Rapid Recognition and Concentration Prediction of Gas Mixtures Based on SMLP

Gas mixtures analysis and concentration measurement is of great significance in the fields of environmental protection, industrial safety, life sciences, and so on, and the nondispersive infrared spectroscopy (NDIR) technique has emerged as a critical measurement method with its high accuracy, selectivity, stability, and fast response. Building upon a multipass optical NDIR measurement system, this article introduces a stepwise multilayer perceptron (SMLP) neural network algorithm for the rapid detection of gas mixtures. This algorithm addresses the limitations of traditional metrology methods that necessitate manual feature selection and extensive data modeling. Through the iterative application of the multilayer perceptron (MLP), this article first classifies gas mixtures and then selectively regresses them against known gas types, enabling the automatic extraction of effective features and simplification of the network structure. The results show that the multipass optical measurement data provide more environmental information, achieving higher classification accuracy and smaller regression root mean square error (RMSE). In addition, achieving the same accuracy and RMSE with fewer training samples is possible. The classification and concentration regression for CO2 and CH4 can be achieved using only 0.5 s of data, with a classification accuracy of 98.21% and normalized RMSE estimates of 0.42% and 0.45%, respectively. In contrast to commonly employed classical machine learning algorithms, the SMLP algorithm demonstrates enhanced classification and regression capabilities, offering the potential to streamline network complexity and simplify the parameterization process.