Hybrid static-sensory data modeling for prediction tasks in basic oxygen furnace process

Abstract

In this paper, we propose a novel data-driven prediction system for Multivariate Time Series (MTS) in an industrial context, where classic relational data contain key information in order to properly interpret the MTS. Particularly we focus on the accurate endpoint prediction of temperature and chemical composition at the basic oxygen furnace, which is a step in the steel production pipeline where liquid iron is refined to steel. The precise prediction of temperature is important for proper process control while reaching the target chemical composition is essential for quality control. Our deep learning methodology employs two modules followed by an aggregation block; a Convolutional Neural Network (CNN) handles the MTS, while in parallel, the static data is processed by a Fully Connected Network (FCN). We enhance the CNN performance by adding two Squeeze-and-excitation (SE) blocks, which act like an attention module over the different channels. By taking the MTS data into account we improve the prediction by up to 10% relative over the models which only consider the static data. The hybrid FCN-CNN-SE architecture slightly improves the state-of-the-art MTS approaches by 2%, with less outliers on the prediction of final temperature and phosphorus concentration, while being easier to implement and more scalable to larger datasets and input space than current solutions.

Dynamic Operation Optimization of Complex Industries Based on a Data-Driven Strategy

“…Additionally, with the growing environmental awareness of industrial processes, wastewater discharge and treatment in the petroleum and chemical industries have become crucial aspects worthy of attention. Stress-strain produced by steel heat Wang [11], He [12], Liu [13], Sala [14], Song [15],…”

Toward learning steelmaking—A review on machine learning for basic oxygen furnace process

“…Figure 4 reveals that static features have been predominantly employed for BOF data‐driven modeling. Despite the significant role of process time‐series features, their availability has been limited to a few BOF datasets [26–28] . In Brämming et al., [29] a comprehensive blend of all available features—process parameters, image, off‐gas data, and vessel vibration audiometry measurements—was utilized to estimate foam height and endpoint phosphorus prediction.…”

“…The application of recurrent neural networks (RNNs), specifically long short‐term memory (LSTM), has been explored for handling time‐series data [53] . A deep learning framework based on fully connected networks (FCN) and CNN has been developed for regression tasks, taking into account both static and multivariate time series information [28] …”

Forecasting Steel Production in the World—Assessments Based on Shallow and Deep Neural Networks

“…For evaluation of the performance of the models, the metrics used are the root mean square error (RMSE) and the mean absolute percentage error (MAPE). RMSE is calculated as in Formula (13) and MAPE is calculated as in Formula (14).…”

“…Figure 1 shows the result after searching the topic under study, which shows that the databases returned four articles in total. They focus on the following issues: hybrid static-sensory data modeling for prediction tasks in the basic oxygen furnace process [14]; temperature prediction for a reheating furnace by the closed recursive unit approach [15]; detecting and locating patterns in time series using machine learning [16]; and deep learning for blast furnaces [17]. Other studies found focus on steel forecasting in China [18], in Japan [19], and in Poland [20].…”