(373e) Soft Sensor Models Based On a Process Variable and Dynamics Selection Method and Support Vector Regression

Plant operators have to monitor the operating conditions of industrial plants and control process variables, such as temperature, pressure, liquid level, and concentration of products. These variables, therefore, must be measured online. However, it is not easy to measure all variables online due to technological limitations, large measurement delays, and high investment costs. Thus, soft sensors have been widely used to estimate process variables that are difficult to measure online. An inferential model is constructed between those variables that are easy to measure online and those that are not, and an objective variable y is then estimated using that model. Through the use of soft sensors, the values of ycan be estimated with a high degree of accuracy in real-time. In addition, soft sensors can give useful information in terms of fault detection by working with hardware sensors in parallel.

Soft sensor models predicting values of y should be constructed with only important explanatory variables X in terms of predictive ability, better interpretation of models and lower measurement costs. Besides, some X-variables can affect y with time-delays. Therefore, we previously proposed the method for selecting important X-variables and optimal time-delays of each variable simultaneously, by modifying the genetic algorithm-based wavelength selection method that is one of the wavelength selection methods in spectrum analysis. This proposed method was named as genetic algorithm-based process variable and dynamics selection (GAVDS). The GAVDS method can select time-regions of X-variables as a unit by using process data that includes process variables that are delayed in the range from zero to a set/given maximum value.

However, a nonlinear relationship between X and y cannot be modeled with GAVDS because a partial least squares (PLS) method is used as a modeling technique. In fact, nonlinearity between X and y has been reported in the fields of soft sensors. We therefore attempted to select appropriate X-variable regions and construct a highly predictive model in the case of a nonlinear relationship between X and y, and then, we proposed a new method for variable region selection based on GAVDS and support vector regression (SVR), which is a nonlinear regression method. The proposed method is named as GAVDS-SVR. Both the representation of nonlinearity between X and yand variable region selection could be achieved employing the GAVDS-SVR method.

To verify the effectiveness of the proposed method, we applied it to simulation data and soft sensor analysis. In the analysis of simulation data, we compared PLS, GAVDS, SVR, and GAVDS-SVR in the case of nonlinearity between X, in which variables have high correlation, and y. From the results of the analysis of the simulation data, it was found that although GAVDS could handle monotonically increasing and decreasing functions, appropriate variable selection was not accomplished by the GAVDS method when the nonlinear function is a quadratic function with a local minimal value. In this case, the proposed GAVDS-SVR method could select appropriate X-variable regions and construct a highly accurate and predictive model.

Afterward, we applied the GAVDS-SVR method to soft sensor analysis and used actual industrial data obtained during an industrial polymer process. The proposed method could construct models predicting polymer quality with predictive ability higher than that of models constructed by other methods. In addition, process variables and the dynamics of the variables could be selected by employing the GAVDS-SVR method.

The proposed method will be able to reduce the change in parameters and effort of maintenance of soft sensor models during model reconstruction, and probably contribute to the optimal positioning of sensors.