Harmonic Amplitude Dispersion: When Production Intrinsic Prior Meets Oxygen Concentration Detection of Pharmaceutical Glass Vials

Abstract

In situ and accurate oxygen concentration detection is drawing great attention in the pharmaceutical manufacturing industry. This article proposes a novel concept of harmonic amplitude dispersion (HAD) under the framework of wavelength-modulation-based tunable diode laser absorption spectroscopy (TDLAS/WMS). Then, a reliable production intrinsic prior to the long-term information entropy distribution pattern is learned from the demodulated second-harmonic signals. Interestingly, the performance in oxygen concentration detection can be improved in two rounds by the HAD concept. First, for accuracy, signal reconstruction is implicitly carried out during the HAD-based parameter optimization process for the horizontal span ( $k_{x}$ ) and the vertical tension ( $k_{y}$ ) in the Voigt spectral line modeling. This method is denoted as HAD signal reconstruction (HADSR). Second, for efficiency, with the guidance of HAD, only approximately 12% of reconstructed signals are fed into the online sequential extreme learning machine (OS-ELM) for further concentration classification, as the data with higher HADs are determined to be uninformative and are discarded early before data training to enable the classifier to be more compact. The detection framework of HADSR + OS-ELM was first verified on an automated vision inspection machine in an actual pharmaceutical manufacturing line. Promisingly, the proposed HAD robustly narrowed the average variance in the signal peak-to-peak values by nearly 30%, and together with the HAD-guided OS-ELM, an average detection accuracy of 98.1% within an acceptable time cost of 300 ms/vial was realized during a long-term test.