Continuous casting breakout forecasting method and system based on deep learning

Abstract

The invention relates to the technical field of continuous casting detection in the metallurgical industry, in particular to a continuous casting breakout forecasting method and system based on deep learning, and the method comprises the steps: obtaining a plurality of samples; obtaining the temperature change deviation of each thermocouple in each sample; analyzing the distribution range of the temperature data of each thermocouple in each sample and the temperature change deviation, and obtaining the temperature change value of each thermocouple in each sample; obtaining the spatial distribution difference of each thermocouple in each sample by comparing the temperature distribution of each thermocouple in each sample and all the other thermocouples and the temperature difference between each thermocouple and the adjacent thermocouple at the position of each thermocouple; obtaining the fitness of each sample; and labels are distributed to all the samples according to the fitness, a neural network is trained through the samples with the labels distributed, and a breakout forecasting model is obtained. According to the method, the distinguishing precision of the bonded bleed-out sample and the normal sample is improved, and the forecasting precision of the bleed-out forecasting model is further improved.

Description

Technical Field

[0001] The present application relates to the field of continuous casting detection technology in the metallurgical industry, and specifically to a method and system for predicting continuous casting leakage based on deep learning. Background Art

[0002] In the continuous casting production process, bonding leakage is the most common form of leakage. Once a leakage accident occurs, it may seriously damage the equipment and even endanger personal safety. Therefore, it is very necessary to develop a real-time and effective leakage prediction model.

[0003] In the process of constructing the steel leakage prediction model, when distinguishing the temperature data under different working conditions, the temperature change characteristics of a single thermocouple and the relationship between the temperature changes of different rows of thermocouples were not fully considered, resulting in insufficient accuracy in distinguishing the temperature data. As a result, the subsequent steel leakage prediction model obtained using the distinguished temperature data is prone to omissions or false alarms. Summary of the invention

[0004] In view of the above, it is necessary to provide a continuous casting breakout prediction method and system based on deep learning. Compared with the traditional continuous casting breakout prediction method, the accuracy of distinguishing between bonding breakout samples and normal samples is improved, thereby improving the prediction accuracy of the breakout prediction model:

[0005] In a first aspect, an embodiment of the present application provides a continuous casting leak prediction method based on deep learning, the method comprising the following steps:

[0006] Acquire multiple samples, each sample containing temperature data of all thermocouples within each preset time period;

[0007] By comparing the change of temperature data of each thermocouple in each sample with that of all thermocouples in all samples, the temperature change deviation of each thermocouple in each sample is obtained;

[0008] Analyze the distribution range of the temperature data of each thermocouple in each sample and the temperature change deviation to obtain the temperature change value of each thermocouple in each sample;

[0009] By comparing the temperature distribution of each thermocouple in each sample with that of all other thermocouples, and the temperature difference between each thermocouple and its adjacent thermocouple, the spatial distribution difference of each thermocouple in each sample is obtained;

[0010] Combining the temperature change values ​​and spatial distribution differences of all thermocouples in each sample, obtaining the fitness of each sample;

[0011] All samples are labeled according to the fitness, and a neural network is trained by the labeled samples to obtain a steel leakage prediction model.

[0012] In one embodiment, the process of obtaining the temperature change deviation is:

[0013] Arrange all temperature data of each thermocouple in each sample in time sequence to form each temperature sequence, and obtain a fitting curve of all temperature data in each temperature sequence;

[0014] Obtaining the curvature calculation formula of each fitting curve, and calculating the definite integral result of each curvature calculation formula in the interval [1, N], wherein N is the total number of data in the temperature sequence; the change is reflected by the definite integral result;

[0015] Calculating the mean of the definite integral results of all thermocouples in all samples;

[0016] The temperature variation deviation is obtained by comparing the definite integral result of each thermocouple in each sample with the mean value.

[0017] In one embodiment, the expression of the temperature variation deviation is: S j,i =I j,i -I j,avg Where S j,i represents the temperature variation deviation of the i-th thermocouple in the j-th sample; I j,i represents the integration result of the i-th thermocouple in the j-th sample; I j,avg represents the mean of the integration results of all thermocouples in all samples.

[0018] In one embodiment, the process of obtaining the temperature change value is:

[0019] Calculate the range of all temperature data for each thermocouple in each sample;

[0020] The temperature change value is positively correlated with the extreme difference and the temperature change deviation respectively.

[0021] In one embodiment, the expression of the spatial distribution difference is:

[0022] In the formula, Z j,i represents the spatial distribution difference of the i-th thermocouple in the j-th sample; norm() represents the normalization function; calculates the mean temperature data of each row of thermocouples in the j-th sample at each acquisition time, and records the difference in the mean temperature data between the row where the i-th thermocouple in the j-th sample is located and the row above it as the temperature value difference, t j,irepresents the sum of the temperature differences of the i-th thermocouple in the j-th sample at all acquisition moments; obtains the stationary point of the fitting curve of each thermocouple, arranges the appearance moments of the stationary points of all thermocouples in the j-th sample in chronological order to form a time series, and arranges the time intervals between any two adjacent moments in the time series to form a time difference series, ΔT avg represents the sum of all values ​​in the first-order difference sequence of the time difference sequence; α represents a preset value greater than 0.

[0023] In one embodiment, the process of obtaining the fitness is as follows:

[0024] Calculate the fusion results of the temperature change value and spatial distribution difference of each thermocouple in each sample;

[0025] The fitness is the sum of the fusion results of all thermocouples in each sample.

[0026] In one embodiment, the process of assigning labels to all samples by using the fitness is:

[0027] A genetic algorithm is used to select initial clustering centers according to the suitability of each sample, and a clustering algorithm is used to cluster all samples based on the initial clustering centers. The bonding leaking steel samples and normal samples are distinguished by the clustering results, and different labels are assigned to the bonding leaking steel samples and normal samples. The label of the normal sample is a first preset value, and the label of the bonding leaking steel sample is a second preset value.

[0028] In one embodiment, the step of training a neural network by assigning labeled samples includes:

[0029] All bonding leak samples are divided into training sets and test sets according to a preset ratio, and normal samples are divided into training sets and test sets according to a preset ratio;

[0030] The neural network is trained and tested using the training set and the test set.

[0031] In one embodiment, when testing the trained neural network, the test result evaluation indicators include prediction accuracy and reporting rate;

[0032] Forecast accuracy σ p The expression is: Report rate r The expression is: Where N r Indicates the number of correct alarms, N f Indicates the number of false alarms, and N1 indicates the number of missed alarms;

[0033] When the output of the neural network is the second preset value, an alarm is issued.

[0034] In the second aspect, an embodiment of the present application also provides a continuous casting leak prediction system based on deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor, and when the processor executes the computer program, it implements the steps of any one of the above-mentioned continuous casting leak prediction methods based on deep learning.

[0035] This application has at least the following beneficial effects:

[0036] The present application fully considers the temperature changes of thermocouples installed at different positions on the crystallizer, including the fact that the temperature of a single thermocouple shows a trend of first rising and then falling when bonding leakage occurs, and the temperature change range of a single thermocouple is large, and the temperature changes of adjacent rows of thermocouples show "time lag" and "temperature inversion" in space, and obtains fitness, which more accurately reflects the working conditions to which the temperature data belongs; compared with the traditional use of the quality index of clustering results as fitness, the calculated fitness does not depend on the intra-cluster distance and the inter-cluster distance, so that under data noise or dimensionality disaster, it can reduce the distance measurement deviation, select the optimal initial clustering center, and then classify the samples based on the initial clustering center, thereby improving the accuracy of distinguishing samples;

[0037] Furthermore, the steel leakage prediction model is trained based on the accurately distinguished samples, so that the steel leakage prediction model can more accurately capture the intrinsic characteristics of the temperature data, and thus have higher prediction accuracy when the steel leakage prediction model is used to predict steel leakage in the continuous casting process. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 A flowchart of the steps of a continuous casting breakout prediction method based on deep learning provided in one embodiment of the present application;

[0040] Figure 2 This is a schematic diagram of the layout of the thermocouple;

[0041] Figure 3 It is a schematic diagram of temperature change when bonding steel leaks out;

[0042] Figure 4 This is a schematic diagram of the temperature change of adjacent rows of thermocouples when bonding steel leaks out;

[0043] Figure 5 This is a schematic diagram of the structure of the Transformer model;

[0044] Figure 6 This is a schematic diagram of the structure of the Transformer-BiLSTM model;

[0045] Figure 7 This is a diagram of the accuracy-loss value of the Transformer-BiLSTM model;

[0046] Figure 8 This is a diagram of the accuracy-loss value of the Kmeans-Transformer-BiLSTM model;

[0047] Fig. 9 The following is a diagram of the accuracy-loss value of the GA-Kmeans-Transformer-BiLSTM model. DETAILED DESCRIPTION

[0048] In the description of the embodiments of the present application, words such as "exemplary", "or", "for example" and the like are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "or", "for example" and the like is intended to present related concepts in a concrete manner.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art in the present application. The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. It should be understood that, unless otherwise specified, " / " means or.

[0050] It should also be noted that the terms "first" and "second" in the present application are used to distinguish similar objects rather than to describe a specific order or sequence.

[0051] The specific scheme of the continuous casting leak prediction method and system based on deep learning provided by the present application is described in detail below with reference to the accompanying drawings.

[0052] See also Figure 1 , which shows a flowchart of a method for predicting continuous casting breakout based on deep learning provided by an embodiment of the present application, the method comprising the following steps:

[0053] Step 1: Acquire multiple samples, each sample containing temperature data of all thermocouples in each preset time period.

[0054] The temperature data is collected using thermocouples buried in the copper plate of the crystallizer, wherein the thermocouples are evenly buried in the copper plate of the crystallizer.

[0055] The size of the crystallizer in this embodiment is 1870 cm long, 400 cm wide, and 1200 cm high. Six rows of thermocouples are installed on the two fixed wide faces, the left narrow face, and the right narrow face of the crystallizer. Among them, eight rows of thermocouples are installed on the two fixed wide faces, and two rows of thermocouples are installed on the left narrow face and the right narrow face, for a total of 120 thermocouples. The layout diagram of the thermocouples is shown in FIG. Figure 2 As shown, Figure 2 In the figure, 101 represents a fixed wide surface, and the surface parallel to 101 is also a fixed wide surface. 102 represents a right narrow surface, and the surface parallel to the right narrow surface is a left narrow surface. 103 represents a thermocouple. The implementer can set the number of rows and columns of thermocouples and the distance between adjacent thermocouples by himself. This application does not impose any special restrictions.

[0056] In this embodiment, during the process of collecting temperature data, the collection time interval is 2s. The value of the collection time interval is preset manually and can be set by the implementer. This application does not impose any special restrictions.

[0057] Due to the complex environment of the continuous casting site, the temperature data collected by the thermocouple will be interfered by various noises, so it is necessary to filter the collected temperature data. The useful information in the temperature signal is concentrated in the low-frequency part, and the noise is in the high-frequency part, so a low-pass filter is used to filter the temperature data to remove the high-frequency noise. Among them, the low-pass filter is a well-known technology and will not be described in detail in this application.

[0058] By analyzing the historical temperature data of any steel plant, after a single thermocouple experiences bonding and steel leakage, the temperature data collected by the thermocouple shows a phenomenon of first rising and then falling. The schematic diagram of temperature change during bonding and steel leakage is shown in the figure below: Figure 3 As shown, Figure 3 In the figure, 104 represents the ascending mode, 105 represents the intermediate transition mode, 106 represents the whole process mode, and 107 represents the descending mode.

[0059] In this embodiment, bonding steel leakage samples and normal samples with large fluctuations that are prone to false alarms are selected from historical temperature data, specifically: 40 groups of bonding steel leakage samples are selected in each mode during the bonding steel leakage process, a total of 160 groups of bonding steel leakage samples, and 140 groups of normal samples are selected. Each group of samples is a matrix, and the matrix has a total of 120 rows. Each row of the matrix is ​​a temperature sequence composed of temperature data collected by a single thermocouple within 30s arranged in time sequence. Among them, 40 and 140 are only an embodiment of the present application, and the implementer can select other numbers of bonding steel leakage samples and normal samples at will, and the present application does not make special restrictions. 30s is due to the fact that it is learned through the analysis of historical temperature data: when bonding steel leakage occurs, the temperature mutation time of a single thermocouple is 30s.

[0060] Step 2: Comprehensively analyze the temperature change characteristics of a single thermocouple and the relationship between the temperature changes of different rows of thermocouples to obtain the fitness.

[0061] This application uses a K-means clustering algorithm based on a genetic algorithm to cluster all samples. By comparing and measuring the similarity between data, all samples are divided into K categories. For the samples in this embodiment, they can be divided into bonding leakage samples and normal samples, so K is set to 2.

[0062] Selecting cluster centers through genetic algorithms can avoid imbalanced distribution of cluster centers and fall into local optimal solutions. However, the traditional GA-Kmeans algorithm usually uses the quality indicators of clustering results as fitness, such as the silhouette coefficient and the Davidson-Bourdin index. This fitness function depends on the intra-cluster distance and inter-cluster distance, and does not take into account the single-pair complete time series and group-pair spatial linkage characteristics of thermocouples. Under data noise or dimensional disaster, the distance measurement deviation is large, which affects the selection of the optimal cluster center, resulting in inaccurate clustering results and affecting subsequent identification. Therefore, the fitness function can be determined according to the temperature change characteristics of the thermocouple.

[0063] The schematic diagram of the temperature change of adjacent rows of thermocouples during bonding leakage is as follows Figure 4 As shown, Figure 4In the figure, a represents the bonding point, b represents the copper plate of the crystallizer, 1, 2, and 3 represent the first, second, and third rows of thermocouples, respectively. In the longitudinal direction, due to the increase in the thickness of the billet shell, the temperature of the lower row of thermocouples is lower than that of the upper row of thermocouples. Under normal conditions, the temperature curve of the thermocouples is stable, as shown in stage ①. When bonding and steel leakage occur, the bonding point first passes through the first row of thermocouples, the molten steel contacts the copper plate of the crystallizer, and the temperature of the thermocouples rises rapidly, as shown in stage ②. When the bonding point continues to move downward, the temperature of the first row of thermocouples drops, and the temperature of the second row of thermocouples rises rapidly, as shown in stage ③. Similarly, as the bonding point moves downward and passes through the third row of thermocouples, the temperature of the third row of thermocouples rises, and the temperatures of the first and second rows of thermocouples drop, as shown in stage ④. Finally, as the bonding point moves away, the temperature of the thermocouples continues to drop, as shown in stage ⑤.

[0064] When bonding and steel leakage occurs, the temperature of the thermocouples shows a trend of first rising and then falling, and the temperature curve of the lower row of thermocouples shows a "time lag" phenomenon compared to the upper row. The phenomenon that the temperature of the lower row of thermocouples is higher than that of the upper row is called "temperature inversion".

[0065] Step 2.1, by comparing the change of temperature data of each thermocouple in each sample with that of all thermocouples in all samples, the temperature change deviation of each thermocouple in each sample is obtained.

[0066] When bonding and leakage occur, the temperature of a single thermocouple will change greatly. In the early stage, the molten steel contacts the copper plate, the heat flow increases suddenly, and the temperature rises sharply. As the bonding point moves, the heat transfer of the embryo shell changes, the heat exchange of the molten steel is more complicated, and the temperature fluctuates or decreases. Under normal working conditions, the temperature of a single thermocouple is relatively stable, and even if there is a temperature fluctuation, it is relatively small. The reason is that under normal working conditions, the heat distribution and transfer inside the crystallizer reach a state of equilibrium, in which the heat is jointly formed through conduction, convection movement of the molten steel, and heat dissipation to the surrounding environment, forming a dynamic thermal balance.

[0067] Based on the above analysis, taking the jth sample as an example, in order to better distinguish between bonding leakage and normal working conditions, the fitting curves of all data in each temperature sequence are obtained, the curvature calculation formulas of each fitting curve are obtained, and the definite integral results of each curvature calculation formula in the interval [1, N] are calculated, where N is the total number of data in the temperature sequence. The curvature calculation formula and the calculation of the definite integral are both well-known technologies and will not be repeated in this application.

[0068] In this embodiment, the least squares method is used to obtain the fitting curve of all the data in each temperature sequence. As other implementation methods, on the basis of obtaining the fitting curve of all the data in each temperature sequence, the implementer can adopt other existing technologies, such as polynomial fitting methods, etc., and this application does not impose any special restrictions.

[0069] Furthermore, by comparing the definite integral results of each thermocouple in each sample with all thermocouples in all samples, the temperature change deviation of each thermocouple in each sample is obtained to characterize the temperature change deviation of each thermocouple in each sample, and the expression is:

[0070] S j,i =I j,i -I j,avg Where S j,i represents the temperature variation deviation of the i-th thermocouple in the j-th sample; I j,i represents the integration result of the i-th thermocouple in the j-th sample; I j,avg represents the mean of the integration results of all thermocouples in all samples.

[0071] It should be noted that: the integration result represents the cumulative effect of the bending degree of the fitting curve in the interval [1, N]. The larger the temperature change deviation is, the greater the temperature change of the i-th thermocouple in the j-th sample is compared with all thermocouples in all samples, and the temperature data of the i-th thermocouple in the j-th sample is more likely to be the temperature data when bonding leakage occurs.

[0072] Step 2.2, analyzing the distribution range of the temperature data of each thermocouple in each sample and the temperature change deviation, and obtaining the temperature change value of each thermocouple in each sample.

[0073] When temperature fluctuations occur or a thermocouple fails, the temperature will jump. At this time, there is no bonding or leakage, and it is still in normal working conditions, but the temperature change deviation will be large. In order to avoid misjudgment, it is necessary to further judge based on the distribution range of the thermocouple temperature data.

[0074] Based on the above analysis, the distribution range of the temperature data of each thermocouple in each sample and the temperature change deviation are analyzed to obtain the temperature change value of each thermocouple in each sample, and characterize the temperature change degree of each thermocouple in each sample. The expression is:

[0075] X j,i =norm(R j,i ×S j,i );where X j,i represents the temperature change value of the i-th thermocouple in the j-th sample; norm() represents the normalization function; R j,i represents the range of the data in the temperature series of the i-th thermocouple in the j-th sample; S j,i Represents the temperature change deviation of the i-th thermocouple in the j-th sample.

[0076] In this embodiment, the Min-Max normalization method is used to normalize R j,i ×S j,iNormalization is performed, as another implementation method, in which R j,i ×S j,i On the basis of normalization processing, the implementer may adopt other existing technologies, such as decimal calibration normalization method, Sigmoid function, etc., and this application does not impose any special restrictions.

[0077] It should be noted that: for a single thermocouple in all samples, even if there is temperature fluctuation or a thermocouple failure under normal working conditions, causing the temperature to jump, the temperature change range is much smaller than the case of bonding and leaking. The reason is that the filtering process will remove the data with excessive deviation in the temperature data; the greater the cumulative effect of the curvature of the fitting curve of the i-th thermocouple, and the larger the temperature change range, the more likely the temperature data of the i-th thermocouple in the j-th sample is the temperature data when bonding and leaking occurs, and the greater the temperature change value. The temperature change value is used as part of the fitness of the genetic algorithm. The larger the fitness value, the more suitable the sample is as the cluster center, otherwise, it is less suitable.

[0078] Step 2.3, by comparing the temperature distribution of each thermocouple in each sample with that of all other thermocouples, and the temperature difference between each thermocouple and its adjacent thermocouple at the same position, the spatial distribution difference of each thermocouple in each sample is obtained.

[0079] The temperature change of a single thermocouple is not enough to reflect the situation of bonding and leakage. Since multiple rows of thermocouples are installed on the crystallizer, the temperature changes and correlation between different rows of thermocouples can reflect the movement characteristics of the bonding point, thereby determining whether bonding occurs. Under normal working conditions, the temperature field in the crystallizer is uniform and stable, multiple thermocouples work together smoothly, the temperature changes on each surface and between each row are small, the heat conduction is uniform, and a stable spatial temperature pattern is formed.

[0080] When bonding and steel leakage occurs, temperature imbalance will occur and a strong heat source will be formed at the bonding point. For each row of thermocouples, the thermocouple closest to the bonding point will heat up first. As the position of the bonding point changes, the lower row of thermocouples will gradually heat up, while the upper row of thermocouples will gradually cool down as they move away from the bonding point, eventually forming a temperature inversion phenomenon in which the temperature of the lower row of thermocouples is higher than that of the upper row of thermocouples.

[0081] Based on the above analysis, by comparing the temperature distribution of each thermocouple in each sample with that of all other thermocouples, as well as the temperature difference between the row where each thermocouple is located and the row above it, the spatial distribution difference of each thermocouple in each sample is obtained to characterize the influence of the change in the position of the bonding point on the temperature change of each row of thermocouples. The expression is:

[0082] In the formula, Z j,irepresents the spatial distribution difference of the i-th thermocouple in the j-th sample; norm() represents the normalization function; calculates the mean temperature data of each row of thermocouples in the j-th sample at each acquisition time, and records the difference in the mean temperature data between the row where the i-th thermocouple in the j-th sample is located and the row above it as the temperature value difference, t j,i represents the sum of the temperature differences of the i-th thermocouple in the j-th sample at all acquisition moments; obtains the stationary point of the fitting curve of each thermocouple, arranges the appearance moments of the stationary points of all thermocouples in the j-th sample in chronological order to form a time series, and arranges the time intervals between any two adjacent moments in the time series to form a time difference series, ΔT avg It represents the sum of all values ​​in the first-order difference sequence of the time difference sequence; α represents a preset value greater than 0, the purpose of which is to avoid the denominator being 0. The value of α is preset manually and can be set by the implementer. In this embodiment, the value of α is 0.1.

[0083] In this embodiment, the difference between the mean values ​​of the temperature data is the absolute value of the difference. As other implementation methods, the implementer may adopt other calculation methods, such as the square of the difference, the ratio, etc., and this application does not impose any special restrictions.

[0084] In this embodiment, the Min-Max normalization method is used to Normalization is performed, as another implementation method, in which On the basis of normalization processing, the implementer may adopt other existing technologies, such as decimal calibration normalization method, Sigmoid function, etc., and this application does not impose any special restrictions.

[0085] It should be noted that: j,i When the value is larger, the difference between the temperature of each row of thermocouples and the temperature of the previous row is greater, indicating that bonding leakage is more likely to occur and temperature inversion will occur; the stagnation point indicates the point where the temperature starts to drop from the highest point, ΔT avg The smaller it is, the more time-lag the temperature changes of different rows of thermocouples are, that is, the temperatures of different rows of thermocouples reach the highest point at different times; the more likely it is that bonding and leakage will occur, and the greater the value of the spatial distribution difference. Among them, for the temperature sequence under normal working conditions and the temperature sequence of thermocouples far away from the bonding point, the fitting curve does not have a stagnation point. If the fitting curve of any thermocouple does not have a stagnation point, then when calculating the spatial distribution difference of any thermocouple, ΔT avg The value of is 0. The spatial distribution difference is used as part of the fitness of the genetic algorithm. The larger the fitness value, the more suitable it is as a cluster center, and vice versa.

[0086] Step 2.4, combining the temperature change values ​​and spatial distribution differences of all thermocouples in each sample to obtain the fitness of each sample.

[0087] In summary, the fusion results of the temperature change value and spatial distribution difference of each thermocouple in each sample are calculated; the fitness of each sample is obtained by combining the fusion results of all thermocouples in each sample, and the expression is:

[0088] In the formula, F j represents the fitness of the jth sample; n represents the number of thermocouples in the jth sample; H j,i It represents the fusion result of the temperature change value and spatial distribution difference of the i-th thermocouple in the j-th sample.

[0089] It should be understood that fusion refers to combining multiple independent variables in a way that enhances the overall effect, such as additive relationship, multiplicative relationship, etc. The implementer can limit it according to the actual situation, and this application does not impose any special restrictions.

[0090] In this embodiment, the fusion result is the sum of the temperature change value and the spatial distribution difference.

[0091] In another embodiment, the fusion result is the product of the temperature change value and the spatial distribution difference.

[0092] The calculated fitness of each sample is used instead of the traditional silhouette coefficient and Davidson-Bourdin index. The calculated fitness does not depend on the intra-cluster distance and inter-cluster distance, and takes into account the temperature change characteristics of a single thermocouple and the connection between the temperature changes of thermocouples in different rows. The larger the fitness of any sample, the more it conforms to the bonding leakage phenomenon and the more suitable it is to become the cluster center.

[0093] Step 3, optimize the GA-Kmeans model and assign labels to all samples according to the fitness.

[0094] In order to cluster all samples and distinguish between bonded steel leakage samples and normal samples based on clustering results, two initial clustering centers need to be obtained, including the initial clustering center of bonded steel leakage samples and the initial clustering center of normal samples. When the initial clustering center of bonded steel leakage samples is selected by genetic algorithm according to the fitness of each sample, the samples selected from the historical temperature data are first encoded and converted into gene form, and then the initial population is randomly generated, and the quality of the samples is determined according to the fitness. Then, selection, crossover and mutation operations are performed until the maximum number of iterations or the convergence threshold is reached, and the sample with the largest fitness output by the genetic algorithm is used as the initial clustering center of bonded steel leakage samples. When the initial clustering center of normal samples is selected by genetic algorithm according to the fitness of each sample, the samples selected from the historical temperature data are first encoded and converted into gene form, and then the initial population is randomly generated, and the quality of the samples is determined according to the opposite number of the fitness. Then, selection, crossover and mutation operations are performed until the maximum number of iterations or the convergence threshold is reached, and the sample with the largest fitness output by the genetic algorithm is used as the initial clustering center of normal samples.

[0095] In this embodiment, when using a genetic algorithm to select the initial clustering center, the population size is set to 200, the crossover probability is set to 0.8, the mutation probability is set to 0.05, the maximum number of iterations is set to 200, and the convergence threshold is set to 0.01. The values ​​of the population size, crossover probability, mutation probability, maximum number of iterations and convergence threshold are preset manually, and the implementer can choose the values ​​according to the actual situation. This application does not impose any special restrictions.

[0096] Based on the initial cluster center determined by the genetic algorithm, a clustering algorithm is used to assign each sample selected from the historical temperature data to the nearest cluster center. The samples in the same cluster cluster as the initial cluster center of the bonding leakage sample are used as bonding leakage samples; the samples in the same cluster cluster as the initial cluster center of the normal sample are used as normal samples. In this process, the cluster center position is continuously updated iteratively until the cluster cluster to which the sample belongs no longer changes or the preset number of iterations is reached, thereby achieving a stable division of temperature data based on similarity, effectively distinguishing the temperature data corresponding to bonding leakage and normal samples, and providing solid data preprocessing support for the subsequent leakage prediction model. After clustering, a label is assigned to each sample, the label of the normal sample is the first preset value, and the label of the bonding leakage sample is the second preset value.

[0097] In this embodiment, the first preset value is 0, indicating a normal working condition, and the second preset value is 1, indicating that bonding leakage occurs.

[0098] In this embodiment, the value of the preset number of iterations is 200. The value of the preset number of iterations is preset manually and can be set by the implementer. This application does not impose any special restrictions.

[0099] Step 4: Build the Transformer-BiLSTM model.

[0100] Bi-directional Long Short-Term Memory (BiLSTM) is an improved version of Long Short-Term Memory (LSTM) and contains two LSTM layers. BiLSTM can process sequences in both the forward and backward directions at the same time, capturing the past and future information of the sequence, which helps to improve the prediction performance of the model. However, due to its own recursive structure, BiLSTM cannot really solve the problem of long-term dependency. The Transformer model based on the self-attention mechanism uses a parallel attention mechanism for feature extraction, which does not have the problem of long-term dependency and can extract distant time series features. The Transformer is combined with BiLSTM to enhance the ability to capture time series data. The core mechanisms of the Transformer model include self-attention mechanism, multi-head attention, position encoding, residual connection and layer normalization. The structural diagram of the Transformer model is shown in the figure. Figure 5 As shown in the figure. In the steel leakage prediction model, the input is the known historical temperature data, and the output is the temperature data classification result. In the above case, the attention layer structure of the decoder is not required. Therefore, the decoder part in the traditional Transformer model is modified to a fully connected layer, and the original attention part is replaced by the BiLSTM part to establish the Transformer-BiLSTM model. The structural diagram of the Transformer-BiLSTM model is shown in the figure. Figure 6 The traditional Transformer-BiLSTM model is a well-known technology.

[0101] In this embodiment, the input is a sample of temperature data of 120 thermocouples, and the classification output is a label. The number of hidden neurons of the Transformer-BiLSTM model is 128, the number of attention heads is 4, the number of key channels is 128, the batch size is 64, the initial learning rate is 0.01, the maximum number of iterations is 100, the optimizer is Adam, and the classifier is Softmax. The values ​​of the number of hidden neurons, the number of attention heads, the number of key channels, the batch size, the initial learning rate, and the maximum number of iterations can be set by the implementer according to the actual situation, and this application does not make special restrictions.

[0102] Step 5, training the neural network by assigning labeled samples to obtain a steel leakage prediction model.

[0103] The optimized GA-Kmeans model is combined with the Transformer-BiLSTM model to form a GA-Kmeans-Transformer-BiLSTM model. In order to improve the adaptability and generalization ability of the steel leakage prediction model, the training samples and test samples are cross-selected to avoid duplication, the bonding steel leakage samples are divided into the training set and the test set according to the first preset ratio, and the normal samples are divided into the training set and the test set according to the second preset ratio. The training set is used to train the Transformer-BiLSTM model, the Kmeans-Transformer-BiLSTM model and the GA-Kmeans-Transformer-BiLSTM model respectively. Among them, the samples in the training set and the test set are all labeled. The loss function is the cross entropy loss function. The accuracy-loss value diagram of the Transformer-BiLSTM model is shown in Figure 7 As shown, the accuracy-loss value diagram of the Kmeans-Transformer-BiLSTM model is as follows Figure 8 As shown, the accuracy-loss value diagram of the GA-Kmeans-Transformer-BiLSTM model is as follows Fig. 9 As shown. Figure 7 It can be seen from the figure that after 100 iterations of the Transformer-BiLSTM model, neither the accuracy nor the loss value reached the best. Figure 8 It can be seen that the Kmeans-Transformer-BiLSTM model has an accuracy of 100% after 44 iterations, and the loss value drops to 0. Fig. 9 It can be seen that the GA-Kmeans-Transformer-BiLSTM model reaches 100% accuracy at 18 iterations, and the loss value drops to 0, reaching the optimal state faster and having the best training performance. Among them, model training and accuracy calculation are both well-known technologies, and this application will not repeat them.

[0104] In this embodiment, the first preset ratio is 5:3, and the second preset ratio is 5:2. The implementer can set the values ​​of the first preset ratio and the second preset ratio at will, and this application does not impose any special restrictions.

[0105] The trained Transformer-BiLSTM model, Kmeans-Transformer-BiLSTM model and GA-Kmeans-Transformer-BiLSTM model are tested using the test set. When the output of the Transformer-BiLSTM model, Kmeans-Transformer-BiLSTM model and GA-Kmeans-Transformer-BiLSTM model is the second preset value, an alarm is issued. Test result evaluation indicators are introduced, including prediction accuracy and reporting rate.

[0106] Forecast accuracy σ p The expression is: Report rate r The expression is: Where N r Indicates the number of correct alarms, N f Indicates the number of false alarms, and N1 indicates the number of missed alarms.

[0107] The test results of the Transformer-BiLSTM model, the Kmeans-Transformer-BiLSTM model and the GA-Kmeans-Transformer-BiLSTM model are shown in Table 1.

[0108] Table 1 Test results of the model

[0109]

[0110] As shown in Table 1, both the Transformer-BiLSTM model and the Kmeans-Transformer-BiLSTM model have the problem of missing reports, while the GA-Kmeans-Transformer-BiLSTM model has 0 missing reports and a report rate of 100%. In addition, the Transformer-BiLSTM model and the Kmeans-Transformer-BiLSTM model have 6 and 4 false reports, respectively, with prediction rates of 93.0% and 96.0%, respectively, while the GA-Kmeans-Transformer-BiLSTM model has only 1 false report, with a prediction accuracy of up to 99.5%, which is a significant improvement compared to the prediction accuracy of the other two models. It can be seen that the GA-Kmeans-Transformer-BiLSTM model has improved the recognition accuracy of both normal samples and bonded steel leakage samples, and has better solved the problems of missing reports and false reports, and has better recognition effect and generalization ability.

[0111] The trained GA-Kmeans-Transformer-BiLSTM model is used as the steel leakage prediction model.

[0112] The samples to be classified are input into the GA-Kmeans-Transformer-BiLSTM model, and the prediction results are output. The prediction results are the labels of the samples. During the continuous casting production process, the manufacturing process of high-quality steel can be guided according to the prediction results.

[0113] Based on the same inventive concept as the above-mentioned method, an embodiment of the present application also provides a continuous casting leak prediction system based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of any one of the above-mentioned continuous casting leak prediction methods based on deep learning are implemented.

[0114] In summary, the present application fully considers the temperature changes of thermocouples installed at different positions on the crystallizer, including the fact that the temperature of a single thermocouple shows a trend of first rising and then falling when bonding leakage occurs, and the temperature change range of a single thermocouple is large, and the temperature changes of adjacent rows of thermocouples show "time lag" and "temperature inversion" in space, and obtains fitness, which more accurately reflects the working conditions to which the temperature data belongs; compared with the traditional use of the quality index of clustering results as fitness, the calculated fitness does not depend on the intra-cluster distance and the inter-cluster distance, so that under data noise or dimensionality disaster, it can reduce the distance measurement deviation, select the optimal initial clustering center, and then classify the samples based on the initial clustering center, thereby improving the accuracy of distinguishing samples;

[0115] Furthermore, the steel leakage prediction model is trained based on the accurately distinguished samples, so that the steel leakage prediction model can more accurately capture the intrinsic characteristics of the temperature data, and thus have higher prediction accuracy when the steel leakage prediction model is used to predict steel leakage in the continuous casting process.

[0116] The flowchart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to the embodiment of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which can depend on the functions involved. In the description corresponding to the flowchart and the block diagram in the accompanying drawings, the operations or steps corresponding to different boxes can also occur in a different order from the order disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, which can depend on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified functions or actions, or may be implemented by a combination of dedicated hardware and computer instructions.

[0117] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the basic features of the present application. Therefore, no matter from which point of view, the above embodiments of the present application should be regarded as exemplary and non-restrictive.

Claims

1. A continuous casting breakout prediction method based on deep learning, characterized in that: The method comprises the following steps: Acquire multiple samples, each sample containing temperature data of all thermocouples within each preset time period; By comparing the change of temperature data of each thermocouple in each sample with that of all thermocouples in all samples, the temperature change deviation of each thermocouple in each sample is obtained; Analyze the distribution range of the temperature data of each thermocouple in each sample and the temperature change deviation to obtain the temperature change value of each thermocouple in each sample; By comparing the temperature distribution of each thermocouple in each sample with that of all other thermocouples, and the temperature difference between each thermocouple and its adjacent thermocouple, the spatial distribution difference of each thermocouple in each sample is obtained; Combining the temperature change values ​​and spatial distribution differences of all thermocouples in each sample, obtaining the fitness of each sample; All samples are labeled according to the fitness, and a neural network is trained by the labeled samples to obtain a steel leakage prediction model.

2. The continuous casting breakout prediction method based on deep learning according to claim 1, characterized in that: The process of obtaining the temperature change deviation is as follows: Arrange all temperature data of each thermocouple in each sample in time sequence to form each temperature sequence, and obtain a fitting curve of all temperature data in each temperature sequence; Obtaining the curvature calculation formula of each fitting curve, and calculating the definite integral result of each curvature calculation formula in the interval [1, N], wherein N is the total number of data in the temperature sequence; the change is reflected by the definite integral result; Calculating the mean of the definite integral results of all thermocouples in all samples; The temperature variation deviation is obtained by comparing the definite integral result of each thermocouple in each sample with the mean value.

3. The continuous casting breakout prediction method based on deep learning as claimed in claim 2, characterized in that: The expression of the temperature variation deviation is: S j,i =I j,i -I j,avg Where S j,i represents the temperature variation deviation of the i-th thermocouple in the j-th sample; I j,i represents the integration result of the i-th thermocouple in the j-th sample; I j,avg represents the mean of the integration results of all thermocouples in all samples.

4. The method for predicting continuous casting breakout based on deep learning according to claim 1, characterized in that: The process of obtaining the temperature change value is as follows: Calculate the range of all temperature data for each thermocouple in each sample; The temperature change value is positively correlated with the extreme difference and the temperature change deviation respectively.

5. The method for predicting continuous casting breakout based on deep learning according to claim 2, characterized in that: The expression of the spatial distribution difference is: In the formula, Z j,i represents the spatial distribution difference of the i-th thermocouple in the j-th sample; norm() represents the normalization function; calculates the mean temperature data of each row of thermocouples in the j-th sample at each acquisition time, and records the difference in the mean temperature data between the row where the i-th thermocouple in the j-th sample is located and the row above it as the temperature value difference, t j,i represents the sum of the temperature differences of the i-th thermocouple in the j-th sample at all acquisition moments; obtains the stationary point of the fitting curve of each thermocouple, arranges the appearance moments of the stationary points of all thermocouples in the j-th sample in chronological order to form a time series, and arranges the time intervals between any two adjacent moments in the time series to form a time difference series, ΔT avg represents the sum of all values ​​in the first-order difference sequence of the time difference sequence; α represents a preset value greater than 0.

6. The continuous casting breakout prediction method based on deep learning according to claim 1, characterized in that: The process of obtaining the fitness is as follows: Calculate the fusion results of the temperature change value and spatial distribution difference of each thermocouple in each sample; The fitness is the sum of the fusion results of all thermocouples in each sample.

7. The continuous casting breakout prediction method based on deep learning according to claim 1, characterized in that: The process of assigning labels to all samples through the fitness is: A genetic algorithm is used to select initial clustering centers according to the suitability of each sample, and a clustering algorithm is used to cluster all samples based on the initial clustering centers. The bonding leaking steel samples and normal samples are distinguished by the clustering results, and different labels are assigned to the bonding leaking steel samples and normal samples. The label of the normal sample is a first preset value, and the label of the bonding leaking steel sample is a second preset value.

8. The method for predicting continuous casting breakout based on deep learning according to claim 7, characterized in that: The step of training a neural network by assigning labeled samples comprises: All bonding leak samples are divided into training sets and test sets according to a preset ratio, and normal samples are divided into training sets and test sets according to a preset ratio; The neural network is trained and tested using the training set and the test set.

9. The method for predicting continuous casting breakout based on deep learning according to claim 8, characterized in that: When testing the trained neural network, the test result evaluation indicators include prediction accuracy and reporting rate; Forecast accuracy σ p The expression is: Report rate r The expression is: Where N r Indicates the number of correct alarms, N f Indicates the number of false alarms, and N1 indicates the number of missed alarms; When the output of the neural network is the second preset value, an alarm is issued.

10. A continuous casting breakout prediction system based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the continuous casting leak prediction method based on deep learning as described in any one of claims 1 to 9 are implemented.

Images