One dimensional convolutional neural networks for improved training time and standardization in spectral classification

Physical Sciences Inc. (PSI) will develop a deep-learning based classification algorithm for detection and classification of trace explosives, opioids and narcotics on surfaces for optical spectroscopic systems. The sensor-customizable algorithm will be trained using a module consisting of a standard desktop CPU and GPU for accelerated training times. The algorithm will be deployable on a smaller, hardened operational module containing a single-board computer with low SWAP that can be integrated with a spectrometer system. The algorithm uses a one-dimensional convolutional neural network architecture (1D-CNN) that is trained using synthetic data produced by a data injector model to negate the need for a large data collection effort.The Phase II program would build off work completed in the Phase I program to extend the algorithms capabilities from infrared (IR) reflectance spectroscopy to include Raman spectroscopy. Feasibility of the algorithm wasestablished in the Phase I program though demonstrations of training models using synthetic IR data produced by the injector, and achieving classification accuracies greater than 90 percent against evaluation data sets comprised of real spectra and synthetic spectra. Identified training and operational module hardware demonstrated classification throughputs less than 3 ms spectra using the trained models. Based on these results, the Phase II operational module prototype is projected to have a classification accuracy of greater than 90 percent for spectral data with an average discriminant SNR greater than 1 and a classification throughput less than 3ms/spectra for spectral data with 600 wavelength channels.