Utilizing Big Data for Development of Soft Sensor

Conventional static soft sensor does not handle the dynamic of processes adequately. Instead the sequential historical data should be considered when developing the soft sensor. It follows that the resultant soft sensor is able to extract dynamic information. With advancement of data storage facilities, there is a deluge of data. Excessive computation burden can result if the complete big data are used. There is thus a need to utilize the data diligently. To this end, a latent variable model (LVM) based strategy to select useful data to overcome the problem of variable correlations and large number of sample is proposed. The uncertainty information obtained from the Gaussian process (GP) model is used as the selection criterion. The combination of LVM and GP model is developed together with an update scheme for advance improvement. This is done in order for the soft sensor to better reflect the current status of the process. The proposed method is applied to the data from an industrial process.