Gray Wolf + NGBoost method for surface defect prediction of hot rolled strip based on genetic algorithm of ingot casting

Abstract

The present invention discloses a Gray Wolf + NGBoost method for predicting surface defects of hot-rolled strip steel based on ingot genetics, which belongs to the field of hot-rolled strip steel surface quality diagnosis. Specifically, the method comprises the following steps: using the surface quality image data of hot-rolled strip steel, hot rolling process control parameters and ingot steel parameters of the same time dimension as the input features of each sample, and outputting the detection results of the hot-rolled strip steel surface defects based on ingot genetics as labels to construct a data set; then, the data set is cleaned, smoothed, filtered and sample homogenized, and divided into a training set and a test set. The Gray Wolf algorithm and the NGBoost integrated model are integrated to construct a "Grey Wolf + NGBoost" model for predicting surface defects of hot-rolled strip steel based on ingot genetics, which is then trained and tested for verification. The results show that the classification prediction accuracy of the model is better than that of the NGBoost model. The present invention solves the problem that the existing NGBoost model has low accuracy due to manual parameter adjustment and the problem that the prediction of hot-rolled strip steel surface defects based on ingot genetics lacks a high-precision intelligent model.

Description

Technical Field

[0001] The present invention belongs to the field of hot-rolled strip surface quality diagnosis, and specifically relates to a "Grey Wolf + NGBoost" method for hot-rolled strip surface defect prediction based on ingot genetics. Background Art

[0002] The surface quality of ingots is one of the key indicators that steel manufacturers focus on, because by testing the surface quality of ingots, the accuracy of various process parameters in the steelmaking and continuous casting links can be traced back, and the quality defects of hot-rolled products caused by ingots can be reduced.

[0003] However, some defects in the ingot cannot be identified and are directly inherited to the hot-rolled strip, resulting in quality defects of the hot-rolled strip products. Such defects are collectively referred to as surface defects of the hot-rolled strip inherited from the ingot, mainly including slag inclusion, warping and cracks.

[0004] With the continuous advancement and application of thin-plate hot rolling technology and the increasing demand for surface quality, image-based hot-rolled strip surface quality inspection has become a hot technology introduced by major steel manufacturers. This is because the traditional method of inspection through local cutting can no longer meet the current market and user needs.

[0005] Image-based methods for inspecting the surface quality of hot-rolled strip primarily utilize high-speed surface-scanning cameras placed above and below the hot-rolling rollers to capture and store images of the strip surface. These cameras then employ filtering and image recognition algorithms to identify defects. However, due to the high computational complexity, low recognition accuracy, and poor stability of existing algorithms, most production sites still rely on manual image recognition for defect identification and classification, limiting the further application of image-based defect detection technology. Summary of the Invention

[0006] The purpose of the present invention is to provide a "Grey Wolf + NGBoost" method for predicting surface defects of hot-rolled strip steel based on ingot genetics, so as to solve the problems of low classification and prediction accuracy of existing hot-rolled strip steel surface defect prediction methods.

[0007] The "Grey Wolf + NGBoost" method for predicting surface defects of hot-rolled strip steel based on ingot genetics has the following specific steps:

[0008] Step 1: For the same time dimension, based on the hot-rolled strip surface quality image data and hot rolling process control parameters, select features and labels to construct a data set;

[0009] The dataset consists of 13,600 samples, each of which has 41 input features and an output label value with four classification characteristics.

[0010] The output label is the actual hot-rolled strip surface quality test result inherited from the ingot, including four categories: slag inclusion, warping, crack or normal.

[0011] Step 2: Clean, smooth, filter and homogenize the data set, and divide it into training set and test set;

[0012] Step 3: Use the Grey Wolf algorithm to collaboratively optimize the parameters of the NGBoost model to form a "Grey Wolf + NGBoost" model for hot-rolled strip surface defect prediction based on ingot genetics, and train the model using the training set;

[0013] The NGBoost model has five parameters, namely, the estimator, the learning rate, the maximum depth, the minimum number of leaf nodes, and the maximum number of splits;

[0014] The specific training process is as follows:

[0015] First, normalize the data set:

[0016]

[0017] X: Feature data input before processing, X * : The data obtained after processing, X min : The minimum value of the feature data input before processing, X max : The maximum value of the feature data input before processing.

[0018] Then, an initial value is set for each parameter of the NGBoost model as the location parameter of the gray wolf; and the population size, individual dimension, parameter optimization range and maximum number of iterations of the algorithm are initialized.

[0019] Next, the classification accuracy is used as the fitness value of each wolf, and the gray wolf population is divided into four levels: α, β, δ, and ω according to the fitness value.

[0020] The hunting process is guided by α, β and δ, and the ω wolf follows these three wolves.

[0021] Finally, the position of the gray wolf is iteratively updated to determine whether the maximum number of iterations has been reached. At the end of the iteration, the optimal gray wolf position is output. The optimal position of the gray wolf is the optimal parameter of the NGBoost model, and the defect prediction "gray wolf + NGBoost" model training is completed.

[0022] Step 4: Use the test set to verify the trained defect prediction "Gray Wolf + NGBoost" model;

[0023] Step 5: For the surface quality image data of hot-rolled strip in the new time period, input the selected features into the verified defect prediction "Grey Wolf + NGBoost" model and directly output the predicted classification label.

[0024] The advantages of the present invention are:

[0025] The Gray Wolf + NGBoost method for predicting surface defects of hot-rolled strip steel based on the inheritance of ingots in the present invention integrates the Gray Wolf algorithm and the NGBoost integrated learning model, can realize intelligent parameter adjustment, and accurately predict the types of surface defects of hot-rolled strip steel based on the inheritance of ingots. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the Gray Wolf + NGBoost method for predicting surface defects of hot-rolled strip steel based on the inherited characteristics of the ingot;

[0027] Figure 2 This is the prediction chart for verifying the “Grey Wolf + NGBoost” model for predicting surface defects of hot-rolled strip steel according to the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation processes to help those skilled in the art have a more complete and clear understanding of the contents of the present invention.

[0029] The "Grey Wolf + NGBoost" method for predicting surface defects of hot-rolled strip steel based on the ingot genetics is as follows: Figure 1 The specific steps are as follows:

[0030] Step 1: For the same time dimension, based on the hot-rolled strip surface quality image data and hot rolling process control parameters, select features and labels to construct a data set;

[0031] The dataset consists of 13,600 samples, each of which has 41 input features and an output label value with four classification characteristics.

[0032] The input features are mainly based on hot rolling process data and ingot parameters, including: smelting number, continuous casting number, heat treatment base, hot flame cleaning base cleaning degree, continuous casting number, tundish molten steel temperature, liquid phase temperature, maximum molten steel volume in the tundish, minimum molten steel volume in the tundish, slab cut length, average temperature in the tundish, maximum and minimum casting speed, average heat exchange rate in the four directions of east, west, south and north, maximum and minimum liquid level, maximum and minimum upper nozzle flow rate, maximum and minimum upper slide flow rate, average upper slide flow rate, maximum, minimum and average upper slide pressure, material length, actual furnace temperature, steel modification or not, slab bottom width, slab top width, maximum and minimum upper nozzle pressure, actual slab width and length, off-line time, transfer time, etc. 41 parameters;

[0033] The output label is the actual hot-rolled strip surface quality test result inherited from the ingot, including four categories: slag inclusion, warping, crack or normal.

[0034] Step 2: Clean, smooth, filter and homogenize the data set, and divide it into training set and test set;

[0035] Step 3: Use the Grey Wolf algorithm to collaboratively optimize the parameters of the NGBoost model to form a "Grey Wolf + NGBoost" model for hot-rolled strip surface defect prediction based on ingot genetics, and train the model using the training set;

[0036] The NGBoost model is an integrated learning model based on natural gradient boosting, which has good prediction accuracy. However, because the prediction accuracy of the NGBoost model is affected not only by the five manually set parameter values, but also by the nonlinearity, correlation and quantity of the sample data input features, it is difficult to directly promote and apply it to industrial production.

[0037] The present invention uses the Gray Wolf algorithm to intelligently optimize the five parameters of the NGBoost model, namely the estimator, learning rate, maximum depth, minimum value of leaf nodes, and maximum number of splits, to construct a "Gray Wolf + NGBoost" prediction model with intelligent parameter adjustment function. The five parameters in the NGBoost model are collaboratively optimized through the Gray Wolf algorithm to find an optimal set of parameter combinations, so that the NGBoost model can achieve the best prediction performance, that is, the highest prediction accuracy.

[0038] As an optimization algorithm for global optimization to find the optimal solution of a function, the Gray Wolf Algorithm is an extremely suitable optimization algorithm for machine learning models that require parameter optimization. Therefore, combining the Gray Wolf Algorithm with the machine learning model can help find the optimal solution for establishing a hot-rolled strip surface defect prediction model and obtain the hot-rolled strip surface defect prediction model with the best prediction effect.

[0039] The specific training process is as follows:

[0040] First, normalize the data set:

[0041]

[0042] X: Feature data input before processing, X * : The data obtained after processing, X min : The minimum value of the feature data input before processing, X max : The maximum value of the feature data input before processing.

[0043] Then, 90% of the hot-rolled strip surface defect dataset of the ingot genetic algorithm was randomly selected as the training set to train the constructed "Grey Wolf + NGBoost" model;

[0044] "Gray Wolf + NGBoost" model parameter optimization and training process flow chart, such as Figure 1 As shown, first set an initial value for each of the five parameters in the NGBoost model:

[0045] The parameter estimator, which is the number of weak estimators in the ensemble, i.e. the number of trees, is set to 1000;

[0046] The learning rate parameter is used to control the learning progress of the model and is set to 0.1;

[0047] The maximum depth parameter controls the maximum depth of the decision tree. The larger the value, the better the fitting effect. However, to prevent overfitting, it is set to 5.

[0048] The minimum parameter of leaf nodes is used to control the minimum number of samples of leaf nodes and is set to 5;

[0049] The maximum number of splits parameter is the minimum number of samples required to control the node re-division, which is used to limit the subtree from continuing to split and is set to 10.

[0050] Using the above preset parameters as the location parameters for the gray wolves, the population size, individual dimensions, parameter optimization range, and maximum number of algorithm iterations were initialized. These five parameters were used to build an NGBoost model on the training set, and the model's classification accuracy was used as the fitness value for each wolf. Based on the fitness values, the gray wolf population was divided into four levels: α, β, δ, and ω. The hunting process was guided by α, β, and δ, with the ω wolf following these three.

[0051] Iterate the position of the gray wolf according to the following formula to determine whether the maximum number of iterations has been reached. At the end of the iteration, output the optimal gray wolf position:

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] Indicates the distance between the gray wolf and its prey. Represents the position of the gray wolf at the t+1th iteration. t is the number of iterations, and is the coefficient vector, is the position of the prey in the tth iteration, is the position of the gray wolf in the tth iteration, Represents the convergence factor, which changes linearly with the number of iterations, decreasing from 2 to 0; and The modulus is a randomly generated number between [0, 1]. Represents the distance between α, β, δ and other gray wolves, Represents the current position of α, β, and δ; is a random vector, It is the current location of the Gray Wolf.

[0060] Finally, the optimal position of the Gray Wolf is the optimal parameter combination of the NGBoost model. The final NGBoost prediction model and the prediction accuracy value at this time are obtained using the optimal parameter combination, and the "Grey Wolf + NGBoost" model for hot-rolled strip surface defect prediction with the optimal parameters is obtained.

[0061] Step 4: Use the test set to verify the trained defect prediction "Gray Wolf + NGBoost" model;

[0062] The remaining 10% of the test set is input into the trained defect prediction "Gray Wolf + NGBoost" model, and the model's predicted labels are compared with the actual labels of the test set to obtain the model prediction accuracy and give the verification results.

[0063] Step 5: For the surface quality image data of hot-rolled strip in the new time period, input the selected features into the verified defect prediction "Grey Wolf + NGBoost" model and directly output the predicted classification label.

[0064] Example:

[0065] The first step is to build a dataset

[0066] The hot-rolled strip surface quality image data of the same time dimension are integrated with the hot rolling process control parameters, 40 control parameters are selected as input features, and the surface quality detection results of the hot-rolled strip inherited from the ingot are output as labels; this dataset is a large 4-category dataset with 13,600 samples and each sample has 41 input features and one output label.

[0067] The second step is to clean, smooth, filter and homogenize the data set; 80% of the samples in the data set are used as the training set and 20% of the samples are used as the test set;

[0068] Step 3: Construction of the "Grey Wolf + NGBoost" model for surface defect prediction of hot-rolled strip based on ingot genetics

[0069] The Gray Wolf Algorithm is used to intelligently optimize the five parameters of the NGBoost model, namely the estimator, learning rate, maximum depth, minimum value of leaf nodes, and maximum number of splits.

[0070] Step 4: Training of the "Grey Wolf + NGBoost" model for surface defect prediction of hot-rolled strip using genetic algorithm

[0071] Based on the training set, the "Grey Wolf + NGBoost" model was trained and constructed; its optimal parameters are: the estimator is 1075, the learning rate is 0.12, the maximum depth is 5, the minimum number of leaf nodes is 8, and the maximum number of splits is 17.

[0072] Step 5: Verification of the "Grey Wolf + NGBoost" model for hot-rolled strip surface defect prediction based on ingot genetics

[0073] Test verification results, such as Figure 2 As shown in Table 1, based on the training set and the test set, the NGBoost model with conventional manual parameter adjustment is used to perform classification prediction, and the results are compared with those obtained by the "Gray Wolf + NGBoost" model. It can be seen that the prediction accuracy of the "Gray Wolf + NGBoost" model is better than that of the NGBoost model, reaching 98.9%.

[0074] Table 1

[0075]

[0076] In the sixth step, for the surface quality image data of hot-rolled strip in the new time period, the selected features are input into the verified defect prediction "Grey Wolf + NGBoost" model, and the predicted classification label is directly output.

[0077] The above specific implementation methods are only descriptions of the technical solutions and model applications of the present invention, and do not limit the scope of the present invention. Without departing from the modeling ideas of the present invention, ordinary technicians in this field can make various modifications and improvements to the technical solutions of the present invention, which should all fall within the scope of protection determined by the claims of the present invention.

Claims

1. The Gray Wolf + NGBoost method for predicting surface defects of hot-rolled strip steel based on genetics of ingot casting is characterized by: The specific steps are as follows: First, based on the surface quality image data of hot-rolled strip and the control parameters of hot rolling process, we select features and labels to construct a dataset in the same time dimension. The dataset includes input samples and an output label value with four classification features; Each sample has 41 input features; The input features include: smelting number, continuous casting number, heat treatment standard, hot flame cleaning standard cleaning degree, continuous casting number, tundish molten steel temperature, liquid phase temperature, maximum molten steel volume in the tundish, minimum molten steel volume in the tundish, slab cut length, average temperature in the tundish, maximum and minimum casting speeds, average heat exchange rate in four directions of east, west, south, and north, maximum and minimum liquid level heights, maximum and minimum upper nozzle flow rates, maximum and minimum upper slide flow rates, average upper slide flow rates, maximum, minimum, and average upper slide pressures, material length, actual furnace tapping temperature, steel modification or not, slab bottom width, slab top width, maximum and minimum upper nozzle pressures, actual slab width and length, off-line time, and transfer time. Then, the dataset is cleaned, smoothed, filtered, and sample homogenized, and divided into training and test sets; The Grey Wolf algorithm was used to collaboratively optimize the parameters of the NGBoost model to form a "Grey Wolf + NGBoost" model for hot-rolled strip surface defect prediction based on ingot genetics, and the model was trained using the training set. The NGBoost model has five parameters, namely, the estimator, the learning rate, the maximum depth, the minimum number of leaf nodes, and the maximum number of splits; Finally, the trained defect prediction "Gray Wolf + NGBoost" model was verified using the test set. For the surface quality image data of hot-rolled strip in the new time period, the selected features are input into the verified defect prediction "Grey Wolf + NGBoost" model, and the predicted classification label is directly output.

2. The slab genetic hot-rolled strip surface defect prediction method based on Gray Wolf + NGBoost according to claim 1, characterized in that: The output label is the actual surface quality test result of the hot-rolled strip inherited from the ingot, including four categories: slag inclusion, warping, cracks or normal.

3. The slab genetic hot-rolled strip surface defect prediction method based on Gray Wolf + NGBoost according to claim 1, characterized in that: The specific training process of the defect prediction "Gray Wolf + NGBoost" model is as follows: First, normalize the data set: X: Feature data input before processing, X * : The data obtained after processing, X min : The minimum value of the feature data input before processing, X max : The maximum value of the feature data input before processing; Then, an initial value is set for each parameter of the NGBoost model as the location parameter of the gray wolf; and the population size, individual dimension, parameter optimization range and maximum number of iterations of the algorithm are initialized. The number of estimators, i.e. the number of trees, is set to 1000; the learning rate is set to 0.1; the maximum depth parameter is set to 5; the minimum value parameter of the leaf node is set to 5; The maximum number of splits parameter is set to 10; Next, the classification accuracy is used as the fitness value of each wolf, and the gray wolf population is divided into four levels: α, β, δ, and ω according to the fitness value; The hunting process is guided by α, β, and δ, and the ω wolf follows these three; Iterate the position of the gray wolf according to the following formula to determine whether the maximum number of iterations has been reached. At the end of the iteration, output the optimal gray wolf position: Indicates the distance between the gray wolf and its prey. Represents the position of the gray wolf at the t+1th iteration, t is the iteration number, and is the coefficient vector, is the position of the prey in the tth iteration, is the position of the gray wolf in the tth iteration, represents the convergence factor; and The modulus is a randomly generated number between [0, 1]. Represents the distance between α, β, δ and other gray wolves, Represents the current position of α, β, and δ; is a random vector, It is the current position of the gray wolf; Finally, the optimal position of the gray wolf is the optimal parameter of the NGBoost model, that is, the defect prediction "Gray Wolf + NGBoost" model training is completed.

4. The slab genetic hot-rolled strip surface defect prediction method based on Gray Wolf + NGBoost according to claim 1, characterized in that: The trained defect prediction "Gray Wolf + NGBoost" model is verified using the test set. Specifically, the test set is input into the trained defect prediction "Gray Wolf + NGBoost" model, and the model's predicted labels are compared with the actual labels of the test set to obtain the model prediction accuracy and give the verification results.

Images