ScholarWorks@경상국립대학교

초록

This paper presents a highly reliable biological detection method to improve the field performance of chemical and biological mass spectrometry. Existing algorithms, which rely on a few pre-selected biomarkers, are susceptible to false alarms owing to external interference in real-world environments. Our approach leverages a deep-learning-based classification algorithm, specifically the ResNet-50 architecture, which is trained to recognize the entire spectral pattern from ion mass spectrometry data. Datasets were secured from both laboratory and field environments, including samples of bacteria, spores, biological toxins, and a “no agent injected” control. Data augmentation techniques were developed to simulate device malfunctions and enhance the robustness of the algorithm. The trained model was deployed on a military-grade single-board computer , which achieved real-time analysis within 5 s. Field testing on new and unseen data resulted in a classification accuracy of 99.348%, thus demonstrating a significant improvement over conventional biomarker-based methods. Additionally, the LIME technique was utilized to provide interpretability, which showed that the model focused on specific biomarker regions for classification. This study provides a robust and high-performance solution for detecting biological agents under challenging field conditions.

키워드