Methane concentration inversion under multiple conditions using feature fusion residual network

Abstract

In the realm of methane concentration measurement using tunable diode laser absorption spectroscopy (TDLAS), relying on a single second harmonic signal peak often leads to information loss. Furthermore, second harmonic peaks exhibit conflicts when subjected to different parameters, thereby posing challenges for accurate methane concentration measurement. This paper presents a novel approach employing a multi-scale feature fusion residual network (MResNet) to measure methane concentration under various operational conditions. To begin, we established a hardware experimental platform based on TDLAS to collect direct absorption spectrum data, and second harmonic data with varying parameters. Then we conducted a comparative analysis between MResNet and conventional models. The results of our experiments demonstrate the effectiveness of MResNet in predicting methane concentration, with a decrease in MAE by 39% and MSE by 70%. Notably, the coefficient of determination for methane concentration prediction using MResNet both reached 99.92% when applied to both direct absorption spectrum and second harmonic data. This capability empowers accurate methane concentration measurement across a wide range of operating conditions.