A method for predicting final rolling temperature of hot rolling based on classification decision tree
Through the classification decision tree-based method, the final rolling temperature prediction model of hot continuous rolling is optimized, combined with the BP neural network, the problem of insufficient prediction accuracy of traditional models is solved, and higher precision temperature prediction is achieved, and product quality control is improved.
Patent Information
- Application Number
- CN202210266243.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-17
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2042-03-17
AI Technical Summary
The traditional hot continuous rolling final rolling temperature prediction model cannot achieve high-precision prediction because it ignores the coupling, nonlinearity, time-varying characteristics of the final rolling temperature, and lacks self-learning function, so it cannot cope with the real-time changing rolling process, resulting in Inadequate prediction accuracy affects product quality.
The model is optimized by a method based on a classification decision tree, and the data set is divided and trained by building a decision tree, combined with the BP neural network model, more accurate nonlinear fit and local optimal problem solving are achieved.
The prediction accuracy of the final rolling temperature is significantly improved. The model's prediction error hits 83.5% within ±10℃ and hits 95.3% within ±15℃, greatly improving the accuracy and reliability of temperature control.
Smart Images
Figure CN115138693B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a hot rolling finishing temperature prediction method based on a classification decision tree, and belongs to the technical field of steel rolling control methods. Background Art
[0002] In the actual production of hot rolling, temperature control is a very important part of hot rolling control. Temperature is interspersed throughout the entire strip production process. The dimensional accuracy and quality indicators of hot-rolled strip are inseparable from the temperature control accuracy. In production, it is necessary to adjust the rolling strategy and process parameters in real time according to the temperature. Therefore, it is necessary to develop an advanced temperature control system to ensure that the strip temperature is within the set range in each production process. The final rolling temperature refers to the temperature of the hot-rolled strip when it leaves the finishing mill during the rolling process. The final rolling temperature directly affects the setting of rolling process factors such as load distribution and rolling force setting. If the final rolling temperature is too low or too high, the product quality will fluctuate. Therefore, the prediction accuracy of the final rolling temperature in the hot rolling production process is particularly important.
[0003] In the past, the final rolling temperature control system was mainly based on the mathematical model established by the traditional rolling theory for prediction. It is necessary to conduct in-depth exploration of the rolling process to establish a more accurate prediction model. This mathematical model can make a preliminary prediction of the final rolling temperature of hot rolling and has been used in rolling production for a long time. However, the actual rolling process is very complicated, and the traditional rolling theory cannot avoid a large number of assumptions and approximate processing. Therefore, it is difficult to further improve the calculation accuracy of the final rolling temperature, and the prediction accuracy can no longer meet the needs of today's products. In addition, the traditional mathematical model ignores the coupling, nonlinearity, and time-varying characteristics of the final rolling temperature in rolling, and has no self-learning function, and cannot cope with the real-time changing rolling process. Therefore, it is difficult for the traditional model to achieve the standard of modern hot rolling temperature control in actual production, which brings huge losses to actual production. Therefore, we need to find a practical and effective method to improve the prediction accuracy of the final rolling temperature in the rolling process. Summary of the invention
[0004] The purpose of the present invention is to provide a hot rolling finishing temperature prediction method based on a classification decision tree, which adopts a decision tree classification algorithm to optimize the model, deeply mines the data set, divides it according to the characteristics, and subdivides the constructed model. According to the different classifications of the data set, independent training is performed and finally the prediction results are uniformly output, which further solves the problem of falling into local optimality in neural network prediction, greatly improves the prediction accuracy of the model, and effectively solves the above-mentioned problems existing in the background technology.
[0005] The technical solution of the present invention is: a method for predicting the final rolling temperature of hot rolling based on a classification decision tree, comprising the following steps:
[0006] Step 1: Send PDI data to finishing process automation;
[0007] Step 2: Trigger the finishing rolling model calculation according to the PDI data, and the rolled piece starts the rolling process according to the finishing rolling procedure;
[0008] Step 3: Normalize the historical production data so that data standards of different magnitudes are mapped between 0 and 1;
[0009] Step 4: Build a classification decision tree
[0010] Step 4.1: First, the optimal features and optimal split points are selected from the data set according to the above method as the root node of the tree, and then the data is divided according to the data split point to split the data set into two subsets;
[0011] Step 4.2: If all the data in the classified subset belong to one category, add a leaf label to the subset. If the data in the subset can still be divided, treat the subset as a new root node and re-find the optimal features and optimal partitioning nodes to classify the data set;
[0012] Step 4.3: Repeat the steps in 4.1 and 4.2 to continuously divide the sub-datasets until all sub-datasets become leaf nodes. At this time, the classification is stopped and the decision tree is built.
[0013] Step 5: Decision tree termination condition
[0014] When the decision tree partitions the data set, the decision tree stops recursively:
[0015] When the decision tree partitioning reaches a certain node set where the elements have no features, the decision tree stops recursively;
[0016] When the number of samples of the current decision tree node is less than the set value, the decision tree stops recursing;
[0017] When the number of data in the sub-dataset is less than the set value when the decision tree is divided, the sub-dataset returns to the previous node and the decision tree stops recursively;
[0018] Calculate the mean square error of each sample. When the mean square error is less than the set value, the decision tree stops recursively.
[0019] Step 6: Decision Tree Pruning
[0020] Post-pruning is adopted, that is, all nodes are first split normally and then pruned. The pruning loss function is used to describe the pruning loss. T represents the node and t represents the subtree. The pruning loss expression is:
[0021] C α (t)=C(t)+α|t| (1)
[0022] In the formula, C(t) is the prediction error, α is the regularization coefficient, and |t| is the number of leaves;
[0023] If pruning operation is performed on node T, its loss function is:
[0024] C α (T) = C(T) + α (2)
[0025] Dividing the node will reduce the prediction error and increase the fitting degree. Therefore, C(t) < C(T). When the regularization coefficient is very small and approaches 0, C α (t) < C α (T), that is, the loss function without pruning is less than the loss function with pruning. At this time, the model tends to retain the leaves; when the regularization coefficient is very large, C α (t) > C α (T), and the loss function with pruning is less than the loss function without pruning. At this time, the model tends to prune;
[0026] Step 7: Cross-validation
[0027] Use the cross-validation method to prune the generated decision tree, find the optimal regularization coefficient α, perform good pruning, and retain a good fitting accuracy while enabling the model to have a certain generalization ability.
[0028] In the said step 1, the PDI data includes the intermediate billet size data of the rolled piece, the finished product size data, the steel grade name, and the chemical composition.
[0029] In the said step 7, the specific method is to solve the regularization coefficient α by combining formula (1) and formula (2):
[0030]
[0031] Calculate the regularization coefficients of all nodes using formula (3), and arrange the regularization coefficients in ascending order. Start verifying from the largest α, and finally find the optimal regularization coefficient value to complete the pruning of the decision tree.
[0032] The beneficial effects of the present invention are: using the decision tree classification algorithm to optimize the model, deeply mining the data set, dividing according to features, subdividing the constructed model, independently training according to different classifications of the data set and finally uniformly outputting the prediction results, further solving the problem of falling into local optimum in neural network prediction, and greatly improving the prediction accuracy of the model. Description of the Drawings
[0033] Figure 1 It is a schematic diagram of the classification decision tree model of the present invention;
[0034] Figure 2 This is a comparison chart of temperature predictions from traditional models;
[0035] Figure 3 It is a prediction comparison diagram of the classification decision tree model of the present invention. DETAILED DESCRIPTION
[0036] In order to make the purpose, technical solutions and advantages of the invention implementation cases clearer, the technical solutions in the implementation cases of the present invention will be clearly and completely described below in conjunction with the drawings in the implementation cases. Obviously, the implementation cases described are only a small part of the implementation cases of the present invention, rather than all the implementation cases. Based on the implementation cases in the present invention, all other implementation cases obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0037] A method for predicting the final rolling temperature of hot rolling based on a classification decision tree comprises the following steps:
[0038] Step 1: Send PDI data to finishing process automation;
[0039] Step 2: Trigger the finishing rolling model calculation according to the PDI data, and the rolled piece starts the rolling process according to the finishing rolling procedure;
[0040] Step 3: Normalize the historical production data. Since the properties and types of the sorted rolling data are different, the numerical differences of different data are large. Therefore, it is necessary to normalize the input data again so that the data standards of different magnitudes are mapped to between 0 and 1.
[0041] Step 4: Build a classification decision tree
[0042] Step 4.1: First, the optimal features and optimal split points are selected from the data set according to the above method as the root node of the tree, and then the data is divided according to the data split point to split the data set into two subsets;
[0043] Step 4.2: If all the data in the classified subset belong to one category, add a leaf label to the subset. If the data in the subset can still be divided, treat the subset as a new root node and re-find the optimal features and optimal partitioning nodes to classify the data set;
[0044] Step 4.3: Repeat the steps in 4.1 and 4.2 to continuously divide the sub-datasets until all sub-datasets become leaf nodes. At this time, the classification is stopped and the decision tree is built.
[0045] Step 5: Decision tree termination condition
[0046] When the decision tree partitions the data set, the decision tree stops recursing:
[0047] When there are no features for the elements in a certain node set during the decision tree partitioning, the decision tree stops recursing;
[0048] When the number of samples in the current decision tree node is less than the set value, the decision tree stops recursing;
[0049] When the number of data in the sub - dataset is less than the set value during the decision tree partitioning, the sub - dataset returns to the previous node and the decision tree stops recursing;
[0050] Calculate the mean squared error of each sample. When the mean squared error is less than the set value, the decision tree stops recursing;
[0051] Step 6: Decision tree pruning
[0052] Pre - pruning evaluates the generalization ability of the model before and after partitioning the current node. By comparing the generalization ability of the model before and after partitioning, it decides whether to split the current node. However, for some nodes, although the current split may lead to poor generalization ability temporarily, the generalization ability will improve in subsequent partitions. Pre - pruning directly preventing the current node from splitting easily leads to underfitting of the model.
[0053] The present invention adopts post - pruning, that is, first let all nodes split normally and then perform pruning operations on the nodes. Use the pruning loss function to describe the pruning loss. Use T to represent the node and t to represent the subtree. Its pruning loss expression is:
[0054] C α (t)=C(t)+α|t| (1)
[0055] In the formula, C(t) is the prediction error, α is the regularization coefficient, and |t| is the number of leaves;
[0056] If pruning operation is performed on node T, its loss function is:
[0057] C α (T)=C(T)+α (2)
[0058] Partitioning the node will reduce the prediction error and increase the fitting degree. Therefore, C(t)<C(T). When the regularization coefficient is very small and approaches 0, C α (t)<C α (T), that is, the loss function without pruning is less than the loss function with pruning. At this time, the model tends to retain the leaves; when the regularization coefficient is very large, C α (t)>C α (T), the loss function with pruning is less than the loss function without pruning. At this time, the model tends to prune;
[0059] Step 7: Cross - validation
[0060] The generated decision tree is pruned using cross-validation to find the best regularization coefficient α and perform good pruning to retain good fitting accuracy while giving the model a certain generalization ability.
[0061] In step 1, the PDI data includes the intermediate billet size data, the finished product size data, the steel grade name and the chemical composition.
[0062] In step 7, the specific method is to solve the regularization coefficient α by combining equations (1) and (2):
[0063]
[0064] Use formula (3) to calculate the regularization coefficients of all nodes, and arrange the regularization coefficients in order of size. Start with the largest α for verification, and finally find the optimal regularization coefficient value to complete the pruning of the decision tree.
[0065] In practical applications,
[0066] 1. Model structure
[0067] Import the decision tree classification module into the neural network model. The optimized model first classifies the data set into multiple data subsets, and then randomly extracts 80% of the data in the data subset as the training set, and the remaining data as the test set. Train each data subset separately, iterate the data in the training set to adjust the weights and bias of the BP neural network model to achieve the purpose of model training, and finally use the data in the test set to test the model and output the predicted value. The optimized model structure is as follows: Figure 1 shown.
[0068] By using decision trees to classify data, the BP neural network model is split into multiple sub-models. The parameters in the sub-models are iteratively updated separately without affecting each other. This allows for a more accurate nonlinear fitting of the final rolling temperature of the hot rolling process. Therefore, the decision tree algorithm effectively improves the prediction accuracy of the BP neural network model. At the same time, due to the use of decision tree classification to conduct a deeper mining of data, the BP neural network optimization interval is reduced, reducing the possibility of the BP neural network falling into a local optimum rather than a global optimum. The optimized model effectively improves the situation of low prediction accuracy of special data.
[0069] 2 Model effect evaluation
[0070] Model optimization is mainly to solve the problem that the predicted values of a few data deviate greatly from the true values. The data features are classified according to the constructed decision tree. After classification, the data impurity of the sub-dataset can be reduced, avoiding the problem of low prediction accuracy of special data in the prediction process.
[0071] Draw a fitting graph of the optimized data and compare it with the data before optimization, such as Figure 2 and Figure 3 It can be found that the points drawn in the optimized model fitting diagram are closer to the 45° standard straight line, and the points outside the red line of the error ±20°C boundary are significantly reduced. Among the extracted prediction data, only one set of data has a prediction error slightly exceeding 20°C. This shows that using decision trees for classification before prediction can effectively improve the reliability of the model.
[0072] Table 1 below is a comparison of the various indicators of the prediction data before and after optimization. After optimization, the model predicted that the relative error within ±5°C was 53.6%, the relative error within ±10°C was 83.5%, the relative error within ±15°C was 95.3%, and the relative error within ±20°C was 99.9%. All indicators predicted by the model have been improved, indicating that the decision tree algorithm optimization has a certain effect.
[0073] After optimization, the prediction error of the final rolling temperature prediction model is within 10℃, with a hit rate of 83.5% and within 15℃, with a hit rate of 95.3%.
[0074] Table 1
[0075]
[0076]
Claims
1. A hot rolling final rolling temperature prediction method based on classification decision tree, characterized in that The following steps are involved: Step 1: Send PDI data to finishing process automation; Step 2: Trigger the finishing rolling model calculation according to the PDI data, and the rolled piece starts the rolling process according to the finishing rolling procedure; Step 3: Normalize the historical production data so that data standards of different magnitudes are mapped between 0 and 1; Step 4: Build a classification decision tree Step 4.1: First, the optimal features and optimal split points are shaped from the data set according to steps 1-3 as the root node of the tree, and then the data is divided according to the data split point to split the data set into two subsets; Step 4.2: If all the data in the classified subset belong to one category, add a leaf label to the subset. If the data in the subset can still be divided, treat the subset as a new root node and re-find the optimal features and optimal partitioning nodes to classify the data set; Step 4.3: Repeat the steps in 4.1 and 4.2 to continuously divide the sub-datasets until all sub-datasets become leaf nodes. At this time, the classification is stopped and the decision tree is built. Step 5: Decision tree termination condition When the decision tree partitions the data set, the decision tree stops recursing: When the decision tree partitioning reaches a certain node set where the elements have no features, the decision tree stops recursively; When the number of samples of the current decision tree node is less than the set value, the decision tree stops recursing; When the number of data in the sub-dataset is less than the set value when the decision tree is divided, the sub-dataset returns to the previous node and the decision tree stops recursively; Calculate the mean square error of each sample. When the mean square error is less than the set value, the decision tree stops recursively. Step 6: Decision Tree Pruning Post-pruning is adopted, that is, all nodes are first split normally and then pruned. The pruning loss function is used to describe the pruning loss. T represents the node and t represents the subtree. The pruning loss expression is: C α (t)=C(t)+α|t| (1) Where C(t) is the prediction error, α is the regularization coefficient, and |t| is the number of leaves; If node T is pruned, its loss function is: C α (T)=C(T)+α (2) Dividing the nodes will reduce the prediction error and increase the fit, so C(t)<C(T). When the regularization coefficient is very small and close to 0, C α (t)<C α (T), that is, the loss function of not pruning is less than the loss function of pruning, and the model tends to retain leaves; When the regularization coefficient is large, C α (t)>C α (T), the loss function of pruning is less than the loss function of not pruning, and the model tends to prune; Step 7: Cross Validation The generated decision tree is pruned using cross-validation to find the best regularization coefficient α and perform good pruning to retain good fitting accuracy while giving the model a certain generalization ability.
2. The method for predicting the final rolling temperature of hot rolling based on a classification decision tree according to claim 1, characterized in that: In step 1, the PDI data includes the intermediate billet size data, the finished product size data, the steel grade name and the chemical composition.
3. The method for predicting the final rolling temperature of hot rolling based on a classification decision tree according to claim 1, characterized in that: In step 7, the specific method is to solve the regularization coefficient α by combining equations (1) and (2): Use formula (3) to calculate the regularization coefficients of all nodes, arrange the regularization coefficients according to size, start verification from the largest α, and finally find the optimal regularization coefficient value to complete the pruning of the decision tree.