Anomaly Detection using 1D Convolutional Neural Networks for Surface Enhanced Raman Scattering

An accurate supervised classification technique requires a large training database with an equal number of samples in each category. However, in practice, data class imbalance is naturally inherent in detection and identification tasks. In an extreme case, one category of data has a majority of training samples (positive class), causing over-classifying. In these circumstances, the negative classes are either absent, poorly sampled or not well defined. Deep one-class classifiers are artificial neural networks developed to overfit the positive class samples. This unique situation constrains the network model to be trained data features just with the knowledge of the positive class. One well-known application of one-class classifiers is for anomaly detection problem, where the model stands out outliers. Recently, convolutional neural networks (CNNs) have outperformed previous machine learning methods in pattern recognition tasks. In this study, we proposed using a one-dimensional CNN model for anomaly detection in surface-enhanced Raman scattering (SERS) data acquired by portable Raman spectrometers. Raman spectroscopy technique has been widely adopted by first responders and military forces for the field analysis and identification of unknown hazardous materials. The performance and accuracy of the recognition method might compromise the success rate of an interrogation operation. Our experimental results revealed that a 1D CNN model could be used as a one-class classifier to distinguish anomalies in SERS data with a successful detection rate of 100 percent.