Explainable convolutional transformer for contaminant classification in recycled coarse aggregates using single-shot laser-induced breakdown spectroscopy

Fast in-line identification of contaminants in recycled coarse aggregates (RCA) is essential for quality control of the processing of RCA. However, robust classification from single-shot spectra remains challenging under practical operating conditions. Laser-induced breakdown spectroscopy (LIBS) offers potential for fast, non-contact material identification, but direct recognition from raw single-shot LIBS spectra is difficult because the signals are high-dimensional and affected by shot-to-shot variability. This challenge becomes more pronounced in multi-material settings, where spectral overlap and local heterogeneity make contaminants less distinguishable across classes. To address these challenges, an explainable spectral-attention network is proposed for single-shot LIBS spectral classification based on an Explainable Convolutional Transformer Network (ECT-Net). The model integrates a convolutional layer for local spectral feature extraction directly from standard normal variate (SNV)-preprocessed spectra and sequence downsampling, a Transformer encoder for modelling longer-range dependencies across spectral regions, and a global average pooling head for compact multi-class prediction. The framework was evaluated on a balanced ten-class dataset comprising 24,000 single-shot LIBS spectra acquired experimentally under realistic conveyor-belt conditions with RCA and contaminants commonly existing in construction and demolition waste recycling. The proposed lightweight Explainable Convolutional Transformer Network (ECT-Lite) achieved a test macro-averaged F1 score of 98.59%, with a corresponding test accuracy of 98.63%. Attention visualization confirmed that the model focused on physically meaningful wavelength regions. These results demonstrate the effectiveness of the proposed ECT-Lite for SNV-preprocessed single-shot LIBS spectra in recycling-oriented multi-material recognition.