Recognition of Mixture Vapors Using SERS Gas Sensor Fabricated by the Sputtering Method

Surface-enhanced Raman scattering (SERS) gas sensing is important in various fields, including environmental monitoring, health diagnosis, and food safety. We used a one-step sputtering method to fabricate an SERS gas sensor to detect and recognize volatile organic aromatic compound (VOAC) vapors using close-set and open-set models. Four pure VOAC vapors (acetophenone, anisole, anethole, and benzaldehyde) and mixtures of the compounds were analyzed using the SERS sensor. The normalized Raman spectra of the target compounds were preprocessed using the principal component analysis (PCA) and discrete wavelet transform (DWT) dimension-reduction methods before analysis. For closed-set recognition, three classical classifiers [naive Bayesian classifier (NBC), support vector machines (SVMs) classifier, and random forest classifier (RFC)] were used to identify three vapor datasets (normalized, PCA-, and DWT-based datasets). The best accuracy (99.4%) was achieved using the PCA-based dataset by SVM classifier. For open-set recognition, the dataset was first imported into a convolutional neural network (CNN) to extract features, and then, three open-set recognition models [Softmax, Openmax, and class anchor clustering (CAC)] were applied. The open-set classifier performances for different open problems were assessed using the dataset openness values of 0.0871, 0.1835, and 0.3333. CAC gave an area under receiver operating characteristic (AUROC) of 0.846 for discriminating between known and unknown samples and an accuracy of 94.5% for classifying the known vapor samples at an openness of 0.0871. The SERS sensor effectively detected vapor-phase compounds and could distinguish between the target compounds in both close-set and open-set recognition scenarios.