摘要
针对传统模式识别算法对混合气体定性和定量检测准确率较低的问题,提出了一种基于机器学习的新型混合气体定性识别和浓度定量检测算法。算法首先构造传感器阵列数据特征图,然后利用卷积神经网络(CNN)提取特征,根据特征提取后的特征图,使用不同分支网络对不同气体进行定性识别,得到气体种类和相应气体所处浓度区间;根据前面的气体识别结果,使用核主成分分析(KPCA)与梯度提升树(GBDT)对混合气体的组成成分进行定量估计;最后采用加州大学机器学习数据库的动态混合气体气体传感器阵列数据集进行对比验证。实验结果表明,算法在乙烯和甲烷定性识别中准确率达到了98.7%,定量检测平均相对误差小于4.1%。通过与传统模式识别算法对比,所提出的基于机器学习的混合气体检测算法具有更高的精度和泛化能力。
In view of the low accuracy of the traditional pattern recognition algorithm for qualitative and quantitative detection of mixed gases, a novel algorithm of hybrid gas qualitative identification and concentration quantitative detection based on machine learning is proposed. The algorithm constructs the feature map of sensor array data first, then uses the convolutional neural network(CNN) to extract features from feature maps. According to the feature map after feature extraction, different branches are used to identify different gases, then the species of gases and their concentration range were obtained;based on the results of gas identification, the kernel principal component analysis(KPCA) and gradient boosting decision tree(GBDT) were used to estimate the composition of the mixed gases quantitatively. Finally, this paper used the dataset of sensor array of mixed gases of Machine Learning Database of the University of California to verify the results. Experimental results show that the accuracy of the algorithm in the qualitative recognition of ethylene and methane reaches 98.7% and the average relative error of quantitative detection was less than 4.1%. Compared with the traditional pattern recognition algorithm, the machine learning based mixed gas detection algorithm that proposed has higher accuracy and stronger generalization ability.
作者
谭光韬
张文文
王磊
Tan Guangtao;Zhang Wenwen;Wang Lei(Sino-German College,Tongji University,Shanghai 201804,China;College of Electronic and Information,Tongji University,Shanghai 201804,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2020年第7期95-102,共8页
Journal of Electronic Measurement and Instrumentation
基金
国家重点研发计划(2018YFE0105000)
国家重点研发计划(2017YFE0100900)
国家留学基金(201906260029)资助项目。
关键词
传感器阵列
卷积神经网络
核主成分分析
梯度提升树
sensor array
convolutional neural networks(CNN)
kernel principal component analysis(KPCA)
gradient boosting decision tree(GBDT)
作者简介
谭光韬,2017年于东华大学获得学士学位,现为同济大学硕士研究生,主要研究方向为模式识别与气体传感器。E-mail:gt_tan@tongji.edu.cn;张文文,现为同济大学博士研究生,主要研究方向为机器学习,智能信息处理与模式识别。E-mail:zhangwenwen_1203@163.com;王磊,现为同济大学教授,博士生导师,主要研究方向为传感器检测技术与测量系统。E-mail:leiwang@tongji.edu.cn。
