In the presented research, machine learning methods were applied to the prediction of longitudinal cracks in steel slabs during continuous casting. We employ a deep learning approach to process 68 thermocouple signals as a multivariate time series (MTS) along with 32 static features, which encompass both chemical composition and process information. Our deep learning approach integrates two distinct parallel modules, followed by an aggregation block; a Convolutional Neural Network (CNN) processes the thermocouple MTS, while in parallel, the static data undergo processing via a Fully Connected Network (FCN). To enhance the performance of the CNN, we incorporate two Squeeze and Excitation (SE) blocks, which act as an attention mechanism across different channels. By integrating chemical information with MTS in the detection system, we improve the performance of defect detection by 15% relatively.

Annealing and galvanization production lines in steel mills run continuously to maximize production throughput. As a part of this process, individual steel coils are joined end-to-end using mash seam welding. Weld breaks result in a production loss of multiple days, so non-destructive, data-driven techniques are used to detect and replace poor quality welds in real-time. Statistical models are commonly used to address this problem as they use data readily available from the welding machine and require no specialized equipment. While successful in finding anomalies, these statistical models do not provide insight into the underlying process and are slow to adapt to changes in the machine’s or material’s behavior. We combine knowledge-based and data-driven techniques to create an incremental grey-box welding current prediction model for detecting anomalous welds, resulting in a powerful and interpretable model. In this work, we detail our approach and show evaluation results on industrial welding data collected over a period of 15 months containing behavioral shifts attributed to machine maintenance. Due to its incremental nature, our model resulted in two-thirds fewer rejected welds compared to statistical models, thus greatly reducing production overhead. Grey-box modeling can be applied to other welding features or domains and results in models that are more desirable for the industry.