At sea commanders must maintain situational awareness that includes a wide range of surface and subsurface contacts with multiple acoustic (AC), radio frequency (RF), optical (VIS), and thermal (IR) signatures. Automatic detection and threat classification can dramatically improve their response time and course of action decisions. This proposal demonstrates the feasibility of artificial intelligence machine learning neural nets by delivering a USV/UUV Naval Abnormal Signal Detection & Classification (NASDC) intelligent system prototype to learn operational normal background signatures and historical signatures for vessels of interest, plus other data from a varying subset of hybrid on-board sensors (e.g., differing combinations of AC, RF, VIS, and IR bands). The NASDC will be tested on historical and simulated data for in-stride detection of unknown abnormal temporal signatures and multi-spectral historical signature classification confidence scores. NASDC will also provide a categorization results trust score for each time window based upon the similarity of the test data to the full training set. NASDC will be trained and tested based upon noise models and acoustic performance simulations to characterize environments plus historical experiment data at Applied Ocean Sciences (AOS).
Benefit: NASDC AI machine learning has the potential to significantly improve the Commanders ability to operate at sea with USV/UUV platforms. Automated, rapid, self-maintaining and improving on-board AI for USV/UUV assets benefit fleet operations by shortening the timeline from detection, to classification to engagement if a threat. Additionally, cognizant of a naval commanders reluctance to adopt an opaque black box solutions in their historically independent operating environments, NASDC will provide decision-aid information with understandable trust scores derived from the neural net computational methods. The key benefit is that the NASDC system will evolve and get better with use especially later when user feedback is incorporated into the system. User integration will enable individualization tailoring of the system with the classification, trust, and association hypothesis evaluation NNs tailored to the operational environment, sensors, scenarios, and entities of interest to each user. NASDC will be based on the current Intelligent System Associate (ISA) toolbox that contains the ANOM Enterprise Satellite as a Sensor (E-SAS) system which currently has the Abnormality Detection Classification Viewer (ADCV) that allows the user to suppress, enhance, and tailor the abnormality detection, tracking, and classification tools to his current operational needs. The ANOM E-SAS is currently being successfully used by Aerospace Corporation and AF sites to detect, track, and characterize unknown unexpected satellite State of Health (SOH) events. These abnormalities are used to detect subtle precursors that enable prediction of abnormal satellite events that support automated condition-based response. These events have been correlated with abnormal space weather events found by ANOM to support abnormality prediction using the ISA Smoking Gun tools. The NASDC intelligent system will extend these successful NN tools to provide emerging threat multi-spectral sensor report-to-track data association hypothesis scoring along with its own trust score. This will be affordable and easily extendable to different threats, sensor combinations, and operational environments since the solutions are data/goal-driven (i.e., need only be trained on the operational application of interest). The price paid for this affordable performance is that the accuracy is less than that of model-driven fusion solutions when the association hypothesis evaluation models are known.
Keywords: abnormality detection & categorization, abnormality detection & categorization, multi-spectral sensor fusion, autonomous retraining, Naval Abnormal Signal Detection & Classification (NASDC) intelligent system, Unmanned Surface Vehicle (USV) and Unmanned Underwater Vehicle (UUV) autonomous behavior, categorization neural networks trust scoring, Deep Multi-Start Residual Training (D-MSRT) neural networks
