A prediction method for mold breakout based on eigenvector and logistic regression model

Abstract

The present invention provides a method for predicting steel leakage in a crystallizer based on a feature vector and a logistic regression model, and belongs to the technical field of continuous casting detection in iron and steel metallurgy. The method for predicting steel leakage in a crystallizer extracts a visualized feature vector from an abnormal area of ​​the temperature rate of a copper plate of a crystallizer, and classifies the feature vector using a logistic regression model, thereby detecting and predicting steel leakage in a crystallizer. The present invention constructs a feature vector containing static and dynamic characteristics of a bonding area through a visualized thermal image of the temperature rate of a copper plate of a crystallizer, and classifies the feature vector through a logistic regression model, thereby realizing detection and prediction of steel leakage in a crystallizer. The present invention performs real-time detection and prediction of steel leakage in a crystallizer based on a logistic regression model, and can significantly reduce the false alarm rate while ensuring that all steel leakage is reported, thereby effectively improving the prediction accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of continuous casting detection in iron and steel metallurgy, and relates to a method for predicting steel leakage in a crystallizer based on a characteristic vector and a logistic regression model. Background Art

[0002] During the continuous casting process, the non-uniformly solidified primary billet shell in the crystallizer cannot withstand the dual effects of the static pressure of the molten steel and the pulling force, and is prone to fracture at weak points, resulting in steel leakage. Steel leakage is a major safety accident in continuous casting production, which not only endangers personal safety and damages equipment, but also causes forced interruption of production, affecting the output and product quality of the casting machine. With the continuous development and progress of continuous casting technology, the probability of steel leakage can be effectively reduced by standardizing operations and maintaining good operating conditions of equipment. Steel leakage is closely related to many factors such as protective slag, pulling speed, liquid level fluctuations, and heat flux density. Although metallurgical workers and scholars at home and abroad have conducted extensive research on the causes of its formation, it is difficult to completely avoid the occurrence of steel leakage. Therefore, reducing and preventing steel leakage accidents has always been the focus of metallurgical workers at home and abroad. Crystallizer steel leakage detection is the core of abnormal prediction of continuous casting process and is of great significance.

[0003] Invention patent CN201710346785.1 discloses a continuous casting leak prediction method, which detects the occurrence of bonding leaks in real time based on logical judgment. The main steps are: obtaining real-time data of the casting process such as thermocouple temperature, pulling speed and crystallizer liquid level, reading the parameters of each logical judgment rule in the system parameter configuration file, and using the model algorithm to judge the light alarm and heavy alarm of bonding leaks. This method reduces the impact of changes in casting process and equipment conditions on the leak prediction algorithm and system, and designs the function of system parameter self-learning to adapt to the actual complex changes in production rules. However, the leak prediction model based on logical judgment is highly dependent on equipment parameters, process conditions, and physical parameters, and requires frequent adjustment of thresholds and parameters, resulting in poor robustness of the prediction algorithm.

[0004] Invention patent CN201811507030.6 discloses a method for predicting continuous casting steel leakage, which provides a method for predicting crystallizer steel leakage based on feature vectors and hierarchical clustering. The main steps are: extracting the temperature feature vectors of bonding steel leakage, normal operating condition historical data, and online measured data, and establishing a feature vector sample set; normalizing the sample set and performing hierarchical clustering; then checking and judging whether the feature vector extracted online belongs to the steel leakage cluster, and then identifying and predicting crystallizer steel leakage. The method based on hierarchical clustering gets rid of the limitations of manually setting and adjusting the prediction parameters, but it is difficult to meet the real-time and online application requirements of steel leakage prediction, so the model is less practical.

[0005] In view of the shortcomings of the existing steel leakage prediction methods, and taking into account the complexity of the prediction algorithm and the cumbersomeness of the sample making process, the present invention proposes to extract the visual feature vector of the bonding area based on the temperature rate thermal image of the crystallizer copper plate, and classify the steel leakage and non-steel leakage samples by combining the logistic regression model, and finally use the trained model to detect and predict the steel leakage of the crystallizer in real time. Summary of the invention

[0006] The purpose of the present invention is to propose a method for predicting mold leakout based on eigenvector and logistic regression model, to detect and predict sticking leakout in real time and accurately, and to provide a new solution for abnormal monitoring of continuous casting process.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for predicting mold breakout based on feature vector and logistic regression model, which extracts visual feature vectors from abnormal temperature rate areas of copper plates in the mold and classifies the feature vectors using a logistic regression model, thereby detecting and predicting mold breakout, and specifically includes the following steps:

[0009] Step 1: Feature extraction of abnormal temperature rate area of ​​the crystallizer copper plate

[0010] (1) Multiple thermocouples are arranged on the copper plates on the inner and outer arc widths of the crystallizer and on the left and right narrow copper plates. The temperature of the thermocouples on the crystallizer copper plates is detected online, and the temperature value of the crystallizer copper plates at the non-thermocouple measuring point is calculated by interpolation algorithm. Using computer graphics and OpenGL technology, the copper plate temperature is mapped to a two-dimensional plane to obtain the corresponding two-dimensional temperature thermal image.

[0011] (2) The temperature change rate at the copper plate measuring point and non-measuring point is calculated by the inter-frame difference algorithm to obtain a two-dimensional temperature rate thermal image corresponding to the copper plate temperature.

[0012] (3) After statistically summarizing the temperature rate data in the sample library, set the temperature rate threshold as T z The threshold segmentation algorithm is used to distinguish the normal temperature fluctuation area from the abnormal temperature fluctuation area, and the areas with a temperature rate less than T are removed from the two-dimensional temperature rate thermal image. z The normal temperature fluctuation area, and the run-length recursive algorithm is used to calculate the temperature rate greater than T z The abnormal points are marked for connectivity and the temperature rate abnormal area is obtained.

[0013] (4) The temperature rate abnormal area includes the bonding leaking area and the normal working condition area, and the height H, width W, area S, and lateral movement rate V of the two are extracted respectively. y , longitudinal movement rate V x Five visualization features.

[0014] Step 2: Construction and processing of feature vectors of abnormal regions

[0015] (1) The regional characteristics of the bonded steel leak extracted in the first step are compared with the pulling speed V under the bonded steel leak. c Combined into feature vector X B , and the normal working area characteristics and the pulling speed V under normal working conditions c 'Combined into feature vector X N :

[0016] X B =[H B ,W B ,S B ,V Bx ,V By ,V C ]

[0017] X N =[H N ,W N ,S N ,V Nx ,V Ny ,V C ']

[0018] (2) Obtain m samples of feature vectors X of bonding leakage areas B and n samples of normal operating condition feature vectors X N , form a sample set Q:

[0019] Q = {(X B1 ,1),(X B2 ,1),...,(X Bm ,1),(X N1 ,0),(X N2 ,0),...,(X Nn ,0)}

[0020] In the formula, 1 and 0 represent the category labels of samples in the bonding leaking area and samples in normal working condition, respectively.

[0021] (3) Normalize the feature vector samples in Q:

[0022]

[0023] In the formula, X imin , X imax Respectively represent the minimum and maximum values ​​of the i-th dimension feature of the feature vector sample X, F i It represents the value corresponding to the normalized feature of the ith dimension of the feature vector X. The normalized sample set is denoted as P:

[0024] P={(F B1 ,1),(F B2 ,1),...,(F Bm ,1),(F N1 ,0),(F N2 ,0),...,(F Nn ,0)}

[0025] Step 3: Adjust hyperparameters and train the logistic regression model

[0026] (1) Sample set division.

[0027] The sample set P is divided K times according to the principle of K-fold cross-validation. Each time, K-1 subsets of P are taken as training sets R, and the remaining subset is taken as validation set E. After division, a total of K different training sets R and validation sets E are obtained.

[0028] (2) Grid search to adjust hyperparameters.

[0029] Based on the original logistic regression model, K rounds of training and validation were performed for each combination of model hyperparameters. In each round, the model was trained on a training set R containing K-1 subsets, and the model was validated on a validation set E containing the remaining subset, that is, the model classification accuracy was calculated. The average classification accuracy of the K-round model was taken as the validation score for this set of hyperparameters.

[0030] The original logistic regression model is as follows:

[0031]

[0032] Where P(Y=1|x) represents the probability that the model predicts the feature vector sample x as leaking steel, w is the weight vector, and b is the bias term.

[0033] (3) Build the final model.

[0034] Compare the validation scores calculated for each combination of the logistic regression model hyperparameters, and select the combination with the highest validation score as the final hyperparameters for the model.

[0035] (4) Based on the finally constructed logistic regression model, the model is trained on the sample set P to obtain the model parameters w and b.

[0036] Step 4: Online detection and prediction of steel leakage based on logistic regression model

[0037] (1) The typical visual features of the abnormal temperature rate area of ​​the copper plate of the crystallizer are extracted in real time. The features are combined and normalized according to the casting speed under real-time working conditions to construct the abnormal area feature vector F. fv ;

[0038] (2) The abnormal region feature vector F fv Input into the logistic regression model and get the model's predicted value y.

[0039] (3) Predict mold leaks based on the predicted value y of the logistic regression model. If y = 1, it is a leak due to bonding, an alarm is issued and the casting speed is quickly reduced; if y = 0, it is a normal operating condition, and the detection and prediction of leaks due to bonding will continue at the next moment.

[0040] The above-mentioned method for predicting steel leakage is applicable to predicting steel leakage of continuous casting billets such as slabs, square billets, round billets, and profiled billets.

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

[0042] The prediction method provided by the present invention constructs a feature vector containing static and dynamic characteristics of the bonding area through a crystallizer copper plate temperature rate visualization thermal image, and classifies the feature vector through a logistic regression model, thereby realizing the detection and prediction of crystallizer leakage. The method performs real-time detection and prediction of crystallizer leakage based on a logistic regression model, and can significantly reduce the false alarm rate and effectively improve the prediction accuracy while ensuring that all leakage is reported. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The invention provides a method and process for predicting mold leakage.

[0044] Figure 2 Schematic diagram of the arrangement of thermocouples on the copper plate of the crystallizer.

[0045] Figure 3 Visualize the characteristic map of the normal operating area. Figure 3 (a) is the initial formation diagram of the normal working condition area; Figure 3 (b) is the longitudinal tensile diagram of the normal working condition area; Figure 3 (c) is the in-situ expansion diagram of the normal operating area; Figure 3 (d) is the longitudinal contraction diagram of the normal operating area.

[0046] Figure 4 This is a visualization feature map of the bonding leakage area. Figure 4 (a) is the initial bonding formation diagram; Figure 4 (b) is the lateral expansion diagram of the bonding area; Figure 4 (c) is the longitudinal propagation diagram of the bonding area; Figure 4 (d) is the “V”-shaped characteristic diagram of bonded steel leakage.

[0047] Figure 5 Visualize feature maps for online detection of temperature rate anomaly areas. Figure 5 (a) is the characteristic diagram of the “V” type of bonding leakage;

[0048] Figure 5(b) is the characteristic diagram of the normal operating area.

[0049] In the figure: 1 thermocouple; 2 outer arc wide copper plate; 3 left narrow copper plate; 4 right narrow copper plate; 5 inner arc wide copper plate. DETAILED DESCRIPTION

[0050] The present invention will be further described below through specific embodiments in conjunction with the accompanying drawings.

[0051] like Figure 1 The figure shows the flow chart of the method for predicting steel leakage in the crystallizer. First, the visual features of the abnormal temperature rate area of ​​the copper plate of the crystallizer are extracted and preprocessed to construct a six-dimensional feature vector; secondly, the logistic regression model is adjusted and trained; finally, the feature vector is classified and steel leakage is predicted through the trained logistic regression model.

[0052] Step 1: Visualization of the temperature rate of the crystallizer copper plate and extraction of abnormal area features

[0053] (1) Figure 2 The figure shows the distribution diagram of the crystallizer copper plate and its thermocouple. The crystallizer is composed of four copper plates with a total height of 900mm and an effective height of 800mm during casting. Three rows of 19 thermocouples are arranged on the inner and outer arc wide copper plates 5 and 2 of the crystallizer, and three rows of 1 thermocouple are arranged on the left and right narrow copper plates 3 and 4, with a total of 120 thermocouples. Among them, the distance between the first row of thermocouples and the top of the crystallizer is 210mm, the spacing between the first and second rows of thermocouples is 115mm, the spacing between the second and third rows of thermocouples is 120mm, and the spacing between two adjacent columns of thermocouples is 150mm. The temperature of all crystallizer copper plate thermocouples is detected online, and the crystallizer copper plate temperature value at the non-thermocouple measuring point is calculated by interpolation algorithm. Using computer graphics and OpenGL technology, the copper plate temperature is mapped to a two-dimensional plane to obtain the corresponding two-dimensional temperature thermal image.

[0054] (2) The temperature change rate at the copper plate measuring point and non-measuring point is calculated by the inter-frame difference algorithm to obtain a two-dimensional temperature rate thermal image corresponding to the copper plate temperature.

[0055] (3) The temperature rate threshold is set to 0.3°C / s, and the threshold segmentation algorithm is used to distinguish the normal temperature fluctuation area from the abnormal temperature fluctuation area. The normal temperature fluctuation area with a temperature rate less than the threshold is removed from the two-dimensional temperature rate thermal image, and the run-length recursive algorithm is used to segment the temperature fluctuation area with a temperature rate greater than T z The abnormal points are marked for connectivity and the temperature rate abnormal area is obtained.

[0056] (4) Figure 3The figure shows the visualization feature map of the normal working condition area. The simulation interval along the casting direction is the first to third rows of thermocouples, distributed on 100 pixels; the horizontal simulation interval is the first to nineteenth rows of thermocouples, distributed on 300 pixels. Figure 3 The typical V-shaped characteristics of bonding leakage did not appear at T1 to T4. Figure 4 The visualization feature map of the bonding leaking area shown in the figure shows that the typical V-shaped feature of bonding leaking gradually appears from time T1 to T4, and the interval between each time is 3s. The height H of the bonding leaking area at time T4 is extracted. B =22.56cm, width W B =35.91cm, shaded area S B =453.97cm 2 , and the lateral movement rate V is calculated based on the change of the centroid coordinates of the bonding leaky steel area over time By =0.33m / min, longitudinal moving speed V Bx =0.31m / min. Similarly, we can get Figure 3 There are five visualization features corresponding to the normal operating area T4, namely: N =16.45cm, W N =22.07cm, S N =243.25cm 2 , V Ny =-0.18m / min, V Nx =0.23m / min.

[0057] Step 2: Construction and processing of feature vectors of abnormal regions

[0058] (1) The characteristics of the bonding leaky area extracted in the first step are combined with the pulling speed V under bonding leaky c Combined into feature vector X B , and the normal working area characteristics and the pulling speed V under normal working conditions c 'Combined into feature vector X N :

[0059] X B =[H B ,W B ,S B ,V Bx ,V By ,V C ]=[22.56,35.91,453.97,0.31,0.33,0.80]

[0060] X N =[H N ,W N ,S N ,VNx ,V Ny ,V C ']=[16.45,22.07,243.25,0.23,-0.18,0.70]

[0061] (2) Obtain 40 samples of feature vectors X of bonding leakage areas B and 40 samples of normal operating characteristic vectors X N , form a sample set Q:

[0062] Q = {(X B1 ,1),(X B2 ,1),...,(X B40 ,1),(X N1 ,0),(X N2 ,0),...,(X N40 ,0)}

[0063] In the formula, 1 and 0 represent the category labels of samples in the bonding leaking area and samples in normal working condition, respectively.

[0064] (4) Normalize the feature vector samples in Q:

[0065]

[0066] Where, X imin , X imax Respectively represent the minimum and maximum values ​​of the i-th dimension of the feature vector X, F i Represents the value corresponding to the normalized feature of the ith dimension of feature vector X. B , X N The normalized eigenvector is F B =[0.96,0.12,0.14,0.21,0.58,0.62], F N =[0.68,0.06,0.08,0.15,0.5,0.46].

[0067] According to the above feature vector construction and processing method, the new sample set after normalization is recorded as P:

[0068] P={(F B1 ,1),(F B2 ,1),...,(F B40 ,1),(F N1 ,0),(F N2 ,0),...,(F N40 ,0)}

[0069] Step 3: Adjust hyperparameters and train the logistic regression model

[0070] (1) Sample set division.

[0071] The sample set P is divided 5 times according to the principle of 5-fold cross-validation. Each time, 4 subsets of P are taken as training sets R, and the remaining 1 subset is taken as validation set E. After division, 5 different training sets R and validation sets E are obtained.

[0072] (2) Grid search to adjust hyperparameters.

[0073] Based on the original logistic regression model, 5 rounds of training and validation are performed for each combination of model hyperparameters. In each round, the model is trained on the training set R containing 4 subsets, and the model is validated on the validation set E of the remaining 1 subset, that is, the model classification accuracy is calculated. The average classification accuracy of the 5-round model is taken as the validation score of this set of hyperparameters. The original logistic regression model is as follows:

[0074]

[0075] Where P(Y=1|x) represents the probability that the model predicts the feature vector sample x as leaking steel, w is the weight vector, and b is the bias term.

[0076] (3) Constructing the final model. Compare the validation scores calculated for each combination of the hyperparameters of the logistic regression model, and select the hyperparameter combination with a validation score of 71.1% as the final hyperparameter of the logistic regression model.

[0077] (4) Based on the finally constructed logistic regression model, the model is trained on the sample set P to obtain the model parameters, where the weight vector w = [-0.60, -0.15, -0.36, -0.83, 0.51, 0.15] and the bias term b = 0.74.

[0078] Step 4: Online detection and prediction of steel leakage based on logistic regression model

[0079] (1) Real-time extraction of typical visualization features of abnormal temperature rate areas of the crystallizer copper plate, such as Figure 5 As shown in the figure, after combining the features and normalizing them in combination with the pulling speed under real-time working conditions, the feature vector of the abnormal area is constructed:

[0080] F fv1 =[0.82,0.12,0.09,0.43,0.66,0.54]; F fv2 =[0.24,0.06,0.03,0.21,0.42,0.31]

[0081] (2) The abnormal region feature vector F fv1 and F fv2 Input the logistic regression model L and get the model's predicted values ​​y1 and y2:

[0082] y1=L(F fv1 )=L([0.82,0.12,0.09,0.43,0.66,0.54])=1

[0083] y2=L(F fv2 )=L([0.24,0.06,0.03,0.21,0.42,0.31])=0

[0084] (3) Predict mold leaks based on the predicted value y of the logistic regression model. y1 = 1, corresponding to the data label of the bonding leak area, leaks occur, an alarm is issued and the casting speed is quickly reduced; y2 = 0, corresponding to the data label of the normal working condition area, no leaks occur, and the leak detection and prediction of the next moment are continued.

[0085] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for predicting mold breakout based on eigenvector and logistic regression model, characterized in that: The method for predicting mold breakout extracts a visual feature vector from an abnormal temperature rate region of a copper plate of a crystallizer, and classifies the feature vector using a logistic regression model, and finally detects and predicts mold breakout, comprising the following steps: Step 1: Feature extraction of abnormal temperature rate area of ​​the crystallizer copper plate (1) Arranging a plurality of thermocouples on the copper plates of the inner and outer arc wide surfaces of the crystallizer and on the left and right narrow copper plates; detecting the temperature of the thermocouples of the crystallizer copper plates online, and calculating the temperature value of the crystallizer copper plates at non-thermocouple measuring points by an interpolation algorithm; mapping the temperature of the copper plates to a two-dimensional plane to obtain a corresponding two-dimensional temperature thermal image; (2) Calculate the temperature change rate at the copper plate measuring point and non-measuring point by using the inter-frame difference algorithm to obtain a two-dimensional temperature rate thermal image corresponding to the copper plate temperature; (3) After statistically summarizing the temperature rate data in the sample library, set the temperature rate threshold as T z The threshold segmentation algorithm is used to distinguish the normal temperature fluctuation area from the abnormal temperature fluctuation area, and the areas with a temperature rate less than T are removed from the two-dimensional temperature rate thermal image. z The normal temperature fluctuation area, and the run-length recursive algorithm is used to calculate the temperature rate greater than T z The abnormal points are marked for connectivity, and the temperature rate abnormal area is obtained; (4) The temperature rate abnormal area includes the bonding leaking area and the normal working condition area, and the height H, width W, area S, and lateral movement rate V of the two are extracted respectively. y , longitudinal movement rate V x Five visualization features; Step 2: Construction and processing of feature vectors of abnormal regions (1) The regional characteristics of the bonded steel leak extracted in the first step are compared with the pulling speed V under the bonded steel leak. c Combined into feature vector X B , and the normal working area characteristics and the pulling speed V under normal working conditions c 'Combined into feature vector X N : X B =[H B ,W B ,S B ,V Bx ,V By ,V C ] X N =[H N ,W N ,S N ,V Nx ,V Ny ,V C '] Among them, H B Indicates the height of the bonding leak area, W B Indicates the width of the bonding leak area, S B Indicates the area of ​​the bonding leak area, V Bx Indicates the lateral movement rate of the bonding leaking area, V By Indicates the longitudinal movement rate of the bonding leaking area; H N Indicates the height of the normal operating area, W N Indicates the width of the normal operating area, S N Indicates the area of ​​the normal operating area, V Nx Indicates the lateral movement rate in the normal operating area, V Ny Indicates the longitudinal movement rate in the normal operating area; (2) Obtain m samples of feature vectors X of bonding leakage areas B and n samples of normal operating condition feature vectors X N , form a sample set Q: Q={(X B1 ,1),(X B2 ,1),...,(X Bm ,1),(X N1 ,0),(X N2 ,0),...,(X Nn ,0)} In the formula, 1 and 0 represent the category labels of samples in the bonding leaking area and samples in normal working condition, respectively; (3) Normalize the feature vector samples in Q: Where, X i represents the i-th dimension feature of the feature vector sample X, X imin , X imax Respectively represent the minimum and maximum values ​​of the i-th dimension feature of the feature vector sample X, F i Represents the value corresponding to the normalized feature of the ith dimension of the feature vector X; the normalized sample set is denoted as P: P={(F B1 ,1),(F B2 ,1),...,(F Bm ,1),(F N1 ,0),(F N2 ,0),...,(F Nn ,0)} Step 3: Adjust hyperparameters and train the logistic regression model (1) Sample set division; Divide the sample set P into K times according to the principle of K-fold cross-validation. Each time, take K-1 subsets from P as the training set R, and the remaining subset as the validation set E. After division, a total of K different training sets R and validation sets E are obtained. (2) Grid search to adjust hyperparameters; Based on the original logistic regression model, K rounds of training and validation are performed for each combination of model hyperparameters. In each round, the model is trained on a training set R containing K-1 subsets, and validated on a validation set E containing the remaining subset, that is, the model classification accuracy is calculated; the average classification accuracy of the K-round model is taken as the validation score for this set of hyperparameters; The original logistic regression model is as follows: Where P(Y=1|x) represents the probability that the model predicts the feature vector sample x as leaking steel, w is the weight vector, and b is the bias term; (3) Build the final model; Compare the validation scores calculated for each combination of the logistic regression model hyperparameters and select the combination with the highest validation score as the final hyperparameters of the model; (4) Based on the final constructed logistic regression model, the model is trained on the sample set P to obtain the model parameters w and b; Step 4: Online detection and prediction of steel leakage based on the logistic regression model (1) The typical visual features of the abnormal temperature rate area of ​​the copper plate of the crystallizer are extracted in real time. The features are combined and normalized according to the casting speed under real-time working conditions to construct the abnormal area feature vector F. fv ; (2) The abnormal region feature vector F fv Input into the logistic regression model and obtain the model's predicted value y; (3) According to the predicted value y of the logistic regression model, mold leakage is predicted; if y = 1, it is bonding leakage, an alarm is issued and the casting speed is quickly reduced; if y = 0, it is a normal operating condition, and the bonding leakage detection and prediction are continued at the next moment.

2. The method for predicting mold breakout based on characteristic vector and logistic regression model according to claim 1, characterized in that: The method for predicting steel leakage is suitable for predicting steel leakage of slabs, square billets, round billets, profiled billets or other continuous casting billets.

Images