神经网络在超声催化气体单传感器的应用研究  

Application Study of Neural Networks in Ultrasonic Catalytic Gas Single Sensor

作　　者：张家铖 孙志峻[1] 张嘉亮 何永胜 ZHANG Jiacheng;SUN Zhijun;ZHANG Jialiang;HE Yongsheng(College of Aerospace Engineering,Nanjing University of Aeronautics and Astronautics)

机构地区：[1]南京航空航天大学航空学院

出　　处：《仪表技术与传感器》2024年第7期1-7,12,共8页Instrument Technique and Sensor

基　　金：国家自然科学基金项目(52275058)。

摘　　要：超声催化气体传感器是一种具备识别气体种类与分析气体浓度能力的新型单气体传感器,现有的数据库识别与K值判别方法存在校准工作量大与识别准确率不高的问题,因此提出一种基于神经网络的气体种类与浓度识别方法。结合实验采集的数据,利用随机森林(RF)与互信息(MI)等特征选择方法选择最优特征组合,设计神经网络结构,实现气体种类和浓度识别。实验表明:气体分类模型预测在1%~20%爆炸下限(LEL)浓度的甲醇、乙醇、丙酮和氢气分类正确率能达到96.88%,通过鲸鱼算法优化后的分类模型分类正确率提高至97.82%;气体浓度预测模型对甲醇、乙醇、丙酮和氢气浓度预测误差分别为3.49%、2.5%、10.12%、9.76%。结果表明,神经网络能有效进行超声催化气体单传感器的气体分类与浓度识别。An ultrasonic catalytic gas sensor is a novel single gas sensor for identifying gas types and analyzing concentration.Existing database recognition and K-value discrimination methods suffer from large calibration workloads and low recognition accuracy.Therefore,this paper proposed a neural network-based method for gas type and concentration recognition.Combining the data collected by the experiment,the optimal feature combination was chosen using methods like random forest(RF)and mutual information(MI).The neural network structure was designed to realize the gas type and concentration recognition.The experiment shows that the gas classification model achieves 96.88%accuracy in classifying methanol,ethanol,acetone and hydrogen at 1%to 20%lower explosion limit(LEL)concentration.Optimized by the whale algorithm,the classification model's accuracy improves to 97.82%.The gas concentration prediction model has prediction errors of 3.49%,2.5%,10.12%,and 9.76%for methanol,ethanol,acetone and hydrogen concentrations,respectively.The results show that the neural network can effectively perform gas classification and concentration recognition for ultrasonic catalytic gas single sensors.

关 键 词：电子鼻 超声催化 单传感器 气体识别 神经网络 智能优化 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置]