An off-line data analysis system for continuous casting quality determination

Abstract

The application discloses a continuous casting quality judgment off-line data analysis system, comprising an off-line data analysis module and an off-line data verification module; the off-line data analysis module comprises a data acquisition module for acquiring various data in a continuous casting process, a data preprocessing module for labeling data based on a set rule, a feature extraction module for classifying the labeled data and acquiring data features, a mechanism model module for model building in different stages of the continuous casting process according to a mechanism model mode and acquiring the mechanism model, and a model training module for inputting the data features into the corresponding mechanism model for training; the off-line data verification module is used for verifying the trained mechanism model with data. The application can comprehensively collect and analyze continuous casting quality data and simultaneously make corresponding quality judgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data analysis for continuous casting quality determination, and in particular to an offline data analysis system for continuous casting quality determination. BACKGROUND

[0002] At present, some of the patents applied in China are mainly directed to offline analysis systems for continuous casting quality determination, but there is no offline analysis system for the whole line. For example, patent 201310130078.0 discloses a test bed for offline detection of the surface of continuous casting billets, which comprises a trolley motion control platform, a test bed motion platform, a mobile trolley platform, and an imaging system. The trolley motion control platform drives the test bed motion; the test bed motion platform is used to fix the mobile trolley and make it slidably installed on the track; the mobile trolley platform is movably installed on the test motion platform and is used to place and level the continuous casting billet sample plate; and the imaging system is installed across the test motion platform and moves synchronously. The test bed can make the trolley move out of the motion track, adjust the bottom of the continuous casting billet horizontally, and ensure that the imaging position meets the depth of field requirements of the camera. The imaging system can move and adjust along the direction of the test bed motion track, and the camera can image within the full length and full width range of the sample plate. The test bed is reliable in operation, easy to use, and can realize offline and online detection of thick plate low magnification. This technology mainly uses an offline test bed to perform surface offline detection, and detects through hardware, which cannot be applied to offline data analysis in the online production line environment.

[0003] Patent 200610117522.5 discloses an offline simulation system for slab continuous casting secondary cooling and dynamic soft reduction, which mainly solves the technical problem of high cost of adjusting dynamic soft reduction parameters in actual production. The computer of the simulation system contains: a simulator kernel unit for simulating and modeling main pouring condition process parameters; a process model parameter setting unit for setting process and control parameters in the model; a model calculation kernel unit for calculating the process model in combination with the simulated pouring condition parameters; a monitoring and display unit for mainly displaying the simulation results; the simulator kernel unit and the model calculation kernel unit realize the interaction of process parameter data through shared memory; the model calculation kernel unit and the monitoring and display unit interact simulation result data through an ETHERNET local area network; and the model calculation kernel unit and the process model parameter setting unit realize the interaction of process model parameters through a model parameter database. The present application is mainly used for simulating and modeling the specific control process of slab continuous casting secondary cooling water distribution and dynamic soft reduction. This technology is mainly directed to offline simulation of the secondary cooling water distribution area and the dynamic soft reduction area, and is not suitable for offline analysis systems for the whole line. SUMMARY

[0004] In view of the deficiencies existing in the above problems, the present application provides a continuous casting quality judgment off-line data analysis system.

[0005] To achieve the above object, the present application provides a continuous casting quality judgment off-line data analysis system, comprising an off-line data analysis module and an off-line data verification module.

[0006] The off-line data analysis module comprises a data acquisition module, a data preprocessing module, a feature extraction module, a mechanism model module and a model training module.

[0007] The data acquisition module is used for acquiring various data in the continuous casting process.

[0008] The data preprocessing module is used for labeling the data based on a set rule, and the label is divided into normal label and abnormal label.

[0009] The feature extraction module is used for classifying the labeled data and obtaining data features.

[0010] The mechanism model module is used for building a model for different stages of the continuous casting process according to the mechanism model, to obtain a mechanism model.

[0011] The model training module is used for inputting the data features into the corresponding mechanism model for training.

[0012] The off-line data verification module is used for verifying the trained mechanism model with the data.

[0013] Preferably, the data includes ladle temperature, ladle molten steel weight, tundish temperature, tundish molten steel weight, casting speed and casting blank defect position.

[0014] Preferably, the data acquisition module acquires data through the message middleware mode of Kafka.

[0015] Preferably, the data preprocessing module is further used for cleaning, denoising and outlier processing of the data.

[0016] Preferably, the mechanism model in the mechanism model module includes a ladle molten steel quality change model, a ladle molten steel temperature change model, a tundish molten steel quality change model, a tundish molten steel temperature change model, a solidification steady-state mechanism model and a secondary cooling water distribution mechanism model.

[0017] Preferably, the off-line data analysis module further comprises a correlation coefficient statistical module, which is used for calculating the correlation coefficient between the data and the quality of the continuous casting machine, and screening target data from the data according to the correlation coefficient.

[0018] Preferably, the offline data verification module verifies accuracy, precision, recall, F1 value of the mechanism model.

[0019] Preferably, the offline data verification module is further configured to optimize the data to the trained mechanism model.

[0020] Preferably, the optimization step of the offline data verification module comprises removing data features that do not contribute to the mechanism model and adjusting parameters of the mechanism model.

[0021] Preferably, adjusting the parameters of the mechanism model comprises adaptive parameter adjustment, incremental parameter adjustment, deep reinforcement learning, target-oriented parameter adjustment and multi-objective parameter adjustment.

[0022] Compared with the prior art, the present application has the following beneficial effects:

[0023] The present application can comprehensively collect and analyze continuous casting quality data, and simultaneously make corresponding quality determination. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a module flow chart of the continuous casting quality determination offline data analysis system of the present application. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0026] With reference to Figure 1 , the present application provides a continuous casting quality determination offline data analysis system, comprising:

[0027] an offline data analysis module and an offline data verification module;

[0028] The offline data analysis module comprises a data acquisition module, a data preprocessing module, a feature extraction module, a mechanism model module, a model training module and a correlation coefficient statistical module.

[0029] The data acquisition module is configured to acquire various data in the continuous casting process.

[0030] Specifically, it is responsible for collecting various data in the continuous casting production process, such as ladle temperature, ladle molten steel weight, tundish temperature, tundish molten steel weight, casting speed and casting blank defect position. The data collection mode of L1, L2 and L3 is collected by adding a camera, and the defect detection and casting blank shape analysis are carried out by the camera collection mode. At the same time, new big data analysis technology is adopted in the collection process, and the message middleware mode of Kafka is added to the data channel collection mode, which provides data gain collection, increases the data collection amount and response degree in the traditional case, and at the same time, wireless communication 5G and cloud data collection are carried out through the sensor network to form a multi-point data collection mode. At the same time, the data between the sensing nodes is verified through the data verification mode to ensure the reliability of the data communication of the sensing nodes. Through the access of the overall sensor network, a new and all-round data collection mode is provided for the offline analysis system.

[0031] The data preprocessing module is used for labeling data based on a set rule, and the label is divided into normal label and abnormal label.

[0032] Specifically, due to the access of the multi-channel sensor network, data preprocessing in the form of multiple data needs to be cleaned, denoised and processed for abnormal values. The offline system adopts a label offline data preprocessing mode in manual and automatic modes. In the process of data entering the offline data system, the data is labeled by manual labeling, and the label function of the data is automatically obtained through the relational verification of the coordinated data, such as: when the molten steel is in the tundish state, the temperature label is automatically set to 1300-1600 degrees, and the label above 1600 degrees is directly set as an abnormal label, and then the next step of abnormal value processing of the label data is carried out. This kind of automatic label adopts a new rule engine mode to automatically determine the label, and the rule is determined in real time to automatically label. Through this rule configuration, the efficiency and accuracy of automatic labeling are greatly improved. The automatic labeling method can also be extended to the online system to record abnormal events. The data denoising module adopts an offset value abnormal denoising algorithm to denoise the algorithm. The abnormal data above the threshold value in the offset value abnormal state is denoised in detail, such as 3Sigma statistical test analysis of data. The data exceeding 3Sigma is analyzed as noise, and the abnormal value is processed by smoothing the data, and the corresponding filtering and denoising are carried out.

[0033] The feature extraction module is used for classifying the labeled data and obtaining data features.

[0034] Specifically, after data preprocessing, the features of the data are extracted. The feature extraction method is based on traditional algorithms. Since a series of label information has been processed by the rule engine in the data preprocessing link, the features of the data are extracted by machine learning and deep learning. The machine learning method classifies various data by algorithm classification, effectively classifies various data containing defects and defects, obtains the features of the data, and increases the information gain of each type of data. For example, in the process of conventional process parameters such as pulling speed, temperature value, and casting speed, the feature information gain related to slab quality is added, such as surface micro-crack area, slag area and other information, which forms a feature information content containing feedback information content for feature extraction of the model. The deep learning method uses convolutional neural network for classification calculation. The convolutional network is constructed by 256 layers of depth, and the classification accuracy is improved by deepening the network.

[0035] A mechanism model module is configured to build a model for different stages of the continuous casting process according to a mechanism model, and obtain a mechanism model.

[0036] Specifically, the module is built by the mechanism model, and the mechanism model of the modules of the ladle, tundish, crystallizer, fan-shaped segment and secondary cooling region is built, such as the module for the ladle, which contains a ladle weight mechanism model, a side reaction of the ladle, and a change in the weight of the molten steel in the ladle during pouring to the tundish,

[0037] The ladle molten steel area mechanism model has a ladle molten steel quality change model and a ladle molten steel temperature change model.

[0038] M LD(τ) = m LD(τ+dr) + m TDout(τ)

[0039] Wherein:

[0040]

[0041] The tundish molten steel area mechanism model has:

[0042] A tundish molten steel quality change model and a tundish molten steel temperature change model.

[0043] For example:

[0044] M TD(τ) = M TD(τ-dr) + m LDout(τ) + m TDout(τ)

[0045] Wherein: M TD(τ) ≤ mTDfull

[0046]

[0047] At the same time, for the crystallizer and the fan-shaped segment and other areas, the solidification steady-state model mechanism model is used to characterize the process section to reflect the on-site process situation, and the solidification steady-state mechanism model is: (the pulling speed)

[0048]

[0049]

[0050] According to the above heat conduction equation, the temperature distribution of the casting blank is predicted, from the initial condition of determining the volume unit in the casting blank from the meniscus of the crystallizer to the downward movement of the pulling speed u, and the surface boundary condition of the casting blank through the crystallizer, the secondary cooling zone and the radiation zone.

[0051] The initial condition is set as follows: it is stipulated that the temperature of the molten steel at the meniscus of the crystallizer is equal to the pouring temperature Tc at the beginning of pouring (t = 0).

[0052] t = 0, 0 <= X <= a, 0 <= Y <= b, T(X, Y) = Tc (5)

[0053] Boundary conditions: the heat flow density in the casting blank is continuous. The boundary conditions of the casting blank surface are different in each cooling zone.

[0054] The secondary cooling water distribution mechanism model has:

[0055]

[0056] Q' = f(V cal )

[0057] Delta T = T Cal- T Aim

[0058] Delta Q i = k i · Delta T i

[0059] Q = Q' + Delta Q

[0060]

[0061] Through the loading of the mechanism model module, the mechanism model modeling can be used to realize the correlation between the key process parameters on site, and the effective derivation form between the parameters is achieved.

[0062] The model training module is used for inputting data features to the corresponding mechanism model for training;

[0063] Specifically, since the corresponding feature mode is used in advance, the corresponding training can be performed for the model, the model training module can optionally perform selection and setting of various models to construct a model network, and the data is trained separately by selecting various models, and the network structure is constructed by constructing the model, for example: the number of layers of the model can be freely defined, and the model sub-item module such as the pooling layer and the convolution layer can be freely defined. The model part can also be effectively managed, and the version management is performed through these ways to train a group of models. Through the model training module, the data modeling can be effectively performed, and the good data model can be effectively applied in the related classification system.

[0064] A correlation coefficient statistical module is configured to calculate a correlation coefficient between the data and the quality of the continuous casting machine, and to filter target data from the data according to the correlation coefficient;

[0065] Specifically, since there is a correlation between the process parameters and the process parameters, the correlation between the process parameters and the parameters is calculated, the module performs correlation statistics on various process parameters, and the module sorts various correlation coefficients, and through sorting, the closely related process parameters are subjected to corresponding statistical analysis to achieve the purpose of correlation parameter extraction comparison. Then, according to the correlation parameters, the system is analyzed to determine which type of process parameter is related to the final product quality such as crack, center porosity and other defects, obtain the process parameters, and then how to monitor and adjust to reduce the generation of quality problems and optimize the output quality result.

[0066] An offline data verification module is configured to verify the data against a trained mechanism model.

[0067] Specifically, the offline data model verification method can be used to verify each type of data offline, and the optimal model can be obtained through the verification data model to perform online analysis. The offline data verification module is mainly used to verify the accuracy, precision, recall rate, F1 value and other information of the model, determine the specific performance of the model through the accuracy of the verification model, and screen and verify the model through the index. For the same model, the specific performance of the model can also be seen by adjusting the model parameters to obtain the index information of the model, so as to achieve the effect of offline data model verification.

[0068] For specific scenarios such as surface quality determination, the evaluation model and verification model module mainly uses confusion matrix, ROC curve, cross-validation, bias and variance, etc. Gain evaluation parameters, through the increase of evaluation parameters to more accurately evaluate the model. Among them, for various surface quality scenarios, multi-class confusion matrix is used for evaluation, for example: for surface slag, surface iron oxide, surface crack, etc. As a one-class confusion matrix evaluation, through this kind of confusion matrix, the correlation between the features of the above three types of defects can be evaluated, and the correlation between the above three types of defects can also be evaluated for surface slag, surface crack, and surface radius crack. The correlation between the above three types of defects is evaluated.

[0069] The offline data verification module is also used to optimize the data for the trained mechanism model. Since adjustment and optimization of the offline data model are required during offline verification, optimization of the offline data model mainly includes the following:

[0070] Feature selection: Offline data model optimization can optimize the model through feature selection, remove features that do not contribute to the model, and improve the accuracy and performance of the model. For example: under the input of a large number of process parameters, a considerable part of the parameters does not contribute to the model. In order to improve the accuracy and performance of the model, expert knowledge and methods are used to screen the features, and features that are not related to the output are excluded, for example: for the main influence output feature parameters such as pulling speed, casting speed, ladle temperature, and molten steel flow, relevant feature selection is performed, while for parameters such as fan-shaped segment temperature, fan-shaped segment curvature, and ladle volume that have no special influence, optimization selection is performed to remove the corresponding feature parameters, thereby improving the performance of the model and the features of the model. The feature selection method combining expert experience and expert knowledge is applied as a new feature extraction method in the offline analysis system.

[0071] Parameter adjustment: By adjusting the parameters of the model, the accuracy and performance of the model can be improved, and the generalization ability of the model can be improved. In the parameter adjustment process, a diverse data set is used for verification, and multi-class data can be used for parameter adjustment and application.

[0072] The purpose of parameter adjustment is to improve the accuracy and performance of the model by adjusting the parameters of the model to achieve the best effect. In traditional parameter adjustment methods, commonly used methods include grid search, random search, Bayesian optimization, etc. However, in practical applications, in order to improve the efficiency and accuracy of parameter adjustment and to cope with some special application scenarios, the offline system uses some innovative parameter adjustment methods, such as:

[0073] Adaptive parameter adjustment: adaptive parameter adjustment refers to automatically adjusting the parameters of the model by monitoring the training process of the model. During the training process of the model, adaptive parameter adjustment will monitor the training error, gradient size and other indicators of the model, and dynamically adjust the parameters of the model according to the monitoring results to improve the training effect of the model. Through the adaptive parameter adjustment method, automatic parameter adjustment can be effectively realized, and the effect of adaptive parameter adjustment can be formed.

[0074] Incremental parameter adjustment: incremental parameter adjustment refers to adjusting the parameters of the model incrementally during the training process to improve the accuracy and performance of the model. In incremental parameter adjustment, the parameters of the model are fine-tuned according to the results of each iteration to gradually optimize the performance of the model. Through incremental parameter adjustment, the goal of gradually improving the performance of the model can be achieved. Since the performance of the model can be adjusted incrementally, this adjustment method can effectively accelerate the performance of the model.

[0075] Deep reinforcement learning: deep reinforcement learning is a parameter adjustment method that combines deep learning and reinforcement learning. In deep reinforcement learning, the parameters of the model are optimized through reinforcement learning to obtain the best action strategy in a specific task. For example, a reward parameter is set for the output indicators of the model training, and the model is reinforced through reinforcement learning to continuously adjust the parameters of the model and optimize the model parameters, thereby achieving the optimal performance of the model.

[0076] Target-oriented parameter adjustment: target-oriented parameter adjustment is a method of adjusting parameters by defining a target function. In target-oriented parameter adjustment, a specific target function can be specified to achieve better adjustment results, such as optimizing the classification accuracy of the model or reducing the error of the model.

[0077] Multi-objective parameter adjustment: multi-objective parameter adjustment is a parameter adjustment method that simultaneously optimizes multiple objective functions. In multi-objective parameter adjustment, multiple performance indicators of the model can be optimized simultaneously, such as accuracy, recall rate, F1 score, etc. By simultaneously optimizing multiple performance indicators, the highest value of model precision and performance can be achieved, and the optimal point of model precision and performance can be found

[0078] The above is only the preferred embodiment of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An offline data analysis system for continuous casting quality assessment, characterized in that, Includes an offline data analysis module and an offline data verification module; The offline data analysis module includes a data acquisition module, a data preprocessing module, a feature extraction module, a mechanism model module, and a model training module; The data acquisition module uses a sensor network model to collect various data in the continuous casting process, specifically: A multi-point data acquisition mode is formed through 5G wireless communication and cloud data acquisition; the data between sensor nodes is verified through data verification to ensure the reliability of sensor node data communication. The data preprocessing module is used to label the data based on set rules. The labels are divided into normal labels and abnormal labels, and the labeling methods include manual labeling and automatic labeling. The automatic labeling methods include: using a rule engine to integrate data, obtain real-time rules, make judgments on the data, and obtain data labels; The feature extraction module is used to classify and obtain data features from the labeled data using machine learning and deep learning methods; wherein, Machine learning methods include: classifying various types of data, including those with defects and those without defects, and performing information gain on the features of each type of data to form feature information content containing feedback information, which is used for feature extraction in the model; Deep learning methods include: using convolutional neural networks for classification calculations, and improving classification accuracy by deepening the network; The mechanism model module is used to build models for different stages of the continuous casting process based on the mechanism model approach, obtain mechanism models, and form the correlation and derivation forms between process parameters; specifically: In the ladle area, construct: a model of molten steel quality variation and a model of molten steel temperature variation; In the tundish area, construct: a model of molten steel quality variation in the tundish and a model of molten steel temperature variation in the tundish; A solidification steady-state mechanism model was built in the crystallizer and sector section areas to reflect the on-site process conditions. A model of the water distribution mechanism in the secondary cooling zone was constructed. The model training module is used to input the data features into the corresponding mechanism model for training; the specific training method includes: Model selection: Choose the appropriate training model based on different data classifications; Model network structure construction: Set the number of layers and sub-modules of the model. Sub-modules include pooling layers and convolutional layers. The offline data verification module verifies the trained mechanism model using methods including: confusion matrix, ROC curve, cross-validation, bias and variance. Specifically, for the surface quality scenario, a confusion matrix is ​​used to evaluate the correlation between the characteristics of different types of defects.

2. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The data includes ladle temperature, ladle molten steel weight, tundish temperature, tundish molten steel weight, casting speed, and billet defect location.

3. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The data acquisition module uses Kafka's message middleware mode to collect data through a data channel.

4. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The data preprocessing module is also used to clean, denoise, and handle outliers in the data.

5. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The method for constructing the solidification steady-state mechanism model includes: under given initial temperature and regional cooling boundary conditions, solving the heat conduction equation, simulating the temperature field evolution of the billet moving downward at a pulling speed in the continuous casting machine, and thereby predicting the temperature distribution of the billet.

6. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The offline data analysis module also includes a correlation coefficient statistics module, which is used to calculate the correlation coefficient between the data and the quality of the continuous casting machine, and to filter out target data from the data based on the correlation coefficient.

7. The offline data analysis system for continuous casting quality assessment according to claim 1, characterized in that, The offline data verification module verifies the accuracy, precision, recall, and F1 score of the mechanistic model.

8. The offline data analysis system for continuous casting quality assessment according to claim 7, characterized in that, The offline data verification module is also used to optimize the trained mechanism model with the data. The optimization method includes: removing data features that do not contribute to the mechanism model through feature selection and adjusting the parameters of the mechanism model.

9. The offline data analysis system for continuous casting quality assessment according to claim 8, characterized in that, The specific method for feature selection includes: using expert experience and knowledge to screen out parameters that affect output features, including: casting speed, pouring speed, ladle temperature, and molten steel flow rate; and removing parameters that have no special influence, including: sector temperature, sector curvature, and ladle volume.

10. The offline data analysis system for continuous casting quality assessment according to claim 9, characterized in that, Adjusting the parameters of the mechanistic model includes: Adaptive parameter tuning: During model training, monitor training error and gradient magnitude metrics, and dynamically adjust model parameters based on the monitoring results to improve model training performance; Incremental parameter tuning: During model training, incremental parameter fine-tuning is performed based on the results of each iteration to gradually optimize the model's performance; Deep reinforcement learning: Set reward parameters for the metrics output by the model training, and use the reward parameters to perform reinforcement learning on the model to optimize its performance; Objective-oriented parameter tuning: Parameters are tuned by defining an objective function to optimize the tuning effect; Multi-objective parameter tuning: In parameter tuning, multiple performance metrics of the model are optimized simultaneously to achieve the optimal balance between model accuracy and performance.

Images