(268g) Detection of Valve Stiction in Closed-Loop Nonlinear Systems Using Volterra and Stable AR Model Identification

Presence of valve stiction in control loops lead to sustained oscillations, resulting in loss of productivity and leading to loss in overall profit. It is therefore, essential to detect the presence of sticky valves in a control loop which can help in performing valve maintenance to reduce the impact of stiction. The characteristic nature of stiction in a valve is that it produces non-linear sustained oscillations in a closedloop system. This non-linear phenomenon introduced by stiction has been analyzed and discussed in several articles. Further, numerous methods have been developed to detect and quantify stiction in linear closed-loop systems. All of these methods assume that stiction is the only source of non-linearity in the closed-loop system. However, industrial process are mostly non-linear and therefore, presence of process non-linearities are inevitable. This necessiates for the development of a technique which can detect the presence of valve stiction in non-linear closed-loop systems (with process being non-linear).

In this current study, a Volterra model based technique has been developed to detect the presence of stiction in closed-loop nonlinear systems. The core idea behind this technique is to approximate the process output using a second-order volterra system along with an auto-regressive component. Since chemical processes are large time constant systems (low pass systems), auto-regressive component of the model plays an important role in the model identification. However, the model identification task can lead to unstable models even though the open-loop process model is stable. This is mainly due to one or more of the following reasons, namely, (i) noise, (ii) availability of limited amount of data and (iii) nonlinear distortions. In this work, in addition to detecting stiction in non-linear closed-loop systems, a novel method has been proposed to identify Volterra models along with a stable auto-regressive component of the model.