近日，來自安徽大學(xué)的周勝副教授團隊發(fā)表了《基于深度神經(jīng)網(wǎng)絡(luò)的無需壓力校準和輪廓擬合的氣體傳感光譜技術(shù)》論文。

Recently, the research team from Associate Professor Zhou Sheng's from Anhui University published an academic papers Pressure calibration- and profile fitting-free spectroscopy technology based on deep neural network for gas sensing.

甲烷（CH₄）是天然氣的主要成分，在工業(yè)生產(chǎn)和日常生活中廣泛用作燃料。此外，甲烷是一種重要的溫室氣體，其濃度對全球氣候產(chǎn)生重要影響。因此，甲烷的測量對環(huán)境監(jiān)測、生物醫(yī)藥和能源研究具有重要意義。氣體濃度通常通過各種微量氣體傳感器進行測量，例如氣相色譜儀、半導(dǎo)體氣體傳感器和電化學(xué)設(shè)備。半導(dǎo)體氣體傳感器在適當?shù)牟僮鳝h(huán)境下具有ppm級別的靈敏度。激光吸收光譜技術(shù)具有高選擇性、高靈敏度、快速和多成分監(jiān)測等優(yōu)勢，目前廣泛用于各種氣體的檢測。激光吸收光譜技術(shù)可以準確測量氣體分子的特征吸收線，并基于可調(diào)諧激光有效降低其他氣體光譜線的干擾。此外，它提供了實時原位氣體檢測的可能性，這對于從工業(yè)過程到環(huán)境變化的各種現(xiàn)象的理解和監(jiān)測至關(guān)重要。氣體分子可以通過其指紋吸收光譜進行有效識別，包括典型的所謂“自展寬"參數(shù)和“空氣展寬"參數(shù)。光譜線參數(shù)是壓力和溫度的函數(shù)。濃度測量的準確性取決于壓力穩(wěn)定性和光譜擬合精度。對于定量光譜分析，傳統(tǒng)上通過準確的模型對光譜進行擬合，同時壓力和溫度必須定期校準，尤其是在相對大的環(huán)境波動情況下。因此，為實現(xiàn)所需的準確性，系統(tǒng)的復(fù)雜性增加了。

Methane (CH₄), which is the main component of natural gas, is widely used as fuel in industrial production and daily life. In addition, CH₄ is an important greenhouse gas whose concentration has a substantial influence on global climate. Therefore, the measurement of CH₄ has significant importance for environmental monitoring, biomedicine, and energy research. The gas concentrations are commonly measured by various trace gas sensors, such as gas chromatographs, semiconductor gas sensors, and electrochemical devices. The semiconductor gas sensors have a sensitivity of ppm level under a suitable operating environment. The laser absorption spectroscopy, which has the advantages of high selectivity, high sensitivity, and fast and multi-component monitoring, is currently widely used in the detection of a variety of gases. Laser absorption spectroscopy technology can accurately measure the characteristic absorption lines of gas molecules and effectively reduce the interference of other gas spectral lines based on the tunable lasers. Moreover, it provides the possibility of real-time in-situ gas detection, which is crucial for understanding and monitoring a variety of phenomena from industrial processes to environmental change. A gas molecule can be effectively identified by its fingerprint absorption spectrum, including typical so-called “self-broadening" parameters and “air-broadening" parameters. The spectral line parameters are functions of pressure and temperature. The accuracy of concentration measurement depends on pressure stability and spectral fitting accuracy. For quantitative spectral analysis, the spectra are traditionally fitted by an accurate model, while the pressure and temperature must be calibrated on time, especially in the case of relatively large environmental fluctuations. Consequently, the complexity of system is increased to achieve the required accuracy.

目前，人工智能的快速發(fā)展為解決這個問題提供了一種新途徑。人工神經(jīng)網(wǎng)絡(luò)已被用于氣體識別，并在足夠訓(xùn)練數(shù)據(jù)的條件下表現(xiàn)出良好性能。基于Hopfield自聯(lián)想記憶算法的神經(jīng)網(wǎng)絡(luò)已用于識別五種類似的醇的紅外光譜。反向傳播神經(jīng)網(wǎng)絡(luò)用于從混合氣體中識別目標氣體，證明了卷積神經(jīng)網(wǎng)絡(luò)（CNN）模型可以有效提高識別準確性。此外，最近的研究表明深度神經(jīng)網(wǎng)絡(luò)也可以應(yīng)用于振動光譜分析。卷積神經(jīng)網(wǎng)絡(luò)和自編碼器網(wǎng)絡(luò)被用于處理一維振動光譜數(shù)據(jù)。與傳統(tǒng)氣體檢測技術(shù)相比，輔以深度學(xué)習(xí)的氣體傳感器可以實現(xiàn)準確的靈敏度測量，并降低異常檢測的魯棒性。深度神經(jīng)網(wǎng)絡(luò)（DNN）可以在經(jīng)過足夠樣本訓(xùn)練后直接從吸收光譜中學(xué)習(xí)特征，實現(xiàn)不需要壓力校準和輪廓擬合的氣體濃度直接識別。這種網(wǎng)絡(luò)為檢索氣體濃度提供了一種新途徑，無需昂貴且復(fù)雜的壓力控制器。為了展示提出的DNN輔助算法的性能，構(gòu)建了一個基于DFB激光二極管的甲烷檢測氣體傳感器系統(tǒng)。預(yù)測的濃度與校準值相當吻合。這項研究表明，基于DNN的激光吸收光譜在大氣環(huán)境監(jiān)測、呼氣檢測等方面具有顯著潛力。

Currently, the rapid development of artificial intelligence provides a new way to solve this problem. The artificial neural network has been used for gas identification and shows a good performance under the condition of sufficient data for training. The infrared spectra of five similar alcohols has been identified by a neural network based on the Hopfield self-associative memory algorithm . A back propagation neural network is used to recognize target gas from the mixtures of gases, which proved that the convolutional neural networks (CNN) model can improve identification accuracy effectively. In addition, recent studies indicate that deep neural networks can also be applied to vibrational spectral analysis. The convolutional neural and auto encoder networks are used to process onedimensional vibrational spectroscopic data. Compared with traditional gas detection technology, the gas sensors assisted with deep learning can achieve accurate sensitivity measurement and reduce the robustness of anomaly detection.

A deep neural network (DNN), which can learn features directly from the absorption spectra after training with sufficient samples, achieves the direct identification of gas concentration free of pressure calibration and profile fitting. This network provides a new way to retrieve gas concentrations without expensive and complicated pressure controllers. To demonstrate the performance of proposed DNN assisted algorithm, a DFB diode laser-based gas sensor system for CH4 detection is constructed. The predicted concentrations are in good agreement with the calibrated values. This study indicates that DNN-based laser absorption  spectroscopy has remarkable potential in atmospheric environmental monitoring, exhaled breath detection and etc..

實驗裝置

用于獲取甲烷（CH₄）氣體吸收光譜的實驗裝置如圖1所示。一臺近紅外DFB激光二極管，最大峰值輸出功率為20毫瓦，被用作光源。通過控制激光溫度和電流，激光可以在6045 cm^-1到6047 cm^-1范圍內(nèi)進行調(diào)諧，寧波海爾欣光電科技有限公司為此項目提供激光驅(qū)動器，型號為QC-1000。所選CH₄在6046.95 cm^-1附近的吸收線在圖2中基于從HITRAN數(shù)據(jù)庫獲取的光譜線參數(shù)進行了模擬。DFB激光二極管經(jīng)過纖維準直器進行準直，然后由一塊CaF2分束器進行對準，分束后的可見紅光（632.8納米）光束用作跟蹤激光。隨后，光束被送入一個7米有效光程的多程傳輸池，并且池內(nèi)的壓力由壓力控制器、流量控制器和隔膜泵協(xié)同控制。一個典型頻率為100赫茲的三角波被用作掃描信號，以驅(qū)動激光二極管。最后，激光通過一個InGaAs光電探測器進行檢測，并被數(shù)據(jù)采集單元卡獲取。信號隨后傳輸?shù)接嬎銠C，并由自制的LabVIEW程序進行分析。

Experimental setup

The experimental setup used to obtain CH4 gas absorption spectra is depicted in Fig. 1. A near-infrared DFB diode laser with a maximum peak output power of 20 mW is used as the optical source. The laser can be tuned from 6045 cm^?1 to 6047 cm^?1 by controlling the laser temperature and current via the controller (QC-1000, Healthy photon Co., Ltd.). The absorption line of selected CH₄ near 6046.95 cm^?1 is simulated based on spectral line parameters obtained from the HITRAN database in Fig. 2. The DFB diode laser is collimated by a fiber collimator and aligned by a CaF2 beam splitter with a beam of visible red light (632.8 nm) as the tracking laser. Subsequently, the beam is sent to a multi-pass cell with a 7 m effective optical length, and the pressure inside the cell is collaborative controlled by a pressure controller, a flow controller, and a diaphragm pump. A triangular wave with a typical frequency of 100 Hz is used as a scanning signal to drive the diode laser. Finally, the laser is detected through an InGaAs photodetector and acquired by a data acquisition unit card. The signal is subsequently transmitted to the computer and analyzed by the homemade LabVIEW program.

 QC-1000, Healthy photon Co., Ltd.

Fig. 1. Experimental device diagram.

Fig. 2. The spectral line intensities of CH₄ in the tuning range of 6046.93–6046.96 cm^?1 and the cross-section of the selected line obtained from the HITRAN database.

結(jié)論

總體而言，本項目開發(fā)了基于DNN算法和激光吸收光譜的概念驗證氣體傳感器，并設(shè)計了基于DFB激光二極管的甲烷檢測傳感器系統(tǒng)。此外，通過計算RMSE和訓(xùn)練時間評估了DNN算法的性能，并優(yōu)化了DNN層、神經(jīng)元數(shù)量和epochs等參數(shù)，以獲取最佳參數(shù)。提出了改進的系統(tǒng)來分析和預(yù)測氣體吸收光譜數(shù)據(jù)，在甲烷濃度預(yù)測方面表現(xiàn)出良好的準確性和穩(wěn)定性。不同濃度的甲烷預(yù)測值與相應(yīng)的理論值線性擬合，證明其在實際領(lǐng)域應(yīng)用中具有巨大潛力，尤其適用于惡劣環(huán)境。

Conclusions

Overall, a proof-of-concept gas sensor based on the DNN algorithm and laser absorption spectroscopy is developed, and a CH4 detection sensor system based on the DFB diode laser is designed in this paper. In addition, the performance of the DNN algorithm is evaluated by calculating RMSE and training times, and the parameters, which include DNN layers, neuron number, and epochs, are optimized to obtain optimal parameters. The modified system is proposed to analyze and predict the gas absorption spectrum data, demonstrating good accuracy and stability in the prediction of CH4 concentrations. The predicted values of methane with different concentrations are linearly fitted with the corresponding theoretical value, which proves it has great potential in practical field applications, especially for harsh environments.

References

Pressure calibration- and profile fitting-free spectroscopy technology based on deep neural network for gas sensing, Measurement 204 (2022) 11207