Real-Time Forecasting of Subsurface Inclusion Defects for Continuous Casting Slabs: A Data-Driven Comparative Study

Chihang Wei, Zhihuan SongSensors 2023 · DOI

TL;DR

This paper tackles the critical challenge of detecting subsurface inclusion defects in continuous casting steel slabs in real-time. The authors introduce two novel data-driven methods, SR-KDLS and SDAE-BPNN, which significantly outperform traditional approaches and common machine learning techniques by accurately and rapidly forecasting these defects, crucial for improving steel quality and preventing costly production issues.

What Problem Does This Paper Address?

Subsurface inclusion defects are a major concern in continuous casting of steel, impacting the quality of final products and potentially leading to costly accidents like breakouts during hot rolling. Traditional methods for detecting these defects are often slow, destructive, or rely on complex and sometimes inaccurate physical models. The key challenge is to detect these defects in real-time as the steel is being cast, before it's too late to intervene or when the defect is deeply embedded and difficult to observe. This paper addresses the need for accurate, timely, and non-destructive methods for subsurface inclusion defect forecasting.

Methods

The paper presents a comparative study of various data-driven methods, introducing two novel approaches:

  • Scatter Regularized Kernel Discriminative Least Squares (SR-KDLS): This method is designed to directly provide forecasting information, bypassing the intermediate step of generating low-dimensional embeddings. It combines a least squares loss function with a regularization term that maximizes between-class scatter and minimizes within-class scatter. This aims to create a more discriminative feature space for accurate classification. It leverages the Reproducing Kernel Hilbert Space (RKHS) for flexibility.
  • Stacked Defect-Related Autoencoder Back Propagation Neural Network (SDAE-BPNN): This approach utilizes a deep learning architecture. It starts with Defect-Related Autoencoders (DAEs) that are designed to extract features specifically related to defects, unlike standard autoencoders that might focus on general input reconstruction. These DAEs are then stacked hierarchically to form a Stacked DAE (SDAE), allowing for the learning of increasingly complex, defect-related features. Finally, a Back Propagation Neural Network (BPNN) is used as a forecasting layer on top of the SDAE's deepest features.

The paper also evaluates several baseline data-driven methods including Linear Discriminant Analysis (LDA), Logistic Regression (LR), Support Vector Machines (SVM), and XGBoost.

Datasets

The study uses data collected from a real-life continuous casting process in China. This dataset comprises 33 process variables and includes eight different categories of casting slabs. The data was collected between October and November 2018. The datasets exhibit significant class imbalance, with a much higher percentage of normal (negative) samples compared to defective (positive) samples. To address this, the SMOTE (Synthetic Minority Oversampling Technique) was employed to augment the training data and mitigate the imbalance problem.

Key Insights and Ideas

  • Real-time forecasting is achievable and crucial: The study demonstrates that data-driven methods can predict subsurface inclusion defects with minimal delay (under 0.01 ms), which is essential for industrial application.
  • Defect-specific feature extraction matters: Standard autoencoders might not be optimal for defect detection because they prioritize general data reconstruction. The DAE approach, which explicitly incorporates defect information during training, shows promise for better feature learning.
  • Deep learning for complex features: Stacking multiple DAEs (SDAE) allows for the extraction of deeper, more abstract defect-related features, leading to improved forecasting performance compared to shallower models.
  • Discriminative information is key: The SR-KDLS method's focus on maximizing inter-class separability and minimizing intra-class compactness in the feature space is a powerful strategy for improving classification accuracy.
  • Class imbalance is a significant challenge: The performance of all models is shown to be inversely related to the imbalance degree of the dataset, highlighting the importance of techniques like SMOTE.

Key Results

  • Superior performance of novel methods: The developed SR-KDLS and SDAE-BPNN methods consistently achieved higher F1 scores across most of the eight casting slab categories compared to traditional methods like LDA, LR, SVM, and XGBoost. For instance, for category 6, SR-KDLS and SDAE-BPNN achieved F1 scores of 90.9% and 91.2% respectively, significantly outperforming SVM's 71.4%.
  • Timely predictions: All tested models, including the novel ones, met the real-time forecasting requirement, with online forecasting times consistently below 0.1 ms per sample.
  • Computational efficiency: While the offline training for SR-KDLS and SDAE-BPNN was more computationally intensive than simpler models, this is acceptable as training is an offline process. The online prediction speed is extremely fast, ensuring no delay in practical applications.
  • Impact of imbalance: When the imbalance degree was large (e.g., category 1), the F1 scores for most methods were lower, underscoring the challenge of limited minority class data.

Limitations and Future Work

  • Activation function selection: The paper used a sigmoid activation function for BPNN, noting that future work should analyze the impact of different activation functions for optimal performance.
  • Time delay estimation: Current data alignment relies on rough estimates of time delays. Developing more sophisticated variable time delay estimation techniques is suggested.
  • Data augmentation for imbalance: While SMOTE was used, the paper acknowledges that randomly augmented information might not always be reliable. Future work could explore new data-augmentation techniques, potentially using Generative Adversarial Networks (GANs) or transferring information from data-rich to data-scarce regions.
  • Integrating process knowledge: The authors suggest that combining data-driven methods with existing process knowledge (metallurgy functions) could further enhance forecasting performance and improve minority class reliability.
  • Generalization of SR-KDLS: The SR-KDLS is currently designed for binary forecasting but can be extended to multi-class scenarios.

Summary updated

Abstract

Subsurface inclusions are one of the most common defects that affect the inner quality of continuous casting slabs. This increases the defects in the final products and increases the complexity of the hot charge rolling process and may even cause breakout accidents. The defects are, however, hard to detect online by traditional mechanism-model-based and physics-based methods. In the present paper, a comparative study is carried out based on data-driven methods, which are only sporadically discussed in the literature. As a further contribution, a scatter-regularized kernel discriminative least squares (SR-KDLS) model and a stacked defect-related autoencoder back propagation neural network (SDAE-BPNN) model are developed to improve the forecasting performance. The scatter-regularized kernel discriminative least squares is designed as a coherent framework to directly provide forecasting information instead of low-dimensional embeddings. The stacked defect-related autoencoder back propagation neural network extracts deep defect-related features layer by layer for a higher feasibility and accuracy. The feasibility and efficiency of the data-driven methods are demonstrated through case studies based on a real-life continuous casting process, where the imbalance degree drastically vary in different categories, showing that the defects are timely (within 0.01 ms) and accurately forecasted. Moreover, experiments illustrate the merits of the developed scatter-regularized kernel discriminative least squares and stacked defect-related autoencoder back propagation neural network methods regarding the computational burden; the F1 scores of the developed methods are clearly higher than common methods.
BibTeX
@inproceedings{wei2023real,
  author = {Chihang Wei and Zhihuan Song},
  title = {{Real-Time Forecasting of Subsurface Inclusion Defects for Continuous Casting Slabs: A Data-Driven Comparative Study}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23125415},
}
Figure