【作者】 王雪；

【导师】 王殊；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2019， 硕士

【摘要】 气体检测技术在工业与生活中都发挥着重要作用,声学气体检测方法相比于现有的光吸收谱、电化学等气体检测方法有价格、实时、耐久等优势,已成为新一代气体传感领域的研究热点。深度学习方法能弥补从声谱到气体成分理论推导上的不足,加速声学弛豫气体传感技术的实际用。本文首先分析基于声弛豫理论的气体声谱重建算法,通过仿真得到气体声吸收谱和声速谱,利用声谱分析声弛豫气体检测的正确性。参考现有相关理论和案例设计适用于声谱分类的一维卷积神经网络模型,并且选择合适的激活函数和优化方法。然后以声弛豫理论和深度学习方法为指导,在Django框架下构建声弛豫气体探测仿真实验系统,实现跨平台、交互式声谱仿真,用TensorFlow.js实现在浏览器端训练深度学习模型。最后在仿真实验系统上进行声学气体检测,并通过平均影响值分析提取可用于气体识别的声谱特征,以及讨论模型给出预测结果的合理性。结合气体特性将检测方法改进为多温度检测,进一步提高检测准确率。结论:本文在声弛豫气体探测仿真实验系统用一维卷积神经网络进行气体检测,CO₂、CH₄、H₂、N₂和Cl₂五种气体单温度定量检测中都能有80%以上的检测准确率,平均影响值分析的结果验证了卷积神经网络有提取声谱拐点特征的能力。实验发现使用温度变化后的声谱增量能提升检测准确率,并且在温度间隔10K以上时效果更明显。据此,改进方案为多温度检测后强弛豫气体的定量检测准确率都能达到90%以上,弱弛豫气体N₂的检测准确率也能达到87%。本文的研究成果为基于声弛豫参数“温度变化率”的定性气体探测方法奠定了仿真预研基础,提供了有别于传统模型推导的思路:利用机器学习方法寻找和验证可用于气体探测的声弛豫特征参数。同时,本文建立的声弛豫气体探测仿真实验系统提供了声弛豫气体探测的开放实验平台。更多还原

【Abstract】 Gas detection plays an important role in industry and our daily life.Acoustic gas detection method has obvious advantages in price,instantaneity,and durability compared with the existing methods based on electrochemical or optical theory.It has become the hotspot in next generation of gas sensing.Deep learning can make up for the insufficiency in theoretical deduction from sound spectrum to gas composition,and accelerate the application of acoustic gas sensing technology.Firstly,this paper analyzed gas acoustic spectrum reconstruction algorithm based on acoustic relaxation theory.The sound absorption spectrum and sound velocity spectrum for a certain gas can be obtained by simulation,and the correctness of acoustic relaxation gas detection is analyzed.A one-dimensional convolutional neural network model with proper activation functions and optimization method suitable for acoustic spectrum classification was designed.Then,guided by acoustic theory and deep learning method,under the Django framework,an acoustic relaxation gas detection simulation experimental system was built to realize cross-platform and interactive sound spectrum simulation,and TensorFlow.js implemented the visualization of deep learning training process on the browser side.Finally,the mean impact value method is used to extract the spectral characteristics which can be used for gas component identification and proving the rationality of the predicted results.According to gas characteristics,the detection method is improved to multi-temperature detection,so as to improve the detection accuracy.Conclusion:In this paper,one-dimensional convolutional neural network performs well in gas detection.In the quantitative detection of several common gases,CO₂,CH₄,H₂,N₂ and Cl₂,the detection accuracy can be more than 80%.The improved scheme of multi-temperature detection can achieve more than 90%detection accuracy for strong relaxation gases.And for the weak relaxation gas N₂,the accuracy reached 87%.When temperature interval greater than 10K,the improvement is more obvious.The mean impact value method verified that the convolutional neural network has the ability to extract the characteristics of sound spectrum.The result of this paper laid foundation for gas detection based on acoustic relaxation parameter“rate change to temperature”.This paper provided an new method different from conventional module derivation:searching for acoustic relaxation characteristic parameters by means of deep learning.Besides,the acoustic relaxation gas detection simulation experimental system offers an open platform for gas detection.更多还原