Chu Jifeng, Yang Aijun, Wang Qiongyuan, Yang Xu, Wang Dawei, Wang Xiaohua, Yuan Huan, Rong Mingzhe
State Key Laboratory of Electrical Insulation and Power Equipment, Xi'an Jiaotong University, 710049 Xi'an, China.
Microsyst Nanoeng. 2021 Mar 1;7:18. doi: 10.1038/s41378-021-00246-1. eCollection 2021.
A difficult issue restricting the development of gas sensors is multicomponent recognition. Herein, a gas-sensing (GS) microchip loaded with three gas-sensitive materials was fabricated via a micromachining technique. Then, a portable gas detection system was built to collect the signals of the chip under various decomposition products of sulfur hexafluoride (SF). Through a stacked denoising autoencoder (SDAE), a total of five high-level features could be extracted from the original signals. Combined with machine learning algorithms, the accurate classification of 47 simulants was realized, and 5-fold cross-validation proved the reliability. To investigate the generalization ability, 30 sets of examinations for testing unknown gases were performed. The results indicated that SDAE-based models exhibit better generalization performance than PCA-based models, regardless of the magnitude of noise. In addition, hypothesis testing was introduced to check the significant differences of various models, and the bagging-based back propagation neural network with SDAE exhibits superior performance at 95% confidence.
限制气体传感器发展的一个难题是多组分识别。在此，通过微加工技术制造了一种装载三种气敏材料的气敏（GS）微芯片。然后，构建了一个便携式气体检测系统，以收集芯片在六氟化硫（SF）各种分解产物下的信号。通过堆叠去噪自动编码器（SDAE），可以从原始信号中总共提取五个高级特征。结合机器学习算法，实现了对47种模拟物的准确分类，五折交叉验证证明了其可靠性。为了研究泛化能力，进行了30组测试未知气体的实验。结果表明，无论噪声大小，基于SDAE的模型都比基于主成分分析（PCA）的模型表现出更好的泛化性能。此外，引入假设检验来检查各种模型的显著差异，具有SDAE的基于装袋法的反向传播神经网络在95%置信度下表现出卓越性能。
