在常温常压下, 利用光电导天线式太赫兹时域光谱仪和自行设计的气室, 在0.1~3.0 THz范围内对甲醇气体进行太赫兹时域光谱测试, 测试结果表明, 甲醇气体在1.0~3.0 THz没有明显的吸收峰, 但是在0.1~1.0 THz波段存在明显的吸收峰。 为了准确测定甲醇气体的浓度, 根据甲醇气体在0.1~1.0 THz范围内的15处不同的位置处的特征吸收峰强度和甲醇气体浓度的关系, 对十五组不同浓度的甲醇气体进行检测, 获得了在特征吸收峰处的差异曲线。 基于误差反向传播(BP)神经网络的函数逼近特点, 并利用遗传算法(GA)收敛速度较快, 不宜陷入局部极值的优点, 采用GA优化BP神经网络的初始的权值和阈值, 构建了以预测甲醇浓度为目的的数学模型。 结果表明, 该网络模型适用于体积浓度范围为0.028 3~0.424 6 m3·L-1的甲醇的浓度预测, 两组样本的平均相对标准误差为1.7%, 平均回收率为98%, 神经网络误差精度10-1, 实测值与期望值的相关系数为0.996 77, 基本达到理想预测结果。 本成果不仅获得了甲醇气体在太赫兹频段的实验数据, 而且发现太赫兹时域光谱法和GA-BP神经网络相结合的方法能有效地检测甲醇气体的体积浓度, 为检测甲醇气体浓度提供新的方法。

At ambient temperature and atmospheric pressure, making use of a photoconductive-antenna terahertz time-domain spectrograph and a self-designed air chamber, the terahertz time-domain spectroscopy (THz-TDS) technique test of methanol gas in the range of 0.1~3.0 THz shows that the methanol gas has no obvious absorption peaks in the range of 0.1~3.0 THz and has obvious absorption peaks in the range of 0.1~1.0 THz. In order to improve the determination accuracy of the concentration of the methanol gas, the author detected 15 groups of methanol gas with different concentrations on the basis of the relationship between the strengths of 15 characteristic absorption peaks of different locations and the concentration of the methanol gas, and obtained the difference curve of the of the characteristic absorption peaks. Based on the function approximation of BP neural network, the author optimized the initial weights and biases of the BP neural network by using the GA the genetic algorithm, which has higher rate of convergence to prevent from getting into local optimum easily, and constructed the mathematical model with the purpose of predicting the methanol gas concentration. The test results show that the neural network is applicable to predict methanol gas in the volume concentration range of 0.028 3~0.424 6 m3·L-1, the average relative standard deviation of the 2 sets of samples is 1.7%, the average recovery rate is 98%, the error precision of the neural network is 10-1, and correlation coefficient of the measured values and the predicted values is 0.996 77. The test basically achieved ideal predicted results. The research results obtained experimental data of methanol gas in the terahertz frequency band and found that the method of combining terahertz time-domain spectroscopy with GA-BP neural network can effectively detect the volume concentration of methanol gas, and provided a new method for the detection of concentration of methanol gas.