Target-Free Multi-Source Domain Adaptation with Data-Augmented Triplet-Aware Learning for Coal Moisture Prediction

Portable near-infrared (NIR) spectroscopy devices offer the advantages of rapid, non-destructive, and versatile coal quality analysis. However, in complex mining environments, variations in the probe–sample distance can cause significant spectral distortions, resulting in severe distribution shifts between the source and target domains and thus limiting model generalization. In practical industrial scenarios, target-domain data are often unavailable, making conventional domain adaptation methods that rely on target samples difficult to apply. To address this challenge, this paper proposes a target-free multi-source domain adaptation framework tailored for portable device distance-shift scenarios to achieve robust prediction of coal air-dried moisture (Mad). Under a multi-source joint learning strategy, the framework aligns cross-domain features through adversarial training and distribution matching, while a spectroscopy-specific data augmentation strategy is designed to simulate realistic measurement disturbances such as noise perturbation, baseline drift, and wavelength shift, thereby enhancing the model’s robustness from the source side. In addition, a Mad-aware triplet loss function is introduced to establish a balanced constraint between task consistency and domain invariance, effectively improving cross-domain generalization capability. Experimental results on multi-distance NIR datasets show that the proposed method significantly outperforms representative comparison algorithms in terms of R2, RMSE, and MAE, verifying that the framework effectively mitigates the effects of probe–sample distance shifts under target-free conditions and achieves high-precision coal moisture prediction.