China is not only a large agricultural country, but also one of the countries suffering from the most serious agricultural disasters in the world. Early fire detection can help to avoid greater losses. Crops have a period of smoldering before the open flame. Due to lack of oxygen, the combustion is extremely insufficient, and a small amount of carbon monoxide (CO) with a relatively stable concentration will be produced. Therefore, the occurrence of early fire can be judged by detecting gas-phase CO. Non-Dispersive Infrared (NDIR) absorption spectroscopy was used in this paper. Based on the absorption band of CO gas molecules at 4.6 μm, a differential infrared CO sensor system for early fire detection was developed by using a broadband infrared thermal light source and a dual-channel pyroelectric detector. This sensor system is mainly composed of a gas pretreatment part, an optical part, an electrical part and an upper computer monitoring part. Firstly, the detection principle of the sensor system was introduced, and then the absorption band of CO in the infrared region was selected by comparing the strength of absorption lines in different absorption bands and excluding the absorption interference of other gas molecules. Through the derivation of the optical matrix, the structure of the gas cell was designed and optimized, and the optical path of gas absorption reached 180 cm . The background noise fluctuation range of the phase-locked amplifier is 38.89 μV~43.23 μV, and the lowest detection limit is 0.15 μV. Finally, the performance of the sensor system was tested through related experiments. The results show that the measurement resolution is less than 2×10^-5 and the response time is 35~38 s. The concentration level of the 0×10^-6 CO standard gas sample was dynamically monitored for 80 minutes, and its concentration fluctuation range is -1.42×10^-5~1.51×10^-5. When the integration time is 0.25 s, the detection limit of the system is 1.54 ×10^-6, and when the integration time is 300 s, the detection limit of the system can reach 3.50×10^-7. Kalman filtering algorithm was used to improve the stability of the system. Similarly, the 0×10^-6 CO standard gas sample was dynamically monitored for 80 minutes. The results show that the relative error is reduced by 40.56 %, and when the integration time is 0.25 s, the detection limit is reduced to 3.60×10^-7. Finally, fire smoldering experiments of cotton, paper and wood were carried out to study the relationship between CO concentration and smoldering time. It is proved that the change of CO concentration can be used to detect the occurrence of fire. The experimental results show that the CO sensor system has a good early fire detection capability and wide application prospects.