Evaluation and improvement of model robustness for plastics samples classification by laser-induced breakdown spectroscopy

The robustness of classification models is important in the real-world application of laser-induced breakdown spectroscopy (LIBS). This study using seven well-known chemometric models (ANN, CART, kNN, LDA, PLS-DA, SVM, and SIMCA) to classify LIBS spectral data from four types of typical plastics samples (ABS, Nylon, 3240, and its improved product FR-4). The robustness of these models for data acquired by different excitation sources (85 mJ pulse @ 1064 nm and 44 mJ pulse @ 532 nm) and over time (collected in November 2015, August 2016, December 2016 and September 2018) were evaluated and compared. The training set was constructed with the spectra acquired by 1064 nm wavelength laser excitation in August 2016. The effect introduced by 5 preprocessing methods (autoscaling, mean-centering, normalized by the total area, normalized by the maximum, and standard normal variate (SNV)) on the robustness of the model was investigated. The performance (accuracy and robustness) of different models was compared and analyzed. The results showed that the robustness of ANN, LDA, and PLS-DA model performed well and the ANN model was best. The experimental results demonstrated that the robustness of the model for LIBS spectra could be improved by using a suitable preprocessing method.