The application of wavelet transform of Raman spectra to facilitate transfer learning for gasoline detection and classification

Abstract

This study proposed an artificial intelligence (AI) approach for rapid, automatic detection of gasoline samples from different ignitable liquid samples based on their Raman spectra. The system integrated Raman spectroscopy, Raman signal transform, and transfer learning with a convolutional neural network (CNN). A hand-held Raman spectrometer was utilized to collect the learning dataset, including 180 gasoline spectra and 170 non-gasoline spectra from 17 various types of liquid samples. Continuous wavelet transform (CWT) was adopted to transform Raman spectra into image representations to facilitate transfer learning using a pre-trained CNN for image recognition. This approach streamlined AI model training, performance verification, and assessment. Experimental results showed that the CNN model could achieve 100% classification performance in precision, sensitivity, specificity, F1 score, and accuracy. The established AI model could be adopted for the Raman spectrum collected by different accessories, such as using the point-and-shoot attachment of liquid samples in glass vials. The experimental results suggested that CWT processing of Raman scattering signals enabled effective CNN transfer learning. The study demonstrated that CWT could be applied for Raman spectra processing to develop edge AI in forensic field testing of physical evidence.