Context- and AI-Based Reasoning for Identification onBOard UUVs (CARIBOU) is a comprehensive, AI-based automatic target recognition (ATR) agent paired with a fundamentally novel reasoning engine. CARIBOU is designed to provide accurate, robust target recognition capabilities that are not easily fooled by novel classes, dim signals, or attempts to deceive, without requiring a human-in-the-loop or burdensome size, weight, and power (SWaP) requirements. The CARIBOU team has already completed some studies under IRAD funding which provide a basis of confidence for the proposed work, demonstrating that off-the-shelf deep neural networks (DNNs) are able to recognize sonar signals at high accuracy. In Phase I, we will expand our work into a full proof-of-concept that demonstrates not only target recognition, but also an ability to reason over the results. This reasoning will improve classification accuracy and identify irregularities such as results with high-uncertainty or indicators of attempts to deceive. Phase I will focus on passive sonar; future phases will consider additional modalities and will include in-water testing using our facilities. CARIBOU is a joint effort between Boston Fusion Corp, which brings expertise in machine learning research for DoD applications, and Advanced Acoustic Concepts, which brings expertise in sonar processing, acoustic modeling, and UxV development.
Benefit: CARIBOU will find immediate application as a tool for ATR on passive sonar, including (1) operationally, both directly and on US Navy platforms such as AN/SQQ-89; (2) commercially, on large cargo ships; and (3) in the context of sonar watch-stander training programs. Integrating CARIBOU onboard UxVs will likely generate high interest both from within the US Navy and from other NATO countries. Finally, the reasoning modules represent a significant advance in shortening the data-to-decision pipeline and improving ATR performance.
Keywords: Sonar Processing, Sonar Processing, UUVs, Deep Learning, Signal processing, Artificial Intelligence, Machine Learning