A mold sticking and breakout prediction method and device, electronic equipment and medium

Abstract

This invention provides a method, apparatus, electronic device, and medium for predicting crystallizer adhesion leakage. The method includes: obtaining the original time series of thermocouple temperatures and the corresponding pulling speed time series for each sample from adhesive and non-adhesive samples; training a model based on the pulling speed time series and the original thermocouple temperature time series to obtain a crystallizer adhesion leakage prediction model; obtaining a predicted value of the adhesion probability of a single row of thermocouples through the crystallizer adhesion leakage prediction model; and providing a crystallizer adhesion leakage prediction method by combining the lateral transmission characteristics of the adhesion point. This method is applicable to scenarios where a certain row of thermocouples in a multi-row thermocouple system is disabled, which helps reduce the missed detection rate of adhesion phenomena and also helps reduce the number of thermocouple rows installed. Compared with the logical threshold judgment rule, it can not only achieve more accurate and timely identification of adhesion phenomena in multiple application scenarios, but also helps reduce the human error and time cost introduced by constantly adjusting threshold parameters based on the actual situation.

Description

Technical Field

[0001] This application relates to the field of continuous casting detection technology in iron and steel metallurgy, specifically to a method, device, electronic equipment, and medium for predicting steel leakage due to crystallizer adhesion. Background Technology

[0002] In continuous casting, molten steel enters the tundish from the ladle via a long nozzle, and then flows into the crystallizer through a submerged entry nozzle. Due to heat transfer between the high-temperature molten steel and the water-cooled copper plates of the crystallizer, it begins to cool and solidify, forming a billet shell of a certain thickness. Under the vibration of the crystallizer and the action of the casting process, the billet shell is continuously pulled out into the secondary cooling zone. However, in actual production, factors such as poor lubrication of the continuous casting flux, excessive fluctuations in the crystallizer liquid level, or improper control of the molten steel's superheat can cause the billet to easily adhere to the copper plates of the crystallizer. If relevant measures are not taken in time, the billet shell is easily torn apart and re-solidified under the vibration of the crystallizer and the action of the casting process, forming new adhesion points. When these adhesion points move down to the crystallizer outlet, the billet shell, due to its insufficient strength to withstand the internal static pressure of the molten steel and the resistance of the casting process, cracks, resulting in a blowout. Blowouts not only cause continuous casting machine shutdowns, leading to production interruptions and reduced steel production efficiency, but also involve energy waste, safety risks, and environmental pollution. Therefore, the establishment of a crystallizer blowout prediction system is of great significance.

[0003] Patent document CN107096899A discloses a crystallizer leakage prediction system based on logic judgment. This system primarily targets all thermocouples on the copper plate of the crystallizer. Based on various logic judgment rules, it performs temperature rise detection, temperature rise rate detection, continuous temperature drop detection, temperature inversion detection, and temperature change delay detection for three thermocouples within the inverted triangle area of ​​abnormal temperature rise thermocouples. The system then comprehensively analyzes the detection results to determine whether to trigger a bonding alarm. However, using temperature inversion as a characteristic parameter for identifying bonding phenomena is inappropriate for two reasons. First, the bonding alarm point corresponding to temperature inversion occurs after the intersection of the temperature time sequence curves of the two rows of thermocouples, resulting in a certain alarm lag and a higher risk of leakage. Second, in actual production, due to factors such as fluctuations in the molten steel level, the performance of the continuous casting protective slag, and the fluidity of the molten steel, the temperature readings of the lower row of thermocouples are often higher than those of the upper row. If the temperature inversion detection rule designed in this patent is used in such cases, false alarms are highly likely to occur. Meanwhile, the crystallizer bonding leakage prediction method based on the logic threshold judgment rule can only achieve theoretically zero false alarms and zero missed alarms when the bonded and non-bonded samples are linearly separable. Furthermore, the design of the logic threshold judgment rule is highly dependent on process conditions, equipment quality, and physical property parameters, requiring continuous adjustment of parameter thresholds based on actual field conditions. This results in poor robustness of the crystallizer bonding leakage prediction method based on the logic threshold judgment rule, manifested in higher maintenance costs and a high susceptibility to false alarms in new production modes. Finally, this invention patent requires consideration of all thermocouples during feature extraction, which does not contribute to the cost reduction resulting from reducing the number of thermocouple rows. Summary of the Invention

[0004] In view of the shortcomings of the above-mentioned related technologies, this application provides a method, apparatus, electronic device and medium for predicting crystallizer sticking and leakage, so as to solve the above problems.

[0005] This application provides a method for predicting crystallizer adhesion leakage. The method includes: acquiring the original time series of thermocouple temperatures and the casting speed time series corresponding to each original sample, wherein the original samples include adhesion samples between the billet and the crystallizer and non-adhesion samples between the billet and the crystallizer; obtaining a feature matrix based on the casting speed time series and the original time series of thermocouple temperatures, and training a model using the feature matrix to obtain a crystallizer adhesion leakage prediction model; inputting the temperature time series of any two rows of each column of the thermocouple to be detected and their corresponding casting speed time series extracted in advance or in real time within a preset distribution range into the crystallizer adhesion leakage prediction model to obtain multiple single-column thermocouple adhesion probability prediction values, and generating a corresponding single-column thermocouple adhesion probability prediction value sequence for each of the multiple single-column thermocouple adhesion probability prediction values ​​based on continuous detection over a preset time period; and determining whether there is adhesion in the crystallizer by combining the single-column thermocouple adhesion probability prediction value in each single-column thermocouple adhesion probability prediction value sequence with the lateral transmission feature recognition criterion of adhesion points.

[0006] In one embodiment of this application, the temperature time sequence of any two rows of thermocouples to be detected in each column within a preset distribution range, extracted in advance or in real time, and their corresponding casting speed time sequence are input into the crystallizer sticking leakage prediction model to obtain multiple single-column thermocouple sticking probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple sticking probability prediction value sequence is generated for each of the multiple single-column thermocouple sticking probability prediction values. This includes: acquiring the temperature time sequence of any two rows of thermocouples in some or all columns within the continuous casting machine within a preset time period, as the temperature time sequence of any two rows of thermocouples to be detected in each column within the preset distribution range. The process involves: inputting the temperature time sequence of any two rows of thermocouples to be detected in each column within the preset distribution range, along with their corresponding pulling speed time sequence, into the crystallizer adhesion leakage prediction model to obtain multiple single-column thermocouple adhesion probability prediction values; combining the single-column thermocouple adhesion probability prediction values ​​corresponding to each moment within the preset time period to form a single-column thermocouple adhesion probability prediction value sequence to obtain multiple single-column thermocouple adhesion probability prediction value sequences; and combining the single-column thermocouple adhesion probability prediction value in each single-column thermocouple adhesion probability prediction value sequence with the adhesion point lateral transmission feature identification criterion to determine whether the crystallizer exhibits adhesion phenomena.

[0007] In one embodiment of this application, obtaining a feature matrix based on the pulling speed time series and the original thermocouple temperature time series includes: filtering outliers in the original thermocouple temperature time series to obtain the thermocouple temperature time series; performing smoothing and noise reduction processing on the thermocouple temperature time series using a Bezier curve to obtain a smoothed and denoised thermocouple temperature time series; obtaining feature values ​​based on the pulling speed time series and the smoothed and denoised thermocouple temperature time series, and constructing a feature matrix using the feature values.

[0008] In one embodiment of this application, a model training method is used to obtain a crystallizer sticking leakage prediction model through the feature matrix, including: setting the label value of the sticking sample to a first preset value and setting the label value of the non-sticking sample to a second preset value; combining the first preset value and the second preset value, the feature matrix is ​​stratified and sampled according to a preset data division ratio to obtain a training set, a validation set, and a test set, wherein the training set and the validation set are used for model training, and the test set is used to test the trained crystallizer sticking leakage prediction model; fitting the initial crystallizer sticking leakage prediction model to the training set, and adjusting the initial crystallizer sticking leakage prediction model through the validation set to obtain the crystallizer sticking leakage prediction model.

[0009] In one embodiment of this application, determining whether adhesion exists by combining the predicted adhesion probability value of a single-column thermocouple in each sequence with the identification criterion of lateral transfer characteristics of adhesion points includes: if all the predicted adhesion probability values ​​of a single-column thermocouple in at least one sequence are greater than a first preset probability, then the thermocouple column corresponding to the latest sequence of predicted adhesion probability values ​​is marked as a third preset value; if the thermocouple column corresponding to the third preset value conforms to the lateral transfer characteristics of adhesion points, then it is determined that adhesion exists in the crystallizer and an alarm message is issued; if the thermocouple column corresponding to the third preset value does not conform to the lateral transfer characteristics of adhesion points, then it is determined that adhesion does not exist in the crystallizer.

[0010] In one embodiment of this application, determining whether adhesion exists by combining the predicted adhesion probability value of a single column of thermocouples in each sequence with the identification criterion of lateral transfer characteristics of adhesion points includes: acquiring the predicted adhesion probability value of each column of thermocouples in consecutive time intervals to obtain the predicted adhesion probability value of thermocouples at a first time interval, a predicted adhesion probability value at a second time interval, and a predicted adhesion probability value of thermocouples at a third time interval for each column of thermocouples; if the predicted adhesion probability value of a single column of thermocouples at the first time interval and the predicted adhesion probability value of thermocouples at the second time interval are both greater than a second preset probability, and the predicted adhesion probability value of a single column of thermocouples at the third time interval is greater than a third preset probability, then the thermocouple column corresponding to the predicted adhesion probability value of a single column of thermocouples at the third time interval is marked with a third preset value; if the thermocouple column corresponding to the third preset value conforms to the lateral transfer characteristics of adhesion points, then it is determined that adhesion exists in the crystallizer, and an alarm message is issued; if the thermocouple column corresponding to the third preset value does not conform to the lateral transfer characteristics of adhesion points, then it is determined that adhesion does not exist in the crystallizer.

[0011] In one embodiment of this application, if the marked value conforms to the lateral transmission characteristics of the bonding point, then it is determined that a bonding phenomenon exists, including: extracting the thermocouple column corresponding to the third preset value; obtaining the lateral propagation rate of the bonding point based on the time difference of the thermocouple column corresponding to the third preset value and the corresponding average pulling speed; obtaining the detection time through the lateral propagation rate of the bonding point and the arrangement spacing between the thermocouple columns corresponding to the third preset value; if the number of thermocouple columns corresponding to the third preset value is greater than or equal to a fourth preset value within the detection time, then it is determined that the crystallizer has a bonding phenomenon; if the number of thermocouple columns corresponding to the third preset value is less than the fourth preset value within the detection time, then it is determined that the crystallizer does not have a bonding phenomenon.

[0012] This application provides a crystallizer adhesion leakage prediction device, the device comprising: a time series acquisition module, used to acquire the original time series of thermocouple temperatures and the casting speed time series corresponding to each original sample; the original samples include adhesion samples of the billet and the copper plate of the crystallizer, and non-adhesion samples of the billet and the copper plate of the crystallizer; a model training module, used to obtain a feature matrix based on the casting speed time series and the original time series of thermocouple temperatures, and to train the model through the feature matrix to obtain a crystallizer adhesion leakage prediction model; a time series detection module, used to input the time series of thermocouple temperatures to be detected and their corresponding casting speed time series of any two rows in each column within a preset distribution range, extracted in advance or in real time, into the crystallizer adhesion leakage prediction model to obtain multiple single-column thermocouple adhesion probability prediction values, and to form a corresponding single-column thermocouple adhesion probability prediction value sequence for each of the multiple single-column thermocouple adhesion probability prediction values ​​based on continuous detection over a preset time period; and an adhesion determination module, used to determine whether there is adhesion in the crystallizer by combining the single-column thermocouple adhesion probability prediction value in each single-column thermocouple adhesion probability prediction value sequence with the lateral transmission feature recognition criterion of adhesion points.

[0013] Beneficial Effects: This invention provides a method, device, electronic device, and storage medium for predicting mold adhesion leakage. It obtains the original thermocouple temperature time series and the corresponding casting speed time series for each sample from bonded and non-bonded samples of the billet and the mold. A feature matrix is ​​obtained based on the casting speed time series and the original thermocouple temperature time series, and a model is trained using the feature matrix to obtain a mold adhesion leakage prediction model. The temperature time series of any two rows of thermocouples to be detected and their corresponding casting speed time series within a preset distribution range are input into the mold adhesion leakage prediction model to obtain multiple single-column thermocouple adhesion probability prediction values. Multiple thermocouple adhesion probability prediction value sequences are formed based on these multiple single-column thermocouple adhesion probability prediction values. The presence of adhesion is determined by combining the single-column thermocouple adhesion probability prediction values ​​in each sequence with the lateral transmission feature recognition criteria of the adhesion point. By using a neural network model to obtain the predicted adhesion probability of a single-row thermocouple, and comprehensively considering the temporal sequence of multiple single-row thermocouple adhesion probability predictions and the lateral transmission characteristics of adhesion points, this approach is suitable for scenarios where a particular row of thermocouples in a multi-row thermocouple system is disabled. This helps reduce the false negative rate of adhesion phenomena and also helps reduce the number of thermocouple rows installed, saving costs. Compared to traditional logical threshold judgment methods, this approach can more accurately and promptly identify adhesion phenomena in various application scenarios, such as different types of temperature change characteristics and multiple production sites, preventing continuous casting machines from shutting down due to adhesion and steel leakage, improving steel production efficiency, and avoiding energy waste, safety risks, and environmental pollution caused by adhesion and steel leakage. Furthermore, it eliminates the need for continuous adjustment of threshold parameters based on actual on-site conditions, helping to reduce human error and time costs. On the other hand, by monitoring and analyzing the temporal sequence of the predicted adhesion probability of a single-row thermocouple, the risk of adhesion and steel leakage can be detected in advance. By combining multiple single-row thermocouple bonding probability prediction value sequences with the identification criteria of lateral transmission characteristics of bonding points, alarm signals can be reasonably designed. The triggering of the alarm signal will remind the operator to take preventive measures in time or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of bonding and steel leakage.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a crystallizer sticking leak prediction method as an exemplary embodiment of this application;

[0016] Figure 2 This is an exemplary embodiment of the present application illustrating the temperature and pulling speed timing of a single-row thermocouple in a bonded sample;

[0017] Figure 3 This is an exemplary embodiment of the present application illustrating the temperature and pulling speed timing of a single-row thermocouple in a bonded sample;

[0018] Figure 4This is an exemplary embodiment of the present application illustrating the temperature and pulling speed timing of a single-row thermocouple in a bonded sample;

[0019] Figure 5 This is an example diagram illustrating the feature set of a crystallizer sticking and steel leakage prediction model, as shown in an exemplary embodiment of this application.

[0020] Figure 6 This is a flowchart illustrating a crystallizer sticking leak prediction method, as shown in an exemplary embodiment of this application.

[0021] Figure 7 This is a comparison diagram of marked time points illustrating the adhesion phenomenon in an exemplary embodiment of this application;

[0022] Figure 8 This is a block diagram illustrating a crystallizer sticking and steel leakage prediction device, as shown in an exemplary embodiment of this application;

[0023] Figure 9 This is a schematic diagram illustrating the structure of an electronic device as shown in an exemplary embodiment of this application. Detailed Implementation

[0024] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0025] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] It should be noted that in this application, terms such as "first" and "second" are merely for distinguishing similar objects, and do not limit the order or sequence of similar objects. The variations of "including" and "having" indicate that the scope covered by the subject of the word is not exclusive, except for the examples shown by the word.

[0027] It is understood that the various numerical designations, step numbers, and other identifiers recorded in this application are for descriptive convenience and are not intended to limit the scope of this application. The size of the identifiers in this application does not imply the order of execution; the execution order of each process should be determined by its function and internal logic.

[0028] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0029] The embodiments of this application respectively propose a method for predicting crystallizer sticking and leakage, a device for predicting crystallizer sticking and leakage, an electronic device, and a storage medium. These embodiments will be described in detail below.

[0030] First, it's important to clarify that continuous casting primarily involves casting refined molten steel into billets using a continuous casting machine. The main process includes: transporting a ladle filled with refined molten steel to a turret; once the turret is in the pouring position, the molten steel is poured into the tundish; and then the tundish distributes the molten steel to the various crystallizers through the nozzle. The crystallizer is one of the core pieces of the continuous casting machine. It allows the casting to take shape and solidify rapidly. Molten steel is poured into the crystallizer from above, and through its cooling effect, it begins to solidify inside, forming a billet shell. As the billet moves downwards, the shell gradually thickens until the desired thickness is reached. The billet is then pulled out of the crystallizer by a leveling machine and enters the secondary cooling zone for further cooling and solidification. The design and operation of the crystallizer significantly impact the quality of the billet and the production efficiency of the continuous casting machine. The crystallizer should possess good thermal conductivity, wear resistance, and structural rigidity to ensure uniform solidification of the billet and prevent problems such as leaks. However, in actual production, due to factors such as poor lubrication effect of continuous casting protective slag, excessive fluctuation of liquid level in the crystallizer, or improper control of molten steel superheat, the billet is prone to sticking to the copper plate of the crystallizer. If relevant measures are not taken in time, the billet shell is easily torn and re-solidified under the action of crystallizer vibration and billet pulling to form new sticking points. When the sticking points move down to the crystallizer outlet, the billet shell is not strong enough to withstand the static pressure of the molten steel inside and the resistance of billet pulling, resulting in rupture and leakage of steel. This phenomenon is known as sticking and leakage, a major form of leakage in continuous casting. Sticking and leakage not only cause continuous casting machine shutdowns, leading to production interruptions and reduced steel production efficiency, but also involve energy waste, safety risks, and environmental pollution. Therefore, this invention uses a neural network model to obtain predicted sticking probabilities, forming a sequence of predicted thermocouple sticking probabilities. Based on the predicted thermocouple sticking probabilities in each sequence, it determines whether sticking occurs. This allows for accurate and timely identification of sticking, preventing continuous casting machine shutdowns due to sticking and leakage, improving steel production efficiency, and avoiding energy waste, safety risks, and environmental pollution caused by sticking and leakage.

[0031] Please see Figure 1 , Figure 1This is a flowchart illustrating an exemplary embodiment of the present application of a method for predicting crystallizer sticking and leakage. For example... Figure 1 As shown, in an exemplary embodiment, the method for predicting crystallizer sticking and steel leakage includes at least steps S110 to S140, which are described in detail below:

[0032] Step S110: Obtain the original thermocouple temperature time series and the pulling speed time series corresponding to each original sample. The original samples include the bonded samples between the billet and the crystallizer and the non-bonded samples between the billet and the crystallizer.

[0033] In one embodiment of this application, the casting speed refers to the speed of the casting rolls of the continuous casting machine. The non-bonded samples of the billet and the crystallizer include samples with large fluctuations in thermocouple temperature timing, samples with irregular small oscillations in temperature timing, and samples of normal casting before the bonding instance. In this embodiment, each sample includes the original temperature timing of the first row of thermocouples and the original temperature timing of the second row of thermocouples. Each sample also includes the casting speed timing. As an exemplary embodiment, the obtained original temperature timing of the first row of thermocouples is P = {p0, p1, ..., p...} n The original time sequence of the second row of thermocouple temperatures is Q = {q0, q1, ..., q}. n The pulling speed sequence is S = {s0, s1, ..., s}. n Specifically, a total of 744 original samples were obtained, and the pulling speed time series corresponding to each original sample was collected to obtain the pulling speed time series S={s0,s1,…,s 59}, collect the original time sequence of the first row of thermocouple temperatures {p0,p1,…,p} corresponding to each original sample. 59 The original time sequence of the second row of thermocouple temperatures {q0,q1,…,q} 59 The data acquisition time is denoted as {t0, t1, ..., t}. 59 The time intervals are approximately 1 second. For example, such as... Figure 2 As shown, Figure 2 This invention illustrates a time-series diagram of temperature and pulling speed of a single-row thermocouple in a bonded sample. In this embodiment, by extracting the temperature and pulling speed time sequence of a single-row thermocouple from the bonded sample, a time-series diagram showing temperature inversion and typical bonding characteristics of the single-row thermocouple temperature measurements is obtained. As an example, such as... Figure 3 As shown, Figure 3 The present invention illustrates a time series diagram of temperature and pulling speed of a single-row thermocouple in a bonded sample. In this embodiment, by extracting the temperature and pulling speed time series of a single-row thermocouple in the bonded sample, a time series diagram is obtained showing the temperature inversion and M-shaped temperature change characteristics of the temperature measurement values ​​of the single-row thermocouple in the bonded sample.

[0034] As an example, such as Figure 4 As shown, Figure 4The present invention illustrates a time series diagram of temperature and pulling speed of a single-row thermocouple in a bonded sample. In this embodiment, by extracting the temperature and pulling speed time series of a single-row thermocouple in the bonded sample, a time series diagram is obtained showing that the temperature measurement values ​​of the single-row thermocouple in the bonded sample exhibit a large temperature difference between the upper row and the middle and lower rows, and that the upper and middle rows show an approximately parallel decreasing characteristic.

[0035] Step S120: Obtain the feature matrix based on the pulling speed time sequence and the original thermocouple temperature time sequence, and train the model using the feature matrix to obtain the crystallizer sticking and leakage prediction model.

[0036] In one embodiment of this application, model training using feature matrices not only extracts the most relevant and representative data from the raw data, but also facilitates batch processing of data by machine algorithms by transforming the raw data into a structured format. Furthermore, model training using feature matrices makes it easier to debug and optimize the model. When model performance is poor, the problem can be identified by examining the feature matrix and model parameters, and appropriate adjustments can be made.

[0037] Step S130: Input the temperature time sequence of any two rows of thermocouples to be detected in each column and their corresponding pulling speed time sequence within the preset distribution range, which are extracted in advance or in real time, into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, generate a corresponding single-column thermocouple bonding probability prediction value sequence for each of the multiple single-column thermocouple bonding probability prediction values.

[0038] In one embodiment of this application, a thermocouple is an important temperature measuring device, generally used to measure the temperature of molten steel. In continuous casting, the temperature of the molten steel is a critical process parameter, directly affecting the quality of the cast billet and the production efficiency of the continuous casting machine. Therefore, real-time monitoring and control of the molten steel temperature is necessary. The working principle of a thermocouple is based on the thermoelectric effect, that is, when the temperature changes, free electrons in a conductor or semiconductor are affected by heat energy, generating a potential difference. By measuring this potential difference, the temperature value can be determined. Since thermocouple temperature data is relatively easy to monitor and accurately measure, and the temperature change of a thermocouple is the most intuitive and representative characteristic when adhesion occurs, this application uses real-time temperature data from thermocouples to detect adhesion phenomena in the crystallizer.

[0039] Step S140: Combine the predicted probability values ​​of single-column thermocouple bonding in each single-column thermocouple bonding probability prediction value sequence with the identification criteria for lateral transfer characteristics of bonding points to determine whether there is bonding phenomenon in the crystallizer.

[0040] In one embodiment of this application, compared with traditional threshold-based methods for detecting adhesion phenomena, predicting the adhesion probability of the crystallizer using a deep learning model is more accurate and reliable. This is because classifying adhered and non-adhesive samples is not necessarily a simple linearly separable problem, and the deep learning-based prediction method is beneficial for uncovering hidden features in the thermocouple temperature time series that are helpful for identifying adhesion phenomena. At the same time, it reduces the human error and time cost introduced by constantly adjusting threshold parameters based on actual field conditions. By generating multiple thermocouple adhesion probability prediction value sequences, a better understanding of the changing trend of adhesion risk can be achieved. This helps to formulate more reasonable maintenance plans, reduce unnecessary maintenance activities, and ensure that appropriate measures are taken when the adhesion risk is high, thereby reducing the number and duration of production interruptions, lowering equipment failure rates, increasing the stability and reliability of the production line, and ultimately improving production efficiency.

[0041] exist Figure 1 In the technical solution shown, the original thermocouple temperature time series and the corresponding casting speed time series for each sample are obtained from the bonded and non-bonded samples of the billet and the crystallizer. A feature matrix is ​​obtained based on the casting speed time series and the original thermocouple temperature time series, and the model is trained using the feature matrix to obtain a crystallizer bonding leakage prediction model. The temperature time series of any two rows of the thermocouples to be detected and their corresponding casting speed time series within a preset distribution range are input into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Multiple thermocouple bonding probability prediction value sequences are formed based on the multiple single-column thermocouple bonding probability prediction values. The presence of bonding phenomenon is determined by combining the single-column thermocouple bonding probability prediction value in each sequence with the lateral transmission feature recognition criterion of the bonding point. By using a neural network model to obtain the predicted adhesion probability of a single-row thermocouple, and comprehensively considering the temporal sequence of multiple single-row thermocouple adhesion probability predictions and the lateral transmission characteristics of adhesion points, this approach is suitable for scenarios where a particular row of thermocouples in a multi-row thermocouple system is disabled. This helps reduce the false negative rate of adhesion phenomena and also helps reduce the number of thermocouple rows installed, saving costs. Compared to traditional logical threshold judgment methods, this approach can more accurately and promptly identify adhesion phenomena in various application scenarios, such as different types of temperature change characteristics and multiple production sites, preventing continuous casting machines from shutting down due to adhesion and steel leakage, improving steel production efficiency, and avoiding energy waste, safety risks, and environmental pollution caused by adhesion and steel leakage. Furthermore, it eliminates the need for continuous adjustment of threshold parameters based on actual on-site conditions, helping to reduce human error and time costs. On the other hand, by monitoring and analyzing the temporal sequence of the predicted adhesion probability of a single-row thermocouple, the risk of adhesion and steel leakage can be detected in advance. By combining multiple single-row thermocouple bonding probability prediction value sequences with the identification criteria of lateral transmission characteristics of bonding points, alarm signals can be reasonably designed. The triggering of the alarm signal will remind the operator to take preventive measures in time or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of bonding and steel leakage.

[0042] In one embodiment of this application, a feature matrix is ​​obtained based on the pulling speed time series and the original thermocouple temperature time series. This includes: filtering outliers from the original thermocouple temperature time series to obtain the thermocouple temperature time series; smoothing and denoising the thermocouple temperature time series using Bézier curves to obtain a smoothed and denoised thermocouple temperature time series; obtaining feature values ​​based on the pulling speed time series and the smoothed and denoised thermocouple temperature time series; and constructing a feature matrix using these feature values. By removing outliers from the original thermocouple temperature time series, the accuracy of the collected data can be ensured. Using Bézier curves to smooth and denoise the thermocouple temperature time series after outlier removal reduces the impact of noise on the collected data, provides more usable temperature features, and facilitates accurate calculation of feature values. This reduces the model's false negative and false positive rates for adhesion phenomena and solves the problem of missing value imputation.

[0043] In one embodiment of this application, the original thermocouple temperature timing sequence includes the original temperature timing sequence of the upper row of thermocouples and the original temperature timing sequence of the lower row of thermocouples. The thermocouple temperature timing sequence includes the temperature timing sequence of the upper row of thermocouples and the temperature timing sequence of the lower row of thermocouples. The thermocouple temperature smoothing and noise reduction timing sequence includes the temperature smoothing and noise reduction timing sequence of the upper row of thermocouples and the temperature smoothing and noise reduction timing sequence of the lower row of thermocouples. The original temperature timing sequence of the upper row of thermocouples is P = {p0, p1, ..., p...} n The original time sequence of the lower row of thermocouple temperatures is Q = {q0, q1, ..., q}. n The pulling speed sequence is S = {s0, s1, ..., s}. n Outliers in the original temperature time series P and Q of the upper and lower rows of thermocouples are filtered out to obtain the temperature time series X and Y of the upper and lower rows of thermocouples. These are then smoothed and denoised using Bézier curves to obtain the smoothed and denoised temperature time series U and V of the upper and lower rows of thermocouples. During thermocouple temperature acquisition, outliers may occur due to equipment delays and other factors. Removing these outliers from the original temperature time series ensures the accuracy of the acquired data. Furthermore, thermocouple temperature acquisition is subject to various noise interferences, such as electromagnetic interference and mechanical vibration. Therefore, using Bézier curves to smooth and denoise the temperature time series after outlier removal reduces the impact of noise on the acquired data, provides more usable temperature features, and facilitates accurate feature value calculation. This reduces the model's false negative and false positive rates for adhesion phenomena and addresses the issue of missing value imputation.

[0044] In one embodiment of this application, the original time sequence of the upper row of thermocouple temperatures is P = {p0, p1, ..., p...} 59 The original time sequence of the lower row of thermocouple temperatures is Q = {q0, q1, ..., q}. 59 The pulling speed sequence is S = {s0, s1, ..., s}.59}. Filter out outliers p in the original time series of the upper row of thermocouple temperatures. 47 The resulting time series can be denoted as X = {x0, x1, ..., x...} 58}. Filter out outliers q in the original time series of the lower row of thermocouple temperatures. 36 The resulting time series can be denoted as Y = {y0, y1, ..., y2}. 58 Without causing ambiguity, the data acquisition times corresponding to the thermocouple temperature time series after outlier removal are all denoted as {w0, w1, ..., w}. 58 The temperature time series X of the upper row of thermocouples and the temperature time series Y of the lower row of thermocouples are smoothed and denoised using Bézier curves to obtain the smoothed and denoised temperature time series U = {u0, u1, ..., u} of the upper row of thermocouples. 59} and the timing sequence for temperature smoothing and noise reduction of the lower row of thermocouples V={v0,v1,…,v 59 During the acquisition of thermocouple temperatures, outliers may occur due to equipment delays and other reasons. Removing outliers from the original thermocouple temperature time series ensures the accuracy of the acquired data. Furthermore, the acquisition process is subject to various noise interferences, such as electromagnetic interference and mechanical vibration. Therefore, using Bézier curves to smooth and denoise the thermocouple temperature time series after outlier removal reduces the impact of noise on the acquired data, provides more usable temperature features, and facilitates accurate feature value calculation. This reduces the model's false negative and false positive rates for adhesion phenomena and also solves the problem of missing value imputation.

[0045] In one embodiment of this application, smoothing and noise reduction of the upper row thermocouple temperature time series X using Bezier curves specifically includes:

[0046]

[0047] Among them, B x (k) represents the smoothing and noise reduction result of the temperature time series X of the upper row of thermocouples. Let be the combination number, x be the temperature time sequence of the upper row of thermocouples, and i be the acquisition time of the vertical axis. Because the value of parameter k in this formula is continuous, the number and specific value of k can be flexibly set according to different needs. In this embodiment, the smoothing and noise reduction processing of the temperature time sequence Y of the lower row of thermocouples using Bezier curves specifically includes:

[0048]

[0049] Among them, B y (k) represents the smoothing and noise reduction result of the temperature time series Y of the lower row of thermocouples. Let be the combination number, y be the temperature time sequence of the lower row of thermocouples, and i be the acquisition time of the vertical axis. Since the value of parameter k in this formula is continuous, the number and specific value of k can be flexibly set according to different needs. In this embodiment, the number of values ​​of parameter k is set to 60, and its specific assignment is based on the following principle: ensuring that the temperature smoothing and noise reduction time sequence is consistent with the original time sequence in the time dimension.

[0050] As an example, feature values ​​related to thermocouple temperature and pulling speed are extracted from the pulling speed time sequence S, the upper row thermocouple temperature smoothing and noise reduction time sequence U, and the lower row thermocouple temperature smoothing and noise reduction time sequence V. These feature values ​​specifically include: upper row smoothing and noise reduction temperature value, upper row temperature slope, upper row temperature directional curvature, lower row temperature slope, lower row temperature directional curvature, difference between upper and lower row smoothing and noise reduction temperature values, pulling speed, upper row temperature rise duration, upper row temperature rise amplitude, ratio of upper row temperature rise amplitude to rise start value, upper row temperature fall duration, upper row temperature fall amplitude, lower row temperature rise duration, lower row temperature rise amplitude, ratio of lower row temperature rise amplitude to rise start value, lower row temperature fall duration, lower row temperature fall amplitude, longitudinal time delay, and the ratio of adhesive longitudinal propagation rate to average pulling speed.

[0051] Specifically, characteristic values ​​related to thermocouple temperature and pulling speed are extracted in the following way: First, the upper row temperature slope is obtained using the following formula:

[0052]

[0053] Among them, B′ X (k) represents the slope of the upper temperature row. Let x be the combination number, i be the temperature time sequence of the upper row of thermocouples, i be the acquisition time of the vertical axis, and k be the number of possible values. The specific assignment of k is based on the following criteria: ensure that the temperature smoothing and noise reduction time sequence is consistent with the original time sequence in the time dimension.

[0054] The slope of the lower temperature row can be obtained using the following formula:

[0055]

[0056] Among them, B′ y (k) represents the slope of the lower temperature row. Let y be the combination number, y be the temperature time sequence of the lower row of thermocouples, i be the acquisition time of the vertical axis, and the number of values ​​for parameter k is set to 60. The specific assignment is based on the following criteria: ensure that the temperature smoothing and noise reduction time sequence is consistent with the original time sequence in the time dimension.

[0057] The directional temperature curvature of the upper row can be obtained using the following formula.

[0058]

[0059] Among them, K x (k) represents the directed curvature of the upper temperature row, w represents the data acquisition time series corresponding to the thermocouple temperature time series after outlier removal, and B″ x (k) represents the temperature slope B′ of the upper row. x (k) is the first derivative of B′ with respect to parameter k. w (k) is the first derivative of the Bézier curve with respect to the parameter k, obtained based on the control point w. For B′ w The squared term of (k), For B′ x The squared term of (k), B″ w (k) is the second derivative of the Bézier curve obtained based on the control point w with respect to the parameter k. The number of values ​​for the parameter k is set to 60. The specific assignment is based on the following criteria: that is, to ensure that the temperature time series and the original time series have the same values ​​in the time dimension.

[0060] The directional temperature curvature of the lower row can be obtained using the following formula.

[0061]

[0062] Among them, K y (k) represents the directed curvature of the lower row temperature, w represents the data acquisition time series corresponding to the thermocouple temperature time series after outlier removal, and B″ y (k) represents the lower row temperature slope B′ y (k) is the first derivative of B′ with respect to parameter k. w (k) is the first derivative of the Bézier curve with respect to the parameter k, obtained based on the control point w. For B′ w The squared term of (k), For B′ x The squared term of (k), B″ w (k) is the second derivative of the Bézier curve obtained based on the control point w with respect to the parameter k. The number of values ​​for parameter k is set to 60. The specific assignment is based on the following criteria: that is, to ensure that the temperature smoothing and noise reduction timing sequence and the original timing sequence have the same value in the time dimension.

[0063] Specifically, B′ is calculated using the following formula. w (k),

[0064]

[0065] Among them, B′ w (k) is the first derivative of the Bézier curve obtained based on control point w with respect to parameter k, and i is the acquisition time of the vertical axis. The number of combinations is w, which is the data acquisition time series corresponding to the thermocouple temperature time series after the outlier. The number of values ​​for parameter k is set to 60. Its specific assignment is based on the following criteria: that is, to ensure that the temperature smoothing and noise reduction time series and the original time series have the same value in the time dimension.

[0066] Calculate B″ using the following formula w (k),

[0067]

[0068] Among them, B″ w (k) is the second derivative of the Bézier curve with respect to parameter k obtained based on the control point w. Let i be the number of combinations, i be the acquisition time of the vertical axis, w be the data acquisition time series corresponding to the thermocouple temperature time series after the outlier, and the number of values ​​for parameter k is set to 60. Its specific assignment is based on the following criteria: that is, to ensure that the temperature smoothing and noise reduction time series and the original time series have the same value in the time dimension.

[0069] Calculate B″ using the following formula x (k),

[0070]

[0071] Among them, B″ x (k) represents the temperature slope B′ of the upper row. x (k) is the first derivative of the parameter k. Let i be the number of combinations, i be the acquisition time of the vertical axis, w be the data acquisition time series corresponding to the thermocouple temperature time series after the outlier, and the number of values ​​for parameter k is set to 60. Its specific assignment is based on the following criteria: that is, to ensure that the temperature smoothing and noise reduction time series and the original time series have the same value in the time dimension.

[0072] The difference in temperature between the upper and lower rows of smoothed noise reduction is calculated using the following formula:

[0073] UV = u i -v i i = 0, 1, ..., 59

[0074] Where U represents the temperature smoothing and noise reduction timing sequence for the upper row of thermocouples, and V represents the temperature smoothing and noise reduction timing sequence for the lower row of thermocouples. i v represents the temperature value at time i in the upper row thermocouple temperature smoothing and noise reduction time series. i This is the temperature value at time i in the temperature smoothing and noise reduction time sequence of the lower row of thermocouples;

[0075] The difference between the peak temperature of the upper row and the start time of the temperature rise of the upper row is determined as the duration of the temperature rise of the upper row; the difference between the peak temperature of the lower row and the start time of the temperature rise of the lower row is determined as the duration of the temperature rise of the lower row.

[0076] The difference between the peak temperature of the upper row and the initial temperature rise of the upper row is determined as the temperature rise rate of the upper row; the difference between the peak temperature of the lower row and the initial temperature rise of the lower row is determined as the temperature rise rate of the lower row.

[0077] The ratio of the temperature rise of the upper row to the initial value of the temperature rise is obtained by dividing the temperature rise of the upper row by the initial value of the temperature rise of the upper row; the ratio of the temperature rise of the lower row to the initial value of the temperature rise is obtained by dividing the temperature rise of the lower row by the initial value of the temperature rise of the lower row.

[0078] t 59 The difference between the time corresponding to the peak temperature of the upper row and the time corresponding to the peak temperature of the upper row is determined as the duration of the temperature drop in the upper row; t 59 The difference between the time corresponding to the peak temperature of the lower row and the time corresponding to the peak temperature of the lower row is determined as the duration of the temperature drop in the lower row.

[0079] The peak temperature of the upper row and the upper row thermocouple at t 59 The difference between the measured temperature values ​​is used to determine the temperature drop of the upper row; the temperature peak of the lower row and the temperature of the lower row thermocouples at t are used to determine the temperature drop. 59 The difference between the measured temperature values ​​is used to obtain the temperature drop of the lower row.

[0080] The longitudinal time delay is obtained by the difference between the start time of the temperature rise in the lower row and the start time of the temperature rise in the upper row.

[0081] The ratio of the longitudinal propagation rate of the bond to the average pulling speed is obtained by using the longitudinal time delay, the average pulling speed during the period from the start of the temperature rise of the upper row to the start of the temperature rise of the lower row, and the distance between the upper and lower row thermocouples. Specifically, the ratio of the distance between the upper and lower row thermocouples to the product of the longitudinal time delay and the average pulling speed during the period from the start of the temperature rise of the upper row to the start of the temperature rise of the lower row is determined as the ratio of the longitudinal propagation rate of the bond to the average pulling speed. In this embodiment, as shown... Figure 5 As shown, Figure 5This is an example diagram of the feature set for the mold sticking and leakage prediction model. In the casting speed time series S, the upper row thermocouple temperature smoothing and noise reduction time series U, and the lower row thermocouple temperature smoothing and noise reduction time series V, feature values ​​related to thermocouple temperature and casting speed time series are obtained through the calculation methods listed above, resulting in an example diagram of the feature value set for the mold sticking and leakage prediction model. As an example, all the calculated feature values ​​are used to construct a 19×60 matrix, obtaining the feature matrix, which serves as the input to the subsequent initial mold sticking and leakage prediction model for training, thus obtaining the mold sticking and leakage prediction model. By organizing the feature values ​​into a feature matrix form, a structured input is provided to the model, facilitating the model to learn more about the impact of features on the accurate identification of mold sticking phenomena and improving the accuracy of model predictions.

[0082] In one embodiment of this application, a model for predicting crystallizer sticking leakage is obtained through model training using a feature matrix. This includes: setting the label values ​​of sticking samples to a first preset value and the label values ​​of non-sticking samples to a second preset value; combining the first and second preset values, stratified sampling of the feature matrix based on a preset data partitioning ratio to obtain a training set, a validation set, and a test set, wherein the training and validation sets are used for model training, and the test set is used to test the trained crystallizer sticking leakage prediction model; fitting the initial crystallizer sticking leakage prediction model to the training set, and adjusting the initial crystallizer sticking leakage prediction model using the validation set to obtain the final crystallizer sticking leakage prediction model. Through clear label definitions, stratified data sampling, and reasonable dataset partitioning, the accuracy, stability, and generalization ability of the crystallizer sticking leakage prediction model are improved. By obtaining the predicted value of sticking probability through the crystallizer sticking leakage prediction model, and determining whether sticking phenomenon exists based on the predicted value of sticking probability of a single thermocouple in each sequence, sticking phenomenon can be accurately and timely identified, avoiding continuous casting machine shutdown due to sticking leakage, improving steel production efficiency, avoiding energy waste, safety risks and environmental pollution caused by sticking leakage, and eliminating the need to constantly adjust threshold parameters based on actual on-site conditions, which helps to reduce human error and time costs.

[0083] As an example, the first preset value is 1, and the second preset value is 0. Each sample is labeled based on whether adhesion occurs. The label value of adhered samples is set to the first preset value "1", and the label value of non-adhesive samples is set to the second preset value "0". The feature matrix extracted in the above embodiment is used to obtain a training set, a validation set, and a test set based on the label value of each sample, using a 7:2:1 stratified sampling method. The training and validation sets are used for model training, and the test set is used to test the trained crystallizer adhesion leakage prediction model. The initial crystallizer adhesion leakage prediction model is fitted to the training set, and then adjusted using the validation set to obtain the final crystallizer adhesion leakage prediction model. Through clear label definitions, stratified data sampling, and reasonable dataset partitioning, the accuracy, stability, and generalization ability of the crystallizer adhesion leakage prediction model are improved. By obtaining the predicted value of sticking probability through the crystallizer sticking leakage prediction model, and determining whether sticking phenomenon exists based on the predicted value of sticking probability of a single thermocouple in each sequence, sticking phenomenon can be accurately and timely identified, avoiding continuous casting machine shutdown due to sticking leakage, improving steel production efficiency, avoiding energy waste, safety risks and environmental pollution caused by sticking leakage, and eliminating the need to constantly adjust threshold parameters based on actual on-site conditions, which helps to reduce human error and time costs.

[0084] In one embodiment of this application, the initial crystallizer adhesion leakage prediction model is a long short-term memory (LSTM) neural network model, and the crystallizer adhesion leakage prediction model is an LSTM crystallizer adhesion leakage prediction model. The loss function uses binary cross-entropy, and the Adam algorithm is selected as the optimizer for model training. The training and validation sets are input into the LSTM model, and the hyperparameters of the model are optimized based on GPGO. The final hyperparameter values ​​of the trained LSTM crystallizer adhesion leakage prediction model are as follows: the number of memory units is set to 512; both L1 and L2 regularization are used, and their weight coefficients are both set to 0.01; the learning rate is set to 0.001; the number of iterations (epochs) is set to 200; and the batch size is set to 128. The trained LSTM crystallizer adhesion leakage prediction model is used to predict the adhesion probability of the test set. The time taken to predict each sample varies from 5 to 35 milliseconds, concentrated around 15 milliseconds. For bonded samples, the predicted bond probability values ​​are all greater than 0.98; for non-bonded samples, the predicted bond probability values ​​are all less than 0.6. This demonstrates that, for this test set, the LSTM crystallizer bond leakage prediction model constructed in this embodiment can completely distinguish between bonded and non-bonded samples.

[0085] In one embodiment of this application, the temperature time sequence of any two rows of thermocouples to be detected in each column within a preset distribution range, extracted in advance or in real time, and their corresponding casting speed time sequence are input into the crystallizer sticking leakage prediction model to obtain multiple single-column thermocouple sticking probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple sticking probability prediction value sequence is generated for each of the multiple single-column thermocouple sticking probability prediction values. This includes acquiring the temperature time sequence of any two rows of thermocouples in some or all columns within the continuous casting machine within a preset time period, as the temperature time sequence of any two rows of thermocouples to be detected in each column within the preset distribution range. The model uses a time-series approach. It inputs the temperature time series of any two rows from each column within a preset distribution range, along with their corresponding pulling speed time series, into the crystallizer adhesion leakage prediction model to obtain multiple predicted values ​​for single-column thermocouple adhesion. These predicted values ​​are then combined to form a sequence of predicted values ​​for single-column thermocouple adhesion probabilities for each moment within a preset time period. Finally, by combining the predicted values ​​for single-column thermocouple adhesion probabilities in each sequence with the lateral transmission characteristic identification criteria of adhesion points, the model determines whether adhesion occurs in the crystallizer. By monitoring and analyzing the time series of predicted values ​​for single-column thermocouple adhesion probabilities, the risk of adhesion leakage can be detected in advance. Furthermore, by comprehensively considering the time series of multiple predicted values ​​and the lateral transmission characteristics of adhesion points, the model more comprehensively reflects the characteristics of adhesion, thus more accurately determining its occurrence.

[0086] In one embodiment of this application, determining whether adhesion exists by combining the predicted adhesion probability value of a single-column thermocouple in each sequence with the lateral transfer characteristic identification criterion of the adhesion point includes: if all predicted adhesion probability values ​​of a single-column thermocouple in at least one sequence are greater than a first preset probability, then the thermocouple column corresponding to the latest time-to-time predicted adhesion probability value sequence is marked as a third preset value; if the thermocouple column corresponding to the third preset value conforms to the lateral transfer characteristic of the adhesion point, then it is determined that adhesion exists in the crystallizer, and an alarm message is issued; if the thermocouple column corresponding to the third preset value does not conform to the lateral transfer characteristic of the adhesion point, then it is determined that adhesion does not exist in the crystallizer. By comprehensively considering the time sequence of multiple predicted adhesion probability values ​​of single-column thermocouples and the lateral transfer characteristic of the adhesion point, the characteristics of adhesion can be more comprehensively reflected, thereby more accurately determining the occurrence of adhesion.

[0087] In one embodiment of this application, an LSTM crystallizer sticking and leakage prediction model trained based on historical production data of steel plant A is used for real-time detection of sticking phenomena at steel plant B. 107 production records from steel plant B are collected, each covering a time range of 5 minutes. For each record, the aforementioned trained LSTM crystallizer sticking and leakage prediction model is used to detect sticking phenomena in real-time for three combinations of thermocouple rows: top-middle row, top-bottom row, and middle-bottom row. The detection of sticking phenomena in different combinations is independent of each other, where the top, middle, and bottom rows of thermocouples correspond to the first, second, and third rows of thermocouples, respectively. As an example, the temperature sequences of the thermocouples to be detected are the temperature sequences of the upper and lower rows of thermocouples. An LSTM crystallizer bonding leakage prediction model is used to detect the temperature sequences of the upper and lower rows of thermocouples in real time. For each column of thermocouples, a single-column bonding probability prediction value is output. In this embodiment, based on the thermocouple bonding probability prediction values ​​that are continuous across three time dimensions, the bonding probability prediction value of each column of thermocouples output at the first time is recorded as P1, the bonding probability prediction value of each column of thermocouples output at the second time is recorded as P2, and the bonding probability prediction value of each column of thermocouples output at the third time is recorded as P3. The predicted probability value is denoted as P3. In this embodiment, the third preset value is 1, and the first preset probability is 0.965. If the predicted probability sequence of thermocouple adhesion values ​​with consecutive values ​​in the three time dimensions is greater than the first preset probability "0.965", then the thermocouple column corresponding to the latest thermocouple adhesion prediction value sequence P3 is marked as the third preset value "1". If the thermocouple column corresponding to the third preset value "1" conforms to the lateral transmission characteristics of the adhesion point, then it is determined that adhesion exists in the crystallizer, and an alarm message is issued. If the thermocouple column corresponding to the third preset value "1" does not conform to the lateral transmission characteristics of the adhesion point, then it is determined that adhesion does not exist in the crystallizer. By combining the time series of predicted probability values ​​of adhesion from multiple single-column thermocouples and the lateral transmission characteristics of the adhesion point, the accuracy of determining adhesion can be improved. Time series data can reveal the trend of adhesion over time, while the lateral transmission characteristics of the adhesion point reflect the spatial expansion of adhesion. By comprehensively considering the time series data of predicted probability values ​​of adhesion from multiple single-column thermocouples and the lateral transmission characteristics of the adhesion point, the characteristics of adhesion can be more comprehensively reflected, thereby more accurately determining the occurrence of adhesion. On the other hand, by monitoring and analyzing the time series of predicted adhesion probability values ​​for single-row thermocouples, the risk of adhesion leakage can be detected in advance. By combining multiple predicted adhesion probability value sequences of single-row thermocouples with the identification criteria for lateral transmission characteristics of adhesion points, alarm signals can be rationally designed. The triggering of the alarm signal will remind operators to take preventive measures in a timely manner or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of adhesion leakage.

[0088] In one embodiment of this application, determining whether adhesion exists by combining the predicted adhesion probability value of a single-column thermocouple in each sequence with the lateral transfer characteristic identification criterion of adhesion points includes: acquiring the predicted adhesion probability value of each column of thermocouples at consecutive time intervals to obtain the predicted adhesion probability value of a single-column thermocouple at a first time interval, a second time interval, and a third time interval for each column of thermocouples; if the predicted adhesion probability value of a single-column thermocouple at the first time interval and the predicted adhesion probability value of a single-column thermocouple at the second time interval are both greater than a second preset probability, and the predicted adhesion probability value of a single-column thermocouple at the third time interval is greater than a third preset probability, then the thermocouple column corresponding to the predicted adhesion probability value of a single-column thermocouple at the third time interval is marked with a third preset value; if the thermocouple column corresponding to the third preset value conforms to the lateral transfer characteristic of adhesion points, then it is determined that adhesion exists in the crystallizer, and an alarm message is issued; if the thermocouple column corresponding to the third preset value does not conform to the lateral transfer characteristic of adhesion points, then it is determined that adhesion does not exist in the crystallizer. By combining the time sequence of multiple predicted adhesion probability values ​​of single-column thermocouples with the lateral transfer characteristic of adhesion points, the accuracy of determining adhesion can be improved. Time-series data can reveal the temporal trend of adhesion phenomena, while the lateral propagation characteristics of adhesion points reflect the spatial expansion of adhesion phenomena. By comprehensively considering the time-series data of multiple single-row thermocouple adhesion probability predictions and the lateral propagation characteristics of adhesion points, the characteristics of adhesion phenomena can be more comprehensively reflected, thereby more accurately determining the occurrence of adhesion phenomena.

[0089] In one embodiment of this application, an LSTM crystallizer sticking and leakage prediction model trained based on historical production data of steel plant A is used for real-time detection of sticking phenomena at steel plant B. 107 production records from steel plant B are collected, each covering a time range of 5 minutes. For each record, the aforementioned trained LSTM crystallizer sticking and leakage prediction model is used to detect sticking phenomena in real-time for three combinations of thermocouple rows: top-middle row, top-bottom row, and middle-bottom row. The detection of sticking phenomena in different combinations is independent of each other, where the top, middle, and bottom rows of thermocouples correspond to the first, second, and third rows of thermocouples, respectively. As an example, the temperature sequences of the thermocouples to be detected are the temperature sequences of the upper and lower rows of thermocouples. An LSTM crystallizer bonding leakage prediction model is used to detect the temperature sequences of the upper and lower rows of thermocouples in real time. For each column of thermocouples, a single-column bonding probability prediction value is output. In this embodiment, based on the thermocouple bonding probability prediction values ​​that are continuous across three time dimensions, the bonding probability prediction value of each column of thermocouples output at the first time is recorded as P1, the bonding probability prediction value of each column of thermocouples output at the second time is recorded as P2, and the bonding probability prediction value of each column of thermocouples output at the third time is recorded as P3. In this embodiment, the third preset value is 1, and the second preset value is... With a probability of 0.95 and a third preset probability of 0.98, if the predicted thermocouple adhesion probability sequence P1 is greater than the second preset probability "0.95", the predicted thermocouple adhesion probability sequence P2 is greater than the second preset probability "0.95", and the predicted thermocouple adhesion probability sequence P3 is greater than the third preset probability "0.98", then the thermocouple column corresponding to the latest predicted thermocouple adhesion probability sequence P3 is marked as the third preset value "1". If the thermocouple column corresponding to the third preset value "1" conforms to the lateral transmission characteristics of the adhesion point, then it is determined that adhesion exists in the crystallizer, and an alarm message is issued. If the thermocouple column corresponding to the third preset value "1" does not conform to the lateral transmission characteristics of the adhesion point, then it is determined that adhesion does not exist in the crystallizer. By combining the time series of multiple single-column predicted thermocouple adhesion probability values ​​and the lateral transmission characteristics of the adhesion point, the accuracy of determining adhesion can be improved. Time series data can reveal the trend of adhesion over time, while the lateral transmission characteristics of the adhesion point reflect the spatial expansion of the adhesion phenomenon. By comprehensively considering the time series data of multiple single-row thermocouple bonding probability predictions and the lateral transmission characteristics of bonding points, the characteristics of bonding phenomena can be reflected more comprehensively, thus more accurately determining the occurrence of bonding phenomena. On the other hand, by monitoring and analyzing the time series of single-row thermocouple bonding probability predictions, the risk of bonding leakage can be detected in advance. Combining the bonding probability prediction sequences of multiple single-row thermocouples with the identification criteria of lateral transmission characteristics of bonding points, alarm signals can be rationally designed. The triggering of the alarm signal reminds operators to take preventive measures in a timely manner or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of bonding leakage.

[0090] In one embodiment of this application, if the marked value matches the lateral propagation characteristics of the bonding point, then the existence of bonding is determined. This includes: extracting the thermocouple column corresponding to the third preset value; obtaining the lateral propagation rate of the bonding point based on the time difference of the thermocouple column corresponding to the third preset value and the corresponding average pulling speed; obtaining the detection time through the lateral propagation rate of the bonding point and the arrangement spacing between the thermocouple columns corresponding to the third preset value; if the number of thermocouple columns corresponding to the third preset value is greater than or equal to a fourth preset value within the detection time, then the crystallizer is determined to have bonding; if the number of thermocouple columns corresponding to the third preset value is less than the fourth preset value within the detection time, then the crystallizer is determined not to have bonding. By combining the time series of multiple single-column thermocouple bonding probability prediction values ​​and the lateral propagation characteristics of the bonding point, the accuracy of determining the bonding phenomenon can be improved. Time series data can reveal the trend of bonding phenomena over time, while the lateral propagation characteristics of the bonding point reflect the spatial expansion of the bonding phenomenon. By comprehensively considering the time series data of multiple single-column thermocouple bonding probability prediction values ​​and the lateral propagation characteristics of the bonding point, the characteristics of the bonding phenomenon can be more comprehensively reflected, thereby more accurately determining the occurrence of the bonding phenomenon.

[0091] In one embodiment of this application, the lateral transmission characteristic of adhesion points refers to the expansion and transmission of this adhesion phenomenon along the width direction (i.e., laterally) of the crystallizer when adhesion occurs between molten steel and the copper wall of the crystallizer. This transmission characteristic can be influenced by various factors, such as the flow of molten steel, the temperature distribution of the copper wall of the crystallizer, and crystallizer vibration. The lateral transmission characteristic of adhesion points can typically be analyzed and judged by observing and monitoring the predicted adhesion probability values ​​of thermocouples at different locations within the crystallizer. When the thermocouple at a certain location detects an adhesion probability exceeding a certain threshold, this adhesion phenomenon may gradually expand along the width direction of the crystallizer, i.e., the adhesion point is transmitted laterally.

[0092] In one embodiment of this application, the fourth preset value is 2. Thermocouple columns marked with the third preset value "1" are extracted. In this embodiment, two thermocouple columns marked with the third preset value "1" are extracted. The average pulling speed Q corresponding to the time difference between the two thermocouple columns marked with the third preset value "1" is calculated. The lateral bonding transmission rate V = 0.9 × Q. The detection time is obtained by using the lateral bonding transmission rate and the arrangement spacing between the thermocouple columns corresponding to the third preset value "1". If, within the detection time, the number of thermocouple columns marked with the third preset value "1" among three adjacent thermocouple columns is greater than or equal to the fourth preset value "2", that is, at least two thermocouple columns are marked with the third preset value "1", then it is determined that the crystallizer has bonding. If the number of thermocouple columns marked with the third preset value "1" is less than the fourth preset value "2", then it is determined that the crystallizer does not have bonding. By combining the timing of multiple single-column thermocouple bonding probability prediction values ​​and the lateral transmission characteristics of the bonding points, the accuracy of determining bonding phenomena can be improved. Time-series data can reveal the temporal trend of adhesion phenomena, while the lateral transmission characteristics of adhesion points reflect the spatial expansion of adhesion phenomena. By comprehensively considering the time-series data of adhesion probability predictions from multiple single-row thermocouples and the lateral transmission characteristics of adhesion points, the characteristics of adhesion phenomena can be more comprehensively reflected, thus more accurately determining the occurrence of adhesion phenomena. On the other hand, by monitoring and analyzing the time-series of adhesion probability predictions from single-row thermocouples, the risk of adhesion leakage can be detected in advance. Combining the adhesion probability prediction sequences of multiple single-row thermocouples with the identification criteria of lateral transmission characteristics of adhesion points, alarm signals can be rationally designed. The triggering of the alarm signal reminds operators to take preventive measures in a timely manner or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of adhesion leakage.

[0093] Please see Figure 6 , Figure 6 This is a flowchart illustrating an exemplary embodiment of the present application of a method for predicting crystallizer sticking and leakage. For example... Figure 6 As shown, in an exemplary embodiment, the method for predicting crystallizer sticking and steel leakage includes at least steps S601 to S610, which are described in detail below:

[0094] Step S601: Obtain the original thermocouple temperature time series and the pulling speed time series corresponding to each original sample. The original samples include the bonded samples of the billet and the crystallizer and the non-bonded samples of the billet and the crystallizer. Then, execute step S602.

[0095] Step S602: Filter out the abnormal values ​​in the original thermocouple temperature time series to obtain the thermocouple temperature time series. Use Bezier curves to smooth and reduce noise in the thermocouple temperature time series to obtain the thermocouple temperature smoothed and noise-reduced time series. Then execute step S603.

[0096] Step S603: Obtain eigenvalues ​​based on the pulling speed timing and thermocouple temperature smoothing and noise reduction timing, construct a feature matrix using the eigenvalues, and then execute step S603.

[0097] Step S603: Set the label value of the bonded sample to the first preset value, set the label value of the non-bonded sample to the second preset value, and then execute step S604.

[0098] Step S604: Combining the first preset value and the second preset value, the feature matrix is ​​stratified and sampled according to the preset data division ratio to obtain the training set, the validation set and the test set, and the model is trained to obtain the crystallizer sticking and leakage steel prediction model, and then step S605 is executed.

[0099] Step S605: Obtain the temperature timing sequence of any two rows of thermocouples in some or all columns of the continuous casting machine as the temperature timing sequence of any two rows of thermocouples to be tested in each column within the preset distribution range. Input the temperature timing sequence of any two rows of thermocouples to be tested in each column within the preset distribution range and their corresponding casting speed timing sequence into the mold sticking leakage prediction model to obtain multiple single-column thermocouple sticking probability prediction values, and then execute step S606.

[0100] Step S606: Obtain the predicted bonding probability value of each column of thermocouples within consecutive time intervals, and obtain the predicted bonding probability value of a single column of thermocouples at the first time interval, the predicted bonding probability value of a single column of thermocouples at the second time interval, and the predicted bonding probability value of a single column of thermocouples at the third time interval for each column of thermocouples, and then execute step S607.

[0101] Step S607: If the predicted probability of single-column thermocouple bonding at the first time and the predicted probability of single-column thermocouple bonding at the second time are both greater than the second preset probability, and the predicted probability of single-column thermocouple bonding at the third time is greater than the third preset probability, then mark the thermocouple column corresponding to the predicted probability of single-column thermocouple bonding at the third time as the third preset value, and then execute step S608.

[0102] Step S608: Determine whether the thermocouple array corresponding to the third preset value conforms to the lateral transfer characteristics of the bonding point. If the thermocouple array corresponding to the third preset value conforms to the lateral transfer characteristics of the bonding point, then proceed to step S609. If the thermocouple array corresponding to the third preset value does not conform to the lateral transfer characteristics of the bonding point, then proceed to step S610.

[0103] Step S609: Confirm that there is adhesion in the crystallizer, issue an alarm message, and then end.

[0104] In step S610, it is confirmed that there is no adhesion in the crystallizer, and then the process ends.

[0105] exist Figure 6In the technical solution shown, the original thermocouple temperature time series and the corresponding casting speed time series for each sample are obtained from the bonded and non-bonded samples of the billet and the crystallizer. A feature matrix is ​​obtained based on the casting speed time series and the thermocouple temperature smoothing and noise reduction time series. The model is trained using the feature matrix to obtain the crystallizer bonding leakage prediction model. The temperature time series of any two rows of the thermocouples to be detected in each column within the preset distribution range and their corresponding casting speed time series are input into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple bonding probability prediction value sequence is formed for the multiple single-column thermocouple bonding probability prediction values. The presence of bonding phenomenon is determined by combining the single-column thermocouple bonding probability prediction value in each sequence with the lateral transmission feature recognition criterion of the bonding point. By using a neural network model to obtain the predicted adhesion probability of a single-row thermocouple, and comprehensively considering the temporal sequence of multiple single-row thermocouple adhesion probability predictions and the lateral transmission characteristics of adhesion points, this approach is suitable for scenarios where a particular row of thermocouples in a multi-row thermocouple system is disabled. This helps reduce the false negative rate of adhesion phenomena and also helps reduce the number of thermocouple rows installed, saving costs. Compared to traditional logical threshold judgment methods, this approach can more accurately and promptly identify adhesion phenomena in various application scenarios, such as different types of temperature change characteristics and multiple production sites, preventing continuous casting machines from shutting down due to adhesion and steel leakage, improving steel production efficiency, and avoiding energy waste, safety risks, and environmental pollution caused by adhesion and steel leakage. Furthermore, it eliminates the need for continuous adjustment of threshold parameters based on actual on-site conditions, helping to reduce human error and time costs. On the other hand, by monitoring and analyzing the temporal sequence of the predicted adhesion probability of a single-row thermocouple, the risk of adhesion and steel leakage can be detected in advance. By combining multiple single-row thermocouple bonding probability prediction value sequences with the identification criteria of lateral transmission characteristics of bonding points, alarm signals can be reasonably designed. The triggering of the alarm signal will remind the operator to take preventive measures in time or directly reduce the pulling speed and stop the operation, thereby avoiding the occurrence of bonding and steel leakage.

[0106] In one embodiment of this application, Figure 7 This is a comparison chart of marked time points illustrating the adhesion phenomenon in an exemplary embodiment of this application. For the test samples collected from steel mill B, the adhesion phenomenon was detected in real time using an LSTM crystallizer adhesion and leakage prediction model trained based on historical production data from steel mill A. The results showed zero missed detections and zero false detections. Furthermore, for... Figures 3-5 The three representative adhesion samples shown illustrate the timing of adhesion phenomena for the LSTM crystallizer adhesion leakage prediction model and the logic threshold judgment rule used by steel plant B. Figure 7 .Depend on Figure 7Analysis reveals the following: Firstly, compared to traditional logical threshold judgment methods, the LSTM crystallizer sticking and leakage prediction model identifies sticking phenomena more promptly. Secondly, although the LSTM crystallizer sticking and leakage prediction model is trained using the temperature information of the first and second rows of thermocouples, it can still accurately and promptly determine sticking phenomena even when the temperature sequence of any two rows of thermocouples is taken. Combining the test results of the LSTM crystallizer sticking and leakage prediction model at steel plant B and its comparison with the logical threshold judgment method for identifying sticking phenomena, it is evident that compared to traditional logical threshold judgment methods, the deep learning-based crystallizer sticking and leakage prediction method provided in this patent can identify sticking phenomena more accurately and efficiently.

[0107] Please see Figure 8 , Figure 8 This is a block diagram of a crystallizer sticking and leaking steel prediction device, which is an exemplary embodiment of this application.

[0108] like Figure 8As shown, the exemplary crystallizer sticking and leakage prediction device includes: a timing acquisition module 810, a model training module 820, a timing detection module 830, and a sticking determination module 840. The system includes the following modules: a timing acquisition module for acquiring the original timing of thermocouple temperatures and the casting speed timing corresponding to each original sample; original samples include bonded samples of the billet and the copper plate of the crystallizer, and non-bonded samples of the billet and the copper plate of the crystallizer; a model training module for obtaining a feature matrix based on the casting speed timing and the original timing of thermocouple temperatures, and training the model using the feature matrix to obtain a crystallizer bonding leakage prediction model; a timing detection module for inputting the timing of any two rows of thermocouple temperatures to be detected and their corresponding casting speed timing from a preset distribution range (extracted in advance or in real time) into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values, and forming a corresponding single-column thermocouple bonding probability prediction value sequence for each of the multiple single-column thermocouple bonding probability prediction values ​​based on continuous detection over a preset time period; and a bonding determination module for determining whether bonding exists in the crystallizer by combining the single-column thermocouple bonding probability prediction value in each single-column thermocouple bonding probability prediction value sequence with the lateral transfer feature recognition criterion of bonding points. The original thermocouple temperature time series and the corresponding casting speed time series for each sample are obtained from the bonded and non-bonded samples of the billet and the crystallizer. A feature matrix is ​​obtained based on the casting speed time series and the original thermocouple temperature time series. The model is trained using the feature matrix to obtain a crystallizer bonding leakage prediction model. The temperature time series of any two rows of the thermocouples to be detected in each column within a preset distribution range and their corresponding casting speed time series are input into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple bonding probability prediction value sequence is formed for each of the multiple single-column thermocouple bonding probability prediction values. The presence of bonding phenomenon is determined by combining the single-column thermocouple bonding probability prediction value in each sequence with the lateral transmission feature recognition criterion of the bonding point. By using a neural network model to obtain the predicted probability of adhesion of a single thermocouple, and combining the predicted probability of adhesion of a single thermocouple in each sequence with the identification criteria of the lateral transmission characteristics of the adhesion point, it is possible to determine whether adhesion occurs. This can accurately and timely identify adhesion, avoid continuous casting machine shutdown due to adhesion and steel leakage, improve steel production efficiency, avoid energy waste, safety risks and environmental pollution caused by adhesion and steel leakage, and eliminate the need to constantly adjust threshold parameters based on actual on-site conditions, which helps to reduce human error and time costs.

[0109] In an exemplary embodiment of the present invention, the timing detection module is used to acquire the temperature timing of any two rows of thermocouples in some or all columns within a preset time period, as the temperature timing of any two rows of thermocouples to be detected in each column within a preset distribution range; input the temperature timing of any two rows of thermocouples to be detected in each column within the preset distribution range into the mold sticking and leakage prediction model to obtain multiple single-column thermocouple sticking probability prediction values; combine the single-column thermocouple sticking probability prediction values ​​corresponding to each moment within the preset time period to form a single-column thermocouple sticking probability prediction value sequence to obtain multiple single-column thermocouple sticking probability prediction value sequences; and combine the single-column thermocouple sticking probability prediction value in each single-column thermocouple sticking probability prediction value sequence with the lateral transmission feature identification criterion of the sticking point to determine whether there is sticking in the mold.

[0110] In an exemplary embodiment of the present invention, the model training module is used to filter out outliers in the original thermocouple temperature time series to obtain the thermocouple temperature time series; the thermocouple temperature time series is smoothed and denoised using a Bezier curve to obtain a smoothed and denoised thermocouple temperature time series; feature values ​​are obtained based on the pulling speed time series and the smoothed and denoised thermocouple temperature time series, and a feature matrix is ​​constructed using the feature values.

[0111] In an exemplary embodiment of the present invention, the model training module sets the label value of the bonded sample to a first preset value and the label value of the non-bonded sample to a second preset value; combining the first preset value and the second preset value, the feature matrix is ​​stratified and sampled according to a preset data division ratio to obtain a training set, a validation set, and a test set, wherein the training set and the validation set are used for model training, and the test set is used to test the trained crystallizer bonded leakage prediction model; the initial crystallizer bonded leakage prediction model is fitted to the training set, and the initial crystallizer bonded leakage prediction model is adjusted through the validation set to obtain the crystallizer bonded leakage prediction model.

[0112] In an exemplary embodiment of the present invention, the adhesion determination module is configured to mark the thermocouple column corresponding to the latest thermocouple adhesion probability prediction value sequence as a third preset value if all the predicted adhesion probabilities of the single-column thermocouples in at least one single-column thermocouple adhesion probability prediction value sequence are greater than a first preset probability; if the thermocouple column corresponding to the third preset value conforms to the lateral transmission characteristics of the adhesion point, it is determined that the crystallizer has an adhesion phenomenon and an alarm message is issued; if the thermocouple column corresponding to the third preset value does not conform to the lateral transmission characteristics of the adhesion point, it is determined that the crystallizer does not have an adhesion phenomenon.

[0113] In an exemplary embodiment of the present invention, the adhesion determination module is further configured to extract the thermocouple array corresponding to the third preset value, and obtain the lateral propagation rate of the adhesion point based on the time difference of the thermocouple array corresponding to the third preset value and the average pulling speed therefore; obtain the detection time by using the lateral propagation rate of the adhesion point and the arrangement spacing between the thermocouple arrays corresponding to the third preset value; if the number of thermocouple arrays corresponding to the third preset value is greater than or equal to the fourth preset value within the detection time, it is determined that the crystallizer has an adhesion phenomenon; if the number of thermocouple arrays corresponding to the third preset value is less than the fourth preset value within the detection time, it is determined that the crystallizer does not have an adhesion phenomenon.

[0114] This embodiment also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the electronic device to implement the crystallizer sticking and steel leakage prediction method provided in the above embodiments.

[0115] Please see Figure 9 , Figure 9 This is a schematic diagram illustrating the structure of an electronic device according to an exemplary embodiment of this application. It should be noted that... Figure 9 The electronic device 900 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0116] like Figure 9 As shown, the electronic device 900 includes a processor 901, a memory 902, and a communication bus 903; the communication bus 903 is used to connect the processor 901 and the memory 902; the processor 901 is used to execute a computer program stored in the memory 902 to implement one or more methods as described in the above embodiments.

[0117] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When executed by a computer processor, the computer program causes the computer to perform the aforementioned method for predicting crystallizer sticking and leakage. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently without being assembled into the electronic device. This embodiment also provides a computer program product or computer program including computer instructions stored in the computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the crystallizer sticking and leakage prediction method provided in the various embodiments above.

[0118] The electronic device provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic device performs the various steps of the above method.

[0119] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0120] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0121] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for predicting steel leakage due to crystallizer adhesion, characterized in that, The method includes: Obtain the original thermocouple temperature time series and the pulling speed time series corresponding to each original sample. The original samples include the bonded samples of the billet and the crystallizer and the non-bonded samples of the billet and the crystallizer. A feature matrix is ​​obtained based on the casting speed time series and the original thermocouple temperature time series. The model is then trained using this feature matrix to obtain a crystallizer sticking and leakage prediction model. The process of obtaining the feature matrix includes: filtering outliers from the original thermocouple temperature time series to obtain the thermocouple temperature time series; smoothing and denoising the thermocouple temperature time series using Bezier curves to obtain a smoothed and denoised thermocouple temperature time series; obtaining eigenvalues ​​from the casting speed time series and the smoothed and denoised thermocouple temperature time series, and constructing a feature matrix using these eigenvalues; and training the model using the feature matrix to obtain a crystallizer sticking and leakage prediction model. A leak-out prediction model for crystallizers includes: setting the label value of the bonded sample to a first preset value and setting the label value of the non-bonded sample to a second preset value; combining the first preset value and the second preset value, stratifying the feature matrix based on a preset data division ratio to obtain a training set, a validation set, and a test set, wherein the training set and validation set are used for model training, and the test set is used to test the trained crystallizer leak-out prediction model; fitting the initial crystallizer leak-out prediction model to the training set, and adjusting the initial crystallizer leak-out prediction model using the validation set to obtain the crystallizer leak-out prediction model; The temperature time sequence of any two rows of thermocouples to be detected in each column and their corresponding pulling speed time sequence within the preset distribution range are extracted in advance or in real time and input into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple bonding probability prediction value sequence is generated for each of the multiple single-column thermocouple bonding probability prediction values. The presence of adhesion in the crystallizer is determined by combining the predicted adhesion probability values ​​of each individual thermocouple adhesion probability sequence with the lateral transfer feature identification criterion of adhesion points.

2. The method for predicting crystallizer sticking and steel leakage according to claim 1, characterized in that, The temperature time series of any two rows of thermocouples to be detected in each column within a preset distribution range, extracted in advance or in real time, and their corresponding pulling speed time series are input into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, a corresponding single-column thermocouple bonding probability prediction value sequence is generated for each of the multiple single-column thermocouple bonding probability prediction values, including: Within a preset time period, the temperature timing sequence of any two rows of thermocouples in some or all columns of the continuous casting machine is obtained, and used as the temperature timing sequence of any two rows of thermocouples to be detected in each column within the preset distribution range. Input the temperature time sequence of any two rows of thermocouples to be detected in each column within the preset distribution range and their corresponding pulling speed time sequence into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. The predicted values ​​of the single-column thermocouple bonding probability at each moment within the preset time period are combined to form a single-column thermocouple bonding probability prediction value sequence, resulting in multiple single-column thermocouple bonding probability prediction value sequences. The presence of adhesion in the crystallizer is determined by combining the predicted adhesion probability values ​​of each individual thermocouple adhesion probability sequence with the lateral transfer feature identification criterion of adhesion points.

3. The method for predicting crystallizer sticking and steel leakage according to claim 1, characterized in that, The presence of adhesion is determined by combining the predicted adhesion probability of a single thermocouple in each sequence with the lateral transfer feature identification criteria of the adhesion point, including: If all the predicted values ​​of single-column thermocouple bonding probability in at least one single-column thermocouple bonding probability prediction value sequence are greater than the first preset probability, then the thermocouple column corresponding to the single-column thermocouple bonding probability prediction value sequence at the latest time is marked as the third preset value. If the thermocouple array corresponding to the third preset value conforms to the lateral transmission characteristics of the bonding point, it is determined that the crystallizer has a bonding phenomenon and an alarm message is issued. If the thermocouple array corresponding to the third preset value does not conform to the lateral transmission characteristics of the bonding point, it is determined that the crystallizer does not have a bonding phenomenon.

4. The method for predicting crystallizer sticking and steel leakage according to claim 1, characterized in that, The presence of adhesion is determined by combining the predicted adhesion probability of a single thermocouple in each sequence with the lateral transfer feature identification criteria of the adhesion point, including: The predicted bonding probability of each column of thermocouples is obtained in consecutive time intervals, and the predicted bonding probability of each column of thermocouples at the first time interval, the second time interval, and the third time interval are obtained. If the predicted probability of single-column thermocouple bonding at the first time and the predicted probability of single-column thermocouple bonding at the second time are both greater than the second preset probability, and the predicted probability of single-column thermocouple bonding at the third time is greater than the third preset probability, then the thermocouple column corresponding to the predicted probability of single-column thermocouple bonding at the third time is marked with the third preset value. If the thermocouple array corresponding to the third preset value conforms to the lateral transmission characteristics of the bonding point, it is determined that the crystallizer has a bonding phenomenon and an alarm message is issued. If the thermocouple array corresponding to the third preset value does not conform to the lateral transmission characteristics of the bonding point, it is determined that the crystallizer does not have a bonding phenomenon.

5. The method for predicting crystallizer sticking and steel leakage according to claim 3 or 4, characterized in that, If the third preset value matches the lateral transfer characteristics of the bonding point, then bonding is determined to exist, including: Extract the thermocouple array corresponding to the third preset value, and obtain the lateral propagation rate of the bonding point based on the time difference of the thermocouple array corresponding to the third preset value and the average pulling speed. The detection time is obtained by measuring the lateral propagation rate at the bonding point and the spacing between the thermocouple rows corresponding to the third preset value. If the number of thermocouple rows corresponding to the third preset value within the detection time is greater than or equal to the fourth preset value, it is determined that the crystallizer has an adhesion phenomenon; if the number of thermocouple rows corresponding to the third preset value within the detection time is less than the fourth preset value, it is determined that the crystallizer does not have an adhesion phenomenon.

6. A crystallizer sticking and leak prediction device, characterized in that, The device includes: The timing acquisition module is used to acquire the original timing of thermocouple temperature and the pulling speed timing corresponding to each original sample; the original samples include the bonded samples of the billet and the copper plate of the crystallizer, and the non-bonded samples of the billet and the copper plate of the crystallizer. The model training module is used to obtain a feature matrix based on the casting speed time series and the original thermocouple temperature time series, and to train the model using the feature matrix to obtain a crystallizer sticking leakage prediction model. Specifically, obtaining the feature matrix based on the casting speed time series and the original thermocouple temperature time series includes: filtering outliers in the original thermocouple temperature time series to obtain the thermocouple temperature time series; smoothing and denoising the thermocouple temperature time series using Bézier curves to obtain a smoothed and denoised thermocouple temperature time series; obtaining eigenvalues ​​based on the casting speed time series and the smoothed and denoised thermocouple temperature time series, and constructing a feature matrix using the eigenvalues; and training the model using the feature matrix to obtain... The method for predicting crystallizer sticking leakage includes: setting the label value of the sticking sample to a first preset value and setting the label value of the non-sticking sample to a second preset value; combining the first preset value and the second preset value, stratifying the feature matrix based on a preset data division ratio to obtain a training set, a validation set, and a test set, wherein the training set and validation set are used for model training, and the test set is used to test the trained crystallizer sticking leakage prediction model; fitting the initial crystallizer sticking leakage prediction model to the training set, and adjusting the initial crystallizer sticking leakage prediction model using the validation set to obtain the crystallizer sticking leakage prediction model; The timing detection module is used to input the temperature timing of any two rows of thermocouples to be detected in each column and their corresponding pulling speed timing from a preset distribution range, which are extracted in advance or in real time, into the crystallizer bonding leakage prediction model to obtain multiple single-column thermocouple bonding probability prediction values. Based on continuous detection over a preset time period, the module forms a corresponding single-column thermocouple bonding probability prediction value sequence for each of the multiple single-column thermocouple bonding probability prediction values. The adhesion determination module is used to determine whether the crystallizer has adhesion by combining the predicted adhesion probability values ​​of each single-row thermocouple in the single-row thermocouple adhesion probability prediction value sequence with the lateral transmission feature identification criterion of the adhesion point.

7. An electronic device, characterized in that, The electronic device includes: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the crystallizer sticking leak prediction method as described in any one of claims 1 to 5.

8. A computer-readable medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the crystallizer sticking leak prediction method according to any one of claims 1 to 5.

Images