Antibiotic SERS spectral analysis based on data augmentation and attention mechanism strategy

The analysis of Raman spectrum data has gradually transitioned into the era of machine learning. However, it is still constrained by the challenge of acquiring large volumes of raw data and the issue of losing characteristic information from spectral data. In this paper, we propose a strategy that combines data amplification and attention mechanisms for analyzing antibiotic spectral data. Firstly, a Generative Adversarial Network was employed to amplify the SERS spectrum of eight antibiotics by 10 times, to augment the dataset to fulfill the requirements of the neural network. Then, the amplified data is input into a one-dimensional convolutional neural network with an attentional mechanism module, which enables a more accurate capture of spectral feature information. The one-dimensional convolutional neural network achieved a 97.5% accuracy in classifying eight antibiotics. The accuracy of the four mixtures within the same class was 89.4%.