An Unsupervised Radiation Anomaly Detection Algorithm Using a Deep Neural Autoencoder

The automated detection and characterization of anomalous radiological signatures is of great importance for persistent radiation detection systems deployed at border crossings and other points of radiological surveillance. Because of the high variability in background radiation and anomalous source signatures, machine learning methods are often employed with the goal of developing anomaly detection algorithms capable of learning from the collected data. An unsupervised radiological anomaly detection algorithm has been developed and tested on a network of persistent radiological sensors deployed around the High Flux Isotope Reactor at Oak Ridge National Laboratory. This algorithm uses a deep autoencoder neural network that generates a reduced encoding of background gamma radiation spectra to learn a unique representation of the background spectra being collected. Autoencoder neural networks work by using dimensionality reduction to generate a smaller representation (encoding) of the input data, such that the autoencoder is then able to reconstruct (decode) the input data from the reduced encoding. After training on a set of input data, the autoencoder’s decoding accuracy is then used to separate typical spectra from spectra containing anomalous radiation signatures, thereby triggering or suppressing anomaly detection alarms. As an unsupervised learning method, this algorithm does not require a labeled training set and can continue to learn throughout the duration of its deployment in the field.