基于ConvGRU融合注意力机制的工业混合气体智能识别方法

doi:10.12068/j.issn.1005-3026.2025.20240164

摘要/Abstract

摘要：

针对传统半导体气敏传感器在气体识别方面存在的数据依赖度高和混合气体识别精度不足的问题，本研究提出了1种融合门控循环单元（GRU）、卷积与注意力机制的ConvGRUAttention网络模型.采用经验小波变换（empirical wavelet transform，EWT）对原始信号进行时频域转换与多尺度分解，抑制噪声并降低数据依赖，提高了模型的鲁棒性.本模型通过卷积层提取局部动态特征，利用GRU捕捉信号的长期依赖，并引入注意力机制动态优化多尺度信号的特征权重，增强模型的特征提取和泛化能力.通过实验验证，定性识别准确率达到了100%，定量识别的均方根误差为3.3×10^-6.与传统方法相比，混合气体检测精度显著提高.

关键词: 半导体气敏传感器, 经验小波变换, 卷积门控循环注意力模型, 定性识别, 定量检测

Abstract:

To address the issue of high data dependency and insufficient accuracy in mixed gas identification for traditional semiconductor gas sensors， a ConvGRUAttention network model that integrates gated recurrent units （GRU）， convolutional layers， and attention mechanism is proposed. Empirical wavelet transform （EWT） is employed to convert raw signals into the time-frequency domain and perform multi-scale decomposition， which suppresses noise， reduces data dependency， and enhances the model’s robustness. The model extracts local dynamic features through convolutional layers， captures long-term dependencies using GRU， and optimizes feature weights across multi-scale signals via the attention mechanism， thereby improving feature extraction and generalization capabilities. Experimental results demonstrate 100% accuracy in qualitative identification and a root mean square error （RMSE） of 3.3×10⁻⁶ in quantitative detection. Compared with the traditional methods， the detection accuracy for mixed gases is significantly improved.

Key words: semiconductor gas sensor, empirical wavelet transform （EWT）, ConvGRUAttention model, qualitative identification, quantitative detection

中图分类号:

TP 249

孟凡利, 李书畅, 王浩, 苑振宇. 基于ConvGRU融合注意力机制的工业混合气体智能识别方法[J]. 东北大学学报（自然科学版）, 2025, 46(7): 37-48.

Fan-li MENG, Shu-chang LI, Hao WANG, Zhen-yu YUAN. Intelligent Identification Method of Industrial Mixed Gases Based on ConvGRU Fusion Attention Mechanism[J]. Journal of Northeastern University(Natural Science), 2025, 46(7): 37-48.

图/表 16

图1 NiO和NiO-Sn的XRD结果

Fig.1 XRD results of NiO and NiO-Sn

图2 气体智能测试平台

Fig.2 Intelligent gas testing platform

图3 阶梯下降加热波形

Fig.3 Step down heating waveform

图4 不同体积分数气体的响应曲线（a）—甲苯；（b）—二甲苯；（c）—二甲苯（10×10-6）-甲苯；（d）—二甲苯（30×10-6）-甲苯；（e）—二甲苯（70×10-6）-甲苯；（f）—二甲苯（100×10-6）-甲苯.

Fig.4 Response curves of gases with different concentrations

图5 甲苯、二甲苯及其混合气体的单边频谱及EWT分解图（a）—10×10-6甲苯的频谱；（b）—10×10-6二甲苯频谱；（c）—10×10-6甲苯-二甲苯混合后的频谱；（d）—甲苯的EWT分解结果；（e）—二甲苯的EWT分解结果；（f）—混合气体的EWT分解结果．

Fig.5 Single-side spectrum and EWT decomposition diagram of toluene, xylene, and their mixed gas

图6 ConvGRU单元结构

Fig.6 ConvGRU cell structure

图7 注意力机制计算流程

Fig.7 Calculation flow of the attention mechanism

表1 ConvGRUAttention模型参数

Table 1 ConvGRUAttention model parameters

模块名称	层次结构	卷积核	输入特征量	步长	填充	输出特征量	droput
ConvGRU	32个ConvGRU单元	3×3		3×3	1		0.1
ConvGRU	64个ConvGRU单元	3×3		3×3	1		0.1
Attention	一维平均池化层	1		1	1		0
	一维最大池化层	1		1	1		0
	线性层		64			4
	激活函数层	ReLu函数
	线性层		4			64
	激活函数层	ReLu函数
输出	一维平均池化层	16		16	1		0
	线性层		1 024			300
	线性层		300			3(类别)/2(浓度)
	激活函数层	ReLu函数

图8 定性识别的损失曲线与准确率曲线（a）—损失曲线；（b）—准确率曲线．

Fig.8 Loss and accuracy curves of qualitative identification

图9 验证集与测试集的混淆矩阵（a）—验证集；（b）—测试集．

Fig.9 Confusion matrices of the verification set and test set

图10 定量识别的结果

Fig.10 Results of quantitative identification

图11 ConvGRU和GRU定性识别的消融实验结果注：G代表GRU;CG代表ConvGRU.

Fig.11 Results of ablation experiments for ConvGRU and GRU qualitative identification

图12 3种模型消融实验的Kappa系数曲线

Fig.12 Kappa coefficient curves of ablation experiments of three models

表2 定性识别的消融实验结果 (identification)

Table 2 Ablation experiment results of qualitative

模型- 气体	准确率/%
模型- 气体	IMF1	IMF2	IMF3	IMF4	IMF5	IMF6
G-T	71.88	68.25	44.12	7.12	2.12	1.12
G-X	75.12	80.88	24.38	0.00	0.00	4.25
G-TX	90.72	90.00	96.78	9.81	100.00	98.41
CG-T	94.75	96.44	38.44	15.50	39.38	48.12
CG-X	86.22	86.88	39.75	0.00	5.81	71.69
CG-TX	95.30	95.70	95.38	97.48	91.33	88.16
CGA-T	92.31	96.81	44.56	12.75	49.31	41.44
CGA-X	86.25	87.62	47.44	0.00	39.56	71.00
CGA-TX	94.77	96.83	95.36	97.89	92.86	89.52

表3 各检测方法的定性比较结果 (detection method)

Table 3 Comparative qualitative results of each

模型	准确率/%			Kappa值
模型	甲苯	二甲苯	甲苯-二甲苯	Kappa值
SVM	87.50	87.50	90.62	0.796
CNN	70.88	85.38	96.59	0.799
GRU	80.00	76.12	94.03	0.768
TCN	87.50	62.50	93.75	0.750
EWTGRU	72.25	76.88	91.03	0.708
ConvGRU	96.00	85.50	97.06	0.898
Ours	100.00	100.00	100.00	1.000

表4 各检测方法的定量比较结果 (10-6)

Table 4 Comparative quantitative results of each detection method

模型	MAE(平均绝对误差)			RMSE（均方根误差）
模型	甲苯	二甲苯	甲苯-二甲苯	甲苯	二甲苯	甲苯-二甲苯
SVM	7.9	6.6	7.2	8.8	7.5	11.6
CNN	9.6	14.6	12.1	13.3	18.1	22.5
GRU	10.6	7.0	8.8	12.8	8.4	10.8
TCN	19.0	7.5	13.3	23.8	9.4	18.1
EWTGRU	8.3	4.9	6.6	10.1	7.5	8.9
ConvGRU	3.9	2.6	3.3	5.4	3.2	4.5
Ours	2.3	2.3	2.3	3.6	3.0	3.3

参考文献 37

[1]	Elansky N F， Ponomarev N A， Verevkin Y M. Air quality and pollutant emissions in the Moscow megacity in 2005—2014［J］. Atmospheric Environment， 2018， 175： 54-64.
[2]	Afrasiabi S， Darghlou M F， Mohammadi Y， et al. Effects of short and long-term exposure to benzene， toluene， ethylbenzene， and xylenes （BTEX） in indoor environment air on human health： a systematic review and meta-analysis［J］. Journal of Air Pollution and Health， 2024， 9（4）： 551-566.
[3]	Soltanpour Z， Mohammadian Y， Fakhri Y. The concentration of benzene， toluene， ethylbenzene， and xylene in ambient air of the gas stations in Iran： a systematic review and probabilistic health risk assessment［J］. Toxicology and Industrial Health， 2021， 37（3）： 134-141.
[4]	Edokpolo B， Yu Q J， Connell D. Health risk assessment of ambient air concentrations of benzene， toluene and xylene （BTX） in service station environments［J］. International Journal of Environmental Research and Public Health， 2014， 11（6）： 6354-6374.
[5]	Litra D， Chiriac M， Ababii N， et al. Acetone sensors based on Al-coated and Ni-doped copper oxide nanocrystalline thin films［J］. Sensors， 2024， 24（20）： 6550.
[6]	Huang L Q， Wu X， Li X B， et al. Influence of sensor array on MS/AE source location accuracy in rock mass［J］. Transactions of Nonferrous Metals Society of China， 2023， 33（1）： 254-274.
[7]	Ji H Y， Mi C J， Yuan Z Y， et al. Multicomponent gas detection method via dynamic temperature modulation measurements based on semiconductor gas sensor［J］. IEEE Transactions on Industrial Electronics， 2023， 70（6）： 6395-6404.
[8]	Liu F， Wu H Y， Zhao Y G， et al. Mapping high resolution national soil information grids of China［J］. Science Bulletin， 2022， 67（3）： 328-340.
[9]	Ji H Y， Yuan Z Y， Zhu H M， et al. Dynamic temperature modulation measurement of VOC gases based on SnO₂ gas sensor［J］. IEEE Sensors Journal， 2022， 22（15）： 14708-14716.
[10]	Gudiksen M S， Lauhon L J， Wang J F， et al. Growth of nanowire superlattice structures for nanoscale photonics and electronics［J］. Nature， 2002， 415（6872）： 617-620.
[11]	Rao Y J. In-fibre Bragg grating sensors［J］. Measurement Science and Technology， 1997， 8（4）： 355-375.
[12]	Le V H D. Method and apparatus for operating a gas sensor： US 3906473［P］. 1975-09-16.
[13]	Wu Z， Zhang H， Ji H， et al. Novel combined waveform temperature modulation method of NiO-In₂O₃ based gas sensor for measuring and identifying VOC gases［J］. Journal of Alloys and Compounds， 2022， 918： 165510.
[14]	Lee A P， Reedy B J. Temperature modulation in semiconductor gas sensing［J］. Sensors and Actuators B： Chemical， 1999， 60（1）： 35-42.
[15]	Cai L X， Miao G Y， Li G， et al. A temperature-modulated gas sensor based on CdO-decorated porous ZnO nanobelts for the recognizable detection of ethanol， propanol， and isopropanol［J］. IEEE Sensors Journal， 2021， 21（22）： 25590-25596.
[16]	Nakata S， Takahara N. Distinction of gaseous mixtures based on different cyclic temperature modulations［J］. Sensors and Actuators B： Chemical， 2022， 359： 131615.
[17]	Kadlec P， Gabrys B， Strandt S. Data-driven soft sensors in the process industry［J］. Computers & Chemical Engineering， 2009， 33（4）： 795-814.
[18]	Villa A， Benediktsson J A， Chanussot J， et al. Hyperspectral image classification with independent component discriminant analysis［J］. IEEE Transactions on Geoscience and Remote Sensing， 2011， 49（12）： 4865-4876.
[19]	Ji H Y， Liu Y， Zhang R Z， et al. Detection and recognition of toluene and butanone in mixture based on SnO₂ sensor via dynamic transient and steady-state response analysis in jump heating voltage mode［J］. Sensors and Actuators B： Chemical， 2023， 376： 132969.
[20]	Khurshid A， Pani A K. An integrated approach combining randomized kernel PCA， Gaussian mixture modeling and ICA for fault detection in non-linear processes［J］. Measurement Science and Technology， 2024， 35（7）： 076208.
[21]	Ji H Y， Zhu H M， Zhang R Z， et al. Semiconductor sensor virtual array： gas detection strategy in Internet of Things to suppress humidity interference［J］. IEEE Internet of Things Journal， 2024， 11（3）： 4934-4941.
[22]	Zhang S W， Yuan Z Y， Zhang W， et al. Strategy to expedite gas sensor calibration based on cyclic temperature operation and data augmentation［J］. Sensors and Actuators B： Chemical， 2024， 416： 136060.
[23]	Daubechies I. The wavelet transform， time-frequency localization and signal analysis［J］. IEEE Transactions on Information Theory， 1990， 36（5）： 961-1005.
[24]	Huang N E， Shen Z， Long S R， et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis［J］. Proceedings of the Royal Society of London，Series A： Mathematical， Physical and Engineering Sciences， 1998， 454（1971）： 903-995.
[25]	Gilles J. Empirical wavelet transform［J］. IEEE Transactions on Signal Processing， 2013， 61（16）： 3999-4010.
[26]	Hui A L， Yu J， Lu W Q. Research on rotor fault diagnosis technology of three-phase asynchronous motor based on NA-MEMD mutual information and SVM［J］. Measurement and Control， 2025， 58（1）： 110-118.
[27]	Liu T， Li J， Cai X F， et al. A time-frequency analysis algorithm for ultrasonic waves generating from a debonding defect by using empirical wavelet transform［J］. Applied Acoustics， 2018， 131： 16-27.
[28]	Al-Musaylh M S， Al-Daffaie K， Prasad R. Gas consumption demand forecasting with empirical wavelet transform based machine learning model： a case study［J］. International Journal of Energy Research， 2021， 45（10）： 15124-15138.
[29]	Cao X G， Guo X Y， Duan Y， et al. Health status recognition method for rotating machinery based on multi-scale hybrid features and improved convolutional neural networks［J］. Sensors， 2023， 23（12）： 5688.
[30]	de Bruin T， Verbert K， Babuška R. Railway track circuit fault diagnosis using recurrent neural networks［J］. IEEE Transactions on Neural Networks and Learning Systems， 2017， 28（3）： 523-533.
[31]	Liu L Z， Hu C X， Ou T S， et al. Selective detection of mixtures via a single nonselective sensor： making the unworkable sensor workable by machine learning［J］. Advanced Intelligent Systems， 2023， 5（1）： 2200136.
[32]	Vlachos D， Avaritsiotis J. Fuzzy neural networks for gas sensing［J］. Sensors and Actuators B： Chemical， 1996， 33（1/2/3）： 77-82.
[33]	Karagulian F， Barbiere M， Kotsev A， et al. Review of the performance of low-cost sensors for air quality monitoring［J］. Atmosphere， 2019， 10（9）： 506.
[34]	Yang M， Li J， Liu J C. Forecast model for gas well productivity based on PSO and SVM［J］. Applied Mechanics and Materials， 2011， 71/72/73/74/75/76/77/78： 1915-1919.
[35]	Kang M G， Cho I， Park J， et al. High accuracy real-time multi-gas identification by a batch-uniform gas sensor array and deep learning algorithm［J］. ACS Sensors， 2022， 7（2）： 430-440.
[36]	Jia P T， Liu H D， Wang S J， et al. Research on a mine gas concentration forecasting model based on a GRU network［J］. IEEE Access， 2020， 8： 38023-38031.
[37]	Ma X， Wu F， Yan J， et al. TF-TCN： a time-frequency combined gas concentration prediction model for E-nose data［J］. Sensors and Actuators A： Physical， 2024， 376： 115654.