Anomaly detection in radiation signals using kernel machine intelligence

Acquisition of consecutive ultra-short interval measurements with a radiation detector serves in detecting anomalies in the ambient environment, potentially caused by hidden radioactive material. An anomaly is defined as the contribution of a man-made radiation source to the measured signal. In this paper a new methodology is presented and tested on detecting anomalies in various radiation signals. The methodology employs machine intelligence tools to perform automated detection of malignant source contribution. In particular, a Gaussian process (GP) modeled as a function of a learning kernel is used for post-processing of measured signals. For detection purposes, a window that slides over the measured signal is divided into three parts, where the two parts are used for training the GP that is subsequently used for predicting the values of the third part. Next, a comparison between predicted and measured values in the third part is conducted by computing the mean average percentage error (MAPE). If MAPE is relatively high, then the likelihood of an anomaly in the part under examination is respectively high. Adoption of a sliding window ensures that no part of signal will remain unchecked. Testing results obtained on a set of gamma-ray signals exhibit the potency of the presented methodology in effectively detecting anomalies.