Leveraging deep learning for optimal methane gas detection: Residual network filter assisted direct absorption spectroscopy

Abstract

Noise interference in gas concentration detection has always been an intractable problem. A novel residual network filter based on deep learning algorithm in this study is proposed to compensate the shortcomings of direct absorption spectroscopy under noisy conditions. Benefiting from the ability of extracting spectral features, the residual network filter established a mapping relationship between the input noisy spectra and the denoised spectra. The signal-to-noise ratio of the residual network filter reached 33.90 dB in the simulation test and its generalization ability was further verified in the experimental test, both of which are superior to widely used filtering algorithms. In the real-time measurement evaluation, the coefficient of determination and standard deviation of the residual network filter-assisted sensor in concentration inversion and stability testing were 0.9999 and 5.82 ppm, respectively. A series of experiments show that this research has high detection accuracy and stability within the full concentration range, which greatly suppresses the noise interference of traditional gas detection. The signal can be effectively extracted even under extremely low signal-to-noise ratio conditions, demonstrating the feasibility of improvement of traditional gas sensors through the proposed deep learning model.