A rough rolling wide spread scattering learning method and system

Abstract

This invention relates to the field of automatic control technology for steel rolling, and particularly to a scattering learning method and system for width spread in roughing mills. The method involves reading the steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer of the current rolling pass; constructing a nine-square grid region consisting of a center point and eight cells in adjacent rows or columns; obtaining the instantaneous width spread learning value corresponding to the center point; and retrieving the original width spread learning values ​​corresponding to the cells with valid check flags within the nine-square grid from the database; and performing scattering learning update calculations on the center point, adjacent points, and diagonal points respectively to obtain the width spread learning update values ​​for each cell within the nine-square grid. The advantages of this invention are: when the entry thickness layer or width reduction layer shifts, the cells of the newly pointed layer have already received a preliminary update through adjacent scattering learning, avoiding a step change in learning values ​​caused by layer switching and significantly suppressing abnormal fluctuations in the width model prediction values.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for steel rolling, and in particular to a scattering learning method and system for roughing mill width expansion. Background Technology

[0002] In hot-rolled strip steel production, the roughing process typically employs a reciprocating multi-pass rolling method. Roughing width learning is one of the core functions of the secondary width model to achieve high-precision width prediction. Roughing width learning is usually determined by multiple dimensions, including steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer. The model updates the width learning value based on the layer cell pointed to by the current rolling pass.

[0003] In existing technologies, the breadth learning adopts a single-point update mode, that is, the learning value is calculated and written only for the single layer cell pointed to by the current rolling pass. However, in production practice, it has been found that although the steel grade layer, finished product width layer, and roughing mill stand number layer have not changed, the entry thickness layer and width reduction layer may shift due to factors such as fluctuations in the incoming slab size, the target size of the finished product, or the model calculation value, causing the layer cell pointed to by the current pass to change. The layers before and after the change are often in adjacent or close positions, but because the original layer cell has been learned frequently enough, while the new layer cell has not been learned enough, there may be a significant difference in the breadth learning values ​​between the two.

[0004] This inadequate learning, whether or not it occurs within the same rolling cycle, can easily lead to abnormal jumps or fluctuations in the predicted value of the roughing mill exit width model, which in turn causes the finished coil width to exceed tolerances or become blocked, resulting in increased rework costs and spot losses, severely hindering the company's goal of reducing costs and increasing efficiency.

[0005] To address the aforementioned issues, while some studies have attempted to improve learning performance by increasing sample size and optimizing filtering coefficients, structural defects such as imbalanced learning values ​​between adjacent layers and slow convergence speed of new layers remain unresolved. Therefore, providing a broad learning method that can simultaneously consider its own layer's learning and radiate outwards to adjacent layers has become a pressing technical challenge for those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a scattering learning method and system for roughing roll width expansion, to solve the technical problems in the prior art caused by the imbalance of learning values ​​between old and new layers and abnormal fluctuations in model predictions due to the offset of entry thickness layer and width reduction layer. By constructing a nine-square grid area composed of entry thickness layer and width reduction layer with the layer cell pointed to by the current rolling pass as the center point, and using differentiated smoothing coefficients to perform scattering learning update calculations based on the positional attributes of the center point, adjacent points, and diagonal points, synchronous updates of the center point and its surrounding cells are achieved. This allows for a smooth transition of width expansion learning values ​​between adjacent layers, accelerates the convergence of new layer learning, improves the accuracy of width model prediction, and effectively reduces rework costs and spot losses caused by width blockage, providing strong technical support for enterprises to reduce costs and increase efficiency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A scattering learning method for rough rolling width expansion includes: S1. Read the steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer of the current rolling pass; S2. Using the cell pointed to by the entry thickness layer and width reduction layer in the two-dimensional table as the center point, construct a nine-square grid area consisting of the center point and eight cells in the adjacent row or column; S3. Obtain the instantaneous value of the width learning corresponding to the center point, and read the original value of the width learning corresponding to each cell in the nine-square grid whose check flag is valid from the database. S4. Based on the instantaneous value of the width learning corresponding to the center point, the original value of the width learning, and the preset smoothing coefficient, perform scattering learning update calculations on the center point, adjacent points, and diagonal points respectively to obtain the width learning update value of each cell in the nine-square grid. S5. Write the expanded learning update value to the database.

[0008] The check marks for each cell within the 3x3 grid are obtained through an AND logical operation between the position rationality check and the learning rationality check; The location validity check is used to determine whether the corresponding cell is within the valid range of the two-dimensional table; The learning rationality check is used to determine whether the corresponding cell has been learned and updated as the center point in other courses of the current volume.

[0009] The formula for calculating the scattering learning update value of the center point is: zle_new[i][j]=zle_old[i][j]+α*(zle_cur-zle_old[i][j])①; in: zle_cur represents the instantaneous value of the breadth learning, in mm; zle_old[i][j] represents the original value of the wide learning curve, in mm; α represents the center point smoothing coefficient; zle_new[i][j] represents the center point width learning update value, in mm, with row and column coordinates [i][j] of (2,2).

[0010] The formula for calculating the scattering learning update value of adjacent points is: zle_new[i][j]=zle_old[i][j]+α1*α*(zle_cur-zle_old[i][j])②; in: α1 represents the smoothing coefficient between adjacent points; zle_new[i][j] represents the learning update value of the width of the adjacent point, in mm, and the row and column coordinates [i][j] are (2,1), (2,3), (1,2), (3,2).

[0011] The formula for calculating the scattering learning update value of the corner points is: zle_new[i][j]=zle_old[i][j]+α2*α*(zle_cur-zle_old[i][j])③; in: α2 represents the diagonal smoothing coefficient, and α1≥α2; zle_new[i][j] represents the learning update value of the diagonal point width, in mm, with row and column coordinates [i][j] of (1,1), (1,3), (3,1), (3,3).

[0012] Before writing the expanded learning update value to the database, the following steps are also included: performing a limit range check on the expanded learning update value; if it exceeds a preset threshold, then taking the boundary value of the preset threshold.

[0013] The check marks for each cell within the 3x3 grid are determined through an AND logical operation between the positional rationality check and the learning rationality check. Specifically: For any cell within the 3x3 grid, its check flag is set to "valid" if and only if the corresponding cell simultaneously satisfies both the position rationality check result and the learning rationality check result as "yes"; otherwise, its check flag is set to "invalid".

[0014] A scattering learning system for rough rolling mills includes a layer reading module, a nine-square grid construction module, an inspection mark calculation module, a data reading module, a scattering learning calculation module, and a data writing module. The layer reading module is used to read multiple layer information for the current rolling pass; The nine-grid building module is used to construct a nine-grid based on the inlet thickness layer and the width reduction layer; The check mark calculation module is used to determine the validity of each cell in the 3x3 grid based on the location rationality criterion and the learning rationality criterion, and generate the corresponding check mark. The data reading module is used to obtain the instantaneous value of the wide learning and the original value of the wide learning corresponding to the cells in the nine-square grid where the check flag is valid; The scattering learning calculation module is used to perform corresponding wide-range learning update calculations based on the positional attributes of the cell's center point, adjacent points, and diagonal points. The data writing module is used to write the updated wide learning values ​​into the database.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By expanding the traditional single-point learning mode to a nine-grid scattering learning mode, the learning value of the current rolling pass not only updates the center point cell, but also radiates and diffuses to adjacent points and diagonal points; when the entry thickness layer or width reduction layer shifts, the new pointing layer cell has already been initially updated through adjacent scattering learning, avoiding the learning value step caused by layer switching and significantly suppressing abnormal fluctuations in the width model prediction value; 2. This invention enables the learning increment of any rolling pass to be applied to nine cells within a nine-square grid simultaneously, significantly increasing the learning frequency of cells in each layer; for edge layers or low-frequency rolling layers that have not been sufficiently learned before, sample information can be quickly accumulated through nine-square grid scattering learning, shortening the learning convergence cycle and improving the overall response speed of the model. 3. Through the nine-square grid scattering learning mechanism, the spatial distribution of the roughing mill width learning value is smoother and more continuous, eliminating the unreasonable phenomenon of excessive differences in learning values ​​between adjacent layers; the width model predicts the exit width based on more accurate and balanced learning values, the prediction accuracy is substantially improved, and the roughing mill width hit rate is significantly improved. 4. The improved accuracy of the width model prediction directly reduces the proportion of finished rolls blocked due to width deviation, avoiding rework costs and losses from spot sales; at the same time, the abnormal fluctuations of the model caused by insufficient layer learning are greatly reduced, reducing the frequency of manual intervention and achieving a dual improvement in production process stability and economic benefits. 5. This invention does not change the original multi-dimensional layer classification system of roughing rolling width learning. It only performs nine-square grid scattering expansion on the two-dimensional table constructed by the entry thickness layer and the width reduction layer. No additional hardware investment or modification of rolling line equipment is required. The calculation logic of this invention is clear and the amount of computation is small. It can be directly embedded into the existing two-level width model software architecture and has good production line adaptability and promotion value. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the scattering learning of the rough rolling mill.

[0017] Figure 2 This is a schematic diagram of the broad learning levels.

[0018] Figure 3 This is a schematic diagram of the scattering learning method for rough rolling.

[0019] Figure 4 This is a diagram illustrating the original values ​​of the learning data for the relevant layers in the database table.

[0020] Figure 5 This is a diagram illustrating the effect of updating the database table for the original single-point learning technology.

[0021] Figure 6 This is a schematic diagram of the original value of the nine-square grid wide learning system of the present invention.

[0022] Figure 7 This is a schematic diagram of the nine-grid wide-area learning update value of the present invention.

[0023] Figure 8 This is a schematic diagram illustrating the effect of updating the database table in this invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings, but it should be noted that the implementation of the present invention is not limited to the following embodiments.

[0025] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0026] Example 1: The meanings of the relevant terms in this invention are as follows: Steel grade tier: A tier number (family) classified according to steel grade, composition, and process. Finished product width layer: Layer number (wrt_idx) based on the target width of the finished product; Roughing mill stand number / layer: The layer number (std_idx) is based on the roughing mill stand. Entry thickness layer: The layer number (eth_idx) is determined based on the entry thickness of each pass in the roughing mill. Width Reduction Layer: The layer number (dft_idx) is determined based on the width reduction of each pass in the roughing mill.

[0027] The value range for the above-mentioned layers can be found in [link to relevant documentation]. Figure 1 The range of values ​​is affected by the division method and group interval, and may vary from production line to production line; the basis for the group interval division of the above layers is as follows: Figure 2 Due to the influence of product structure, the production lines may also differ.

[0028] A 3x3 grid: a 9-cell area symmetrically expanded outwards from the center point by two rows and two columns, with row and column coordinates defined in the range of (1~3, 1~3). See [link / reference]. Figure 1 .

[0029] Center point: The unique cell pointed to by the steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer, with row and column coordinates (2,2). See [link / reference]. Figure 1 The cell containing the black dot.

[0030] Adjacent points: Within the 3x3 grid, the four adjacent cells in the same row or column as the center point, with row and column coordinates (2,1), (2,3), (1,2), (3,2), see [link to relevant documentation]. Figure 1 The cell indicated by the short black arrow.

[0031] Diagonal points: Within the 3x3 grid, the four diagonal cells that are different from the center point in both row and column, with row and column coordinates of (1,1), (1,3), (3,1), and (3,3). See [link / reference]. Figure 1 The cell indicated by the long black arrow.

[0032] The relevant formulas of this invention are explained as follows: Formula for scattering update value of center point of nine-square grid: zle_new[i][j]=zle_old[i][j]+α*(zle_cur-zle_old[i][j])①; in zle_cur represents the instantaneous value of the breadth learning, in mm; zle_old[i][j] represents the original value of the wide learning curve, in mm; α represents the center point smoothing coefficient; zle_new[i][j] represents the center point width learning update value, in mm, with row and column coordinates [i][j] of (2,2); Formula for updating the scattering learning value of adjacent points in a 3x3 grid: zle_new[i][j]=zle_old[i][j]+α1*α*(zle_cur-zle_old[i][j])②; in: α1 represents the smoothing coefficient between adjacent points; zle_new[i][j] represents the learning update value of the width of the adjacent point, in mm, and the row and column coordinates [i][j] are (2,1), (2,3), (1,2), (3,2); Formula for updating the diagonal scattering values ​​in a 9x9 grid: zle_new[i][j]=zle_old[i][j]+α2*α*(zle_cur-zle_old[i][j])③; in: α2 represents the diagonal smoothing coefficient, and α1≥α2; zle_new[i][j] represents the learning update value of the diagonal point width, in mm, with row and column coordinates [i][j] of (1,1), (1,3), (3,1), (3,3); Formula notes: zle_cur: Instantaneous value of wide-range learning, ranging from -20 to 20 [mm].

[0033] The existing technology is known, and the instantaneous value of the wide-range learning is the difference between the exit width measured by the width gauge and the exit width recalculated by the model.

[0034] i: Thickness layer value at the entrance of the nine-grid entrance, ranging from 1 to 3; j: Value of the 3x3 grid width compression amount, ranging from 1 to 3; zle_old[i][j]: The original value of the 3x3 grid wide learning space, with a value range of -15 to 15[mm]; zle_new[i][j]: The 3x3 grid width learning update value, ranging from -15 to 15[mm]; α: The learning smoothing coefficient of the center point of the nine-square grid is 50%, with a value range of 0~100 [%]; α1: The learning smoothing coefficient for the width expansion of adjacent points in the 3x3 grid is 70%, with a value range of 0~100 [%]; α2: The learning smoothing coefficient for the diagonal points of the nine-square grid is 50%, with a value range of 0~100 [%]; Usually α1 is greater than or equal to α2.

[0035] As is known in existing technology, the roughing zone of a hot rolling production line typically employs a reciprocating multi-pass rolling method, and the roughing width learning is also performed in passes. The roughing width learning layers are usually determined by multiple dimensions, including steel grade layer, finished product width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer.

[0036] This invention is based on a two-dimensional width model of a hot rolling production line. In a two-dimensional table constructed from the entry thickness layer and width reduction layer, the current learning point is used as the center point, radiating outwards to a nine-square grid. Through specific width learning calculations, the process of synchronously calculating and updating the width learning of the center point and its surrounding cells is achieved. This invention is illustrated using a single rolling pass in the roughing mill region as an example; the process for other passes is the same.

[0037] A scattering learning method for rough rolling is described. For specific steps, see [link to documentation]. Figure 3 : S1, Read the layer value; Read the values ​​of steel grade layer, finished product width layer, roughing mill stand layer, entry thickness layer, and width reduction layer.

[0038] S2. Calculate the nine-square grid check marks; The check flags for each point within the 3x3 grid are stored in a three-row, three-column two-dimensional array enable[3][3]. The formula for calculating each check flag is as follows: Location rationality check "and" learning rationality check.

[0039] The check flag representing the center point is set to 1 by default.

[0040] S21. Location rationality check: In the two-dimensional table constructed by the thickness layer and width reduction layer, if the center point is located at the boundary, there will be abnormal points in the nine-square grid that exceed the range of the two-dimensional table. Therefore, it is necessary to determine the reasonableness of the position of each point in the nine-square grid except for the center point. If the position is reasonable, the check result is 1, otherwise it is 0.

[0041] S22. Check the rationality of learning: The learning rationality of each point in the nine-square grid, excluding the center point, is judged. If the point has not been studied as the center point in other courses of the current volume, the result is 1; otherwise, it is 0.

[0042] S3. Read relevant data; Read the instantaneous value of the current point (i.e., the center point). Read the original value of the wide learning corresponding to the point with the nine-square check flag of 1 from the wide learning database table and store it in a three-row, three-column two-dimensional array zle_old[3][3].

[0043] Read the width and smoothness coefficients of the center point, adjacent points, and diagonal points of the 3x3 grid.

[0044] S4. Two-layer loop completes scattering learning; S41: Use a double loop to traverse all points in the 3x3 grid. If all points have been traversed, proceed to step S5; otherwise, proceed to step S42.

[0045] S42: Determine the check flag of the current point. If the check flag is 0, return to step S41 to traverse the next point; otherwise, execute step S43.

[0046] S43: Calculate the position attributes of the current point and store the result in the variable cell_pos (center point = 0, adjacent points = 1, diagonal points = 2). Then execute step S44.

[0047] S44: Determine the current point's position attribute. If it is an adjacent point, perform the adjacent point scattering learning update value calculation and return to step S41; otherwise, proceed to step S45.

[0048] S45: Determine the current point's position attribute. If it is a diagonal point, perform the diagonal point scattering learning update value calculation and return to step S41; otherwise, proceed to step S46.

[0049] S46: Perform the calculation of the center point scattering learning update value, and return to step S41.

[0050] S5, Limit check and processing of nine-square grid learning update value; The learning update values ​​of the nine-square grid are stored in a three-row, three-column two-dimensional array zle_new[3][3]. Limit checks are required for each element of the array. If the value exceeds the preset range, the boundary value is taken.

[0051] S6, Write the 3x3 grid learning update value; Write the width learning update value corresponding to the point marked as 1 in the nine-square grid check to the database table.

[0052] Example 2: In this embodiment, a scattering learning method and system for rough rolling is the same as in Embodiment 1, but with the addition of a dynamic update mechanism for inspection marks.

[0053] A hot-rolled strip steel plant has two stands, R1 and R2, in its roughing section to perform a 3+3 pass combination rolling process on a certain rolled product. The product has a steel grade layer of 5 and a finished width layer of 1. During the fourth pass of roughing, the roughing mill stand layer is 2, the entry thickness layer is 7, and the width reduction layer is 4. The model calculates the instantaneous width spread learning value for the current pass to be 3.0 mm.

[0054] See the original data of the relevant strata learning in the database table. Figure 4 : I. The effects of the original technology; The values ​​for steel grade layer, finished product width layer, roughing mill stand layer, entry thickness layer, and width reduction layer are 5, 1, 2, 7, and 4 respectively, pointing to cells with a value of 1.0 mm. Single-point learning updates are completed through learning calculations. See [link / reference]. Figure 5 .

[0055] zle_new[7][4]=zle_old[7][4]+α*(zle_cur-zle_old[7][4]); zle_new[7][4]=1.00+0.5*(3.0-1.0); zle_new[7][4]=2.00.

[0056] II. Technical Effects of the Invention; 1. Read the layer value; The values ​​for steel grade layer, finished product width layer, roughing mill stand layer, entry thickness layer, and width reduction layer are 5, 1, 2, 7, and 4, respectively.

[0057] 2. Calculate the nine-square grid inspection marks; The check flags for each point within the nine-square grid are stored in a three-by-three two-dimensional array enable[3][3]. The calculation formula for each check flag is: position rationality check "AND" learning rationality check.

[0058] (1) Location rationality check: Centered on this point, the nine-square grid extending outward does not exceed the range of the two-dimensional table constructed by the entrance thickness layer and the width reduction layer. Therefore, the result of the reasonable position check for each point is 1.

[0059] (2) Check the rationality of learning: In this example, apart from the center point, none of the other points in the nine-square grid have been studied as center points in the learning process of other lessons in the current volume. Therefore, the result of the rationality check for each point is 1.

[0060] Therefore, all data in the enable[3][3] check flags of each point in the 3x3 grid are 1.

[0061] 3. Read relevant data; The current width learning instantaneous value is read as 3.0mm.

[0062] Since the check flag for the nine-square grid is all 1, the original values ​​of the expansion learning corresponding to all points of the nine-square grid are read from the expansion learning database table and stored in a three-row, three-column two-dimensional array zle_old[3][3]. The values ​​of each element of the array are shown in the figure. Figure 6 .

[0063] The learning smoothing coefficients for the center point, adjacent points, and diagonal points of the 3x3 grid are 0.5, 0.7, and 0.5, respectively.

[0064] 4. Scattering learning is completed through a two-layer loop; The system iterates through the 3x3 grid using a double loop. Each point is classified as a center point, adjacent points, or diagonal points based on its attribute calculation results, and its update value is calculated for each point.

[0065] Center point: In the above image, the cell with a value of 1.00 has row and column coordinates of (2,2); Adjacent points: In the above image, the value is cell 1.01, with row and column coordinates of (2,1), (2,3), (1,2), (3,2); Diagonal point: In the above image, the value is cell 1.02, with row and column coordinates of (1,1), (1,3), (3,1), (3,3); Calculation of scattering at the center point of a 3x3 grid: zle_new[i][j]=zle_old[i][j]+α*(zle_cur-zle_old[i][j]); zle_new[i][j]=1.0+0.5*(3.0-1.0); zle_new[i][j]=2.0; Nine-square grid adjacent point scattering learning calculation: zle_new[i][j]=zle_old[i][j]+α1*α*(zle_cur-zle_old[i][j]); zle_new[i][j]=1.0+0.7*0.5*(3.0-1.01); zle_new[i][j]≈1.7; Nine-square grid diagonal scattering learning calculation: zle_new[i][j]=zle_old[i][j]+α2*α*(zle_cur-zle_old[i][j]); zle_new[i][j]=1.0+0.5*0.5*(3.0-1.02); zle_new[i][j]≈1.5; After the calculation is completed, update the values ​​of each element in the value array zle_new[3][3], see below. Figure 7 .

[0066] 5. Limit check and processing of the nine-square grid learning update value; The learning update values ​​of the nine-square grid are stored in a three-row, three-column two-dimensional array zle_new[3][3]. Each element of the array needs to be checked for limits. Since the value range is -15~15mm, there are no out-of-limit values ​​in this example.

[0067] 6. Write the 3x3 grid learning update value; Because the nine-square grid check flag is all 1, the entire learning update value array zle_new[3][3] is written into the wide learning database table. The learning value data for the relevant layers in the database table after writing can be found in [link to database table]. Figure 8 .

[0068] III. A wide-ranging scattering learning system for rough rolling; A scattering learning system for rough rolling mills is used to implement the above-mentioned scattering learning method, including a layer reading module, a nine-square grid construction module, an inspection mark calculation module, a data reading module, a scattering learning calculation module, and a data writing module; The layer reading module is used to read the steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer information for the current rolling pass; The nine-grid construction module is connected to the layer reading module and is used to construct a nine-grid area consisting of eight cells of the center point and its adjacent rows or columns, with the cell pointed to by the entry thickness layer and the width pressing amount layer in the two-dimensional table as the center point. The check mark calculation module is connected to the nine-square grid construction module. It is used to determine the validity of each cell in the nine-square grid based on the position rationality criterion and the learning rationality criterion, and generate the corresponding check mark. Among them, the location rationality criterion refers to whether the nine-square grid area constructed with the current cell as the center point is entirely within the effective range of the two-dimensional table formed by the entrance thickness layer and the width reduction layer; the learning rationality criterion refers to whether the current cell has not been learned and updated as the center point in other rolling passes of the current roll. The data reading module is connected to the check mark calculation module to obtain the instantaneous value of the center point's wide learning, and to read the original value of the wide learning corresponding to the cell in the nine-square grid where the check mark is valid from the wide learning database table; The scattering learning calculation module is connected to the data reading module. It is used to perform corresponding wide-range learning update calculations based on the position attributes of the center point, adjacent points, and diagonal points of the cell, respectively, using formulas ①, ②, and ③, to obtain the wide-range learning update value of each cell in the nine-square grid. The data writing module is connected to the scattering learning calculation module and is used to write the updated wide learning values ​​into the wide learning database table.

[0069] IV. System Workflow; When this system is working, the layer reading module first reads the five-dimensional layer information of the current rolling pass; the nine-square grid construction module constructs a nine-square grid area based on the entry thickness layer and the width reduction layer; the inspection mark calculation module judges the position rationality and learning rationality of each cell in the nine-square grid and generates inspection marks; the data reading module obtains the instantaneous width learning value and the original width learning value of the corresponding cell based on the inspection marks; the scattering learning calculation module performs scattering learning update calculations for the center point, adjacent points, and diagonal points according to the cell position attributes; finally, the data writing module writes the updated width learning value into the database, completing the scattering learning process of the current pass.

[0070] The modular design enabled the engineering implementation of the scattering learning method. This system, combined with the scattering learning method, demonstrated its effectiveness in actual production on two roughing mills, R1 and R2: A dynamic update mechanism for inspection markers prevented excessive and repeated updates of cells during the same coil rolling process, improving the stability of the learned values; the use of a nine-grid scattering learning calculation ensured that the instantaneous learning value of 3.0mm for the fourth pass effectively radiated to the surrounding eight cells, with the center point updated to 2.0mm, adjacent points updated to approximately 1.7mm, and diagonal points updated to approximately 1.5mm, achieving a smooth transition of the learned values; the decoupled design between system modules facilitated maintenance and upgrades, and allowed for rapid adaptation to different production line layer division methods and group spacing criteria; limit checks ensured that the learned values ​​remained within a safe range of -15 to 15mm, guaranteeing the stable operation of the model.

[0071] In summary, the technology has successfully achieved a technological leap from "single-point isolated update" to "regional collaborative diffusion" in roughing width learning, solving the long-standing problem of insufficient layer learning in the field of hot rolling width control, and providing strong technical support for enterprises to reduce costs and increase efficiency.

[0072] This invention extends the traditional single-point learning mode to a nine-grid scattering learning mode, enabling the learning value of the current rolling pass to not only update the center cell but also radiate and diffuse to adjacent and diagonal points. When the entry thickness layer or width reduction layer shifts, the cell of the new pointing layer has already been initially updated through adjacent scattering learning, avoiding the learning value jump caused by layer switching and significantly suppressing abnormal fluctuations in the width model prediction value. This invention allows the learning increment of any rolling pass to simultaneously act on nine cells within the nine-grid range, greatly increasing the learning frequency of each layer cell. For previously insufficiently learned edge layers or low-frequency rolling layers, nine-grid scattering learning can quickly accumulate sample information, shorten the learning convergence cycle, and improve the overall response speed of the model. Through the nine-grid scattering learning mechanism, the spatial distribution of roughing width learning values ​​becomes smoother and more continuous, eliminating the impact of adjacent layers. This invention addresses the unreasonable phenomenon of excessively large differences in learning values ​​between different layers. The width model, based on more accurate and balanced learning values, predicts the exit width, resulting in a substantial improvement in prediction accuracy and a significant increase in the roughing width hit rate. This improved prediction accuracy directly reduces the proportion of finished coil blockages caused by width deviations, avoiding rework costs and losses from spot sales. Simultaneously, it significantly reduces abnormal model fluctuations caused by insufficient layer learning, lowering the frequency of manual intervention and achieving a dual improvement in production process stability and economic benefits. This invention does not alter the original multi-dimensional layer division system for roughing width learning; it only performs a nine-square grid scattering expansion on the two-dimensional table constructed from the entry thickness layer and the width reduction layer. No additional hardware investment or modification of rolling mill equipment is required. The invention has clear computational logic, low computational load, and can be directly embedded into existing two-level width model software architectures, possessing good production line adaptability and promotional value.

Claims

1. A scattering learning method for rough rolling mills, characterized in that, include: S1. Read the steel grade layer, finished width layer, roughing mill stand number layer, entry thickness layer, and width reduction layer of the current rolling pass; S2. Using the cell pointed to by the entry thickness layer and width reduction layer in the two-dimensional table as the center point, construct a nine-square grid area consisting of the center point and eight cells in the adjacent row or column; S3. Obtain the instantaneous value of the width learning corresponding to the center point, and read the original value of the width learning corresponding to each cell in the nine-square grid whose check flag is valid from the database. S4. Based on the instantaneous value of the width learning corresponding to the center point, the original value of the width learning, and the preset smoothing coefficient, perform scattering learning update calculations on the center point, adjacent points, and diagonal points respectively to obtain the width learning update value of each cell in the nine-square grid. S5. Write the expanded learning update value to the database.

2. The scattering learning method for rough rolling widening according to claim 1, characterized in that, The check flags for each cell within the nine-square grid are obtained through an AND logical operation between the position rationality check and the learning rationality check; The location validity check is used to determine whether the corresponding cell is within the valid range of the two-dimensional table; The learning rationality check is used to determine whether the corresponding cell has been learned and updated as the center point in other courses of the current volume.

3. The scattering learning method for rough-rolled widening as described in claim 1, characterized in that, The formula for calculating the scattering learning update value of the center point is: zle_new[i][j]=zle_old[i][j]+α*(zle_cur-zle_old[i][j])①; in: zle_cur represents the instantaneous value of the breadth learning, in mm; zle_old[i][j] represents the original value of the wide learning curve, in mm; α represents the center point smoothing coefficient; zle_new[i][j] represents the center point width learning update value, in mm, with row and column coordinates of [i][j](2,2).

4. The scattering learning method for rough-rolled widening as described in claim 1, characterized in that, The formula for calculating the scattering learning update value of the adjacent points is: zle_new[i][j]=zle_old[i][j]+α1*α*(zle_cur-zle_old[i][j])②; in: α1 represents the smoothing coefficient between adjacent points; zle_new[i][j] represents the learning update value of the width of the adjacent point, in mm, and the row and column coordinates [i][j] are (2,1), (2,3), (1,2), (3,2).

5. The scattering learning method for rough-rolled widening as described in claim 1, characterized in that, The formula for calculating the scattering learning update value of the diagonal point is as follows: zle_new[i][j]=zle_old[i][j]+α2*α*(zle_cur-zle_old[i][j])③; in: α2 represents the diagonal smoothing coefficient, and α1≥α2; zle_new[i][j] represents the learning update value of the diagonal point width, in mm, with row and column coordinates [i][j] of (1,1), (1,3), (3,1), (3,3).

6. The scattering learning method for rough-rolled widening as described in claim 1, characterized in that, Before writing the expanded learning update value to the database, the method further includes: performing a limit range check on the expanded learning update value; if it exceeds a preset threshold, then taking the boundary value of the preset threshold.

7. The scattering learning method for rough-rolled widening as described in claim 1, characterized in that, The check flags for each cell within the 3x3 grid are determined through an AND operation between positional rationality checks and learning rationality checks. Specifically: For any cell within the 3x3 grid, its check flag is set to "valid" if and only if the corresponding cell simultaneously satisfies both the position rationality check result and the learning rationality check result as "yes"; otherwise, its check flag is set to "invalid".

8. A scattering learning system for implementing the method of any one of claims 1-7 in roughing mill expansion, characterized in that, It includes a layer reading module, a nine-square grid construction module, a check mark calculation module, a data reading module, a scattering learning calculation module, and a data writing module; The layer reading module is used to read multiple layer information for the current rolling pass; The nine-grid building module is used to construct a nine-grid based on the inlet thickness layer and the width reduction layer; The check mark calculation module is used to determine the validity of each cell in the 3x3 grid based on the location rationality criterion and the learning rationality criterion, and generate the corresponding check mark. The data reading module is used to obtain the instantaneous value of the wide learning and the original value of the wide learning corresponding to the cells in the nine-square grid where the check flag is valid; The scattering learning calculation module is used to perform corresponding wide-range learning update calculations based on the positional attributes of the cell's center point, adjacent points, and diagonal points. The data writing module is used to write the updated wide learning values ​​into the database.

Images