Robust BOD5 soft sensor design using local learning

In wastewater treatment plants, It's difficult to acquire online data of BOD5 (Biochemical Oxygen Demand for 5 days) due to its characteristic and unreliability of on-line sensors. Furthermore, although soft sensors models are widely used in wastewater treatment, only a few approaches for soft sensors models are designed to address the problems currently existing in the wastewater treatment. In such situations, there is a really need to develop a reliable, robust and real time soft sensor. To facilitate this soft sensor design, first, this paper presents a robust Principal Component Analysis (PCA), which combines PCA with statistical Jolliffe Parameters, to detect outliers and increases the robustness of soft sensors. Second, advanced local learning methods including Radial basis function (RBF) and Locally weighted projection regression are introduced to address the modeling issues. These strategies have the potentials to significantly improve the measurement of BOD5. © 2010 IEEE.