Autoregressive model based feature extraction method for time shifted chromatography data

This study aims to demonstrate the distinct advantages of a novel feature extraction method based on autoregressive (AR) model for the classification of the chromatography data with time shifts. Two typical classification methods, principal component regression (PCR) and neural networks (NN), were employed to compare the classification effects of three types of feature spaces: the chromatogram AR coefficients obtained through an AR model, the chromatogram wavelet coefficients, and the original chromatography data. The results indicate: (1) for the normal non-time shifted chromatography data, three types of feature spaces show almost equally good classification results; (2) for the time shifted chromatography data, the classification effect based on the AR coefficients is significantly better than those based on the other two feature spaces, especially when an NN model is employed for the nonlinearity in the classification system; and (3) without time alignment, the damage caused by the time shifts to the classification based on the chromatogram wavelet coefficients and the original chromatography data is not easily overcome even by employing an NN model. This study not only demonstrates the distinct ability of this proposed feature extraction method to free us from time alignment for the chromatography data with time shifts, but also illustrates the robustness of this feature extraction method in terms of the AR model order selection. The AR model based feature extraction method provides a novel, fast, and reliable strategy for chromatogram characterization and classification, with a great potential to benefit the development of automated instrumentation systems.