With the acceleration of industrialization and urbanization, the global generation of waste glass continues to rise. Due to its stable chemical properties and resistance to natural degradation, glass often persists in the environment for extended periods if not properly recycled. Particularly, some specialized glass products often contain heavy metal elements such as lead (Pb) or barium (Ba), which pose serious threats to the ecological environment and human health. Meanwhile, the recycling rate of glass is low in many countries, partly due to the inefficiency of the sorting process. Traditional sorting methods, such as manual sorting or color-based classification, often fail to distinguish glass samples with similar appearance but different chemical compositions. To address these issues, this study proposes a LIBS-LIPA-ML system that integrates laser-induced breakdown spectroscopy (LIBS), laser-induced plasma acoustic (LIPA) signals, and machine learning (ML) to achieve real-time, in-situ detection and classification of waste glass. This method aims to enhance classification performance, accurately identify trace hazardous elements, and promote intelligent large-scale glass recycling.

Four types of typical commercial glass were selected, including quartz glass, soda-lime glass, borosilicate glass, and lead-barium glass. These samples differed in their elemental composition, and the lead-barium glass in particular contained heavy metal elements that posed potential environmental risks. The system employed a high-power neodymium-doped yttrium aluminum garnet (Nd:YAG) laser as the excitation source. The laser was focused onto the surface of the glass sample fixed on a platform. The resulting plasma was collected through an optical fiber and transmitted to a four-channel LIBS spectrometer for analysis. Simultaneously, the plasma acoustic signals were acquired by an electret condenser microphone, displayed in real-time on a digital oscilloscope, and digitized. The raw spectral data were processed using a screening algorithm based on the intensity of the sodium emission line to ensure data reliability. Principal component analysis (PCA) was then employed to reduce the dimensionality of the normalized high-dimensional spectral data, extract dominant features, and retain the most significant variance. The first three principal components were selected as input features for training a back-propagation neural network (BPNN) model. The number of hidden neurons was optimized via grid search to achieve an effective balance between learning complexity and generalization capability. The model performance was evaluated using ten-fold cross-validation to reduce the risk of overfitting. To further improve classification robustness, the study adopted feature-level fusion, concatenating the principal components of the LIBS data and LIPA signals into a unified feature matrix that served as the new input to the model.

The LIBS spectral analysis revealed distinct elemental characteristics for each glass type. Multiple characteristic spectral lines of lead and barium were clearly observed in the lead-barium glass spectra (Fig.2), highlighting the high sensitivity of the system for heavy metal detection. Soda-lime glass exhibited prominent spectral line intensities for calcium and magnesium (Fig.3c), while borosilicate glass displayed unique boron spectral lines (Fig.3b). In contrast, all characteristic spectral lines of quartz glass were generally weaker (Fig.3a). PCA reduced the dimensionality of the high-dimensional normalized spectral data, and the first three principal components collectively explained over 98% of the variance (Fig.4). The score plot derived from the principal components (Fig.5) showed well-separated clusters among the glass types, visually confirming the distinct separability of the four glass samples. When using only LIBS features for classification, the BPNN model with the optimal number of hidden neurons determined by grid search achieved an average classification accuracy of 86.88% under ten-fold cross-validation (Fig.6). However, due to the similar composition of quartz glass and borosilicate glass, the confusion matrix (Fig.9a) showed occasional misclassifications between them. To improve the robustness and classification accuracy of the model, LIPA signals were further incorporated. Features extracted from the first part of the acoustic waveform (Fig.7) were combined with the LIBS spectral features to form a new dataset (Fig.8). The BPNN model trained on this fused-feature dataset achieved a classification accuracy of 98.12%, demonstrating a significant improvement. As observed from the new confusion matrix (Fig.9b), misclassification was nearly eliminated. As listed in Table 2, the precision, recall, and F₁-score for all glass types were significantly improved. The combined use of LIBS and LIPA features enabled the system to extract more comprehensive information from the samples, overcoming the limitations of relying on a single data source.

This study demonstrates the feasibility and effectiveness of the LIBS-LIPA-ML system for the online and in-situ classification of waste glass. The system is capable of detecting both major and trace elements, including heavy metal elements such as Pb and Ba, and can also capture material acoustic features, thereby enhancing classification robustness. Through feature extraction via PCA and modeling with BPNN, the system maintains good adaptability while ensuring efficiency. More importantly, the multimodal fusion strategy significantly improves the classification performance, with an accuracy exceeding 98%. By integrating laser diagnostic technology and machine learning, the system offers a scalable, non-contact, and intelligent solution for industrial glass recycling. It supports the early-stage detection of hazardous substances and the precise sorting of complex waste streams. This approach contributes to sustainable waste management and opens new possibilities for real-time material classification in the environmental and manufacturing sectors.