Simultaneous prediction of bauxite quality parameters using TC-Unet and near-infrared spectroscopy
摘要
Near-infrared (NIR) spectroscopy, renowned for its rapid and non-destructive analytical capabilities, faces substantial hurdles in bauxite quality assessment due to the material’s complex composition and inherent spectral noise. These factors lead to overlapping absorption bands and highly nonlinear relationships between spectra and quality parameters. Existing chemometric methods, such as partial least squares regression (PLSR), fail to model such nonlinearity, while other conventional machine learning methods encounter difficulties in spectral feature extraction. This paper introduces TC-Unet, a novel deep learning framework for simultaneous prediction of multiple parameters, which successfully addresses these limitations through the technical integration of self-attention mechanisms, multi-scale feature extraction, and multi-task learning. The framework is specifically equipped with a transformer-based self-attention module that establishes long-range dependencies between spectral bands, a U-Net architecture that captures local absorption peak structures through encoder-decoder pathways with skip connections, and a multi-task decoupled head that ensures parameter independence leveraging inter-task correlations. Additionally, a channel-squeeze-and-excitation network (cSENet) dynamically adjusts spectral feature importance, optimizing shared representations for multi-target regression. The preprocessing pipeline, which includes an iterative Mahalanobis distance method for robust outlier detection and a standard normal variate (SNV) algorithm for scatter correction, significantly enhances the quality of spectral data. Evaluated on 424 NIR spectra, the TC-Unet outperformed classical machine learning and other deep learning methods, achieving average