本文针对化工过程中在线检测丙烯的需求，研究了基于调制吸收光谱技术（TDLAS）的检测技术，提出了一种独立于光谱线型特征的数值仿真方法，考虑实际激光光源宽线宽对吸光度的影响，通过对比仿真和实验的光谱幅度变化规律，确定了丙烯气体分析装置的设计参数和技术方案，选择中心波长为1 628.5 nm的宽调谐DFB激光器，采用差分方案去除解调光谱的直流偏置，采用多元回归模型降低化工过程的背景气体光谱干扰。在模拟实际环境的气体实验中，该装置在0~1% 量程内的最大相对误差为0.55%。对0.2% 的丙烯进行3小时连续测量，标准差为9.3×10^?6；Allen方差分析发现在积分时间为221.9 s 时，极限标准差可达1.33×10^?6。在抗干扰测试中，当背景气体甲烷、乙烯的浓度变化时，丙烯的测量误差最大仅为19.17×10^?6。调制吸收光谱技术克服了色谱和软测量等传统方法的不足，TDLAS装置可检测有复杂光谱特征的重烃分子，展示了测量精度高、稳定性好、抗背景光谱干扰能力强等优点。

To satisfy the need for propylene measurement in the olefin production process, Tunable Diode Laser Absorption Spectroscopy (TDLAS) was studied to improve analytical performance. In this paper, a numerical simulation approach is proposed using absorbance from a spectral database to obtain the optimized design parameters, which is independent of spectral features. In the simulation, the effect of a wider linewidth laser on the absorbance profile was considered. Through the comparison of simulation results and experimental collection, the TDLAS-based propylene analysis apparatus was developed correspondingly. It has a 1 628.5 nm center wavelength broad-tuning DFB laser. A differential method was utilized in demodulated spectral acquisition to eliminate bias voltage. The multivariate linear regression model was employed to reduce the strong spectral interference from the background components in the analysis. Based on the simulated field test, the max relative error is 0.55% in the 0~1% range for the step test. For the long-term test, the standard deviation (1σ) is 9.3×10^-6 for 0.2% propylene concentration. The best standard deviation is 1.33×10^-6 at 221.9 s of integration time through Allen variance analysis. In the anti-interference test, the max error of 19.17×10^-6 is demonstrated for 0.2% propylene concentration while methane and ethylene concentrations vary. The disadvantages of traditional methods such as the Gas Chromatogram (GC) and soft measurement methods are overcome by modulated absorption spectroscopy. The TDLAS system for heavy hydrocarbon detection with complex spectral features was demonstrated to have distinct advantages in precision, stability and interference suppression through multivariate regression modeling.