融合多尺度注意力与领域对抗学习的铝土矿近红外光谱跨设备建模OA

Near-infrared(NIR)spectroscopy,owing to its advantages of rapid detection,non-de-structiveness,and reagent-free analysis,holds great potential for mineral composition analysis.However,in multi-device industrial scenarios,discrepancies in light source intensity,detector sen-sitivity,optical configuration,and sampling distance among different spectrometers lead to signifi-cant distribution shifts in spectral data of the same sample across devices.Consequently,the predic-tive performance of quantitative models trained on one device deteriorates markedly when deployed on another.Traditional chemometric calibration transfer methods(e.g.,direct standardization,piece-wise direct standardization,and slope/bias correction)rely on linear mapping assumptions,making them inadequate for complex nonlinear domain shifts.Moreover,they require repeated measure-ments of standard samples,thereby increasing application costs.To address these issues,this paper proposes a cross-device modeling approach that integrates multi-scale attention mechanisms with do-main-adversarial learning.In the feature extraction stage,a one-dimensional encoder-decoder net-work is constructed by combining convolutional block attention modules with multi-scale feature fu-sion,enabling simultaneous capture of global trends and local spectral details while suppressing noise.In terms of transfer strategy,domain-adversarial learning is introduced,where adversarial training with a gradient reversal layer and a domain classifier achieves end-to-end alignment of fea-ture distributions between source and target devices.Additionally,standard normal variate transfor-mation and Savitzky-Golay convolution smoothing are applied to enhance spectral consistency and sig-nal-to-noise ratio at the input level.On a dataset of 1 330 bauxite spectra collected using two porta-ble NIR spectrometers,eight-fold cross-validation experiments demonstrate that the proposed method achieves a coefficient of determination(R²)of 0.860 3 and a root mean square error(RMSE)of 1.752 1 on the target device,significantly outperforming traditional calibration transfer methods and several deep learning baselines.Feature distribution visualization and ablation studies further validate the ef-fectiveness of multi-scale feature fusion,attention mechanisms,and domain-adversarial strategies in feature alignment and performance improvement.