have defects such as pitting, cracking, and delamination, which seriously affect the mechanical properties.

[0052] High-pressure water descaling uses high-pressure water jet impact to make the oxide layer locally rapidly cool and shrink, crack and warp to peel off. It can efficiently clean the surface of the blank, improve the surface quality of the steel, reduce the wear of the roll, prolong the service life of the roll, reduce the difficulty and acid consumption of pickling, save production cost, and also avoid uneven steel microstructure and performance caused by the oxide layer, and improve the overall quality of the product.

[0053] S2, a first visual detection unit is arranged above the conveying roller way between the outlet of the heating furnace and the blooming mill, the edge position of the blank is identified, the center line offset of the blank is calculated through three-dimensional contour reconstruction, when the centering deviation is greater than the first threshold value, the hydraulic correction is triggered, and the centering data is transmitted to the roughing mill in real time as the reference for the roughing mill pass positioning.

[0054] The first visual detection unit has a machine vision function of deep learning with double viewing angles, can identify the edge position of the blank, and perform three-dimensional contour reconstruction according to the edge position of the blank, calculate the offset of the center line of the blank based on the center line of the roughing mill pass, when the offset Δx is greater than 0.5mm, trigger the hydraulic correction, and calculate the stroke X of the hydraulic correction:

[0055] X = mΔx,

[0056] Wherein, m is the coefficient of compensating mechanical clearance.

[0057] Because of the inevitable mechanical clearance of the rolling line conveying roller, the hydraulic deviation correction pushing amount needs to cover this part of the idle stroke, and the stroke of the hydraulic deviation correction needs to be greater than the deviation amount; and as the temperature of the blank increases, the blank will also expand laterally, so the actual width of the heated blank is greater than the cold width, and therefore the pushing amount needs to offset this expansion amount in the opposite direction.

[0058] This closed-loop control mechanism ensures that the roll and the rolled piece are always accurately matched, reduces the size deviation and surface defects caused by the misalignment of the pass, and lays a quality foundation for the subsequent finishing process.

[0059] Three-dimensional profile reconstruction can also calculate the cross-sectional asymmetry of the rolled piece. If the cross-sectional asymmetry exceeds the sixth threshold value, the rolled piece is sent back to the heating furnace for reheating. This effectively intercepts blanks with serious internal defects, avoids rolling steel and strip breakage accidents caused by cross-sectional asymmetry, and reduces material waste and rework costs caused by product rejection.

[0060] The first visual detection unit identifies the edge position of the blank in real time, calculates the center line deviation, and automatically triggers hydraulic deviation correction when the centering deviation is greater than the first threshold value. The centering data is transmitted to the roughing mill in real time as the pass alignment reference, greatly improving the accuracy of the blank centering, ensuring accurate pass alignment during subsequent rolling, and laying a foundation for high-quality rolling.

[0061] S3, a second visual detection unit is arranged at the outlet side of the roughing mill, and an image segmentation algorithm is used to extract the residual area of the rolled piece surface scale in real time. When the detected residual area is greater than the second threshold value, secondary descaling is triggered. The spread amount is calculated synchronously, and if the measured spread amount deviates from the theoretical value by more than the third threshold value, the intermediate mill stand roll side pressure is adjusted.

[0062] In actual application, the second visual detection unit is located 5m away from the outlet of the roughing mill, avoiding the cooling water splashing area, and has a 45° downward angle with the rolling line of the roughing mill, covering the rolled piece. The second visual detection unit includes a high-speed industrial camera and a laser profile sensor.

[0063] A second visual detection unit is arranged at the outlet side of the roughing mill, and an image segmentation algorithm is used to extract the residual area of the rolled piece surface scale in real time. This effectively solves the problem of residual scale and ensures the surface quality of the rolled piece. The spread amount is calculated in real time to accurately control the spread amount and improve the size accuracy of the rolled piece.

[0064] Specifically, the following operations can be implemented: first, the rolling piece image is denoised by Gaussian filtering, and then the oxide scale area is dynamically segmented based on the Otsu algorithm, the oxide scale coverage rate, that is, the residual area, is calculated by dividing the number of oxide scale pixels by the total area pixel number, and according to the different compositions of the blank, the second threshold value is set between 3%-5%, if greater than the second threshold value, the high-pressure water is triggered to perform secondary descaling, and the nozzle angle of the secondary descaling is 40°±5°.

[0065] Synchronization calculation of the spread amount, the spread amount representing the difference between the actual spread amount and the theoretical spread amount of the rolling piece in the rolling process is a key indicator for measuring the rolling size control accuracy, and the specific calculation method is as follows:

[0066] ΔW=W 实测 —W 理论 ,

[0067] Wherein, ΔW is the deviation between the measured spread amount and the theoretical value, W 实测 is the value obtained by the second visual detection unit by deep learning to measure the actual width of the rough rolling rolling piece, and W 理论 is the width variation of the rolling piece in the ideal rolling state calculated in advance.

[0068] The spread amount of the rough rolling output directly affects the width of the blank at the entrance of the medium rolling, and the medium rolling vertical roll actively controls the subsequent spread amount through the side pressure amount, so that the spread amount is too large, and the medium rolling mill vertical roll side pressure amount is adjusted, and the specific operation is to calculate the vertical roll side pressure adjustment amount:

[0069] ΔS=αΔW+C,

[0070] Wherein, ΔS is the vertical roll side pressure adjustment amount, α is the spread compensation coefficient, and C is the system error compensation term.

[0071]

[0072] Wherein, σ s is the yield strength of the rolling piece material, and β is the equalization coefficient, usually 300.

[0073] Further, due to the excessive rough rolling reduction, the insufficient medium rolling vertical roll side pressure, or the rolling temperature is too low, the roll wear causes the pass to be too wide, which will cause the spread amount to be too small, which is also harmful to the quality of the convex steel, and the medium rolling vertical roll pressure is reduced to eliminate the harm.

[0074] S4, a third visual detection unit is arranged on the outlet side of the medium rolling mill, and the temperature distribution of the rolling piece is monitored in real time, and when the temperature distribution is detected to be uneven, the flow distribution of the cooling water of the finishing mill is adjusted; the transition area between the flange and the waist of the rolling piece is scanned, and if the measured value of the fillet radius deviates from the design value by more than the fourth threshold value, the side pressure amount of the finishing mill vertical roll is corrected.

[0075] The third visual inspection unit is composed of a high-precision infrared camera and a high-speed visible light camera, which work in coordination with a lighting system. The infrared camera has high sensitivity in temperature detection, enabling accurate capture of temperature data on the surface of the rolled piece. The high-speed visible light camera is used to obtain clear appearance images of the rolled piece, assisting in subsequent geometric size analysis. The detection unit is fixed through a stable mechanical structure, ensuring accurate detection position and angle during production, to obtain accurate and reliable data.

[0076] During the continuous passage of the rolled piece through the exit of the rolling mill, the third visual inspection unit continuously and real-time monitors the temperature distribution of the rolled piece, quickly obtains a large number of temperature data points, and integrates and analyzes these data using advanced image processing algorithms. By constructing a temperature distribution thermal map of the surface of the rolled piece, the temperature differences in different regions of the rolled piece are intuitively presented.

[0077] Temperature uniformity is one of the key factors for the stability of the rolling process. If the temperature distribution of the rolled piece is uneven, the metal deformation resistance in different regions will differ, leading to fluctuations in rolling force, which may cause problems such as rolling mill vibration, rolled piece deviation, affecting the continuity and stability of production. Real-time monitoring of temperature distribution can help identify potential problems in a timely manner, providing a basis for subsequent adjustments. Temperature non-uniformity can also cause inconsistent deformation of different parts of the rolled piece, leading to defects such as uneven thickness, waves, and warping in subsequent rolling, seriously affecting the mechanical properties and appearance quality of the product. By monitoring the temperature distribution, measures can be taken in advance to avoid the occurrence of these defects, improving product qualification rate.

[0078] When the third visual inspection unit detects uneven temperature distribution of the rolled piece through data analysis, it will quickly transmit relevant data to the finishing mill to control the cooling water flow of each cooling area of the finishing mill. For areas with higher temperatures, increase the cooling water flow to speed up heat dissipation and reduce the temperature in that area; for areas with lower temperatures, appropriately reduce the cooling water flow to avoid excessive cooling. During the adjustment process, the system will monitor the cooling water flow and temperature changes of the rolled piece in real time, and dynamically adjust the flow distribution based on feedback information to ensure that the temperature of the rolled piece gradually tends to be uniform.

[0079] Specifically, the third visual inspection unit is provided with an infrared camera to detect the temperature difference of the rolled piece. If the temperature difference between the lowest temperature and the highest temperature of the rolled piece is greater than the maximum temperature difference threshold, the flow of cooling water in the high-temperature area corresponding to the highest temperature is increased to Q new :

[0080] Q new = Q0 × [1 + 0.02(ΔT - 50)],

[0081] where Q new is the adjusted cooling water flow, Q0 is the original set cooling water flow, and ΔT is the temperature difference of the rolled piece.

[0082] The rough rolling mill rolling speed is increased to v new :

[0083] v new = v0 x [1 + 0.01 (50 - Delta T)],

[0084] Wherein, v new is the adjusted rolling speed, and v0 is the original set rolling speed.

[0085] The third visual detection unit monitors the temperature distribution of the rolled piece in real time, ensures that the temperature of each part of the rolled piece is uniform, makes the deformation of the rolled piece more uniform, and is beneficial to improving the plate shape quality.

[0086] If the temperature difference of the head and tail, edge and core of the rolled piece is too large, the metal deformation resistance distribution is uneven, the rolling force fluctuation further causes the product thickness and shape deviation, and if the local area is over-cooled, cracks are prone to occur in the local area, and the grain in the overheated area is coarse.

[0087] The traditional solution is to manually adjust the cooling water by experience, the response is lagging, and the high-temperature area cannot be accurately positioned. Therefore, the third visual detection unit is adopted, which includes an infrared camera and can generate a temperature field cloud map in real time to identify the accurate position exceeding the maximum temperature difference threshold.

[0088] The method adopted by the present application can dynamically adjust the cooling water and rolling speed, and the temperature difference is controlled within ±15 DEG C in actual use.

[0089] The high-temperature area increases the cooling water flow, targets to reduce the high-temperature point temperature, avoids local over-burning, increases the rough rolling mill rolling speed, reduces the residence time of the rolled piece in the roll gap in the low-temperature area, and suppresses the rapid temperature drop. The linkage fine rolling pre-compensation transmits the temperature difference data to the fine rolling model in advance, optimizes the edge temperature detected at the exit of the intermediate rolling in the distribution of the side pressure of the vertical roll.

[0090] For example, in production, the third visual detection unit detects that the edge temperature of the rolled piece is 1050 DEG C and the core temperature is 980 DEG C at the exit of the intermediate rolling, and the temperature difference Delta T is 70 DEG C. Then the specific operation of the present application is: the cooling water flow is increased from 100 L / min to 140 L / min, the rolling speed is increased from 2.0 m / s to 2.4 m / s, the fine rolling receives the temperature difference data, and the reduction is adjusted from 12 mm to 10 mm in advance. The ideal result finally obtained is that the temperature difference is reduced to 30 DEG C, the finished product edge has no cracks, and the thickness tolerance is improved from ±0.5 mm to ±0.2 mm.

[0091] Uniform temperature distribution helps the rolled piece to realize uniform deformation during fine rolling, reduces the size deviation and internal stress concentration caused by temperature difference, thereby improving the size precision and mechanical properties of the product, and meeting the high requirements of different users on product quality.

[0092] While monitoring the temperature distribution, the third visual inspection unit uses a high-speed visible light camera to perform high-precision scanning on the transition zone between the flange and the waist of the rolled piece, ensuring that every subtle geometric feature can be captured.

[0093] Using an edge detection algorithm based on deep learning and a three-dimensional reconstruction technique, the profile information of the transition zone of the rolled piece is accurately extracted, and the measured value of the fillet radius is calculated. By comparing with the pre-set design value, the system can accurately determine whether the fillet radius meets the requirements.

[0094] If the measured value of the fillet radius deviates from the design value by more than the fourth threshold value, the third visual inspection unit will immediately send a correction instruction to the finishing mill control system. After receiving the instruction, the finishing mill control system quickly calculates the adjustment amount of the roll side pressure according to the pre-set correction model and algorithm.

[0095] Ensuring product size consistency, timely correction of roll side pressure can quickly correct the deviation of the rolled piece fillet radius, restore the normal geometric shape of the rolled piece in the subsequent rolling process, reduce the rework, scrap and other situations caused by size deviation, improve production efficiency and reduce production cost. At the same time, avoid unnecessary wear of rolling mill and other equipment caused by excessive adjustment of rolling parameters, prolong the service life of the equipment, and reduce the maintenance cost of the equipment.

[0096] S5, a fourth visual inspection unit is arranged at the outlet side of the finishing mill, a flange cross-sectional profile is generated by three-dimensional point cloud reconstruction, and the filling rate is compared with the theoretical model. When the filling rate is less than the fifth threshold value, the roll gap compensation amount is calculated, and the finishing mill pressure is corrected in real time; at the same time, the rolled piece cross-sectional size tolerance is detected, if the flange deviation is greater than the sixth threshold value, the rolled piece is sent back to the finishing mill for re-rolling until the output rolled piece with flange deviation within the qualified range.

[0097] As a comprehensive evaluation index, the flange filling rate can consider the influence of various rolling factors, providing a unified reference standard for process optimization and control, ensuring that high-quality rolled products can be obtained under different rolling processes, so this index is used as the evaluation standard in the final finishing step.

[0098] Specifically, the operation of correcting the finishing mill pressure is to adjust the finishing mill pressure to Ah:

[0099]

[0100] Where Ah is the total pressure after compensation, h0 is the original set pressure, η is the flange filling rate, and K is the material deformation compensation coefficient.

[0101] The larger reduction requires a longer time to transfer the deformation to the core of the rolled piece. Reducing the speed can avoid internal cracks caused by uneven deformation of the surface and the core. Therefore, after adjusting the reduction of the finishing mill, the rolling speed is reduced by 10-15% of the original rolling speed. Reducing the speed can reduce dynamic rolling force fluctuations, prevent motor overload or roll damage, slow down the temperature drop rate of the rolled piece, maintain the final rolling temperature above the phase transition point, improve size accuracy, and reduce scrap rates.

[0102] By generating the flange cross-sectional profile through the above three-dimensional point cloud reconstruction, the filling rate is compared with the theoretical model calculation to ensure that the flange shape of the rolled piece meets the design requirements. At the same time, the rolled piece with excessive error is sent back to the finishing mill for re-rolling until the flange deviation of the output rolled piece is within the qualified range, effectively avoiding the production of unqualified products and improving production efficiency and product quality.

[0103] At the same time, the cross-sectional size tolerance of the rolled piece is detected. If the flange deviation is greater than the sixth threshold value, the rolled piece is sent back to the finishing mill for re-rolling until the flange deviation of the output rolled piece is within the qualified range.

[0104] The flange size deviation directly affects the assembly accuracy and structural strength. Immediate rework only increases 5-10% energy consumption, while scrapping and re-rolling costs 3-5 times higher.

[0105] It is required to have a reverse rolling mode to support the rolled piece return and automatic parameter backtracking.

[0106] The fourth vision detection unit identifies the scale indentation defects on the rolled piece based on the spectral characteristics of the scale. If the scale indentation defect rate rises, the second threshold value and the pressure of the high-pressure water descaling are dynamically adjusted. Scale indentation can cause micro-cracks or reduced coating adhesion during subsequent processing. Early control can improve the yield of finished products. Hard scale accelerates roll wear, reduces its residue, prolongs the roll surface maintenance cycle, and reduces production costs. For example, in the production verification, when the fourth vision detection unit detects scale indentation defects and the scale indentation defect rate rises from 1.9% to 2.1%, the second threshold value is adjusted from 5% to 3%, and the high-pressure water descaling pressure is adjusted from 18 MPa to 22 MPa. The scale indentation defect rate can be reduced from 2.1% to 1.3%.

[0107] After the second threshold value of the secondary descaling is reduced, high-pressure water cleaning can be triggered earlier, reducing the total amount of scale entering the finishing mill and thus reducing the risk of indentation.

[0108] The fourth vision detection unit can also detect surface defects of the rolled piece with flange deviation within the qualified range in real time. If there are surface defects, the rolled piece is automatically marked for rejection.

[0109] Prevent the rolling piece with surface defects from flowing into the subsequent process or market, and ensure the stability of product quality. Automatic code marking and rejection can establish a unique "identity" for the problem rolling piece, which is convenient for subsequent accurate tracing of the cause of defects, such as raw material problems and abnormal rolling parameters, and is beneficial to targeted process improvement. At the same time, rejecting the defective rolling piece can avoid occupying subsequent processing resources, improve the continuity and efficiency of the production process, and reduce the cost of rework and scrap caused by defective products.

[0110] The convex letter steel rolling shape control system based on deep learning comprises:

[0111] A heating furnace heats the billet to a preset rolling temperature range and removes the surface oxide layer by high-pressure water descaling.

[0112] A blooming mill preliminarily rolls the heated billet into an intermediate billet.

[0113] A rough rolling mill rolls the intermediate billet to preliminarily form the basic shape profile of the rolling piece.

[0114] A medium rolling mill continues to roll the rolling piece on the basis of rough rolling.

[0115] A finishing mill performs the final rolling of the rolling piece.

[0116] A first visual detection unit identifies the edge position of the billet, calculates the center line offset of the billet through three-dimensional contour reconstruction, triggers hydraulic correction when the centering deviation is greater than a first threshold value, and transmits the centering data to the rough rolling mill in real time as the reference for the hole pattern alignment of the rough rolling mill.

[0117] A second visual detection unit extracts the residual area of the rolling piece surface scale in real time through an image segmentation algorithm, triggers secondary descaling when the detected residual area is greater than a second threshold value, and simultaneously calculates the spread amount, and adjusts the side pressure amount of the medium rolling mill vertical roll if the measured spread amount deviates from the theoretical value by more than a third threshold value.

[0118] A third visual detection unit monitors the temperature distribution of the rolling piece in real time, adjusts the flow distribution of the cooling water of the finishing mill when the detected temperature distribution is uneven, and scans the flange and waist transition area of the rolling piece, and corrects the side pressure amount of the finishing mill vertical roll if the measured value of the fillet radius deviates from the design value by more than a fourth threshold value.

[0119] A fourth visual detection unit generates a flange cross-sectional profile through three-dimensional point cloud reconstruction, calculates the filling rate by comparing the theoretical model, calculates the roll gap compensation amount when the filling rate is less than a fifth threshold value, and corrects the rolling amount of the finishing mill in real time. At the same time, the cross-sectional size tolerance of the rolling piece is detected, and if the flange deviation is greater than a sixth threshold value, the rolling piece is sent back to the finishing mill for re-rolling until the rolling piece with the flange deviation within the qualified range is output.

[0120] The method and system are based on deep learning technology, realize real-time monitoring, intelligent analysis and automatic adjustment of the whole process of convex character steel rolling, reduce manual intervention, improve the automation and intelligent level of production, reduce labor intensity, and improve production efficiency and product quality stability.

Claims

1. A deep learning-based method for controlling the shape of rolled convex steel sheets, characterized in that, Includes the following steps: S1. Heat the billet to reach the preset rolling temperature range, and use high-pressure water to descale and remove the surface oxide layer. S2. A first vision detection unit is set above the conveying roller table between the heating furnace outlet and the billet mill to identify the edge position of the billet. The offset of the billet centerline is calculated by three-dimensional contour reconstruction. When the centering deviation is greater than the first threshold, hydraulic correction is triggered, and the centering data is transmitted to the roughing mill in real time as a reference for the roughing mill pass alignment. S3. A second vision detection unit is set up on the exit side of the roughing mill. The residual oxide scale area on the surface of the rolled piece is extracted in real time through the image segmentation algorithm. When the residual area is detected to be greater than the second threshold, secondary descaling is triggered. The width spread is calculated simultaneously. If the deviation between the measured width spread and the theoretical value is greater than the third threshold, the pressure on the vertical roll side of the intermediate mill is adjusted. S4. A third vision detection unit is set up on the exit side of the intermediate mill to monitor the temperature distribution of the rolled piece in real time. When uneven temperature distribution is detected, the flow rate distribution of the cooling water in the finishing mill is adjusted. The transition area between the flange and the waist of the rolled piece is scanned. If the measured value of the fillet radius deviates from the design value by more than the fourth threshold, the side pressure of the vertical roll of the finishing mill is corrected. S5. A fourth vision detection unit is set up on the exit side of the finishing mill. The flange cross-sectional profile is generated by reconstructing the three-dimensional point cloud. The filling rate is calculated by comparing it with the theoretical model. When the filling rate is less than the fifth threshold, the roll gap compensation is calculated and the finishing mill reduction is corrected in real time. At the same time, the cross-sectional dimensional tolerance of the rolled piece is detected. If the flange deviation is greater than the sixth threshold, the rolled piece is sent back to the finishing mill for rolling again until the output rolled piece with flange deviation within the qualified range is obtained.

2. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, The specific operation for adjusting the vertical roll side pressure of the rolling mill in S3 is as follows: calculate the vertical roll side pressure adjustment amount: ， in, This is the adjustment amount for the side pressure of the vertical roller. For the width compensation coefficient, To account for the deviation between the measured width and the theoretical value, C This is the system error compensation term; ， in, The yield strength of the rolled material. This is the equilibrium coefficient.

3. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, The third vision inspection unit is also equipped with an infrared camera to detect the temperature difference of the rolled piece. If the temperature difference between the lowest and highest temperatures of the rolled piece exceeds the maximum temperature difference threshold, the flow rate of cooling water is increased in the high-temperature zone corresponding to the highest temperature. Q new : ， in, Q new The adjusted cooling water flow rate, Q 0 represents the original set cooling water flow rate, Δ T This refers to the temperature difference of the rolled parts.

4. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, In S2, the three-dimensional contour reconstruction can also calculate the cross-sectional asymmetry of the rolled piece. If the cross-sectional asymmetry exceeds the sixth threshold, the rolled piece is sent back to the heating furnace for reheating.

5. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, The specific operation for correcting the finishing mill reduction as described in S5 is to adjust the finishing mill reduction to Δ h : ， Where, Δ h The total reduction after compensation. h 0 represents the original set pressure amount. η For flange fill rate, K This is the material deformation compensation coefficient.

6. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, The fourth vision inspection unit can also detect surface defects of rolled parts in real time if the flange deviation is within the acceptable range. If surface defects are found, the rolled parts will be automatically marked and rejected.

7. The deep learning-based method for controlling the shape of rolled convex steel plates according to claim 1, characterized in that, The fourth vision detection unit identifies oxide scale indentation defects on the rolled piece based on the spectral characteristics of the oxide scale. If the oxide scale indentation defect rate increases, the second threshold and the pressure of high-pressure water descaling are dynamically adjusted.

8. A deep learning-based control system for the shape of rolled convex steel sheets, characterized in that, include: The heating furnace heats the billet to the preset rolling temperature range and uses high-pressure water descaling to remove the surface oxide layer. The billet mill is used to initially roll heated billets into intermediate billets. The roughing mill rolls intermediate billets, initially forming the basic shape and outline of the rolled piece. The intermediate rolling mill continues to roll the workpiece based on the roughing mill. The finishing mill performs the final rolling of the workpiece; The first vision detection unit identifies the edge position of the billet, calculates the offset of the billet centerline through three-dimensional contour reconstruction, and triggers hydraulic correction when the centering deviation is greater than the first threshold. The centering data is then transmitted to the roughing mill in real time as a reference for the roughing mill pass alignment. The second vision detection unit extracts the residual oxide scale area on the surface of the rolled piece in real time through image segmentation algorithm. When the detected residual area is greater than the second threshold, it triggers secondary descaling. Simultaneously, it calculates the width spread. If the deviation between the measured width spread and the theoretical value is greater than the third threshold, it adjusts the side pressure of the vertical roll of the intermediate mill. The third vision inspection unit monitors the temperature distribution of the rolled piece in real time. When uneven temperature distribution is detected, it adjusts the flow distribution of the cooling water in the finishing mill. It scans the transition area between the flange and the waist of the rolled piece. If the measured value of the fillet radius deviates from the design value by more than the fourth threshold, it corrects the side pressure of the vertical roll of the finishing mill. The fourth vision inspection unit generates the flange cross-sectional profile through 3D point cloud reconstruction, calculates the filling rate by comparing it with the theoretical model, and calculates the roll gap compensation amount when the filling rate is less than the fifth threshold, and corrects the reduction amount of the finishing mill in real time. At the same time, it detects the cross-sectional dimensional tolerance of the rolled piece. If the flange deviation is greater than the sixth threshold, the rolled piece is sent back to the finishing mill for rolling again until the output rolled piece with flange deviation within the qualified range is obtained.

Images