From Event: SPIE Defense + Commercial Sensing, 2023
A machine learning based approach has been developed to classify Raman spectroscopic data. The algorithm is based on a one dimensional neural network (1D-CNN) architecture which is trained with synthetic data that can incorporate sensor specific characteristics such as spectral range, spectral resolution and noise. The synthetic spectra are based on high SNR measurements which are then augmented by mixing target and background signatures. The CNN is trained to consider target representations in the presence of certain background materials including glass and HDPE. These additional target representations allow the CNN to make detections for materials taken through a container. Within this paper the performance of CNNs trained for Raman sensor systems has been evaluated using real data collected using the ThermoFisher FirstDefender. The evaluation data consists of various target chemicals (including explosives) and interferents (including household materials) collected through glass and plastic vials. The data was acquired with a controlled range of collection settings, including integration time and laser power, available on the unit. The performance of the 1D-CNN approach has demonstrated high classification accuracies, high probability of detection and low false alarm rates. Specifically, these metrics have been calculated as a function of signal to noise ratio. Additionally, a sensitivity analysis was conducted using an acetonitrile standard diluted in water which demonstrates the CNN’s capability of detecting all dilutions of acetonitrile down to weight concentrations of <1%. This sensitivity analysis was mirrored using a mixture of potassium chlorate and Vaseline. The CNN demonstrated detections down to 10% by weight of potassium chlorate.
© (2023) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Michael S. Primrose, Gabriel H. Weedon, and Jay Giblin, "One-dimensional convolutional neural networks for Raman spectral analysis," Proc. SPIE 12541, Chemical, Biological, Radiological, Nuclear, and Explosives (CBRNE) Sensing XXIV, 125410L (Presented at SPIE Defense + Commercial Sensing: May 03, 2023; Published: 14 June 2023); https://doi.org/10.1117/12.2664042.