基于卷积神经网络的气体传感器阵列识别算法研究及应用  

Research and Application on Recognition Algorithm Based on Gas Sensor Array and Convolutional Neural Network

作　　者：李震 王秀玲 王传玺 罗志华 王雪 董玉华 孙炎辉 LI Zhen;WANG Xiu-ling;WANG Chuan-xi;LUO Zhi-hua;WANG Xue;DONG Yu-hua;SUN Yan-hui(School of Information and Communication Engineering,Dalian Minzu University,Dalian Liaoning 116605,China;Kede Numerical Control Co.Ltd,Dalian Liaoning 116600,China)

机构地区：[1]大连民族大学信息与通信工程学院,辽宁大连116605 [2]科德数控股份有限公司,辽宁大连116600

出　　处：《大连民族大学学报》2023年第5期431-436,共6页Journal of Dalian Minzu University

基　　金：辽宁省自然科学基金项目(2022-BS-102)。

摘　　要：为解决混合气氛中气体浓度识别问题,常利用气体传感器阵列配合模式识别算法进行检测。设计了基于嵌入式处理器的传感器阵列,并利用识别算法对设备采集的混合气体进行分类识别及浓度预测。建立了以氨气、丙酮、甲醇气体为目标的混合气体数据集。使用最邻近分类算法(KNN)、三层全连接反向神经网络(BPNN)和三层卷积神经网络(CNN)分别对混合气体中的氨气、丙酮、甲醇气体进行识别分析。测试结果表明:改进的BPNN和CNN对测试数据集的分类识别率最高均可达100%,对混合气体的浓度预测均方差最低可达3.89和2.47,三层卷积层的CNN算法相对于BPNN和KNN在识别精度上提高明显。通过迁移学习思想,将该算法移植到树莓派中,并进行实际测试,实现了电子鼻应用。In order to solve the problem of gas concentration identification in mixed atmosphere,gas sensor arrays and pattern recognition algorithms are often used for detection.A sensor array based on the embedded processor is designed,and a recognition algorithm is used to classify and predict the gas mixture collected by the device.The mixed gas dataset targeting ammonia,acetone and methanol gas is established.The nearest neighbor classification algorithm(KNN),three-layer fully connected back propagation neural network(BPNN),and three-layer convolutional neural network(CNN)are used to identify and analyze ammonia,acetone,and methanol gases in the gas mixture.The test results show that the improved BPNN and CNN have the highest classification recognition rate of 100%for the test dataset,and the lowest mean square error of concentration prediction of mixed gas can reach 3.89 and 2.47.Compared with BPNN and KNN,the CNN with three convolutional layers has a significant improvement in recognition accuracy.Through transfer learning,the algorithm is transplanted into Raspberry PI and tested in practice to realize the application of electronic noses.

关 键 词：混合气体 全连接反向神经网络 卷积神经网络 气体传感器阵列 

分 类 号：TP212.2[自动化与计算机技术—检测技术与自动化装置]