Development of a Centralized Classifier for Decentralized Decision Making

With the current development pace of IoT technology and analytical instrumentation, the acquisition of information from devices or operations that are geographically distributed is becoming increasingly common. In some applications, these devices/operations are essentially alike but executed with slightly distinct machines/measurement systems. Such differences have been handled by constructing local models, in a fully decentralized and independent way. However, it is now possible to conceive improved centralized schemes to derive a unified model that takes advantage of the entire data lake. In this work, we demonstrate how such an endeavor can be achieved in the scope of a real case study of predicting an important property (coagulation) of waste lubricant oil (WLO) using FTIR spectra collected at different locations (laboratories). The unified classifier uses a compound mapping of Partial Least Squares for Discriminant Analysis (PLS-DA) and the Bayesian linear classifier. The best prediction models were selected by screening over 36 potential models. The best candidate models to be implemented in practice showed an accuracy greater than 98 %, while retaining less than four latent variables.