<p>Near-infrared (NIR) spectroscopy is a powerful tool for rapid, non-destructive analysis of materials, but multivariate calibration models often lose accuracy when applied to new conditions due to domain shifts from instrumental differences, environmental factors, or sample matrix variations. Conventional methods for calibration transfer or updating, such as spectral standardization or full recalibration, typically demand substantial reference data, transfer standards, or high-dimensional computations, which can be impractical in resource-limited settings. Building on a descriptive geometric framework for the multivariate calibration model vector b-derived through orthogonal projection to eliminate structured non-analyte interference and least-squares scaling to ensure unit analyte sensitivity-this study introduces two efficient vector-based strategies for calibration updating using only a small number of target-domain samples. The first strategy forms an updated vector as a convex linear combination of source and target descriptive vectors, while the second employs weighted augmentation of source and target data before computing the descriptive vector. In both approaches, the mixing parameter is optimized by minimizing prediction error on a target-domain validation set, preserving the geometric properties of the model (orthogonality to interferents and analyte selectivity) while adapting to target-specific variations without latent variables or complex transformations. The proposed methods were tested on the standard NIR corn dataset (78 samples measured on two instruments, with analytes moisture, oil, protein, and starch), utilizing 54 source calibration samples, 24 source validation samples, and only 10 labeled target samples for updating. Results show marked reductions in prediction errors on held-out target samples, often approaching source-domain performance, with acceptable trade-offs in calibration error due to balanced integration of target variability.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Efficient calibration updating using descriptive model vectors for domain adaptation in multivariate spectroscopy

  • Somaye Vali Zade

摘要

Near-infrared (NIR) spectroscopy is a powerful tool for rapid, non-destructive analysis of materials, but multivariate calibration models often lose accuracy when applied to new conditions due to domain shifts from instrumental differences, environmental factors, or sample matrix variations. Conventional methods for calibration transfer or updating, such as spectral standardization or full recalibration, typically demand substantial reference data, transfer standards, or high-dimensional computations, which can be impractical in resource-limited settings. Building on a descriptive geometric framework for the multivariate calibration model vector b-derived through orthogonal projection to eliminate structured non-analyte interference and least-squares scaling to ensure unit analyte sensitivity-this study introduces two efficient vector-based strategies for calibration updating using only a small number of target-domain samples. The first strategy forms an updated vector as a convex linear combination of source and target descriptive vectors, while the second employs weighted augmentation of source and target data before computing the descriptive vector. In both approaches, the mixing parameter is optimized by minimizing prediction error on a target-domain validation set, preserving the geometric properties of the model (orthogonality to interferents and analyte selectivity) while adapting to target-specific variations without latent variables or complex transformations. The proposed methods were tested on the standard NIR corn dataset (78 samples measured on two instruments, with analytes moisture, oil, protein, and starch), utilizing 54 source calibration samples, 24 source validation samples, and only 10 labeled target samples for updating. Results show marked reductions in prediction errors on held-out target samples, often approaching source-domain performance, with acceptable trade-offs in calibration error due to balanced integration of target variability.