(255c) Building Deep Learning Based Predictive Model and Advisory Control System for a Blast Furnace Operation

Authors: 
Lee, Y. M., IBM T.J. Watson Research Center
Yeo, K., IBM T.J. Watson Research Center
Nguyen, N., IBM T.J. Watson Research Center
Melnyk, I., IBM T.J. Watson Research Center
Kalagnanam, J., IBM T.J. Watson Research Center
Kim, Y. S., POSCO
Developing adaptive, data-driven predictive models by employing machine-learning (ML) [1] or deep learning (DL) [2] to make inferences / decisions from real-time sensor data is emerging as a key technology of cognitive manufacturing, or Industry 4.0. Internet of Things (IoT) is a new technological foundation for connectivity and real time messaging of data coming from many sensors, devices, equipment and unit operations (stages) in complex manufacturing production processes. Advancement and availability of ML and DL algorithms along with IoT enable the development of an intelligent plant advisory system that adaptively generates predictive models and executes an appropriate predictive model in real time to predict future state of the process and compute optimal control set point.

In this work, we applied various ML and DL techniques to develop predictive models that can accurately predict status of a complex manufacturing process, a blast furnace operation of steel manufacturing process. A blast furnace is a complex operation that involves multiple chemical reactions and phase transitions of materials, which are difficult to model using first principle equations. At the same time, because of the complex multiscale nature of the process, in which the response time of the input materials, such as iron ore, coke, oxygen, water, pulverized coal (PC), etc., have wide variations from order of minutes to hours, it is very difficult to develop a data-driven model in the conventional machine-learning approaches. Here, a time-series prediction DL model, called Recurrent Neural Network (RNN), is employed to build a predictive model. Particularly, we use the Long Short-Term Memory (LSTM) network [1], which is capable of learning multi-scale temporal dependency structures, to build models for predicting key state variables of the blast furnace operation. The LSTM seems to capture complex non-linear dynamics well and is shown to outperform conventional ML algorithms, such as Sparse Linear Model (LASSO), Decision Tree, Gradient Boosting, and Gaussian Processes, in the prediction of blast furnace status.

We describe the modeling approach and architectures of LSTM models for predicting several key response variables of the blast furnace operation, and prediction accuracy. A formulation of model predictive control (MPC) model that computes the optimal set points of key control variables of the blast furnace is also presented.

References:

[1] C. M. Bishop, Pattern recognition and machine learning, Springer, 2006.

[2] Y. LeCun, Y. Bengio & G. Hinton, Deep learning. Nature14539, 2015.

[3] S. Hochreiter & J. Schmidhuber. Long short-term memory. Neural Comput. 9(8): 1735-1780, 1997.