CHARACTERIZING SEISMIC EVENTS USING ARTIFICIAL NEURAL SYSTEMS

THIS STUDY WILL INVESTIGATE THE FEASIBILITY OF APPLYING NEURAL NETWORK METHODS FOR THE EXTRACTION OF SEISMIC WAVEFORM CHARACTERISTICS AND THE IDENTIFICATION OF SEISMIC EVENTS. THE SPECIFIC OBJECTIVES IN THIS INITIAL STUDY WILL BE TO DEVELOP A NEURAL NETWORK WHICH WILL RECOGNIZE THE SEISMIC SIGNAL CHARACTERISTICS OF DELAYED BLASTING (RIPPLE FIRING) IN ECONOMIC EXPLOSIONS AND CHARACTERIZED THE RIPPLE-FIRE PATTERNS, IN TERMS OF NUMBER OF EXPLOSIONS, RELATIVE DELAY TIME, AND RELATIVE YIELDS, FROM THE SPECTRA AND CEPSTRA OF THE SIGNALS. A CONSTRAINED NEURAL NETWORK ARCHITECTURE WITH HIDDEN UNITS WILL BE USED. THE INPUT NODES WILL BE THE RIPPLE FIRED BLAST. TRAINING WILL BE ACCOMPLISHED BY PRESENTING AS INPUT TO THE NEURAL NETWORK CEPSTRAL FEATURES FOR SEISMIC SIGNALS FROM BLASTS WITH KNOWN RIPPLE FIRE CHARACTERISTICS AND CONSTRAINING THE OUTPUT TO BE THE RIPPLE FIRE PARAMETERS. THE WIDROW BACKPROPAGATION METHOD WILL BE USED TO SET THE WEIGHTS ON THE INTERCONNECTED HIDDEN NODES IN THE TRAINING PROCESS. AFTER TRAINING, THE NEURAL NETWORK WILL PRODUCE THE DESIRED RIPPLE FIRE PATTERN WHEN PRESENTED THE SAME CEPSTRAL FEATURES. INITIAL TESTING OF THIS ALGORITHM WILL BE ACCOMPLISHED WITH SYNTHETIC SIGNAL FEATURES FOR ASSUMED REALISTIC RIPPLE FIRING CONFIGURATIONS.