A method for predicting mold breakout based on eigenvectors and neural networks

Abstract

The present invention provides a method for predicting mold breakout based on feature vectors and neural networks, belonging to the technical field of continuous casting detection in iron and steel metallurgy. The mold breakout prediction method extracts the shape features and extended features of the abnormal temperature rate region of the mold copper plate, and uses a neural network to classify the constructed feature vectors, so as to detect and predict mold breakout. It is applicable to the breakout prediction of continuous casting billets such as slab, square billet, round billet, and special-shaped billet. The present invention obtains the measured temperature of the thermocouple online, visually characterizes the abnormal temperature rate region, and then extracts and constructs the feature vectors of the abnormal region. Through the neural network model, the feature vector sample library is learned and trained, and finally the online prediction of mold breakout is realized; this method is based on the neural network model to detect and predict mold breakout in real time, and can reduce the false alarm rate while ensuring a 100% reporting rate of bonded breakout, effectively improving the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of continuous casting detection in iron and steel metallurgy, and relates to a method for predicting breakout of a mold based on eigenvectors and neural networks. Background Art

[0002] Continuous casting is a process in which molten steel at high temperature is cooled under forced cooling to form a casting blank. Although China has made great progress in high-efficiency continuous casting technology, the excessive steel throughput has led to a continuous increase in the thermal load of the continuous casting mold. A series of casting blank defects and abnormal problems will also occur during the casting process under high load. Among them, breakout is the most destructive production accident in continuous casting. Breakout is a phenomenon in which the primary shell ruptures and high-temperature molten steel flows out. It will not only seriously hinder the smooth production process of continuous casting, but also damage the copper plates of the mold and the equipment in the secondary cooling zone, causing huge economic losses and serious safety hazards. Therefore, developing an accurate and practical breakout prediction model is of great significance for ensuring the quality of casting blanks and promoting the smooth production of continuous casting.

[0003] Patent CN202010349356.1 discloses a method for predicting breakout in continuous casting. This method provides a method for predicting breakout of a mold based on image processing. It mainly includes: obtaining the temperature, temperature rise rate, and two-dimensional plane coordinates of the temperature measurement points arranged on the mold to be measured; obtaining a temperature section and a temperature rise rate section with preset standard temperature and preset standard temperature rise rate; and performing an early warning prompt according to the appearance of the temperature section and the temperature rise rate section and the change of the temperature section parameters according to a preset early warning rule. Although the above method can solve the bottleneck in the false alarm rate and missed alarm rate of breakout, the logical judgment parameters and early warning rules in the model are relatively complex, and the alarm threshold needs to be frequently adjusted and set under different working conditions, and the transferability and robustness of the model are relatively low.

[0004] Patent CN201110293406.X discloses a method for predicting breakout in continuous casting. This method relates to a method for predicting breakout of a slab continuous casting mold based on drawing resistance. The drawing resistance is calculated through the production data on the continuous casting site, and the resistance characteristics during normal continuous casting and abnormal conditions are extracted to construct an eigenvector; the eigenvector and a support vector machine are used to train an identification model of the resistance signal to achieve breakout prediction. Although this method can solve the problem of time lag in the breakout prediction method, the change of the drawing resistance is easily masked by other process factors, and the friction condition in a local area of the mold cannot be accurately detected. Thus, the application of this method is limited.

[0005] In view of the limitations of the above-mentioned breakout prediction model, and considering the practicality of the prediction method and the complexity of model parameter adjustment, the present invention proposes to obtain a copper plate temperature thermal image through temperature mapping based on the measured temperature of the thermocouples on the copper plate of the mold, extract the shape features and expansion features characterizing the breakout bonding area, construct a feature vector and preprocess it, and then use a neural network model for training, and finally online predict the breakout of the mold through the trained neural network. Summary of the Invention

[0006] The object of the present invention is to propose a method for predicting the breakout of the mold based on feature vectors and neural networks, which can detect and predict the bonding breakout in real time and accurately, and provide technical support for the abnormal monitoring of the continuous casting process.

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

[0008] A method for predicting the breakout of the mold based on feature vectors and neural networks, which extracts the shape features and expansion features of the abnormal temperature rate area of the copper plate of the mold, and uses a neural network to classify the constructed feature vector, so as to detect and predict the breakout of the mold. The specific steps are as follows:

[0009] The first step: Visual characterization of the abnormal temperature rate area of the copper plate of the mold

[0010] (1) Arrange a plurality of thermocouples on the wide-face copper plates of the inner and outer arcs of the mold and the narrow-face copper plates on the left and right sides according to a certain layout. Online detect the temperature of each thermocouple, and calculate the copper plate temperature value at the non-thermocouple measurement point through the interpolation algorithm.

[0011] (2) Through computer graphics and OpenGL technology, establish the mapping relationship between the copper plate temperature and the two-dimensional plane, and obtain a two-dimensional temperature thermal image characterizing the copper plate temperature distribution. Calculate the adjacent two-dimensional temperature thermal images through the frame difference method to obtain a two-dimensional temperature rate thermal image characterizing the copper plate temperature change distribution.

[0012] (3) Observe a large amount of temperature rate data in the sample library, and try to use different segmentation thresholds to distinguish the abnormal temperature fluctuation area from the normal temperature fluctuation area. After tracing back a large number of samples, select T z as the segmentation threshold for temperature rate binarization, and extract the abnormal temperature fluctuation area through the threshold segmentation algorithm, that is, eliminate the normal temperature fluctuation area with a temperature rate less than T z from the two-dimensional temperature rate thermal image, and judge the connectivity of the abnormal temperature rate points through the two-pass scan algorithm to obtain the distribution area of the abnormal temperature rate points of the copper plate.

[0013] The second step: Extraction of the shape features and expansion features of the abnormal area

[0014] (1) Extract the shape features of the abnormal temperature fluctuation region

[0015] (1.1) Starting from the upper left corner point of the abnormal temperature fluctuation region obtained in step (3) of the first step, traverse and extract all the outer contour points of the region in a clockwise direction according to the boundary tracking method, and represent them in the form of two-dimensional plane coordinates: (x0, y0), (x1, y1)…(x K-1 , y K-1 ), a total of K contour points in total.

[0016] (1.2) Represent the contour points with a coordinate sequence s(k) = [x(k), y(k)], k = 0, 1,...K-1, and convert each coordinate point pair into a complex number for processing, that is, s(k) = x(k) + jy(k), k = 0, 1, 2,..., K-1.

[0017] (1.3) Perform a discrete Fourier transform on s(k): The transformed complex coefficient a(u) is called the Fourier descriptor of the boundary of the abnormal temperature fluctuation region. Subsequently, s(k) can be restored through the inverse Fourier transform of these coefficients.

[0018] (1.4) Observe the Fourier coefficients calculated in (1.3). If the value of the Pth Fourier coefficient has approached 0, then take the first P Fourier coefficients to approximately reconstruct the contour and perform normalization processing on it to obtain the Fourier coefficients characterizing the contour of the abnormal temperature fluctuation region:

[0019] (1.5) Calculate the vector fd k = [fd0, fd1,...fd P-1 and take the modulus ||fd k || as the shape feature value Fourier_Descriptor of the current abnormal region.

[0020] (2) Extract the dynamic expansion features of the abnormal temperature fluctuation region

[0021] (2.1) Observe the continuous expansion behavior of the abnormal temperature fluctuation region within n seconds, extract its area value S t , t = 0, 1...n at each moment, and calculate the sum of the distances of the centroid of the region at each moment from the centroids of the regions at t = 0 and t = n

[0022] (2.2) Calculate the dynamic expansion coefficient of the abnormal region according to the following formula:

[0023]

[0024] Trace back a large number of abnormal temperature fluctuation regions in the sample library and calculate their values according to the method in (2.1). After summarization, select the minimum value D min and the maximum value D max as the lower and upper limits of the distance range to be satisfied; if the distance requirement is not met, set Sticking_Expansion to the invalid value Invalid.

[0025] Step 3: Feature combination and data preprocessing

[0026] (1) Refer to the production data report at the continuous casting site, extract the abnormal temperature fluctuation regions of the samples according to the method in the first step, and divide the abnormal temperature fluctuation regions into sticking breakout regions and normal working condition regions according to the report statistics. Extract their shape features and dynamic expansion features respectively according to the steps described in the second step, and combine them into a two-dimensional feature vector V B and V N , that is:

[0027] V B = [FD B , SE B

[0028] V N = [FD N , SE N

[0029] Among them, FD B and SE B are the shape feature and dynamic expansion feature of the sticking breakout region respectively, and FD N and SE N are the shape feature and dynamic expansion feature of the normal working condition region respectively.

[0030] (2) Construct a feature vector sample library D, which contains m feature vectors V B of the sticking breakout region and n feature vectors V N of the normal working condition region. D = {(V B1 , 1), (V B2 , 1),..., (V Bm , 1), (V N1 , 0), (V N2 , 0),..., (V Nn , 0)}, where 1 and 0 represent the sample labels of the sticking breakout region and the normal working condition region respectively.

[0031] (3) Normalize the samples in D:

[0032] ​​

[0033] Among them, V imin , V imax respectively represent the minimum value and the maximum value of the i-th dimension feature of the feature vector sample V, and Fv i represents the normalized value of the i-th dimension feature of the feature vector V. The sample library after normalization processing is denoted as P:

[0034] P = {(Fv B1 , 1), (Fv B2 , 1),...,(Fv Bm , 1), (Fv N1 , 0), (Fv N2 , 0),...,(Fv Nn , 0)}

[0035] Step 4: Construction and training of the neural network model

[0036] (1) Construction of network layers and parameter setting.

[0037] Construct a 3-layer BP neural network with a single hidden layer, where the network layers respectively contain d input neurons, l output neurons and q hidden layer neurons; the thresholds of the j-th neuron in the output layer and the h-th neuron in the hidden layer are θ j and γ h respectively; the connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is v ih , and the connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is w hj . Denote the input of the h-th neuron in the hidden layer as and the input of the j-th neuron in the output layer as where b h is the output of the h-th neuron in the hidden layer, and the activation function f of the network layer all adopts the sigmoid function.

[0038] (2) Train the neural network model based on the existing sample set P, and iteratively update all the weights and thresholds in the network according to the backpropagation algorithm.

[0039] (2.1) Input: sample set P, learning rate η.

[0040] (2.2) Parameter update process: Initialize all the weights and thresholds in the network layer. Perform the following iterative process for all samples in the sample set P:

[0041] (2.2.1) Obtain the output value

[0042] after forward propagation through the network layer according to the current input sample Among them is the actual label value of the current sample, i.e., 0 or 1.

[0043] (2.2.3) Calculate the gradients of the hidden layer neurons

[0044] (2.2.4) Update the network layer weights and thresholds according to the following formulas:

[0045] w hj ←w hj +Δw hj ,Δw hj =ηg j b h

[0046] θ j ←θ j +Δθ j ,Δθ j =-ηg j

[0047] v ih ←v ih +Δv ih ,Δv ih =ηe h Fv i

[0048] γ h ←γ h +Δγ h ,Δγ h =-ηe h

[0049] (2.3) Reach the iteration termination condition and stop updating the parameters.

[0050] (2.4) Output: A BP neural network with determined network weights and thresholds.

[0051] Step 5: The neural network model predicts breakout in real time

[0052] (1) Obtain the temperature data of the mold copper plate thermocouple in real time, obtain the two-dimensional temperature thermal image of the copper plate through temperature-color mapping, and then obtain the spatial distribution of the abnormal area of the copper plate temperature rate through inter-frame difference, threshold segmentation, and two-pass scanning method.

[0053] (2) Extract the shape features and dynamic expansion features of the abnormal area of the temperature rate, combine them and perform normalization processing to obtain the feature vector Fv characterizing the abnormal area of the temperature rate.

[0054] (3) Input the feature vector Fv into the BP neural network model obtained in the fourth step to obtain the predicted value of the model

[0055] (4) Based on the predicted value Determine whether the mold leaks steel. If then sticking breakout occurs, alarm and quickly reduce the casting speed of the casting machine; if then it is a normal working condition, go to step (1) to continue collecting and processing temperature data and monitoring breakout.

[0056] The above breakout prediction method is applicable to breakout prediction of continuous casting billets such as slab, square billet, round billet, and special-shaped billet.

[0057] The beneficial effects of the present invention are:

[0058] The prediction method provided by the present invention obtains the measured temperature of the thermocouple online, visually characterizes the abnormal temperature rate area, and then extracts and constructs the feature vector of the abnormal area. The neural network model is used to learn and train the feature vector sample library, and finally realizes the online prediction of mold breakout. This method is based on the neural network model to detect and predict mold breakout in real time. It can reduce the false alarm rate while ensuring a 100% sticking breakout reporting rate, and effectively improve the prediction accuracy. Description of the Drawings

[0059] Figure 1 is the flow chart of the mold breakout prediction method.

[0060] Figure 2 is the visual thermal image of the copper plate temperature and temperature rate.

[0061] Figure 3 is the schematic diagram of feature extraction in the normal working condition area. t = 0 - 5 represents the dynamic expansion and propagation process of the normal working condition area.

[0062] Figure 4 is the schematic diagram of feature extraction in the sticking breakout area. t = 0 - 5 represents the dynamic expansion and propagation process of the sticking breakout area.

[0063] Figure 5 is the BP neural network architecture diagram.

[0064] Figure 6 is the visual thermal image of the online detected abnormal area. Figure 6 (a) is the visual thermal image of sticking breakout; Figure 6 (b) is the visual thermal image of the normal working condition area. Detailed Embodiments

[0065] The following further elaborates the present invention through specific embodiments in combination with the drawings

[0066] Such as Figure 1The figure shows the flow chart of the mold breakout prediction method. First, the measured temperature of the thermocouple is obtained online, the abnormal temperature rate area of the copper plate is visually characterized, and its shape features and expansion features are extracted to construct a two-dimensional feature vector. Secondly, the neural network model is trained to obtain the prediction model. Finally, the neural network model is used to classify the feature vectors of sticking breakout and normal conditions in real time and predict the breakout.

[0067] Step 1: Visual characterization of the abnormal temperature rate area of the mold copper plate

[0068] (1) Three rows and 19 columns of thermocouples are arranged on the inner and outer arc wide-face copper plates of the mold, and three rows and 1 column of thermocouples are arranged on the left and right narrow-face copper plates. The temperature values at the thermocouple measurement points of the copper plate are obtained through the measured temperature of the thermocouple, and the copper plate temperature values at non-measurement points are calculated through the interpolation algorithm.

[0069] (2) As Figure 2 shown, the copper plate temperature is mapped onto a two-dimensional plane through computer graphics and OpenGL technology to obtain the two-dimensional temperature thermal image of the copper plate. The horizontal simulation interval is from the 1st column to the 19th column of thermocouples along the slab width direction, and the vertical simulation interval is from the 1st row to the 3rd row of thermocouples along the casting direction. The two-dimensional temperature rate thermal image of the copper plate is obtained by calculating adjacent two-dimensional temperature thermal images through the frame difference method.

[0070] (3) A large number of sample temperature rate data are statistically collected, and different segmentation thresholds are used to try to distinguish the abnormal temperature fluctuation area from the normal temperature fluctuation area through multiple tests. After tracing back a large number of samples, the temperature rate value divided by the threshold is set to 0.3 °C / s. The temperature rate thermal image is binarized through the threshold segmentation method, the normal temperature fluctuation area with a temperature rate less than 0.3 °C / s is removed, and all connected areas are extracted through the two-pass scanning algorithm to obtain the spatial distribution of the abnormal temperature fluctuation area of the copper plate. According to the production data report statistics on the continuous casting site, the abnormal temperature fluctuation area can be further divided into the sticking breakout area and the normal condition area.

[0071] Step 2: Extraction of the shape features and expansion features of the abnormal area

[0072] As Figure 3 shown is the dynamic expansion process of the normal condition area from t = 0 to t = 5. It can be seen from the figure that the normal condition area only has a longitudinal expansion trend along the casting direction and has no obvious lateral movement characteristics, that is, it is always in the in-situ movement state. And Figure 4 the sticking breakout area in

[0073] (1) Extraction of the shape features of the abnormal area

[0074] (1.1) Starting from the upper left corner point of the abnormal region at t = 5 in Figure 3 according to the boundary tracking method, all the peripheral contour points are extracted in sequence and represented in the form of two-dimensional coordinates: (0, 134), (0, 135)…(1, 134), a total of 106 contour points.

[0075] (1.2) Represent the contour point list in the form of a coordinate sequence: s(k) = [x(k), y(k)], k = 0, 1,...105, and convert each coordinate point pair into a complex number for processing, that is, s(k) = x(k) + jy(k), k = 0, 1, 2,..., 105.

[0076] (1.3) Perform a discrete Fourier transform on s(k): Take the complex coefficient a(u) as the Fourier descriptor characterizing the boundary shape of the abnormal temperature fluctuation region, and s(k) can be restored through the inverse Fourier transform of these coefficients subsequently.

[0077] (1.4) Observe the Fourier coefficients calculated in (1.3), and find that the 30th Fourier coefficient value has approached 0. Therefore, take the first 30 Fourier coefficients to approximately reconstruct the contour, and after normalization, obtain the Fourier coefficients of the abnormal temperature fluctuation region contour with translational, rotational, and scale invariance:

[0078] (1.5) Calculate the modulus length 4.632 of the vector fd k = [8.29, 2.46,...0.05] as the shape eigenvalue Fourier_Descriptor of the current normal working condition region. Similarly, it can be calculated that Figure 4 the shape feature Fourier_Descriptor of the bonding breakout region at t = 5 in

[0079] (2) Extraction of dynamic expansion characteristics of the abnormal region

[0080] (2.1) Extract the area value S of the normal working condition region in Figure 3 from t = 0 to t = 5 at each moment t = [S0, S1, S2, S3, S4, S5] = [10.68, 31.23, 54.74, 81.37, 122.46, 141.03].

[0081] And calculate the sum of the distances from the centroid of the region at each moment to the centroids of the regions at t = 0 and t = 5

[0082] (2.2) Calculate the dynamic expansion coefficient of the normal working condition region according to the following formula:

[0083]

[0084] Trace back a large number of samples in abnormal temperature fluctuation regions and calculate their values in the way of (2.1). After summarization and arrangement, select the minimum value 2 and the maximum value 47 as the lower and upper limits of the distance range to be satisfied. If the distance requirement is not met, set Sticking_Expansion to the invalid value Invalid = 10000. Calculate that in the normal working condition region, Sticking_Expansion = -1915.816; similarly, it can be obtained that Figure 4 in the sticking breakout region, Sticking_Expansion = 0.97

[0085] Step 3: Feature combination and data preprocessing

[0086] (1) Based on the sticking breakout region and the normal working condition region obtained in step 1 (3), extract their shape features and dynamic expansion features respectively according to the steps described in step 2 and combine them into a feature vector V B and V N , that is:

[0087] V B = [FD B , SE B = [1.47, 0.97]

[0088] V N = [FD N , SE N = [4.632, -1915.816]

[0089] Among them, FD B and SE B represent the shape feature and the dynamic expansion feature of the sticking breakout region respectively, and FD N and SE N represent the shape feature and the dynamic expansion feature of the normal working condition region respectively.

[0090] (2) Obtain 53 feature vectors V B of the sticking breakout region and 736 feature vectors V N of the normal working condition region in the way described in step 3 (1), and construct a feature vector sample set D = {(V B1 , 1), (V B2 , 1),..., (V B53 , 1), (V N1 , 0), (V N2 , 0),..., (V N736 , 0)}, where 1 and 0 represent the sample labels of the sticking breakout region and the normal working condition region respectively.

[0091] (3) Normalize the samples in D:

[0092]

[0093] Among them, V imin , V imax are the minimum and maximum values of the i-th dimension feature of the feature vector sample V, and Fv i represents the normalized value of the i-th dimension feature of the feature vector V. V B and V N The normalized feature vector is Fv B = [0.08, 0.99], Fv N = [0.25, 0.37]. After the normalization process of the sample set D, it is denoted as P:

[0094] P = {(Fv B1 , 1), (Fv B2 , 1),...,(Fv B53 , 1), (Fv N1 , 0), (Fv N2 , 0),...,(Fv N736 , 0)}

[0095] Step 4. Neural network model construction and training

[0096] (1) Network layer construction and parameter setting. As Figure 5 shown, construct a 3-layer BP neural network, with the input layer, hidden layer, and output layer containing 2, 5, and 1 neurons respectively. The threshold of the output layer neuron y j and the h-th neuron in the hidden layer is set to θ j and γ h ; the connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is v ih , and the connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is w hj . Denote the input of the h-th neuron in the hidden layer as The input of the output layer y j is denoted as where b h is the output of the h-th neuron in the hidden layer, and the activation function f of the network layer all uses the sigmoid function.

[0097] (2) Train the neural network model based on the preprocessed sample set P, and iteratively update all the weights and thresholds in the network according to the backpropagation algorithm.

[0098] (2.1) Input: Sample set P, learning rate η = 0.05.

[0099] (2.2) Parameter update process: Initialize all weights and thresholds in the network layer. θ j = 0, that is, no threshold is set;

[0100] γ = [γ1, γ2, γ3, γ4, γ5] = [0.16, 0.002, -0.09, 0.10, 0.05]

[0101] v = [[v 11 , v 12 , v 13 , v 14 , v 15 , [v 21 , v 22 , v 23 , v 24 , v 25 =

[0102] [[0.14, 0.10, -0.03, -0.10, 0.005], [-0.04, 0.11, -0.08, -0.01, 0.03]]

[0103] w = [w 11 , w 21 , w 31 , w 41 , w 51 [-0.99, 1.64, 1.93, 1.24, 1.61].

[0104] Perform the following iterative process on the first sample [0.08, 0.99] in the sample set P:

[0105] (2.2.1) Input it into the network for forward propagation to obtain the output value

[0106] (2.2.2) Backpropagate to calculate the gradient term of the output layer neurons where

[0107] (2.2.3) Similarly calculate the gradient term of the hidden layer neurons:

[0108]

[0109] e = [e1, e2, e3, e4, e5] = [-0.0009, 0.002, 0.002, 0.001, 0.002]

[0110] (2.2.4) Substitute the calculated gradient terms into the following formula to update the weights and thresholds of the network layer according to the gradient descent method:

[0111] w ← w + Δw, Δw = ηgj ×[b1, b2, b3, b4, b5] = [1.15×10 -4 , 1.05×10 -4 , 0.94×10 -4 , 1.00×10 -4 , 1.03×10 -4

[0112] v ← v + Δv, Δv = ηe[Fv1, Fv2] =

[0113] [[-4.46×10 -5 , 9.9×10 -5 , 9.9×10 -5 , 4.95×10 -5 , 9.9×10 -5 , [-2.25×10 -6 , 5×10 -6 , 5×10 -6 , 2.5×10 -6 , 5×10 -6 γ ← γ + Δγ, Δγ = -ηe = [0.45×10 -4 , -1×10 -4 , -1×10 -4 , -0.5×10 -4 , -1×10 -4

[0114] (2.2.4) Repeat the above parameter update process for the remaining samples in the sample set P.

[0115] (2.3) Repeat the above process until the maximum number of iterations reaches 10,000 times, then stop the iteration.

[0116] (2.4) Output: The trained BP neural network.

[0117] Step 5, Real-time prediction of breakout based on the neural network model

[0118] (1) According to the temperature data of the mold copper plate thermocouples obtained in real time, two two-dimensional temperature thermal images of the copper plate as shown in Figure 6 (a) and (b) are obtained through temperature-color mapping, and then the thermal image of the abnormal temperature fluctuation area of the copper plate is obtained through image processing.

[0119] (2) Extract the shape features and dynamic expansion features of the abnormal temperature fluctuation area, combine and normalize them to obtain the feature vector characterizing this area: Fv1 = [0.16, 0.99], Fv2 = [0.04, 0.47]

[0120] ​​(3) Predict Fv1 and Fv2 respectively through the BP neural network model to obtain the predicted values of the model. and

[0121] (4) Based on and Make a decision on the prediction of mold breakout. If it is adhesive breakout, an alarm is issued and the casting speed of the casting machine is quickly reduced; If it is normal working condition, go to step (1) to continue collecting and processing temperature data and monitoring breakout.

[0122] The above embodiments only represent the implementation manners of the present invention, but should not be construed as limiting the scope of the present invention patent. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for predicting mold breakout based on feature vectors and neural networks, characterized in that, The described method detects and forecasts mold breakout by extracting the shape features and expansion features of the abnormal temperature rate area of the mold copper plate and classifying the constructed feature vectors using a neural network, including the following steps: First step: Visual characterization of the abnormal temperature rate area of the mold copper plate to obtain the fluctuation area of the abnormal copper plate temperature rate points; Second step: Extraction of the shape features and expansion features of the abnormal area (1) Extract the shape features of the abnormal temperature fluctuation area (1.1) Starting from the upper left corner point of the abnormal temperature fluctuation region obtained in the first step according to the boundary tracking method, traverse and extract all the peripheral contour points of the region in a clockwise direction, and represent them in the form of two-dimensional plane coordinates: (x0, y0), (x1, y1) … (x K-1 , y K-1 ), a total of K contour points; (1.2) Represent the contour points with a coordinate sequence s(k) = [x(k), y(k)], k = 0, 1,... K - 1, and transform each coordinate point pair into a complex number for processing, i.e., s(k) = x(k) + jy(k), k = 0, 1, 2,..., K - 1; (1.3) Perform the discrete Fourier transform on s(k): The transformed complex coefficient a(u) is called the Fourier descriptor of the boundary of the abnormal temperature fluctuation region, and s(k) can be recovered by the inverse Fourier transform of these coefficients subsequently; (1.4) Observe the Fourier coefficients calculated in (1.3). If the value of the P-th Fourier coefficient has approached 0, then take the first P Fourier coefficients to approximately reconstruct the contour and perform normalization on it to obtain the Fourier coefficients characterizing the contour of the abnormal temperature fluctuation region: (1.5) Calculate the vector fd k = [fd0, fd1,... fd P-1 of the modulus ||fd k || as the shape eigenvalue Fourier_Descriptor of the current abnormal area; (2) Extract the dynamic expansion features of the abnormal temperature fluctuation area (2.1) Observe the continuous expansion behavior of the abnormal temperature fluctuation region within n seconds, and extract the area value S at each moment t , t = 0, 1... n, and calculate the sum of the distances of the center of gravity of the region at each moment from the center of gravity of the region at t = 0 and t = n (2.2) Calculate the dynamic expansion coefficient of the abnormal area according to the following formula: Backtrack a large number of abnormal temperature fluctuation regions in the sample library and calculate their values in the manner of (2.1). After summarization, select the minimum value D min and the maximum value D max as the lower and upper limits of the distance range that needs to be satisfied; if the distance requirement is not met, set Sticking_Expansion to the invalid value Invalid; Third step: Feature combination and data preprocessing (1) Refer to the production data report on the continuous casting site, extract the abnormal temperature fluctuation regions of the samples in the manner of the first step, and divide the abnormal temperature fluctuation regions into sticking breakout regions and normal working condition regions according to the report statistics. Extract their shape features and dynamic expansion features respectively according to the steps described in the second step, and combine them into a two-dimensional feature vector V B and V N , that is: V B = [FD B , SE B ​ V N = [FD N , SE N ​ Among them, FD B and SE B are respectively the shape feature and the dynamic expansion feature of the sticking breakout area, and FD N and SE N are respectively the shape feature and the dynamic expansion feature of the normal working condition area; (2) Construct a feature vector sample library D, which contains m feature vectors V of the sticking breakout area B and n feature vectors V of the normal working condition area N ; D = {(V B1 , 1), (V B2 , 1),..., (V Bm , 1), (V N1 , 0), (V N2 , 0),..., (V Nn , 0)}, where 1 and 0 represent the sample labels of the sticking breakout and normal working condition areas respectively; (3) Normalize the samples in D: Among them, V imin and V imax respectively represent the minimum value and the maximum value of the i-th dimension feature of the feature vector sample V. Fv i represents the normalized value of the i-th dimension feature of the feature vector V; the sample library after normalization is denoted as P: P = {(Fv B1 , 1), (Fv B2 , 1),..., (Fv Bm , 1), (Fv N1 , 0), (Fv N2 , 0),..., (Fv Nn , 0)} Fourth step: Construction and training of the neural network model (1) Construct a 3-layer BP neural network with a single hidden layer and set the parameters; (2) Train the neural network model based on the existing sample set P to output a BP neural network with determined network weights and thresholds; Fifth step: Real-time forecasting of mold breakout by the neural network model (1) Obtain the temperature data of the thermocouples on the mold copper plate in real time, obtain the two-dimensional temperature thermal image of the copper plate through temperature-color mapping, and then obtain the spatial distribution of the abnormal area of the copper plate temperature rate through inter-frame difference, threshold segmentation, and two-pass scanning method; (2) Extract the shape features and dynamic expansion features of the abnormal temperature rate area, combine them and perform normalization processing to obtain the feature vector Fv representing the abnormal temperature rate area; (3) Input the feature vector Fv into the BP neural network model obtained in the fourth step to obtain the predicted value of the model (4) Based on the predicted value judge whether the mold leaks steel; if then adhesive breakout occurs, alarm and quickly reduce the casting speed of the casting machine; if then it is a normal working condition, go to step (1) to continue collecting and processing temperature data and monitoring breakout.

2. The mold breakout prediction method based on feature vectors and neural networks according to claim 1, wherein The described method for mold breakout forecasting is applicable to mold breakout forecasting of slab, bloom, round billet, special-shaped billet or other continuous casting billets.

3. A mold breakout prediction method based on feature vectors and neural networks according to claim 1, characterized in that, The specific content of the fourth step is as follows: (1) Network layer construction and parameter setting; Construct a three-layer BP neural network with a single hidden layer, where the network layers include d input neurons, l output neurons, and q hidden neurons respectively; the thresholds of the j-th neuron in the output layer and the h-th neuron in the hidden layer are θ j and γ h respectively; the connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is v ih , and the connection weight between the h-th neuron in the hidden layer and the j-th neuron in the output layer is w hj respectively; denote the input of the h-th neuron in the hidden layer as and denote the input of the j-th neuron in the output layer as where b h is the output of the h-th neuron in the hidden layer, and the activation function f of the network layer all adopts the sigmoid function; (2) Train the neural network model based on the existing sample set P and iteratively update all the weights and thresholds in the network according to the backpropagation algorithm; (2.1) Input: Sample set P, learning rate η; (2.2) Parameter update process: Initialize all the weights and thresholds in the network layer; Perform the following iterative process for all samples in the sample set P: (2.2.1) The output value is obtained after the forward propagation of the current input sample through the network layer (2.2.2) Calculate the gradient of the output layer neurons where is the actual label value of the current sample, which is either 0 or 1; (2.2.3) Calculate the gradients of the hidden layer neurons (2.2.4) Update the network layer weights and thresholds according to the following formula: w hj ←w hj +Δw hj ,Δw hj =ηg j b h θ j ←θ j +Δθ j ,Δθ j =-ηg j v ih ←v ih +Δv ih ,Δv ih =ηe h Fv i γ h ←γ h +Δγ h ,Δγ h =-ηe h (2.3) Reach the iteration termination condition and stop parameter update; (2.4) Output a BP neural network with determined network weights and thresholds.

4. A mold breakout prediction method based on feature vectors and neural networks according to claim 1, characterized in that The first step is specifically as follows: (1) Arrange multiple thermocouples on the inner and outer arc wide-face copper plates and the left and right narrow-face copper plates of the mold; Online detect the temperature of each thermocouple and calculate the copper plate temperature value at the non-thermocouple measurement point through interpolation algorithm; (2) Establish the mapping relationship between the copper plate temperature and the two-dimensional plane to obtain the two-dimensional temperature thermal image representing the copper plate temperature distribution; Calculate the adjacent two-dimensional temperature thermal images through the inter-frame difference method to obtain the two-dimensional temperature rate thermal image representing the copper plate temperature change distribution; (3) Observe a large number of temperature rate data in the observation sample library, and use T z as the segmentation threshold for temperature rate binarization. Extract the abnormal temperature fluctuation region through the threshold segmentation algorithm, that is, remove the normal temperature fluctuation region with a temperature rate less than T z from the two-dimensional temperature rate thermal image, and judge the connectivity of the abnormal temperature rate points through the two-pass scanning algorithm to obtain the distribution region of the abnormal temperature rate points of the copper plate.

Images