Aiming at the problem of low accuracy of non-dispersive infrared carbon dioxide(CO

2

) sensor affected by optical path length and temperature

a long-optical-path non-dispersive infrared CO

2

sensor is designed

and a new method of compensating the gas chamber temperature using a domain adaptation generative adversarial network(DAGAN) is proposed.Firstly

the gas absorption chamber is designed with a combination of four spherical mirrors

where the spher

ical mirrors are placed in pairs facing each other with their centers coinciding

forming a conjugated system to help more infrared light reach the detector.Secondly

the relationships between the incident angle of the infrared light source

the radius of the spherical mirror

and the number of reflections of the light rays are analyzed

as well as the relationship between the additional optical path and the different numbers of reflections and incident angles.Lastly

the impact of temperature on CO

2

is compensated based on DAGAN.Experiments show that as the radius of the spherical mirror

the arc length of the spherical mirror

and the distance between the spherical mirrors increase

the CO

2

gas absorption rate increases and then begins to decrease after reaching an extreme value.DAGAN makes the distributions of data in the source domain and target domain closer in the feature space

resulting in a smaller absolute error of the compensation results.Compared with back propagation neural network(BPNN)

improved BPNN(IBPNN)

radial basis function neural network(RBFNN)

generative adversarial network(GAN)

and continuous conditional GAN(CCGAN) algorithms

the proposed method reduces the absolute error by 0.014%

0.012%

0.009%

0.008%

and 0.004%

respectively.