Spectral preprocessing combined with feature selection improve model robustness for plastics samples classification by LIBS

Introduction: Nowadays, the widespread use of plastic products has significantly contributed towards environmental pollution caused by waste plastics. Laser-induced breakdown spectroscopy (LIBS), an emerging spectroscopic technology, has shown great potential for rapid sorting and recycling of plastics. However, the poor robustness of the classification model severely limits the large-scale application of LIBS technology in plastic sorting and recycling. Methods: In this research, we used spectral preprocessing combined with feature selection to improve the robustness of the support vector machine (SVM) classification model for four typical plastic samples (ABS, nylon, 3240, and its modified product FR-4). LIBS spectral data were collected under different experimental conditions, then we defined robustness over time (ROT), robustness over time and different focusing lenses (ROT&RFL), and robustness over time and different manufacturers (ROT&RDM) to assess model performance. The feature importance of the preprocessed spectra was evaluated using the Relief-F algorithm, and the maximum accuracy of the validation set was 92.6% when inputting the first 19 most important features. Eventually, the optimal model was used for the prediction of the test set. Results and discussion: The ROT of the original spectrum, spectrum preprocessing, and spectral preprocessing combined with feature selection were 58.4%, 79.1%, and 98.47%, respectively. Similarly, ROT&RFL for the same methods were 65.54%, 75%, and 95.25%, respectively. ROT&RDM were 65.5%, 67%, and 93.92%, respectively. The results demonstrate that spectral preprocessing combined with feature selection can significantly improve the robustness of the classification model, and the proposed method is feasible for plastic sorting and recycling.