Continuous casting process working condition mode mining method based on data driving

Abstract

The invention discloses a continuous casting process working condition mode mining method based on data driving, and belongs to the technical field of metallurgical production quality management. The method comprises the following steps of: directly acquiring time sequence data from a sensor of a production process monitoring system, analyzing the correlation between the time sequence data and slag inclusion defects by adopting a Spearman correlation method, selecting out the time sequence data with relatively strong correlation with the slag inclusion defects on the surface of the continuous casting billet, then converting the time sequence data into character string data, and finally performing clustering analysis on the character string data to obtain a working condition mode. The method has the advantages that the time sequence data are coded into the character string data, so that the phenomenon of distortion of the curve type data in the extraction process can be prevented, and the data reliability is improved; the working condition mode can help quality management personnel to improve the production process of the continuous casting billet in the continuous casting process and provide reference for necessary quality control.

Description

Technical Field

[0001] The present invention is a data-driven continuous casting process operating mode mining technology, and specifically relates to the technical field of operating mode mining of slag inclusion defects on the surface of continuous casting billets. Background Art

[0002] With the booming domestic economy, the demand for steel is growing, which not only requires quantity, but also puts forward higher standards in terms of quality, especially in the field of process manufacturing technology. How to effectively locate the surface slag inclusion defects of continuous casting billets is an important step in the quality control of continuous casting billets. At present, there are two main ways to effectively locate the surface slag inclusion defects of continuous casting billets: data-driven models and experience judgment. The mobility of experienced personnel is high, and the training data of the current data-driven model has been processed, which will cause data distortion problems.

[0003] At present, all large domestic steel enterprises have realized enterprise automation systems, in which production data is mainly stored in the production process monitoring system (primary system) and the production planning and scheduling system (secondary system). The primary system is the most basic level in the automation system of metallurgical enterprises, and is mainly used for real-time monitoring of production process parameters, including parameters measured by various sensors of the production process monitoring system, such as temperature, pressure, humidity and vibration, etc. The data stored are all time series data with a large amount of data. The secondary system is a system that is further integrated and optimized on the basis of the primary system, and is mainly used for the designation and scheduling of production plans, with the purpose of improving production efficiency and reducing production costs. The data stored in the secondary system is single-value data after the time series data in the primary system is processed with characteristic values ​​such as mean and median, and cannot accurately reflect the changing characteristics of parameters.

[0004] At present, the data used by researchers in the steel field in the quality research of continuous casting billets mostly come from the data of the secondary system, and the machine learning models trained by these data, such as random forest, knowledge vector machine or genetic algorithm and their combination, are used to analyze the quality problems of continuous casting billets. Through the on-site investigation of steel plants, it is found that the accuracy of the current research on the slag inclusion defects on the surface of continuous casting billets in the actual application of the field is quite different from the accuracy of the offline test set. When the quality problems of continuous casting billets occur, it is impossible to quickly and accurately trace the time and place of the problem, so as to recall the relevant batches of products or adjust the process. Summary of the invention

[0005] The purpose of the present invention is to provide a method for mining the working condition pattern of continuous casting process based on data drive, which solves the problem that when the continuous casting billet has quality problems, it is not possible to quickly and accurately trace the time and location of the problem, recall the relevant batches of products or adjust the process. It can prevent the distortion of curve data during the extraction process, improve the reliability of data, and study the working condition pattern mining of continuous casting process based on the data directly collected by the sensors of the production process monitoring system. The specific steps are as follows:

[0006] (1) Collect time series data directly from sensors in the production process monitoring system;

[0007] (2) Data transmission and data storage are performed through the publish-subscribe tool Kafka, and the collected time series data is stored in the database. Kafka is a message queue, which is divided into producers and consumers. The message producer publishes the data to the Kafka message queue. Kafka has multiple topics, and different types of message data will be stored in different topics. The message consumer will subscribe to the topic of message storage, so as to read the time series data sent by the producer and store it in the database;

[0008] (3) Preprocessing the stored time series data, including noise reduction, downsampling, and standardization of the stored time series data;

[0009] (4) The Spearman correlation method is used to analyze their correlation with slag inclusion defects, and the time series data with strong correlation with slag inclusion defects on the surface of continuous casting billets are selected;

[0010] (5) Use the Sax algorithm to encode the preprocessed time series data and convert it into string data;

[0011] (6) Cluster analysis is used to analyze the string data to find a fixed sequence, which is the fixed operating mode substring, referred to as the operating mode.

[0012] The data transmission and data storage described above obtain the necessary parameters for subsequent research from the data center through the publish-subscribe tool Kafka; Kafka, as a message queue, is divided into producers and consumers. Message producers publish data to the Kafka message queue. Kafka has multiple Topics, and different types of message data will be stored in different Topics. Message consumers will subscribe to the Topic where the message is stored and read the message data. The data collection method is the consumer role. After determining the necessary parameters, you can directly subscribe to the corresponding Topic to obtain data in Json format, parse it into structured data, and store it in the corresponding business database.

[0013] The preprocessed time series data is encoded and converted into string data:

[0014] Select the symbol set. First, you need to select a set of symbols to map the string data collected by the continuous casting process to these symbols. The 26 English letters are used as symbols for mapping.

[0015] Symbol mapping: Mapping time series data of a certain period of time to a symbol in the selected symbol set, which is done by mapping the data value with the range of the symbol set;

[0016] The 26 English letters are used as symbols for mapping, and the Sax algorithm is used for mapping. Before PAA (Piecewise Aggregate Approximation) processing, it is ensured that the time series data has met the requirements of a mean of 0 and a standard deviation of 1 in the standard normal distribution. At this time, PAA can be used for dimensionality reduction; then the split point of the interval is found in the normal distribution table, and the mean represented by PAA is mapped to the corresponding letters based on the Sax algorithm idea, and finally the time series data is discretized into string data.

[0017] The cluster analysis method is the K-means clustering algorithm:

[0018] (1) Data preparation: prepare the string data;

[0019] (2) Determine the number of clusters, i.e., the number of clusters; use the Elbow Method to select the appropriate number of clusters, and set the number of clusters to 1, so that the fixed operating mode substring mined by the K-means algorithm is used as the operating mode of the parameter;

[0020] (3) Initialize the cluster centers. In K-means clustering, the cluster centers need to be initialized and randomly selected from the data. These centers will be continuously updated during the iteration process to find the center of the cluster.

[0021] (4) Iterative clustering process: K-means clustering is an iterative process that calculates the distance of each data point to the nearest cluster center and assigns the data point to the closest cluster;

[0022] (5) Update the center of each cluster and repeat this process until convergence or the specified number of iterations is reached, thereby obtaining a fixed sequence that appears in the defect data, namely, a fixed operating mode substring, referred to as the operating mode.

[0023] The advantages of the present invention are:

[0024] (1) Time series data is encoded as string data, which can prevent the distortion of curve data during the extraction process and improve the reliability of the data;

[0025] (2) Operating mode mining: mining string data to obtain operating mode, which subverts the previous quality event mining;

[0026] (3) Directly use the data collected by the primary sensor to explore the working condition mode, which subverts the problem of using the secondary system data processed by the secondary system or relying on experience to conduct quality research in the past. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is the overall flow chart of the method for mining slag inclusion patterns on the surface of continuous casting billets.

[0028] Figure 2 Flowchart for time series data preprocessing.

[0029] Figure 3 Schematic diagram of Spearman correlation analysis of continuous casting process parameters.

[0030] Figure 4 Encode time series data as a string data flow diagram.

[0031] Figure 5 Operating mode mining flow chart. DETAILED DESCRIPTION

[0032] The implementation steps of the present invention mainly include three modules, namely data collection, time series data encoding and working mode mining. The specific steps are as follows:

[0033] 1. Data Collection

[0034] like Figure 1As shown in the figure, the parameters that affect the quality of slag inclusion on the surface of the continuous casting billet are collected directly from the sensors of the production process monitoring system, as described below: Slicing in the crystallizer moment: Slicing refers to the process of cutting the continuous casting billet into a specific length from the liquid molten steel. The moment records the time point of the slicing operation, which may be related to slag inclusion. Large bag weight: The large bag is a device for storing and transporting liquid molten steel, and its weight may affect the quality of the continuous casting billet and the surface slag inclusion defect. Stopper rod argon sealing pressure and stopper rod argon sealing flow: These two parameters involve the argon sealing operation during the continuous casting process, which is very important for the control of the billet wall and liquid molten steel. Middle bag weight: The middle bag is an intermediate device for storing and transporting liquid molten steel, and its weight may also affect the quality of the continuous casting billet. Crystallizer liquid level height and fluctuation value: The crystallizer liquid level controls the cooling process of the continuous casting billet and has an important influence on the quality. Superheat: Superheat refers to the degree to which the temperature of the liquid molten steel exceeds its liquid phase equilibrium point. Superheat can affect the cooling rate of the billet wall. Crystallizer taper: The crystallizer taper indicates the shape of the crystallizer, which may affect the liquid flow and cooling of the continuous casting billet. Crystallizer inlet and outlet water temperature difference: The inlet and outlet water temperature difference in different parts of the crystallizer may also affect the cooling rate of the billet wall. Crystallizer water volume: The water volume in different parts of the crystallizer controls the flow of liquid steel and may affect the slag inclusion situation. Crystallizer heat flux density: The heat flux density in the crystallizer describes the heat distribution and has an important influence on the temperature and quality of the continuous casting billet. Drawing speed: The drawing speed of the continuous casting billet also has an important influence on its cooling and quality. Bending section drawing speed: The bending section refers to the drawing section of the continuous casting billet during the bending process, and its drawing speed will also affect the formation of the billet wall.

[0035] 2. Data is transmitted through the publish-subscribe tool Kafka, and the collected time series data is stored in the corresponding business database. Kafka is a message queue, which is divided into producers and consumers. The message producer publishes the data to the Kafka message queue. Kafka has multiple Topics, and different types of message data will be stored in different Topics. Message consumers will subscribe to the Topic where the message is stored to read the time series data. The data transmission and storage method is the consumer role. You can directly subscribe to the corresponding Topic to obtain data in Json format, and then parse it into structured data, and store the parsed structured data in the corresponding business database.

[0036] 3. Preprocess the time series data stored in the database. The preprocessing includes noise reduction, downsampling, and standardization of the stored time series data. Data preprocessing: Figure 2As shown in the figure, the Loss average method is used to smooth the time series data in order to reduce the noise in the data and make the change trend of the time series data more obvious. Since the data collection frequency of each parameter is different, in order to ensure that their timestamps are aligned and facilitate the subsequent encoding of the time series data into string data, the time series data is downsampled and the downsampling frequency is selected according to the different distribution of the parameter data points. The downsampled time series data is then standardized, and the standardized time series data meets the requirements of a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0037] 4. There are many parameters that affect the quality of slag inclusion on the surface of continuous casting billet. Figure 3 The Spearman correlation method shown is used to analyze their correlation with slag inclusion defects, and select the time series data with a strong correlation with the slag inclusion defects on the surface of the continuous casting billet.

[0038] 5. Convert time series data into string data.

[0039] like Figure 4 As shown in the figure, converting time series data into string data usually involves mapping the data points in the time series to a series of characters or symbols for easier analysis or representation. There are many ways to convert time series data into string data, such as encoding time series data into string data through the Sax algorithm. The main steps are as follows:

[0040] Select the symbol set: First, you need to select a set of symbols or characters to map the time series data collected by the continuous casting process to these symbols. These symbols can be letters, numbers, special characters or any character set, using the 26 English letters for mapping.

[0041] Symbol mapping: Map the time series data of a certain time period to a symbol in the selected symbol set. This is accomplished by mapping the data value to the range of 26 English letters in the symbol set. The Sax algorithm used in the mapping method of the present invention ensures that the time series data has met the requirements of a mean of 0 and a standard deviation of 1 in the standard normal distribution before PAA (Piecewise Aggregate Approximation) processing. At this time, dimensionality reduction can be performed through PAA. Simply put, PAA is to first divide the original sequence into w subsequences of equal length, replace each subsequence with the mean of each subsequence, and then determine the Σ of the character set, that is, how many letters are used to represent the entire time series. For example, if three letters 'a', 'b', and 'c' are selected, then Σ=3. The present invention uses 26 characters, that is, Σ=26, and then searches for the split point of the interval in the normal distribution table. Based on the idea in the Sax algorithm, the corresponding time series data is mapped to letters according to the breakpoint value of the normal distribution, and the mean represented by PAA is mapped to the corresponding letters, and the time series data is converted into string data.

[0042] The advantages of this step include the following two aspects: data stability. By encoding time series data into strings, noise and frame drops in the data can be eliminated, improving data stability; compressibility: string encoding can compress the original data into a smaller representation, which can save a lot of storage space for businesses that need to process large amounts of time series data.

[0043] 6. Perform K-means clustering analysis on string data

[0044] After string encoding the time series data, you can start mining the operating mode, such as Figure 5 As shown, since it is difficult to directly obtain the fixed working mode substring in the original string, cluster analysis is adopted here: cluster analysis is performed on the time series data represented by the string to find the fixed working mode substring. There are many clustering methods, such as K-means clustering analysis, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) and hierarchical clustering. However, considering that working mode mining has extremely high requirements on data quality, if a more complex method is used for data mining, the effect may not be ideal. Therefore, after comparing the application effects, the K-means clustering algorithm is used. Its main steps are as follows:

[0045] Prepare data: First, prepare the string data converted from time series data for K-means clustering analysis.

[0046] Select the number of clusters: Before using K-means clustering, you first need to determine the number of clusters. In order to find a fixed operating mode substring, the Elbow Method is used to select the appropriate number of clusters. The number of clusters is set to 1, so that the K-means algorithm can be used to mine a fixed operating mode substring.

[0047] Initialize cluster centers: In K-means clustering, you need to initialize cluster centers, usually randomly selected from the data. These centers will be continuously updated during the iteration process to find the center of the cluster.

[0048] Iterative clustering process: K-means clustering is an iterative process that calculates the distance from each data point to the nearest cluster center and assigns the data point to the closest cluster. Then, the center of each cluster is updated and the process is repeated until convergence or the specified number of iterations is reached, thereby obtaining a fixed sequence in the defect data, which is a fixed operating mode substring, referred to as the operating mode.

[0049] The advantages of mining fixed operating mode substrings include three aspects: quality improvement. By identifying fixed operating mode substrings of key parameters, we can better understand how the operating conditions affect product quality, so as to take measures to improve the production process; predict and prevent problems, quickly find the problem of continuous casting defects, and assist process personnel to take preventive measures to reduce the occurrence of product defects and quality problems; research reference. The mined fixed operating mode substrings can provide useful reference information for subsequent quality research and optimization, helping researchers to better understand the production process.

[0050] Interpretation and application of clustering results: Based on the clustering results, fixed operating mode substrings can be identified. After converting the time series data into string data, if a fixed operating mode substring containing slag inclusion defect data on the surface of the continuous casting billet is included, it is considered that a quality problem has occurred. This involves further data interpretation and subsequent actions, such as helping quality management personnel improve the production process of the continuous casting billet in the continuous casting process and perform necessary quality control as a reference.

Claims

1. A method for mining continuous casting process operating mode based on data drive, characterized in that: The specific steps include: (1) Collect time series data directly from sensors in the production process monitoring system; (2) Data transmission and data storage are performed through the publish-subscribe tool Kafka, and the collected time series data is stored in the data center. Kafka is a message queue, which is divided into producers and consumers. The message producer publishes the data to the Kafka message queue. Kafka has multiple topics, and different types of message data will be stored in different topics. The message consumer will subscribe to the topic of message storage, read the time series data sent by the producer, and store it in the database; (3) Preprocessing the stored time series data, including noise reduction, downsampling, and standardization of the stored time series data; (4) The Spearman correlation method is used to analyze their correlation with slag inclusion defects, and the time series data with strong correlation with slag inclusion defects on the surface of continuous casting billets are selected; (5) Use the Sax algorithm to encode the preprocessed time series data and convert it into string data; (6) Cluster Analysis method is used to analyze the string data to find a fixed sequence that appears in the string, which is the fixed operating mode substring, referred to as the operating mode for short.

2. A method for mining continuous casting process operating mode based on data drive according to claim 1, characterized in that: The data transmission and data storage described above obtain the necessary parameters for subsequent research from the data center through the publish-subscribe tool Kafka; Kafka, as a message queue, is divided into producers and consumers. Message producers publish data to the Kafka message queue. Kafka has multiple Topics, and different types of message data will be stored in different Topics. Message consumers will subscribe to the Topic where the message is stored and read the message data. The data collection method is the consumer role. After determining the necessary parameters, you can directly subscribe to the corresponding Topic to obtain data in Json format, parse it into structured data, and store it in the corresponding business database.

3. The method for mining continuous casting process operating mode based on data drive according to claim 1, characterized in that: The preprocessed time series data is encoded and converted into string data: Select the symbol set. First, you need to select a set of symbols to map the string data collected by the continuous casting process to these symbols. The 26 English letters are used as symbols for mapping. Symbol mapping: Mapping time series data of a certain period of time to a symbol in the selected symbol set, which is done by mapping the data value with the range of the symbol set; The 26 English letters are used as symbols for mapping, and the Sax algorithm is used for mapping. Before PAA (Piecewise Aggregate Approximation) processing, it is ensured that the time series data has met the requirements of a mean of 0 and a standard deviation of 1 in the standard normal distribution. At this time, PAA can be used for dimensionality reduction; then the split point of the interval is found in the normal distribution table, and the mean represented by PAA is mapped to the corresponding letters based on the Sax algorithm idea, and finally the time series data is discretized into string data.

4. A method for mining continuous casting process operating mode based on data drive according to claim 3, characterized in that: The cluster analysis method is the K-means clustering algorithm: (1) Data preparation: prepare the string data; (2) Determine the number of clusters, i.e., the number of clusters; use the Elbow Method to select the appropriate number of clusters, and set the number of clusters to 1, so that the fixed operating mode substring mined by the K-means algorithm is used as the operating mode of the parameter; (3) Initialize the cluster centers. In K-means clustering, the cluster centers need to be initialized and randomly selected from the data. These centers will be continuously updated during the iteration process to find the center of the cluster. (4) Iterative clustering process: K-means clustering is an iterative process that calculates the distance of each data point to the nearest cluster center and assigns the data point to the closest cluster; (5) Update the center of each cluster and repeat this process until convergence or the specified number of iterations is reached, thereby obtaining a fixed sequence that appears in the defect data, namely, a fixed operating mode substring, referred to as the operating mode.

Images