An Antibiotic Terahertz Spectrum Recognition Method Based on CNN and Attention-BiLSTM

Abstract

The terahertz spectrum has the characteristics of a fingerprint spectrum, which can realize the identification of antibiotics. Antibiotic identification based on traditional learning methods has achieved certain results. Models based on deep learning can automatically extract kernel features, and convolutional neural network (CNN) models rely on a convolution layer to extract features. In view of the high characteristic dimension of antibiotic spectrum sequences, we propose a new CNN model and attention bidirectional long short-term memory (BiLSTM). We use CNN to reduce the dimension of an antibiotic sequence. However, the convolution kernel limits a CNN's long-term dependence in processing time-series signal data. BiLSTM can effectively solve this problem and capture the dependence before and after the timing signal. The attention mechanism is added to BiLSTM to further extract the subtle features of the antibiotic spectrum, and can better capture the most important local information. A full connection network achieves the purpose of antibiotic identification. In experiments, the F1 score of the proposed model is 0.98, confirming its strong recognition ability and good interpretability.