Interference Fringe Suppression for Oxygen Concentration Measurement Using Adaptive Harmonic Feeding Generative Adversarial Network

This paper proposes an efficient interference fringe suppression method for the oxygen concentration measurement system by adopting emerging machine learning techniques. First, the interfered and interference-free signal datasets are generated on HITRAN molecular spectroscopic database after a transmission factor is considered in the wavelength-modulation-based TDLAS (TDLAS/WMS) theory. Then, an adaptive harmonic feeding generative adversarial network (AHFGAN) is developed to deal with the task of interference fringe suppression, where a novel adaptive weighted scheme is proposed to guide the weight learning process based on the data prior knowledge of dispersion degree refined from a large number of harmonic signals. Based on the AHFGAN, nearly perfect interference-free harmonic signals are directly learnt from the real-world TDLAS system, with an average absolute oxygen concentration inversion error of 0.57% when applied in an actual pharmaceutical production line, which performs better than other five recent state-of-the-arts.