Feature selection and online discrimination for weak oxygen absorption spectrum

Real-time online discrimination of the residual oxygen concentration in pharmaceutical glass vials is drawing great attention in the pharmaceutical industry. The features of the demodulated 2nd harmonic signals in the wavelength modulation spectroscopy (WMS) supervision system, the data basis of the inversion of oxygen concentration, are inevitably destroyed by various time-varying industrial noises. In this work, we propose an improved relief method based on the minimum redundancy (MR-Relief) to select effective features from the 2nd harmonic signal to reduce the influence of fast time-varying noise on the system. Then, the selected effective features are fed into the online adaptive updating radial basis function neural network (AU-RBFNN) to further eliminate the slow time-varying noise in the WMS supervision system. Experimental results show that the proposed framework can resist various time-varying industrial noises with an average discrimination rate of 96.17% under the discrimination speed of 300 vails/min during a long-term test.