Nanowire-Based Sensor Array for Detection of Cross-Sensitive Gases Using PCA and Machine Learning Algorithms
Nanowire-Based Sensor Array for Detection of Cross-Sensitive Gases Using PCA and Machine Learning Algorithms
复制标题
基于PCA和机器学习算法的纳米线传感器阵列交叉敏感气体检测
DOI：
10.1109/jsen.2020.2972542
复制
发表时间：
2020-06-01
影响因子：
4.3
通讯作者：
Rao, Mulpuri, V
中科院分区：
文献类型：
--
作者：
Khan, Md Ashfaque Hossain;Thomson, Brian;Rao, Mulpuri, V
在这项工作中，采用行业标准自上而下的制造方法设计和开发了一种气体传感器阵列，该阵列由 Pt、Cu 和 Ag 装饰的 TiO2 和 ZnO 功能化 GaN 纳米线组成。阵列内的受体金属/金属氧化物组合是根据我们之前的分子模拟结果使用基于密度泛函理论（DFT）的第一原理计算确定的。在室温下，在 H2O 和 O-2 气体存在下，在紫外光下收集 NO2、乙醇、SO2 和 H-2 的单一气体和混合物的气体传感数据。每种气体在阵列内的传感器上产生独特的响应模式，通过这种模式可以精确识别交叉敏感气体。对原始数据进行预处理后，对阵列响应应用无监督主成分分析（PCA）技术。结果发现，每种分析气体在所有目标气体及其混合物的得分图中形成一个单独的簇，表明它们之间有明显的区别。然后，使用决策树、支持向量机 (SVM)、朴素贝叶斯（内核）和 k-最近邻 (k-NN) 等四种监督机器学习算法，使用其重要参数和我们的阵列数据集进行训练和优化，以进行气体类型分类。结果表明，优化后的 SVM 和 NB 分类器模型在测试数据集上表现出 100% 的分类准确率。还讨论了所考虑算法的实际适用性。此外，与高功耗的商用金属氧化物传感器相比，该阵列设备在室温下工作，使用非常低功耗和低成本的紫外发光二极管（LED）。
In this work, a gas sensor array has been designed and developed comprising of Pt, Cu and Ag decorated TiO2 and ZnO functionalized GaN nanowires using industry standard top-down fabrication approach. The receptor metal/metal-oxide combinations within the array have been determined from our prior molecular simulation results using first principle calculations based on density functional theory (DFT). The gas sensing data was collected for both singular and mixture of NO2, ethanol, SO2 and H-2 in presence of H2O and O-2 gases under UV light at room temperature. Each gas produced a unique response pattern across the sensors within the array by which precise identification of cross-sensitive gases is possible. After pre-processing of raw data, unsupervised principal component analysis (PCA) technique was applied on the array response. It is found that, each analyte gas forms a separate cluster in the score plot for all the target gases and their mixtures, indicating a clear discrimination among them. Then, four supervised machine learning algorithms such as- Decision Tree, Support Vector Machine (SVM), Naive Bayes (kernel) and k-Nearest Neighbor (k-NN) were trained and optimized using their significant parameters with our array dataset for the classification of gas type. Results indicate that the optimized SVM and NB classifier models exhibited 100% classification accuracy on test dataset. Practical applicability of the considered algorithms has been discussed as well. Moreover, this array device works at room-temperature using very low power and low-cost UV light-emitting diode (LED) as compared to high power consuming commercially available metal-oxide sensors.