Inference detection and classification of illicit drugs by a modest Raman spectrometer with a convolutional neural network analyzer

Faced with widespread of emerging new psychoactive substances (NPS) and illicit drugs, real-time and on-site detection techniques are urgently in need but remain a challenge. For this purpose, we present a simple but accurate in-house developed portable Raman spectral imaging system. The system consists of a compact, costeffective and redesigned Raman spectrometer and a responsive analyzer for chemical compound identification based on convolutional neural network (CNN) techniques. To test the system, twelve substances are selected and divided into the library (established in database) and the challenge (unknown) groups. The challenge validation shows that the system achieves excellent predictive accuracy and high sensitivity even on low spectral resolution. The illicit drugs in the challenge set are identified as suspicious illicit compounds at an accuracy rate of about 92%. This accuracy results from that the spectral signatures of the functional groups or molecular structural similarities of the chemical compounds are recognized by the CNN model.