Application of Neural Network Methodology to the Modelling of the Yield Strength in a Steel Rolling Plate Mill

Ah Chung TsoiNeurIPS 1991

Introduction

In the steel manufacturing industry, the "hot rolling" process is essential for transforming thick, raw steel slabs into flat, usable plates. To ensure the final product meets rigorous quality and safety standards, engineers must accurately predict the steel's yield strength—the specific stress threshold at which the metal begins to deform plastically—as it passes through the heavy rollers.

Traditionally, predicting yield strength has relied on theoretical models grounded in physics or empirical "best-fit" equations, such as the inverse hyperbolic sine law. However, this paper demonstrates that a flexible, data-driven approach using Multivariate Adaptive Regression Splines (MARS)—a tree-based neural network architecture—can significantly outperform these traditional metallurgical models and standard machine learning classifiers.

The Challenge of Hot Rolling Models

Theoretical and empirical models face two major hurdles in real-world industrial applications:

  1. Measurement Noise: Unlike controlled laboratory environments, real-world steel mills are chaotic. Measuring variables like strain and strain rate during active production is notoriously imprecise, leading to noisy data that can throw off rigid equations.
  2. Model Rigidity: Traditional models are often highly "tuned" to specific, narrow datasets. When mill conditions change over time or when processing steel at different temperatures, these rigid models frequently fail to generalize, leading to inaccurate predictions.

The author of this study sought a more "intelligent" modeling approach capable of handling the noisy, nonlinear nature of industrial data without requiring a perfect, comprehensive theoretical understanding of every metallurgical phase change.

A Data-Driven Solution: MARS

The core of this research is the application of MARS (Multivariate Adaptive Regression Splines). MARS acts as a hybrid between a decision tree and a regression model. Instead of using simple "if-then" splits that result in constant values at the leaves (like standard decision trees), MARS uses "knots" to join different polynomial functions, known as splines, together.

For this specific application, the author utilized an additive model featuring piecewise cubic polynomials. This architecture allows the model to approximate highly complex, nonlinear relationships for each input variable independently, offering a much smoother and more accurate fit than linear approximations.

To prove its efficacy, the MARS approach was benchmarked against several established methods:

  • Conventional Tuning: An inverse hyperbolic sine law model updated via Kalman filters (both batch and recursive).
  • Standard Classifiers: CART (Classification and Regression Trees), C4.5 (a decision tree algorithm), and Bayesian classification.

Experimental Setup

The study leveraged a robust, real-world industrial dataset to train and test the models:

  • Scale: 2,877 data points derived from 180 individual steel plates.
  • Complexity: Each plate underwent approximately 16 distinct "passes" through the mill rollers.
  • Input Variables: Temperature, strain, strain rate, and interpass time (the duration between roller passes).
  • Target Variable: Yield stress.
Cyclic data representing individual passes through the steel rolling mill.

For the final performance comparison, the dataset was partitioned into a training set of 78 plates (1,242 passes) and a testing set of 16 plates (252 passes).

Key Insights

The application of the MARS model yielded several fascinating insights into the hot rolling process:

  • The Dominance of Strain Rate: The model revealed that strain rate has the most significant influence on yield stress, followed by temperature and then strain. This provides highly actionable feedback to process engineers regarding which parameters require the tightest operational control.
  • Automatic Feature Selection: Interestingly, for this specific dataset, interpass time played no significant role in the model's predictions. While this might be a peculiarity of the specific mill's data, it highlights MARS's inherent ability to perform automatic feature selection, ignoring irrelevant variables.
  • Smooth Nonlinearity: Piecewise cubic polynomials provided a vastly superior fit compared to linear ones. This confirms that the underlying physics of metal deformation is smoothly nonlinear, rather than jagged or simplistic.
  • Generalization over Interpretation: While MARS provided superior predictions, its "breakpoints" (the temperatures where the model's mathematical behavior changes) did not perfectly align with traditional metallurgical theory. For instance, the model found breakpoints at 1017°C and 1129°C, rather than the theoretically expected 925°C. This highlights a common machine learning trade-off: the model is exceptionally good at predicting what will happen, even if its internal math doesn't perfectly map to why classical physics says it happens.

Results and Performance

The MARS model consistently and significantly outperformed all other methods, particularly when evaluated on unseen testing data.

Noisy measurement data typical of industrial hot rolling processes.

When yield stress was divided into 10 classes for a standardized comparison, the results were stark:

  • MARS: Achieved a remarkably low prediction error of only 6.2%.
  • CART: 12.99% error.
  • C4.5: 16.14% error.
  • Bayesian Classifier: 65.4% error.

Furthermore, the standard deviation of prediction errors for MARS was considerably lower than that of the Kalman filter-tuned models. This indicates that MARS is not just more accurate, but also far more reliable and stable across varying industrial conditions.

Limitations and Future Directions

The primary limitation of the MARS approach is its lack of physical interpretability. Because it is a purely data-driven neural network methodology, its internal parameters and breakpoints do not easily translate back into standard, universally applicable metallurgical formulas.

To bridge this gap, the author suggests a promising path for future research: Constrained MARS. By artificially forcing the model's "breakpoints" to exist only within specific temperature regions where known microstructural changes occur, researchers could potentially create a "gray-box" model. This future model would combine the unparalleled predictive power and flexibility of neural networks with the physical validity and interpretability of traditional metallurgy.

Summary updated

BibTeX
@inproceedings{tsoi1991application,
  author = {Ah Chung Tsoi},
  title = {{Application of Neural Network Methodology to the Modelling of the Yield Strength in a Steel Rolling Plate Mill}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1991},
}
Figure