模拟退火优化的径向量核支持向量回归算法在人工嗅觉系统的应用  

Application of Radial Basis Function Kernel Support Vector Regression Optimized by Simulated Annealing Algorithm in Electronic Nose System

作　　者：汪雨晴 赵庆贺 WANG Yu-qing;ZHAO Qing-he(Shenyang Academy of Instrumentation Science Co.,Ltd.,Shenyang 110043,China;College of Electrical Engineering and Information,Northeast Agricultural University,Harbin 150030,China)

机构地区：[1]沈阳仪表科学研究院有限公司,辽宁沈阳110043 [2]东北农业大学电气与信息学院,黑龙江哈尔滨150030

出　　处：《仪表技术与传感器》2022年第7期111-116,共6页Instrument Technique and Sensor

摘　　要：提出了一种结合了机器学习和模拟退火算法的建模方法,有效解决了检测非甲烷总烃(NMHC)的人工嗅觉系统在复杂环境下的气体干扰和传感器漂移问题。通过采用多通道仪器响应构成多维度数据组抑制复杂气体环境对NMHC传感器的干扰,数据中进一步结合记忆时序消除随时间变化的传感器漂移误差。针对设计的多维度记忆数据结构,设计了带有径向量核函数的支持向量回归对多维数据进行建模得到拟合模型,并且通过模拟退火方法完成超参数优化得到sa-rbf-svr模型。通过130组以1 h为间隔的数据样本训练后,在相隔1 d后的独立24个样本中完成了和线性模型的对比实验,结果表明:带有记忆的多维度数据结合sa-rbf-svr模型可将R~2从0.924 8提升至0.984 1,MAE指标从41.057 5提升至15.244 4。In order to solve the problem of gas interference and sensor drift of detecting non⁃methane hydrocarbons(NMHC)in complex environment by the electronic nose system,a modeling method was proposed combining with machine learning and simulated annealing algorithms.Multi⁃channel instrument response formed a multi⁃dimensional data to suppress the interference of NMHC sensor from the complex environment.Memory⁃sequence was further combined to eliminate the drift error of the sensor over time.For the designed multidimensional memory data structure,a support vector regression with radial kernel function was designed to fit the multi⁃dimensional data,the sa⁃rbf⁃svr models were gotten through the simulated annealing algorithms optimized by hyperparameter.Through training of 130 groups of data samples with hourly intervals,twenty⁃four independent samples were compared with the linear model after one day.The results show the multi⁃dimensional data with memory combined with sa⁃rbf⁃svr model can improve R2 from 0.9248 to 0.9841,and improve MAE from 41.0575 to 15.2444.

关 键 词：人工嗅觉 支持向量机 模拟退火 

分 类 号：TH701[机械工程—仪器科学与技术]