Search
Fig.
1.
Temperature and thickness variation curve of strip in hot-rolling production
| Cite this article as: |
Feifei Li, Anrui He, Yong Song, Zheng Wang, Xiaoqing Xu, Shiwei Zhang, Yi Qiang, and Chao Liu, Deep learning for predictive mechanical properties of hot-rolled strip in complex manufacturing systems, Int. J. Miner. Metall. Mater., 30(2023), No. 6, pp.1093-1103. https://dx.doi.org/10.1007/s12613-022-2536-y
|
Steel production is the basic industry to promote national economic development. Driven by the concept of intelligent manufacturing, the key technologies of steel production efficiency, process optimization, and quality control have attracted wide attention in the industry [1]. Hot rolled strip as one of the important products in iron and steel enterprise is widely used in shipbuilding, automobile manufacturing, construction, household appliance manufacturing, and other industries [2–3]. Its production and processing process involves many process links such as heating, rolling, cooling, and heat treatment. In this process of industrial manufacturing, materials will undergo continuous and complex chemical and physical transformations [4], and there are coupling effects between different chemical components, between chemical components and processes, and between different processes. So it is difficult to establish an accurate mathematical model to express the relationship between chemical composition, production process parameters, and final mechanical properties. Furthermore, the accurate prediction of mechanical properties is the core problem of hot-rolled strip product development and quality control [5].
In order to improve the prediction accuracy of mechanical properties of hot-rolled strip products, many researchers introduced machine learning algorithms to construct the prediction model of mechanical properties, such as gradient boosting decision tree (GBDT) [6], random forest (RF) [7], support vector machine (SVM) [8], and logistic regression (LR) [9], and they have achieved good prediction results [10–11]. However, in traditional mechanical property prediction modeling based on machine learning, the data features are mainly chemical composition and several typical process parameters. These features selected by experience are difficult to capture the abstract and implicit cross information in the data. Moreover, the complexity of traditional machine learning is low, and it is difficult to accurately express the complex relationship between chemical composition, production process, and final mechanical properties of finished products.
Driven by computer technology, deep learning has become an important field of artificial intelligence research, and it is also one of the effective means to solve the above problems. Deep learning is a feature learning algorithm with a multi-layer structure, which has higher complexity than traditional machine learning algorithms. Due to the deep feature extraction and strong nonlinear fitting ability of deep learning [12–13], its performance in complex tasks such as machine vision, natural language processing, and speech recognition has far exceeded that of traditional machine learning algorithms [14–17]. The multi-layer structure of deep learning can extract advanced abstract features from the original data layer by layer, which is consistent with the organizational evolution process of multiple process links involved in the prediction of strip mechanical properties and the progressive relationship between chemical composition, process, organization, and performance. In this direction, researchers introduced deep learning methods such as multilayer back propagation neural network, convolution neural network, deep feedforward neural network, and long-term and short-term memory neural network to establish the relationship model between chemical composition, process, and mechanical properties [18–22]. Among them, Li et al. [18] based on the collected more than 50000 hot-rolling data including 20 influencing factors, converted the one-dimensional numerical data of strip steel into two-dimensional data to express the complex interaction between influencing factors, and then used the deep convolution network to establish the tensile strength prediction model, which achieved good results. Song et al. [22] combined with the actual process flow of compact strip production (CSP) hot continuous rolling, designed multi-layer stacked self-encoder networks representing different processes respectively for feature extraction of production process data, and used SVM to predict the mechanical properties, which has good comprehensive performance. Therefore, combined with the inherent advantages of deep learning and the characteristics of the strip production process, deep learning can be better applied to the prediction of mechanical properties and improve the overall performance of the model. However, the above deep learning modeling scheme still has some shortcomings. (1) Incomplete data information. The selection of model input features depends on human experience and knowledge. The limited discrete process parameters selected subjectively are difficult to accurately reflect the actual processing process of strip, and the selected discrete process parameters also lose the characteristic information of time sequence. (2) Large demand for data. These deep learning schemes need to rely on a large amount of label data, while the label data of mechanical properties is difficult to obtain and costly, and many production sites do not have the data conditions required for deep neural network modeling. (3) Complex parameter adjustment. The parameter adjustment of these deep learning models is very complex and requires a lot of time and energy.
With the development of artificial intelligence technology, Professor Zhou’s team proposed a deep learning algorithm [23–24] with multi-layer forest structure in 2017. The algorithm draws lessons from the deep neural network structure and adopts the multi-grained scanning structure and cascade forests structure for data feature extraction and fitting respectively. Multi-grained cascade forest (gcForest) algorithm requires few parameters, is easy to train and is not sensitive to the setting of super parameters, and the same super parameter setting can get good results on different data sets. GcForest is composed of multiple groups of random forests, which inherits the advantages of random forests, and it can process high-dimensional data and achieve good performance on small data sets. Professor Zhou’s team conducted experiments in image classification, face recognition of only 400 pictures, music classification, and other tasks, and the results show that gcForest has a very good performance [23]. For this reason, Ding et al. [25] carried out bearing fault diagnosis based on gcForest, which verified that the algorithm has less training time on the premise of ensuring accuracy and avoids the complex parameter adjustment process. The author also carried out the model performance comparison and stability analysis of the algorithm under complex working conditions, and verified the influence of the diagnosis accuracy, generalization ability, and parameter setting of the model on the stability of the model. Yao and Chen [26] use the gcForest algorithm to solve the problem of power equipment image recognition under small samples. They compared K-nearest neighbor (KNN), SVM, and convolutional neural network (CNN), and the results show that the classification accuracy of gcForest is significantly higher than the other three algorithms. These results show that gcForest has great advantages in model tuning, model stability, prediction accuracy, and training data volume. Based on the gcForest framework, according to the complex process flow of hot-rolled strip production and the abnormal sensitivity of process path and parameters to product quality in hot-rolled strip production, this paper proposes a three-dimensional continuous time series process data sampling method based on time-temperature-deformation, and uses the multi-grained scanning structure of the gcForest model to extract the characteristics of the hot-rolled strip production process. So as to better reflect the influence of process parameters such as temperature change, deformation degree, deformation temperature, and deformation rate on the law of microstructure evolution in different stages of the whole processing process. Compared with the existing deep learning model, this method considers the collaborative change process of the strip in the three dimensions of time, thickness, and temperature at the same time, and makes full use of gcForest to carry out continuous multi-dimensional feature learning along the production sequence process, rather than learning a kind of state information described by a limited number of isolated process parameters. Moreover, the gcForest model is easy to tune and can perform well on small data sets. Using the relevant data collected from a steel plant, experiments were carried out on three steel grades to verify the effectiveness of gcForest in actual production.
As shown in Fig. 1, the upper part is the schematic diagram of a hot-rolling production line. In the hot continuous rolling production, the slab needs to go through heating furnace, roughing mill, finishing mill, cooling, coiling, and other links to finally obtain the finished strip steel. The microstructure evolution of steel includes three key processes in the time dimension. First, the growth process of steel micro grains in the heating furnace [27–28], then the refinement of grains [29–30] and the strain-induced precipitation of second phase particles [31–32] in the rolling process, and finally the phase transformation of steel microstructure caused by the cooling process, resulting in ferrite, pearlite, and even bainite, martensite, and other structures [33–34]. The composition, proportion, and distribution of these microstructures directly determine the mechanical properties of steel products [35]. Therefore, in this paper, the model input data is designed as three groups of sequence data, including time and steel thickness and temperature at the corresponding time, which reflects the historical information of steel thickness and temperature changes in the time dimension. At the same time, the local historical information obtained by multi-grained scanning is fused with the basic information of the slab (chemical composition and typical process parameters), making the input of the next link have both local and global features. This dynamic sampling method along the process changes the traditional data input form based on isolated and discrete process parameters. Compared with the sample construction method using spatial dimension in CNN modeling in previous literature [18], the model construction scheme based on time dimension corresponds to the continuity of steel processing process and internal organization evolution, which not only increases the time sequence information related to the process, but also realizes the continuity characterization of adjacent processes. So as to reflect the microstructure evolution law of steel in the whole processing process.
The input sequence data takes the slab entering the heating furnace as the starting point and coiling as the end point. The lower half of Fig. 1 corresponds to the curve of the thickness (THK), surface temperature (Surface), core temperature (Core), and average temperature (Average) of the strip steel production process under each production link with time.
In addition, the output data are yield strength (YS), tensile strength (TS), and elongation (EL), respectively. For each sample in the dataset, the input data is composed of time–temperature–thickness sequence data and chemical composition, and the output data is composed of YS, TS, and EL. These output data needs to be locally sampled for steel products, and then obtained by destructive tests. The above time–temperature–thickness sequence data selects the process flow of steel sampling location.
The gcForest is also known as the deep forest [24], which includes two parts: multi-grained scanning structure and cascade forest structure. Multi-grained scanning structure can be understood as the feature extraction of the original data, which is equivalent to the convolution layer, while the cascade forest structure adopts a hierarchical structure similar to deep neural network, and the output results of the previous layer will be used as the input of the next layer of forest until the last layer.
In the task of sequence data modeling, the multi-grained scanning structure uses multiple windows with different lengths for sliding sampling, and multiple groups of interrelated and different subsequences can be obtained. Fig. 2 is a schematic diagram of the principle of feature conversion of sequence data with length d after multi-grained scanning.
In this part, first, the sequence data with length of d is sampled by sliding window. After sampling, P groups of sub sequences with equal length can be obtained, in which P can be calculated by the following formula:
|
P=(d−k)/s+1 |
(1) |
where k represents the length of subsequence data, and s represents the moving step length of the sliding window.
After that, P groups of sequence data obtained by sliding sampling will be input into random forest A and completely random forest B models for prediction, and P groups of prediction results will be obtained by A and B forest models respectively. The prediction results are spliced in order to obtain a set of row vectors with a length of 2P, which is the output of a multi-grained scanning structure.
The output data of the multi-grained scanning structure will be input into the cascade forest structure as the feature. Fig. 3 is the schematic diagram of the cascade forest structure. The input data of the first layer directly comes from the multi-grained scanning structure, while the input of other layers consists of the output of the previous layer and the output of the multi-grained scanning structure. In this structure, the output of each forest in each layer is consistent with the output of each forest in the multi-grained scanning structure, which is the target label. For the regression task, the output results of each forest in the last layer are calculated by the integration strategy to obtain the final prediction results, such as average value, median value, maximum value.
The production process of hot continuous rolling involves many technological links. The change of time, temperature, and thickness in the production process of hot-rolled strip represents the evolution of steel micro-organization. Therefore, it is the key factor affecting the mechanical properties of finished products. In hot-rolling production, due to different process control requirements of different strips, there will be differences in the number of data points in heating and rough rolling, which leads to inconsistent data dimensions in the production process of each coil of strip. Therefore, the inconsistency of data characteristics is also one of the characteristics of hot-rolling production data.
In the gcForest framework, the number of output features of the multi-grained scanning structure is 2P. According to Eq. (1), when the input data length d changes, the number of output results 2P will also change, resulting in the change of the length of the output feature vector of the multi-grained scanning structure, which will cause the problem that the subsequent cascade forest structure cannot be adapted, because the input of the cascade forest structure is similar to the full connection layer in the neural network. It requires that the length of the eigenvector of the input data is fixed. Therefore, this paper improves the multi-grained scanning structure in gcForest and adds a sub sampling link of variable window to ensure the consistency of the number of output features.
Fig. 4 is a schematic diagram of the principle of the sub sampling link of the variable window, which will be inserted into the multi-grained scanning structure. In the sub sampling link, the output results of A and B forests are spliced respectively to obtain two groups of feature sequences with length P. Then, sub sampling is carried out for each feature sequence. The sub sampling windows can be X1, X10, X20, etc., or different window combinations. Where X1 indicates that the feature sequence with the original length of P is sub sampled once, that is, the average value of the features is calculated respectively to obtain one feature; X20 means that the feature sequence of the original length P is sub sampled 20 times. The window size of each sub sampling is the same, and the step length is the same. For the results of each sub sampling, the average value is calculated respectively, and a total of 20 features are obtained. Assuming that the goal of sub sampling is to obtain N groups of features, under the original feature sequence with length P, the following equation needs to be satisfied:
|
n=P+m(N−1)N |
(2) |
where n represents the length of each sub sampling window, and n is a positive integer; m represents the overlapping length of two adjacent sub sampling windows, m is a non-negative integer, and m is required to be as small as possible.
The sub sampling link is similar to the pooling process in convolutional neural network. It aggregates and counts the data in different windows to extract different kinds of information from the data. When there are fewer sub sampling windows (such as X1), it tends to extract global information, while when there are more sub sampling windows (such as X50), it tends to extract local information. In addition, the average method and maximum method can be used in aggregation statistics. After the sub sampling of the variable window, the original features with different lengths will be processed into new feature sequences with the same number of features. For example, after the sub sampling schemes of X1, X5, X10, X20, and X50 are used to process the original feature sequences with length P, the new feature sequences with length 86 can be obtained. Therefore, even if the feature lengths of the original data are inconsistent, the new feature sequences with the same length will be output. This ensures the consistency of the length of the output feature sequence of the scanning structure.
This paper proposes a gcForest model that can learn from complex continuous production process data with a small number of sample. The gcForest framework in this paper includes two parts: multi-grained scanning structure and cascade forest structure. YS, TS, and EL are the target outputs.
The production of hot-rolled strip is composed of a series of orderly process, and the whole process can be described by the change of temperature and thickness of the rolled piece in the time dimension. Multi-grained scanning is used to extract features from the original data. Since the microstructure evolution of steel is closely related to its own chemical composition, the attribute information of the rolled piece itself is added to the multi-grained scanning process in this paper. In addition, the microstructure evolution of steel changes synchronously with the production process, so this paper extracts the original information in the sequence data along the production process path, which makes the proposed model no longer be limited to specific production processes and can be adapted to various production data sets with different process characteristics.
Fig. 5 is a designed multi-grained scanning structure, which includes a random forest and a completely random forest. The first is the input data, which contains three groups of sequence data and a group of raw material information data. The sequence data are time, temperature, and thickness respectively. For the same coil of strip steel, the length d of the three groups of information is the same. In the first step, the window with the length of k is used to scan the three information of a single coil of strip steel at the same time, and the window slides in steps of s to obtain P groups subsequences containing time, temperature, and thickness information. Each subsequence is combined with the raw material information with the length of b to obtain P groups subsequence data with global and local characteristics at the same time. Therefore, each group of characteristic data contains 3k+b characteristic values, and the calculation of P is shown in Eq. (1).
Then, the 3k+b characteristic data corresponding to each subsequence is used as the input of A forest and B forest respectively, and the output of YS, TS, and EL. The output of each forest contains P groups of results, in which A and B forests adopt random forest and completely random forest respectively. For the output results of the random forest model, the sub sampling method of the variable window is used for processing, and the data in the window is aggregated and counted by the average value. Finally, the output characteristic sequence with a certain dimension can be obtained.
The specific configuration of gcForest is as follows: In order to ensure the extraction ability of local features and the overall operation efficiency of the model, one scanning granularity is adopted, and the scanning granularity k and the scanning step are set to 10 and 1, respectively.
The production of hot-rolled strip involves multiple complex process links, and there are complex nonlinear relationships between the chemical components and the process parameters. In order to accurately fit the correlation between the input characteristics and the target label, a cascade forest structure is designed as shown in Fig. 6. This structure includes two levels of forests, each consisting of two random forests and two completely random forests. The mechanical properties of steel include YS, TS, and EL. There is usually a certain coupling between these three indexes. In order to enable the training process of gcForest to couple the information between these indicators, the outputs of forests in the cascade forest structure except the end layer are set as 3 outputs, which are YS, TS, and EL. The output of the end layer of the cascade forest structure is a single mechanical property.
The input of this structure is the characteristic data output from the multi-grained scanning structure, which is recorded as m. First, it enters the first level forest structure, including two random forests and two completely random forests. The outputs are YS, TS, and EL, with a total of 12 outputs. Then, these 12 outputs are spliced with the original input data to obtain a new characteristic data with the length m+12, which is used as the input of the second level forest. The output of the last stage forest is a single mechanical performance label such as YS, TS, or EL.
GcForest is a depth model based on decision tree. Because the decision tree is a non-differentiable module, the training process of the gcForest model does not involve gradient adjustment, and its training and prediction process will also be different from neural network algorithm. In gcForest, the random forest model at each level needs to be trained separately according to the order of model structure.
In order to fully verify the performance of gcForest, in this paper, the data of three kinds of steel grades in a steel plant was selected for the experiment, namely Q235B, 510L, and 610L. Among them, the Q235B dataset has 952 samples, the 510L dataset has 749 samples, and the 610L dataset has 557 samples. Each sample contains the time–temperature–thickness change information of the production process and the chemical composition (C, Si, Mn, P, S, Ti, Nb, V, N, Ni, Cu) of the material itself. The data labels of each roll of strip steel are YS, TS, and EL.
In order to more accurately reflect the application effect of the model, this paper adopts the way of rolling training to simulate the real production rhythm as much as possible. Fig. 7 is a schematic diagram of the data set division during rolling training. Each box represents the data for one day. Green represents the test set selected for each prediction, blue represents the training set selected for each training, and gray represents unused data. In the simulation experiment, according to the steel coil sampling sequence, select the data of one day as the test set and the earlier data as the training set. After completing the prediction of the first day, select the data of the second day as the test set, add the data of the first day to the training set, and repeat the above process until the prediction of all M-day test samples is completed.
In order to reflect the performance of the model as accurately as possible, in this paper, for the three datasets of Q235B, 510L, and 610L steels, 30% of the samples were used to evaluate the model performance, and 70% of the samples were used for model training and verification. In addition, in the training dataset, 20% of the data is selected as the validation dataset to optimize the model parameters. In this process, we use the above rolling training method to retrain the model by selecting the previous day and previous samples as training sets when predicting the production samples for a certain day. This rolling training method enables new sampling data to participate in model training in a timely manner, and each sample to be predicted does not appear in the training set.
In the gcForest model, the maximum depth and maximum number of leaf nodes of trees in each forest in the multi-grained scanning structure and cascade forest structure are related to the fitting ability of the model and are very important to the improvement of the final accuracy. In addition, the sub sampling scheme of variable window in the multi-grained scanning structure is related to the accuracy of feature extraction and the efficiency of the model. Through experimental comparison and analysis, the maximum depth of the tree in the forest is set to 6, the maximum number of leaf nodes and sub sampling scheme are determined by grid search, and other parameters adopt default values.
In gcForest model training, in order to avoid training falling into overfitting or underfitting, 20% validation set is used to evaluate the fitting of the model. When the prediction error of the trained model on the training set and the validation set is consistent, it can be considered that the model has been trained to the best state. Because this paper adopts rolling training, the data set used by each training model will change. Fig. 8 shows the average absolute error (MAE) changes of the gcForest model in the training set (Train), validation set (Validation), and test set (Test) during the prediction process. Fig. 8(a), (b), and (c) denotes the YS, TS, and EL of Q235B, respectively; Fig. 8(d), (e), and (f) represents the YS, TS, and EL of 510L, respectively; Fig. 8(g), (h), and (i) denotes the YS, TS, and EL of 610L, respectively. Each step predicts a day’s test samples, and the previous samples are used as the Train and Validation. The MAE of the test set is calculated by all the predicted samples before the current step. From the perspective of fluctuation, the model under each step has reached a good fitting state.
In the multi-grained scanning structure, the consistency of the number of output features can be guaranteed by agreeing on the sub sampling scheme, and different sub sampling schemes will have a certain impact on the feature extraction ability, and then affect the final prediction accuracy. Table 1 records the optimal accuracy (Accr) of three steel grades obtained through grid search and the corresponding model parameter configuration. For ease of expression, the maximum number of leaf nodes of trees in the forest in the multi-grained scanning structure and cascade forest structure are recorded as α and β respectively, and the sub sampling schemes of the selected five variable windows are represented by codes A (X1), B (X1, X10), C (X1, X10, X20), D (X1, X10, X20, X50), E (X1, X10, X20, X50, X80). In addition, the YS, TS, and EL are respectively ±30 MPa, ±30 MPa, and ±3% absolute error for accuracy statistics.
| Grade | YS | TS | EL | |||||||||||
| Code | α | β | Accr / % | Code | α | β | Accr / % | Code | α | β | Accr / % | |||
| Q235B | C | 4 | 9 | 94.97 | C | 4 | 9 | 96.36 | B | 3 | 9 | 94.97 | ||
| 510L | D | 5 | 9 | 96.22 | D | 5 | 4 | 93.78 | A | 5 | 6 | 94.00 | ||
| 610L | B | 3 | 6 | 91.92 | A | 3 | 6 | 92.57 | A | 3 | 8 | 91.67 | ||
DownLoad:
CSV
From the prediction results in Table 1, gcForest has high accuracy in predicting the mechanical properties of the three kind of steel grades, indicating that gcForest can still play an excellent role in the scene with a small sample size. From the perspective of model parameters, the model parameters corresponding to different steel grades and different mechanical property indexes are different. Therefore, in practical application, the model parameters need to be selected according to the specific data.
In order to further prove the superiority of gcForest, the common artificial neural network (NN), bayesian neural networks (BNN), RF, and complete random forest (CRF) algorithms were selected for comparison. The model parameter configuration of RF and CRF algorithms refer to the research of Wang and An [11]. Because the small amount of data is one of the problems to face in this paper, the number of samples in all the three datasets is within 1000, In order to prevent the neural network from falling into over fitting, after debugging, the models of the three steels are set as follows: (1) For NN, the number of hidden layers is 1, the number of neurons in the hidden layer is set to 6, 7, and 3, respectively, the learning rate is set to 0.0001, the number of epochs is 1000, and the size of the validation dataset is 20%. (2) For BNN, the number of hidden layers is 1, the number of neurons in the hidden layer is set to 6, 7, and 3, respectively, the learning rate is set to 0.008, the number of epochs is 400, and the size of the validation dataset is 20%.
Fig. 9 is the loss curve in training process of NN and BNN.
Table 2 shows the comparison of the prediction accuracy of the 5 algorithms, in which the absolute errors of ±30 MPa, ±30 MPa, and ±3% are used for YS, TS, and EL, respectively. Obviously, RF and CRF are better than NN, which is consistent with the experimental results of Wang et al. [11]. Because gcForest has high-dimensional feature input and deep data mining ability, its comprehensive prediction performance is the best of the 5 algorithms.
| Model | Accr for Q235B / % | Accr for 510L / % | Accr for 610L / % | ||||||||
| YS | TS | EL | YS | TS | EL | YS | TS | EL | |||
| gcForest | 94.97 | 96.36 | 94.97 | 96.22 | 93.78 | 94.00 | 91.92 | 92.57 | 91.67 | ||
| CRF | 93.22 | 93.92 | 91.82 | 95.78 | 92.89 | 93.11 | 89.10 | 89.75 | 91.67 | ||
| RF | 86.22 | 89.72 | 92.17 | 94.89 | 91.11 | 91.33 | 90.23 | 88.05 | 91.10 | ||
| NN | 87.62 | 87.62 | 84.83 | 92.22 | 73.33 | 73.11 | 83.45 | 73.36 | 77.54 | ||
| BNN | 86.36 | 81.82 | 79.02 | 77.78 | 81.78 | 62.67 | 67.66 | 71.26 | 56.29 | ||
DownLoad:
CSV
In addition, this paper also compares the MAE and the standard deviation (S) of the error distribution of the prediction results of the 5 algorithms. Fig. 10(a), (b), and (c) denotes the YS, TS, and EL of Q235B, respectively; Fig. 10(d), (e), and (f) represents the YS, TS, and EL of 510L, respectively; Fig. 10(g), (h), and (i) denotes the YS, TS, and EL of 610L, respectively. gcForest has the best comprehensive performance, followed by CRF and RF, and NN and BNN has the worst performance, which is mainly due to the small amount of data used for modeling. Especially in the data sets of 510L and 610L steel grades, the indexes of the NN model and BNN model are much worse than those of the other three models. Therefore, in this small sample data set, it is very difficult to construct a deep neural network model with neurons as the basic unit, and the final accuracy is also difficult to be satisfactory.
Although the total number of samples of 610L steel is only 557, the gcForest model can still achieve better prediction effect than the other four models, and the prediction accuracy of the three mechanical property indexes has reached more than 90%, which fully shows the applicability of gcForest in small sample data sets. Moreover, because gcForest adopts the model construction scheme based on time dimension data, corresponding to the continuity of steel processing history and internal organization evolution, it not only increases the time sequence information related to the process, but also realizes the continuous characterization of adjacent processes. Compared with the traditional modeling method using finite isolated process parameters, the prediction accuracy of the gcForest model is significantly improved, Although the total number of features input by gcForest is close to one thousand, it does not affect the training of gcForest. Although only a few hundred training samples, it can still get very excellent comprehensive performance, which is sufficient to prove that gcForest has strong competitiveness in modeling tasks of small sample datasets.
In this study, the idea of extracting features from the data of the strip production process is realized through gcForest, and the input data is designed as three-dimensional sequence data of time–temperature–thickness. Our experimental results show that using the algorithm to adaptively extract features from the production process is more accurate and reasonable than several discrete features selected according to experience. Since the three-dimensional sequence data of time–temperature–thickness contains the most detailed information in the rolling production process and is consistent with the progressive relationship of steel composition–process–structure–performance, it is more conducive to the final accuracy of the model. In addition, the time–temperature–thickness sequence data describes the state change process of the rolled piece itself, and the rules learned from this data will not change due to the change in production environment. Therefore, this algorithm can be used to realize the migration learning of data from different production lines. On the one hand, it can solve the problem of insufficient training data for a single production line. On the other hand, it can also improve the generalization ability of the model by learning different processing processes.
As a deep learning algorithm, gcForest has the characteristics of easy parameter adjustment and high precision and is suitable for small sample scenarios, so it has a broad application prospect. In the field of steel production, the quality of steel products can be predicted timely and accurately, and the sampling quantity of steel products can be reduced; moreover, the optimization algorithm can be combined to improve the production process of steel products to achieve cost reduction and efficiency increase; finally, it can be used for the development and design of new products to assist relevant personnel to quickly develop reasonable production paths, reduce product development cycle, and save research and development costs. In addition, it can also be extended to other industrial scene manufacturing process modeling.
In this study, a three-dimensional continuous time series process data sampling method based on time–temperature–deformation was designed. The improved gcForest algorithm was used to extract features and build model along the production process to improve the prediction accuracy of mechanical properties of hot-rolled strip. Based on the data set of three kinds of steel grades collected from the production line, it is compared with a variety of common modeling methods in practical application. The following conclusions can be drawn.
(1) Among the five modeling schemes, the gcForest model shows the best performance with the smallest error, and the prediction accuracy of each strength index exceeds 90%. This feature learning method along the production process can obtain more complete and accurate information about strip steel, which can significantly improve the performance of the model.
(2) In the scenario of small sample size, the prediction effect of the gcForest model is significantly better than that of the neural network model, and it has the smaller prediction error and variance. This deep learning algorithm based on random forest is more suitable for complex learning tasks under small samples and has more advantages in accuracy and prediction stability.
This work was financially supported by the National Natural Science Foundation of China (No. 52004029) and the Fundamental Research Funds for the Central Universities, China (No. FRF-TT-20-06).
The authors declare no competing financial interests.
| [1] |
B. Sun, J.T. Dai, K.K. Huang, C.H. Yang, and W.H. Gui, Smart manufacturing of nonferrous metallurgical processes: Review and perspectives, Int. J. Miner. Metall. Mater., 29(2022), No. 4, p. 611. DOI: 10.1007/s12613-022-2448-x
|
| [2] |
C.C. Tasan, M. Diehl, D. Yan, et al., An overview of dual-phase steels: Advances in microstructure-oriented processing and picromechanically guided design, Annu. Rev. Mater. Res., 45(2015), No. 1, p. 391. DOI: 10.1146/annurev-matsci-070214-021103
|
| [3] |
J.X. Li, W. Wang, Y. Zhou, et al., A review of research status of hydrogen embrittlement for automotive advanced high-strength steels, Acta Metall. Sin., 56(2020), No. 4, p. 444. DOI: 10.11900/0412.1961.2019.00427
|
| [4] |
J. Kaczmarczyk and A. Grajcar, Numerical simulation and experimental investigation of cold-rolled steel cutting, Materials, 11(2018), No. 7, art. No. 1263. DOI: 10.3390/ma11071263
|
| [5] |
H.D. Fu, H.T. Zhang, C.S. Wang, W. Yong, and J.X. Xie, Recent progress in the machine learning-assisted rational design of alloys, Int. J. Miner. Metall. Mater., 29(2022), No. 4, p. 635. DOI: 10.1007/s12613-022-2458-8
|
| [6] |
T.Q. Cheng and G.C. Chen, Prediction of mechanical properties of hot-rolled strip steel based on PCA-GBDT method, J. Phys. Conf. Ser., 1774(2021), art. No. 012002. DOI: 10.1088/1742-6596/1774/1/012002
|
| [7] |
W. Yang, W.G. Li, Y.T. Zhao, B.K. Yan, and W.B. Wang, Mechanical property prediction of steel and influence factors selection based on random forests, Iron Steel, 53(2018), No. 3, p. 44.
|
| [8] |
C.X. Zhang and L. Wu, SVM-based mechanical model used to predict mechanical properties of cold rolled stainless steel sheet, J. Donghua Univ. Nat. Sci., 36(2010), No. 5, p. 559.
|
| [9] |
F.F. Li, Y. Song, C. Liu, R.J. Jia, and B. Li, Research on error distribution modeling of mechanical performance prediction model for hot rolled strip, Metall. Ind. Autom., 43(2019), No. 6, p. 28.
|
| [10] |
J. Yang and S.W. Wu, Property prediction of steel rolling process based on machine learning, Iron Steel, 56(2021), No. 9, p. 1.
|
| [11] |
X.D. Wang and R.D. An, Mechanical properties prediction models based on machine learning for hot rolled strip and application, J. Plast. Eng., 28(2021), No. 10, p. 155. DOI: 10.3969/j.issn.1007-2012.2021.10.020
|
| [12] |
I.J. Ding, N.W. Zheng, and M.C. Hsieh, Hand gesture intention-based identity recognition using various recognition strategies incorporated with VGG convolution neural network-extracted deep learning features, J. Intell. Fuzzy Syst., 40(2021), No. 4, p. 7775. DOI: 10.3233/JIFS-189598
|
| [13] |
T. Mazda, Y. Kajita, T. Akedo, and T. Hazama, Recognition of nonlinear hysteretic behavior by neural network using deep learning, IOP Conf. Ser. Mater. Sci. Eng., 809(2020), art. No. 012010. DOI: 10.1088/1757-899X/809/1/012010
|
| [14] |
G.E. Hinton, S. Osindero, and Y.W. Teh, A fast learning algorithm for deep belief nets, Neural Comput., 18(2006), No. 7, p. 1527. DOI: 10.1162/neco.2006.18.7.1527
|
| [15] |
B. SravyaPranati, D. Suma, C. ManjuLatha, and S. Putheti, Large-scale video classification with convolutional neural networks, [in] T. Senjyu, P.N. Mahalle, T. Perumal, and A. Joshi, eds, International Conference on Transportation Information and Safety (ICTIS), Springer, Singapore, 2020, p. 689.
|
| [16] |
D. Liu, S. Li, and Z.D. Cao, State-of-the-art on deep learning and its application in image object classification and detection, Comput. Sci., 43(2016), No. 12, p. 13. DOI: 10.11896/j.issn.1002-137X.2016.12.003
|
| [17] |
L.N. Zhang and X.L. Sun, Study on Speech Recognition Method of Artificial Intelligence Deep Learning, J. Phys. Conf. Ser., 1754(2021), art. No. 012183. DOI: 10.1088/1742-6596/1754/1/012183
|
| [18] |
W.G. Li, L. Xie, Y.T. Zhao, Z.X. Li, and W.B. Wang, Prediction model for mechanical properties of hot-rolled strips by deep learning, J. Iron Steel Res. Int., 27(2020), No. 9, p. 1045. DOI: 10.1007/s42243-020-00450-9
|
| [19] |
H. Xu, Z.W. Xu, and K. Zhang, Mechanical properties prediction for hot roll steel Using convolutional neural network, [in] L.Q. Pan, J. Liang, and B.Y. Ou, eds, Bio-inspired Computing: Theories and Applications. BIC-TA 2019. Communications in Computer and Information Science, Vol. 1160, Springer, Singapore, 2020, p. 56.
|
| [20] |
B.B. Wang, Y. Song, and J. Wang, A deep learning model for the accurate prediction of the microstructure performance of hot rolled steel, Int. J. Manuf. Res., 16(2021), No. 3, p. 262. DOI: 10.1504/IJMR.2021.117918
|
| [21] |
J.X. Chen, F. Zhao, Y.G. Sun, L Zhang, and YL Yin, Prediction model based on XGBoost for mechanical properties of steel materials, Int. J. Model. Ident. Control, 33(2020), No. 4, p. 322. DOI: 10.1504/IJMIC.2019.107482
|
| [22] |
Y. Song, B. Li, C. Liu, and F.F. Li, Prediction model of mechanical properties of hot rolled strip based on improved stacked self-encoder, Metall. Ind. Autom., 44(2020), No. 6, p. 2.
|
| [23] |
Z.H. Zhou and F. Ji, Deep forest: Towards an alternative to deep neural networks, [in] International Joint Conference on Artificial Intelligence (IJCAI), Melbourne, 2017, p. 3553.
|
| [24] |
Z.H. Zhou and F. Ji, Deep forest, Natl. Sci. Rev., 6(2019), No. 1, p. 74. DOI: 10.1093/nsr/nwy108
|
| [25] |
J.M. Ding, Y.H. Wu, Q.B. Luo, and D. Yi, A fault diagnosis method of mechanical bearing based on the deep forest, J. Vib. Shock, 40(2021), No. 12, p. 107.
|
| [26] |
N. Yao and K. Cheng, Electric power equipment image recognition based on deep forest learning model with few samples, J. Phys. Conf. Ser., 1732(2021), art. No. 012025. DOI: 10.1088/1742-6596/1732/1/012025
|
| [27] |
X.Z. Zhang, J.S. Liu, W.W. He, X.X. Yan, and C.P. Yang, Effect of heating parameters on grain growth behavior of 12%Cr steel, Heat Treat. Met., 45(2020), No. 9, p. 77.
|
| [28] |
D. Xu, C. Ji, H.Y. Zhao, D.Y. Ju, and M.Y. Zhu, A new study on the growth behavior of austenite grains during heating processes, Sci. Rep., 7(2017), art. No. 3968. DOI: 10.1038/s41598-017-04371-8
|
| [29] |
K.M. Agarwal, R.K. Tyagi, V.K. Chaubey, and A. Dixit, Comparison of different methods of Severe Plastic Deformation for grain refinement, IOP Conf. Ser. Mater. Sci. Eng., 691(2019), art. No. 012074. DOI: 10.1088/1757-899X/691/1/012074
|
| [30] |
B. Li, Q.Y. Liu, S.J. Jia, and J. Chen, Grain ultrafining of low-carbon microalloyed steel by intercritical deformation, Heat Treat. Met., 43(2018), No. 5, p. 133.
|
| [31] |
Y.J. Hui, Y. Yu, L. Wang, C. Wang, and J. Chen, Strain induced precipitation of second phase in titanium microalloyed IF steel, Iron Steel, 51(2016), No. 1, p. 82.
|
| [32] |
X.L. Wu, Z.Z. Zhao, Y. Tian, et al., Phase transformation and second-phase precipitation behavior of Nb–Ti microalloyed steel during cooling after deformation, Iron Steel Vanadium Titanium, 29(2008), No. 1, p. 66.
|
| [33] |
J.F. Cheng, Z.D. Liu, and G.B. Tang, Numerical simulation of phase transformation process for carbon steel during Run-out table cooling, Iron Steel, 47(2012), No. 1, p. 55.
|
| [34] |
M.J. Sun, X.M. Xia, and L.Z. Long, Finite element analysis of temperature field and phase transition in laminar-like cooling and coiling process of strip steel, Hot Work. Technol., 49(2020), No. 10, p. 131.
|
| [35] |
S.Z. Wang, Z.J. Gao, G.L. Wu, and X.P. Mao, Titanium microalloying of steel: A review of its effects on processing, microstructure and mechanical properties, Int. J. Miner. Metall. Mater., 29(2022), No. 4, p. 645. DOI: 10.1007/s12613-021-2399-7
|
| [1] | Rui-ming Wang, Zhi-ying Gao, Wen-rui Wang, Yang Xue, De-yi Fu. Dynamic characteristics of the planetary gear train excited by time-varying meshing stiffness in the wind turbine [J]. International Journal of Minerals, Metallurgy and Materials, 2018, 25(9): 1104-1112. DOI: 10.1007/s12613-018-1661-0 |
| [2] | Teng-yue Gao, Kui-ren Liu, Qing Han, Bin-shi Xu. Enrichment of copper and recycling of cyanide from copper–cyanide waste by solvent extraction [J]. International Journal of Minerals, Metallurgy and Materials, 2016, 23(11): 1258-1263. DOI: 10.1007/s12613-016-1347-4 |
| [3] | Ke Xu, Yong-hao Ai, Xiu-yong Wu. Application of multi-scale feature extraction to surface defect classification of hot-rolled steels [J]. International Journal of Minerals, Metallurgy and Materials, 2013, 20(1): 37-41. DOI: 10.1007/s12613-013-0690-y |
| [4] | Jiao Luo, Bin Wu, Miao-quan Li. 3D finite element simulation of microstructure evolution in blade forging of Ti-6Al-4V alloy based on the internal state variable models [J]. International Journal of Minerals, Metallurgy and Materials, 2012, 19(2): 122-130. DOI: 10.1007/s12613-012-0526-1 |
| [5] | Guifang Wu, Ke Xu, Jinwu Xu. Application of a new feature extraction and optimization method to surface defect recognition of cold rolled strips [J]. International Journal of Minerals, Metallurgy and Materials, 2007, 14(5): 437-442. DOI: 10.1016/S1005-8850(07)60086-3 |
| [6] | Tichang Sun. Action time effect of lime on its depressive ability for pyrite [J]. International Journal of Minerals, Metallurgy and Materials, 2004, 11(3): 193-196. |
| [7] | Shixiong Zhang, Guozhu Zeng, Tao Peng. Space-time principles of reducing stripping in furrow pits [J]. International Journal of Minerals, Metallurgy and Materials, 2003, 10(1): 16-17. |
| [8] | Lequan Min, Xisong Dong. Analytical criteria for local activity of CNN with five state variables and application to Hyper-chaos synchronization Chua's circuit [J]. International Journal of Minerals, Metallurgy and Materials, 2002, 9(4): 311-318. |
| [9] | Berthe Kya, Yang Yang. Improve Fractal Compression Encoding Speed Using Feature Extraction and Self-organization Network [J]. International Journal of Minerals, Metallurgy and Materials, 2001, 8(4): 306-310. |
| [10] | Qing Li, Deling Zheng, Jianlong Zhou. Stabilization of Chaotic Time Series by Proportional Pulse in the System Variable Based on Genetic Algorithm [J]. International Journal of Minerals, Metallurgy and Materials, 1999, 6(3): 228-229. |
| 1. | Bingtao Ren, Chenchong Wang, Yuqi Zhang, et al. Industrial big data analysis strategy based on automatic data classification and interpretable knowledge graph. Journal of Materials Informatics, 2025, 5(2) DOI:10.20517/jmi.2024.85 |
| 2. | Tetiana Vlasenko, Szymon Glowacki, Vitaliy Vlasovets, et al. Increasing Exploitation Durability of Two-Layer Cast Mill Rolls and Assessment of the Applicability of the XGBoost Machine Learning Method to Manage Their Quality. Materials, 2024, 17(13): 3231. DOI:10.3390/ma17133231 |
| 3. | Qiuna Wang, Jiquan Sun, Jiaxuan Yang, et al. Online Partition-Cooling System of Hot-Rolled Electrical Steel for Thermal Roll Profile and Its Industrial Application. Processes, 2024, 12(2): 410. DOI:10.3390/pr12020410 |
| 4. | Xiaoyang Luo, Rongping Guo, Qiwen Zhang, et al. Prediction of Mechanical Properties of Cold-Rolled Steel Based on Improved Graph Attention Network. Symmetry, 2024, 16(2): 188. DOI:10.3390/sym16020188 |
| 5. | Qiwen Zhang, Wenkui Wu, Xingchang Tang, et al. Prediction of mechanical properties of rolled steel based on dual-attention multiscale convolutional neural network. Materials Today Communications, 2024, 41: 110432. DOI:10.1016/j.mtcomm.2024.110432 |
| 6. | Xiangyu Xu, Lu Zhang, Zifei Wang, et al. Critical precipitation behavior of MnTe inclusions in resulfurized steels during solidification. International Journal of Minerals, Metallurgy and Materials, 2024, 31(8): 1849. DOI:10.1007/s12613-023-2757-8 |
| 7. | Cheng-Yan Ding, Jun-Cheng Ye, Jia-Wei Lei, et al. An interpretable framework for high-precision flatness prediction in strip cold rolling. Journal of Materials Processing Technology, 2024, 329: 118452. DOI:10.1016/j.jmatprotec.2024.118452 |
| 8. | Jianye Xue, Kang Li, Tongshuai Zhang, et al. Data-Driven Soft Sensor for Hot Strip Mill Process Based on Multiscale Information-Fusion Attention Network. IEEE Transactions on Instrumentation and Measurement, 2024, 73: 1. DOI:10.1109/TIM.2024.3450083 |
| 9. | Qiwen Zhang, Yichao Wang. Multi-task prediction of hot rolled steel mechanical properties based on CA-MMCE. Materials Today Communications, 2024, 39: 108594. DOI:10.1016/j.mtcomm.2024.108594 |
| 10. | Cheng-Yan Ding, Jun-Cheng Ye, Long-Jun Wang, et al. A novel cost-sensitive quality determination framework in hot rolling steel industry. Information Sciences, 2024, 678: 121054. DOI:10.1016/j.ins.2024.121054 |
| 11. | Hongyi Wu, Borui Zhang, Zhiwei Li. Small Sample-Oriented Prediction Method of Mechanical Properties for Hot Rolled Strip Steel Based on Model Independent Element Learning. IEEE Access, 2024, 12: 197300. DOI:10.1109/ACCESS.2024.3517752 |
| 12. | Zida Liu, Diyuan Li, Quanqi Zhu, et al. Intelligent method to experimentally identify the fracture mechanism of red sandstone. International Journal of Minerals, Metallurgy and Materials, 2023, 30(11): 2134. DOI:10.1007/s12613-023-2668-8 |
| 13. | M. Prem Anand, M. Anand, M. Adams Joe, et al. Lightweight Bi-LSTM method for the prediction of mechanical properties of concrete. Multimedia Tools and Applications, 2023, 83(18): 54863. DOI:10.1007/s11042-023-17796-3 |
| 14. | Kaizheng Wang, Yitong Fu, Shunzhen Zhou, et al. Cloud-fog-based approach for smart wildfire monitoring. Simulation Modelling Practice and Theory, 2023, 127: 102791. DOI:10.1016/j.simpat.2023.102791 |
| 15. | Chengyan Ding, Jie Sun, Xiaojian Li, et al. A high-precision and transparent step-wise diagnostic framework for hot-rolled strip crown. Journal of Manufacturing Systems, 2023, 71: 144. DOI:10.1016/j.jmsy.2023.09.007 |
| 16. | Suyang Wang, Qingqi Zhu, Jiangshan Hao, et al. Predicting Mechanical Properties of Steel Based on Ensemble Learning and Bayesian Optimization. 2024 Twelfth International Conference on Advanced Cloud and Big Data (CBD), DOI:10.1109/CBD65573.2024.00013 |
| 17. | Robert Gibbs, Cinzia Giannetti, Thomas Baynes, et al. Leveraging Machine Learning Insights at the Hot Strip Rolling Mill, Tata Steel, Port Talbot. 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), DOI:10.1109/I2MTC60896.2024.10560883 |
| 18. | Jianhui Wu. Implementation of Power Grid Fault Diagnosis and Prediction Platform Based on Deep Learning Algorithms. 2023 IEEE 15th International Conference on Computational Intelligence and Communication Networks (CICN), DOI:10.1109/CICN59264.2023.10402185 |
| 19. | Deepak Banerjee, Neha Sharma, Deepak Upadhyay, et al. Advancing Metallurgical Quality Assurance: Predictive Modeling for Defect Detection using CNN-Random Forest Fusion. 2024 4th Asian Conference on Innovation in Technology (ASIANCON), DOI:10.1109/ASIANCON62057.2024.10838169 |
Figures(10) / Tables(2)
Copyright©2019 Editorial Office of International Journal of Minerals, Metallurgy and Materials 京ICP备13030111号
Supported by:
Beijing Renhe Information Technology Co., Ltd.
Email:
info@rhhz.net
| Grade | YS | TS | EL | |||||||||||
| Code | α | β | Accr / % | Code | α | β | Accr / % | Code | α | β | Accr / % | |||
| Q235B | C | 4 | 9 | 94.97 | C | 4 | 9 | 96.36 | B | 3 | 9 | 94.97 | ||
| 510L | D | 5 | 9 | 96.22 | D | 5 | 4 | 93.78 | A | 5 | 6 | 94.00 | ||
| 610L | B | 3 | 6 | 91.92 | A | 3 | 6 | 92.57 | A | 3 | 8 | 91.67 | ||
DownLoad:
CSV
| Model | Accr for Q235B / % | Accr for 510L / % | Accr for 610L / % | ||||||||
| YS | TS | EL | YS | TS | EL | YS | TS | EL | |||
| gcForest | 94.97 | 96.36 | 94.97 | 96.22 | 93.78 | 94.00 | 91.92 | 92.57 | 91.67 | ||
| CRF | 93.22 | 93.92 | 91.82 | 95.78 | 92.89 | 93.11 | 89.10 | 89.75 | 91.67 | ||
| RF | 86.22 | 89.72 | 92.17 | 94.89 | 91.11 | 91.33 | 90.23 | 88.05 | 91.10 | ||
| NN | 87.62 | 87.62 | 84.83 | 92.22 | 73.33 | 73.11 | 83.45 | 73.36 | 77.54 | ||
| BNN | 86.36 | 81.82 | 79.02 | 77.78 | 81.78 | 62.67 | 67.66 | 71.26 | 56.29 | ||
DownLoad:
CSV
