CN112926622A

CN112926622A - Crystallizer breakout prediction method for generating countermeasure network based on feature vector and SWGAN-GP

Info

Publication number: CN112926622A
Application number: CN202110087095.5A
Authority: CN
Inventors: 王旭东; 王砚宇; 段海洋; 姚曼
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2021-01-22
Filing date: 2021-01-22
Publication date: 2021-06-08
Anticipated expiration: 2041-01-22
Also published as: CN112926622B

Abstract

The invention provides a mold breakout prediction method based on feature vector and SWGAN-GP generation confrontation network, which belongs to the technical field of continuous casting detection of iron and steel metallurgy. The invention constructs a feature vector including static and dynamic features of the bonding area by visualizing the thermal image of the temperature rate of the mold copper plate, and classifies the feature vector through the discrimination model of the SWGAN-GP generation adversarial network, thereby realizing the detection and detection of the mold breakout. forecast. Based on the SWGAN-GP model, the present invention performs real-time detection and prediction on mold breakout, which can significantly reduce the false alarm rate and effectively improve the prediction accuracy on the premise of ensuring that all the breakouts are reported.

Description

Crystallizer breakout prediction method for generating countermeasure network based on feature vector and SWGAN-GP

Technical Field

The invention belongs to the technical field of ferrous metallurgy continuous casting detection, and relates to a crystallizer bleed-out forecasting method for generating a countermeasure network based on a feature vector and SWGAN-GP.

Background

In the continuous casting process, the non-uniformly solidified primary blank shell in the crystallizer cannot bear the dual functions of the static pressure of molten steel and the withdrawal force, and the weak part is easy to break so as to form steel leakage. Breakout 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 and influences the yield and the product quality of a casting machine. With the continuous development and progress of the continuous casting technology, the occurrence probability of bleed-out can be effectively reduced by standardizing operation and maintaining the good running state of equipment. The breakout is closely related to a plurality of factors such as covering slag, drawing speed, liquid level fluctuation, heat flux density and the like, and although metallurgical workers and scholars at home and abroad carry out extensive research on the formation reason, the breakout is difficult to be thoroughly avoided. Therefore, the reduction and prevention of breakout accidents are always the key points of attention of metallurgical workers at home and abroad, and the detection of the breakout of the crystallizer is the core of the abnormal prediction in the continuous casting process, so that the method has important significance.

The invention patent 200710093907.7 discloses a continuous casting breakout prediction method, which predicts the occurrence of bonding breakout based on logic judgment and according to the temperature change condition of a thermocouple of a crystallizer. The method mainly comprises the following steps: capturing typical temperature characteristics, determining breakout probability and controlling casting speed. The method effectively solves the technical problems that interference factors in the existing breakout prediction method are not considered comprehensively, and small-range adhesion and slag entrapment cannot be reported timely. However, the breakout prediction model based on logic judgment has high dependence on equipment parameters, process conditions and physical parameters, and needs frequent adjustment of threshold values and parameters, resulting in poor robustness of the prediction algorithm.

The invention patent 201010207115.X discloses a continuous casting breakout prediction method, which uses a genetic algorithm to initialize a time sequence neural network breakout prediction model of a single thermocouple. The method mainly comprises the following steps: thermocouple temperature data are collected on line, data are preprocessed, and a model is forecasted to forecast breakout. However, the breakout prediction model based on the single neural network has strict limitations on the quality and quantity of training samples, and the process of making the samples is cumbersome, and the practicability and the applicability are low.

In view of the defects of the existing breakout prediction method and the complexity of the prediction algorithm and the complexity of the sample making process, the invention provides a method for detecting and predicting breakout of a crystallizer in real time by extracting visual characteristic vectors of a bonding area based on a temperature rate thermograph of a copper plate of the crystallizer, generating a countermeasure network model by combining SWGAN-GP, and training breakout and non-breakout samples of the crystallizer.

Disclosure of Invention

The invention aims to provide a crystallizer bleed-out forecasting method for generating a countermeasure network based on a characteristic vector and SWGAN-GP, which can timely and accurately detect and forecast the bonded bleed-out and provides a reliable means for monitoring the abnormity of the continuous casting process.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a crystallizer breakout prediction method for generating a countermeasure network based on a feature vector and SWGAN-GP is characterized in that visual feature vectors are extracted from an abnormal region of a crystallizer copper plate temperature rate, and the countermeasure network is generated by the SWGAN-GP to classify the feature vectors, so that the crystallizer breakout is detected and predicted, and the method specifically comprises the following steps:

first, extracting the characteristics of the abnormal area of the temperature rate of the copper plate of the crystallizer

(1) 3 rows of 19 rows of thermocouples are arranged on the wide-surface copper plates of the inner arc and the outer arc of the crystallizer, and 3 rows of 1 row of thermocouples are arranged on the narrow-surface copper plates at the left side and the right side. And detecting the temperature of the thermocouple of the crystallizer copper plate on line, and calculating the temperature value of the crystallizer copper plate at the position of the non-thermocouple measuring point by an interpolation algorithm.

(2) And calculating the temperature change rate of each point of the copper plate through an interframe difference algorithm, and mapping the temperature rate of the copper plate to a two-dimensional plane by using computer graphics and OpenGL technology to obtain a two-dimensional temperature rate thermal image corresponding to the temperature of the copper plate.

(3) After counting and summarizing temperature rate data of a plurality of breakout samples, setting a temperature rate threshold value as T_zUsing threshold segmentationThe method comprises the steps of removing a normal temperature fluctuation area with the temperature rate smaller than a threshold value from a two-dimensional temperature rate thermal image, and performing connectivity search on temperature rate abnormal points by using a run-length recursive algorithm to obtain a temperature rate abnormal area.

(4) Extracting the height H, width W, area S and transverse moving speed V of the abnormal temperature speed region_yLongitudinal moving speed V_xAnd (6) visualizing the features.

Second step, abnormal region feature vector construction and processing

(1) Combining the abnormal region features extracted in the first step into a feature vector X_BSimultaneously combining the normal working condition region features into a feature vector X_N：

X_B＝[H_B,W_B,S_B,V_Bx,V_By]

X_N＝[H_N,W_N,S_N,V_Nx,V_Ny]

(2) Pulling speed V under combined abnormal and normal working conditions_cAnd V_c', for feature vector X_BAnd X_NPerforming continuous processing to construct continuous feature vector Z_BAnd Z_N：

(3) For continuous feature vector Z_BAnd Z_NAnd (3) carrying out normalization treatment:

in the formula, Z_min、Z_maxRespectively representing the minimum and maximum values of the continuous-type eigenvector Z, F_iNormalized pair of ith dimension feature representing continuous feature vector ZThe corresponding values.

(4) Respectively obtaining m abnormal area bleed-out feature vector samples F according to the feature vector construction and processing modes of the steps (1), (2) and (3)_BAnd n normal working condition area non-breakout eigenvector samples F_NAnd establishing a sample set Q:

Q＝{(F_B1,1),(F_B2,1),…,(F_Bm,1),(F_N1,0),(F_N2,0),…,(F_Nn,0)}

in the formula, m and n respectively represent the number of the breakout sample and the non-breakout sample. 1 and 0 represent class labels for the breakout and non-breakout specimens, respectively, as (1, 0, 0) and (0, 1, 0) in the form of one-hot codes.

Thirdly, constructing an SWGAN-GP generation confrontation network model

(1) Constructing a generative model G comprising 1 noise input layer and 2 full-connection layer neural networks, wherein the concrete structure of the generative model G sequentially comprises the following steps: noise input layer → first fully connected layer → second fully connected layer. The input noise generates pseudo samples with the same dimension as the original feature vector after G.

(2) Constructing a discrimination model D containing 1 sample input layer and 3 full-connection layer neural networks, wherein the specific structure of the discrimination model D is as follows in sequence: sample input layer → first fully connected layer → parallel branch second fully connected layer → parallel branch third fully connected layer. And finally D, judging the authenticity of the input sample and classifying the sample.

(3) And combining the generation model G with the discrimination model D to construct a SWGAN-GP generation confrontation network model.

Fourthly, training SWGAN-GP to generate a confrontation network model

(1) And (3) training the discrimination model D constructed in the step (2) in the third step. And freezing the generated model when the discriminant model is trained, namely setting the parameters of the generated model to be not updatable.

1.1) randomly obtaining a breakout sample F from a sample set Q_BAnd non-breakout specimen F_NEach train _ samples has a single sample denoted x, and x belongs to the real sample set P_rI.e. x to P_r(ii) a Simultaneous acquisition of Gaussian distributed noiseSet of samples z, i.e. z-P_z。

1.2) inputting the noise sample z into the generation model G constructed in the step (1) in the third step to generate a pseudo sample

The label of (a) is represented in a one-hot coded form as (0, 0, 1):

in the formula, a pseudo sample

Belonging to a set of dummy samples P_GI.e. by

1.3) obtaining P_rSample x and P in (1)_GSample of (1)

And interpolation is carried out to obtain a new sample

In which ε follows a uniform distribution over [0,1 ].

1.4) calculating the Total loss L of the discriminant model D_D：

In the formula, m is the batch sample number batch _ size, i is 1, and 2 … m is the sample number index.

D (x) are respectively

x corresponds to the output value of the discriminator D. λ represents a gradient penalty coefficient.

Is a gradient penalty term. C denotes the number of sample classes, j is 1, and 2 … C is the class index. y is_jFor the true label corresponding to the jth category, f_j(x) In order for the arbiter to predict the value for that sample,

representing the loss of cross-entropy part of the multivariate classification.

(2) Training the generative model G constructed in the step (1) in the third step. And (4) freezing the discrimination model when the generation model is trained, namely setting the parameters of the discrimination model to be not updatable.

2.1) obtaining a set of noise samples z' -P obeying Gaussian distribution_z′。

2.2) inputting the noise sample z' into the generation model G constructed in the step (1) in the third step to generate a pseudo sample

Then will be

Calculating the model loss L of the discriminant model D_G：

(3) Training discriminant model n in each round of training_criticThen, generating a model n_genRepeating the training to judge model D and generate model G, and observing L_GAnd L_DAs a function of the number of training rounds, up to L_GAnd L_DThe loss curve gradually flattens and fluctuates steadily. At this time, the generative model G and the discriminant model D can be determinedNash balance is achieved, and the training is finished.

Fifthly, detecting and forecasting bleed-out on line based on SWGAN-GP model

(1) Extracting typical visual characteristics of the crystallizer copper plate temperature rate abnormal area in real time, preprocessing the typical visual characteristics, and constructing to obtain an abnormal area characteristic vector F_fv；

(2) Abnormal region feature vector F_fvInputting the data into a discrimination model D of the SWGAN-GP generated countermeasure network to obtain a predicted value y of the model:

y＝D(F_fv)

(3) and forecasting the crystallizer bleed-out according to the output result y of the discrimination model.

Expressing y as (y) in a form of one-hot coding₁，y₂，y₃): if y₁＝max(y₁,y₂,y₃) If the data label corresponding to y is (1, 0, 0), the alarm is given and the casting machine pulling speed is rapidly reduced; if y₂＝max(y₁,y₂,y₃) And if the detected steel leakage is normal, continuing to detect and forecast the steel leakage at the next moment corresponding to the data label with the value of y being (0, 1, 0). Feature vector x of abnormal region_fvAll derived from real samples detected on-line rather than pseudo-samples generated by generative model G

So that y cannot occur₃The case of the maximum value.

The method for forecasting the breakout is suitable for forecasting the breakout of continuous casting billets such as plate blanks, square blanks, round blanks, special blanks and the like.

The invention has the beneficial effects that: the method constructs a characteristic vector containing static and dynamic characteristics of a bonding area through a visualized thermograph of the temperature rate of the crystallizer copper plate, and classifies the characteristic vector through a discrimination model of a countermeasure network generated by SWGAN-GP so as to realize detection and forecast of the breakout of the crystallizer. The method is used for detecting and forecasting the crystallizer bleed-out in real time based on the SWGAN-GP model, so that the false alarm rate can be obviously reduced and the forecasting accuracy rate can be effectively improved on the premise of ensuring that the bleed-out is completely reported.

Drawings

FIG. 1 is a flow of a crystallizer breakout prediction method.

Fig. 2 is a schematic diagram of arrangement of a thermocouple of a copper plate of the crystallizer.

Fig. 3 is a temperature rate abnormal region visualization characteristic diagram. FIG. 3(a) is a diagram of initial formation of bonding; FIG. 3(b) is a transverse propagation diagram of the bond region; FIG. 3(c) is a longitudinal propagation diagram of the bond region; FIG. 3(d) is a V-shaped characteristic diagram of the bonded breakout.

FIG. 4 is a normal operating condition region visualization characteristic diagram. FIG. 4(a) is a diagram of initial formation of bonding; FIG. 4(b) is a longitudinal contraction of the bonded area; FIG. 4(c) is an in-situ expansion diagram of the bonding region; FIG. 4(d) is a transverse contraction diagram of the bonded area.

FIG. 5 is SWGAN-GP generative model G.

FIG. 6 is a SWGAN-GP discriminant model D.

FIG. 7 is a visual characteristic diagram of an online detection temperature rate abnormal region. FIG. 7(a) is a V-shaped characteristic diagram of the bonded breakout; FIG. 7(b) is a normal condition bonding area diagram.

In the figure: 1, a thermocouple; 2 outer arc wide copper plate; 3, a left narrow-face copper plate; 4, a right narrow-face copper plate; 5 inner arc wide copper plate.

Detailed Description

The invention will be further elucidated by means of specific embodiments, in conjunction with the drawing

Fig. 1 shows a flow chart of a method for predicting breakout of a crystallizer. Firstly, extracting visual characteristics of a crystallizer copper plate temperature rate abnormal area and preprocessing the visual characteristics to construct and obtain a five-dimensional characteristic vector; secondly, constructing and training a SWGAN-GP model; and finally, classifying the feature vectors and forecasting the breakout through a discrimination model of SWGAN-GP.

First step, visualization of temperature rate of crystallizer copper plate and extraction of abnormal area characteristics

(1) Fig. 2 shows the distribution diagram of the copper plate of the mold and its thermocouple. The crystallizer is formed by combining four copper plates, the total height is 900mm, and the effective height during casting is 800 mm. 3 rows of 19 rows of thermocouples 1 are arranged on the inner and outer arc wide-

surface copper plates

5 and 2 of the crystallizer, 3 rows of 1 row of thermocouples are arranged on the left and right narrow-surface copper plates 3 and 4, and the total number of the thermocouples is 120. The distance between the upper openings of the first row of thermocouple data crystallizers is 210mm, the distance between the first row of thermocouple data crystallizers and the second row of thermocouple data crystallizers is 115mm, the distance between the second row of thermocouple data crystallizers and the third row of thermocouple data crystallizers is 120mm, and the distance between two adjacent rows of thermocouples is 150 mm. And (3) detecting the temperature of all the thermocouples (1) of the crystallizer copper plate on line, and calculating the temperature value of the crystallizer copper plate at the position of the non-thermocouple measuring point by an interpolation algorithm.

(3) Setting a temperature rate threshold value to be 0.3 ℃/s, removing a normal temperature fluctuation area with the temperature rate smaller than the threshold value from the two-dimensional temperature rate thermal image by using a threshold segmentation algorithm, and performing connectivity search on temperature rate abnormal points by using a run recursion algorithm to obtain a temperature rate abnormal area.

(4) Fig. 3 is a graph showing the visualization characteristic of the abnormal temperature rate region. The simulation interval from the first row to the third row along the casting direction is distributed to 100 pixel points; the horizontal simulation interval is from the first row to the nineteenth row of thermocouples distributed to 300 pixel points. T is₁～T₄Representing 4 moments in time corresponding to the bond region from initial formation to the appearance of a distinct "V" shaped feature, each moment being 3s apart. Extraction of T₄Height H of time anomaly region_B20.45cm wide W_B56.11cm, shaded area S_B＝607.08cm²And calculating to obtain the transverse movement velocity V according to the change of the barycentric coordinates of the abnormal region along with the time_ByLongitudinal moving speed V of-0.06 m/min_Bx＝0.25m/min。

Fig. 4 is a normal condition area visualization characteristic diagram. The visual characteristics of the region, height H, can be obtained by the same method_N5.64cm, width W_N21.5cm, shaded area S_N＝92.36cm²Transverse moving velocity V_Ny0m/min, longitudinal moving speed V_Nx＝-0.39m/min。

Second step, abnormal region feature vector construction and processing

(1) Combining the abnormal and normal working condition region characteristics extracted in the first step into a characteristic vector X_BAnd X_N：

X_B＝[H_B,W_B,S_B,V_Bx,V_By]＝[20.45,56.11,607.08,0.25,-0.06]

X_N＝[H_N,W_N,S_N,V_Nx,V_Ny]＝[5.64,21.5,92.36,-0.39,0]

(2) Pulling speed V under combined abnormal and normal working conditions_c0.9m/min and V_c' 0.65m/min, for feature vector X_BAnd X_NPerforming continuous processing to construct continuous feature vector Z_BAnd Z_N：

in the formula, Z_min、Z_maxRespectively representing the minimum and maximum values of the continuous-type eigenvector Z, F_iAnd representing the corresponding numerical value after the ith dimension feature normalization of the continuous feature vector Z. Normalized feature vector F_B＝[0.84,0.71,0.86,0.73,0.13]，F_N＝[0.19,0.18,0.02,0,0.25]。

Respectively obtaining 50 bleed-out eigenvector samples F according to the above eigenvector construction and processing mode_BAnd 50 non-breakout eigenvector samples F_NAnd establishing a sample set Q:

Q＝{(F_B1,1),(F_B2,1),…,(F_B50,1),(F_N1,0),(F_N2,0),…,(F_N50,0)}

in the formula, 1 and 0 represent class labels of the breakout sample and the non-breakout sample, respectively, and are expressed as (1, 0, 0) and (0, 1, 0) in the form of one-hot codes.

Thirdly, constructing an SWGAN-GP generation confrontation network model

(1) As shown in fig. 5, a generative model including a noise input layer and a fully connected layer neural network is constructed for generating a pseudo sample, denoted as generative model G.

(2) As shown in fig. 6, a discrimination model including a sample input layer and a full connection layer neural network is constructed, and is used for discriminating an authentic sample and classifying the sample, and the discrimination model is recorded as a discrimination model D.

Fourthly, training SWGAN-GP to generate a confrontation network model

(1) And (5) training a discrimination model. And freezing the generated model when the discriminant model is trained, namely setting the parameters of the generated model to be not updatable.

1.1) randomly obtaining a breakout sample F from a sample set Q_BAnd non-breakout specimen F_NEach 30, a single sample is denoted x, and x belongs to the set of true samples P_rI.e. x to P_r(ii) a Simultaneously obtaining a set of noise samples z, i.e. z-P, which obey a Gaussian distribution_z。

1.2) input of noise samples z into the generative model G to generate pseudo samples

The label of (a) is represented in a one-hot coded form as (0, 0, 1):

in the formula, a pseudo sample

Belonging to a set of dummy samples P_GI.e. by

1.3) obtaining P_rSample x and P in (1)_GSample of (1)

And interpolation is carried out to obtain a new sample set

In which ε follows a uniform distribution over [0,1 ].

1.4) calculating the Total loss L of the discriminant model D_D：

In the formula, m is the batch sample number 32, i is 1, and 2 … m is the sample number index.

D (x) are respectively

x corresponds to the output value of the discriminator D. The gradient penalty factor lambda is 10,

is a gradient penalty term. The number of sample types C is 3, and represents the breakout sample F_BNon-bleed-out sample F_NAnd a dummy sample

j is 1,2, 3 is a category index. y is_jFor the true label corresponding to the jth category, f_j(x) In order for the arbiter to predict the value for that sample,

representing the loss of cross-entropy part of the multivariate classification.

(2) And training the generated model. And (4) freezing the discrimination model when the generation model is trained, namely setting the parameters of the discrimination model to be not updatable.

2.2) inputting the noise sample z' into the generative model G to generate a pseudo sample

Then will be

Calculating the model loss L of the discriminant model D_G：

Training the discriminant model 5 times and generating the model 1 time in each round of training, repeating the training of discriminant model D and generating model G, and observing L_GAnd L_DWhen the number of training rounds reaches 20000 times as the number of training rounds changes, L is observed_GAnd L_DThe loss curve gradually flattens and fluctuates steadily. At this time, it can be judged that nash balance is achieved between the generated model G and the discriminant model D, and the training is completed.

Fifthly, detecting and forecasting bleed-out on line based on SWGAN-GP model

(1) Extracting typical visual characteristics of the crystallizer copper plate temperature rate abnormal area in real time, as shown in fig. 7, preprocessing the typical visual characteristics, and constructing to obtain an abnormal area characteristic vector:

F_fv1＝[0.78,0.30,0.46,0.84,0.73]；F_fv2＝[0.06,0.23,0.08,0.63,0.56]

(2) will be abnormalRegion feature vector F_fv1And F_fv2Inputting the predicted value y of the model into a discrimination model D of the SWGAN-GP generated countermeasure network₁And y₂：

y₁＝D(F_fv1)＝D([0.78,0.30,0.46,0.84,0.73])＝(0.81,0.19,0)

y₂＝D(F_fv2)＝D([0.06,0.23,0.08,0.63,0.56])＝(0.12,0.88,0)

(3) And forecasting the crystallizer bleed-out according to the output result y of the discrimination model. y is₁If the data label corresponding to y is (1, 0, 0) is breakout, (0.81,0.19,0) gives an alarm and rapidly reduces the casting machine pulling speed; y is₂And if the value is equal to (0.12,0.88 and 0), the data label corresponding to (0, 1 and 0) is a normal working condition, and the steel leakage detection and prediction at the next moment are continued. Feature vector x of abnormal region_fvAll derived from real samples detected on-line rather than pseudo-samples generated by generative model G

So that y cannot occur₃The case of the maximum value.

The above-mentioned embodiments only express the embodiments of the present invention, but not should be understood as the limitation of the scope of the invention patent, it should be noted that, for those skilled in the art, many variations and modifications can be made without departing from the concept of the present invention, and these all fall into the protection scope of the present invention.

Claims

1. A mold breakout prediction method based on eigenvectors and SWGAN-GP generative adversarial network, is characterized in that, the method extracts visual feature vector by abnormal area of mold copper plate temperature rate, and utilizes SWGAN-GP to generate adversarial network The classification of feature vectors to detect and predict mold breakout includes the following steps:

The first step, the feature extraction of the abnormal area of the temperature rate of the copper plate of the mold

(1) Arrange thermocouples on the inner and outer arc wide-surface copper plates and the left and right narrow-surface copper plates; detect the thermocouple temperature of the mold copper plate online, and calculate the non-thermocouple measuring point position of the mold copper plate through interpolation algorithm temperature value;

(2) Calculate the temperature change rate of each point of the copper plate by the difference algorithm between frames, and map the temperature rate of the copper plate to a two-dimensional plane to obtain a two-dimensional temperature rate thermal image corresponding to the temperature of the copper plate;

(3) After statistics and summarizing the temperature rate data of multiple breakout samples, set the temperature rate threshold as T _z , use the threshold segmentation algorithm to remove the normal temperature fluctuation area with the temperature rate less than the threshold from the two-dimensional temperature rate thermal image, and use the run recursion The algorithm searches the connectivity of the abnormal temperature rate points to obtain the abnormal temperature rate areas;

(4) Extracting the visualization features of height H, width W, area S, lateral movement velocity V _y , and longitudinal movement velocity V _x of the abnormal temperature rate area;

The second step, abnormal area feature vector construction and processing

(1) Combine the abnormal area features extracted in the first step into a feature vector X _B , and at the same time combine the normal operating conditions region features into a feature vector X _N :

X _B =[H _B ,W _B ,S _B ,V _Bx ,V _By ]

X _N =[H _N ,W _N ,S _N ,V _Nx ,V _Ny ]

(2) Combine the pulling speeds V _c and V _c ' under abnormal and normal working conditions, perform continuous processing on the eigenvectors X _B and X _N , and construct continuous eigenvectors Z _B and Z _N :

(3) Normalize the continuous eigenvectors Z _B and Z _N :

In the formula, Z _min and Z _max represent the minimum and maximum values of the continuous eigenvector Z, respectively, and F _i represents the corresponding value of the i-th dimension feature of the continuous eigenvector Z after normalization;

(4) According to the above (1)(2)(3) feature vector construction and processing methods, respectively obtain m cases of abnormal area breakout feature vector samples _FB and _n cases of non-breakout feature vector samples in normal working conditions area, respectively, Form a sample set Q:

Q={(F _B1 ,1),(F _B2 ,1),…,(F _Bm ,1),(F _N1 ,0),(F _N2 ,0),…,(F _Nn ,0)}

In the formula, m and n represent the number of breakout samples and non-breakout samples, respectively; 1 and 0 represent the category labels of breakout samples and non-breakout samples, respectively, which are expressed in the form of one-hot encoding as (1, 0, 0 ) and (0, 1, 0);

The third step is to build a SWGAN-GP generative adversarial network model

(1) Construct a generative model G that includes one noise input layer and two fully connected layer neural networks, and the input noise passes through G to generate pseudo samples with the same dimension as the original feature vector;

(2) Construct a discriminant model D including a sample input layer and 3 fully connected layer neural networks, and use D to discriminate the authenticity of the input samples and classify the samples;

(3) Combining the generative model G and the discriminative model D to construct a SWGAN-GP generative adversarial network model;

The fourth step, training the SWGAN-GP generative adversarial network model

(1) Train the discriminant model D constructed in the third step (2); freeze the generative model when training the discriminant model, that is, the parameters of the generative model are set to be non-updateable;

1.1) Randomly obtain each _{train_samples} of breakout samples F _B and non-breakout samples F _N from the sample set Q, a single sample is denoted as x, and x belongs to the real sample set Pr , namely x ~ _Pr ; The distributed noise sample set z, that is, z～P _z ;

1.2) Input the noise sample z into the generative model G constructed in the third step (1) to generate pseudo samples

The labels are represented as (0, 0, 1) in one-hot encoded form:

In the formula, the pseudo sample

belongs to the false sample set P _G , namely

1.3) Get the sample x in P _r and the sample in P _G

and interpolate to get new samples

In the formula, ε obeys a uniform distribution on [0,1];

1.4) Calculate the total loss LD of the discriminant model _D :

In the formula, m is the number of batch samples batch_size, i=1, 2...m is the index of the number of samples;

D(x) are respectively

The output value of the discriminator D corresponding to x; λ represents the gradient penalty coefficient;

is the gradient penalty term; C represents the number of sample categories, j=1,2...C is the category index; y _j is the real label corresponding to the jth category, and f _j (x) is the discriminator's predicted value for the sample,

represents the loss of the multivariate classification cross-entropy part;

(2) Train the generative model G constructed in the third step (1); freeze the discriminant model when training the generative model, that is, the parameters of the discriminant model are set to be non-updateable;

2.1) Obtain a set of noise samples z′～P _z′ obeying a Gaussian distribution;

2.2) Input the noise sample z' into the generative model G constructed in the third step (1) to generate pseudo samples

again

Input into the discriminant model D to calculate the generative model loss L _G :

(3) In each round of training, train the discriminant model n _critic times and the generative model n _gen times, repeat the training of the discriminant model _D and the generative model G, and observe the changes of LG and LD with the number of training rounds until _LG and _L The _D loss curve gradually flattens and fluctuates steadily; at this time, it can be judged that the Nash equilibrium has been reached between the generative model G and the discriminant model D, and the training is over;

The fifth step, online detection and prediction of steel breakout based on SWGAN-GP model

(1) Real-time extraction of typical visualization features of the abnormal region of the temperature rate of the copper plate of the mold, and preprocessing it, and constructing the abnormal region feature vector F _fv ;

(2) Input the abnormal area feature vector F _fv into the discriminant model D of the SWGAN-GP generative adversarial network, and obtain the predicted value y of the model:

y=D(F _fv )

(3) Predict mold breakout according to the output result y of the discriminant model;

Represent y as (y ₁ , y ₂ , y ₃ ) in the form of one-hot encoding: if y ₁ =max(y ₁ , y ₂ , y ₃ ), then the data corresponding to y=(1, 0, 0) The label is for breakout, which will issue an alarm and rapidly reduce the casting speed; if y ₂ =max(y ₁ , y ₂ , y ₃ ), the data label corresponding to y=(0, 1, 0) is the normal operation. Continue with the breakout detection and prediction at the next moment; because the characteristic vectors x _fv of the abnormal area are all derived from the real samples detected online rather than the fake samples generated by the generation model G

Therefore, there cannot be a situation where y ₃ is the maximum value.

2 . The mold breakout forecasting method according to claim 1 , wherein the breakout forecasting method is suitable for breakout forecasting of slabs, square billets, round billets, special-shaped billets or other continuous casting slabs. 3 .

3. The mold breakout prediction method according to claim 1, wherein the specific structure of the generated model G in the third step (1) is in turn: noise input layer→first fully connected layer→ The second fully connected layer.

4. The method for predicting mold breakout according to claim 1, wherein the specific structure of the discriminant model D in the third step (2) is in turn: sample input layer→first fully connected layer→ Parallel branch to the second fully connected layer → Parallel branch to the third fully connected layer.