An adaptive soft sensor based on multi-state partial least squares regression

Soft sensor is widely used in chemical processes to monitor the product´s quality which is unmeasurable or measured with low frequency. There are many kinds of methods to develop validated soft sensors. One of most popular methods is Partial Least Square (PLS) algorithm. Although it works well, the traditional PLS cannot satisfy the process with multiple operating regimes. To remove deviation among different operating regimes, an adaptive Multi-State PLS (MSPLS) algorithm is proposed to build a soft sensor. The proposed algorithm includes key variable selection, operating state division, adaptive scheme, etc. Applications on a continuous stirred tank reactor and a industrial process demonstrate the performance of the preset soft sensor.