An online prediction method for annealing strip hardness based on multi-scale data fusion
Through multi-scale data fusion and one-dimensional convolutional neural network combined with adaptive parameter initialization Adaboost_R2 technology, the problem of online detection of strip steel hardness in cold rolling mills has been solved, the prediction accuracy and generalization capabilities are improved, production needs are met, and economics are enhanced. benefit.
Patent Information
- Application Number
- CN202210025657.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-11
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-01-11
AI Technical Summary
The existing technology cannot realize the online detection of strip hardness during continuous annealing of cold rolling mills, resulting in product quality failure and economic benefits affected. Traditional prediction methods based on macro data fail to consider mesoscopic data, resulting in low prediction accuracy and Poor generalization ability.
Using a multi-scale data fusion method, the tissue change process mechanism of strip steel during continuous annealing is introduced, the mesoscopic scale characteristics are calculated and the macroscopic scale data is fused. One-dimensional convolutional neural network and improved Adaboost_R2 technology are used to build the belt Steel hardness online prediction model.
It significantly improves the prediction accuracy of strip hardness, meets production needs, increases the economic benefits of the enterprise, and prevents model training from falling into local optimization through differential evolution algorithms, improving the model convergence speed and prediction accuracy.
Smart Images

Figure CN114564879B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of automatic control of cold rolling and annealing production processes in steel enterprises, and in particular to an online prediction method for the hardness of annealed strip steel based on multi-scale data fusion. Background Art
[0002] In the actual production process of continuous annealing in the cold rolling mill of steel enterprises, the hardness of the strip is one of the core indicators for measuring product quality and guiding production. In the actual production process, online detection of the hardness of the strip has not yet been realized. The hardness of the strip is mainly obtained by cutting the head and tail of the strip after annealing, and then conducting offline manual testing to determine the product quality. However, this method of obtaining the hardness of the strip by offline detection generally has a certain time lag, which brings obstacles to the precise control of the continuous annealing production process. Therefore, the hardness of the strip often fluctuates greatly, resulting in substandard product quality or even waste quality problems, which seriously affects the economic benefits of the cold rolling mill.
[0003] The paper "Design and Implementation of Strip Quality Prediction and Process Monitoring System for Continuous Annealing Unit Based on PLS" (Wang Yuan, Northeastern University, 2009) proposed a method based on partial least squares (PLS) The patent "A method for predicting the hardness of continuous annealing products online with integrated learning" (ZL201410843307.8) and the patent "A method for detecting the quality of continuous annealing strip based on hybrid integrated learning" (ZL201710159565.8) both proposed methods for predicting the quality of strip based on integrated learning. However, the process parameters used in these technologies are all macro production data, without considering the mesoscopic data with a greater correlation with the hardness of the strip in the actual production process, such as the initial grain size, recrystallization size and grain size after growth of the strip. When the macro continuous annealing process parameters are applied to the strip, it will cause the recrystallization and grain growth of the grains inside the strip material, thereby forming a new grain distribution, which directly determines the quality indicators such as the organization and hardness of the strip. The above traditional prediction technologies based on macro process parameters do not consider the mesoscopic information such as the internal organization structure of the strip, which leads to the inability of these traditional technologies to accurately predict the quality indicators such as the hardness of the strip, and it is also difficult to obtain a good generalization ability. Summary of the invention
[0004] In view of the shortcomings of the existing strip steel hardness prediction technology, the present invention provides an online prediction method for the hardness of continuously annealed strip steel based on multi-scale data fusion, which introduces the process mechanism of the organizational change of the strip steel in the continuous annealing process to calculate the mesoscopic scale characteristics related to the internal organizational structure of the strip steel, and then fuses them with the macroscopic scale production process data to form a macro-mesoscopic multi-scale fused data sample, and then uses a one-dimensional convolutional neural network as a sub-learning machine, combined with the improved Adaboost_R2 technology with an adaptive parameter initialization strategy, to construct an online prediction method for the hardness of continuously annealed strip steel based on multi-scale data fusion. The method can realize the online prediction of the hardness of the strip steel in the continuous annealing production process, greatly improve the prediction accuracy of the hardness of the continuously annealed strip steel, meet the actual production needs of the enterprise, and increase the economic benefits of the enterprise;
[0005] The technical solution of the present invention is as follows:
[0006] An online prediction method for the hardness of a strip steel strip based on multi-scale data fusion comprises the following steps:
[0007] S1: For a tempering degree of H j ,j∈{1,2,…,Z} strip steel products, where Z is the number of existing strip steel products with different quenching and tempering degrees, obtain the actual continuous annealing production data of the strip steel with this quenching and tempering degree in the recent T time period from the enterprise database used to store the actual continuous annealing production data, and obtain a set D1 containing N continuous annealing production samples, each of which contains p macro-scale information data items and an actual value of the strip steel hardness;
[0008] S2: for the continuous annealing production sample set D1 in S1, the strip mesoscopic scale information data items are supplemented and the data is normalized to obtain the sample set D2;
[0009] S3: Based on the sample set D2 processed by S2, an integrated learning prediction model for the hardness of continuous annealing strip steel based on macro-mesoscopic multi-scale data fusion is established, and the established prediction model is stored in the enterprise's continuous annealing strip steel product hardness prediction model library;
[0010] S4: Select the scheduling quality H of the strip steel to be continuously annealed from the continuous annealing strip steel product hardness prediction model library j Consistent strip product hardness prediction model to start continuous annealing production;
[0011] S5: Set the sampling period to T max, read the continuous annealing production status data from each sampling point on the continuous annealing production line in real time, and synthesize it with the corresponding strip information in the database into continuous annealing production process data, the production process data includes macro-scale data items such as strip information and continuous annealing production status, use the corresponding process mechanism model in S2 to calculate the meso-scale data items related to the internal grain size of the strip, merge the meso-scale data items with the macro-scale data items, and form a continuous annealing production process input vector containing macro-mesoscopic data items;
[0012] S6: Use the maximum and minimum values of each data item in the sample set D2 obtained in S2 to normalize the values of the corresponding data items in the process input vector;
[0013] S7: Based on the normalized process input vector, the strip steel product hardness prediction model selected in S4 is used to predict the strip steel product hardness;
[0014] S8: After the continuous annealing production of the strip steel product is completed, the strip steel head is intercepted for offline manual testing to obtain the actual hardness value of the strip steel product, and the actual hardness value is matched and synthesized with the continuous annealing process input vector containing macro data items recorded in S5 for the strip steel product to obtain a new continuous annealing production sample, which includes p macro-scale data items and an actual hardness value of the strip steel product;
[0015] S9: Update the scheduling quality to H j The method is as follows: for the scheduled quality strip steel, after the continuous annealing production samples in the latest O time period are collected, the continuous annealing production samples in the latest O time period are used to replace the continuous annealing production samples in the earliest O time period of the scheduled quality strip steel in the enterprise database;
[0016] S10: For the strip steel products of scheduling quality in Z, set the threshold G to judge each scheduling quality H i (i=1,2,…,Z) Whether the number of updated continuous annealing samples is greater than G, if it is satisfied, then according to the methods S1 to S3, based on the updated continuous annealing production samples of the scheduling quality strip steel, the hardness prediction model of the scheduling quality strip steel product is retrained and updated to ensure that it can track the latest operating conditions of the continuous annealing unit producing the scheduling quality strip steel product;
[0017] S11: When the continuous annealing production of a new strip steel product starts, S4 to S11 are repeatedly executed to realize the online prediction of the hardness of the existing Z types of continuously annealed strip steel products of different scheduling qualities;
[0018] The specific process of S2 includes the following steps:
[0019] S2.1: For each continuous annealing production sample in the sample set D1, use the corresponding mechanism model to calculate q mesoscopic data items related to the internal grain size of the strip steel, and fuse the q mesoscopic data items with the p macroscopic data items to form a continuous annealing macro-mesoscopic multi-scale fusion data sample set D2 containing p+q macro-mesoscopic information input items and 1 actual strip steel hardness output item;
[0020] S2.2: A new set D2 containing N continuous annealing production samples is obtained from S2.1, wherein each production sample contains p+q macro-mesoscopic information input items and 1 actual strip hardness output item; each input item in D2 is normalized to ensure the dimensional consistency between different input items, and the maximum and minimum values of the p+q input items in the sample set D2 are saved;
[0021] The specific process of S3 includes the following steps:
[0022] S3.1: One Dimensional Convolutional Neural Networks (OD-CNN) is used as the sub-learning machine, and the improved Adaboost_R2 technology with adaptive parameter initialization strategy is used to generate K OD-CNN models. The specific steps are as follows:
[0023] S3.1.1: Initialize the model counter k to 1, and assign the same weight value to all samples in the sample set D2 after S2 data processing, that is, the weight of each sample is
[0024] S3.1.2: Use resampling methods incorporating sample weights Perform sampling with replacement on the sample set D2 to obtain a new set containing N production samples
[0025] S3.1.3: Construct the kth OD-CNN model and initialize the parameters of the kth OD-CNN model based on the adaptive parameter initialization strategy. The specific steps are as follows:
[0026] S3.1.3.1: Determine whether the kth OD-CNN is the first model, that is, whether k is equal to 1. If k is equal to 1, use the random initialization method to initialize the parameters of the kth OD-CNN model and jump to S3.1.4. If k is not equal to 1, jump to S3.1.3.2;
[0027] S3.1.3.2: Determine whether k can be tested by the trend frequency T trend+1 is divisible. If it is not divisible, jump to S3.1.3.3. If it is divisible, use the Mann-Kendall trend test method to analyze the newly generated kT trend T to the k-1th trend Whether the performance of the OD-CNN meets the improvement conditions. If so, jump to S3.1.3.4; otherwise, jump to S3.1.3.5;
[0028] The specific steps of determining whether the performance meets the improvement conditions are as follows:
[0029] S3.1.3.2.1: For the newly generated kT trend T to the k-1th trend OD-CNN models, respectively, the sample sets obtained in S3.1.2 Calculate the root mean square error of each OD-CNN model and construct a trend A vector V of elements, each element in V is the root mean square error value of the corresponding OD-CNN model;
[0030] S3.1.3.2.2: Based on the vector V constructed in S3.1.3.2.1, calculate the test statistic Z zk ;
[0031] S3.1.3.2.3: Set the confidence level α and obtain the threshold Z by querying the standard normal distribution table 1-α / 2 , when |Z zk |>Z 1-α / 2 And Z zk >0, the newly generated kT trend T to the k-1th trend The performance of the OD-CNN model meets the improvement condition, otherwise the performance does not meet the improvement condition;
[0032] S3.1.3.3: Use the model parameters optimized by the trained k-1th OD-CNN model to initialize the kth OD-CNN model parameters, that is, directly use the model parameter values optimized by the k-1th OD-CNN as the initial values of the model parameters corresponding to the kth OD-CNN;
[0033] S3.1.3.4: When the newly generated T trend The performance of the kth OD-CNN meets the improvement conditions, and the parameter initialization method in S3.1.3.3 is continued to be used to initialize the parameters of the kth OD-CNN model;
[0034] S3.1.3.5: When the newly generated T trendIf the performance of the kth OD-CNN does not meet the improvement conditions, the crossover mutation operation in the differential evolution algorithm (DE) is used to generate new model parameters to initialize the kth OD-CNN model parameters. The specific steps are as follows:
[0035] S3.1.3.5.1: Select three different models from the first k-1 trained OD-CNNs, including the model with the best prediction performance, that is, The OD-CNN model with the smallest RMS error calculated above and the two randomly selected models are used to obtain the optimized parameters M of each model. best 、M e and M h ;
[0036] S3.1.3.5.2: The model parameters M after the three OD-CNNs are optimized based on S3.1.3.5.1 best 、M e and M h , using DE / best / 1 mutation mode and binomial crossover to generate new model parameters
[0037] S3.1.3.5.3: Use the model parameters generated by S3.1.3.5.2 Initialize the corresponding parameter values of the kth OD-CNN model;
[0038] S3.1.4: For the kth OD-CNN model after parameter initialization in S3.1.3, the sample set obtained by resampling in S3.1.2 Train on it and establish an OD-CNN model for hardness prediction of strip steel products;
[0039] S3.1.5: Update the model counter k=k+1, and adjust the sample weights according to the training errors of each production sample for the sample set D2 processed by S2.2 The specific method is: first increase the weight of samples with larger training errors, reduce the weight of samples with smaller training errors, and then normalize the new sample weights;
[0040] S3.1.6: Repeat S3.1.2 to S3.1.5 to establish K OD-CNN models for hardness prediction of continuous annealing strip products;
[0041] S3.2: Based on the K OD-CNN models established in S3.1, first use the "k-means" clustering algorithm to cluster, and then randomly select an OD-CNN model (i.e., a sub-learning machine model) from each cluster to build an integrated learning machine model; the specific steps are as follows:
[0042] S3.2.1: Based on the sample set D2 processed by S2, the K OD-CNN models established in S3.1 are evaluated using the two indicators of accuracy and dispersion, and then a data set D3 containing K samples is constructed, where the kth sample is a two-dimensional vector containing the two attribute values of "accuracy" and "dispersion"; the accuracy formula (1) and dispersion formula (2) of the kth OD-CNN model are calculated as follows:
[0043]
[0044]
[0045] Where N is the total number of samples in D2, K is the total number of OD-CNN models constructed in S3.1, is the hardness prediction value of the strip steel product of the i-th sample in D2 by the k-th OD-CNN model, r i is the true value of the hardness of the strip steel product of the i-th sample in D2, represents the average value of the hardness prediction of the strip steel product on the i-th sample in D2 by K OD-CNN models;
[0046] S3.2.2: Based on the dataset D3 constructed in S3.2.1, the K OD-CNN models established in S3.1 are clustered using the “k-means” algorithm to obtain the cluster partition C = {C1, C2, …, C l}, l is the number of clusters obtained after clustering;
[0047] S3.2.3: Cluster C obtained from S3.2.2 i (i=1,2,…,l), when C i When the number of OD-CNN models in is greater than m, m models are randomly selected as sub-learning machines of the integrated model; i When the number of OD-CNN models is less than or equal to m, C i All OD-CNN models in are used as sub-learning machines of the integrated model;
[0048] S3.3: For the S sub-learning machine models selected in S3.2, use the sub-learning machine weight optimization method based on differential evolution to optimize the weight v of each sub-learning machine model. i(i=1,2,…,S) is optimized, and according to the S sub-learning machine models and their respective weights, the S sub-learning machine models are integrated into an integrated learning machine model by the linear weighting method shown in formula (3). The integrated learning machine model is the hardness prediction model of the tempered strip steel product, and its output is the weighted composite value of the strip steel product hardness prediction values output by each sub-learning machine model;
[0049]
[0050] Among them, P fimal (x) is the predicted value of the strip product hardness of the integrated learning machine model for the production sample x, v i and p i (x) are the weight value of the i-th sub-learning machine model and the predicted value of the hardness of the strip steel product output by the i-th sub-learning machine model for the production sample x;
[0051] Beneficial technical effects
[0052] 1. The present invention fuses the macroscopic scale data of the continuous annealing production process (mainly including strip information data and continuous annealing production status data) with the mesoscopic scale data (mainly including grain size data) which has a great influence on the hardness of the strip to obtain multi-scale fusion data, uses a one-dimensional convolutional neural network as a sub-learning machine, and combines the improved Adaboost_R2 technology with an adaptive parameter initialization strategy to construct an online prediction method for the hardness of the continuous annealing strip based on multi-scale data fusion. The method of the present invention can realize the online prediction of the hardness of the strip in the continuous annealing production process, greatly improve the prediction accuracy of the hardness of the continuous annealing strip, meet the actual production needs of the enterprise, and increase the economic benefits of the enterprise;
[0053] 2. The present invention uses the crossover mutation operation in the differential evolution algorithm to generate new model parameters to initialize the model parameters of OD-CNN, which can prevent the OD-CNN model training from falling into the local optimum, accelerate the model convergence speed, and improve the model prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 It is a schematic diagram of the principle of an online prediction method for the hardness of a strip steel strip based on multi-scale data fusion according to an embodiment of the present invention;
[0055] Figure 2 This is a flow chart of an online prediction method for hardness of annealed strip steel based on multi-scale data fusion according to an embodiment of the present invention;
[0056] Figure 3 A one-dimensional convolutional network structure diagram of an online prediction method for hardness of annealed strip steel based on multi-scale data fusion according to an embodiment of the present invention;
[0057] Figure 4 This is a comparison result diagram of the predicted values of the online prediction method for the hardness of the continuous annealing strip steel based on multi-scale data fusion according to an embodiment of the present invention and the actual values of the corresponding test samples. DETAILED DESCRIPTION
[0058] The present invention will be further described below in conjunction with the accompanying drawings and embodiments;
[0059] Continuous annealing production is an important process in the cold rolling mill of steel enterprises. Figure 1 As shown in the figure, the continuous annealing production line can be divided into the following 7 main stages according to its function: heating furnace (HF), soaking furnace (SF), slow cooling furnace (SF), fast cooling furnace (FCF), over-aging furnace (OA), water quenching furnace (WQ) and skin pass mill; in the actual continuous annealing production process, the strip will pass through each furnace zone at a certain speed and complete the annealing treatment according to the set annealing process route, thereby eliminating the internal stress generated during the cold rolling process, and then pass through the skin pass mill to finally obtain high-quality strip steel;
[0060] Aiming at the current situation that the online detection of strip hardness cannot be realized in the actual continuous annealing production process, and the existing methods usually only consider the macroscopic scale data such as production process parameters and strip information parameters related to the strip hardness, while ignoring the problem of mesoscopic scale data with greater correlation with the strip hardness, the present invention proposes an online prediction method for the hardness of continuous annealing strip based on multi-scale data fusion; the working principle of the method is as follows Figure 1As shown in the figure: Firstly, the historical continuous annealing production data samples of the enterprise are used, and the corresponding mechanism model is used to supplement the mesoscopic data to obtain the continuous annealing macro-mesoscopic multi-scale fusion data. OD-CNN is used as the sub-learning machine, combined with the improved Adaboost_R2 technology with adaptive parameter initialization strategy, and the continuous annealing strip hardness integrated learning prediction model based on multi-scale data fusion is established for strips with different tempering degrees, and an offline continuous annealing strip product hardness prediction model library is formed; secondly, in the actual production process, the corresponding product hardness prediction model is selected from the offline established prediction model library according to the tempering degree of the strip to be continuously annealed; then, the real-time production process at each sampling point of the continuous annealing production line is obtained from the continuous annealing production process sampling and control system. The data is preprocessed and synthesized with the strip information to obtain the macro data of the continuous annealing production process, and then the mesoscopic data related to the internal grain size of the strip is calculated using the corresponding mechanism model, and then synthesized to obtain the real-time macro-mesoscopic multi-scale data of the continuous annealing production process; then, the real-time data is passed to the selected strip product hardness prediction model, so as to output the strip product hardness prediction value under the current working conditions to the on-site operator; in addition, the information parameters and production process data at the head of each strip roll will be stored, and when the offline hardness test results of the strip product come out, they will be synthesized with them to form a new continuous annealing production data sample, which will be stored in the enterprise database; when the number of new samples meets the requirements, the model will be retrained to ensure that the prediction model can track the latest working conditions of the continuous annealing unit;
[0061] This implementation is based on Figure 1 The continuous annealing production process shown in the figure produces a strip steel product of a certain scheduling quality, and according to the principle of the continuous annealing strip steel hardness prediction method based on multi-scale data fusion, the present invention proposes an online prediction method for the continuous annealing strip steel hardness based on multi-scale data fusion, such as Figure 2 As shown, the following steps are included:
[0062] S1: For a type of strip steel with scheduling quality H1, continuous annealing production samples within the last 30 days are collected from the enterprise database, and a set D1 containing 5945 continuous annealing production samples is obtained; each sample consists of the following 24 macro data items, including 5 strip steel specification information data items, namely, inlet width, inlet thickness, strip steel carbon content, strip steel nitrogen content, strip steel silicon content, and 5 hot rolling stage information data items, namely, slab furnace temperature, average curling temperature, average finishing rolling temperature, average final rolling temperature and continuous pickling-rolling extension temperature. elongation (CDCM elongation); 14 continuous annealing related production status data items, namely, central section speed, HF furnace average temperature, SF furnace average temperature, SCF furnace average temperature, CF furnace cooling gas temperature, 1OA furnace zone average temperature, 2OA furnace zone 1 average temperature, 2OA furnace zone 2 average temperature, WQ furnace water temperature, leveler mill entrance strip tension, leveler mill intermediate strip tension, leveler mill exit strip tension, leveler mill 1# stand rolling force and 2# stand rolling force; and 1 output item, namely, the actual value data item of strip product hardness;
[0063] S2: Use the corresponding process mechanism model to calculate two mesoscopic data items related to the grain size inside the strip, namely the recrystallized grain size formula (4) and the grain size formula (5) after growth, and fuse them with the previous 24 macroscopic data items to form a continuous annealing macro-mesoscopic multi-scale fusion data sample set containing 26 input items and 1 output item; as shown in Table 1, each row represents a sample, each column is a data item, and they are arranged in the order of strip specification information data item, continuous annealing related production status data item, continuous annealing strip mesoscopic data item to the last strip product hardness actual value data item; S i represents the sample, S i =(s i1 ,s i2 ,...,s i26 ), where i = 1, 2, 3, ..., 5945, s ik Represents the kth data item of the i-th sample; the process mechanism model used is as follows:
[0064]
[0065] Where d is the recrystallized grain size, ε cr is the rolling pressure, d0 is the initial grain size, A, p1 and q1 are constants, which are 1.5, 0.67 and -0.5 respectively;
[0066]
[0067] where d θis the grain size after growth, ξ is the holding time, T is the average temperature of each furnace zone, R is a constant, which is taken as 16.5;
[0068] Table 1. Macro-mesoscopic data sample set of continuous annealing of H1 tempered strip steel products
[0069]
[0070] Each input item in Table 1 is normalized to obtain the data set D2, ensuring the dimensional consistency between different input items, and saving the maximum and minimum values of each input item in Table 1;
[0071] S3: Based on the data set D2 obtained in S2, an integrated learning prediction model for the hardness of continuous annealing strip steel based on macro-mesoscopic multi-scale fusion data is established, and the established prediction model is stored in the enterprise's offline continuous annealing strip steel product hardness prediction model library. The specific steps are as follows:
[0072] S3.1: One-dimensional convolutional neural network (OD-CNN) is used as the sub-learning machine, and the improved Adaboost_R2 technology with adaptive parameter initialization strategy is used to generate 200 OD-CNN models. The network structure of the OD-CNN model is as follows: Figure 3 As shown, other parameter settings are shown in Table 2;
[0073] Table 2. One-dimensional convolutional neural network hyperparameter settings
[0074] parameter Settings parameter Settings Learning Rate 0.001 Batch Normalization Method BatchNorm1d Number of training sessions 500 Activation Function ReLU Batch quantity 50 Loss Function MSELoss Gradient Descent Method Adam
[0075] S3.2: For the 200 OD-CNN models established in S3.1, first use the "k-means" clustering algorithm to cluster them, and then randomly select the OD-CNN model (i.e., the sub-learning machine model) used to build the integrated learning machine model from each cluster. The specific steps are as follows:
[0076] S3.2.1: For each OD-CNN model generated in S3.1, the following formulas are used to calculate the accuracy (6) and dispersion (7) of the model:
[0077]
[0078]
[0079] Among them, N = 5945 is the total number of samples, K = 200 is the total number of OD-CNN models, is the hardness prediction value of the strip steel product of the i-th sample by the k-th OD-CNN model, r i is the true value of the hardness of the strip steel product of the i-th sample, represents the average value of the hardness prediction value of the strip steel product on the i-th sample of the 200 OD-CNN models; construct an indicator data set D3 containing 200 samples, as shown in Table 3;
[0080] Table 3. Sample set of evaluation indicators for one-dimensional convolutional neural model
[0081] serial number accuracy Dispersibility 1 0.116 3.790 2 0.113 3.719 3 0.122 3.961 … … … 199 0.107 3.398 200 0.107 3.290
[0082] S3.2.2: Based on the indicator dataset D3 constructed in S3.2.1, the 200 OD-CNN models are clustered using the “k-means” clustering algorithm to obtain 5 clusters C = {C1, C2, …, C5};
[0083] S3.2.3: For each cluster C i (i=1,2,…,5), when C i When the number of OD-CNN models in is greater than 3, three of them are randomly selected as sub-learning machines of the integrated model; i When the number of OD-CNN models is less than or equal to 3, C i All OD-CNN models in the ensemble are used as sub-learning machines of the ensemble model; finally, 13 OD-CNN models are selected from 200 OD_CNN models as sub-learning machines of the ensemble model through this method;
[0084] S3.3: For the 13 sub-learning machine models selected by clustering, the weight v of each sub-learning machine model is optimized using the sub-learning machine weight optimization method based on differential evolution. i (i=1,2,…,13) is optimized based on the 13 sub-learning machine models established and the weight values of the corresponding models after optimization, according to the formula The strip hardness prediction results of the integrated learning machine model are obtained, where p k (x) is the predicted value of sample x calculated by the kth OD-CNN model, v k is the weight after optimization of the kth OD-CNN model; the numerical simulation results show that the root mean square error of the hardness prediction model of the quenched and tempered strip steel product obtained after the integration of 13 sub-learning machine models on the training set is 0.947, and the relative average error is 0.055%;
[0085] In addition to the 5945 training samples used in the last 30 days, 50 strip steel product hardness samples that were not involved in the training were collected from the enterprise database as test samples for the hardness prediction model of the quenched and tempered strip steel products to verify the effectiveness of the proposed integrated learning prediction model for the hardness of the annealed strip steel based on macro-mesoscopic data fusion; the comparison results between the predicted values and the actual values of the 50 test samples are shown in Tables 5 and Figure 4As shown, it can be seen that the method proposed in the present invention has a root mean square error of 6.97 and a relative average error of 0.924% in the prediction performance of the hardness of the continuous annealing product for the test samples that do not participate in the learning, and the change trend of the predicted value is basically consistent with the actual hardness value of the strip steel product, indicating that the integrated learning prediction model for the hardness of the continuous annealing strip steel based on macro-mesoscopic data fusion proposed in the present invention has good generalization ability and robustness;
[0086] Table 5 Comparison of prediction results of the test sample and the ensemble learning prediction model of annealing strip hardness based on macro-mesoscopic data
[0087]
[0088] S4: In the actual cold rolling continuous annealing production process, a strip steel product hardness prediction model consistent with the current production strip scheduling quality (for example, H1) is selected from the continuous annealing strip steel product hardness prediction model library to start continuous annealing;
[0089] S5: Read the production status data generated when the head of each strip passes through from each sampling point of the continuous annealing production process sampling and control system with a sampling period of 2 seconds, and synthesize it with the strip information into continuous annealing production process data, that is, 24 macro-scale data items of the continuous annealing sample. Then, use the corresponding process mechanism model, that is, formula (4) and formula (5), to calculate two meso-scale data items related to the grain size inside the strip, namely the recrystallized grain size and the grain size after growth, and merge them with the previous 24 macro-data items to form a continuous annealing macro-mesoscopic data online sample containing 26 input items, and record the sample at the same time;
[0090] S6: using the maximum and minimum values of each input item saved in S2 to normalize the values of each input item of the newly collected online sample;
[0091] S7: Based on the normalized online samples, the strip steel product hardness prediction model with scheduling quality H1 is used to predict the corresponding strip steel product hardness value, and the hardness value is displayed to the on-site operator through the operating platform display device;
[0092] S8: After the strip steel is produced, the strip steel head is cut off for offline testing to obtain the actual hardness value of the strip steel product, and the hardness value is matched and synthesized with the 24 macro data items of the strip steel product recorded in S5 to obtain a new continuous annealing production sample, which is saved in the enterprise server;
[0093] S9: Update the continuous annealing production sample of the scheduling quality H1. The specific method is: for the strip steel of scheduling quality H1, use the continuous annealing production sample of the latest month to replace the continuous annealing production sample of the earliest month of the scheduling quality in the enterprise database;
[0094] S10: For three strip products with different scheduling qualities, determine the H i (i=1,2,3) whether the number of updated continuous annealing samples is greater than 5500. If so, the hardness prediction model of the quenched and tempered strip steel product is retrained and updated according to the updated continuous annealing production samples according to the methods S1 to S3, so as to ensure that the hardness prediction model of the continuous annealing strip steel product of the scheduling quality can track the latest operating conditions of the continuous annealing unit producing the scheduling quality strip steel product;
[0095] S11: When the continuous annealing production of a new scheduling quality strip steel product begins, S4 to S11 are repeatedly executed to realize the online prediction of the hardness of three continuously annealed strip steel products of different scheduling qualities.
Claims
1. An online prediction method for the hardness of annealed strip steel based on multi-scale data fusion, characterized in that: It is divided into the following steps: S1: For a tempering degree of H j ,j∈{1,2,…,Z} strip steel products, where Z is the number of existing strip steel products with different tempering degrees, and the actual continuous annealing production data of the strip steel with this tempering degree in the recent T time period is obtained from the enterprise database used to store the actual continuous annealing production data, and a set D1 containing N continuous annealing production samples is obtained, each production sample contains p macro-scale information data items and an actual value of the strip steel hardness; S2: for the continuous annealing production sample set D1 in S1, the strip mesoscopic scale information data items are supplemented and the data is normalized to obtain the sample set D2; S3: Based on the sample set D2 processed by S2, an integrated learning prediction model for the hardness of continuous annealing strip steel based on macro-mesoscopic multi-scale data fusion is established, and the established prediction model is stored in the enterprise's continuous annealing strip steel product hardness prediction model library; S4: Select the scheduling quality H of the strip steel to be continuously annealed from the continuous annealing strip steel product hardness prediction model library j Consistent strip product hardness prediction model to start continuous annealing production; S5: Set the sampling period to T max , read the continuous annealing production status data from each sampling point on the continuous annealing production line in real time, and synthesize it with the corresponding strip steel information in the database into continuous annealing production process data, the production process data includes strip steel information and continuous annealing production status macro-scale data items, use the corresponding process mechanism model in S2 to calculate the mesoscopic scale data items related to the internal grain size of the strip steel, merge the mesoscopic scale data items with the macroscopic scale data items, and form a continuous annealing production process input vector containing macro-mesoscopic data items; S6: Use the maximum and minimum values of each data item in the sample set D2 obtained in S2 to normalize the values of the corresponding data items in the process input vector; S7: Based on the normalized process input vector, the strip steel product hardness prediction model selected in S4 is used to predict the strip steel product hardness; S8: After the continuous annealing production of the strip steel product is completed, the strip steel head is intercepted for offline manual testing to obtain the actual hardness value of the strip steel product, and the actual hardness value is matched and synthesized with the continuous annealing process input vector containing macro-mesoscopic data items recorded by the strip steel product in S5 to obtain a new continuous annealing production sample; the production sample includes p macro-scale data items and an actual hardness value of the strip steel product; S9: Update the scheduling quality to H j The method is as follows: for the scheduled quality strip steel, after the continuous annealing production samples in the latest O time period are collected, the continuous annealing production samples in the latest O time period are used to replace the continuous annealing production samples in the earliest O time period of the scheduled quality strip steel in the enterprise database; S10: For the strip steel products of scheduling quality in Z, set the threshold G to judge each scheduling quality H i , i=1,2,…,Z Whether the number of updated continuous annealing samples is greater than G, if so, then according to the methods S1 to S3, based on the updated continuous annealing production samples of the scheduling quality strip steel, the hardness prediction model of the scheduling quality strip steel product is retrained and updated to ensure that it can track the latest operating conditions of the continuous annealing unit producing the scheduling quality strip steel product; S11: When the continuous annealing production of a new strip steel product begins, S4 to S11 are repeatedly executed to realize the online prediction of the hardness of the existing Z types of continuously annealed strip steel products of different scheduling qualities.
2. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 1, characterized in that: The specific process of S2 includes the following steps: S2.1: For each continuous annealing production sample in the sample set D1, use the corresponding mechanism model to calculate q mesoscopic data items related to the internal grain size of the strip steel, and fuse the q mesoscopic data items with the p macroscopic data items to form a continuous annealing macro-mesoscopic multi-scale fusion data sample set D2 containing p+q macro-mesoscopic information input items and 1 actual strip steel hardness output item; S2.2: From S2.1, a new set D2 containing N continuous annealing production samples is obtained, where each production sample contains p+q macro-mesoscopic information input items and 1 actual strip hardness output item; each input item in D2 is normalized to ensure the dimensional consistency between different input items, and the maximum and minimum values of the p+q input items in the sample set D2 are saved.
3. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 1, characterized in that: The specific process of S3 includes the following steps: S3.1: One Dimensional Convolutional Neural Networks (OD-CNN) is used as the sub-learning machine, and the improved Adaboost_R2 technology with adaptive parameter initialization strategy is used to generate K OD-CNN models: S3.2: Based on the K OD-CNN models established in S3.1, first cluster them using the "k-means" clustering algorithm, and then randomly select OD-CNN models from each cluster for building the ensemble learning machine model; S3.3: For the S sub-learning machine models selected in S3.2, use the sub-learning machine weight optimization method based on differential evolution to optimize the weight v of each sub-learning machine model. i , i = 1, 2, ..., S for optimization. According to the S sub-learning machine models and their respective weights, the S sub-learning machine models are integrated into an integrated learning machine model by the linear weighting method shown in formula (3). The integrated learning machine model is the hardness prediction model of the tempered strip steel product. Its output is the weighted composite value of the strip steel product hardness prediction values output by each sub-learning machine model. Among them, P final (x) is the predicted value of the strip product hardness of the integrated learning machine model for the production sample x, v i and p i (x) are the weight value of the i-th sub-learning machine model and the predicted value of the hardness of the strip steel product output by the i-th sub-learning machine model for the production sample x.
4. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 3, characterized in that: The specific process of step 3 is divided into the following steps: S3.1.1: Initialize the model counter k to 1, and assign the same weight value to all samples in the sample set D2 after S2 data processing, that is, the weight of each sample is S3.1.2: Use resampling methods incorporating sample weights Perform sampling with replacement on the sample set D2 to obtain a new set containing N production samples S3.1.3: Construct the kth OD-CNN model and initialize the parameters of the kth OD-CNN model based on the adaptive parameter initialization strategy. The specific steps are as follows: S3.1.4: For the kth OD-CNN model after parameter initialization in S3.1.3, train it on the sample set obtained by resampling in S3.1.2 to establish an OD-CNN model for hardness prediction of strip steel products; S3.1.5: Update the model counter k=k+1, and adjust the sample weights according to the training errors of each production sample for the sample set D2 processed by S2.2 The specific method is: first increase the weight of samples with larger training errors, reduce the weight of samples with smaller training errors, and then normalize the new sample weights; S3.1.6: Repeat S3.1.2 to S3.1.5 to establish K OD-CNN models for hardness prediction of continuous annealing strip products.
5. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 4, characterized in that: The specific process of step 3.1.3 is divided into the following steps: S3.1.3.1: Determine whether the kth OD-CNN is the first model, that is, whether k is equal to 1. If k is equal to 1, use the random initialization method to initialize the parameters of the kth OD-CNN model and jump to S3.1.
4. If k is not equal to 1, jump to S3.1.3.2; S3.1.3.2: Determine whether k can be tested by the trend frequency T trend +1 is divisible. If it is not divisible, jump to S3.1.3.
3. If it is divisible, use the Mann-Kendall trend test method to analyze the newly generated kT trend T to the k-1th trend Whether the performance of the OD-CNN meets the improvement conditions. If so, jump to S3.1.3.4; otherwise, jump to S3.1.3.5; S3.1.3.3: Initialize the parameters of the k-th OD-CNN model using the optimized parameters of the trained k-1th OD-CNN model, that is, directly use the model parameter values optimized by the k-1th OD-CNN as the initial values of the model parameters corresponding to the k-th OD-CNN; S3.1.3.4: When the newly generated T trend The performance of the kth OD-CNN meets the improvement conditions, and the parameter initialization method in S3.1.3.3 is continued to be used to initialize the parameters of the kth OD-CNN model; S3.1.3.5: When the newly generated T trend If the performance of the kth OD-CNN does not meet the improvement conditions, the crossover mutation operation in the differential evolution algorithm (DE) is used to generate new model parameters to initialize the kth OD-CNN model parameters.
6. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 5, characterized in that: The specific process of step 3.1.3.5 is divided into the following steps: S3.1.3.5.1: Select three different models from the first k-1 trained OD-CNNs, including the model with the best prediction performance, that is, The OD-CNN model with the smallest RMS error calculated above and the two randomly selected models are used to obtain the optimized parameters M of each model. best 、M e and M h ; S3.1.3.5.2: The model parameters M after the three OD-CNNs are optimized based on S3.1.3.5.1 best 、M e and M h , using DE / best / 1 mutation mode and binomial crossover to generate new model parameters S3.1.3.5.3: Use the model parameters generated by S3.1.3.5.2 Initialize the corresponding parameter values of the kth OD-CNN model.
7. The method for online prediction of annealing strip hardness based on multi-scale data fusion according to claim 3, characterized in that: The specific process of step 3.2 is divided into the following steps: S3.2.1: Based on the sample set D2 processed by S2, the K OD-CNN models established in S3.1 are evaluated using the two indicators of accuracy and dispersion, and then a data set D3 containing K samples is constructed, where the kth sample is a two-dimensional vector containing the two attribute values of "accuracy" and "dispersion"; the accuracy formula (1) and dispersion formula (2) of the kth OD-CNN model are calculated as follows: Where N is the total number of samples in D2, K is the total number of OD-CNN models constructed in S3.1, is the hardness prediction value of the strip steel product of the i-th sample in D2 by the k-th OD-CNN model, r i is the true value of the hardness of the strip steel product of the i-th sample in D2, represents the average value of the hardness prediction of the strip steel product on the i-th sample in D2 by K OD-CNN models; S3.2.2: Based on the dataset D3 constructed in S3.2.1, the K OD-CNN models established in S3.1 are clustered using the "k-means" algorithm to obtain the cluster partition C = {C1, C2, ..., C l }, l is the number of clusters obtained after clustering; S3.2.3: Cluster C obtained from S3.2.2 i , i=1,2,…,l, when C i When the number of OD-CNN models in is greater than m, m models are randomly selected as sub-learning machines of the integrated model; i When the number of OD-CNN models is less than or equal to m, C i All OD-CNN models in the ensemble are used as sub-learning machines of the ensemble model.
8. The method for online prediction of hardness of annealed strip steel based on multi-scale data fusion according to claim 5, characterized in that: The specific steps of determining whether the performance meets the improvement conditions are as follows: S3.1.3.2.1: For the newly generated kT trend T to the k-1th trend OD-CNN models, respectively, the sample sets obtained in S3.1.2 Calculate the root mean square error of each OD-CNN model and construct a trend A vector V of elements, each element in V is the root mean square error value of the corresponding OD-CNN model; S3.1.3.2.2: Based on the vector V constructed in S3.1.3.2.1, calculate the test statistic Z zk ; S3.1.3.2.3: Set the confidence level α and obtain the threshold Z by querying the standard normal distribution table 1-α / 2 , when |Z zk |>Z 1-α / 2 And Z zk >0, the newly generated kT trend T to the k-1th trend Whether the performance of an OD-CNN meets the improvement conditions, otherwise the performance does not meet the improvement conditions.
