Thick Plate Shape Prediction Method Based on Industrial Big Data

Abstract

The present invention discloses a method for predicting the shape of thick plates based on industrial big data, comprising the steps of: 1. Generating a shape quality prediction model through offline training of the model and storing it in the model database (4); 2. Inputting the thick plate number, and the data preprocessing module (2) retrieves the steel type information from the industrial database (6); 3. Inputting the production time, the model training module (3) retrieves the available model ID from the model database and selects the model ID; 4. The model prediction module (5) retrieves the shape quality prediction model from the model database, loads and initializes the shape quality prediction model; 5. The data preprocessing module retrieves multi-source heterogeneous industrial data from the industrial database and transmits it to the model prediction module; 6. The model prediction module performs the prediction of the shape quality of the thick plate. The present invention can accurately predict the final shape quality of the thick plate during the production process of the thick plate, be used to understand the product performance of the thick plate in advance and guide the production process, greatly improve the product quality, and increase the qualified rate of the product.

Description

Technical Field

[0001] The present invention relates to a method for controlling the production quality of steel plates, and particularly to a method for predicting the shape of thick plates based on industrial big data. Background Art

[0002] Thick plates are steel plates with a thickness of 20 - 60 mm in strip steel products. The production process mainly includes several stages such as slab selection, slab heating, rolling, post-rolling accelerated cooling, hot straightening, cooling bed cooling, shearing, and cold straightening. The cooling of the steel plate on the cooling bed is a slow process. If the final shape quality of the steel plate can be predicted in time before the cooling bed cooling process, the shape quality of the steel plate can be corrected in time in the subsequent production processes, and the subsequent production process can be improved by adjusting parameters, which can effectively ensure the product quality of the steel plate and improve production efficiency.

[0003] In the production process of thick plates, the industrial big data comes from multiple sensors and has the characteristics of multi-source heterogeneity. Due to reasons such as different dimensions between the data samples collected by sensors, different data storage formats, different specification parameters of different steel plate samples, and different running speeds of the conveyor belts, the data dimensions are not unified. Therefore, the industrial big data in thick plate production includes one-dimensional data (such as index data of process variables, etc.), two-dimensional data (such as information of temperature field, shape meter, and profiler, etc.), and three-dimensional data (such as information of flaw detector, rolling time sequence, etc.).

[0004] Chinese Patent CN 108637020 A discloses a method for predicting the convexity of strip steel using an adaptive mutation PSO-BP neural network. This method uses actual data to correct the shape prediction model, and predicts the convexity of the strip steel by determining the input, output, and modifying the prediction algorithm and control parameters. The input layer neurons of this method include mill roll gap, work roll speed, bending roll force, position feedback, rolling force, rolling force difference, tension, and temperature. However, due to the different dimensions of the multi-source heterogeneous data samples and the lack of unification, the prediction accuracy of the steel plate shape is reduced, affecting the subsequent production quality of thick plates.

[0005] Chinese Patent CN 108268979 A discloses a method for predicting the quality of medium-thick plates based on evolutionary fuzzy association rules. This method uses a fitting function to replace the consequent of the fuzzy association rule to achieve classification and regression prediction functions. However, this method is only applicable to the production of medium-thick plates with a thickness of 4.5 - 25 mm, and is only applicable to the index data analysis of steel plates. It cannot process the multi-source heterogeneous data generated in the thick plate production process, which has a greater impact on the accuracy of the steel plate shape prediction result.

[0006] Chinese Patent CN 103418619 A discloses a shape prediction and control method for cold-rolled strip steel. This method combines the changes in rolling force and mill tension to establish a fuzzy inference model, and online adjusts the work roll bending device, thereby controlling two shape defects of center wave and edge wave in strip steel products. This method only considers the influence of the rolling mill on the shape quality of the steel plate, does not introduce other processes in the modeling process, loses some effective information, reduces the accuracy of shape prediction, and cannot ensure the production quality of subsequent strip steel. Summary of the Invention

[0007] The object of the present invention is to provide a shape prediction method for heavy plates based on industrial big data, which can accurately predict the final shape quality of heavy plates during the production process of heavy plates, be used to understand the product performance of heavy plates in advance, and be used to guide the production process, greatly improving the product quality and increasing the product qualification rate.

[0008] The present invention is implemented as follows:

[0009] A shape prediction method for heavy plates based on industrial big data, which is implemented based on a shape prediction system. The shape prediction system includes:

[0010] A front-end interface for inputting parameters and displaying the shape quality prediction results;

[0011] A model database for storing the shape quality prediction model and its model ID;

[0012] An industrial database for storing multi-source heterogeneous industrial data during the production process of heavy plates;

[0013] A model training module for offline training the shape quality prediction model and storing the shape quality prediction model after the training is completed in the model database;

[0014] And a model prediction module for retrieving the shape quality prediction model from the model database and performing shape quality prediction through the shape quality prediction model, and sending the shape quality prediction result to the front-end interface;

[0015] The shape prediction method for heavy plates described above includes the following steps:

[0016] Step 1: Generate a shape quality prediction model for a specified heavy plate through offline model training and store it in the model database;

[0017] Step 2: The user inputs the heavy plate number through the front-end interface. The data preprocessing module retrieves the steel grade information corresponding to the heavy plate number from the industrial database and displays it through the front-end interface;

[0018] Step 3: The user inputs a continuous production time through the front-end interface. The model training module retrieves the model ID of the available flatness quality prediction model during this continuous production time from the model database and displays it through the front-end interface. The user selects the model ID.

[0019] Step 4: The model prediction module retrieves the flatness quality prediction model corresponding to the model ID selected in Step 3 from the model database, loads the flatness quality prediction model and initializes it.

[0020] Step 5: The data preprocessing module retrieves the multi-source heterogeneous industrial data during the production process of the heavy plate corresponding to the heavy plate number in Step 2 from the industrial database, passes it to the model prediction module, and displays it through the front-end interface at the same time.

[0021] Step 6: The model prediction module performs flatness quality prediction on the heavy plate corresponding to the heavy plate number in Step 2 through the flatness quality prediction model loaded in Step 4, and displays the flatness quality prediction result of the heavy plate through the front-end interface.

[0022] The described Step 1 includes:

[0023] Step 1.1: Store the historical data generated during the production of different heavy plates in the industrial database.

[0024] Step 1.2: The user issues a quality prediction execution instruction for a specified heavy plate through the front-end interface. The data preprocessing module retrieves the historical data of the specified heavy plate from the industrial database.

[0025] Step 1.3: The data preprocessing module performs information interception and data filling on the historical data to form a multi-source heterogeneous data block, and passes the multi-source heterogeneous data block into the model training module.

[0026] Step 1.4: The model training module classifies the flatness quality anomalies and establishes anomaly criteria for each flatness quality anomaly category respectively.

[0027] Step 1.5: The model training module establishes a flatness quality prediction model for each flatness quality anomaly category through the multi-source heterogeneous data block.

[0028] Step 1.6: Save the flatness quality prediction model corresponding to the specified heavy plate in the model database and assign a corresponding model ID.

[0029] The method for the data preprocessing module to perform information interception on the historical data is:

[0030] S1.3.1: Respectively intercept the temperature measurement intervals of several temperature sensors at the corresponding positions through the position information of the heavy plate.

[0031] S1.3.2: Set the temperature jump reference value as the criterion for judging temperature jump, and intercept the temperature data information of each temperature sensor within the temperature measurement range according to the temperature jump reference value.

[0032] When intercepting the temperature measurement range of each temperature sensor, a certain margin is reserved, and the length of this margin is 5% of the length of the temperature measurement range.

[0033] In the temperature measurement range of each temperature sensor, the temperature jump reference value is 50% of the temperature difference between the highest temperature and the lowest temperature in this temperature measurement range.

[0034] In step 1.3, data filling includes data filling for the heating process, data filling for the rolling process, and data filling for the cooling process;

[0035] The method for data filling in the heating process is: extract the in-furnace time of a thick plate sample, and intercept the furnace temperature data corresponding to several temperature sensors according to the in-furnace time, and arrange the furnace temperature data in chronological order into two-dimensional furnace temperature data for this thick plate sample; taking the maximum amount of furnace temperature data in all thick plate samples as the standard, fill in zeros around the furnace temperature data in the remaining thick plate samples to make the dimensions of the furnace temperature data in all thick plate samples unified;

[0036] The method for data filling in the rolling process is: taking the maximum amount of the pass dimension in all thick plate samples as the standard, fill in random data in the pass dimension of the remaining thick plate samples to make the pass dimensions in all thick plate samples unified;

[0037] The method for data filling in the cooling process is: taking the maximum amount of the data dimension in all thick plate samples as the standard, fill in random data in the data dimension of the remaining thick plate samples to make the data dimensions in all thick plate samples unified.

[0038] In step 1.4, the categories of abnormal plate shape quality include: head and tail warping, thickness abnormality, center wave, left wave, and right wave;

[0039] The abnormal criteria for the head and tail warping are: the FQC data of the head of the thick plate is over the limit, or the FQC data of the tail of the thick plate is over the limit, or there is a wave shape at the head of the thick plate, or there is a wave shape at the tail of the thick plate; the abnormal criteria for the thickness abnormality are: the FQC data of the middle area of the thick plate is over the limit; the abnormal criteria for the center wave are: the FQC data of the middle area of the thick plate is not over the limit, but there is a wave shape in the longitudinal direction of the thick plate; the abnormal criteria for the left wave are: the FQC data of the left side area of the thick plate is over the limit or there is a wave shape; the abnormal criteria for the right wave are: the FQC data of the right side area of the thick plate is over the limit or there is a wave shape.

[0040] Step 1.5 includes:

[0041] S1.5.1: Split a multi-source heterogeneous data block into several small data blocks according to the different characteristics of the data under different processes;

[0042] S1.5.2: Characterize each small data block using the maximum value, minimum value, average value, and variance to form statistical data;

[0043] S1.5.3: Combine the statistical data into a one-dimensional new statistical index data according to the actual association order of the processes, and use it as an input to the prediction model;

[0044] S1.5.4: Use the maximum information coefficient algorithm to calculate the correlation coefficient m between the new statistical index data and each type of strip shape quality anomaly category, where the maximum information coefficient MIC ∈ [0, 1];

[0045] S1.5.5: Sum the correlation coefficients of each small data block to obtain the correlation coefficient M between different data blocks and the strip shape quality anomaly category, M = {m 1 , m 2 , …, m n}, where m n is the correlation coefficient between the nth small data block and the strip shape quality anomaly category, m n = m n ave + m n std + m n max + m n min ; where m n ave is the correlation coefficient between the average value of the nth small data block and the strip shape quality anomaly category, m n std is the correlation coefficient between the variance of the nth small data block and the strip shape quality anomaly category, m n max is the correlation coefficient between the maximum value of the nth small data block and the strip shape quality anomaly category, m n min is the correlation coefficient between the minimum value of the nth small data block and the strip shape quality anomaly category;

[0046] S1.5.6: Sum the correlation coefficients between several small data blocks belonging to the same multi-source heterogeneous data block and the strip shape quality anomaly category to obtain the correlation coefficient G between different multi-source heterogeneous data blocks and the strip shape quality anomaly category;

[0047] S1.5.7: Feature extraction is performed on each small data block. The statistical data of all small data blocks form one-dimensional time series data, and a convolutional neural network model with the small data block as the input and the strip shape quality anomaly category as the output is established; S1.5.8: According to the hierarchical structure of the convolutional neural network model, the features in the feature extraction layer of the convolutional neural network model are selected as the new features of the small data block;

[0048] S1.5.9: Repeat S1.5.7 to S1.5.8, select all multi-source heterogeneous data blocks as the input of the convolutional neural network model respectively, and extract the new features of the small data blocks in the multi-source heterogeneous data blocks;

[0049] S1.5.10: Merge all the new features extracted from the multi-source heterogeneous data blocks into the data input layer, and assign feature weights to each new feature in the data input layer according to the correlation coefficient G. The feature weight assigned to the i-th new feature is λ i , where n is a natural number and j is an integer in the interval [1, n];

[0050] S1.5.11: Establish a strip shape quality prediction model based on a single-layer neural network, multiply the extracted new features by the feature weight λ i and use it as the input of the strip shape quality prediction model. The prediction probability of the strip shape quality anomaly category of the thick plate is used as the output, and the maximum value of the prediction probability is selected as the final strip shape quality prediction result of the strip shape quality anomaly category;

[0051] S1.5.12: Repeat S1.5.4 to S1.5.11 to establish strip shape quality prediction models for all strip shape quality anomaly categories respectively.

[0052] In the said step 2, the steel type information includes the steel type, target width, target length, and target thickness of the thick plate.

[0053] The strip shape quality prediction result of the thick plate includes the head and tail warping probability, the middle thickness anomaly probability, the middle wave probability, the left wave probability, and the right wave probability.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention fills in the data of the thick plate sample data, unifies the dimensions of the multi-source heterogeneous data, and solves the problem of low strip shape prediction accuracy caused by the inability to fuse the multi-source heterogeneous data generated in the thick plate production process. Compared with manual experience inference or existing prediction methods, the present invention effectively avoids problems such as information redundancy and loss of effective information, thereby improving the prediction accuracy of the thick plate strip shape.

[0056] 2. The present invention precisely intercepts the thick plate sample data, can extract the effective information in the temperature measurement interval data of the entire rolling line cooling process, and reserves a certain margin interval. By setting the judgment criteria for temperature jumps, it avoids the confusion between rolling line disturbance data and effective data, which is beneficial to the subsequent data filling and the establishment of the flatness quality prediction model, thereby further improving the prediction accuracy of the thick plate flatness.

[0057] 3. The present invention divides the multi-source heterogeneous data, cuts it into different data blocks according to its different data dimensions, and establishes a neural network based on deep learning through statistical characterization, correlation coefficient, and feature extraction to form a flatness quality prediction model, effectively fusing data such as one-dimensional index data, two-dimensional temperature data, and three-dimensional rolling data in the thick plate production process. Compared with the prior art, the present invention solves the problems of limited data processing volume and lack of flexibility, and the flatness quality prediction model has good robustness.

[0058] In the production process of thick plates, the present invention can establish a flatness quality prediction model based on complex multi-source heterogeneous industrial data using deep learning methods, accurately predict the final flatness quality of thick plates, be used to understand in advance the product performance of thick plates of various specifications, and be used to guide the production process. Thus, it can greatly improve the product quality by timely correcting parameters, etc., improve the qualified rate of bad reviews, and is applicable to the rolling production of thick plates with a thickness of 20 - 60 mm. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is the flowchart of the thick plate flatness prediction method based on industrial big data of the present invention;

[0060] Figure 2 is the flatness prediction system architecture diagram of the thick plate flatness prediction method based on industrial big data of the present invention;

[0061] Figure 3 is the temperature interval diagram of the cooling process data without interception in the thick plate flatness prediction method based on industrial big data of the present invention;

[0062] Figure 4 is the temperature interval diagram of the cooling process data after interception in the thick plate flatness prediction method based on industrial big data of the present invention;

[0063] Figure 5 is the statistical schematic diagram of the prediction accuracy of the thick plate flatness prediction method based on industrial big data of the present invention.

[0064] In the figure, 1 is the front-end interface, 2 is the data preprocessing module, 3 is the model training module, 4 is the model database, 5 is the model prediction module, and 6 is the industrial database. DETAILED DESCRIPTION OF THE INVENTION

[0065] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0066] Please refer to the attached Figure 2 , a thick plate shape prediction method based on industrial big data, which is implemented based on a shape prediction system. The shape prediction system includes:

[0067] A front-end interface 1 for inputting parameters and displaying the results of shape quality prediction;

[0068] A model database 4 for storing the shape quality prediction model and its model ID;

[0069] An industrial database 6 for storing multi-source heterogeneous industrial data in the thick plate production process;

[0070] A model training module 3 for offline training of the shape quality prediction model and storing the trained shape quality prediction model in the model database 4;

[0071] A model prediction module 5 for retrieving the shape quality prediction model from the model database 4 and performing shape quality prediction through the shape quality prediction model, and sending the shape quality prediction result to the front-end interface 1;

[0072] And a data preprocessing module 2 for retrieving data samples from the industrial database 6 and preprocessing the data samples before passing them into the model training module 3 and the model prediction module 5.

[0073] Please refer to the attached Figure 1 , the thick plate shape prediction method includes the following steps:

[0074] Step 1: Generate a shape quality prediction model for a specified thick plate through offline model training and store it in the model database 4.

[0075] Step 1.1: Store the historical data (i.e., multi-source heterogeneous industrial data) generated during the production of different thick plates in the industrial database 6, and use the SSM (Spring + SpringMVC + MyBatis) framework to extract the historical data from the industrial database 6 and form an API interface. Other historical data extraction methods can also be used, which will not be elaborated here.

[0076] Step 1.2: The user issues a quality prediction execution instruction for a specified thick plate through the front-end interface 1, that is, enters the thick plate number through the front-end interface 1. Each thick plate number corresponds to a type of thick plate. The data preprocessing module 2 retrieves the historical data of the specified thick plate from the industrial database 6 through the API interface.

[0077] Step 1.3: The data preprocessing module 2 performs information interception and data filling on historical data to form multi-source heterogeneous data blocks, and transmits the multi-source heterogeneous data blocks to the model training module 3.

[0078] During the cooling process of the thick plate, the actual effective data of the thick plate is mixed with the disturbance data on the rolling line, and there may be sliding when the thick plate is conveyed on the roller table. Deviations will occur if the temperature data is intercepted only based on the position information of the thick plate on the cooling line. Therefore, it is necessary to accurately locate and extract the temperature information of the steel plate from a continuous segment of temperature data.

[0079] The method by which the data preprocessing module 2 intercepts information from historical data is as follows:

[0080] S1.3.1: Intercept the temperature measurement intervals of several temperature sensors at the corresponding positions respectively through the position information of the thick plate. As shown in the attachment, in the figure, the abscissa is the position data of the thick plate, the ordinate is the temperature data of the thick plate, and p1 - p5 are temperature sensors. Preferably, the temperature sensors can use pyrometers of existing technologies. Figure 3 As shown, in the figure, the abscissa is the position data of the thick plate, the ordinate is the temperature data of the thick plate, and p1 - p5 are temperature sensors. Preferably, the temperature sensors can use pyrometers of existing technologies.

[0081] Preferably, a certain margin is reserved when intercepting the temperature measurement intervals of each of the temperature sensors, and the length of this margin is 5% of the length of the temperature measurement interval. By reserving a certain margin at one or both ends of the temperature measurement interval, the effectiveness and integrity of the data during the cooling process of the entire rolling line can be ensured, thereby ensuring the accurate interception of the temperature information during the cooling process.

[0082] S1.3.2: Set a temperature jump reference value as the judgment criterion for temperature jump, and intercept the temperature data information of each temperature sensor within the temperature measurement interval according to the temperature jump reference value, as shown in the attachment. By setting the temperature jump judgment criterion, the interference of data information such as rolling line disturbance data on the temperature data can be effectively avoided, thereby ensuring the accurate interception of the temperature information during the cooling process. Figure 4 As shown, in the figure, the abscissa is the position data of the thick plate, the ordinate is the temperature data of the thick plate, and p1 - p5 are temperature sensors. Preferably, the temperature sensors can use pyrometers of existing technologies. By setting the temperature jump judgment criterion, the interference of data information such as rolling line disturbance data on the temperature data can be effectively avoided, thereby ensuring the accurate interception of the temperature information during the cooling process.

[0083] Preferably, within the temperature measurement interval of each temperature sensor, the temperature jump reference value is 50% of the temperature difference between the highest temperature and the lowest temperature in this temperature measurement interval.

[0084] During the cooling process, due to the different data dimensions of different samples and inaccurate temperature information acquisition, data filling needs to be performed before prediction, including data filling for the heating process, data filling for the rolling process, and data filling for the cooling process.

[0085] During the heating process, several pyrometers are distributed in the heating furnace, jointly characterizing the furnace temperature change in each heating stage. The method for complementing the heating process data is as follows: extract the in-furnace time of a thick plate sample, and respectively intercept the furnace temperature data corresponding to several temperature sensors according to the in-furnace time, and arrange the furnace temperature data in chronological order to form two-dimensional furnace temperature data for this thick plate sample. Since the in-furnace time of each thick plate sample is different, the amount of furnace temperature data extracted from different thick plate samples is different. In order to facilitate subsequent model processing, it is necessary to complement the furnace temperature data of different thick plate samples. Taking the maximum amount of furnace temperature data among all thick plate samples as the standard, zeros are complemented around the furnace temperature data in the remaining thick plate samples to unify the dimensions of the furnace temperature data in all thick plate samples. Since the zero data does not contain any features, the data dimensions can be unified without affecting the characteristics of the original data itself.

[0086] The rolling process data can be represented from three dimensions, namely the pass dimension (i.e., as production time progresses, the data generated by the thick plate in each rolling pass), the index dimension (i.e., on the spatial scale, the control indices generated simultaneously in each pass, such as rolling force, roll bending force, etc.), and the measurement dimension (i.e., multiple original measurement data of the thick plate under each index), forming a three-dimensional data block. Different thick plates have the same indices and measurement data in each pass, but the number of rolling passes for different specifications of thick plates is different, resulting in different dimensions of rolling data among different thick plate samples. The method for complementing the rolling process data is as follows: taking the maximum amount in the pass dimension among all thick plate samples as the standard, random data is complemented in the pass dimension of the remaining thick plate samples to unify the pass dimension of all thick plate samples and ensure that the characteristics of the original data itself are not lost. Preferably, the random data is randomly generated data within the range of 0 - 1, used to complement the data to unify the pass dimension of the thick plate sample. The data between 0 - 1 hardly affects the characteristics of the original data and is essentially the same as zero; zeros can also be directly used to complement the data.

[0087] During the cooling process, the data includes temperature data, water flow rate data of the upper and lower ducts, etc. Since the running speeds of different specifications of thick plates are different, the dimensions of the data intercepted during the steel plate cooling process are not unified. The method for complementing the cooling process data is as follows: taking the maximum amount of data dimensions among all thick plate samples as the standard, random data is complemented in the data dimensions of the remaining thick plate samples to unify the data dimensions of all thick plate samples and not affect the characteristics of the original data. Preferably, the random data is randomly generated data within the range of 0 - 1, used to complement the data to unify the pass dimension of the thick plate sample. The data between 0 - 1 hardly affects the characteristics of the original data and is essentially the same as zero; zeros can also be directly used to complement the data.

[0088] Step 1.4: The model training module 3 classifies the shape quality anomalies according to production requirements such as shape quality rules and final shape quality data (abbreviation of FQC, Finish or Final Quality Control, i.e., finished product quality inspection), and establishes anomaly criteria for each shape quality anomaly category respectively.

[0089] The two ends of the thick plate in the length direction are taken as the head and the tail; the two ends of the thick plate in the width direction are taken as the left side and the right side. Both the left side and the right side are located between the head and the tail and are symmetrically arranged; the remaining area of the thick plate is the middle area. The categories of the shape quality anomalies include: head and tail warping, thickness anomaly, center wave, left wave and right wave.

[0090] The anomaly criterion for the head and tail warping is: the FQC data of the thick plate head is over the limit, or the FQC data of the thick plate tail is over the limit, or there is a wave shape at the thick plate head, or there is a wave shape at the thick plate tail. The anomaly criterion for the thickness anomaly is: the FQC data of the middle area of the thick plate is over the limit (for example, the thickness deviation of the middle area exceeds -6 to 2 mm). The anomaly criterion for the center wave is: the FQC data of the middle area of the thick plate is not over the limit (for example, the thickness deviation of the middle area is within the range of -6 to 2 mm), but there is a wave shape in the longitudinal direction of the thick plate. The anomaly criterion for the left wave is: the FQC data of the left side area of the thick plate is over the limit (for example, the thickness deviation of the left side area exceeds -6 to 2 mm) or there is a wave shape. The anomaly criterion for the right wave is: the FQC data of the right side area of the thick plate is over the limit (for example, the thickness deviation of the right side area exceeds -6 to 2 mm) or there is a wave shape.

[0091] Step 1.5: The model training module 3 establishes a shape quality prediction model for each shape quality anomaly category through multi-source heterogeneous data blocks.

[0092] S1.5.1: According to the different characteristics of the data under different processes, a multi-source heterogeneous data block is sliced into several small data blocks.

[0093] The number of small data blocks sliced from each multi-source heterogeneous data block is determined according to the actual production process of the thick plate. For example: for the furnace temperature data of the thick plate during the heating stage, the horizontal direction represents the furnace temperature distribution in the heating furnace from different orientations at the same moment during the same heating stage of the thick plate. The vertical direction represents the change of the furnace temperature over time during the time the thick plate is in the heating furnace. The furnace temperature data of the thick plate is divided into 5 parts in the vertical direction; for the rolling data generated during the rolling process of the thick plate, since the rolling process includes the pass dimension and is divided into rough rolling and finish rolling, according to the mechanism experience, the influence of rough rolling on the plate shape is relatively small, while the influence of the last three stages of finish rolling on the plate shape is relatively large. Therefore, the rough rolling is regarded as one part, the first 3 pass stages of finish rolling are regarded as one part, and the last 3 pass stages and the intermediate pass of finish rolling are regarded as one part; for the two-dimensional temperature field data of the thick plate after cooling, the horizontal direction represents the temperature distribution of the thick plate in the width direction, and the vertical direction represents the temperature distribution of the steel plate in the length direction. Therefore, the two-dimensional temperature field data of the thick plate is divided into 3 parts horizontally and 5 parts vertically.

[0094] S1.5.2: Characterize each small data block using the maximum value, minimum value, average value, and variance to form statistical data.

[0095] S1.5.3: According to the actual association order of the processes, combine the statistical data into a one-dimensional new statistical index data, and use it as an input to the prediction model. By setting and combining the statistical quantities of the data, it can ensure that the association information hidden between different small data blocks is not lost, ensuring the integrity of the data.

[0096] S1.5.4: Use the maximum information coefficient (MIC) algorithm to calculate the correlation coefficient m (MIC ∈ [0, 1]) between the new statistical index data and each type of plate shape quality anomaly category.

[0097] S1.5.5: Sum the correlation coefficients of each small data block to obtain the correlation coefficient M between different data blocks and the plate shape quality anomaly category, M = {m 1 , m 2 , …, m n}, where m n is the correlation coefficient between the nth small data block and the plate shape quality anomaly category, m n = m n ave + m n std + m n max + m n min ; among them, m n aveis the correlation coefficient between the average value of the nth small data block and the strip shape quality anomaly category, m n std is the correlation coefficient between the variance of the nth small data block and the strip shape quality anomaly category, m n max is the correlation coefficient between the maximum value of the nth small data block and the strip shape quality anomaly category, m n min is the correlation coefficient between the minimum value of the nth small data block and the strip shape quality anomaly category.

[0098] S1.5.6: Sum the correlation coefficients between several small data blocks belonging to the same multi-source heterogeneous data block and the strip shape quality anomaly category to obtain the correlation coefficient G between different multi-source heterogeneous data blocks and the strip shape quality anomaly category.

[0099] S1.5.7: Perform feature extraction on each small data block respectively. The statistical data of all small data blocks form one-dimensional time series data, which respectively represent the statistical quantities of the surface temperature, center temperature, temperature uniformity, furnace outlet temperature, and furnace temperature during the heating stage of the thick plate at each stage, and the statistical quantities of the rolling force, bending roll force, roll gap distance, torque, output thickness, output temperature, etc. of each pass during the rolling process of the thick plate; the statistical quantities of the thick plate temperature of each one-dimensional temperature sensor and two-dimensional temperature field during the cooling process of the thick plate, etc. For each small data block, establish a convolutional neural network model with the small data block as the input and the strip shape quality anomaly category as the output. Preferably, the convolutional neural network can adopt the GoogleNet-Inception V4 network of the existing technology.

[0100] S1.5.8: According to the hierarchical structure of the convolutional neural network model, select the features in the feature extraction layer of the convolutional neural network model as the new features (i.e., variables) of the small data block, which can extract the data information hidden in the original data, filter redundant and interference information, and play the role of feature filtering and generation.

[0101] S1.5.9: Repeat S1.5.7 to S1.5.8, select all multi-source heterogeneous data blocks as the input of the convolutional neural network model respectively, and extract the new features of the small data blocks in the multi-source heterogeneous data blocks.

[0102] S1.5.10: Combine all the new features extracted from the multi-source heterogeneous data blocks into the data input layer, and assign feature weights to each new feature in the data input layer according to the correlation coefficient G. The feature weight assigned to the ith new feature is λ i , where n is a natural number and j is an integer in the interval [1, n].

[0103] S1.5.11: Establish a flatness quality prediction model based on a single-layer neural network. Multiply the newly extracted features by the feature weight λ, and use the result as the input of the flatness quality prediction model. The prediction probability of the abnormal flatness quality category of the thick plate is used as the output, and the maximum value of the prediction probability is selected as the final flatness quality prediction result of the flatness quality abnormal category. i Multiply them and use the result as the input of the flatness quality prediction model. The prediction probability of the abnormal flatness quality category of the thick plate is used as the output, and the maximum value of the prediction probability is selected as the final flatness quality prediction result of the flatness quality abnormal category.

[0104] S1.5.12: Repeat S1.5.4 to S1.5.11 to establish flatness quality prediction models for all flatness quality abnormal categories respectively.

[0105] Step 1.6: Save the flatness quality prediction model corresponding to the specified thick plate in the model database 4 and assign a corresponding model ID.

[0106] Step 2: The user inputs the thick plate number through the front-end interface 1. The data preprocessing module 2 retrieves the steel type information corresponding to the thick plate number from the industrial database 6 and displays it through the front-end interface 1.

[0107] The steel type information includes the steel type, target width, target length, target thickness, etc. of the thick plate, which is convenient for distinguishing thick plate categories.

[0108] Step 3: The user inputs a continuous production time through the front-end interface 1. The model training module 3 retrieves the model IDs of the available flatness quality prediction models within the continuous production time from the model database 4 and displays them through the front-end interface 1. The user selects a model ID.

[0109] Step 4: The model prediction module 5 retrieves the flatness quality prediction model corresponding to the model ID selected in Step 3 from the model database 4, loads the flatness quality prediction model and initializes it.

[0110] Step 5: The data preprocessing module 2 retrieves the multi-source heterogeneous industrial data during the production process of the thick plate corresponding to the thick plate number in Step 2 from the industrial database 6, passes it to the model prediction module 5, and displays it through the front-end interface 1 at the same time.

[0111] Step 6: The model prediction module 5 performs flatness quality prediction on the thick plate corresponding to the thick plate number in Step 2 through the flatness quality prediction model loaded in Step 4, and displays the flatness quality prediction result of the thick plate through the front-end interface 1.

[0112] The flatness quality prediction result of the thick plate includes the head and tail warping probability, the middle thickness abnormal probability, the middle wave probability, the left wave probability, and the right wave probability.

[0113] Example 1:

[0114] Perform flatness quality prediction on multiple thick plates that did not participate in offline training. The multiple thick plates are thick plate samples from March to May 2019, including thick plates of multiple steel grades. The flatness quality abnormal categories include five types: head and tail warping abnormality, middle thickness abnormality, middle wave, left wave, and right wave. Establish abnormal standards for each flatness quality abnormal category. Use the thick plate samples from April 1, 2019 to May 31, 2019 as the training set, and generate a flatness quality prediction model for the specified thick plates through offline training of the model and store it in the model database 4. Take the thick plate samples from June 1, 2019 to June 10, 2019 for ten days as the test set and divide them into 5 small test sets to verify the flatness quality prediction results of the thick plates.

[0115] Step 1: Generate a flatness quality prediction model for the thick plate samples from April 1, 2019 to May 31, 2019 through offline training of the model and store it in the model database 4.

[0116] Step 1.1: Store the historical data generated during the production of all thick plates from April 1, 2019 to May 31, 2019 in the industrial database 6, and use the SSM framework to extract the historical data from the industrial database 6 and form an API interface.

[0117] Step 1.2: The user issues a quality prediction execution instruction for the specified thick plate through the front-end interface 1, and the data preprocessing module 2 retrieves the historical data of the specified thick plate from the industrial database 6 through the API interface.

[0118] Step 1.3: The data preprocessing module 2 performs information interception and data filling processing on the historical data, constructs a multi-source heterogeneous data block, and passes the multi-source heterogeneous data block into the model training module 3.

[0119] The data preprocessing module 2 performs information interception on the historical data:

[0120] S1.3.1: Intercept the temperature measurement intervals of 5 pyrometers at the corresponding positions through the position information of the thick plate. In this embodiment, the positions of the 5 pyrometers are as follows: the first one is at the cooling inlet, the second one is at the middle position of the cooling, and the last three are at the upper, lower, and right sides of the cooling outlet. The abscissa interval of the cooling process data during this production is [0s, 133s], and the ordinate interval is [97℃, 793℃].

[0121] S1.3.2: According to the standard that 50% of the temperature difference between the highest temperature and the lowest temperature in the temperature measurement interval is used as the temperature jump reference value, intercept the temperature data information of each one. Taking the first pyrometer as an example, the temperature difference between the highest temperature and the lowest temperature is 696℃, so the data with the ordinate value in the range of [445℃, 793℃] is retained during interception.

[0122] Data filling for the heating process: Extract the in-furnace time of a thick plate sample, and intercept the furnace temperature data corresponding to 5 pyrometers according to the in-furnace time. Arrange the furnace temperature data in chronological order to form two-dimensional furnace temperature data for this thick plate sample. Taking the maximum amount of furnace temperature data among all thick plate samples as the standard, fill in zeros around the furnace temperature data in the remaining thick plate samples to unify the dimensions of the furnace temperature data in all thick plate samples.

[0123] The rolling process data includes pass dimension, index dimension, and measurement dimension. Data filling for the rolling process: Taking the maximum amount of pass dimension among all thick plate samples as the standard, fill in 0.1 in the pass dimension of the remaining thick plate samples to unify the pass dimension in all thick plate samples.

[0124] The data in the cooling process includes temperature data and water flow data of the upper and lower ducts. Data filling for the cooling process: Taking the maximum amount of data dimension among all thick plate samples as the standard, fill in 0.1 in the data dimension of the remaining thick plate samples to unify the data dimension in all thick plate samples.

[0125] Step 1.4: The shape quality model training module 3 classifies the shape quality anomalies according to the shape quality rules and the final shape quality data, including: head and tail warping, thickness anomaly, center wave, left wave, and right wave, and establishes anomaly criteria for each shape quality anomaly category respectively.

[0126] Step 1.5: The shape quality model training module 3 establishes a shape quality prediction model for each shape quality anomaly category through multi-source heterogeneous data blocks.

[0127] S1.5.1: For the furnace temperature data in the heating stage of the training set, in the longitudinal direction, it represents the change of furnace temperature over time during the time of the thick plate in the heating furnace. Divide the furnace temperature data of the thick plate into 5 parts in the longitudinal direction. In the transverse direction, for the rolling data in the selected data, regard the data in the rough rolling stage as one part, the first 3 pass stages in the finish rolling stage as one part, and the last 3 pass stages and the middle pass in the finish rolling stage as one part, that is, divide it into 3 parts.

[0128] S1.5.2: Characterize each small data block using the maximum value, minimum value, average value, and variance to form statistical data.

[0129] S1.5.3: Combine the statistical data into a one-dimensional new statistical index data according to the actual association order of the processes, and use it as an input to the prediction model.

[0130] S1.5.4: Use the maximum information coefficient algorithm to calculate the correlation coefficient m (MIC ∈ [0,1]) between the new statistical index data and each shape quality anomaly category.

[0131] S1.5.5: Sum the correlation coefficients of each small data block to obtain the correlation coefficient M between different data blocks and the strip shape quality anomaly category, M = {m 1 , m 2 , …, m n}, where m n is the correlation coefficient between the nth small data block and the strip shape quality anomaly category, m n = m n ave + m n std + m n max + m n min ; where m n ave is the correlation coefficient between the average value of the nth small data block and the strip shape quality anomaly category, m n std is the correlation coefficient between the variance of the nth small data block and the strip shape quality anomaly category, m n max is the correlation coefficient between the maximum value of the nth small data block and the strip shape quality anomaly category, m n min is the correlation coefficient between the minimum value of the nth small data block and the strip shape quality anomaly category.

[0132] S1.5.6: Sum the correlation coefficients between several small data blocks belonging to the same multi-source heterogeneous data block and the strip shape quality anomaly category to obtain the correlation coefficient G between different multi-source heterogeneous data blocks and the strip shape quality anomaly category.

[0133] S1.5.7: Perform feature extraction on each small data block respectively. The statistical data of all small data blocks form one-dimensional time series data, which respectively represent the statistical data of the surface temperature, center temperature, temperature uniformity, furnace outlet temperature, and furnace temperature during the heating stage of the heavy plate, and the statistical data of indicators such as rolling force, bending roll force, roll gap distance, torque, output thickness, and output temperature in each pass during the rolling process of the heavy plate; the statistical data of the heavy plate temperature of each one-dimensional temperature sensor and two-dimensional temperature field during the cooling process of the heavy plate, etc. For each small data block, establish a GoogleNet-Inception V4 convolutional neural network model with the data block as the input and the strip shape quality anomaly category as the output.

[0134] S1.5.8: According to the hierarchical structure of the convolutional neural network model, select the features in the feature extraction layer of the convolutional neural network model as the new features of the small data block.

[0135] S1.5.9: Repeat S1.5.7 to S1.5.8, select all multi-source heterogeneous data blocks as the inputs of the convolutional neural network model respectively, and extract the new features of the small data blocks in the multi-source heterogeneous data blocks.

[0136] S1.5.10: Merge all the new features extracted from the multi-source heterogeneous data blocks into the data input layer, and assign weights λ to each new feature in the data input layer according to the correlation coefficient G. i ,

[0137] S1.5.11: Establish a flatness quality prediction model based on a single-layer neural network, multiply the extracted new features by the feature weights λ, and use them as the inputs of the flatness quality prediction model. The prediction probability of the abnormal flatness quality category of the thick plate is used as the output, and the maximum value of the prediction probability is selected as the final flatness quality prediction result of the flatness quality abnormal category. i and use them as the inputs of the flatness quality prediction model. The prediction probability of the abnormal flatness quality category of the thick plate is used as the output, and the maximum value of the prediction probability is selected as the final flatness quality prediction result of the flatness quality abnormal category.

[0138] S1.5.12: Repeat S1.5.4 to S1.5.11 to establish flatness quality prediction models for 5 flatness quality abnormal categories respectively.

[0139] S1.6: Save the flatness quality prediction models corresponding to all thick plates in the training set during the period from April 1, 2019 to May 31, 2019 in model database 4, and assign corresponding model IDs.

[0140] Step 2: For the 5 small test sets from June 1, 2019 to June 10, 2019, on June 1, 2019, the user inputs the thick plate number #1 through the front-end interface 1. The data preprocessing module 2 retrieves the steel grade information corresponding to the thick plate number from the industrial database 6. The target length of this steel grade is 38.24 meters, the target width is 2.43 meters, and the target thickness is 0.028 meters, and it is displayed through the front-end interface 1.

[0141] Step 3: The user inputs from 00:00:00 on April 1, 2019 to 00:00:00 on May 31, 2019 through the front-end interface 1. The model training module 3 retrieves the model IDs of the available flatness quality prediction models from April 1, 2019 to May 31, 2019 from the model database 4 and displays them through the front-end interface 1. The user selects the model ID, and the model ID is 2019-05-03 00:00:00.

[0142] Step 4: The model prediction module 5 retrieves the flatness quality prediction model corresponding to the model ID selected in Step 3 from the model database 4, loads the flatness quality prediction model and initializes it.

[0143] Step 5: The data preprocessing module 2 retrieves the multi-source heterogeneous industrial data in the thick plate production process corresponding to the thick plate number in Step 2 from the industrial database 6, and transmits it to the model prediction module 5, and simultaneously displays it through the front-end interface 1.

[0144] Step 6: The model prediction module 5 performs a flatness quality prediction on the thick plate corresponding to the thick plate number in Step 2 by using the flatness quality prediction model loaded in Step 4. The head and tail warping anomaly probability of this thick plate is output as 38% by the flatness quality prediction model, the middle thickness anomaly probability is 22%, the middle wave probability is 82%, the left wave probability is 41%, and the right wave probability is 35%. The maximum value of the prediction probability is selected as the final flatness quality prediction result of this flatness quality anomaly category. Then the final flatness quality prediction result is output as: the middle wave probability is 82%.

[0145] Before production, the alarm threshold can be set according to the production quality requirements. In this embodiment, the alarm threshold is 30%. Then the middle wave probability 82% > 30%, that is, the final flatness quality prediction result is a middle wave defect. According to the actual FQC data of this thick plate, it is obtained that the middle wave defect is prominent during the production of this thick plate, which is consistent with the flatness quality prediction result.

[0146] Embodiment 2:

[0147] Flatness quality predictions are respectively performed on multiple thick plates that did not participate in the offline training. The multiple thick plates are thick plate samples from April 2019 to June 2019 for three months, including thick plates of multiple steel grades. The flatness quality anomaly categories include five types: head and tail warping anomaly, middle thickness anomaly, middle wave, left wave, and right wave. Anomaly standards are respectively established for each flatness quality anomaly category. The thick plate data from April 1, 2019 to June 12, 2019 is used as the training set, and the model is offline trained to generate a flatness quality prediction model for the specified thick plate and store it in the model database 4. The thick plate data from June 13, 2019 to June 30, 2019 is taken as the test set and divided into 9 small test sets for verifying the flatness quality prediction results of the thick plates.

[0148] Step 1: Generate a flatness quality prediction model for the thick plate samples from April 1, 2019 to June 12, 2019 through model offline training and store it in the model database 4.

[0149] Step 1.1: Store the historical data generated during the production process of all thick plates from April 1, 2019 to June 12, 2019 in the industrial database 6, and use the SSM framework to extract the historical data from the industrial database 6 and form an API interface.

[0150] Step 1.2: The user issues a quality prediction execution instruction for the specified thick plate through the front-end interface 1, and the data preprocessing module 2 retrieves the historical data of the specified thick plate from the industrial database 6 through the API interface.

[0151] Step 1.3: The data preprocessing module 2 performs information interception and data filling on the historical data to form multi-source heterogeneous data blocks, and transmits the multi-source heterogeneous data blocks to the model training module 3.

[0152] The data preprocessing module 2 performs information interception on the historical data:

[0153] S1.3.1: Intercept the temperature measurement ranges of 5 pyrometers at the corresponding positions through the position information of the thick plate. In this embodiment, the positions of the 5 pyrometers are as follows: the first one is at the cooling inlet, the second one is at the middle position of the cooling, and the last three are at the upper, lower, and right sides of the cooling outlet. The abscissa range of the cooling process data in this production is [0s, 145s], and the ordinate range is [95°C, 807°C].

[0154] S1.3.2: According to the standard of taking 50% of the temperature difference between the highest temperature and the lowest temperature in the temperature measurement range as the temperature jump reference value, intercept the temperature data information of each pyrometer. Taking the first pyrometer as an example, the temperature difference between the highest temperature and the lowest temperature is 712°C, so the data with the ordinate value in the range of [551°C, 807°C] is retained during interception.

[0155] Data filling for the heating process: Extract the in-furnace time of a thick plate sample, and intercept the corresponding furnace temperature data of 5 pyrometers according to the in-furnace time, and arrange the furnace temperature data in chronological order into two-dimensional furnace temperature data for the thick plate sample. Taking the maximum amount of furnace temperature data among all thick plate samples as the standard, fill in zeros around the furnace temperature data in the remaining thick plate samples to make the dimensions of the furnace temperature data in all thick plate samples unified.

[0156] The rolling process data includes pass dimension, index dimension, and measurement dimension. Data filling for the rolling process: Taking the maximum amount of the pass dimension among all thick plate samples as the standard, fill in 0.2 in the pass dimension of the remaining thick plate samples to make the pass dimensions of all thick plate samples unified.

[0157] The data of the cooling process includes temperature data and water flow data of the upper and lower ducts. Data filling for the cooling process: Taking the maximum amount of data dimensions among all thick plate samples as the standard, fill in 0.2 in the data dimensions of the remaining thick plate samples to make the data dimensions of all thick plate samples unified.

[0158] Step 1.4: The model training module 3 classifies the shape quality anomalies according to the shape quality rules and the final shape quality data, including: head and tail warping, thickness anomaly, middle wave, left wave, and right wave, and establishes anomaly standards for each shape quality anomaly category respectively.

[0159] Step 1.5: The model training module 3 establishes a flatness quality prediction model for each flatness quality anomaly category through multi-source heterogeneous data blocks.

[0160] S1.5.1: For the furnace temperature data in the heating stage in the training set, which represents the change of the furnace temperature over time during the time of the thick plate in the heating furnace in the longitudinal direction, the furnace temperature data of the thick plate is divided into 5 parts in the longitudinal direction. In the transverse direction, for the rolling data in the selected data, the data in the rough rolling stage is regarded as one part, the data in the first 3 passes of the finish rolling stage is regarded as one part, and the data in the last 3 passes and the intermediate passes of the finish rolling stage is regarded as one part, that is, it is divided into 3 parts.

[0161] S1.5.2: Use the maximum value, minimum value, average value and variance to characterize each small data block to form statistical data.

[0162] S1.5.3: According to the actual association order of the processes, combine the statistical data into a one-dimensional new statistical index data and use it as an input of the prediction model.

[0163] S1.5.4: Use the maximum information coefficient algorithm to calculate the correlation coefficient m (MIC ∈ [0,1]) between the new statistical index data and each flatness quality anomaly category.

[0164] S1.5.5: Sum the correlation coefficients of each small data block to obtain the correlation coefficient M between different data blocks and the flatness quality anomaly category, M = {m 1 ,m 2 ,…,m n}, where m n is the correlation coefficient between the nth small data block and the flatness quality anomaly category, m n = m n ave + m n std + m n max + m n min ; where m n ave is the correlation coefficient between the average value of the nth small data block and the flatness quality anomaly category, m n std is the correlation coefficient between the variance of the nth small data block and the flatness quality anomaly category, m n max is the correlation coefficient between the maximum value of the nth small data block and the flatness quality anomaly category, m n min is the correlation coefficient between the minimum value of the nth small data block and the flatness quality anomaly category.

[0165] S1.5.6: Sum the correlation coefficients between several small data blocks belonging to the same multi-source heterogeneous data block and the strip shape quality anomaly categories to obtain the correlation coefficient G between different multi-source heterogeneous data blocks and the strip shape quality anomaly categories.

[0166] S1.5.7: Extract features from each small data block separately. The statistical data of all small data blocks form one-dimensional time series data, which respectively represent the statistical quantities of the surface temperature, center temperature, temperature uniformity, furnace outlet temperature, and furnace temperature during the heating stage of the heavy plate at each stage, and the statistical quantities of indicators such as rolling force, bending roll force, roll gap distance, torque, output thickness, and output temperature at each pass during the rolling process of the heavy plate; the statistical quantities of the heavy plate temperature of each one-dimensional temperature sensor and two-dimensional temperature field during the cooling process of the heavy plate, etc. For each small data block, establish a GoogleNet-Inception V4 convolutional neural network model with the data block as the input and the strip shape quality anomaly category as the output.

[0167] S1.5.8: According to the hierarchical structure of the convolutional neural network model, select the features in the feature extraction layer of the convolutional neural network model as the new features of the small data blocks.

[0168] S1.5.9: Repeat S1.5.7 to S1.5.8, select all multi-source heterogeneous data blocks as the input of the convolutional neural network model respectively, and extract the new features of the small data blocks in the multi-source heterogeneous data blocks.

[0169] S1.5.10: Merge all the new features extracted from the multi-source heterogeneous data blocks into the data input layer, and assign weights λ to each new feature in the data input layer according to the correlation coefficient G i ,

[0170] S1.5.11: Establish a strip shape quality prediction model based on a single-layer neural network, multiply the extracted new features by the feature weight λ i and use it as the input of the strip shape quality prediction model. The prediction probability of the strip shape quality anomaly category of the heavy plate is used as the output, and the maximum value of the prediction probability is selected as the final strip shape quality prediction result of the strip shape quality anomaly category.

[0171] S1.5.12: Repeat S1.5.4 to S1.5.11 to establish strip shape quality prediction models for 5 strip shape quality anomaly categories respectively.

[0172] S1.6: Save the strip shape quality prediction models corresponding to all heavy plates in the training set during the period from April 1, 2019 to June 12, 2019 in the model database 4, and assign corresponding model IDs.

[0173] For the 9 small test sets from June 13, 2019 to June 30, 2019:

[0174] Step 2: On June 14, 2019, the user inputs the heavy plate number #2 through the front-end interface 1. The data preprocessing module 2 retrieves the steel type information of this heavy plate from the industrial database 6, including: target width 2.428 m, target length 38.248 m, target thickness 0.01211 m, and displays it through the front-end interface 1.

[0175] Step 3: The user inputs from 00:00:00 on April 1, 2019 to 00:00:00 on June 12, 2019 through the front-end interface 1. The model training module 3 retrieves the model ID of the available flatness quality prediction model from April 1, 2019 to June 12, 2019 from the model database 4 and displays it through the front-end interface 1. The user selects the model ID, and the model ID is 2019-05-01 00:00:00.

[0176] Step 4: The model prediction module 5 retrieves the flatness quality prediction model corresponding to the model ID 2019-05-01 00:00:00 from the model database 4, loads the flatness quality prediction model and initializes it.

[0177] Step 5: The data preprocessing module 2 retrieves the multi-source heterogeneous industrial data in the production process of the heavy plate corresponding to the heavy plate number in Step 2 from the industrial database 6, passes it to the model prediction module 5, and simultaneously displays it through the front-end interface 1.

[0178] Step 6: The model prediction module 5 performs flatness quality prediction on the heavy plate corresponding to the heavy plate number in Step 2 through the flatness quality prediction model loaded in Step 4. Through the flatness quality prediction model, the head and tail warping abnormality probability of this heavy plate is 67%, the middle thickness abnormality probability is 62%, the middle wave probability is 97%, the left wave probability is 51%, and the right wave probability is 55%. The maximum value of the prediction probability is selected as the final flatness quality prediction result of the flatness quality abnormality category. Then the final flatness quality prediction result is output: middle wave probability 97%.

[0179] Before production, the alarm threshold can be set according to the production quality requirements. In this embodiment, the alarm threshold is 30%. Then the middle wave probability is 97% > 30%, that is, the final flatness quality prediction result is a middle wave defect. According to the actual FQC data of this heavy plate, the middle wave defect is prominent during the production of this heavy plate, which is consistent with the flatness quality prediction result.

[0180] Similarly, on June 30, 2019, the user entered the thick plate number #3 through the front-end interface 1. The data preprocessing module 2 retrieved the steel type information corresponding to the thick plate number from the industrial database 6. The target length of this steel type was 37.53 meters, the target width was 3.12 meters, and the target thickness was 0.022 meters, and it was displayed through the front-end interface 1.

[0181] The user entered the time from 00:00:00 on April 1, 2019 to 00:00:00 on June 12, 2019 through the front-end interface 1. The model training module 3 retrieved the model ID of the available flatness quality prediction model from April 1, 2019 to June 12, 2019 from the model database 4 and displayed it through the front-end interface 1. The user selected the model ID, and the model ID was 00:00:00 on April 15, 2019.

[0182] The model prediction module 5 retrieved the flatness quality prediction model corresponding to the model ID 00:00:00 on April 15, 2019 from the model database 4, loaded the flatness quality prediction model and initialized it.

[0183] The data preprocessing module 2 retrieved the multi-source heterogeneous industrial data during the production process of the thick plate corresponding to the thick plate number in step 2 from the industrial database 6, passed it to the model prediction module 5, and at the same time displayed it through the front-end interface 1.

[0184] The model prediction module 5 performed flatness quality prediction on the thick plate corresponding to the thick plate number in step 2 through the loaded flatness quality prediction model. Through the flatness quality prediction model, the head and tail warping anomaly probability of this thick plate was 27%, the middle thickness anomaly probability was 17%, the middle wave probability was 19%, the left wave probability was 13%, and the right wave probability was 17%. The maximum value of the prediction probability was selected as the final flatness quality prediction result of this flatness quality anomaly category. Then the final flatness quality prediction result was output: the head and tail warping anomaly probability was 27%.

[0185] Before production, the alarm threshold can be set according to the production quality requirements. In this embodiment, the alarm threshold is 30%. Then the head and tail warping anomaly probability 27% < 30%, that is, it is considered that the final flatness quality prediction is normal. According to the actual FQC data of this thick plate, no defects occurred in the production of this thick plate, which is consistent with the flatness quality prediction result.

[0186] Statistical analysis was performed on the flatness quality prediction accuracy of 9 small test sets, as shown in the appendix Figure 5 shown, in the appendix Figure 5 In it, the abscissa is the test set serial number, and the ordinate is the flatness quality prediction accuracy (acc%). From the appendix Figure 5It can be seen that the accuracy rate of the shape prediction of thick plates by the thick plate shape prediction method of the present invention is stable at 75%-95%, with good prediction function, can be used to guide the production process of thick plates, and plays a certain role in improving product quality.

[0187] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Therefore, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the shape of thick plates based on industrial big data, characterized in that: This method is implemented based on a shape prediction system, and the shape prediction system includes: A front-end interface (1) for inputting parameters and displaying the results of shape quality prediction; A model database (4) for storing the shape quality prediction model and its model ID; An industrial database (6) for storing multi-source heterogeneous industrial data during the production process of thick plates; A model training module (3) for offline training the shape quality prediction model and storing the shape quality prediction model after the training is completed in the model database (4); A model prediction module (5) for retrieving the shape quality prediction model from the model database (4) and performing shape quality prediction through the shape quality prediction model, and sending the shape quality prediction results to the front-end interface (1); And a data preprocessing module (2) for retrieving data samples from the industrial database (6) and preprocessing the data samples and then passing them into the model training module (3) and the model prediction module (5); The method for predicting the shape of thick plates includes the following steps: Step 1: Generate a shape quality prediction model for a specified thick plate through offline model training and store it in the model database (4); Step 2: The user inputs the thick plate number through the front-end interface (1), and the data preprocessing module (2) retrieves the steel type information corresponding to the thick plate number from the industrial database (6) and displays it through the front-end interface (1); Step 3: The user inputs a continuous production time through the front-end interface (1), and the model training module (3) retrieves the model ID of the available shape quality prediction model during the continuous production time from the model database (4) and displays it through the front-end interface (1), and the user selects the model ID; Step 4: The model prediction module (5) retrieves the shape quality prediction model corresponding to the model ID selected in Step 3 from the model database (4), loads the shape quality prediction model and initializes it; Step 5: The data preprocessing module (2) retrieves the multi-source heterogeneous industrial data during the production process of the thick plate corresponding to the thick plate number in Step 2 from the industrial database (6) and passes it into the model prediction module (5), and at the same time displays it through the front-end interface (1); Step 6: The model prediction module (5) performs shape quality prediction on the thick plate corresponding to the thick plate number in Step 2 through the shape quality prediction model loaded in Step 4, and displays the thick plate shape quality prediction results through the front-end interface (1).

2. The method for predicting the shape of thick plates based on industrial big data according to claim 1, characterized in that: The said Step 1 includes: Step 1.1: Store the historical data generated during the production process of different thick plates in the industrial database (6); Step 1.2: The user issues a quality prediction execution instruction for a specified thick plate through the front-end interface (1), and the data preprocessing module (2) retrieves the historical data of the specified thick plate from the industrial database (6); Step 1.3: The data preprocessing module (2) performs information interception and data filling processing on the historical data, forms a multi-source heterogeneous data block, and passes the multi-source heterogeneous data block into the model training module (3). Step 1.4: The shape quality anomaly classification module (3) classifies the shape quality anomalies and establishes anomaly criteria for each shape quality anomaly category respectively; Step 1.5: The shape quality anomaly classification module (3) establishes a shape quality prediction model for each shape quality anomaly category through multi-source heterogeneous data blocks; Step 1.6: Save the shape quality prediction model corresponding to the specified thick plate in the model database (4) and assign a corresponding model ID.

3. The thick plate shape prediction method based on industrial big data according to claim 2, characterized in that: The method for the data preprocessing module (2) to intercept information from historical data is: S1.3.1: Intercept the temperature measurement intervals of several temperature sensors at the corresponding positions respectively through the position information of the thick plate; S1.3.2: Set a temperature jump reference value as the judgment criterion for temperature jump. Intercept the temperature data information of each temperature sensor in the temperature measurement interval according to the temperature jump reference value.

4. The thick plate shape prediction method based on industrial big data according to claim 3, characterized in that: When intercepting the temperature measurement interval of each temperature sensor, a certain margin is reserved, and the length of the margin is 5% of the length of the temperature measurement interval.

5. The thick plate shape prediction method based on industrial big data according to claim 3, characterized in that: In the temperature measurement interval of each temperature sensor, the temperature jump reference value is 50% of the temperature difference between the highest temperature and the lowest temperature in the temperature measurement interval.

6. The thick plate shape prediction method based on industrial big data according to claim 2, characterized in that: In step 1.3, data filling includes data filling for the heating process, data filling for the rolling process, and data filling for the cooling process; The method for data filling in the heating process is: extract the in-furnace time of a thick plate sample, and intercept the furnace temperature data corresponding to several temperature sensors respectively according to the in-furnace time. Arrange the furnace temperature data in chronological order into two-dimensional furnace temperature data for the thick plate sample; take the maximum amount of furnace temperature data among all thick plate samples as the standard, and fill in zeros around the furnace temperature data in the remaining thick plate samples to make the dimensions of the furnace temperature data in all thick plate samples unified; The method for data filling in the rolling process is: take the maximum amount in the pass dimension of all thick plate samples as the standard, and fill in random data in the pass dimension of the remaining thick plate samples to make the pass dimensions of all thick plate samples unified; The method for data filling in the cooling process is: take the maximum amount in the data dimension of all thick plate samples as the standard, and fill in random data in the data dimension of the remaining thick plate samples to make the data dimensions of all thick plate samples unified.

7. The thick plate shape prediction method based on industrial big data according to claim 2, characterized in that: In step 1.4, the categories of shape quality anomalies include: head and tail warping, thickness anomaly, middle wave, left wave, and right wave; The abnormal criteria for head and tail warping are as follows: the FQC data of the head of the thick plate exceeds the limit, or the FQC data of the tail of the thick plate exceeds the limit, or there is a waviness at the head of the thick plate, or there is a waviness at the tail of the thick plate; the abnormal criteria for thickness abnormality are: the FQC data of the middle area of the thick plate exceeds the limit; the abnormal criteria for middle waviness are: the FQC data of the middle area of the thick plate does not exceed the limit, but there is a waviness in the longitudinal direction of the thick plate; the abnormal criteria for left waviness are: the FQC data of the left side area of the thick plate exceeds the limit or there is a waviness; the abnormal criteria for right waviness are: the FQC data of the right side area of the thick plate exceeds the limit or there is a waviness.

8. The method for predicting the shape of thick plates based on industrial big data according to claim 2, characterized in that: The step 1.5 includes: S1.5.1: According to the different characteristics of the data under different processes, a multi-source heterogeneous data block is sliced into several small data blocks; S1.5.2: Use the maximum value, minimum value, average value and variance to characterize each small data block to form statistical data; S1.5.3: According to the actual association order of the processes, the statistical data is combined into a one-dimensional new statistical index data, and it is used as an input of the prediction model; S1.5.4: Use the maximum information coefficient algorithm to calculate the correlation coefficient m between the new statistical index data and each type of plate shape quality abnormality category, where the maximum information coefficient MIC ∈ [0, 1]; S1.5.5: Sum the correlation coefficients of each small data block to obtain the correlation coefficient M between different data blocks and the strip shape quality anomaly category, M = {m 1 , m 2 , …, m n}, where m n is the correlation coefficient between the nth small data block and the strip shape quality anomaly category, m n = m n ave + m n std + m n max + m n min ; where m n ave is the correlation coefficient between the average value of the nth small data block and the strip shape quality anomaly category, m n std is the correlation coefficient between the variance of the nth small data block and the strip shape quality anomaly category, m n max is the correlation coefficient between the maximum value of the nth small data block and the strip shape quality anomaly category, m n min is the correlation coefficient between the minimum value of the nth small data block and the strip shape quality anomaly category; S1.5.6: Sum the correlation coefficients between several small data blocks belonging to the same multi-source heterogeneous data block and the plate shape quality abnormality category to obtain the correlation coefficient G between different multi-source heterogeneous data blocks and the plate shape quality abnormality category; S1.5.7: Perform feature extraction on each small data block respectively. The statistical data of all small data blocks form one-dimensional time series data, and a convolutional neural network model with the small data block as the input and the plate shape quality abnormality category as the output is established; S1.5.8: According to the hierarchical structure of the convolutional neural network model, select the features in the feature extraction layer of the convolutional neural network model as the new features of the small data blocks; S1.5.9: Repeat S1.5.7 to S1.5.8, select all multi-source heterogeneous data blocks as the input of the convolutional neural network model respectively, and extract the new features of the small data blocks in the multi-source heterogeneous data blocks; S1.5.10: Merge all the newly extracted features from the multi-source heterogeneous data blocks into the data input layer, and assign feature weights to each new feature in the data input layer according to the correlation coefficient G. The feature weight assigned to the i-th new feature is λ i , where n is a natural number and j is an integer in the range of [1, n]; S1.5.11: Establish a flatness quality prediction model based on a single-layer neural network, multiply the extracted new features by the feature weight λ i and use the result as the input of the flatness quality prediction model, with the prediction probability of the flatness quality abnormal category of the thick plate as the output. Select the maximum value of the prediction probability as the final flatness quality prediction result of this flatness quality abnormal category; S1.5.12: Repeat S1.5.4 to S1.5.11 to establish a plate shape quality prediction model for each plate shape quality abnormality category respectively.

9. The method for predicting the shape of thick plates based on industrial big data according to claim 1, characterized in that: In the step 2, the steel type information includes the steel type, target width, target length, and target thickness of the thick plate.

10. The method for predicting the shape of thick plates based on industrial big data according to claim 1, characterized in that: The prediction result of the thick plate shape quality includes the probability of head and tail warping, the probability of abnormal middle thickness, the probability of middle waviness, the probability of left waviness, and the probability of right waviness.

Images