A fault detection technique for the steel manufacturing process based on a normal pattern library

Fault detection is increasingly important in the steel manufacturing process. If faults occur, the consequences can be extremely serious in terms of human life, environmental impact and/or economic loss. Therefore, it is necessary to prevent potential accidents by detecting faults as early as possible. However, there are various kinds of faults, and most faults have only a few or even no precedents in the past. Since it is very difficult to obtain time-series data during accidents, the key is to detect the difference from the normal state by utilizing only data under normal operation. As another problem in the steel manufacturing process, each process comprises many thousands of pieces of equipment, and it would be very difficult to construct a fault detection model manually for each of these devices. Therefore, a technique which makes it possible to construct each model and adjust its parameters automatically is required. In order to solve these two major problems, this study proposes a new fault detection concept based on the pattern library. Furthermore, a combination of a static pattern library and a dynamic pattern library is discussed as a means of covering as many pieces of equipment as possible. Offline numerical simulations for both static and dynamic equipment were conducted, and the possibility of the proposed technique was confirmed.