Real-time classification of aluminum metal scrap with laser-induced breakdown spectroscopy using deep and other machine learning approaches
Dillam Jossue Díaz-Romero,Simon Van den Eynde,Wouter Sterkens,Alexander Eckert,Isiah Zaplana,Toon Goedemé,Jef Peeters
DOI: https://doi.org/10.1016/j.sab.2022.106519
2022-10-01
Abstract:In the recycling industry, the use of deep spectral convolutional networks for the purpose of material classification and composition estimation is still limited, despite the great opportunities of these techniques. In this study, the use of Laser-Induced Breakdown Spectroscopy (LIBS), Machine Learning (ML), and Deep Learning (DL) for the three-way sorting of Aluminum (Al) is proposed. Two sample sets of Al scrap are used: one containing 733 pieces for pre-training and validation with a ground truth of X-Ray Fluorescence (XRF), and the second containing 210 pieces for testing for unknown compositions. The proposed method comprises a denoising system combined with a method that extracts 145 features from the raw LIBS spectra. Further, three ML algorithms are assessed to identify the best-performing one to classify unknown pieces of aluminum post-consumer scrap into three commercially interesting output classes. The classified pieces are weighed, melted, and analyzed using spark analysis. Finally, to optimize the best-performing ML system, three state-of-the-art denoising and three feature extraction networks are pre-trained for learning the baseline correction and the proposed feature extraction. Transfer Learning from the six pre-trained networks is applied to create and evaluate 24 end-to-end DL models to classify Al in real-time from >200 spectra simultaneously. The end-to-end DL scheme shows the advantages of learning and denoising the spectra, allowing the transfer of traditional spectral analysis knowledge and the proposed feature extraction into DL, where the network learns from the entire spectrum. The best results for ML and DL were obtained with Random Forest processing one spectrum in 150 ms and BPNN+GHOSTNET(Fine-tuning) processing 200 spectra in 9 ms, which achieved 0.80 Precision, 0.81 Recall, 0.80 F1-score, and 0.80 Precision, 0.79 Recall, 0.79 F1-score, respectively.