Neural Network-Based Data Mining Techniques for Steel Making Institute of Technology, Cambridge, USA

CONTENTS 24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402 24.2 Productivity from Information Technology (PROFIT) Initiative . . . . . . . . . . . 403 24.3 Description of Predictive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406 24.4 NNRUN – ANN Training Suite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 24.5 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409 24.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412

The blast furnace is the heart of any steel mill, where raw materials, such as coke and sinter, are combined to produce pig iron, the precursor of steel. The hot metal temperature (HMT) is an important indicator of the state of the blast furnace, as well as the quality of pig iron produced. Traditionally, due to the highly complex and nonlinear relationships between the various chemical inputs and HMT, models based on conventional statistical forecasting techniques have provided very poor results. To mitigate this problem, the paper highlights a neural network based approach to modeling a blast furnace to generate accurate HMT predictions. Issues such as data processing and augmentation, as well as optimal neural network architectures, are discussed. This paper presents in detail a neural network-training suite, NNRUN, which automates the search for the optimal neural network configuration given userdefined boundaries. The paper presents some of the research findings including data pre-processing, modeling and prediction results based on HMT data received from the industrial blast furnace of a leading steel manufacturer in Asia. Feed-forward neural networks with a single hidden layer have been found to be very accurate predictors of HMT. A prototype system based on the neural network based methodology is currently being implemented at the blast furnace. The relevant research for this paper was conducted by members of the Productivity from Information Technology Initiative at MIT, where research is being conducted into the use of neural networks for a wide variety of application domains, ranging from optimization of inventory operations to automated reading of handwritten bank checks.