Stellite alloy rolling parameter adaptive optimization method based on deep learning

Abstract

The invention discloses a stellite alloy rolling parameter self-adaptive optimization method based on deep learning, and belongs to the technical field of automatic plate shape control of metal plate rolling. The method comprises the following steps: firstly, converting an acquired thickness difference distribution curve into a two-dimensional grayscale image, and extracting features and quantifying taper deviation trend intensity by using a convolutional neural network; when the strength exceeds the threshold value, the deformation resistance fluctuation type is identified through a classification model, and a nonlinear coupling influence coefficient is calculated. And then, based on the coefficient, fusing data of a lateral force sensor, determining optimized rolling reduction distribution along the plate width direction, and solving a real-time inclination angle correction value by adopting an optimized search algorithm in combination with a historical inclination angle sequence. And finally, predicting the deviation risk of the next pass according to the index of the current pass, and implementing preventive parameter fine adjustment. According to the method, real-time sensing, coupling relation decoupling and prospective control of the transverse thickness unevenness are achieved, and the thickness uniformity of the plate is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of automatic shape control technology for metal sheet rolling, and specifically to a deep learning-based adaptive optimization method for Stellite alloy rolling parameters. Background Technology

[0002] Stellite alloys, as high-performance corrosion-resistant and wear-resistant materials, play a crucial role in aerospace, chemical equipment, and marine engineering. The quality of its sheet metal products directly affects the reliability and service life of high-end equipment. Sheet metal rolling is one of the most critical processes in Stellite alloy processing. The precise matching of rolling parameters determines the flatness, thickness uniformity, and shape accuracy of the final product. Among these parameters, sheet metal taper control is a crucial indicator for measuring rolling quality.

[0003] In current industrial practice, Stellite alloy rolling largely relies on operator experience and fixed parameter tables to set key variables such as roll tilt angle, reduction, and lateral force. This approach struggles to achieve dynamic coordination between parameters when dealing with characteristics such as large fluctuations in the material's resistance to high-temperature deformation and strong microstructural inhomogeneity. Especially when the plate width is large or the length is long, the thickness difference between the inlet and outlet gradually accumulates during rolling, leading to a significant tapered deviation along the width direction—a phenomenon where one side is thicker than the other. Once this deviation forms, it is difficult to completely eliminate in subsequent straightening and finishing processes, ultimately resulting in material waste and performance instability.

[0004] Stellite alloys exhibit strong strain rate and temperature sensitivity during high-temperature rolling. Even small changes in reduction or adjustments to the roll tilt angle can significantly alter the flow pattern of the metal along the width direction.

[0005] Meanwhile, the magnitude and distribution of lateral forces further affect the actual deformation state and contact pressure distribution of the rolls, and there is a highly coupled nonlinear relationship between the two. This complex interaction makes it difficult for traditional methods to accurately predict and correct the evolution trend of the sheet taper in a timely manner, often resulting in lag or over-adjustment of the control, causing the taper to fluctuate repeatedly in continuous multi-pass rolling, and even suddenly deteriorating to an unacceptable degree in some passes.

[0006] Therefore, how to sense minute changes in the taper of the plate in real time during the rolling process, and dynamically adjust multi-dimensional parameters such as roll tilt angle, reduction distribution, and lateral force according to the actual deformation behavior of the current pass, so as to achieve stable control of the taper of the entire plate at the micrometer level, has become a key problem that Stellite alloy high-precision plate rolling urgently needs to overcome. Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive optimization method for Stellite alloy rolling parameters based on deep learning, which solves the technical problem that the nonlinear fluctuation of material resistance and the complex coupling effect of rolling parameters during the rolling process of Stellite alloy plates make it impossible to accurately sense and dynamically compensate for the tapered deviation (uniform thickness in the transverse direction) along the width direction in real time.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A deep learning-based adaptive optimization method for Stellite alloy rolling parameters, the method comprising:

[0010] Step 1: Collect the inlet and outlet thickness sequences of the plate using a sensor array, calculate the thickness reduction for each transverse sampling position, and obtain the thickness difference distribution curve along the width direction;

[0011] Step 2: Convert the thickness difference distribution curve into a two-dimensional grayscale image, process the two-dimensional grayscale image using convolution operation, obtain the channel response representing the lateral non-uniform deformation gradient in the multi-layer feature map, and determine the intensity of the taper deviation trend along the width direction.

[0012] Step 3: If the intensity of the taper deviation trend exceeds the preset threshold, extract the stress distribution features and central symmetry deviation features of the edge region from the multi-layer feature map, classify the deformation resistance fluctuation type of the current pass through the classification model, and calculate the nonlinear coupling influence coefficient based on the classification results (such as category probability).

[0013] Step 4: The real-time readings of the lateral force sensor at the corresponding lateral position are weighted and fused using the nonlinear coupling influence coefficient to determine the adjustment requirements of the reduction amount along the width direction of the plate, and the optimized reduction amount distribution target value is obtained.

[0014] Step 5: Based on the optimized reduction distribution target value and the historical sequence of the current roll tilt angle, an optimization search algorithm is used to iteratively determine the real-time tilt angle correction value that makes the transverse thickness profile more uniform.

[0015] Step 6: Update the control system command through the real-time tilt angle correction value, collect the feedback sequence of the adjusted lateral force sensor, calculate the full width thickness fluctuation range of the plate, and obtain the taper stability index of the current pass.

[0016] Step 7: Based on the taper stability index of the current pass and the width variation curve data of the previous passes, use a regression model to predict the probability of deviation risk along the width direction of the next pass, and determine the fine adjustment range of the preventive pressing amount and the tilt angle.

[0017] Step 8: Correct the input parameters of the deformation sensitivity model by adjusting the preventive reduction amount and the tilt angle, calculate the final transverse thickness profile and geometric deviation field of the entire length plate, determine whether the shape accuracy meets the uniformity requirements, and obtain the final rolling quality evaluation conclusion.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This invention uses a sensor array to collect the thickness sequence of the plate at the inlet and outlet in real time, calculates the thickness reduction distribution along the width direction and converts it into a two-dimensional grayscale image. Convolution operations are used to extract the channel response representing the transverse uneven deformation gradient in the multi-layer feature map, accurately quantifying the intensity of the taper deviation trend. When the intensity exceeds a threshold, the edge stress distribution and central symmetry deviation features are further extracted. A classification model is used to identify the nonlinear coupling fluctuation type of the deformation resistance in the current pass and generate an influence coefficient. Lateral force sensor readings are weighted and fused to determine the optimization adjustment target for the reduction along the plate width. Combined with the historical sequence of roll tilt angles, iterative optimization searches obtain real-time tilt angle correction values, achieving closed-loop control. After adjustment, the full-width thickness fluctuation and taper stability are evaluated, and the deviation risk for the next pass is predicted based on a regression model, allowing for preventative fine-tuning. Finally, a deformation sensitivity model is used to calculate the transverse thickness profile and geometric deviation field of the entire plate length, outputting a shape accuracy uniformity assessment conclusion.

[0020] This invention effectively integrates multi-source sensor data with deep learning feature analysis, enabling dynamic perception, real-time identification, precise intervention, and proactive prevention of transverse thickness non-uniformity during the rolling process, significantly improving the consistency of the thickness across the entire width of the plate and the accuracy of geometric shape control. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of the method described in this invention.

[0023] Figure 2 This is a schematic diagram of the image processing and feature quantization process of the present invention.

[0024] Figure 3 This is a flowchart illustrating the forward-looking risk prediction and prevention process of this invention. Detailed Implementation

[0025] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Example 1: See Figures 1-3 This embodiment discloses an adaptive optimization method for Stellite alloy rolling parameters based on deep learning, which may specifically include:

[0028] Step 1: Collect the inlet and outlet thickness sequences of the plate material using a sensor array, calculate the thickness reduction for each transverse sampling position, and obtain the thickness difference distribution curve along the width direction.

[0029] Data is collected at the inlet and outlet of the sheet metal using a sensor array to obtain thickness sequence data at corresponding locations, yielding initial thickness measurement results. Based on the collected thickness sequence data, each lateral sampling location is processed to calculate the thickness difference between the inlet and outlet, determining the thickness reduction amount. Using the calculated thickness reduction amount data, combined with the lateral sampling location information, thickness difference distribution data along the width direction is generated. This distribution data is smoothed to eliminate noise interference during the acquisition process, resulting in a smoothed difference distribution. A distribution curve along the width direction is constructed based on the smoothed difference distribution, providing a visualized thickness change trend. If certain locations on the thickness difference distribution curve exceed a preset threshold range, the data at the corresponding lateral sampling locations are marked to determine the specific distribution of abnormal areas. Based on the marked abnormal area data, combined with the thickness reduction amount data, a targeted analysis file is generated to determine the detailed characteristic information of the abnormal locations.

[0030] In practical implementation, a laser thickness sensor array is used to... The thickness sequence is continuously collected at the transverse spacing at the plate entry point, and the sampling frequency is set to [value missing]. , obtain the first entrance side Thickness sequence at each sampling location Simultaneously, data were collected at the corresponding location on the exit side using the same frequency and spacing. For each horizontal position The sliding window averaging algorithm is used to analyze the most recent The thickness data within the inlet is filtered to calculate the average thickness at the inlet. ,in The number of valid sampling points within the window ( (number), average thickness of export The same calculation applies;

[0031] Then calculate the thickness reduction at that location. For example, when a certain position , hour, ; All horizontal positions Arranged in order of width to form a thickness reduction sequence , The total number of horizontal sampling points is, for example, 201 points;

[0032] A cubic spline interpolation algorithm is applied to the sequence to generate a continuous thickness difference distribution curve, with the interpolation point spacing refined to [value missing]. To improve resolution; then the standard deviation of the curve is calculated. and maximum and minimum difference For example, to obtain , Further spectral analysis of the curve was performed, and the main spatial frequency components were extracted using Fast Fourier Transform (FFT). It was found that the dominant wavelength was approximately... The periodic fluctuations correspond to the influence of roll eccentricity or thermal crown; finally, based on... The threshold is used to determine whether a plate shape warning is triggered, and then... If the difference exceeds 1.5 times the historical average, it is automatically recorded as an abnormal event and associated with the current roll number and rolling parameters, realizing a fully automated process for monitoring and identifying thickness difference distribution.

[0033] Step 2: Convert the thickness difference distribution curve into a two-dimensional grayscale image, process the two-dimensional grayscale image using convolution operation, obtain the channel response representing the lateral non-uniform deformation gradient in the multi-layer feature map, and determine the intensity of the taper deviation trend along the width direction.

[0034] A two-dimensional grayscale image is constructed using a thickness difference distribution curve. The two-dimensional grayscale image is processed using convolution operations to obtain a multi-layer feature map. A specific channel representing the lateral uneven deformation gradient is selected from the multi-layer feature map. The response value sequence of the specific channel is obtained. Based on the change of the response value sequence along the width direction, the gradient magnitude at each position is calculated. If the gradient magnitude exceeds a preset threshold, the corresponding position is marked as having a taper deviation. The strength of the taper deviation trend is determined by the correspondence between the marked positions and the response value sequence.

[0035] In practice, the thickness difference distribution curves are stacked along the rolling direction in time series to construct a two-dimensional grayscale image, where the horizontal axis corresponds to the width of the plate (e.g., 201 sampling points), the vertical axis represents time (or the continuous sampling time within a pass), and the grayscale value (0-255) normalized corresponds to the thickness reduction amount (e.g., 0.0mm to 2.0mm).

[0036] Specifically, thickness data is collected at a fixed frequency (e.g., 500Hz), and at fixed time intervals (e.g., 0.5 seconds), the thickness data of all horizontal positions within the current time and past time window are combined to form a single image frame.

[0037] For example, if each frame of an image contains 1200 time sampling points vertically, and the sampling frequency is 500Hz, then the actual time window length corresponding to a single frame of an image is... The system can capture a 2.4-second data window every 0.5 seconds (i.e., every 250 sampling points) to generate a new image frame, thus achieving image sliding updates. A 3×3 Sobel operator is applied to this grayscale image for convolution to extract the horizontal and vertical gradients. After obtaining the gradient magnitude map, a 5×5 Gaussian smoothing filter is applied to reduce noise. This data is then input into a pre-trained lightweight convolutional neural network. The network is trained by collecting a large number of "thickness difference distribution curves-grayscale images" from historical rolling processes as input, and using manually labeled or high-precision offline shape gauge measurements of the corresponding "taper deviation level" (e.g., no taper, slight left taper, severe right taper, etc.) as supervision labels. The network structure contains four convolutional blocks, each using a 3×3 convolutional kernel, a stride of 1, and 'same' padding, followed by batch normalization and ReLU activation. The first layer outputs 32 channels, the second 64 channels, the third 128 channels, and the fourth 256 channels. Channels specifically responding to lateral gradient changes are separated from the feature maps of the third and fourth layers. The average activation intensity of each channel along the width direction is calculated using global average pooling. Vectors are constructed based on the response values ​​of these channels, and after applying softmax normalization to the vectors, a weighted sum is calculated to obtain the taper deviation trend intensity index.

[0038] For example, when the activation of channel 17 in the third layer and channel 41 in the fourth layer is significantly higher than that of other channels, the calculated trend intensity value is 0.724, indicating a clear unilateral taper shape with the right side being thicker and the left side being thinner. At the same time, compared with the historical average trend intensity of 0.312 for 50 rolls, if the current value exceeds 2.1 times the average, it is judged as a serious lateral uneven deformation, and the rolling speed, roll gap value and bending roll force parameters at the corresponding moment of the image are automatically extracted and associated for storage, thereby realizing fully automatic quantification of taper deviation and abnormal trend identification based on image features.

[0039] Step 3: If the intensity of the taper deviation trend exceeds a preset threshold, extract the stress distribution features and central symmetry deviation features of the edge region from the multi-layer feature map, classify the deformation resistance fluctuation type of the current pass through the classification model, and calculate the nonlinear coupling influence coefficient based on the classification results (such as category probability).

[0040] If the taper deviation trend intensity exceeds a preset threshold, edge stress distribution data and central symmetry deviation data are extracted from the multi-layer feature map. A classification model is used to analyze the deformation resistance fluctuation type, and the classification result of the fluctuation type is obtained. Based on the fluctuation type classification result, the nonlinear coupling influence coefficient related to deformation resistance is obtained, and the priority of the main influencing factors is determined based on the coefficient value. For the main influencing factors, the corresponding edge stress distribution subset is extracted from the multi-layer feature map, and the data within the subset is processed into layers using statistical tools to obtain the layered stress distribution intervals. If there are abnormally high value areas in the layered stress distribution intervals, the central symmetry deviation data of the area is extracted a second time. By comparing the change amplitude of the deviation data, it is determined whether there is a local symmetry imbalance. Based on the judgment result of the local symmetry imbalance, the deformation resistance fluctuation type subclass corresponding to the imbalance area is obtained. By feature mapping of the subclass data, the refined classification label of the fluctuation type is determined. For the refined classification label, the resistance fluctuation records related to taper deviation are extracted from historical data, and the current data is matched with the historical records using a comparison tool to obtain the reference fluctuation pattern with the highest matching degree. If the matching degree of the reference fluctuation pattern is higher than the preset threshold, the current deformation resistance fluctuation type is marked according to the pattern, and a targeted fluctuation type distribution map is generated based on the marking results.

[0041] If the taper deviation trend intensity exceeds a preset threshold (e.g., 0.6), then the top 20% of the channels with the highest activation values ​​are selected from the 256 channels of the fourth-layer feature map. The mean activation values ​​of the pixel bands occupying 15% of the total width on the left and right sides of the edge region and the middle 70% of the pixel band in the center region are extracted. Simultaneously, the absolute deviation ratio of the difference in the mean activation values ​​of the left and right edge regions to the mean activation value of the center region is calculated to obtain the edge asymmetry coefficient. For example, the mean activation value of the left edge is 142.6, the right edge is 87.3, and the center is 105.8. The edge asymmetry coefficient is (142.6-87.3) / 105.8=0.522. Based on these features, a three-dimensional vector containing edge stress gradient difference, center symmetry deviation, and left and right edge activation ratio is constructed and input into a classifier consisting of two fully connected layers. The first layer of the classifier has 64 nodes and uses tanh activation, while the second layer has 5 nodes and uses softmax output, corresponding to five types of deformation resistance fluctuations: periodic oscillation type, sudden step type, slow drift type, local impact type, and composite coupling type.

[0042] The classifier was trained under supervised instruction using 1200 historically labeled lessons. After optimization with cross-entropy loss, the accuracy on the validation set reached 89.6%. When the predicted category is a composite coupling type, the nonlinear coupling influence coefficient... Calculated using the following formula: .in, The output probability for the dominant category (composite coupled type) is 0.783. The aforementioned edge asymmetry coefficient is 0.522. and The weighting coefficients, determined based on regression from historical data, are 0.6 and 0.4 in this example, respectively, and are calculated as follows: This coefficient indicates that the deformation resistance of the current pass is strongly coupled with the transverse temperature gradient and the inlet thickness fluctuation.

[0043] Subsequently, this coefficient was fitted with the real-time acquired inlet plate shape index of 0.021, the measured outlet crown of -18.4μm, and the cumulative reduction rate of intermediate passes of 42.7% using multiple linear regression. The regression coefficients showed that every 0.1 increase in the coupling effect coefficient would cause the outlet crown to shift negatively by about 3.2μm. This automatically adjusted the setting value of the bending roller force for the next pass by 12.6kN to compensate for the intensified center thinning trend caused by nonlinear coupling, thus realizing online classification and adaptive correction of deformation resistance fluctuations.

[0044] Step 4: The real-time readings of the lateral force sensor at the corresponding lateral position are weighted and fused using the nonlinear coupling influence coefficient to determine the adjustment requirements of the reduction amount along the width direction of the plate, and the optimized reduction amount distribution target value is obtained.

[0045] The real-time readings of the lateral force sensor at each lateral position are weighted using the nonlinear coupling influence coefficient to obtain the target distribution of the reduction amount adjustment along the plate width direction. The reduction amount deviation value at each lateral position is calculated based on the target distribution, resulting in a sequence of deviation values ​​for each position. If any deviation value in the sequence exceeds a preset range, the reading gradient features and reading non-uniformity features between lateral positions are extracted from the lateral force sensor readings. A classification model is used to process the reading gradient features and reading non-uniformity features to obtain the current reduction amount deviation pattern category label. Similar deviation sequences are retrieved from historical reading records based on the deviation pattern category label to obtain the most closely matched similar deviation sequence. The difference vector between the current deviation value sequence and the similar deviation sequence is calculated using a sequence comparison method, resulting in a set of difference vectors. The local abrupt change locations of the lateral force readings are determined based on the set of difference vectors, resulting in a set of local abrupt change locations. If multiple locations are consecutively arranged in the set of local abrupt change locations, the reading fluctuation amplitude features and fluctuation period features are extracted for the consecutive abrupt change segments. By comparing the thresholds of the fluctuation amplitude features and fluctuation period features, it is determined whether the current consecutive abrupt change segment belongs to the periodic misalignment type. If the judgment result is a periodic imbalance type, then the distribution of the reduction amount adjustment target is periodically compensated and corrected according to the fluctuation period characteristics to obtain the corrected reduction amount distribution value.

[0046] When weighting and fusing the real-time readings of the lateral force sensors at the corresponding lateral positions using the aforementioned nonlinear coupling influence coefficient, the coefficient 0.758 is first multiplied point-by-point with the instantaneous reading sequence of the 11 lateral force sensors along the plate width direction to obtain the weighted force vector. For example, the original readings of the first sensor on the left are 428.7 kN, the second is 396.2 kN, the sixth in the center is 512.4 kN, and the eleventh on the right is 341.9 kN. After weighting, the first sensor on the left becomes 325.0 kN, the sixth in the center becomes 388.4 kN, and the eleventh on the right becomes 259.2 kN. Then, the quadratic polynomial fitting curve of this vector along the plate width is calculated. The fitted coefficients a, b, and c are -0.041, 3.27, and -18.6, respectively, indicating that the force distribution exhibits a clear parabolic downward opening shape. Then, based on the fitted curve... Compared with the ideal uniformly distributed force value Deviation along the width direction of the plate The full-width comprehensive adjustment coefficient is obtained by integration. for:

[0047] ;

[0048] in, : Comprehensive adjustment coefficient, a dimensionless coefficient characterizing the overall degree of unevenness;

[0049] : Board width;

[0050] The coordinates along the width of the plate, with values ​​ranging from... ;

[0051] Ideal uniformly distributed force value, here it is 512.0 kN;

[0052] : The quadratic polynomial fitting curve function of the lateral force along the width of the plate;

[0053] Subsequently, this comprehensive adjustment coefficient With a pre-defined distribution pattern along the width of the plate based on the shape of a quadratic curve Multiplying these together yields the adjustment ratio of the pressure at each lateral position. ,Right now The distribution mode function is mentioned above. A normalized quadratic function: , For the width of the board, the coefficient is... and Pre-set according to plate width and alloy grade to ensure The integral over the plate width range is 1, thus the overall adjustment amount is... Distribute them proportionally to different horizontal positions; : Each horizontal position Adjust the amount of pressure applied at the point.

[0054] Based on this ratio, it is calculated that the left edge area needs to increase the pressure by about 0.9%, the center area needs to decrease the pressure by about 1.4%, and the right edge area needs to increase the pressure by about 0.7%. Finally, the above relative adjustment ratios are multiplied by the current pass setting of 3.25mm to obtain the optimized pressure distribution target value vector. For example, the left side zone 1 is adjusted to 3.279mm, the center zone 6 is adjusted to 3.204mm, and the right side zone 11 is adjusted to 3.273mm, forming a non-uniform pressure target distribution along the plate width. This is then executed in real time by the hydraulic pressure system to achieve precise compensation for plate shape deviation caused by nonlinear coupling.

[0055] Step 5: Based on the optimized reduction distribution target value and the historical sequence of the current tilt angle of the roll, an optimization search algorithm is used to iteratively determine the real-time tilt angle correction value that makes the transverse thickness profile more uniform.

[0056] The optimized reduction distribution target value and the historical sequence of the current roll tilt angle are obtained. A particle swarm optimization algorithm is used to iteratively search the real-time tilt angle correction value to obtain candidate tilt angle correction values ​​that minimize the sum of squares of the transverse thickness profile deviations. The corresponding simulated thickness profile data is calculated using the candidate tilt angle correction values. The difference between the thickness value at each transverse position and the target thickness value is extracted from the simulated thickness profile data to obtain a thickness deviation sequence. The difference in thickness deviation between adjacent positions is calculated for the thickness deviation sequence to obtain a deviation difference sequence. If there are elements in the deviation difference sequence with an absolute value greater than a preset difference threshold, the corresponding position is marked as a potential mutation point, resulting in a potential mutation point set. Based on the position index in the potential mutation point set, the tilt angle values ​​at the same transverse position in the most recent time period are extracted from the historical roll tilt angle sequence to obtain a tilt angle subsequence. The angle change rate sequence is obtained by calculating the angle change amount at adjacent sampling times and dividing it by the time interval using the tilt angle subsequence. Based on the numerical relationship between the angle change rate sequence and the deviation difference sequence at the corresponding transverse positions, the compensation direction and compensation magnitude at each potential mutation point are determined, resulting in a compensation amount sequence. The compensation sequence is superimposed on the candidate tilt angle correction value to obtain the final real-time tilt angle correction value.

[0057] Specifically, based on the optimized reduction distribution target value and the historical sequence of the current roll tilt angle, the target value vector is first read, for example, 3.28mm for zone 1 on the left, 3.19mm for zone 5 in the center, and 3.27mm for zone 10 on the right. Simultaneously, historical tilt angle data for the past 10 rolling cycles is acquired, assuming the most recent value is 0.35 degrees, the previous value is 0.32 degrees, and the historical average is 0.34 degrees. Next, a particle swarm optimization algorithm is used as the optimization search tool, with a particle swarm size of 50 and 100 iterations. Using the transverse thickness profile uniformity as the objective function, the root mean square value of the thickness deviation in each iteration is calculated, with an initial deviation of 0.12mm. Through iterative updates using inertia weights, individual learning factors, and social learning factors, the minimum deviation of 0.03mm is obtained at the 78th iteration, corresponding to an initial tilt angle correction value of 0.41 degrees. To further optimize and consider the dynamic process, the system calculates a dynamic compensation amount (e.g., 0.01 degrees) based on the current rolling speed (2.5 m / s) and the response characteristics of the roll tilting mechanism. This compensation is then superimposed on the initial correction value to obtain a final real-time tilt angle correction of 0.42 degrees. To ensure logical rigor, the system also uses the current roll temperature distribution data as an auxiliary input. Assuming the temperature on the left is 85.3 degrees Celsius and on the right is 87.1 degrees Celsius, the potential impact of the temperature difference on thickness uniformity is analyzed, resulting in a temperature compensation factor of 1.002. The correction value is then fine-tuned to 0.422 degrees, forming a complete data processing chain to ensure that the adjusted angle value can adapt to the dynamic rolling environment.

[0058] Step 6: Update the control system command through the real-time tilt angle correction value, collect the feedback sequence of the adjusted lateral force sensor, calculate the full width thickness fluctuation range of the plate, and obtain the taper stability index of the current pass.

[0059] Real-time data from the lateral force sensor is acquired to form an adjusted lateral force sequence. The maximum and minimum values ​​across the full width range are extracted from this sequence, and their difference yields the thickness fluctuation amplitude. This amplitude is then compared to historical multi-pass fluctuation amplitude sequences to obtain the current pass's taper stability index. If the current pass's taper stability index is higher than the median value of the historical sequence, the original genetic algorithm optimization step size is maintained, resulting in a standard optimization interval. If the current pass's taper stability index is lower than the median value of the historical sequence, the standard optimization interval is compressed based on the thickness fluctuation amplitude, resulting in a contracted optimization interval. The genetic algorithm iteratively calculates candidate correction values ​​within the contracted optimization interval to obtain a real-time tilt angle correction value that further reduces the thickness fluctuation amplitude. This real-time tilt angle correction value drives the control system to perform angle adjustments, and a new round of lateral force sensor feedback sequences is acquired.

[0060] In practice, the control system command is first automatically updated based on the calculated real-time tilt angle correction value of 0.38 degrees. The correction value is then transmitted to the mill execution unit through the built-in communication module to complete the issuance of the angle adjustment command. The command update cycle is set to once per second to ensure real-time performance. At the same time, the issuance timestamp and command value are recorded to form a traceable data log.

[0061] Next, the adjusted feedback data sequence was collected by lateral force sensor arrays deployed on both sides of the rolling mill. Assuming a collection period of 0.5 seconds, 20 data points were continuously acquired. The average lateral force on the left side was 15.6 kN, and on the right side it was 16.2 kN. The force difference between the two sides was calculated to be 0.6 kN. The force difference fluctuation trend was analyzed in combination with historical data. It was found that the average force difference over the last 10 cycles was 0.5 kN, and the standard deviation was 0.1 kN. It was determined that the current adjusted force distribution was in a stable range.

[0062] Using a full-width thickness detection device, thickness data is collected. Assuming the width of the board is divided into 8 detection areas, the thickness of the first area on the left is 2.85mm, the fourth area in the middle is 2.88mm, and the eighth area on the right is 2.83mm. The full-width thickness fluctuation range is calculated to be 0.05mm. By comparing it with the target fluctuation range of 0.03mm, the current thickness control accuracy is evaluated, and the fluctuation range data is stored in the database for subsequent analysis.

[0063] Based on thickness fluctuation amplitude and lateral force difference data, a weighted average algorithm is used to calculate the taper stability index of the current pass. :

[0064] ;

[0065] in, This represents the current thickness fluctuation range. This represents the largest historical variation in thickness. Given the current lateral force difference, This represents the largest lateral force difference in history. , .

[0066] Substituting into the formula, we can calculate the result. .

[0067] Based on the historical index sequence (the average of the last 5 times is 0.78), the analysis shows that the current taper stability index (0.68) is lower than the historical average, indicating that the control effect needs to be further optimized. The system automatically feeds this index back to the control model as a reference input for parameter optimization in the next cycle, forming a closed-loop adjustment logic to ensure continuous optimization of the rolling process.

[0068] Step 7: Based on the taper stability index of the current pass and the width variation curve data of the previous passes, use a regression model to predict the probability of deviation risk along the width direction of the next pass, and determine the fine adjustment range of the preventive pressing amount and the tilt angle.

[0069] Obtain the data sequence of the width variation curves from the previous passes. Extract the positional deviation sequence in the width direction for each pass from the width variation curve data sequence. Fit the positional deviation sequence in the width direction using a regression model to obtain regression fitting parameters. Calculate the predicted deviation trend value for each position in the width direction for the next pass using the regression fitting parameters. Compare the predicted deviation trend value with the allowable range, and calculate the percentage of positions exceeding the allowable range to obtain the deviation risk probability value. If the deviation risk probability value exceeds a preset threshold, increase the preventive reduction adjustment range while decreasing the tilt angle fine-tuning range, resulting in the first set of adjustment range combinations. If the deviation risk probability value is lower than the preset threshold, decrease the preventive reduction adjustment range while increasing the tilt angle fine-tuning range, resulting in the second set of adjustment range combinations. Based on the comparison between the deviation risk probability value and the taper stability index of the current pass, select one from the first or second set of adjustment range combinations as the final adjustment range. Drive the control system to execute the pre-adjustment actions of the preventive reduction and tilt angle for the next pass using the final adjustment range.

[0070] In practice, the taper stability index of the current pass (0.82) and the width variation curves of the previous four passes are first obtained, with widths of 1250mm, 1242mm, 1238mm, and 1235mm respectively. The decreasing trend of the width is fitted using the least squares method, and the linear regression equation is approximately... ,in The sequence number is the number of the path (starting from 1). This corresponds to the board width (mm).

[0071] Calculate the next step ( The predicted plate width is Subsequently, the deviation sequence along the width direction of each pass was extracted. The maximum edge deviations for the first four passes were 0.07mm, 0.09mm, 0.11mm, and 0.14mm, respectively. A regression model (e.g., support vector regression) was used, with the plate width change rate, pass reduction, and stability index as input features. The kernel function was RBF, and the penalty parameter C was set to 15. After training, the probability of the next pass having a deviation greater than 0.12mm along the width direction was predicted to be 0.67. The system then uses this probability... Based on the current rolling force setting of 3200kN and the target thickness of 2.80mm, a preventative control strategy is employed, and the preventative reduction fine-tuning value is calculated using the following formula. With tilt angle fine-tuning value : ,in This is the empirical gain coefficient (0.4 in this example). The reduction is set to 3.25 mm for this pass, and the calculation is as follows: . ,in This is an empirical coefficient (0.5 in this example). The average of the absolute values ​​of the tilt angles of the most recent passes (0.34 degrees in this example) is calculated as follows: The negative sign indicates a slight adjustment in the negative direction to balance differences in metal flow across the width.

[0072] After the two fine-tuning parameters are verified by fuzzy logic, the control command for the next pass is automatically generated. The pressure is adjusted to +0.64% of the original set value, the tilt angle is corrected to -0.22 degrees, and the simulation results of the predicted risk probability of 0.67 and the expected deviation after adjustment reduced to 0.09mm are stored in the optimization database together as data support for the self-learning of the subsequent pass model, forming a continuous closed-loop mechanism of risk prediction and parameter feedforward.

[0073] Step 8: Correct the input parameters of the deformation sensitivity model by adjusting the preventive reduction amount and the tilt angle, calculate the final transverse thickness profile and geometric deviation field of the entire length plate, determine whether the shape accuracy meets the uniformity requirements, and obtain the final rolling quality evaluation conclusion.

[0074] The process begins by acquiring data on the preventative reduction amount and tilt angle fine-tuning, extracting key adjustment parameters, and combining this with initial data across the entire length of the plate to generate a preliminary set of input parameters. This sets the basis for subsequent input parameter corrections. Based on this correction, the initial settings of the deformation sensitivity model are updated, and the corrected parameters are used to simulate the deformation response across the entire length of the plate, yielding preliminary distribution data for the transverse thickness profile. Based on this preliminary distribution data, the geometric deviation field distribution across the entire length of the plate is calculated, and the offsets at key locations within the deviation field are extracted to determine the characteristic values ​​of the geometric deviation field distribution. If the characteristic values ​​exceed a preset uniformity standard, the local parameters of the deformation sensitivity model are adjusted a second time to obtain adjusted transverse thickness profile data. Using this adjusted data, the shape accuracy distribution across the entire length of the plate is recalculated to determine if the shape accuracy meets the uniformity standard, thus establishing the final shape accuracy assessment result. Finally, based on the shape accuracy assessment result and the preset rules for rolling quality evaluation, a comprehensive analysis of the processing quality across the entire length of the plate is conducted, leading to the final rolling quality assessment conclusion.

[0075] The preventative reduction increment of 0.18 mm and the tilt angle adjustment value of -0.22 degrees, corrected from the previous step, are read and substituted into the input parameter layer of the deformation sensitivity model. Combined with the current rolling temperature of 1080℃, the material's thermal expansion coefficient of 1.18×10⁻⁵ / ℃, and the friction coefficient of 0.28, the finite element deformation sensitivity calculation module is activated. The model employs the three-dimensional elastoplastic finite element method, using hexahedral reduced integral elements to mesh the entire length of the plate, resulting in a total of 168,000 elements. A penalty function contact algorithm is used for the contact between the roll and the plate. Subsequently, the calculated transverse thickness profile data is extracted. The plate's center thickness is 2.802 mm, and the thicknesses at the two side edges are 2.765 mm, 2.771 mm, 2.778 mm, and 2.784 mm, respectively. The maximum transverse thickness difference along the entire length is calculated to be 0.037 mm. Simultaneously, a geometric deviation field distribution is generated, with a peak edge waviness of 0.021 mm and a center indentation depth of 0.014 mm. The system then applies the shape accuracy uniformity evaluation criteria, using the threshold conditions of a lateral thickness standard deviation not exceeding 0.025 mm and an edge waviness less than 0.030 mm for automatic judgment. The lateral thickness standard deviation is then calculated. for:

[0076] ;

[0077] in, For the first The thickness value at each sampling point for The average thickness of each sampling point The total number of transverse thickness sampling points (e.g.) ).

[0078] Calculated standard deviation of transverse thickness Less than the preset threshold And the peak value of edge wavyity Less than its preset threshold All samples met the uniformity requirements. The final rolling quality assessment concluded that the shape accuracy was acceptable, the geometric deviation was within a controllable range, and no secondary intervention was required.

[0079] All calculation results, along with input correction parameters, temperature field and deviation field cloud map data, are written into the quality traceability database for online parameter updates and accuracy adaptive optimization of subsequent batch deformation sensitivity models, forming a closed-loop quality feedforward control link.

[0080] Traditional methods directly use thickness difference sequences for control, failing to fully explore the mechanistic information inherent in the spatial distribution patterns. This invention is based on the fact that the essence of taper deviation is uneven lateral metal flow, which manifests as a specific spatial gradient pattern in thickness distribution. By reconstructing the one-dimensional thickness difference distribution curve into a two-dimensional spatiotemporal grayscale image and applying a convolutional neural network for feature extraction, the abstract physical phenomenon of "lateral uneven deformation gradient" is quantified and visualized. This enables the system to go beyond simple threshold alarms, accurately identifying the intensity and trend of the deviation (such as unilateral taper, mid-wave) and its dynamic evolution, providing high-information-density input for subsequent precise intervention.

[0081] In this invention, a strong nonlinear, time-varying coupling relationship exists between lateral force, reduction, tilt angle, and final thickness distribution during the rolling process. This is a bottleneck that traditional model control struggles to overcome. This invention does not rely on explicit physical models that are difficult to establish precisely. Instead, it uses a deep learning classification model to directly identify "deformation resistance fluctuation types" (such as periodic oscillations and localized impacts) from high-dimensional feature maps. This classification result is used to calculate the "nonlinear coupling influence coefficient," which is essentially a data-driven coupling compensation factor for the specific fluctuation condition. This enables the system to dynamically correct the mapping relationship between lateral force readings and reduction adjustment requirements, achieving adaptive matching between the control strategy and the real-time deformation behavior of the material.

[0082] Traditional closed-loop control corrects for deviations that have already occurred, resulting in a lag. This invention integrates real-time optimization search (iterative optimization based on the current target) with multi-track regression prediction (risk prediction based on historical sequences) to construct a two-layer control architecture of "real-time correction + proactive prevention." Its first principle is that the formation of taper is cumulative and transferable. It not only calculates the optimal tilt angle correction value to homogenize the current profile, but also predicts the probability of deviation risk in the next track through a regression model, and accordingly implements preventative adjustments to the reduction amount and tilt angle. This breaks the passive mode of "detecting deviations - readjusting" and shifts to proactive control of "predicting trends - compensating first," effectively suppressing the accumulation of deviations across tracks.

[0083] This invention achieves refined perception of the taper deviation mechanism by imagerizing the thickness distribution and extracting depth features; it achieves adaptive decoupling of control decisions and complex operating conditions by calculating nonlinear coupling coefficients based on wave type identification; and it realizes a control paradigm shift from passive correction to active leveling by integrating real-time optimization and multi-step prediction. Ultimately, at the mechanistic level, it enhances the system's ability to cope with nonlinear and time-varying uncertainties in Stellite alloy rolling, thereby significantly improving the uniformity of plate thickness and shape control accuracy across the entire width at the engineering level.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based adaptive optimization method for Stellite alloy rolling parameters, characterized in that, The method includes: Step 1: Collect the inlet and outlet thickness sequences of the plate using a sensor array, calculate the thickness reduction for each transverse sampling position, and obtain the thickness difference distribution curve along the width direction; Step 2: Convert the thickness difference distribution curve into a two-dimensional grayscale image, process the two-dimensional grayscale image using convolution operation, obtain the channel response representing the lateral non-uniform deformation gradient in the multi-layer feature map, and determine the intensity of the taper deviation trend along the width direction. Step 3: If the intensity of the taper deviation trend exceeds the preset threshold, extract the stress distribution features and central symmetry deviation features of the edge region from the multi-layer feature map, classify the deformation resistance fluctuation type of the current pass through the classification model, and calculate the nonlinear coupling influence coefficient based on the classification results. Step 4: The real-time readings of the lateral force sensor at the corresponding lateral position are weighted and fused using the nonlinear coupling influence coefficient to determine the adjustment requirements of the reduction amount along the width direction of the plate, and the optimized reduction amount distribution target value is obtained. Step 5: Based on the optimized reduction distribution target value and the historical sequence of the current roll tilt angle, an optimization search algorithm is used to iteratively determine the real-time tilt angle correction value that makes the transverse thickness profile more uniform. Step 6: Update the control system command through the real-time tilt angle correction value, collect the feedback sequence of the adjusted lateral force sensor, calculate the full width thickness fluctuation range of the plate, and obtain the taper stability index of the current pass. Step 7: Based on the taper stability index of the current pass and the width variation curve data of the previous passes, use a regression model to predict the probability of deviation risk along the width direction of the next pass, and determine the fine adjustment range of the preventive pressing amount and the tilt angle. Step 8: Correct the input parameters of the deformation sensitivity model by adjusting the preventive reduction amount and the tilt angle, calculate the final transverse thickness profile and geometric deviation field of the entire length plate, determine whether the shape accuracy meets the uniformity requirements, and obtain the final rolling quality evaluation conclusion.

2. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 1 includes: Data is collected at the inlet and outlet of the plate using a sensor array to obtain thickness sequence data at the corresponding locations, thus obtaining the initial thickness measurement results. Based on the collected thickness sequence data, each lateral sampling position is processed to calculate the thickness difference between the inlet and outlet, and the thickness reduction amount is determined. Using the calculated thickness reduction data and the lateral sampling location information, thickness difference distribution data along the width direction is generated; By smoothing the thickness difference distribution data, noise interference during the acquisition process is eliminated, and the smoothed difference distribution result is obtained. Based on the smoothed differential distribution results, a distribution curve along the width direction is constructed to obtain a visualized trend of thickness variation; If certain locations on the thickness difference distribution curve exceed the preset threshold range, the data at the corresponding lateral sampling locations are marked to determine the specific distribution of the abnormal area. Based on the marked abnormal area data and combined with the thickness reduction data, a targeted analysis file is generated to determine the detailed characteristic information of the abnormal location.

3. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 2 includes: A two-dimensional grayscale image is constructed using the thickness difference distribution curve; The two-dimensional grayscale image is processed by convolution to obtain a multi-layer feature map. Select specific channels representing the gradient of lateral non-uniform deformation from the multi-layer feature map; Obtain the response value sequence of the specific channel; Calculate the gradient magnitude at each location based on the change of the response numerical sequence along the width direction; If the gradient magnitude exceeds the preset threshold, then mark the corresponding position as having a taper deviation; The intensity of the taper deviation trend is determined by the correspondence between the marked positions and the response numerical sequence.

4. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 3 includes: If the taper deviation trend intensity exceeds the preset threshold, edge stress distribution data and center symmetry deviation data are extracted from the multi-layer feature map, and a classification model is used to analyze the deformation resistance fluctuation type to obtain the classification result of the fluctuation type. Based on the classification results of wave types, obtain the nonlinear coupling influence coefficients related to deformation resistance, and determine the priority of the main influencing factors based on the coefficient values; For the main influencing factors, the corresponding edge stress distribution subsets are extracted from the multi-layer feature map, and statistical tools are used to perform layered processing on the data within the subsets to obtain the layered stress distribution intervals. If there are abnormally high value areas in the stress distribution range after stratification, the central symmetry deviation data of the area is extracted a second time. By comparing the change range of the deviation data, it is determined whether there is a local symmetry imbalance. Based on the judgment results of local symmetry imbalance, obtain the deformation resistance fluctuation type subclass corresponding to the imbalance area, and determine the refined classification label of the fluctuation type by feature mapping of the subclass data; For the refined classification labels, resistance fluctuation records related to taper deviation are extracted from historical data. A comparison tool is used to match the current data with the historical records to obtain the reference fluctuation pattern with the highest matching degree. If the matching degree of the reference fluctuation pattern is higher than the preset threshold, the current deformation resistance fluctuation type is marked according to the pattern, and a targeted fluctuation type distribution map is generated based on the marking results.

5. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 4 includes: The real-time readings of the lateral force sensor at each lateral position are weighted by the nonlinear coupling influence coefficient to obtain the target distribution of the pressure adjustment along the width of the plate. Based on the adjustment of the target distribution of the reduction amount, the reduction amount deviation value at each lateral position is calculated to obtain the deviation value sequence at each position; If there are deviation values ​​in the deviation value sequence that exceed the preset range, then extract the reading gradient characteristics and reading non-uniformity characteristics between the lateral positions from the lateral force sensor readings. A classification model is used to process the reading gradient features and reading non-uniformity features to obtain the current compression deviation pattern category label; Based on the deviation pattern category label, similar deviation sequences are retrieved from historical reading records to obtain the most matching similar deviation sequence; The difference vector between the current deviation value sequence and similar deviation sequences is calculated using a sequence alignment method, resulting in a set of difference vectors. The local abrupt change locations of the lateral force readings are determined based on the set of difference vectors, thus obtaining the set of local abrupt change locations; If there are multiple locations arranged consecutively in the set of local mutation locations, then extract the reading fluctuation amplitude characteristics and fluctuation period characteristics for the consecutive mutation segments. By comparing the thresholds of fluctuation amplitude characteristics and fluctuation period characteristics, it can be determined whether the current continuous abrupt change segment belongs to the periodic disorder type. If the judgment result is a periodic imbalance type, then the distribution of the reduction amount adjustment target is periodically compensated and corrected according to the fluctuation period characteristics to obtain the corrected reduction amount distribution value.

6. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 5 includes: Obtain the optimized target value of the reduction distribution and the historical sequence of the current tilt angle of the roll; The particle swarm optimization algorithm is used to iteratively search for the real-time tilt angle correction value to obtain the candidate tilt angle correction value that minimizes the sum of squares of the transverse thickness profile deviation. The simulated thickness profile data is obtained by calculating the corresponding simulated thickness profile data using candidate tilt angle correction values; The difference between the thickness value at each lateral position and the target thickness value is extracted from the simulated thickness profile data to obtain the thickness deviation sequence; The difference in thickness deviation between adjacent positions is calculated for the thickness deviation sequence to obtain the deviation difference sequence; If there are elements in the deviation difference sequence whose absolute value is greater than the preset difference threshold, the corresponding position is marked as a potential mutation point, and a set of potential mutation points is obtained. Based on the position index in the potential mutation point set, the tilt angle values ​​of the same lateral position in the most recent time period are extracted from the historical sequence of roll tilt angles to obtain the tilt angle subsequence. The angle change rate sequence is obtained by dividing the angle change at adjacent sampling times by the time interval, calculated using the tilt angle subsequence. Based on the numerical relationship between the angle change rate sequence and the deviation difference sequence at the corresponding lateral positions, the compensation direction and compensation magnitude at each potential mutation point are determined, and the compensation amount sequence is obtained. The compensation sequence is superimposed on the candidate tilt angle correction value to obtain the final real-time tilt angle correction value.

7. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 6 includes: Acquire real-time data from the lateral force sensor to form an adjusted lateral force sequence; The maximum and minimum values ​​of the full width range are extracted from the adjusted lateral force sequence, and the thickness fluctuation range is obtained by subtracting the two values. The thickness fluctuation amplitude is numerically compared with the historical multi-pass fluctuation amplitude sequence to obtain the taper stability index of the current pass. If the current trace taper stability index is higher than the median value of the historical sequence, then the original genetic algorithm optimization step size is maintained to obtain the standard optimization interval; If the taper stability index of the current pass is lower than the median value of the historical sequence, the standard optimization interval is compressed according to the magnitude of the thickness fluctuation to obtain the contracted optimization interval. A genetic algorithm is used to iteratively calculate candidate correction values ​​within the shrinkage optimization interval to obtain real-time tilt angle correction values ​​that further reduce the thickness fluctuation amplitude. The control system is driven to adjust the angle by real-time tilt angle correction value, and a new round of lateral force sensor feedback sequence is collected.

8. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 7 includes: Obtain the data sequence of the plate width variation curves from the previous few passes; Extract the sequence of positional deviations in the width direction for each pass from the data sequence of the plate width variation curve; A regression model was used to fit the sequence of positional deviations in the width direction, and the regression fitting parameters were obtained. The trend value of the prediction deviation at each position in the width direction of the next pass is calculated by regression fitting parameters; By comparing the predicted deviation trend value with the allowable range, the percentage of locations exceeding the allowable range is statistically analyzed to obtain the deviation risk probability value; If the probability value of deviation risk exceeds the preset threshold, the adjustment range of the preventive reduction amount is increased, while the adjustment range of the tilt angle is decreased to obtain the first set of adjustment range combinations. If the deviation risk probability value is lower than the preset threshold, the adjustment range of the preventive reduction amount is reduced, while the adjustment range of the tilt angle is increased to obtain the second set of adjustment range combinations. Based on the comparison between the deviation risk probability value and the current track taper stability index, select one from the first set of adjustment range combinations or the second set of adjustment range combinations as the final adjustment range. The final adjustment amplitude drives the control system to execute the next preventative pressing and tilt angle pre-adjustment action.

9. The adaptive optimization method for Stellite alloy rolling parameters based on deep learning according to claim 1, characterized in that, Step 8 includes: Acquire the amplitude data of preventive compression and tilt angle fine-tuning, extract key adjustment parameters from them, combine them with the initial data over the entire length of the plate, generate a preliminary set of input parameters, and determine the basis for input parameter correction for subsequent processing; By correcting the basic parameters and updating the initial settings of the deformation sensitivity model, the deformation response of the plate over its entire length is simulated using the corrected parameters, and preliminary distribution data of the transverse thickness profile are obtained. Based on the preliminary distribution data of the transverse thickness profile, the geometric deviation field distribution of the plate is calculated over the entire length, the offset of each key position in the deviation field is extracted, and the characteristic value of the geometric deviation field distribution is determined. If the characteristic value of the geometric deviation field distribution exceeds the preset uniformity standard value, the local parameters of the deformation sensitivity model are adjusted a second time to obtain the adjusted transverse thickness profile data. By using the adjusted transverse thickness profile data, the shape accuracy distribution over the entire length of the plate is recalculated to determine whether the shape accuracy meets the uniformity standard value, and the final shape accuracy judgment result is determined. Based on the shape accuracy judgment results and combined with the preset rules for rolling quality assessment, a comprehensive analysis of the processing quality across the entire length of the plate is conducted to arrive at the final rolling quality assessment conclusion.

Images