In operating chemical plants, an operator has to monitor the state of plants and control process variables, such as temperature, pressure, liquid level, concentration of products, and so on. Therefore, these variables need to be measured online, but all of them are not easy to measure online because of technical difficulty, measurement delays, high investment cost, and so on.
In chemical plants, soft sensors are widely used to estimate a process variable which is difficult to measure online. The soft sensor is an inferential model constructed between variables which are easy to measure online and one which is difficult to measure online, and a value of an objective variable is estimated by the model. By using soft sensors, a value of objective variables can be estimated with high accuracy.
We have developed new soft sensor methods in order to construct predictive models with high accuracy. It is conceivable that changing factors of each process variable measured in chemical plants, such as changes of external temperature and feed composition, are mutually independent. Therefore, we have tried to extract these reasons from measured data by using statistical methods. We have succeeded in selecting an appropriate model from several models by using the ICA based fault detection and classification model. This produced a prediction of an objective value with higher accuracy than traditional methods. In addition, by analyzing real industrial data measured in several chemical plants, we are developing general soft sensor methods.