Forecastable Component Analysis and Partial Least Squares Applied on Process Monitoring

Forecastable Component Analysis (ForeCA) is a new feature extraction method for multivariate time series. ForeCA can find an optimal transformation to dig out the potential forecastable information structure from large amounts of data. This paper combines ForeCA with PLS for industrial process monitoring. This method overcomes the drawback that partial least squares(PLS) rarely use dynamic timing characteristics of system, so it can reflect the dynamic nature of industrial processes better. We use PLS for regression after appropriate forecastable components selected Finally, we construct CUSUM statistic and SPE statistic for monitoring industrial processes. Simulation results on the Tennessee Eastman (TE) process illustrate the effectiveness of the proposed method for detecting slow drift fault.