IED pattern recognition using sinusoidal models

Abstract

More than half of all U.S. casualties in Iraq and Afghanistan were caused by improvised explosive devices (IEDs). Despite the spending of over $75 billion to combat this threat, intelligence analysts still lack efficient tools to conduct IED pattern analysis. This thesis evaluates sinusoidal models for effectiveness in assisting in the identification of IED patterns. We formulate three models to test against IED patterns encountered in Iraq and Afghanistan: the Hawkes point process, the non-linear optimization of a sine function, and discrete Fourier transforms (DFT). Non-linear optimization and DFT models both out-perform a mean inter-arrival model when applied to representative IED patterns. We also applied these models against portions of an Iraq IED dataset using a rolling horizon forecast. Lastly, we test model performance when applied to patterns identified from the Iraq dataset. We conclude that although there is not a silver bullet for IED pattern detection, the use of these models in IED environments has the potential to reduce the amount of time and effort intelligence analysts expend when identifying IED patterns. We recommend incorporating these models into a graphic user interface usable by intelligence analysts responsible for IED pattern recognition.