Terahertz Coal Ash Prediction Method Based on Dual-Channel Convolutional Neural Network

Objective Coal plays a crucial role in China's economy and energy strategy as one of the main sources of energy and an important component of energy security. The ash content has always been a challenging issue for coal preparation plants to control product quality during the coal production process. By collecting and analyzing ash content detection data and maintaining stable ash content, the quality of coal washing products can be ensured, energy utilization can be improved, carbon emissions can be reduced, and environmental protection can be promoted. In China, rapid or slow ash methods are mainly used to detect ash content. This process takes 2-3 h, resulting in long detection cycles, low efficiency, and significant delays in obtaining detection results from coal sampling to analysis. In recent years, breakthroughs and progress have been made in online ash content detection technology. Natural gamma-ray measurement-based online detection technology has poor adaptability to different coal types, while X-ray absorption-based online detection technology offers high measurement precision and accuracy but is inconvenient for management and safety production. Therefore, there is a demand for a fast, accurate, safe, and real-time monitoring method for coal ash content in industrial production. Methods Terahertz spectroscopy is an emerging spectral technique that bridges the gap between microwave and infrared spectroscopy. It encompasses the physical, structural, and chemical information of substances within its frequency range, thus meeting the practical technological requirements of the coal industry. In this study, to address the prediction of coal ash content, 46 samples were tested using a terahertz spectrometer to extract the absorption spectrum and refractive index spectrum of the coal samples. The absorption characteristics and refractive properties of different ash content samples in the terahertz frequency range were investigated. To eliminate the influence of sample thickness on the absorption coefficient, a method based on thickness model correction was proposed for extracting the absorption coefficient features. To improve the prediction accuracy and obtain different feature information, a dual-channel convolutional neural network was established to extract refractive index features and absorption features for coal ash prediction. This research provides the theoretical basis and technical support for intelligent detection in the coal mining industry. Results and Discussions First, we obtained the refractive index and absorption coefficient of coal samples and explored the correlation law between them and the increase in coal ash content in the frequency range of 0. 5-3 THz (Fig. 3). By taking into account the influence of sample thickness on the spectrum, a method based on thickness model correction for extracting the absorption coefficient features was proposed, which improved the data distinguishability of low-ash coal sample absorption curves (Fig. 4). In order to learn and predict the feature vectors of coal ash content samples, a dualchannel convolutional neural network was constructed for feature extraction, weighted fusion, and prediction of coal ash content samples (Fig. 5). The loss function of the network training process gradually decreased with the increase in iteration times, and no overfitting occurred (Fig. 6). A 10-fold cross-validation was used to evaluate the accuracy of the feature fusion network, with alpha=0. 4, and the algorithm achieved the highest prediction accuracy (Fig. 7). The fitting degree and prediction accuracy of the model training process in the training set were R-2=98. 21% and E-RMS=0. 1442, respectively, while in the prediction set, R-2=93. 56% and E-RMS=0. 2037 (Fig. 8), outperforming traditional methods such as PLSR, BP, and LSSVM (Tab. 1). Conclusions In this article, the THz time-domain spectroscopy technique was used to analyze the spectral characteristics of coal samples in the frequency range of 0. 5-3 THz. The results showed that the refractive index and absorption of the coal samples increased with the increase in ash content. By considering the differences in the absorption coefficient of coal samples with different thicknesses within the frequency range of 0. 5-3 THz, we proposed a method for extracting absorption coefficient features based on thickness correction, which could better separate and disperse the original absorption spectra of coal samples, thereby facilitating accurate prediction of different ash content. The experimental results demonstrated significant advantages of the proposed dual-channel convolutional neural network regression model in predicting coal ash content compared with traditional CNN, PLSR, BP, and LSSVM models. Compared with traditional ash content detection methods, this method can reduce the detection time by approximately 80% and greatly improve work efficiency. Additionally, it can be applied to ash content detection of different coal types to meet practical demands.