A Machine Learning Approach for Polymer Classification Based on the Thermal Response under Data Scarcity─Tested on PMMA

Abstract

An important application of machine learning techniques is the intelligent nondestructive testing of polymers. However, data scarcity and class imbalance (for real applications) shape some of the involved challenges. In order to tackle these challenges, an intelligent screening framework for poly­(methyl methacrylate) (PMMA) samples is studied here. An efficient thermal and experimental test is designed and coupled with several machine learning tools for quick classification with four features. Furthermore, a set of sampling techniques are employed/compared; these techniques are Random oversampling (ROS), synthetic minority oversampling technique (SMOTE), and adaptive synthetic sampling (ADASYN). Then, Linear Discriminant Analysis, K-Nearest Neighbor, Naive Bayes, decision tree, random forest, pattern recognition network, support vector machine, and ensemble learning are employed for classification. For assessing these algorithms, their performances are evaluated using a collection of metrics (i.e., Geometric-mean, F1 score, Matthews correlation coefficient, accuracy, true positive rate, true negative rate, positive predictive value, and negative predictive value). Among others, the average G-mean measures, which is a paramount measure for assessing the imbalance data, are increased from 75.97% (original data) to 94.24% (ROS), 93.49% (SMOTE) and 91.27% (ADASYN). That is a clear proof of successful oversampling. The final results show that ROS oversampling coupled with Ensemble classification methods can significantly improve all performance metrics for PMMA classification.