Main > Chinese Optics Letters >  Volume 17 >  Issue 6 >  Page 063001 > Article
Chinese Optics Letters, Vol. 17, Issue 6, 063001 (2019)

High-resolution frequency-domain spectroscopy for water vapor with coherent and continuous terahertz wave

Cunjun Ruan1,2,*, Deyin Kong1, Jun Dai1, Kanglong Chen1, Sujie Guo1, and Xiaojun Wu1

Author Affiliations

  • 1School of Electronic and Information Engineering, Beihang University, Beijing 100191, China
  • 2Beijing Key Laboratory for Microwave Sensing and Security Applications, Beihang University, Beijing 100191, China

Abstract

High-resolution frequency-domain spectroscopy (FDS) is set up using a coherent and continuous wave terahertz (THz) emitter and receiver. THz waves are generated and detected by two photomixers with two distributed feedback (DFB) lasers. Atmospheric water vapor with different relative humidity is systematically investigated by the FDS. A high-frequency resolution of ∼14 MHz is obtained with the help of Hilbert transformation, leading to a well resolved and distinct transmittance characterization of water vapor. Compared with conventional THz time-domain spectroscopy, the high-resolution continuous wave THz spectrometer is one of the most practical systems in gas-phase molecular sensing, identification, and monitoring.

1 引言

火灾是一种常见且容易发生的灾害,在各类工业控制过程中火灾对公众的生命与财产构成了重大威胁。由于火灾具有突发性而且危害大,所以构建准确有效的火焰识别算法对预防火灾拥有重要的意义。

传统的火焰检测算法大致可归纳为两类:基于传感器的火焰检测算法和基于图像的火焰检测算法。基于传感器的火焰检测算法是通过感温和感烟的方式进行火焰检测,如徐炀[1]构建了智能火灾自动报警系统,通过使用火灾报警控制器、感温感烟探测器等物理装置侦测周围温度和空气变换进行火焰检测;胡幸江[2]针对常规的感温感烟探测器等物理装置探测距离短、响应慢等缺点,提出使用多波段红外火焰探测器进行火焰检测。但是,这类方法对环境依赖性强、覆盖范围小、成本较高。随着高清摄像机、图像处理技术的不断发展,基于图像的火焰检测算法逐渐成为主流,张进华等[3]根据火焰的物体特性,利用火焰区域边缘点到中心质点的高度,并使用离散傅里叶变换排除非火焰区域;荣建忠等[4]和Chen等[5]利用RGB和HIS两个颜色空间获取疑似火焰区域,并通过判断其面积增长和中心稳定情况获取火焰的动态像素,引入并改进了统计地形特征的纹理描述方法,采用神经网络算法对火焰候选区域进行分类;李文辉等[6]使用日夜两用型红外摄像机,通过获取火焰在红外和可见光两种状态下的颜色模型进行火焰检测;严云洋等[7]通过RGB彩色空间建立了火焰的彩色模型,并提出了基于连通区域面积阈值化的单像素宽度目标轮廓特征的抽取方法。但是上述传统基于图像的火焰检测识别方法受制于手工设计火焰特征,随着场景变换和场景复杂程度加深,手工设计的火焰特征无法很好描述火焰目标并有效区分疑似火焰目标和火焰目标,造成识别精度的下降。

针对上述基于传统图像火焰检测算法的缺陷,本文提出一种通过训练特征提取网络自动提取火焰特征,并在该特征上使用区域全卷积网络(R-FCN)定位火焰目标,最后使用残差网络(ResNet)进行二次分类端到端的火焰检测算法。

2 R-FCN+ResNet的火焰检测模型介绍

图1所示,R-FCN+ResNet算法模型分为三个部分:第一部分为自动提取特征的深度卷积神经网络,目的是自动提取特征,并将网络最后一层的卷积特征谱图(高×宽为35 pixel×63 pixel,约为原图的1/16)提供给R-FCN;第二部分为R-FCN直接用在特征谱图上检测火焰,并回归出火焰在原图中的位置;第三部分为二次分类器,根据R-FCN给出的火焰位置,将原图中的目标截取下来并调整成224 pixel×224 pixel的图片,再对该目标进行二次分类,以进一步降低误报率。

Figure 1. Flame detection algorithm model diagram

Download full sizeView all figures

2.1 特征提取网络模型结构

本文算法的特征提取网络为ResNet。其图像输入尺寸为600 pixel×1000 pixel,图像通道数为3,网络层数为50,并且已经在ImageNet数据集上做过预训练。

2.2 R-FCN模型介绍

R-FCN效仿R-CNN[10],采用流行的物体检测策略,分为两步:1) 由RPN[11]根据特征提取网络提供的特征谱图求出疑似火焰区域集合(ROIS); 2) 对RPN求出的ROIS逐个进行火焰识别(ROI为ROIS中任意一个疑似火焰区域)。

图1中位于特征谱图右边彩色卷积层是对其进行卷积操作后得出的卷积层,并用来生成位置敏感分数图。位于特征谱图右边彩色卷积层在整幅图像上为每类生成k2个位置敏感分数图,有C类物体外加1个背景,因此其通道数为k2(C+1)。与PASCAL VOC数据集中的20个类别相比,火焰的特征比较单一,若k值太大,则位置敏感分数图的检测速度必然会下降,但是识别精度未必有明显的提高,反而可能造成过拟合。所以,本文算法模型针对火焰检测设置k=3。检测类别分别是疑似火焰目标和火焰目标,其类别数C=2。

图1所示,RPN给出感兴趣的ROIS后,对每个ROI做位置敏感池化得出通道数为k2(C+1)位置敏感分数图。当k=3时,ROI可分割成9个矩形块,则对应的位置敏感分数图也就分割成9块,分别为{top-left,top-center,top-right,…,bottom-right}。再对位置敏感分数图做均值池化,从而得出一个长度为C+1的向量,并对向量进行softmax分类以得出该ROI的类别。

位置敏感ROI池化具体操作是将1个尺寸为w×h的ROI分割成k2个大小为w/k×h/k的矩形方块,并使用(i,j)表示ROI和位置敏感分数图中每个分块的位置,且(i,j)的取值范围为(0≤ik-1,0≤jk-1)。再将图1中top-left箭头指向的通道数为k2的特征图对应的ROI中第(0,0)的方块放到位置敏感分数图的第(0,0);top-center箭头指向的通道数为k2的特征图对应的ROI中第(1,0)的方块放到位置敏感分数图的第(1,0),以此类推,最后将bottom-right箭头指向的通道数为k2的特征图对应的ROI中第(2,2)的方块放到位置敏感分数图的第(2,2),得到完整的位置敏感分数图。得到位置敏感分数图中第(i,j)个矩形方块的具体操作为

In order to verify the sensitivity of THz-FDS, we measure the atmospheric water vapor absorption of 7.5%±0.6%, 12.0%±0.9%, 26.0%±0.8%, 33.5%±0.7%, and 40.2%±1.6%. The temperatures of them are 24.8°C±0.2°C, 25.0%±0.4°C, 25.0°C±0.3°C, 25.5°C±0.2°C, and 26.5°C±0.2°C. The calculated p-p is shown in Table 1, and the results are shown in Fig. 5. Due to crowded absorption lines and strong absorption coefficients at higher frequencies, it is very difficult to obtain obviously resolved water vapor absorption lines for the frequencies. The other reason is the output power of the THz emitter at higher frequencies is relatively low, less than 0.1 μW. The water vapor absorption signatures reduce when the RH decreases. For the three lower frequencies of 0.558, 0.753, and 0.989THz, they can hardly be detected when the RH is 7.5%. Figure 5(b) exhibits the enlarged absorption lines at 0.558 THz measured for RHs of 40.2%, 33.5%, 26%, 12%, and 7.5%, respectively. The frequency difference between the simulated results and the measured ones may be caused by the influence of water clusters[9]. The water clusters in high humidity may have special properties.

  • Table 1. Partial Pressure

    View tableView all Tables

    Table 1. Partial Pressure

    Temperature (°C)Relative Humidity (%)Partial Pressure (Pa)
    24.87.5234.9
    25.012.0380.4
    25.026.0824.2
    25.533.51094.0
    26.540.21392.8

Figure 5. (a) Water vapor transmittance spectra measured by THz-FDS at RHs of 40.2%, 33.5%, 26%, 12%, and 7.5%, respectively. (b) The enlarged view of the transmittance spectra at 0.558 THz for different RHs. Green, measured results; red, simulated results; blue, smoothed result.

Download full sizeView all figures

When the RH decreases, the interference signals from the standing waves of the system caused by the FP effect, for example, from the interface between the air and the silicon lens, turn out to be dominant. This noise limits the system’s ability for trace substance. Up to now, there is no good way to eliminate these phenomena. In our case, we try to remove the oscillations by converting the spectrum to a calculated time-domain signal, where the oscillations and main signal can be distinguished by the intensity. This method is like Fourier self-deconvolution (FSD)[27]. The differences are that FSD is suitable for the absorption spectroscopy, but, in our work, the targeted data is the original data, which contains the transmission coefficient spectrum. The results are shown in Figs. 3, 4, and 5(b).

During our experiments, we find that the THz-FDS is very sensitive to the RH. Therefore, it may become a highly sensitive remote monitoring system for water vapor, especially in some specific conditions, such as vacuum chambers, high altitudes, and outer space water vapor detection. For the fact that the temperature difference of the five data is small, we can get the peaks’ transmittance changed with RH at 0.558, 0.753, and 0.989 THz.

In Fig. 6, for the three frequencies, the transmittance decreases with RH and frequency. The error at RH 7.5% is caused by the low SNR, which is due to FP interference. RH cannot directly represent the amount of water vapor, but, in our results, the deviation of temperature is low, so the RH is approximately in direct proportion to the p-p. In our system, 0.558 THz is near the peak intensity of the source output energy, so the signal is strong, and the absorption of water is large enough for identification under the effect of FP interference. It means that the absorption peak at 0.558 THz is more appropriate for remote sensing of atmospheric water vapor in this system.

Figure 6. Transmittance at 0.558, 0.753, and 0.989 THz for different RHs of 40.2%, 33.5%, 26%, 12%, and 7.5%, respectively.

Download full sizeView all figures

In order to verify the absorption lines of water vapor and compare the capability between pulsed THz sources and CW THz sources, a conventional THz-TDS is applied to measure the water vapor. Figure 7 provides the typical THz temporal waveforms and their corresponding spectrum. For a better resolution, we choose the longest time, where the reflection signal can also be measured, as shown in the temporal waveform of the vacuum in Fig. 7(a). Due to limited resolution, the lower-frequency absorption lines at 0.558, 0.753, and 0.989 THz can be observed but not well resolved. Further comparison between THz-FDS and THz-TDS is summarized in Fig. 8. The spectrum of the THz-FDS is measured at the RH of 33.5%. We can see in Fig. 8(a) that almost all the absorption lines can be measured in the two systems. However, the THz-FDS has obviously much higher frequency resolution, resulting in more points and accurate measurements, referring to Fig. 8(b). This is very important for gas-phase spectroscopy, in which high-resolution is needed for material identification.

Figure 7. (a) Measured THz temporal waveform of the atmospheric water vapor (blue curve) and vacuum (red curve), and (b) the corresponding Fourier transform spectrum of the two kinds of environments, respectively.

Download full sizeView all figures

Figure 8. (a) Merged spectra of atmospheric water vapor measured by THz-TDS and THz-FDS. (b) Comparison of the zoom in graph for the absorption line at 0.558 THz.

Download full sizeView all figures

References

[1] P. F.-X. Neumaier, K. Schmalz, J. Borngräber, R. Wylde, H.-W. Hübers. The Analyst, 140, 213(2015).

[2] D. Bigourd, A. Cuisset, F. Hindle, S. Matton, R. Bocquet, G. Mouret, F. Cazier, D. Dewaele, H. Nouali. Appl. Phys. B Lasers Opt., 86, 579(2007).

[3] K. P. V. Rivera, A. M. Gallego, J. M. Hernández, A. Monroy-Guzmán, Y. Mares-Gutiérrez, J. A. Christen, L. G. Ruiz-Suárez, S. Alavez, A. M. Juárez, 1747, 030001(2016).

[4] H. Cai, D. Wang, J. Shen. Sci. China Phys. Mech. Astron., 56, 685(2013).

[5] H. Zhang, Z. Zhang, X. Zhao, X. Zhang, T. Zhang, C. Cao, Y. Yu. Chin. Opt. Lett., 16, 103001(2018).

[6] M. van Exter, C. Fattinger, D. Grischkowsky. Opt. Lett., 14, 1128(1989).

[7] J. S. Melinger, Y. Yang, M. Mandehgar, D. Grischkowsky. Opt. Express, 20, 6788(2012).

[8] A. S. Pine, R. D. Suenram, E. R. Brown, K. A. McIntosh. J. Mol. Spectrosc., 175, 37(1996).

[9] Y. D. Wu, T. Zhou, Z. W. Yao, J. C. Cao. J. Appl. Spectrosc., 83, 362(2016).

[10] Y. Yang, A. Shutler, D. Grischkowsky. Opt. Express, 19, 8830(2011).

[11] Y. Yang, M. Mandehgar, D. Grischkowsky. Opt. Express, 20, 26208(2012).

[12] X. Wu, S. Chai, J. Ma, B. Zhang, C. Xia, Z. Fang, D. Kong, J. Wang, H. Liu, C. Zhu, X. Wang, C. Ruan, Y. Li. Chin. Opt. Lett., 16, 041901(2018).

[13] F. S. Vieira, F. C. Cruz, D. F. Plusquellic, S. A. Diddams. Opt. Express, 24, 30100(2016).

[14] Z. Mihoubi, K. G. Wilcox, S. Elsmere, A. Quarterman, R. Rungsawang, I. Farrer, H. E. Beere, D. A. Ritchie, A. Tropper, V. Apostolopoulos. Opt. Lett., 33, 2125(2008).

[15] G. Mouret, F. Hindle, A. Cuisset, C. Yang, R. Bocquet, M. Lours, D. Rovera. Opt. Express, 17, 22031(2009).

[16] W. Shi, Y. J. Ding. Laser Phys. Lett., 1, 560(2004).

[17] A. J. Deninger, A. Roggenbuck, S. Schindler, S. Preu. J. Infrared Millim. Terahertz Waves, 36, 269(2014).

[18] A. Roggenbuck, H. Schmitz, A. Deninger, I. C. Mayorga, J. Hemberger, R. Güsten, M. Grüninger. New J. Phys., 12, 043017(2010).

[19] D. W. Vogt, R. Leonhardt. Opt. Express, 25, 16860(2017).

[20] G. L. Carr, R. J. Smith, L. Mihaly, H. Zhang, D. H. Reitze, D. B. Tanner. Infrared Phys. Technol., 51, 404(2008).

[21] T. Seta, H. Hoshina, Y. Kasai, I. Hosako, C. Otani, S. Loßow, J. Urban, M. Ekström, P. Eriksson, D. Murtagh. J. Quant. Spectrosc. Radiat. Transf., 109, 144(2008).

[22] Y. Yang, M. Mandehgar, D. Grischkowsky. J. Infrared Millim. Terahertz Waves, 36, 97(2015).

[23] C. Lin, I. Ho, X. C. Zhang. Chin. Opt. Lett., 10, 043001(2012).

[24] I. Cámara Mayorga, E. A. Michael, A. Schmitz, P. van der Wal, R. Güsten, K. Maier, A. Dewald. Appl. Phys. Lett., 91, 031107(2007).

[25] G. P. Tolstov. Fourier Series(2012).

[26] C. N. Mikhailenko, Y. L. Babikov, V. F. Golovko. Atmos. Ocean. Opt., 18, 685(2005).

[27] J. K. Kauppinen, D. J. Moffatt, H. H. Mantsch, D. G. Cameron. Appl. Spectrosc., 35, 271(1981).