Autogenerated Multilocal PLS Models without Pre-classification for Quality Monitoring of Nonlinear Processes with Unevenly Distributed Data

To achieve the desired product qualities for an operating process, theoperational status must be monitored. Operating conditions are changed because of thevariability in the quality of raw materials and equipment characteristics. The operatingconditions are also changed to produce various product grades and to meet time-to-marketdemands. This makes the process data collected within a wide region to be distributedunevenly, so conventional multivariate statistical process control techniques that use asingle global model are not suited to monitoring. Data is partitioned to construct each localmodel, but partitioning data without considering variable correlations can cause theinformation to be lost. This study uses a multilocal partial least-squares (ML-PLS) modelto monitor a wide operation process. Without pre-selecting the data for each local model, ML-PLS automatically reinforces the datathat pertains to each local model and weakens the data that does not pertain to other local models. The entire range of operation isclustered, and multiple PLS models are constructed using this clustered data. Statistical indices for quality-based process monitoringin the latent and observed variable spaces are derived using the learned ML-PLS models. The effectiveness and accuracy of theproposed method are demonstrated using a numerical example and a practical application for a chemical process.

Automated summary

To achieve desired product qualities, monitoring the operational status is necessary. Traditional multivariate statistical process control techniques are not suitable for monitoring unevenly distributed process data. This study proposes the use of a multilocal partial least-squares (ML-PLS) model to monitor a wide operation process.