Machine Learning Applications in the Steel Production Industry

The blast furnace is a countercurrent chemical metallurgical reactor, and monitoring the blast furnace is of paramount importance in producing a quality product. In this work, the behavior of the main chemical elements that make up the metallic charge was studied in detail. The elements affecting the final quality of the steel can be limited to carbon (C), silicon (Si), manganese (Mn), phosphorus (P), and sulfur (S). Manganese receives less attention because it often plays only a minor role in the production process. Carbon is the main raw material used as a reducing agent for iron oxides in blast furnaces. Metallurgical coke is the main supplier of carbon and accounts for a significant percentage of the cost of cast iron. Therefore, keeping consumption low has always been an important goal. In recent decades, pulverized coal injection (PCI) has become increasingly popular in the global steel industry to reduce specific coke consumption per ton produced. During production, it is important to control the thermal condition of the reactor, and the amount of silicon dissolved in the hot metal is used as an indicator. If the amount of dissolved silicon is high, it indicates an excess of energy, probably an excess of metallurgical coke in the reactor. However, if the silicon level is low, the reactor may need to be warmed up and countermeasures taken to correct the problem. Phosphorus is an element that causes little change in metallurgical processes (sintering, coke ovens, and blast furnaces) but is extremely detrimental to quality, productivity, and cost, especially in steel production, which requires a low phosphorus content in its chemical composition. The amount of phosphorus dissolved in the metal depends on the chemical composition of the metallic charge (sinter, pellets and lump ore). It is possible to reduce the phosphorus content in hot metal by controlling the reactor temperature and the residence time of the molten metal in the reactor. Sulfur in steel is an undesirable residue that negatively affects properties such as ductility, toughness, weldability and corrosion resistance. The production of low-sulfur steel is extremely important for shipbuilding and for pipes for the oil industry. The application of neural network solutions is very popular due to its versatility and reliability. The application of neural network technology in steel production is new and there are few works on this topic. The database of this work consists of Big Data corresponding to 3450 operating days and divided into seven groups: Air Injection, Top gas, Thermal Control, Fuels, Minerals, Hot metal and Slag. The Big Data were divided into three parts, comprising a total of 3 different samples, each with 1150 variables, with 9 variables of interest divided into 3 groups. The variables studied were silicon, phosphorus, sulfur, carbon, temperature, daily production, coke rate, PCI and fuel rate. The artificial neural networks were trained using the Levenberg-Marquardt algorithm with sigmoidal activation functions in the hidden layers. The committee machine has 12 classifiers for each element. Each classifier has 3 neural networks with the same architecture and the same number of neurons, but the input variables used to train the neural network are different. The artificial neural networks have 10, 20, 25, 30, 40, 50, 75, 100, 125, 150, 175, and 200 neurons in each layer. In this study, the quality of the neural network model was evaluated using Pearson's root mean square error (RMSE) and correlation coefficient (R). The result obtained with the committee machine is better than the results of the individual models, but still the models alone are able to provide good and convergent results. Looking at the RMSE values between training, validation and testing, no differences were found that could indicate overfitting, nor do the results obtained indicate underfitting. Neural networks operating in a committee system have been shown to have better predictive power than models from the literature. In conclusion, neural networks working in a committee system can be used in practice because the model is both a predictive tool and an action guide due to the excellent correlations between the real values and the values calculated by the neural network.