CN112036626B - Online forecasting method for edge linear defects of hot rolled strip steel - Google Patents

Abstract

The invention discloses an online prediction method for linear defects at the edge of hot-rolled strip steel. The main steps include: 1. Collecting production data of hot-rolled strip steel as sample data and preprocessing the sample data; 2. According to GA‑DNN Network establishment of linear defect prediction model of hot-rolled strip edge; 3. Training and verification of the hot-rolled strip edge linear defect prediction model; 4. GA-DNN neural network hot-rolled strip edge linear defect prediction Model online forecasting and analysis. This method has the characteristics of high prediction accuracy, fast response speed, real-time online participation in control, etc., and is of great significance to the surface quality control of hot-rolled strip steel.

Description

An Online Prediction Method for Edge Defects of Hot-rolled Strip Steel

technical field

The invention belongs to the field of surface quality control of hot-rolled steel strips in metallurgical rolling technology, and in particular relates to an intelligent online prediction method for linear defects at the edges of hot-rolled steel strips.

Background technique

With the rapid development of my country's industry, the demand for high-quality hot-rolled strip steel is increasing, and the surface quality is one of the important indicators that affect the quality of hot-rolled strip steel products. Edge linear defect is a kind of surface quality defect of hot-rolled strip steel, which is manifested as warped skin or black line according to the severity. This defect not only seriously affects the yield, but may also affect the production process of the downstream processes of hot rolling (such as the acid rolling unit).

At present, there are some relevant literatures on the control and analysis of linear defects in hot-rolled strip. For example: "Using SSP module to improve the linear defects of hot-rolled strip edges" ("Metallurgical Equipment" 2011, No. 2: 28-30+49), "Formation mechanism and research status of hot-rolled strip edge defects" ( "Henan Metallurgy" 2008, Volume 16, No. 3: 1-4+27) all believe that the side wrinkles of the intermediate billet are the main reason for the formation of edge linear defects, and the severity of the narrow surface wrinkles is directly related to the heating temperature and heating time. Corresponding compensatory solutions are put forward; "Study on the Formation Mechanism of Edge Linear Defects in the Rolling Process" ("Journal of Chongqing University of Technology" 2018, Vol. 32, No. 4: 107-110) thinks The temperature distribution in the width direction of the slab is uneven, and the temperature at the edge and corner of the slab is too low, it is easy to enter the two-phase region in advance, and the deformation resistance becomes smaller, resulting in the deformation of this region exceeding that of other positions. During the subsequent rolling process, the side of the middle slab The metal is continuously flattened to the upper and lower surfaces, eventually leading to the formation of wrinkles and linear defects; The generation of linear defects is directly related to the original defects in the steel area.

According to the current research results, the influencing factors of edge linear defects are complex and changeable, but they can be generally divided into two parts: the steel area and the rolling area. For the rolling area, although the specific mechanism of edge linear defects is not completely the same, it is generally believed that the heating temperature, heating time and rough rolling temperature are the key influencing factors. In addition, although various solutions have been proposed at home and abroad, the optimization of equipment or process parameters is basically based on experience, and the online prediction of edge linear defects has not been carried out so far.

In summary, the present invention takes the heating temperature of the rolling area, the heating time and the exit temperature of the rough rolling as the main influencing factors, and adopts an intelligent method to establish an intelligent online prediction method for the linear defects of the hot-rolled strip edge according to the actual production data. According to the current process parameters, real-time prediction of the occurrence of defects can be carried out, and parameter optimization and adjustment can be carried out on this basis, which is of great significance for the surface quality control of hot-rolled strip steel in the production process.

Contents of the invention

The object of the present invention is to provide an online prediction method for linear defects at the edge of hot-rolled strip steel. According to the deep neural network model and based on the actual production data, the present invention establishes an intelligent online prediction model of the neural network neural network for edge linear defects of the hot-rolled strip steel. The model mainly considers three aspects of heating temperature, heating time and rough rolling exit temperature, including: furnace entry temperature, preheating section temperature, first-addition temperature, second-addition temperature, third-addition temperature, furnace exit temperature, R2 feedback temperature, and preheating temperature. The 13 variable parameters of heating time, first adding time, second adding time, third adding time, soaking time and product steel grade are used to predict the occurrence of linear defects at the edge of hot-rolled strip steel. This method has the characteristics of high prediction accuracy, fast response speed, and real-time online participation in control, etc., which is of great significance to actual production.

A method for online prediction of linear defects at the edge of a hot-rolled strip, comprising the following steps:

S1, collecting hot-rolled strip steel production data as sample data and sample data preprocessing:

S11. Collect the actual production data on site and establish the original data set. The production data include furnace entry temperature T _f , temperature T _p in the preheating section, temperature T ₁ for the first addition, temperature T ₂ for the second addition, T ₃ for the third addition, and Temperature T _o , R2 feedback temperature T _R , preheating time S _R , first-adding time S ₁ , second-adding time S ₂ , third-adding time S ₃ , soaking time S _a , steel type m and its corresponding linear defects Occurrence of situation O, select the number of n groups, and get the original dataset dataset as follows:

{(T _f ,T _p ,T ₁ ,T ₂ ,T ₃ ,T _o ,T _R ,S _R ,S ₁ ,S ₂ ,S ₃ ,S _a ,m)|O} _i (i=1,2 ,3···n);

S12. Preprocessing the original data set dataset, after processing the uneven distribution of linear defect data and eliminating abnormal data, the data set dataset1 with the number of n' groups obtained is:

{(T _f ,T _p ,T ₁ ,T ₂ ,T ₃ ,T _o ,T _R ,S _R ,S ₁ ,S ₂ ,S ₃ ,S _a ,m)|O} _j (j=1,2 ,3···n′);

S13. Perform One-Hot one-hot encoding processing on the non-digital steel type m in the dataset dataset1 to convert it into a digital type;

S14. Divide the dataset dataset1 into training samples and test sets, randomly extract 86% of the data as the dataset dataset2 of the training samples, and use the rest as the test set testset;

S2. Based on the GA-DNN neural network, an online prediction model for linear defects at the edge of the hot-rolled strip is established:

S21. Determine the structure of the DNN neural network, specifically including the following steps:

S211. First, determine the input layer variables of the neural network, including furnace entry temperature T _f , preheating section temperature T _p , primary heating temperature T ₁ , secondary heating temperature T ₂ , third heating temperature T ₃ , furnace output temperature T _o , and R2 feedback temperature T _R , preheating time S _R , first adding time S ₁ , second adding time S ₂ , third adding time S ₃ , soaking time S _a and steel type m, so the number of neuron nodes in the input layer is A ₁ =13 ;

S212, the output of the neural network is the occurrence situation O of strip steel linear defects, so the number of nodes in the output layer of the neural network is A _k =1;

S213. Determine the number of hidden layers and the number of nodes in each layer A ₂ , A ₃ ... A _k-1 ;

S214. Select activation function activation, error loss function loss, optimizer optimizer and matrix function of each layer;

S215, setting the learning rate lr;

S216. Determine the small batch training sample batch and the number of training steps Epoch;

S22, setting GA algorithm parameters, the GA algorithm parameters include initial population size Q, maximum number of iterations N, crossover probability _p1 and mutation probability _p2 ;

S3. Train the online prediction model of linear defects on the edge of hot-rolled strip steel:

S31. Divide the dataset dataset2 of the training samples into a training set and a verification set, randomly select 80% of the dataset dataset2 as the training set trainset of the GA-DNN neural network, and use the remaining 20% as the verification set validationset;

S32, train the GA-DNN neural network, and stop training when the network model reaches the number of training steps;

S33. After the model training is finished, make an error loss map and an accuracy map of the training set trainset and the verification set validation set, and judge whether the average loss error of the network model is less than 0.5, and whether the accuracy can meet the production requirement of 85%;

S34. Using the trained GA-DNN neural network, according to the parameters T _f , T _p , T ₁ , T ₂ , T ₃ , T _o , T _R , S _R , S ₁ in the training set trainset and the test set testset, S ₂ , S ₃ , S _a , m predict the data value of the occurrence of linear defects, compare with the data values of the actual occurrence of linear defects in the training set and the test set, and make a difference. A difference of 0 means that the prediction is correct If the difference is ±1, it means that the prediction is wrong, so as to obtain the actual error distribution map of the training set and the test set, and judge whether it meets the 85% accuracy production requirement;

S35. Determine whether the model meets the accuracy requirements. If the conditions of S33 and S34 are met at the same time, save the GA-DNN network model as the online prediction model M of linear defects at the edge of the hot-rolled strip steel, and print the weight threshold coefficients of each layer of the model. value; if one of S33 and S34 is not satisfied, then return to S2 to adjust the network structure and optimize parameters, and retrain the network;

S4. Intelligent online prediction and analysis of linear defects at the edge of hot-rolled strip:

S41, first load the hot-rolled strip edge linear defect online prediction model M saved in step S35, and embed it in the control system of the hot-rolled strip production site;

S42. During the production process, according to the current rolling steel type, according to the automatic feedback of the system, the furnace entry temperature, the temperature of the preheating section, the first addition temperature, the second addition temperature, the third addition temperature, the exit temperature, the R2 feedback temperature, and the preheating time , the value of one plus time, two plus time, three plus time, and soaking time, real-time forecasting of the occurrence of linear defects in the current strip;

S43. According to the online prediction model M of linear defects at the edge of the hot-rolled strip, analyze the numerical range of linear defects induced by various factors of different steel types, and optimize the process based on this to guide actual production.

Preferably, the preprocessing of the dataset dataset in step S12 specifically includes the following steps:

S121. Using an automatic screening method, the collected original dataset dataset is processed separately according to occurrence of linear defects and non-occurrence of linear defects;

S122. Randomly extract the data with linear defects and the data without linear defects at a ratio of 1:1; combine them in random order;

S123. Eliminate data containing missing items, and obtain a new usable dataset dataset1 after sorting.

Preferably, optimizing and adjusting the GA-DNN neural network structure and parameters in step S35 specifically includes the following steps:

S351. Adjust the number of hidden layers of the neural network and the number of neural network nodes in each hidden layer;

S352. Adjust the activation function activation of the output layer of each neural network layer;

S353. Adjust the network training optimizer optimizer;

S354, adjusting the small batch training sample batch and the number of training steps Epoch;

S355. Introduce regularization to each hidden layer of the GA-DNN neural network;

S356. Perform dropout processing on the input layer of the GA-DNN neural network.

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

The online prediction method for edge linear defects of hot-rolled steel strip proposed by the present invention can perform online prediction on the occurrence of linear defects at the edge of strip steel in real time according to actual production data. This method has high accuracy, strong generalization, fast execution speed, high portability, and meets the requirements of industrial production. It can be well embedded in the production site control system to participate in guiding production. Governance improvement has a good effect, which is of great significance to reduce the tailoring rate, increase the yield rate, and enhance the core competitiveness of products.

Description of drawings

Fig. 1 is the GA-DNN neural network flow diagram of the edge linear defect of hot-rolled steel strip of the present invention;

Figure 2 is a structural diagram of the GA-DNN neural network;

Fig. 3 is the error loss figure of neural network training set and verification set of the embodiment of the present invention;

Fig. 4 is the accuracy figure of neural network training set and verification set of the embodiment of the present invention;

Fig. 5 is the actual error distribution figure of the training set of the embodiment of the present invention;

Fig. 6 is the actual error distribution diagram of the test set of the embodiment of the present invention; and

Fig. 7 is a flow chart of the online prediction method for edge linear defects of hot-rolled strip steel according to the present invention.

Detailed ways

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

As shown in Figure 1, it is a flow chart of a method for intelligent online prediction of linear defects at the edge of a hot-rolled strip according to the present invention, which specifically includes the following steps:

S1. Collect production data of hot-rolled strip steel as sample data and preprocess the sample data:

S11. Collect the actual data on site and establish the original data set, specifically including: furnace entry temperature T _f , preheating section temperature T _p , first heating temperature T ₁ , second heating temperature T ₂ , third heating temperature T ₃ , and furnace output temperature T _o , R2 feedback temperature T _R , preheating time S _R , first-adding time S ₁ , second-adding time S ₂ , third-adding time S ₃ , soaking time S _a and steel type m and the corresponding occurrence of linear defects O, where 0 represents the occurrence of linear defects, and 1 represents the absence of linear defects. There are a total of n=12538 groups, and the original data set dataset is obtained as follows:

{(T _f , T _p , T ₁ , T ₂ , T ₃ , T _o , T _R , S _R , S ₁ , S ₂ , S ₃ , S _a , m)|O} _i i=1, 2, 3...12538.

S12. Preprocessing the original dataset dataset, specifically including: processing of uneven distribution of linear defect data, and processing of abnormal data elimination. The specific process is:

S121. Using an automatic screening method, the collected original dataset dataset is processed separately according to occurrence of linear defects and non-occurrence of linear defects;

S122. Randomly extract the data with linear defects and the data without linear defects at a ratio of 1:1; combine them in random order;

S123. Eliminate data containing missing items, and obtain a dataset dataset1 after sorting, the number of which is n′=4200 groups, namely {(T _f , T _p , T ₁ , T ₂ , T ₃ , T _o , T _R , S _R , S ₁ , S ₂ , S ₃ , S _a , m)|O} _j j=1, 2, 3...4200, and its data structure is shown in Table 1.

S13. Perform One-Hot one-hot encoding processing on the non-numeric type of steel type m in the data set dataset1 to convert it into a digital type.

S14. Divide the dataset dataset1 into training samples and test sets, randomly select 86% of the dataset dataset2 as the model training samples, and round up the hundreds to a total of 3600 groups, and the remaining 600 groups are used as the model test set sample data testset.

Part of the data in Table 1 dataset1

S2. Based on the GA-DNN neural network, an intelligent online prediction model for linear defects at the edge of hot-rolled strip steel is established:

S21. Determine the structure of the DNN neural network, specifically including the following steps:

S211. First, determine the input layer variables of the neural network, including furnace entry temperature T _f , preheating section temperature T _p , primary heating temperature T ₁ , secondary heating temperature T ₂ , third heating temperature T ₃ , furnace output temperature T _o , and R2 feedback temperature T _R , preheating time S _R , first adding time S ₁ , second adding time S ₂ , third adding time S ₃ , soaking time S _a and steel type m, so the number of neuron nodes in the input layer is A ₁ =13 ;

S212. Because there are only 13 input variables, the number of neuron nodes in the input layer is small, and the number of neuron nodes in the hidden layer should not be too many, otherwise overfitting will occur. After multiple model trainings, the results are compared with the model After structural adjustment, it was finally decided to use a k=5-layer neural network, set the number of neuron nodes in the hidden layer A ₂ =200, A ₃ =80, A ₄ =50, and the activation functions of the three hidden layers are all Relu functions ;

S213, the output of the neural network is the steel strip linear defect occurrence situation O, so the output layer node number of the neural network is _Ak =1, and the activation function of the output layer selects the Softmax function;

S214. Because the output layer is the occurrence of strip steel linear defects, there are only two possibilities, so the loss function selects Binary-crossentropy, the optimizer optimizer adopts Adam, the matrix function adopts metrics, and the optimizer learning rate is set to 0.001 to determine the small batch training The sample batch is 32, the number of training Epoch is 200 times, L2 regularization is introduced to the hidden layer of the neural network, the setting value is 0.005, and the dropout function is added to the hidden layer of the neural network, the setting value is 0.02.

S22. Setting GA algorithm parameters, specifically including the following steps:

S221. Set the population size Q to 80 according to the weight threshold scale of the neural network, and encode the randomly generated initial value;

S222. Use the randomly generated initial value as the initial weight and threshold of the neural network to predict the output after training the neural network, and use the error between the predicted output and the predicted output and the expected output as the individual fitness value F, Among them, y is the expected output of the i-th node of the neural network; O _i is the predicted output of the i-th node; k is the coefficient;

S223. Perform crossover and mutation operations on the selected individuals with higher fitness values, and set the crossover probability p ₁ to 0.4 and the mutation probability p ₂ to 0.2;

S224. Calculate the fitness value of the reproduced individual, set the maximum number of iterations N to 100, if the number of iterations is reached, transmit the weight threshold obtained from the final optimization to the neural network, otherwise execute S223.

S3. Training of intelligent online prediction model for linear defects at the edge of hot-rolled strip:

S31. Divide the training sample dataset2 into a training set and a validation set, randomly select 80% of the training samples as the GA-DNN neural network training set sample trainset, a total of 2880 sets of data, and the remaining 720 sets of data as the validation set sample validationset.

S32, train the GA-DNN neural network, and set the number of model training times;

S321. Put the data variables into the furnace temperature T _f , the temperature of the preheating section T _p , the temperature of the first addition T ₁ , the temperature of the second addition T ₂ , the temperature of the third addition T ₃ , the output temperature T _o , the R2 feedback temperature T _R , preheating The time S _R , one plus time S ₁ , two plus time S ₂ , three plus time S ₃ , soaking time S _a and steel type m are input into the input layer of the model. When training starts, neurons in each layer are randomly assigned Give an initial weight coefficient between each input and the neuron, then calculate the sum of the products of all inputs and the weight coefficient, and then pass the activation function calculation, and the output of the previous time is used as the input of the next layer to pass the variable to the next hidden Layer and activation function, and so on, propagate forward to the output layer and output layer activation function to complete the first round of training process;

S322. After the first round of forward propagation is over, the model will take the occurrence of linear defects O as the output of the network, and then compare it with the actual occurrence of defects in the training data, and calculate the training value of this round through the set loss function The loss error between the prediction of strip linear defects and the actual occurrence;

S323. Set the optimizer, perform backpropagation through the error obtained by the loss function, and update the weight coefficient of each layer of neurons through chain derivation to minimize the loss value. After the parameter update is completed, continue to import the training data into the model Carry out the next round of training until the set number of training times is reached, and stop training.

S33. After the model training is finished, the error loss diagram of the training set trainset and the validation set is made as shown in Figure 3, and the accuracy diagram is shown in Figure 4. The results show that the average loss error of the network model is less than 0.5, and the average accuracy of the model can Reach 90%, meet the production requirements.

S34. Using the trained GA-DNN neural network, according to the parameters T _f , T _p , T ₁ , T ₂ , T ₃ , T _o , T _R , S _R , S ₁ on the training set trainset and the test set testset, S ₂ , S ₃ , S _a , m predict the occurrence data value of linear defects, compare with the data values of real linear defect occurrence data values on the training set and test set, and make a difference. A value of 0 means that the prediction is correct. The value of ±1 means that the prediction is wrong, and the actual error distribution diagram of the 3600 sets of data in the training set is shown in Figure 5, and the actual error distribution of the 600 sets of data in the test set is shown in Figure 6. The results show that the accuracy of the model reaches 90%, meet the production requirements.

S35. The model satisfies both the conditions of S33 and S34. Therefore, the GA-DNN neural network model is saved as an intelligent online prediction model M of linear defects at the edge of the hot-rolled strip, and the weight threshold coefficient values of each layer of the model are printed.

S4. On-line prediction and analysis of linear defects on the edge of hot-rolled strip steel:

S41. Load the intelligent online prediction model M of linear defects at the edge of the hot-rolled strip saved in S35, and embed it into the control system at the production site of the hot-rolled strip.

S42. During the production process, according to the current rolling steel type, according to the automatic feedback of the system, the furnace entry temperature, the temperature of the preheating section, the first addition temperature, the second addition temperature, the third addition temperature, the exit temperature, the R2 feedback temperature, and the preheating time , 1st addition time, 2nd addition time, 3rd addition time, soaking time and other parameter values, real-time forecasting of the occurrence of linear defects in the current strip.

S43. According to the intelligent online prediction model M of linear defects at the edge of the hot-rolled strip, analyze the numerical intervals of the linear defects induced by various factors of different steel types. Taking silicon steel 270 as an example, the results are obtained through the intelligent online prediction model M of linear defects at the edge of hot-rolled strip steel. The predicted value ranges of the three heating times and soaking time are shown in Table 2. For this reason, the actual production of silicon steel 270 should be optimized according to this heating process.

Table 2 Prediction of parameter intervals for the generation of linear defects on the edge of silicon steel 270

The above-mentioned embodiments are only descriptions of preferred implementations of the present invention, and are not intended to limit the scope of the present invention. All such modifications and improvements should fall within the scope of protection defined by the claims of the present invention.

Claims (3)

1. A method for online prediction of hot-rolled strip edge linear defects, characterized in that it comprises the following steps: S1, collecting hot-rolled strip steel production data as sample data and sample data preprocessing: S11. Collect the actual production data on site and establish the original data set. The production data include furnace entry temperature T _f , temperature T _p in the preheating section, temperature T ₁ for the first addition, temperature T ₂ for the second addition, T ₃ for the third addition, and Temperature T _o , R2 feedback temperature T _R , preheating time S _R , first-adding time S ₁ , second-adding time S ₂ , third-adding time S ₃ , soaking time S _a , steel type m and its corresponding linear defects Occurrence of situation O, select the number of n groups, and get the original dataset dataset as follows: {(T _f ,T _p ,T ₁ ,T ₂ ,T ₃ ,T _o ,T _R ,S _R ,S ₁ ,S ₂ ,S ₃ ,S _a ,m)|O} _i (i=1,2 ,3···n); S12. Preprocessing the original data set dataset, after processing the uneven distribution of linear defect data and eliminating abnormal data, the data set dataset1 with the number of n' groups obtained is: {(T _f ,T _p ,T ₁ ,T ₂ ,T ₃ ,T _o ,T _R ,S _R ,S ₁ ,S ₂ ,S ₃ ,S _a ,m)|O} _j (j=1,2 ,3···n′); S13. Perform One-Hot one-hot encoding processing on the non-digital steel type m in the dataset dataset1 to convert it into a digital type; S14. Divide the dataset dataset1 into training samples and test sets, randomly extract 86% of the data as the dataset dataset2 of the training samples, and use the rest as the test set testset; S2. Based on the GA-DNN neural network, an online prediction model for linear defects at the edge of the hot-rolled strip is established: S21. Determine the structure of the DNN neural network, specifically including the following steps: S211. First, determine the input layer variables of the neural network, including furnace entry temperature T _f , preheating section temperature T _p , primary heating temperature T ₁ , secondary heating temperature T ₂ , third heating temperature T ₃ , furnace output temperature T _o , and R2 feedback temperature T _R , preheating time S _R , first adding time S ₁ , second adding time S ₂ , third adding time S ₃ , soaking time S _a and steel type m, so the number of neuron nodes in the input layer is A ₁ =13 ; S212, the output of the neural network is the occurrence situation O of strip steel linear defects, so the number of nodes in the output layer of the neural network is A _k =1; S213. Determine the number of hidden layers and the number of nodes in each layer A ₂ , A ₃ ... A _k-1 ; S214. Select activation function activation, error loss function loss, optimizer optimizer and matrix function of each layer; S215, setting the learning rate lr; S216. Determine the small batch training sample batch and the number of training steps Epoch; S22, setting GA algorithm parameters, the GA algorithm parameters include initial population size Q, maximum number of iterations N, crossover probability _p1 and mutation probability _p2 ; S3. Train the online prediction model of linear defects on the edge of hot-rolled strip steel: S31. Divide the dataset dataset2 of the training samples into a training set and a verification set, randomly select 80% of the dataset dataset2 as the training set trainset of the GA-DNN neural network, and use the remaining 20% as the verification set validationset; S32, train the GA-DNN neural network, and stop training when the network model reaches the number of training steps; S33. After the model training is finished, make an error loss map and an accuracy map of the training set trainset and the verification set validation set, and judge whether the average loss error of the network model is less than 0.5, and whether the accuracy can meet the production requirement of 85%; S34. Using the trained GA-DNN neural network, according to the parameters T _f , T _p , T ₁ , T ₂ , T ₃ , T _o , T _R , S _R , S ₁ in the training set trainset and the test set testset, S ₂ , S ₃ , S _a , m predict the data value of the occurrence of linear defects, compare with the data values of the actual occurrence of linear defects in the training set and the test set, and make a difference. A difference of 0 means that the prediction is correct If the difference is ±1, it means that the prediction is wrong, so as to obtain the actual error distribution map of the training set and the test set, and judge whether it meets the 85% accuracy production requirement; S35. Determine whether the model meets the accuracy requirements. If the conditions of S33 and S34 are met at the same time, save the GA-DNN network model as the online prediction model M of linear defects at the edge of the hot-rolled strip steel, and print the weight threshold coefficients of each layer of the model. value; if one of S33 and S34 is not satisfied, then return to S2 to adjust the network structure and optimize parameters, and retrain the network; S4. Intelligent online prediction and analysis of linear defects at the edge of hot-rolled strip: S41, first load the hot-rolled strip edge linear defect online prediction model M saved in step S35, and embed it in the control system of the hot-rolled strip production site; S42. During the production process, according to the current rolling steel type, according to the furnace input temperature, preheating section temperature, first heating temperature, second heating temperature, third heating temperature, output temperature, R2 feedback temperature, and preheating time automatically fed back by the system , the values of first adding time, second adding time, third adding time and soaking time, real-time forecasting of the occurrence of linear defects in the current strip; S43. According to the online prediction model M of linear defects at the edge of the hot-rolled strip, analyze the numerical range of linear defects induced by various factors of different steel types, and optimize the process based on this to guide actual production. 2. The hot-rolled strip edge linear defect online prediction method according to claim 1, characterized in that, in step S12, dataset preprocessing specifically includes the following steps: S121. Using an automatic screening method, the collected original dataset dataset is processed separately according to occurrence of linear defects and non-occurrence of linear defects; S122. Randomly extract the data with linear defects and the data without linear defects at a ratio of 1:1; combine the two out of order; S123. Eliminate data containing missing items, and obtain a new usable dataset dataset1 after sorting. 3. The method for online prediction of hot-rolled strip edge linear defects according to claim 1, wherein in step S35, optimizing and adjusting GA-DNN neural network structure and parameters specifically includes the following steps: S351. Adjust the number of hidden layers of the neural network and the number of neural network nodes in each hidden layer; S352. Adjust the activation function activation of the output layer of each neural network layer; S353. Adjust the network training optimizer optimizer; S354, adjusting the small batch training sample batch and the number of training steps Epoch; S355. Introduce regularization to each hidden layer of the GA-DNN neural network; S356. Perform dropout processing on the input layer of the GA-DNN neural network.