Ensemble Approaches for Quartz and Silica Gel Identification using EDX spectra

During a manufacturing process, hard disk drive (HDD) might be contaminated by some materials such as Quartz and Silica Gel. These contaminations induce a major failure in the HDD. Identifying them is a necessary approach to track back to the sources and prevent this issue to happen again in the future. In this work, we are interested in identifying Quartz and Silica Gel by analyzing their spectra obtained from using an Energy Dispersive X-ray Spectroscopy (EDX) instrument. A challenge is that these two materials are from the same functional group (Silicon Dioxide) and have quite similar EDX spectra which cause a classifying problem to experts. In this work, we proposed a parallel ensemble learning approach based on EDX spectra to improve the classification between Quartz and Silica Gel. Four ensemble methods were studied and compared: Hard Voting, Soft Voting, Weighted Hard Voting, and Weighted Soft Voting. We applied and chose the base models of the ensemble methods from 12 well-known machine learning (ML) models. Experimental results show that the Soft Voting and Weighted Soft Voting techniques achieved a maximum accuracy of 82.5% and significantly outperformed individual ML models in classifying Quartz and Silica Gel.