The classification of FTIR plastic bag spectra via label spreading and stacking

• Main Authors: Almanifi, Omair Rashed Abdulwareth, Ng, Jee Kwan, Anwar P. P., Abdul Majeed
• Format: Article
• Language: English
• Published: Penerbit UMP 2021
• Subjects: TJ Mechanical engineering and machinery; TS Manufactures
• Online Access: http://umpir.ump.edu.my/id/eprint/34001/1/The%20classification%20of%20FTIR%20plastic%20bag%20spectra.pdf; http://umpir.ump.edu.my/id/eprint/34001/; https://doi.org/10.15282/mekatronika.v3i2.7390; https://doi.org/10.15282/mekatronika.v3i2.7390
• Institution: Universiti Malaysia Pahang

Whereas plastics are a group of the most useful materials, widely used in all walks of life, the plastic waste that is produced daily poses a great threat towards wildlife and the planet as a whole. The use of biodegradable plastics is an important step in combating the plastic crisis. FTIR spectroscopy is a non-destructive method used for identifying different types of materials, however interpreting spectra produced by such spectrometers is both susceptible to human error, and time-consuming, not to mention that the industry suffers from a great of specialists, in the field of spectroscopy. Utilising machine learning as a method of filling the mentioned issue is suggested by this paper. Four pipelines were investigated, consisting of two machine learning algorithms, a stacked model that stacks the KNN, SVM and RF algorithms together, and Label spreading, as well as two different dimensionality reduction methods namely; SVD and UMAP. The pipelines studied seemed to show great predictivity at 100% classification accuracy acquired by the SVD-Stacked pipeline when data was sampled using an Agilent Cary 660 FTIR Spectrometer, and 99.18% by the same model when IDIR BP10 spectrometer was employed for sampling instead. The semi-supervised learning model (Label Spreading) seemed to achieve close enough accuracy at 99.82% in the case of the former dataset, and 97.54% for the latter, at a labelling rate of only 10% of the full datasets.