(568ab) Process Control Method Based on the Inverse Analysis of Soft Sensors Considering Controllability

In process industry, effective control methods for set point change and disturbance suppression are required to improve productivity and stabilize product quality. Monitoring of process variables is essential for efficient control, but it is not always possible to measure variables such as viscosity and density online because of technological limitations, large measurement delays and high investment costs. In order to monitor these variables, soft sensors are widely used for online estimation. Soft sensors are statistical models constructed between variables that are easy to measure online and variables that are not, by using process data.

Effective control for process variables that are difficult to measure can be accomplished by using online estimated values from soft sensors. PID controllers and model predictive controllers are widely used for process control. Optimization of control parameters, however, is difficult when the process is large and complex, resulting in many processes being operated inefficiently. Therefore, the purpose of our research is to perform more efficient control using information obtained from inefficient daily operation data.

In order to achieve this goal, we have been developing a process control method using soft sensor models. We refer to this method as inverse soft sensor-based feedback control (ISFB). In this work, we only covered set point changes of a controlled variable y. In this method, a soft sensor model is constructed, where y is the output variable, and both U and X are the input variables. U stands for manipulated variables and Xstands for process variables. This soft sensor model is constructed as a dynamic model by considering time delayed input variables.

The optimal control strategy of U is determined by conducting inverse analysis of the soft sensor model in the transition interval of U, by using a pre-defined shape. This shape is created using the knowledge of the process, where the candidates for the transition of U are generated by varying the parameters of the shape.

The soft sensor calculates the estimated y values for each candidate, where the one whose objective function is the most optimal is selected, for example Integral of Squared Error (ISE). The soft sensor model, however, has an estimation error associated and, thus, a deviation from the set point remains. Therefore, the selected candidate is recursively corrected through feedback by the difference between the observed value of yand the estimated value from the soft sensor.

The proposed method was applied to control set point changes of outlet flow concentration in a simulated CSTR system. The assumption of this case study is that the PI controller data for set point change has been obtained, where this data is used to construct the soft sensor. As a result, it was confirmed that the proposed control method was faster and more stable than the PI controller. Furthermore, in practical terms the parameters of the proposed method are automatically determined, what reduces the work load of the operator.