Open Access
Add to BibSonomy
Abstract
A novel mid-infrared methane remote sensor integrated on a movable platform based on a 3.291-µm interband cascade laser (ICL) and wavelength modulation spectroscopy (WMS) is proposed. A transmitting-receiving coaxial, visualized optical layout is employed to minimize laser energy loss. Using a hollow retro-reflector remotely deployed as a cooperative target, the atmospheric average methane concentration over a 100-meter optical range is measured with high sensitivity. A deep neural network (DNN) filter is used for second harmonic (2f) signal denoising to compensate for the performance shortcomings of conventional filtering. Allan deviation analysis indicated that after applying the DNN filter, the limit of detection (LOD) of methane was 86.62 ppb with an average time of 1 s, decreasing to 12.03 ppb with an average time of 229 s, which is a significant promotion compared to similar work reported. The high sensitivity and stability of the proposed sensor are shown through a 24-hour continuous monitoring experiment of atmospheric methane conducted outdoors, providing a new solution for high-sensitivity remote sensing of atmospheric methane.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Methane (CH4) is one of the most important trace gases in the atmosphere, whose sources are mainly fossil fuel extraction, oceans, wetlands, and agricultural land. CH4 plays an important role in global warming, and its concentration variation in the atmosphere is severely affected by human activities [1,2]. As a flammable and explosive gas, CH4 leakage is a safety hazard in many industries including coal mining, natural gas, and chemical [3]. Therefore, highly sensitive continuous monitoring of atmospheric CH4 concentration is essential for the study of global warming, environmental change, and public safety. Compared to traditional monitoring methods including catalytic combustion [4,5], electrochemistry [6], gas chromatography [7], and semiconductors [8], spectroscopic sensors exhibit better sensitivity, response time, dynamic range, and selectivity. Common spectroscopic techniques for gas detection include photoacoustic spectroscopy [9], cavity ring-down spectroscopy (CRDS) [10], cavity-enhanced absorption spectroscopy (CEAS) [11], Fourier transform infrared spectroscopy (FTIR) [12] and tunable diode laser absorption spectroscopy (TDLAS) [13–18]. Owing to high reliability, relatively simple system structure, and high resistance to environmental changes, the TDLAS method has been widely used for atmospheric and ground-based monitoring of trace gases.
Conventional in-situ TDLAS gas sensing requires pumping air into a gas cell for sampling. Although it is possible to increase the optical length of the gas cell to promote the LOD of the sensor to the ppb level or even lower [19], only the gas concentration in a small area near the sampling point can be measured. In the application scenario of gas detection, it is usually essential to obtain the average gas concentration over a large atmospheric area, so multiple sampling is required with in-situ sensors, which is difficult to realize in some harsh environments such as mines and pipelines. However, the remote gas sensor directly measures the long-range gas concentration in the atmosphere, overcoming the mentioned drawbacks of in-situ TDLAS sensors. Due to the lower atmospheric stability and higher complexity of the optical path, remote sensing is limited by more interfering factors than in-situ sensing, yet the convenience of measurement and mobile deployment implies a greater utility. The strongest stretching and vibration absorption band of CH4 is centered at 3.3 µm, whose absorption intensity is approximately two orders higher than that of the near-infrared 1.6 µm overtone absorption band. Therefore, aligning the sensor to the mid-infrared absorption line realizes a higher sensitivity. The Interband cascade lasers (ICL) covering the 3–6 µm band are the excitation source for the proposed sensor, which owns low power consumption and high tuning accuracy.
Based on the TDLAS technique, there are two methods using infrared lasers as excitation sources: direct absorption spectroscopy (DAS) and wavelength modulation spectroscopy (WMS). The DAS gas sensor usually contains electrical and optical noise. Electrical noise is generated by electronic components such as laser drive and amplifier circuits [20], which roughly shows pink noise (∼ 1/f) and white noise. Optical noise mainly includes multibeam interference noise from light reflection by multiple surfaces in the optical path (etalon effect), baseline drift noise caused by long-time laser operation, and useless power modulation [21]. Unlike electronic noise, optical noise has no well-defined statistical properties. The WMS technology superimposes a high-frequency modulating wave on a low-frequency scanning wave to shift the absorbed signal to the high-frequency domain and effectively suppress pink noise, but still cannot avoid the effect of white noise and optical noise. It is reported that some conventional filtering algorithms show efficiency in suppressing electronic noise in TDLAS techniques [22–24], whereas the filtering of spectrally disordered optical noise is ineffective. The deep learning algorithm has been widely used in various fields such as remote sensing [25], materials science [26], biochemistry [27], and gas sensing [28,29]. Compared with conventional filters, deep neural network (DNN) filters are well adapted to multiple noises, especially optical noises which are difficult to be filtered by conventional filters [30]. Therefore, DNN filters are a better choice than conventional filters for atmospheric remote sensing with complex noise compositions.
This study focuses on a realistic application scenario of methane detection and proposes a novelty long-range CH4 remote sensor using DNN filtering. The modular architecture, optical design for remote sensing, and system building of the proposed sensor are introduced in section 2. In section 3, we describe the superiority and optimization of DNN filters, as well as a methodology for building datasets with simulated 2f signals. In section 4, we show the effect of DNN filtering and the detailed evaluation of sensor performance, as well as the observed cyclic change in atmospheric CH4 concentration over 24 hours. The proposed sensor achieved ppb-level sensitivity at 100 m atmospheric range and showed high stability during 24 hours of continuous monitoring, which demonstrates the broad applicability to high-sensitive measurements for methane and other trace gases.
2. Sensor configuration
2.1 Sensor architecture
The structure of the proposed CH4 sensor is shown in Fig. 1(a). It mainly consists of optical, electronic, gas, and filtering parts. In the optical part, a Nanoplus continuous-wave (CW) thermoelectrically cooled (TEC) ICL was employed as the excitation source. The laser beam was collimated by a plano-convex lens (Lens 1, f = 5.98 mm) and reduced by a set of plano-convex lenses (Lens 2, f = 50 mm; Lens 3, f = 20 mm). The output beam is then combined with the visible alignment diode laser beam via a dichroic mirror (DM, Spectrogon, model NB3300-150) to enter a straight-through calibration gas cell with an optical length of 10 cm (Ganwei Technology, model GW-1000B). After leaving the gas cell, the laser beam was reflected twice, passed through an off-axis parabolic reflector (PM, f = 4 in, Thorlabs, model MPD249V-M01) with a through hole, and propagated into the atmosphere. A hollow retroreflector (RR, Thorlabs, Ø2.0 in, model HRR203-M03) was placed 50 m away to reflect the laser beam to the PM, and then the received laser was focused onto a mercury cadmium telluride detector (MCT, VIGO System, model PVI-2TE-3.4).
Fig. 1. (a) Block diagram of the proposed CH4 sensor, showing the optical part, gas part, electrical part, and filtering part. (b) Photograph of the optical part of the proposed CH4 sensor. (c) Photograph of the proposed CH4 sensor integrated into a mobile platform.
For remote sensors with long optical range, reducing the loss of beam energy in the optical path can effectively increase the signal amplitude and thus improve the measuring accuracy. We used two tricks to optimize the optical design of the proposed sensor. Firstly, we indirectly adjusted the invisible infrared optical path through a visible alignment laser. After collimation and beam reduction, we measured the output beam to have a diameter of 2 mm and a divergence angle of about 3 mrad. Secondly, we used the RR as a target for the laser beam, whose retro-reflective properties cause the diverging laser beam to re-converge, thus reducing the energy loss due to the long optical range. It is deduced from the calculation that the laser beam can be effectively confined in the optical system of the proposed sensor. Figure 1(b) shows the optical part of the proposed CH4 sensor. In the gas part, N2 was mixed with CH4 in a pre-defined ratio using a digital mass flow controller (MFC, Laifeng Technology, model LFIX-1) and fed into the calibration gas cell with an accuracy better than 1%.
The electronic part of the proposed CH4 sensor consists of an FPGA board (Altera, EP4CE22F17C8N), a DAQ card (National Instruments, model USB-6361), and a laptop. A direct digital frequency synthesizer (DDS) in the FPGA board generates a 1 Hz sawtooth signal with a 20 kHz sine wave superimposed on it, which is converted to a current signal to drive the ICL laser diode by the DAC8830 chip and constant current source. The FPGA collects the laser diode’s temperature and voltage signals through the ADC7680 chip and then adjusts the direction of the current flowing through the TEC to control the temperature of the laser diode. Absorption spectrum signals output from the MCT detector and the previously generated high-frequency modulation signals are each fed into a digital lock-in amplifier (LIA) within the FPGA board via an SMA port. The digital LIA is based on a Butterworth low-pass filter, with a pass bandwidth of 100 Hz. The absorption spectrum signals are demodulated by digital LIA to 2f signals, which are then sent to the DAQ card for sampling via an SMA port. The DAQ card has 2000 samples per trigger and a sampling frequency of 6.6 kHz, using the trigger signal of the drive-current module as the sampling clock. The laser, detector, and DAQ card were all controlled by software on the laptop. Figure 1(c) shows a photograph of the sensor integrated into the mobile platform. In the filtering part, the DAQ card inputs the acquired 2f signals into the laptop through a USB port, and then the 2f signals are quickly filtered by a DNN filter (for details see Section 3).
2.2 Selection of the CH4 absorption line
Figure 2 depicts the absorption spectra based on the HITRAN database at 296 K, 1 atm, and 100-meter effective optical length in the wavenumber range 3036–3041 cm−1 for CH4 and the two main interfering gases in air, H2O and CO2. Typical atmospheric concentrations of 2 ppm and 385 ppm are used for CH4 and CO2, respectively, and 50% relative humidity at 296 K and 1 atm (∼13866 ppm) for H2O.
Fig. 2. HITRAN-based absorption spectra of CH4 (2 ppm), CO2 (385 ppm), and H2O (relative humidity = 50%) in the wave number range 3036–3041 cm−1, at a temperature of 296 K, a pressure of 1 atm, and an effective optical length of 100 m.
For remote sensing of atmospheric CH4, the selection of independent absorption lines that are free from interference is critical. In the wavenumber range 3036–3041 cm−1, the CH4 absorption line centered at 3038.5 cm−1 has high intensity and no interference, while the absorption effect of CO2 can be neglected, and H2O has a spectral line centered at 3037.1 cm−1. Since the absorption line of H2O is low in intensity and flat near 3038.5 cm−1, it can be regarded as a background signal. Therefore, the CH4 spectral line at 3038.5 cm−1 was selected as the target spectral line for detection.
2.3 ICL and MCT detector characterization
The Nanoplus CW TEC ICL has a central wavelength of 3.291 µm, an operating temperature of 15–25°C, and a drive current regulation range of 20–80 mA. The emission wavenumber of the ICL was measured with a mid-infrared high-precision wavelength meter (Bristol, model 771A-MIR). The function of drive current and temperature as shown in Fig. 3(a), and its emission power was measured with a wide-band power meter (Newport, model PMKIT-21-01). The function of drive current with a wavelength under the operating temperature of 20°C is shown in Fig. 3(b), with a tuning factor is −0.0675 cm−1/mA. The output laser wavelength with a drive current of 60.14 mA was aligned with the target absorption line, and the laser emission power was 4.02 mW.
Fig. 3. (a) The wavenumber variation curve of emission laser with ICL temperature and drive current. (b) The measured emission power for the 3.291 µm CW TEC ICL at 20°C with different driving currents.
A VIGO PVI-2TE-3.4 MCT detector was used to measure the received power after absorption. The detector was packaged in TO8 with a window material of wAl2O3 and an active area of 1 × 1 mm2. The frequency response of the detector is 10 Hz–10 MHz, as well as the detectivity and responsivity of the detector at the central wavelength of 3.4 µm are respectively ∼6.0 × 1011 cmHz1/2/W and ∼0.8 A/W.
3. DNN filter configuration
3.1 Superiority of DNN filtering
DNN is a basic deep learning framework, which learns the abstract features of the input data through the forward propagation of data and backpropagation of errors to form a mapping from training data to labeled data. In both published the methane sensor based on the in-situ DAS technique [30] and the filtering experiment based on simulated Monte Carlo samples [31], DNNs have shown their effectiveness for denoising 1-D time-series signals. The frequency principle (F-principle) indicates that DNNs prefer low-frequency components of the input signal [32], i.e., the frequency components of the input signal are extracted from low to high frequencies in decreasing amplitude, which was proved to be caused by the regularity of the activation function and the initialization of the model [33]. The training behavior of DNNs in the frequency domain is like that of conventional filters. Compared to the denoising mechanism of conventional filters with high-frequency cutoffs, DNN filters can remove not only most of the high-frequency noise in the original signal but also most of the low-frequency optical noise mixed in the 2f signal baseline. Therefore, the application of DNN filters will bring a greater improvement in measuring accuracy for the sensors.
3.2 Methodology for dataset construction
The performance of DNN filters is highly dependent on the quality of their training datasets. While obtaining many high-quality experimental 2f signals of CH4 absorption spectra requires a high cost of time and labor, so we came up with a methodology for constructing a 2f signal dataset with simulated data:
Step 1. A simulation program was used to generate WMS absorption spectra based on the HITRAN database and the Lorentzian broadening mechanism of the absorption spectra, each containing 2000 sample points.
Step 2. A Simulink-based LIA is used to demodulate the pure 2f signal from the simulated absorption spectrum.
Step 3. The electronic noise characterized by the Gaussian function and the optical noise characterized by the Airy function is superimposed on the simulated 2f signal to obtain the noisy 2f signal as shown in Fig. 4(a)–(b).
Fig. 4. (a) The comparison of pure 2f signal with superimposed interference noise and average experimental data. (b) The comparison of pure 2f signal and noisy 2f signal. (c) The pure 2f signal with five different degrees of wavelength drift.
Step 4. A total of 20,000 values were selected uniformly over the concentration range of 0–280 ppm·m and multiplied as coefficients with the noisy 2f signals to obtain the simulated 2f signals corresponding to different CH4 concentrations.
Step 5. Considering the wavelength drift of the laser under prolonged operation that results in absorption peak translation [34] as shown in Fig. 4(c), we add five different levels of wavelength drift (translation units: -4, -2, 0, + 2, + 4) to the dataset to enhance the adaptability of the proposed filter to 2f signal translation.
Finally, the training and labeling datasets were constructed from noisy and pure 2f signal data, respectively, both with the shape of 100000 × 2000. Since the dataset was large enough and well-balanced, a hold-out cross-validation method was used for model training, whereby 20% of the training dataset was randomly selected as the dev dataset.
3.3 Training and optimization
After the construction of the filter based on the simulated dataset and DNN, its hyperparameters need to be tuned for optimal performance on the test dataset. Multiple hidden layers of a DNN filter can sequentially extract different features of the input data. If the number of hidden layers is too large, the neural network may extract useless features, which will lead to overfitting and reduce the effectiveness of denoising real 2f signals. To find the optimal number of hidden layers, we gradually increased the number starting with one hidden layer, and evaluated the SNR of the filtered signal. The DNN filter with 3 hidden layers shows the best performance. Then, we optimize the unit number of the three hidden layers by grid searching as 1000, 500, and 900 respectively. The unit number in both the input and output layers was set at 2000 to correspond to the number of columns in the dataset constructed in section 3.2.
To enhance the generalizability of the DNN model, we added a dropout step to the last hidden layer to randomly discard units and weaken the dependence of the model on specific units. Searching within dropout rates of 0–0.6 showed that the SNR of the predicted signal was the highest when the dropout rate was 0.1. The structure of the DNN model is schematically shown in Fig. 5. The learning rate was searched in the range of 10−1–10−6 and finally determined to be 10−4. To make the loss function converge faster, the activation function was chosen to be CELU. The batch size is set to 1000, i.e., each epoch contains 100 batches. The number of epochs is set to 3500, and Adam is selected as the optimizer. Figure 6 shows the complete schematic of the DNN filter.
Fig. 6. Schematic diagram of the DNN filter in three parts: (a) The dataset generation process. (b) The model training and optimization process. (c) The experimental data filtering process.
4. Sensor performance
4.1 Calibration and data fitting
Following the discussion on the performance of the DNN filtering, a series of tests on the proposed sensor were performed in the laboratory. Firstly, a calibration is required to obtain a linear correspondence between the CH4 concentration and the amplitude of its 2f signal. To minimize the calibration error, the CH4 concentration was calibrated in an empty, airless laboratory environment. Given an atmospheric background CH4 concentration of approximately 2.2 ppm, the concentration calibration range for the 100-m optical length sensor was 0–200 ppm. The MFC was used to prepare CH4 calibration concentrations at 0, 400, 800, 1200, 1600, and 2000ppm, which was passed into a calibration gas cell with a 10-cm optical length, followed by data acquisition using the DAQ card. The sampling frequency was set to 1 Hz and the 2f signal at each CH4 concentration was recorded for 5 min. A total of 1800 sets of experimental data were directly fed into the DNN filter for denoising without normalization with respect to the laser power.
Figure 7(a) shows the amplitude of the 2f signal after denoising. The amplitudes were averaged for each concentration level and are plotted as a function of CH4 concentration in the calibration gas cell in Fig. 7(b). A linear relationship was observed between the amplitude of the 2f signal ($am{p_{2\textrm{f}}}$) and the CH4 concentration ($C$):
The fitting curve indicates a good linear relationship between the 2f signal amplitude and CH4 concentration, with R-square = 99.90%. According to the above equation, the atmospheric CH4 concentration can be calculated using the value of 2f signal amplitude.
Fig. 7. (a) The experimentally observed 2f signal amplitude versus time for CH4 at concentration levels of 0, 40, 80, 120, 160, and 200 ppm·m in a calibration gas cell. (b) CH4 concentration in the calibration gas cell versus the average amplitude of the experimentally observed 2f signal.
4.2 Sensor stability
To evaluate the stability of the sensor during long-term operation, we filled the calibration gas cell with 200 ppm·m of CH4 standard gas to reach an average CH4 concentration of 2 ppm over a 100-meter optical path. Then, we measured the CH4 concentration continuously in the laboratory for one hour. The observed 2f signal was denoised using the DNN filter, then the amplitude of the signals was calculated and the CH4 concentration was inverted using Eq. (1).
Figure 8(a) shows the average CH4 concentrations over 100-m optical length of 3600 sets of 2f signals measured in one hour. The total variation in CH4 concentration over one hour relative to the standard value of 2 ppm ranges from -329–299 ppb, from which the measuring accuracy of the sensor can be calculated to be 97.93 ppb. In Fig. 8(b), The Allan deviation was plotted on a log-log scale as a function of the averaging time, which can be used to characterize the sensitivity of the sensor by calculating the LOD. The LOD of the proposed sensor is 86.62 ppb for an average time of 1 s. With increasing averaging time, the lowest LOD of 12.03 ppb with a 229 s averaging time was obtained. As shown in Table 1, the proposed sensor has a lower LOD than previously reported CH4 remote sensors.
Fig. 8. (a) Distribution of the measured average CH4 concentrations over 100-m optical range relative to a central value of 2 ppm. (b) Allan-Werle deviation plot based on the one-hour experimental data of average CH4 concentrations over a 100-m optical range.
Table 1. Comparison of several CH4 remote sensors.
4.3 DNN filter performance
To improve the measuring accuracy, many studies have been conducted on the application of conventional filtering algorithms to inversion of gas concentration for TDLAS [22,39,40]. Based on 3600 sets of experimental 2f signal data, we compared five widely used classical filters, including Kalman Filter (KF), Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) [41], Wavelet Transform (WT), S-G smoothing filter (SG) and Singular Value Decomposition (SVD) with DNN filter in terms of both denoising ability and improvement in accuracy for sensor. To ensure a fair comparison, all filtering algorithms are hyper-parametrically optimized for best performance.
Figure 9 shows the time and frequency domain distribution of the experimental 2f signal after denoising by different filtering algorithms. Compared to the listed conventional filtering algorithms, the DNN-filtered signal is much smoother in the time domain and has a frequency domain distribution that is closer to the pure signal. Figure 10 shows the comparison of the signal before and after DNN filtering with the pure signal. Obviously, not only Gaussian fluctuations at high frequency but also significant wavelength drift and baseline fluctuations at low frequency are mixed in the experimental signal compared to the pure signal. Denoised by the DNN filter, the noise in the experimental signal is greatly suppressed and the experimental signal almost overlaps with the pure signal.
Fig. 9. (a)–(g) The experimental 2f signal from 3291.03 nm to 3291.14 nm after denoising by several filtering algorithms. (h)–(n) Frequency domain distribution of the experimental 2f signal after denoising by different filtering algorithms.
Fig. 10. The comparison among the experimental signal without filtering, the pure signal, and the DNN filtered signal.
Table 2 shows the average SNR and measuring accuracy (1σ) of 3600 sets of experimental data after denoising by several filtering algorithms, where the measuring accuracy is calculated based on the 2f signal amplitude and calibration results after denoising by each filtering algorithm, respectively. In terms of average SNR, none of the five conventional filtering algorithms improves the SNR of the real 2f signal by more than 0.2 dB, whereas DNN filtering improves the SNR of the real 2f signal to 14.9563 dB, which far exceeds other filtering algorithms. In terms of measuring accuracy, DNN filtering improves the measuring accuracy of the sensor by about 7 ppb, which is 43% better than other algorithms. Thus, the performance of DNN filtering exceeds that of several classical algorithms for both metrics evaluated.
Table 2. Comparison of the average SNR and measuring accuracy obtained by different filtering algorithms for 3600 sets of experimental data from the CH4 remote sensor.
4.4 Atmospheric methane detection
To test the continuous operational stability of the proposed sensor, a mobile platform with integrated sensor system was deployed at an open-air site for 24-hour continuous atmospheric CH4 monitoring as shown in Fig. 11(a). Two experiments were conducted in total, starting at 11:00 on May 7, 2023, and 17:00 on June 17, 2023, respectively. The calibration of the sensor was performed at the experimental site 2 hours before the start of the experiments. For a shorter calibration time, we reduced the sampling time to 2 minutes for each CH4 concentration value, based on the calibration method described in section 4.1. To reduce the amount of data, the sensor was set to sample every 3 seconds, yielding 28,800 sets of 2f signal data. The atmospheric CH4 background concentration can be calculated from the calibration curve and the measured 2f signal amplitude.
Fig. 11. (a) The photograph of the proposed sensor deployed at an open-air site. (b) The 24-hour atmospheric CH4 monitoring results for May 7–8, 2023. (c) The 24-hour atmospheric CH4 monitoring results for June 17–18, 2023.
Figure 11(b)–(c) shows the background concentrations of atmospheric CH4 measured in the two experiments. In the first experiment, the CH4 concentration reaches a minimum (∼2 ppm) in the evening (∼17:00), then rises steadily, and reaches a maximum (∼2.3 ppm) in the morning (∼7:00) before rebounding. In the second experiment, the CH4 concentration reaches a maximum (∼2.4 ppm) in the morning (∼7:00), then declines steadily and reaches a minimum (∼1.8 ppm) in the evening (∼16:00) before rebounding. Despite the different meteorological conditions of the two experiments, the cyclic trends in CH4 concentrations over 24 hours were roughly the same, consistent with other reports [18,24]. The mean atmospheric CH4 concentrations measured in the two experiments were 2.16 ppm ± 108.06 ppb (1σ) and 2.16 ppm ± 185.81 ppb (1σ), respectively.
5. Conclusion and future work
To overcome the drawbacks of traditional in-situ TDLAS gas sensors which require sampling and low detection range, this paper proposes a highly sensitive mid-infrared CH4 remote sensor based on ICL-WMS integrated on a removable platform, with a sensitivity up to the ppb level for 100-meter atmospheric field. In this paper, a conventional in-situ TDLAS gas sensor is revised in terms of both optical design and filtering algorithms.
The off-axis parabolic mirror with a through-hole is key to realizing the coaxiality of the receiving and transmitting optical path, simplifying system architecture and sensor deployment. The hollow retroreflector reduces the facula of the reflected beam and reduces the beam loss for remote sensing to a large extent. The aligning laser visualizes the infrared optical path, which facilitates optical alignments and measuring path positioning. In addition, we denoised the 2f signal using a DNN filter, which yields higher measuring accuracy compared to several classical filters. The calibration results show that the proposed sensor exhibits a good linear response to CH4 concentration. A 1-h stability experiment was conducted in the laboratory, and the Allan-Werle deviation analysis showed that after DNN filtering, the LOD of the sensor was 86.62 ppb with an averaging time of 1 s and could reach 12.03 ppb with an optimal averaging time of 229 s, superior to previous results of CH4 remote sensing. Further, we carried out a 24-hour atmospheric CH4 monitoring to verify the high stability and sensitivity of the proposed sensor, showing its application prospects for high-precision continuous monitoring of atmospheric trace gases.
However, the proposed sensor system still has the potential for optimization in terms of filter and system design, and our future work will build on the foundation of this study. For filtering, more effective deep-learning-based filtering algorithms can still be explored due to the fixed structure and easy overfitting of DNNs. Moreover, the design of a host computer with DNN real-time filtering is also a promising improvement of the proposed sensor. For system design, research on gas sensing based on non-cooperative targets will further simplify deployment and enhance utility. The integration, packaging, and miniaturization of the system can also be further explored to broaden the application scenarios of the proposed sensor.
Funding
Civil space technology pre-research of the 14-th five-years plan (D040107); Key basic research projects of the Basic Strengthening Program (2021-173ZD-025).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. S. Schwietzke, O. A. Sherwood, L. M. P. Bruhwiler, et al., “Upward revision of global fossil fuel methane emissions based on isotope database,” Nature 538(7623), 88–91 (2016). [CrossRef]
2. H. Schaefer, S. E. M. Fletcher, C. Veidt, et al., “A 21st-century shift from fossil-fuel to biogenic methane emissions indicated by 13CH4,” Science 352(6281), 80–84 (2016). [CrossRef]
3. C. Zheng, B. Jiang, S. Xue, et al., “Coalbed methane emissions and drainage methods in underground mining for mining safety and environmental benefits: A review,” Process Saf. Environ. Prot. 127, 103–124 (2019). [CrossRef]
4. N. Lu, S. Fan, Y. Zhao, et al., “A selective methane gas sensor with printed catalytic films as active filters,” Sens. Actuators, B 347, 130603 (2021). [CrossRef]
5. Y. Xia, J. Wang, L. Xu, et al., “A room-temperature methane sensor based on Pd-decorated ZnO/rGO hybrids enhanced by visible light photocatalysis,” Sens. Actuators, B 304, 127334 (2020). [CrossRef]
6. T. Qin, Y. Niu, X. Qiao, et al., “Integrated coaxial graphene-based yarn with multidimensional architecture for self-powered photoelectrochemical methane sensor,” Sens. Actuators, B 365, 131965 (2022). [CrossRef]
7. T. K. Mishra, A. Datta, and A. Mukhopadhyay, “Concentration measurements of selected hydrocarbons in methane/air partially premixed flames using gas chromatography,” Int. J. Therm. Sci. 44(11), 1078–1089 (2005). [CrossRef]
8. S. Barazzouk, R. P. Tandon, and S. Hotchandani, “MoO3-based sensor for NO, NO2 and CH4 detection,” Sens. Actuators, B 119(2), 691–694 (2006). [CrossRef]
9. C. Li, M. Guo, B. Zhang, et al., “Miniature single-fiber photoacoustic sensor for methane gas leakage detection,” Optics and Lasers in Engineering 149, 106792 (2022). [CrossRef]
10. L. E. Mchale, B. Martinez, T. W. Miller, et al., “Open-path cavity ring-down methane sensor for mobile monitoring of natural gas emissions,” Opt. Express 27(14), 20084 (2019). [CrossRef]
11. L. Lechevallier, R. Grilli, E. Kerstel, et al., “Simultaneous detection of C2H6, CH4, and δ 13C-CH4 using optical feedback cavity-enhanced absorption spectroscopy in the mid-infrared region: towards application for dissolved gas measurements,” Atmos. Meas. Tech. 12(6), 3101–3109 (2019). [CrossRef]
12. T. S. Jones, J. E. Franklin, J. Chen, et al., “Assessing urban methane emissions using column-observing portable Fourier transform infrared (FTIR) spectrometers and a novel Bayesian inversion framework,” Atmos. Chem. Phys. 21(17), 13131–13147 (2021). [CrossRef]
13. A. Zifarelli, A. Sampaolo, P. Patimisco, et al., “Methane and ethane detection from natural gas level down to trace concentrations using a compact mid-IR LITES sensor based on univariate calibration,” Photoacoustics 29, 100448 (2023). [CrossRef]
14. N. M. Davis, D. Francis, J. Hodgkinson, et al., “Compact methane sensor using an integrating sphere and interband cascade laser at 3313 nm,” Sens. Actuators, B 389, 133866 (2023). [CrossRef]
15. R. Bi, M. Pi, C. Zheng, et al., “A niobium pentoxide waveguide sensor for on-chip mid-infrared absorption spectroscopic methane measurement,” Sens. Actuators, B 382, 133567 (2023). [CrossRef]
16. L. Zhang, Z. Zhang, P. Sun, et al., “A dual-gas sensor for simultaneous detection of methane and acetylene based on time-sharing scanning assisted wavelength modulation spectroscopy,” Spectrochim. Acta, Part A 239, 118495 (2020). [CrossRef]
17. N. Liu, L. Xu, S. Zhou, et al., “Simultaneous Detection of Multiple Atmospheric Components Using an NIR and MIR Laser Hybrid Gas Sensing System,” ACS Sens. 5(11), 3607–3616 (2020). [CrossRef]
18. G. Gomółka, G. Stępniewski, D. Pysz, et al., “Highly sensitive methane detection using a mid-infrared interband cascade laser and an anti-resonant hollow-core fiber,” Opt. Express 31(3), 3685 (2023). [CrossRef]
19. J. Xia, C. Feng, F. Zhu, et al., “A sensitive methane sensor of a ppt detection level using a mid-infrared interband cascade laser and a long-path multipass cell,” Sens. Actuators, B 334, 129641 (2021). [CrossRef]
20. B. Lins, P. Zinn, R. Engelbrecht, et al., “Simulation-based comparison of noise effects in wavelength modulation spectroscopy and direct absorption TDLAS,” Appl. Phys. B 100(2), 367–376 (2010). [CrossRef]
21. P. Werle, “Accuracy and precision of laser spectrometers for trace gas sensing in the presence of optical fringes and atmospheric turbulence,” Appl. Phys. B 102(2), 313–329 (2011). [CrossRef]
22. C.-T. Zheng, W.-L. Ye, J.-Q. Huang, et al., “Performance improvement of a near-infrared CH4 detection device using wavelet-denoising-assisted wavelength modulation technique,” Sens. Actuators, B 190, 249–258 (2014). [CrossRef]
23. R. Yang and Y. Zhang, “A method of low concentration methane measurement in tunable diode laser absorption spectroscopy and Levenberg-Marquardt algorithm,” Optik 224, 165657 (2020). [CrossRef]
24. F. Song, C. Zheng, W. Yan, et al., “Interband cascade laser based mid-infrared methane sensor system using a novel electrical-domain self-adaptive direct laser absorption spectroscopy (SA-DLAS),” Opt. Express 25(25), 31876 (2017). [CrossRef]
25. P. Zhang, J. Tang, H. Zhong, et al., “Self-Trained Target Detection of Radar and Sonar Images Using Automatic Deep Learning,” IEEE Trans. Geosci. Remote Sensing 60, 1–14 (2022). [CrossRef]
26. C. Vrancken, P. Longhurst, and S. Wagland, “Deep learning in material recovery: Development of method to create training database,” Expert Systems with Applications 125, 268–280 (2019). [CrossRef]
27. Y. Jing, Y. Bian, Z. Hu, et al., “Deep Learning for Drug Design: an Artificial Intelligence Paradigm for Drug Discovery in the Big Data Era,” AAPS J. 20(3), 58 (2018). [CrossRef]
28. S. Zhou, C.-Y. Shen, L. Zhang, et al., “Dual-optimized adaptive Kalman filtering algorithm based on BP neural network and variance compensation for laser absorption spectroscopy,” Opt. Express 27(22), 31874 (2019). [CrossRef]
29. J. Sun, J. Chang, Y. Zhang, et al., “CH4/C2H6 dual gas sensing system using a single mid-infrared laser,” Spectrochim. Acta, Part A 291, 122368 (2023). [CrossRef]
30. L. Tian, J. Sun, S. Zhang, et al., “Near-infrared methane sensor with neural network filtering,” Sens. Actuators, B 354, 131207 (2022). [CrossRef]
31. L. Y. Wang, G. Yin, and Q. Zhang, “Deep filtering,” Communications in Information and Systems 21(4), 651–667 (2021). [CrossRef]
32. Z.-Q. J. Xu, Y. Zhang, Y. Xiao, et al., eds., Lecture Notes in Computer Science (Springer International Publishing, 2019), 11953, pp. 264–274.
33. Z.-Q. J. Xu, Y. Zhang, T. Luo, et al., “Frequency Principle: Fourier Analysis Sheds Light on Deep Neural Networks,” CiCP 28(5), 1746–1767 (2020). [CrossRef]
34. G. Li, X. Zhang, Z. Zhang, et al., “Mid-infrared telemetry sensor based calibration gas cell for CO detection using a laser wavelength locking technique,” Measurement 208, 112445 (2023). [CrossRef]
35. L. Dong, Y. Yao, Y. Zhao, et al., “A high-sensitive methane measuring telemetry system based on a direct optical absorption technique,” Laser Phys. 30(12), 126201 (2020). [CrossRef]
36. G. Li, K. Ma, Y. Jiao, et al., “Performance enhancement of DFBL based near-infrared CH 4 telemetry system using a focus tunable lens,” Microw Opt Technol Lett 63(4), 1147–1151 (2021). [CrossRef]
37. A. Diaz, B. Thomas, P. Castillo, et al., “Active standoff detection of CH4 and N2O leaks using hard-target backscattered light using an open-path quantum cascade laser sensor,” Appl. Phys. B 122(5), 121 (2016). [CrossRef]
38. Ding Wu-Wen, Sun Li-Qun, and Yi Lu-Ying, “High sensitive scheme for methane remote sensor based on tunable diode laser absorption spectroscopy,” Acta Phys. Sin. 66(10), 100702 (2017). [CrossRef]
39. Y. Meng, T. Liu, K. Liu, et al., “A Modified Empirical Mode Decomposition Algorithm in TDLAS for Gas Detection,” IEEE Photonics J. 6(6), 1–7 (2014). [CrossRef]
40. C. Li, X. Guo, W. Ji, et al., “Etalon fringe removal of tunable diode laser multi-pass spectroscopy by wavelet transforms,” Opt. Quantum Electron. 50(7), 275 (2018). [CrossRef]
41. M. A. Colominas, G. Schlotthauer, and M. E. Torres, “Improved complete ensemble EMD: A suitable tool for biomedical signal processing,” Biomedical Signal Processing and Control 14, 19–29 (2014). [CrossRef]
