Fast Locally Weighted PLS Modeling for Large-Scale Industrial Processes

Locally weighted partial least-squares (LW-PLS) is an efficient just-in-time (JIT) modeling method, which can handle process collinearity, nonlinearity, and time-varying characteristics. However, it is not suitable for modeling large-scale industrial processes due to its huge computational cost. To solve this issue, the present work proposes a novel fast LW-PLS (FLW-PLS) method that can handle large-scale process data well. FLW-PLS is designed based on the exact Euclidean locality-sensitive hashing algorithm. Unlike standard LW-PLS which implements a brute-force linear search of similar samples over a reference data set, FLW-PLS searches for similar samples in sublinear time. Thus, significant computational savings can be obtained by FLW-PLS. The effectiveness of the proposed FLW-PLS method was validated through the case studies of predicting the silicon content and phosphorus content in the ironmaking process. The application results have shown that when faced with large-scale data, FLW-PLS can significantly reduce the prediction time without significantly reducing the prediction accuracy in comparison to the conventional LW-PLS method.