Gains of a roll force AGC (Automatic Gain Controller) have been calculated at the first locked-on-time by FSU (Finishing-mill Set-Up model) in hot strip mills and usually these values are not adjusted during the operating time. Consequently, this conventional scheme cannot cope with the continuous variation of system parameters and circumstance, though the gains can be changed manually with the aid of experts to prevent a serious situation such as inferior mass production. Hence, partially uncontrolled variation still remains on delivery thickness. This paper discusses an effective online algorithm which can adjust the gains of the existing control system by considering the effect of time varying variables. This algorithm improves the performance of the system without additional cost and guarantees the stability of the conventional system. Specifically, this paper reveals the major factors that cause the variation of strip and explores the relationship between AGC gains and the effects of those factors through the analysis of thickness signal which occupy different frequency bands. The proposed tuning algorithm is based on the above relationship and realized through ANFIS (Adaptive-Neuro-based Fuzzy Interface System) which is a very useful method because its fuzzy logics reflect the experiences of professionals about the uncertainty and the nonlinearity of the system. The effectiveness of the algorithm is shown by several simulations which are carried out by using the field data of POSCO corporation (South Korea).

Obtaining a linearizing feedback and a coordinate transformation map is very difficult, even though the system is feedback linearizable. It is known that finding a desired transformation map and feedback is equivalent to finding an integrating factor for an annihilating one-form for single input nonlinear systems. It is also known that such an integrating factor can be approximated using the simple C.I.R method and tensor product splines. In this paper, it is shown that m integrating factors can always be approximated whenever a nonlinear system with m inputs is feedback linearizable. Next, m zero-forms can be constructed by utilizing these m integrating factors and the same methodology in the single input case. Hence, the coordinate transformation map is obtained.