Online Prediction of Mechanical Properties of the Hot Rolled Steel Plate Using Time-series Deep Neural Network

Zhao Yang, Yifan Wang, Feng Xu, Xiaoqiang Li, Kai Yang, Weihao Xia, Jiajia Cai, Qian Xie, Qiyan XuISIJ 2023 · DOI

Abstract

The prediction of steel’s properties by the process parameters has been of great interest because they can effectively reduce production costs and improve product quality. At present, the extensively used artificial neural network (ANN) can only reveal the correlation between parameters and mechanical properties from the perspective of statistics but loses the critical information on time-series correlation in the steel production process. In this work, time-series neural networks based on long short-term memory (LSTM) were established to predict the steel plate’s yield strength (YS), ultimate tensile strength (UTS), and elongation (EL). The results verified that the proposed LSTM model exhibited sufficient accuracy and outperformed the classical algorithms (SVM, random forest, ANN), with mean squared error reduced by 30% when compared with ANN. Also, the sensitivity analysis proved that LSTM was more capable to make full use of the information contained in input parameters and achieved better generalization performance. Notably, the interpretation of the results was more consistent with the real physical metallurgical process of a hot-rolled steel plate and can provide a better understanding of the physical metallurgical process.

Introduction

In the modern steel industry, accurately predicting mechanical properties like Yield Strength (YS), Ultimate Tensile Strength (UTS), and Elongation (EL) is vital for quality control and cost reduction. Traditionally, this was achieved through physical metallurgical models. While theoretically sound, these models are often computationally intensive and difficult to implement for real-time industrial applications.

Recent shifts toward Machine Learning (ML) have introduced Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) to the field. However, these models typically treat input data—such as chemistry, furnace temperatures, and rolling speeds—as a single, flat vector. This approach ignores the logical, temporal sequence inherent in steel production. For instance, the temperature in the heating furnace directly influences how the steel behaves during the subsequent rolling stage. By ignoring this "time-series" relationship, standard ML models lose critical physical context, leading to lower accuracy and poor interpretability.

Methods

The core innovation of this research is the application of a Long Short-Term Memory (LSTM) network. Unlike standard ANNs, LSTMs are designed to "remember" information from earlier steps in a sequence, making them ideal for modeling manufacturing processes.

Schematic of the hot rolling production process showing the sequential stages from slab reheating to cooling.

The researchers divided the production data into three logical stages to mirror the physical process:

  1. Element Composition: The chemical "DNA" of the steel.
  2. Heating Stage: Parameters including furnace temperature and heating time.
  3. Rolling Stage: Parameters such as plate thickness, finish rolling temperature, and cooling rates.

By feeding these into the LSTM sequentially, the model "carries over" the influence of chemical elements into the heating phase, and the combined influence of both into the final rolling phase. This architecture was compared against three benchmarks: SVM, Random Forest (RF), and a standard fully connected ANN.

Datasets

The study utilized a substantial industrial dataset comprising 19,008 steel plate records from a manufacturing plant in China. The model utilized 50 input parameters, including chemical elements (C, Mn, Si, Nb, Ti) and process variables (furnace discharge temperature, rolling thickness ratio).

Overview of the industrial dataset, including (a) chemical composition ranges, (b) temperature distribution, and (c) distribution of steel grades.

Data was normalized using Z-scores to ensure feature parity, and categorical data, such as specific steel grades, was handled via One-hot encoding to be processed by the neural network.

Key Insights and Results

The researchers found that a 3-stage division (Chemistry -> Heating -> Rolling) provided the optimal balance of temporal context and data density, achieving a high accuracy of $R^2 = 0.960$.

Key performance metrics include:

  • Superior Accuracy: The LSTM outperformed ANN (0.942), Random Forest (0.940), and SVM (0.935).
  • Error Reduction: Mean Squared Error (MSE) was reduced by 30% compared to the standard ANN.
  • Physics-Consistent Sensitivity: Using Mean Impact Value (MIV) analysis, the LSTM prioritized parameters like Finish Rolling Start Temperature (FRST) and Heating Time. In physical metallurgy, these are known to control grain size and precipitation, the primary drivers of steel strength.
  • Robustness: The model maintained high performance across various steel grades (e.g., Q345B, S355J0), demonstrating its utility in diverse production environments.

Conclusion

By treating steel production as a sequential process rather than a static snapshot, the LSTM model provides a more "physics-aware" approach to property prediction. This method not only improves statistical accuracy but also aligns more closely with established metallurgical theory, offering a robust tool for real-time industrial quality control.

Summary updated

Abstract

The prediction of steel’s properties by the process parameters has been of great interest because they can effectively reduce production costs and improve product quality. At present, the extensively used artificial neural network (ANN) can only reveal the correlation between parameters and mechanical properties from the perspective of statistics but loses the critical information on time-series correlation in the steel production process. In this work, time-series neural networks based on long short-term memory (LSTM) were established to predict the steel plate’s yield strength (YS), ultimate tensile strength (UTS), and elongation (EL). The results verified that the proposed LSTM model exhibited sufficient accuracy and outperformed the classical algorithms (SVM, random forest, ANN), with mean squared error reduced by 30% when compared with ANN. Also, the sensitivity analysis proved that LSTM was more capable to make full use of the information contained in input parameters and achieved better generalization performance. Notably, the interpretation of the results was more consistent with the real physical metallurgical process of a hot-rolled steel plate and can provide a better understanding of the physical metallurgical process.
BibTeX
@inproceedings{yang2023online,
  author = {Zhao Yang and Yifan Wang and Feng Xu and Xiaoqiang Li and Kai Yang and Weihao Xia and Jiajia Cai and Qian Xie and Qiyan Xu},
  title = {{Online Prediction of Mechanical Properties of the Hot Rolled Steel Plate Using Time-series Deep Neural Network}},
  booktitle = {ISIJ},
  year = {2023},
  doi = {10.2355/isijinternational.isijint-2022-383},
}
Figure