Adaptive feedback linearizing control with neural-network-based hybrid models for MIMO nonlinear systems

The difficulties associated with the control of nonlinear systems are especially profound when it involves MIMO systems. One possible approach to tackle the system nonlinearities is to employ the input-output feedback linearizing control strategy. However, this controller can only perform well when the exact knowledge of the system is known. To alleviate this problem, it is proposed here to use neural-network-based hybrid models to model the system nonlinear functions. Particularly, multilayer feedforward networks are used to model the unknown parts of the system nonlinear functions, and then the network outputs are combined with the available knowledge to form the hybrid models. Simulation studies are shown on set point tracking and disturbance rejection studies of two continuous stirred tank reactors, one with single reaction, and another one with multiple reactions. The results showed that the control systems were able to track the set points and reject disturbances with only slight overshoot during the transient period.