人体呼气中的丙酮含量可作为糖尿病的标志物.为实现无创糖尿病诊断,设计以金属氧化物半导体气敏传感器阵列为核心的人工嗅觉系统,对完成痕量丙酮的快速检测具有重要意义.通过多个气体流量控制器MFC(Mass Flow Controller)分别配制出模拟糖尿病患者呼气样本(30×10-6丙酮)与另两种干扰气体样本(30×10-6乙醇样本、15×10-6丙酮与15×10-6乙醇混合样本)进行实验,基于BP神经网络算法对3种气体定性识别,并通过主成分分析PCA(Principal Component Analysis)算法对原始的高维特征子集进行降维优化.实验表明:PCA与BP算法相结合,可降低BP神经网络的复杂性、减少预测的误差,同时能够解决单个气体传感器交叉敏感问题,进而提高对气体的选择性.对痕量丙酮样本与另两种干扰气体样本进行分析识别,识别的结果显示:对3种样本的识别准确率为91％.该研究为准确识别糖尿病标志物实现无创诊断技术提供了理论指导.%The content of acetone in human exhalation can be used as a marker of diabetes. In order to achieve the noninvasive diagnosis of diabetes,we design a metal oxide semiconductor gas sensors array as the core of the artifi-cial olfactory system,,which is of great significance in rapid detection of trace acetone. Through multiple Mass Flow Controller(MFC),we prepared simulated diabetic patients breath samples(30×10-6 acetone)and two interference gas samples(30×10-6 ethanol/composition of 15×10-6 acetone and 15×10-6 ethanol),Three kinds of gas qualitative identifications were carried out based on BP neural network algorithm and optimization of the original high dimen-sional feature subset was achieved through PCA algorithm. The experiment shows that the combination of PCA and BP algorithm can reduce the complexity of BP neural network and reduce the error of prediction. At the same time, the cross sensitivity of individual gas sensors can be solved,thus improve the selectivity of gas d. The identification results of trace acetone and the two interference gas samples show that the accuracy of the recognition of the three gases reaches 91％. This study provides theoretical guidance for accurate identification of diabetes markers and non-invasive diagnosis.
展开▼
