Method for Identifying Causes of Head Wave Shape Defects in Hot-Rolled Strip Steel by Improving Partial Least Squares

Abstract

The present invention provides a method for identifying the causes of head waviness defects in hot-rolled strip steel by improving partial least squares, belonging to the field of hot-rolling automation technology. This method diagnoses and evaluates the process parameters related to the head waviness defects of hot-rolled strip steel through an improved partial least squares algorithm, including the following steps: 1) Obtain a data set and perform preprocessing on the sample data to obtain a sample set after dimensionality reduction without data missing and without data anomalies; 2) Use the improved partial least squares algorithm to establish a feasibility test model, and calculate the squared prediction error statistic and control limit of the samples; 3) Compare the contribution values of each process parameter of the defect samples to the squared prediction error statistic, give the risk coefficient of each process parameter, and combine the risk coefficient to give the dominant cause of the defect. The present invention is applicable to the analysis of defect samples, can quickly and automatically analyze the risk coefficients of the process parameters related to the head waviness defects of hot-rolled strip steel, and helps to improve the shape quality and analysis efficiency of hot-rolled strip steel.

Description

Technical Field

[0001] The present invention relates to the technical field of hot rolling automation, and particularly to a method for identifying the causes of head waviness defects in hot rolled strip steel by improving partial least squares. Background Art

[0002] Hot rolling is an important part of the steelmaking process. As one of the most important quality indicators of the product, the shape quality of hot rolled strip steel has always been a hot issue studied by scholars at home and abroad.

[0003] Currently, waviness defects often occur at the head of hot rolled thin strip steel, seriously affecting the quality of the strip steel. At present, the cause analysis and diagnosis of head waviness defects mostly rely on expert experience or adopt the method of analyzing the numerical distribution of a single process parameter. The defect analysis method relying on expert experience is time-consuming and laborious, with relatively low analysis efficiency and difficult to achieve automation. For the method of analyzing the numerical distribution of a single process parameter, the influence degree of process parameters on quality parameters is not considered, and there is a situation where process parameters with a large deviation have a small influence on quality parameters. Therefore, the credibility of the diagnosis and analysis results is relatively low, and the risk levels of various process parameters cannot be quantified.

[0004] Currently, most of the defect analysis and diagnosis of hot rolled strip steel shape focus on expert experience. The risk assessment of the head shape defects of hot rolled strip steel is mostly through mechanism analysis and expert experience. For the research results driven by big data, the work in the data preprocessing stage is not sufficient, and the subsequent risk quantification is not well processed.

[0005] Under the background of big data, data-driven methods have developed rapidly. This method does not take rolling theory as the main basis and can effectively use the production data of strip steel rolling for modeling analysis, quickly diagnosing the causes of head shape defects of strip steel and helping on-site personnel discover production problems early.

[0006] Currently, the commonly used defect cause analysis methods based on multivariate statistical analysis mainly include: partial least squares method, Fisher discriminant method, opposite element analysis method, and principal component analysis method. The partial least squares method is a supervised learning method that incorporates the ideas of principal component analysis and canonical correlation analysis. Compared with the principal component analysis method, the partial least squares method pays more attention to the correlation between process parameters and quality parameters and has better explanatory ability for the dependent variable. Therefore, the partial least squares method is selected as the modeling method of the present invention.

[0007] However, since the internal model of the partial least squares algorithm is a linear regression model, it has good applicability to linear problems. For non-linear problems, the partial least squares algorithm often cannot extract effective latent variables, resulting in low accuracy of the output results when applied to non-linear problems. Therefore, it needs to be improved.

[0008] To this end, the present invention proposes a method for identifying the cause of the wavy defect on the head of the hot-rolled strip based on improved partial least squares. Under the premise of sufficient data preprocessing, the relevant process parameters of the wavy defect on the head of the hot-rolled strip are analyzed and converted into risk coefficients. By analyzing the risk coefficients of each process parameter, the causes of the defects are found, and online automatic analysis of defective samples is realized. The risk coefficients of each process parameter of the defective samples and the possible causes of the defects are output, which can effectively replace manual work, improve analysis efficiency, and save analysis costs. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide an improved partial least squares method for identifying the cause of the wavy defect on the head of the hot-rolled strip, so as to solve the problem that the on-site identification of the cause of the wavy defect on the head of the hot-rolled strip is too dependent on expert experience and is time-consuming and labor-intensive, so as to make risk assessment more convenient and concise.

[0010] The method comprises the following steps:

[0011] S1: The temperature, roll gap, rolling force and other relevant process parameter data of excellent samples with qualified head wave shape and defective samples with unqualified head wave shape and the flatness hit rate of strip head are obtained through the on-site data acquisition system as quality data, and the obtained excellent samples and defective samples are preprocessed by variance test, correlation test, local abnormal factor test and other data to eliminate outlier samples and reduce dimension;

[0012] S2: Substitute the sample data set containing excellent samples and defective samples into the partial least squares algorithm (PLS algorithm) improved by Mahalanobis distance transformation, calculate the sample square prediction error statistic, i.e., the sample SPE statistic, and the sample square prediction error statistic control limit, i.e., the sample SPE statistic control limit, and adjust the value of the latent variable k according to the degree of separation between the defective sample and the excellent sample SPE statistic and the elbow rule. The defective sample SPE statistic exceeds the sample SPE statistic control limit and the relative maximum separation between the defective sample SPE statistic and the excellent sample SPE statistic is used as the applicability test standard between the algorithm and the sample set. If there is a situation where the defective sample cannot be detected or the excellent sample is detected as a defective sample, it means that the process parameter selection is unreasonable, and it is necessary to re-select parameters, collect data and preprocess data;

[0013] S3: Contribution value of each process parameter to the SPE statistic of defective samples obtained in S2 Calculate and express it in the form of a histogram as the initial risk coefficient of each process parameter;

[0014] S4: Reprocess the initial risk coefficients of each process parameter obtained in S3, unify the risk coefficient evaluation index, and sum up the contribution values of each process parameter to the SPE statistic of the defect samples to calculate the proportion of the contribution value of the process parameter, and use this as the final risk coefficient of each process parameter;

[0015] S5: Compare the magnitudes of the risk coefficients of each process parameter. The process parameter with the largest risk coefficient is the dominant defect factor for the occurrence of defects.

[0016] Among them, the specific elimination of outlier samples in S1 is as follows: Retrieve missing values for excellent samples, eliminate the excellent samples with missing values, and then use the local outlier factor algorithm to detect outliers in the samples after eliminating missing values. By calculating the local outlier factor LOF z (p) and comparing it with 1, the closer it is to 1, the closer the density of the sample and its neighboring samples, and it is very likely to belong to the same cluster. When LOF z (p) is greater than 1, it indicates that the density of the sample is less than the density of its neighboring samples, and the sample is more likely to be an outlier sample point and is eliminated;

[0017]

[0018] Among them, N z (p) represents the z-distance neighborhood of the sample point p, that is, all the sample points within the z-distance of the sample point p. |N z (p)| represents the number of all sample points within the z-distance neighborhood of the sample point p. lrd z (p) represents the reciprocal of the average reachable distance from the sample points within the z-neighborhood of the sample point p to the sample point p. lrd z (o) represents the reciprocal of the average reachable distance from the points within the z-neighborhood of the sample point o to the sample point o. The closer the local outlier factor is to 1, the closer the density of the sample point p and its neighboring sample points, and it is very likely to belong to the same cluster. When the local outlier factor value is greater than 1, it indicates that the density of the sample point p is less than the density of its neighboring sample points, and p is more likely to be an outlier sample point.

[0019] S2 is specifically as follows:

[0020] Transform the sample data processed in S1 using the Mahalanobis distance transformation formula. The Mahalanobis distance D(X) is calculated as follows:

[0021]

[0022] Among them, X is the sample vector, s is the covariance matrix of the samples, and u represents the sample mean; the Mahalanobis distance transformation is a non-linear amplification transformation that can extract more representative latent variables and enhance the applicability of the conventional partial least squares algorithm to non-linear problems.

[0023] Perform partial least squares diagnostic analysis on the data samples after the Mahalanobis distance transformation, and decompose them using the idea of recombining characteristic variables as follows:

[0024] u = u y +e = PR T u + e

[0025] e = (I m*m -PR T )u

[0026] Among them, u ∈ R m*l is the data sample set, m is the number of samples, l is the number of process parameters, u y is the fitting sample, e is the corresponding residual matrix, P is the load matrix, and R T is the transpose matrix of the load matrix, and I m*m is the identity matrix;

[0027] Then calculate the SPE statistic for each sample, and the formula is as follows:

[0028] SPE = ||e|| 2 = ||((I m*m -PR T )u)|| 2

[0029] The control limit SPE of the SPE statistic k is calculated as follows:

[0030]

[0031] g = s SPE / 2μ

[0032] Among them, s SPE is the variance of the sample SPE statistic, μ is the mean of the sample SPE statistic, is the chi-square value, obtained by looking up the table according to the confidence level and degrees of freedom, and g is an intermediate variable.

[0033] Calculate the measure S of the number of latent variables k:

[0034]

[0035]

[0036] S = S error,k -Syes,k

[0037] Among them, S yes,k is the variance of the excellent sample SPE statistic when the number of latent variables is k, is the mean of the SPE statistic of the excellent sample when the number of latent variables is k, SPE j,k is the SPE statistic of the jth sample when the number of latent variables is k, and n is the number of excellent samples. error,k is the difference between the mean SPE statistic of defective samples and the mean SPE statistic of excellent samples when the number of latent variables is k, SPE error,k is the SPE statistic of defect samples when the number of latent variables is k.

[0038] Considering that too large a value of the number of latent variables k will cause overfitting, and too small a value will cause insufficient precision, this method selects the number of latent variables in the range of 2-1 / 2 (l is the number of process parameters). After defining the value range of the number of latent variables k, the elbow rule is applied to the S curve of the measurement index to select the number of latent variables k.

[0039] S3 The calculation method is as follows:

[0040]

[0041]

[0042] in, is the contribution of each variable to the sample SPE statistic, is the i-th row of the unit matrix, and l is the number of process parameters.

[0043] In S4, the ratio of the process parameter contribution value to the total process parameter contribution value is used as the final risk coefficient of each process parameter, and the processing process is as follows:

[0044]

[0045] Among them, RPN i is the risk coefficient of process parameter i, is the contribution value of process parameter i to the SPE statistics of defective samples.

[0046] The beneficial effects of the above technical solution of the present invention are as follows:

[0047] In the above scheme, starting from the perspective of sample data, the correlation between process parameters and quality parameters is fully considered, which can realize online automatic analysis of defective samples, quantify the risk coefficient of each process parameter, and give the dominant process parameters of the defects, thereby improving the analysis efficiency and the credibility of the analysis results, which can effectively replace manual labor, improve analysis efficiency and save analysis costs. Description of the Drawings

[0048] Figure 1 It is a process flow chart of a method for identifying the causes of head waviness defects in hot-rolled strip by improved partial least squares of the present invention;

[0049] Figure 2 It is a graph of the local outlier factor detection results of the remaining samples after abnormal sample rejection in the embodiment of the present invention;

[0050] Figure 3 It is a graph of the relationship between the number of latent variables and evaluation indexes in the embodiment of the present invention;

[0051] Figure 4 It is a graph of the SPE statistic results of each sample in the embodiment of the present invention;

[0052] Figure 5 It is a graph of the contribution results of each process parameter to the defective samples in the embodiment of the present invention;

[0053] Figure 6 It is a graph of the risk assessment results of each process parameter in the embodiment of the present invention;

[0054] Figure 7 It is a sample distribution graph of the F1 rolling force head deviation in the embodiment of the present invention;

[0055] Figure 8 It is a sample distribution graph of the F5 work roll shifting amount in the embodiment of the present invention. Detailed Embodiment

[0056] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0057] The present invention provides a method for identifying the causes of head waviness defects in hot-rolled strip by improved partial least squares.

[0058] As Figure 1 shown, the method includes the following steps:

[0059] S1: Obtain the temperature, roll gap, rolling force process parameter data and the flatness hit rate of the strip head as quality data of excellent samples with qualified head waviness and defective samples with unqualified head waviness through the on-site data acquisition system, and perform data preprocessing on the obtained excellent sample and defective sample data to achieve the rejection of outlier samples with abnormal values and dimensionality reduction;

[0060] S2: Construct a feasibility test model: bring the sample data set into the improved partial least squares algorithm, calculate the square prediction error statistic of the sample, that is, the sample SPE statistic, and the control limit of the square prediction error statistic, that is, the sample SPE statistic control limit, and adjust the value of the latent variable k according to the degree of separation between the defective sample and the excellent sample SPE statistic and the elbow rule, and use the defective sample SPE statistic exceeding the sample SPE statistic control limit and the obvious separation between the defective sample SPE statistic and the excellent sample SPE statistic as the applicability test standard of the sample and the method;

[0061] S3: Contribution value of each process parameter of the sample SPE statistic obtained by the S2 feasibility test model Calculate and express it in the form of a histogram as the initial risk coefficient of each process parameter;

[0062] S4: Reprocess the risk coefficients of each process parameter obtained in S3, unify the risk coefficient evaluation index, and calculate the contribution value of each process parameter to the SPE statistic of the defective sample. Perform summation to calculate the proportion of process parameter contribution values, and use this as the final risk coefficient of each process parameter;

[0063] S5: Compare the risk coefficients of various process parameters. The process parameter with the largest risk coefficient is the dominant defect factor.

[0064] This will be described below with reference to specific embodiments.

[0065] Taking the production data of the 2250 hot rolling production line of a steel plant as an example, the cause of the wave-shaped defect at the head of the hot-rolled strip is identified. The steel grade is MRTRG00502, the thickness is 3.5mm, and the width is 1722mm. The production sample is selected. The quality parameter is the symmetric flatness hit rate of the head of the hot-rolled strip. The process parameters are mainly related parameters such as temperature, rolling force, and roll gap. The specific implementation steps are as follows:

[0066] Step 1:

[0067] The data obtained on site was sorted out and samples with missing values ​​were removed. The Pearson correlation coefficient matrix was calculated. By analyzing the correlation coefficient, the process parameters were reduced from the original 74 dimensions to 15 dimensions including furnace temperature, rolling force deviation, bending roll, etc. The samples after dimensionality reduction were tested for outliers and abnormal samples in the sample set were excluded. The local outlier factor LOF was calculated. k (p), samples with local outlier factors greater than 1 are removed, and the local outlier factors LOF of the removed samples are k (p) Review results such as Figure 2 As shown, Figure 2In the middle is the retest result after processing. The absence of outliers indicates no abnormal points, suggesting that the processing is effective.

[0068] The process parameters of the sample set and each sample after elimination are shown in the following table:

[0069]

[0070] Step 2: Apply the Mahalanobis distance transformation formula to the data after being processed in Step 1 for conversion.

[0071] The sample data after Mahalanobis distance conversion is as follows:

[0072]

[0073]

[0074] The Pearson correlation coefficients between the process parameters and quality parameters in the original samples and the samples after Mahalanobis distance conversion are as follows:

[0075]

[0076] Perform partial least squares diagnostic analysis on the data samples after Mahalanobis distance transformation. Use the idea of recombining characteristic variables to decompose it, then calculate the SPE statistic for each sample. After optimizing the number of latent variables through the elbow rule, it is found that the effect is best when the number of latent variables k = 4. The results are as Figure 3 and Figure 4 shown. By comparing with the actual data set, it is found that the defective samples mined in this step are the same as the actual defective samples, indicating that this model can identify defective samples through the numerical distribution of sample process parameters, and also shows that the risk coefficients of each process parameter can be analyzed through the numerical distribution of the process parameters of defective samples.

[0077] Step 3: After comparing and confirming that the positions of the defective samples determined in Step 2 are correct with the actual positions, calibrate the defects in the sample set, and then calculate the contribution values of each process parameter in the sample to the SPE statistic of the defective samples. The results are as Figure 5 shown.

[0078]

[0079] Step 4: According to the contribution values of each process parameter to the SPE statistic of the defective samples calculated in Step 3, perform the transformation from the statistic to the risk coefficient. Take the ratio of the process parameter contribution value to the total value of the process parameter contribution values as the final risk coefficient of each process parameter. The risk coefficient RPN i has the following conversion relationship:

[0080]

[0081] After calculation, the risk coefficient is presented, such as Figure 6 shown.

[0082] Figure 6 It shows the risk coefficients of various process parameters of the defective samples, and the two process parameters with the largest risk coefficients can be found: the head deviation of F1 rolling force and the working roll shifting amount of F5. Their risk coefficients are (41.3, 22.3) respectively. Combining the analysis of the large deviation of the parameter values of the head deviation of F1 rolling force and the working roll shifting amount of F5 of the defective samples from those of the excellent samples, such as Figure 7 、 Figure 8 , and the high correlation degree between the two parameters of the head deviation of F1 rolling force and the working roll shifting amount of F5 and the quality parameters, it can be known that the analysis result is accurate.

[0083] Using the defect assessment model established by this method process, the analysis result can be obtained in 3.8 seconds, greatly improving the defect analysis efficiency.

[0084] The above is the preferred implementation manner of the present invention. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the principle described in the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An identification method for the causes of head waviness defects in hot-rolled strip steel by improving partial least squares, characterized in that: It includes the following steps: S1: Obtain the temperature, roll gap, rolling force process parameter data of excellent samples with qualified head waviness and defective samples with unqualified head waviness, and the flatness hit rate of the strip steel head as quality data through the on-site data acquisition system, and perform data preprocessing on the obtained excellent sample and defective sample data to achieve the elimination of outlier samples and dimensionality reduction; S2: Substitute the sample data set containing excellent samples and defective samples into the partial least squares algorithm improved by Mahalanobis distance transformation, calculate the squared prediction error statistic of the sample, that is, the sample SPE statistic, and the control limit of the squared prediction error statistic of the sample, that is, the control limit of the sample SPE statistic. Adjust the value of the latent variable k according to the separation degree between the defective sample and the excellent sample SPE statistic and the elbow rule. Taking that the defective sample SPE statistic exceeds the control limit of the sample SPE statistic and achieving the relative maximum separation between the defective sample SPE statistic and the excellent sample SPE statistic as the applicability test standard between the algorithm and the sample set. If the defective sample cannot be detected or the excellent sample is detected as a defective sample, it indicates that the process parameter selection is unreasonable, and the parameter selection, data acquisition and data preprocessing are carried out again; S3: Calculate the contribution values of each process parameter to the defect sample SPE statistic obtained in S2, and present them in the form of a histogram as the initial risk coefficients of each process parameter; ​ S4: Reprocess the initial risk coefficients of each process parameter obtained in S3, unify the risk coefficient evaluation index, and sum up the contribution values of each process parameter to the SPE statistic of the defect sample to calculate the proportion of the contribution value of the process parameter, and use this as the final risk coefficient of each process parameter; S5: Compare the magnitudes of the risk coefficients of each process parameter, and the process parameter with the largest risk coefficient is the dominant defect factor for the occurrence of defects; The specific content of S2 is: The sample data after being processed by S1 is transformed by using the Mahalanobis distance transformation formula, and the Mahalanobis distance D(X) is calculated as follows: where X is the sample vector, s is the covariance matrix of the samples, and μ A represents the sample mean; Perform partial least squares diagnostic analysis on the data samples after Mahalanobis distance transformation, and decompose it by using the idea of recombining characteristic variables, as follows: u = u y +e = PR T u + e e = (I m*m - PR T )u where \(u\in R\) m*l is the data sample set, \(m\) is the number of samples, \(l\) is the number of process parameters, and \(u\) y is the fitting sample, \(e\) is the corresponding residual matrix, \(P\) is the load matrix, and \(R\) T is the transpose matrix of the load matrix, and \(I\) m*m is the identity matrix; Then calculate the SPE statistic of each sample, and the formula is as follows: SPE = ||e|| 2 = ||((I m*m - PR T )u)|| 2 Control limit of the SPE statistic, SPE k It is calculated as follows: g = s SPE / 2μ where s SPE is the variance of the sample SPE statistic, μ is the mean of the sample SPE statistic, is the chi-square value, obtained by looking up the table according to the confidence level and degrees of freedom, and g is an intermediate variable; Calculate the measurement index S of the number of latent variables k: S = S error,k -S yes,k Among them, S yes,k is the variance of the SPE statistic of excellent samples when the number of latent variables is k, is the mean of the SPE statistic of excellent samples when the number of latent variables is k, and SPE j,k is the SPE statistic of the j-th sample when the number of latent variables is k, and n is the number of excellent samples; S error,k is the difference between the SPE statistic of defective samples and the mean of the SPE statistic of excellent samples when the number of latent variables is k, and SPE error,k is the SPE statistic of defective samples when the number of latent variables is k; After defining the value range of the number of latent variables k, apply the elbow rule to the S curve of the measurement index to select the number of latent variables k, where the value range of k is within the range of 2 - l / 2, and l is the number of process parameters; In the above S3 The calculation method is as follows: Among them, is the contribution value of each variable to the sample SPE statistic, is the i-th row of the identity matrix, u ∈ R m*l is the data sample set, m is the number of samples, P is the load matrix, R T is the transpose matrix of the load matrix, I m*m is the identity matrix, and l is the number of process parameters.

2. The identification method for the causes of head waviness defects in hot-rolled strip steel by improving partial least squares according to claim 1, characterized in that: The data preprocessing in S1 includes variance test, correlation test, and local outlier factor test.

3. The identification method for the causes of head waviness defects in hot-rolled strip steel by improving partial least squares according to claim 1, characterized in that: The specific elimination of outlier samples in S1 is as follows: retrieve missing values for excellent samples, eliminate the excellent samples with missing values, and then use the Local Outlier Factor algorithm to detect outliers in the samples after eliminating missing values. By calculating the Local Outlier Factor LOF z (p) and compare it with 1. When LOF z (p) is greater than 1, eliminate it; Among them, N z (p) represents the z-distance neighborhood of the sample point p, that is, all the sample points within the z-distance of the sample point p, |N z (p) represents the number of all sample points within the z-distance neighborhood of the sample point p, lrd z (p) represents the reciprocal of the average reachability distance from the sample points within the z-neighborhood of the sample point p to the sample point p, lrd z (o) represents the reciprocal of the average reachability distance from the points within the z-neighborhood of the sample point o to the sample point o.

4. The identification method for the causes of head waviness defects in hot-rolled strip steel by improving partial least squares according to claim 1, characterized in that: The reprocessing process in S4 is as follows: Among them, RPN i is the risk coefficient of process parameter i, is the contribution value of process parameter i to the SPE statistic of the defect sample, and l is the number of parameter terms.

Images