Rapid Detection of Trace Nitro-Explosives under UV Irradiation by Electronic Nose with Neural Networks

Abstract

The development of an electronic nose (E-nose) for rapid explosive trace detection (ETD) has been extensively studied. However, the extremely low saturated vapor pressure of explosives becomes the major obstacle for E-nose to be applied in practical environments. In this work, we innovatively combine the decomposition characteristics of nitro explosives when exposed to ultraviolet light into gas sensors for detecting explosives, and an Enose consisting of a SnO 2 /WO 3 nanocomposite-based chemiresistive sensor array with an artificial neural network is utilized to identify trace nitro-explosives by detecting their photolysis gas products, rather than the explosive molecules themselves or their saturated vapor. The ultralow detection limits for nitro-explosives can be achieved, and the detection limits toward three representative nitro-explosives of trinitrotoluene, pentaerythritol tetranitrate, and cyclotetramethylene tetranitroamine are as low as 500, 100, and 50 ng, respectively. Moreover, by extracting the features of sensor responses within 15 s, a classification system based on convolutional neural network (CNN) and long short-term memory network (LSTM) is introduced to realize fast and accurate classification. The 5-fold cross-validation results demonstrate that the CNN-LSTM model exhibits the highest classification accuracy of 97.7% compared with those of common classification models. This work realizes the detection of explosives photolysis gases using sensor technology, which provides a unique insight for the classification of trace explosives.

Ultrasensitive Online NO Sensor Based on a Distributed Parallel Self-Regulating Neural Network and Ultraviolet Differential Optical Absorption Spectroscopy for Exhaled Breath Diagnosis

“…Gas sensors using neural network analysis have been widely used to detect gases at the ppb level. − The convolutional neural network (CNN) is a widely used deep learning model, and its special hierarchical structure can be well adapted to the problem of spectral drift in gas spectral data compared with the artificial neural network (ANN). Meanwhile, it can effectively improve the efficiency and accuracy of detection. − Some researchers in their studies have trained the preprocessed absorption spectra as a whole through CNN models. Meanwhile, different absorption characteristic regions of the spectrum contributing differently to the concentration prediction are not considered. − To better combine deep learning technology and spectroscopy technology, in this paper, we construct a distributed parallel self-regulating neural network (DPSRNN) model structure based on the absorption characteristics of UV absorption spectroscopy.…”