Multi-Source Spectral Learning without target-data via Dual Regression Heads and Performance-Driven Domain Adaptive Weighting

Near-Infrared Spectroscopy (NIRS) has become one of the most widely used non-destructive techniques for rapid coal quality assessment due to its efficiency, low cost, and suitability for on-site industrial deployment. However, NIRS data collected under different measurement conditions often exhibit significant spectral distribution shifts, which substantially weaken the cross-domain generalization ability of conventional prediction models. To address this challenge, this paper proposes a multi-source spectral learning method tailored for NIRS-based coal analysis. The method integrates a unified multi-source domain adaptation framework, a dual-regression-head architecture, and a performance-driven dynamic weighting strategy, enabling effective cross-domain learning without requiring target-domain data. Experiments conducted on real coal NIRS datasets demonstrate that the proposed approach consistently outperforms existing baselines across multiple measurement scenarios, highlighting its strong potential for accurate and robust rapid coal quality detection.