Towards Real-Time Machine Learning for Anomaly Detection

We are developing a hardware platform based on FPGA with highly parallel neuromorphic computing architecture that can be utilized in a drone with radiation detector for anomaly detection. Driven by national security threat landscape, there is a need for detectors with readout circuits with low size, weight, and power consumption that are capable of implementing realtime machine learning. Anomalous radioactive source detection task has the limitation of far detecting distance, limited dwell time, and weak source activity, which introduces challenges to detect extremely sparse data dominated by background radiation. However, a detector capable of real-time learning can learn the background signatures and detect anomaly when it appears. In this work, we have performed a simulation study with measured background data and injected simulated source to identify the memory and number of neurons required for anomaly detection in the FPGA platform. The background data was collected with a Kromek D3S detector at the University of Illinois at Urbana-Champaign campus and source data simulated with hyperparameters specified by characteristics of the detector. The simulation result shows that for learning 2048 spectra, 239 neurons with 256 bytes is required. For 1024 spectra anomaly detection, we achieved 99.80% true positive rate and 0.08% false positive rate.