CO is a highly toxic gas that causes worldwide accidental deaths. So there is an extensive demand to achieve real-time and highly sensitive detection of CO for many applications. Photoacoustic spectroscopy has become a popular optical method for trace CO detection due to its advantages of zero-background detection, high sensitivity, fast response, and wide dynamic range. Hence, a CO gas sensor based on photoacoustic spectroscopy technology is demonstrated in this paper. According to the principle, the detection limit of the gas sensors based on photoacoustic spectroscopy can be improved by selecting a stronger absorption line, increasing the optical power, optimizing the structure of a photoacoustic cell, and choosing an acoustic detector with higher sensitivity. Firstly, according to the absorption spectrum of CO gas, the second overtone band locates at optical fiber communication window is chosen as the detecting waveband. The excitation laser and the optical components have been well developed and show more stable performance than other wavebands. In order to compensate the drawback of the weak absorption coefficient and increase the strength of the photoacoustic signal, a commercial erbium-doped fiber amplifier can be employed to boost the optical power to ~ 300 mW. Secondly, considering that the flow noise, the ambient noise, the background noise generated by cell window and cell wall absorption and the electronic device noise existing in the photoacoustic spectroscopy system seriously reduce the system performance. A differential photoacoustic cell is designed and employed, which can effectively strengthen the photoacoustic signal and suppress the noise. Finally, the data processing module of the system has been optimized to obtain the optimal performance. Additionally, a data processing algorithm can be used to increase the signal-to-noise ratio and improve the system performance without increasing the complexity of hardware system. So far, several existing denoising algorithms for photoacoustic spectroscopy are suitable for dealing with linear and non-stationary signals or nonlinear and stationary signals, but the photoacoustic signal is a nonlinear signal due to the gas absorption is a non-linear and non-stationary process with various noise. Hence, the performance of these algorithms is limited when dealing with the photoacoustic signal. On the contrary, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is a well-known algorithm for non-stationary and non-linear signal, which decomposes the signal according to the time scale characteristics of the data. It is regarded as an intuitive and adaptive signal processing method and holds the ability of making up for the inadequacies of other denoising algorithms. The CEEMDAN algorithm decomposed the second harmonic signal into a finite number of intrinsic mode function components and residual component, and then Savitxky-Golay filter was used to denoise each component. In fact, the key feature of the Savitxky-Golay filter is that it can keep the shape and width of the signal unchanged, and it is widely used in spectrum analysis. Finally, the effective components used to reconstruct the signal are selected by evaluating the cross-correlation coefficient between the denoised components and the original signal. The experimental results show that when the integration time is 100 ms at atmospheric pressure and room temperature, the signal-to-noise ratio of CO detection is increased to 4.6 times of that of the original signal, and the minimum detection level is reduced to 2.6×10^-6 by algorithm. The sensor has a good linear response to gas concentration. The experimental results verify the feasibility and effectiveness of this algorithm in improving the detection performance of photoacoustic spectroscopy system.