Lawrence, A.J. and Harris, C.J.
Auto-tuning of Low Order Controllers by Direct Manipulation of Closed Loop Time Domain Measures.
I Mech E Proc. Part I, J. Systems and Control Engineering, 210, .
Full text not available from this repository.
Many controller tuners are based on linear models of both the controller and process. Desired performance is often predetermined or adjusted in a manner which is not directly related to the desired response. All physical processes contain nonlinearities, commonly of the actuator saturating type, and many controllers contain heuristics for implementation in real systems, such as anti-integral wind up in PID controllers. For different processes a range of closed loop response shapes are desired, often described by features of the response such as rise time, over-shoot and settling time. This paper investigates the possibility of basing controller tuning on closed loop system response data such that desired performance is incorporated directly in terms of familiar time domain features or labels thus eliminating the need for a mathematical process model and repeated tuning reformulations to achieve the desired performance. A controller tuning method named Label Based Neuro-Tuning (LBNT) is developed and analysed by application to PID controller tuning for process models indicative of real process behaviour. Simulations and numerical investigation indicate that LBNT is a viable technique for tuning low order controllers for SISO processes. Tuning is straightforward, flexible and copes well with process parametric chances and performance specification reformulation. The drawbacks are a complicated pretune phase, a limited selection of suitable labels and a difficulty in defining general classes of tuning problems for its application. The technique is not based on the assumption of process linearity but due to the inability to characterise classes of input signals and operating points the types of process nonlinearity are restricted. The controller may be nonlinear, but must be structurally predetermined, and an input/output process model of arbitrary structure is required.
Actions (login required)