联合特征子空间分布对齐的标定迁移方法

近红外光谱分析技术近年来在各种领域的定性、定量分析等方面得到广泛的应用。多元标定技术则是光谱分析领域中最先进的技术，而环境条件、测量仪器或测量物质自身的变化，都可能导致多元标定模型不再适用于新样本的预测。重新标定和重新建模必然会浪费大量时间和资源。一种解决方案是标定迁移，将源域已有的标定模型扩展到目标域中，避免重复建模的代价。在化学计量学的相关文献中，绝大多数迁移方法都需要在两台仪器相同条件下都测量一组迁移标准样品，但在近红外光谱测量技术中，由于标准样品具有挥发等特性，使得构建仪器标定迁移方法的标准样品难以获得和保存。针对这些问题，提出了一种联合特征子空间分布对齐（JSDA）的标定迁移方法，此方法可以在从仪器没有标准样本的情况下建立标定迁移模型。JSDA首先建立源域和目标域数据特征的联合主成分分析（PCA）子空间；然后通过对齐映射在联合特征子空间中的源域特征分布和目标域特征分布来校正标定模型；最后，应用最小二乘模型构建校正后源域上的标定模型，该模型可直接用于目标域的标定。实验结果表明与已有成熟的标定迁移方法相比，JSDA在公开的真实数据集上的预测性能比较有优势，验证了该模型在实际应用中的有效性和优越性。

Research on Calibration Transfer Method Based on Joint Feature Subspace Distribution Alignment

Near-infrared spectroscopy analysis technology has the advantages of low cost, high efficiency, and pollution-free. In recent years, it has been widely used in qualitative and quantitative analysis in various fields. Multivariate calibration technology is the most advanced technology in the field of spectroscopy. Changes in conditions, instruments, or substances may cause the multivariate calibration model to no longer be suitable for the prediction purposes of newly measured samples. Re-calibration and re-modeling will inevitably waste a lot of time and resources; another option is calibration transfer, which extends the existing calibration model in the source domain to the target domain to avoid the cost of repeated modeling. In the related chemometrics literature, most transfer methods need to measure a set of transfer standard samples under the same conditions of two instruments. However, in the near-infrared spectroscopy measurement technology, due to the characteristics of volatilization of the standard samples, It is not easy to obtain and save the standard samples for constructing the transfer method for instrument calibration. This paper proposes a joint feature subspace distribution alignment (JSDA) calibration transfer method in response to these problems. This method can establish a calibration transfer model without a standard sample from the instrument. JSDA first establishes the joint PCA subspace (Principal component analysis) of the data features of the source and target domains; then corrects the calibration model by aligning the source domain feature distribution and target domain feature distribution mapped in the joint feature subspace; Finally, the least squares model is used to build a calibration model on the corrected source domain, which can be directly used for the calibration of the target domain. The experimental results show that compared with the existing mature calibration transfer methods, JSDA has more advantages in predicting performance on public real data sets, which verifies the effectiveness and superiority of the model in practical applications.