Robust Multiscale Partial Least Square Regression
Noor Hassan Mohammed Amin Riad · OakTrust (Texas A&M University Libraries) · 2020
Process industries always need to be monitored and controlled to optimize their performances, ensure their long-term operation reliability, and provide safe operation. Therefore, models that can accurately predict the process behavior are needed. Partial Least Square (PLS) model is a well-established modeling technique for multivariate data; however, there are limitations that affect its accuracy. For example, the presence of outliers can bias the regression coefficients and degrade the prediction accuracy of the model. In addition, the presences of measurement noise/error degrade the adequacy of the constructed PLS model. Several methods have been developed to deal with these drawbacks. Partial robust estimator (PRM) was developed to deal with the presence of outliers, and the Integrated multiscale PLS (IMSPLS) method was developed to deal with the presence of Gaussian noise. However, these methods cannot deal with the presence of both Gaussian noise and outliers. Therefore, in this work, a method that integrates the advantages of both PRM and IMSPLS is developed and is called Robust Multiscale PLS (RMSPLS) method. RMSPLS optimizes the weighting scheme for each observation, the decomposition depth of filtered Gaussian noise, and number of principal components of model simultaneously. The performance of the RMSPLS is assessed and compared to both PRM and PLS, and it is shown that it can effectively deal with both Gaussian noise and outliers, and that it is more robust than these methods at different noise levels and types. These advantages are due to the RMSPLS ability to simultaneously suppress the effect of both Gaussian noise and outliers on the model’s prediction accuracy.