Research On Soft Sensor Modeling Based On Multiway Partial Least Square For Batch Process

The focus of many industries has now shifted to the manufacturing of higher-value-added products that are mainly produced through batch processes to meet today's rapidly changing market. In most batch processes, some key production quality indexes mostly obtained by off-line analysis are difficult to measure on-line. So the requirements of online real time control and optimizing operation always can not be met. Soft sensor is developed greatly to resolve this kind of problems.Multiple statistic regression modeling method is an effective tool for soft sensor modeling which takes partial least square(PLS) as the core. As a derivative method of PLS, multiway partial least square(MPLS) method is widely used in batch process modeling. However, the traditional MPLS method presents a bad prediction accuracy in multi-operation stage batch process modeling, and it also encounters the following problems in practical application:uneven-length modeling data, complex model structure, poor stability, and prediction accuracy is seriously affected by the unavailable future process data.Considering that multiplicity of operation stage and repetition are inherent characteristics of many batch processes, combining with advantages of MPLS in dealing with high dimensional and coupled data, new methods for batch process modeling and model updating based on MPLS have been developed as follows:(1) Considering the multiplicity of operation stage, a new MPLS modeling method based on phase-specific average trajectory is proposed. Meanwhile, a complementary method for unknown variables future observation is given to realize on-line forecasting of quality index.(2) To overcome the problem that traditional PLS cannot be updated on-line, a new block-wised recursive MPLS, which combines recursive PLS and phase-specific average trajectory MPLS, is proposed and improved. This paper also presents modeling and model updating steps for quality index prediction.(3) The proposed methods are applied to piercing process to realize energy consumption on line prediction and model updating. Simulation results with real industrial data show that they are effective.