An End-to-end Deep Learning Model for Predicting Total Incombustible Contents in Coal/stone Dust

• Published in: Civil and environmental engineering reports Vol. 35; no. 3; pp. 255 - 270
• Main Authors: WANG, CHENHAO, Zeng, Shevaune, Widzyk-Capehart, Eleonora, Kennedy, Gareth
• Format: Journal Article
• Language: English
• Published: 21.07.2025
• ISSN: 2080-5187; 2450-8594; 2450-8594
• DOI: 10.59440/ceer/205778

Coal dust explosions are a significant threat to underground coal mines. To reduce this risk, the combustible coal dust is mixed with stone dust to increase the total incombustible contents (TIC). Conventional TIC measurement methods rely on time-consuming laboratory analyses, involving extensive sample preparation and chemical testing. In contrast, near-infrared (NIR) spectroscopy has emerged as a rapid, non-destructive alternative for TIC prediction. However, existing machine learning models for analysing high-dimensional spectral data often require extensive preprocessing, increasing the analysis complexity. In this study, we present a residual - Convolutional Neural Network (CNN) based method for end-to-end analysis of raw near-infrared (NIR) spectral data to reduce preprocessing requirements while accurately classifying the TIC levels. The model was evaluated using 300 coal/stone dust samples, with 100 coal samples sourced from various Australian coal mines. The deep learning model, configured with optimal nine residual blocks, demonstrated high accuracy in predicting high TIC samples (TIC ≥ 85%), achieving misclassification rates of 0.05 on the training set and 0.14 on the testing set, respectively. Two challenges were identified: class imbalance and spectral overlap. The low TIC samples (TIC < 70%) accounted for only 9% of the total dataset (27 out of 300), resulting in poor prediction for this underrepresented class. Additionally, significant spectral similarity with distinct TIC values reduced the model’s generalization ability. Despite these challenges, our study demonstrates the potential for developing a reliable and efficient end-to-end deep learning framework for TIC prediction, which would allow for a significant reduction in preprocessing efforts.