In recently, industry has been characterized as a highly complex process and equipped with modern control systems. Advanced process control (APC) is widely used by control engineers to optimize the process operation in order to produce a product with specified quality. The success of the on-spec quality production is based on the model between the quality and process variables. It is difficult to develop their relationship using the first principle due to the complexity of the process. Many of the environmental processes, the quality is measured in analytical laboratory at irregular intervals and it involves resources and considerable time delay. Reliable on-line prediction of quality is extremely beneficial in this environment.
A soft sensor is an inferential model based on software technique to determine the value of a process variable. This is in contrast to physical sensor that directly measures the value of the process variable. Some kinds of the case of complex system where the process mechanism is not well-understood, empirical models developed by system identification techniques are used to derive the implication among the variables. Among the techniques, neural network technology and multiple linear regressions have been employed as a data-driven approach to modeling problem.
Neural Networks are powerful tool for nonlinear system identification and they have variety applications such as process control, signal processing, pattern recognition, process monitoring, finance and chemistry. Many advantages can be identified from their applications to inferential modeling. They can easily grasp the nonlinear relationship among the variables and approximate it to the desired degree. Within the structure of neural nets, a model can be obtained with little prior detailed knowledge of process behavior. However, their successful use and applications have to be driven by expertise and a well sounded development procedures in order to avoid failures and disappointment. Multiple linear regressions can also be used for inferential model development for their straightforwardness and simpleness of relationship explanation.
However, they suffer from the numerical problems when a strong correlation exists among process variables, and they may degrade the model with the collinearity. Partial least squares method has established itself as a robust alternative to the ordinary least squares technique in the analysis of correlated data. It will provide a transparent model and a unique solution, in contrast to the “"black box”" and non-uniqueness nature of Neural Nets.
We use soft-sensor techniques that already explained to overcome the problems with environmental processes. The applications will be taken to approach various environmental processes and that constructed soft-sensors lead optimum condition to make the best products.