The ro-vibrational spectra of the gas molecules is located in mid-infrared waveband, and then the information of the gas type and its concentration can be detected with a high precision based on the non-dispersive infrared technology. In this paper, a SF6 gas sensor was designed with the optical structure of the double light path of a single light, by utilizing a 2～20 μm electrically modulated thermal radiation source and a dual wavelength pyroelectric detector with the central wavelengths of 3.95 μm and 10.55 μm. The method of a radial basis function neural network algorithm was proposed to compensate the detection error caused by the variation of the ambient temperature. The experimental results show that the detection accuracy of this sensor is less than ±1.5%FS within the ambient temperature range of 10℃ to 35℃ and the gas concentrations from 0 to 0.200%. The relative standard deviation is 1.56%. It can effectively eliminate the nonlinear effects caused by the environmental temperature changing in measuring the gas concentration. Compared with the traditional compensation methods with the empirical formula or the temperature control scheme, our method has a better measuring accuracy and stability. Moreover, by using this method, the gas sensor doesn't need any temperature control module, which is beneficial to miniaturize the device size and reduce its cost.