Geometric Feature Mining for the Residual Oxygen Measurement of Pharmaceutical Glass Vials on a Filling Production Line

Oxygen invasion into pharmaceutical glass vials directly affects the quality of medicine, and noncontact inversion of the residual oxygen concentration can effectively determine whether the encapsulated medicine is qualified. Different from studies in an airtight chamber, it is impractical to build a linear inversion model of an open-path optical environment, but this scenario is inevitable for the online residual oxygen measurement of glass-bottled sterile preparations. In this article, the novel concept of a harmonic characteristic triangle (HCT) is proposed. Then, four groups of available geometric features are discovered in unstable second harmonic signals, and the optimal features are selected after a quantitative evaluation of their discriminative abilities. Based on the HCT concept, 2-D, 3-D, and 4-D nonlinear models are constructed for residual oxygen concentration inversion. Experimental results prove that the proposed multidimensional nonlinear models achieved better oxygen inversion effects, with at least a 20% accuracy increase over traditional models.