Design of Data-Oriented GMDH-Based Controller

PID control schemes have been widely used in most industrial control systems, but it has been difficult to determine a suitable set of PID gains as most industrial systems are nonlinear. Although there have been proposals for Cerebellar Model Articulation Controller (CMAC) classified as neural networks, and a design scheme for an intelligent PID controller that uses a CMAC-PID tuner, CMAC-PID tuners have two problems. One issue is that a CMAC must be trained on-line in order to obtain their optimum weights. Another issue is that the CMAC has high computational costs and memory reqirements for some micro controllers. In order to train a CMAC off-line, a CMAC-FRIT (a combination of CMAC and Fictious Reference Iterative Tuning) scheme has been proposed in previous research. FRIT is a scheme to determine control parameters by using a set of experimental data. According to the CMAC-FRIT scheme, CMAC-PID tuners can be trained offline by using a set of operating data. This paper proposes to address the problems of memory requirements and computational costs with a method that expresses a CMAC-PID tuner as a simple nonlinear function by using a Group Method of Data Handling (GMDH). According to the proposed method, a network of N-Adaline (units expressed by a simple nonlinear function) replaces a CMAC-PID tuner (which is trained in advance with a set of operating data), enabling the proposed algorithm to be easily programmed on a micro controller, even if it is a commodity micro controller. The effectiveness of the proposed method is validated by an experiment in order to demonstrate the proposed method, the algorithm is programmed on a general purpose micro controller, which is applied to a magnetic levitation device.