Big Data Tools for High-speed Threat Detection and Classification

We are developing a deep-learning approach for detecting and classifying torpedo-like threats present in passive sonar data. Interference from merchant ships and biologics often obscure threat signals evident on sonar display surfaces. Experienced sonar operators can distinguish signals of interest (SOIs) such as torpedoes and rogue surface craft from clutter in beam-level displays, but this manual search leads to high operator workloads and unacceptably long detect-to-engage timelines. To address this problem, Black River Systems will develop FRONT ROW: Fast Recognition Of Naval Threats for Reducing Operator Workload for the Navy’s AN/SQQ-89 undersea warfare system. FRONT ROW will automatically detect and classify potential threats and call up the beam-level displays of interest for the operator to review. Our approach uses a hybrid convolutional neural network (H-CNN) to classify short (e.g. 0.5 sec) frames of beam-level spectrogram data followed by a sequential Bayesian algorithm to fuse the H-CNN outputs over time. Compared to full-frame techniques, FRONT ROW’s short frame-based approach will utilize a simpler network, require less training data, and readily implement into existing Navy systems. FRONT ROW will deliver >70% correct classification with <1 false alert per hour in a semi-cluttered environment.