Statistical Evaluation of Sinter Quality Indices Based on Data Mining & Machine Learning Approaches
Steel making using the integrated blast furnace (BF) process route is still the dominant steel production method and sinter is one of the most important iron sources within this process. Along this route, sinter plants offer unique capability to recycle residues and fines from iron ores for the steel production. Nevertheless, the current practice of maintenance and sintering quality management has not yet developed into an Industry 4.0 standard and is still controlled manually. Current industrial practice for monitoring sinter quality (tumbler index and particle size distribution) is still based on time-consuming, manual sampling methods.The major focus of this work is the optimization of the sinter plant, in particular the continuous monitoring of sinter quality indices by applying data driven methods and leveraging big data technologies and machine learning (ML) approaches to predict sinter quality indices, like the particle size of the target fraction. This will involve the use of sinter plant operational data and the exploration of new measurement techniques to acquire additional high-resolution data of the sintering process. The new measurement techniques include:Optical & acoustical measurement, at the discharge of the sinter, consideration of the power consumption of the hot crusher and continuous particle size distribution (PSD) measurement along the transfer route to the BF, Figure 1.The research will provide implementation of new data collection methods, data preparation, ML model creation and training supported by metallurgical analysis and concluded with the ML model predicting and monitoring sinter quality indices in real-time.Different supervised ML methods like Decision Tree methods (Gradient Boost), Random Forests, Neural Networks, and other models will be compared based on their prediction accuracy.To determine the accuracy of the new online measurement techniques and ML models, the generated data is validated using the conventionally provided quality data and new online particle size measurements of the sinter plant (sinter strength, screen analysis) to calculate the training, test error (MSE) and correlation coefficient.Model-based predicted quality information along the sinter production process should enable combined optimization of the sintering plant and BF operation by providing expected sinter quality (target fraction size) in real time. This significant time advantage should allow faster response to current operating conditions at the sinter plant and BF before the results of the conventional quality assessment are available. The predicted high-resolution quality indices should further allow quantification of sinter quality variations during short-term events at the sinter plant and could partially replace conventional quality monitoring, resulting in lower sampling costs.
Further, the BF operator can decide whether to burden the current sinter qualities depending on the actual course of the BF. This could result in lower operating costs and to a more ecological mode of operation e.g., in lower consumption of raw materials (energy) decrease of return fines and, as well as lower emissions of greenhouse gases and other pollutants.