Local deep learning of principal component regression model for spectroscopic calibration of time-varying spectra data

Abstract

The motivation of this paper is to improve the calibration performance of the widely used principal component regression model (PCR), aiming at time-varying spectra data. Inspired by the idea of layer-by-layer information extraction and processing in deep learning algorithms, the basic PCR model is firstly extended to the deep form by designing a layer-wise residual learning strategy. Then, a local deep learning framework is further formulated for calibration modeling of time-varying spectra data. Compared to traditional deep learning models, the lightweight deep PCR model has much lower computation burden and only several turning parameters, making it perfectly fitted to the local modeling framework. Besides, the engineering applicability of the simple PCR method can also be well reserved, such as easy implementation, transparent model structure, and superior interpretability of calibration results. Based on a detailed simulation case study on a public spectra dataset, both feasibility and effectiveness of the local deep learning model are confirmed.