基于最小协方差行列式分布对齐的模型转移的研究

Research on Model Transfer Based on Minimum Covariance Determinant Distribution Alignment

摘要: 针对近红外光谱中无监督模型转移方法在目标域存在异常样本时预测性能不佳、鲁棒性不足的问题，本文提出一种基于最小协方差行列式分布对齐的无监督模型转移方法(MCD-uDAR)。该方法通过MCD实现对源域与目标域预测值的稳健分布参数估计，自动识别并抑制异常样本的干扰，完成无监督条件下的跨域分布对齐与模型转移。仿真实验结果表明，无异常样本时，该方法可实现与源域理想模型相当的预测精度；在10%~40%比例的分布漂移异常样本干扰下，其预测性能显著优于uDOP、di-PLS经典方法，为复杂场景下的无监督模型转移提供了可靠方案。

Abstract: To address the issues of unstable prediction performance and insufficient robustness in unsupervised model transfer methods for near-infrared spectroscopy when anomalous samples exist in the target domain, this paper proposes an unsupervised model transfer method based on minimum covariance determinant alignment (MCD-uDAR). This method employs MCD to achieve robust distribution parameter estimation for predictions in both source and target domains, automatically identifying and suppressing interference from anomalous samples to accomplish cross-domain distribution alignment and model transfer under unsupervised conditions. Simulation results demonstrate that in the absence of anomalous samples, this method achieves prediction accuracy comparable to an ideal source domain model. Under interference from distribution-drifted anomalous samples at rates of 10% to 40%, its predictive performance significantly outperforms classical methods such as uDOP and di-PLS, providing a reliable solution for unsupervised model transfer in complex scenarios.