A Modular, 3D-Printed, and Cost-Effective Near-Infrared Spectrometer with Machine Learning for Commodity Plastic Identification

Abstract

Efficient segregation of post-consumer plastics remains a critical challenge in achieving a sustainable circular economy because current sorting technologies are limited in cost, adaptability, and accessibility for community-level use. This study presents the development of a modular, cost-effective, and 3D-printed transmission near-infrared (NIR) spectrometer that operates in the 850 nm to 1100 nm range. The spectrometer was developed using open-source tools and resources such as Theremino and Scikit-learn, as well as commercially available optical components such as a desktop webcam, a transmission grating, and an incandescent penlight. Validation of the spectrometer revealed minor wavelength deviations between theoretical and measured values. Measured values deviated by 5.1% with respect to the 546 nm mercury characteristic peak and 7.6% with respect to the 436 nm characteristic mercury peak emitted from a fluorescent lamp. This deviation is due to the low-cost optical components, but it did not impair the classification performance of the polymers. 504 plastic samples from seven material classes namely acrylonitrile butadiene styrene (ABS), low-density polyethylene (LDPE), polyethylene terephthalate (PET), polymethyl methacrylate (PMMA), polypropylene (PP), green polystyrene (GPS), and red polystyrene (RPS), together with an empty baseline reference (MT) sample were collected and processed using Principal Component Analysis (PCA), where 12 principal components were retained to capture 95% of the cumulative spectral variance. Classification performance was subsequently evaluated using Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and K-Nearest Neighbors (KNN) models. After hyperparameter tuning, both the SVM and KNN models achieved the highest held-out test accuracy (82.24%). SVM was the most reliable overall classifier, with slightly better class-balanced performance (F1-score of 82.36%) than KNN (81.98%) and LDA (73.71%). The empty baseline reference spectrum was classified perfectly with a precision, recall, and F1-score of 100%, confirming the consistency of the baseline penlight source spectrum. Materials like RPS showed greater feature overlap with GPS, and LDPE with PP. This study demonstrated that a low-cost, transmission NIR spectrometer operating between 850 nm and 1100 nm, combined with machine learning, can accurately identify commodity plastics, offering a promising community-level solution.

ABSTRAK: Pengasingan plastik pasca guna secara berkesan masih merupakan cabaran utama dalam mencapai ekonomi kitaran lestari. Kekurangan teknologi pengasingan plastik sedia ada dari segi kos, kebolehsuaian, dan kebolehcapaian gunaan di peringkat komuniti telah menjadi penyumbang kepada cabaran tersebut. Kajian ini menumpukan kepada pembangunan spektrometer inframerah dekat (NIR) jenis transmisi melalui pencetakan 3D bersifat modular, kos efektif, dan mempunyai julat operasi dari 850 nm hingga 1100 nm. Spektrometer ini dibina menggunakan platform sumber terbuka seperti Theremino dan Scikit-learn. Selain itu, komponen optik yang mudah diperoleh seperti kamera web desktop yang diubah suai, kisi difraksi transmisi, dan lampu pijar jenis pen turut digunakan. Semasa proses pengesahan fungsi spektrometer ini, sisihan panjang gelombang kecil antara nilai teori dan nilai yang diukur telah dikesan. Berdasarkan dapatan pengukuran, terdapat sisihan sebanyak 5.1% bagi puncak ciri merkuri pada panjang gelombang 546 nm dan 7.6% bagi puncak ciri merkuri pada panjang gelombang 436 nm yang dipancarkan oleh lampu pendarfluor. Penggunaan komponen optik kos rendah mengakibatkan berlakunya sisihan tersebut, tetapi keupayaan spektrometer ini dalam pengelasan polimer tidak terjejas. Sebanyak 504 sampel plastik daripada tujuh kelas bahan iaitu akrilonitril butadiena stirena (ABS), polietilena berketumpatan rendah (LDPE), polietilena tereftalat (PET), polimetil metakrilat (PMMA), polipropilena (PP), polistirena hijau (GPS), dan polistirena merah (RPS), bersama satu garis dasar rujukan kosong (MT) telah dikumpulkan dan diproses menggunakan Analisis Komponen Utama (PCA), di mana 12 komponen utama dikekalkan bagi mewakili 95% varians spektrum kumulatif. Prestasi pengelasan kemudiannya dinilai menggunakan model Mesin Vektor Sokongan (SVM), Analisis Diskriminan Linear (LDA), dan K-Jiran Terdekat (KNN). Selepas penalaan hiperparameter, model SVM dan KNN mencapai ketepatan ujian ketahanan tertinggi sebanyak 82.24%. SVM dikenal pasti sebagai pengelas keseluruhan yang paling boleh percaya dengan skor F1 sebanyak 82.36%, berbanding KNN (81.98%) dan LDA (73.71%). Spektrum sampel MT berjaya diklasifikasi dengan sempurna dengan nilai ketepatan, penarikan semula, dan skor F1 sebanyak 100%, sekaligus mengesahkan konsistensi spektrum asas yang dihasilkan daripada cahaya lampu pijar. Namun, bahan seperti RPS menunjukkan pertindihan ciri spektrum yang lebih tinggi dengan GPS, manakala spektrum LDPE pula bertindih dengan spektrum PP. Kajian ini berjaya membuktikan bahawa spektrometer NIR transmisi kos rendah yang dihasilkan melalui pencetakan 3D dan pembelajaran mesin dapat beroperasi pada julat 850 nm hingga 1100 nm dan mampu mengenal pasti plastik komoditi dengan tepat di samping menawarkan penyelesaian terbaik pada peringkat komuniti.