APPLICATIONS OF DATA FUSION TO SIGNAL PROCESSING

A PROBLEM OF INTEREST TO THE U.S. NAVY IS THE DEVELOPMENT OF SIGNAL PROCESSING TECHNIQUES TO DETECT AND CLASSIFY THREAT SOURCES OF SHORT-DURATION, ACOUSTIC TRANSIENT EVENTS IN UNDERWATER ACOUSTIC DATA. THE PROBLEM IS VERY DIFFICULT BECAUSE (1) EVERY THREAT SOURCE HAS MANY CLASSES OF ASSOCIATED ACOUSTIC EVENTS, AND (2) THE UNDERWATER ENVIRONMENT HAS AN EVEN GREATER NUMBER OF INTERFERENCE ACOUSTIC EVENTS, SUCH AS THOSE FROM BIOLOGICS. ORINCON HAS FOUND THROUGH EXPERIEMENTS WITH ACTUAL TRANSIENT DATA THAT NO SINGLE SIGNAL PROCESSING TECHNIQUE CAN CREATE SUFFICIENT FEATURES TO ENABLE ONE TO DISTINGUISH UNIQUELY AMONG THE LARGE CLASSES OF TRANSIENTS. RATHER, SEVERAL TECHNIQUES ARE REQUIRED TO GENERATE METRICS FROM WHICH TRANSIENTS CAN BE CLASSIFIED. THE DATA FUSION OF THESE FEATURES IS DIFFICULT BECAUSE (1) EACH FEATURE-VERSUS-TIME PLOT FOUND FROM PROCESSING REAL DATA IS USUALLY A COMPLEX PATTERN OT THE HUMAN EYE, AND (2) THE AGGREGATE PATTERN FROM ALL THE FEATURES IS EVEN MORE COMPLEX TO PERCEIVE. THIS PROPOSAL WILL DETERMINE THE FEASIBILITY OF USING A REAL-TIME MASSIVE NEURAL NETWORK WITH OVER 10,000 INPUT NODES (BEYOND CURRENT REAL-TIME NEURAL NETWORK HARDWARE TECHNOLOGY) TO PERFORM FUSING OF MULTIPLE FEATURE-VERSUS-TIME GRAMS. IN PHASE II, THE ALGORITHMIC APPROACHES DETERMINED FEASIBLE WILL BE IMPLEMENTED BY EXPLOITING NEW NEURAL NETWORK HARDWARE TECHNOLOGY UNDER DEVELOPMENT BY DARPA. ANTICIPATED BENEFITS/POTENTIAL COMMERCIAL APPLICATIONS - MASSIVE NEURAL NETWORK DATA FUSION OF MULTIPLE COMPLEX FEATURE EXTRACTION ALGORITHM OUTPUTS WILL SHORTEN THE CYCLE TIME TO FIELD AN OPERATIONAL TRANSIENT PROCESSING SYSTEM FOR NAVY APPLICATIONS. THE MULTIFEATURE APPROACH WILL INCREASE THE PROBABILITY OF DETECION AND CORRECT CLASSIFICATION, DECREASE THE PROBABILITY OF FALSE ALARM AND MISCLASSIFICATION, AND BROADEN THE NUMBER OF CLASSES THAT CAN BE CLASSIFIED. THIS WILL SIMPLIFY THE DOWNSTREAM INFORMATION