Browsing Physics by Subject "Acoustics"
Now showing items 1-2 of 2
Neural Network Approach To Classification Of Infrasound SignalsAs part of the International Monitoring Systems of the Preparatory Commissions for the Comprehensive Nuclear Test-Ban Treaty Organization, the Infrasound Group at the University of Alaska Fairbanks maintains and operates two infrasound stations to monitor global nuclear activity. In addition, the group specializes in detecting and classifying the man-made and naturally produced signals recorded at both stations by computing various characterization parameters (e.g. mean of the cross correlation maxima, trace velocity, direction of arrival, and planarity values) using the in-house developed weighted least-squares algorithm. Classifying commonly observed low-frequency (0.015--0.1 Hz) signals at out stations, namely mountain associated waves and high trace-velocity signals, using traditional approach (e.g. analysis of power spectral density) presents a problem. Such signals can be separated statistically by setting a window to the trace-velocity estimate for each signal types, and the feasibility of such technique is demonstrated by displaying and comparing various summary plots (e.g. universal, seasonal and azimuthal variations) produced by analyzing infrasound data (2004--2007) from the Fairbanks and Antarctic arrays. Such plots with the availability of magnetic activity information (from the College International Geophysical Observatory located at Fairbanks, Alaska) leads to possible physical sources of the two signal types. Throughout this thesis a newly developed robust algorithm (sum of squares of variance ratios) with improved detection quality (under low signal to noise ratios) over two well-known detection algorithms (mean of the cross correlation maxima and Fisher Statistics) are investigated for its efficacy as a new detector. A neural network is examined for its ability to automatically classify the two signals described above against clutter (spurious signals with common characteristics). Four identical perceptron networks are trained and validated (with >92% classification rates) using eight independent datasets; each dataset consists of three-element (each element being a characterization parameter) feature vectors. The validated networks are tested against an expert, Prof. Charles R. Wilson, who has been studying those signals for decades. From the graphical comparisons, we conclude that such networks are excellent candidate for substituting the expert. Advantages to such networks include robustness and resistance to errors and the bias of a human operator.
The Characterization Of The Infrasonic Noise Field And Its Effects On Least Squares EstimationLocalization of the source of an acoustic wave propagating through the atmosphere is not a new problem. Location methods date back to World War I, when sound location was used to determine enemy artillery positions. Since the drafting of the Comprehensive Nuclear-Test-Ban Treaty in 1996 there has been increased interest in the accurate location of distant sources using infrasound. A standard method of acoustic source location is triangulation of the source from multi-array back azimuth estimates. For waves traveling long distances through the atmosphere, the most appropriate method of estimating the back azimuth is the least squares estimate (LSE). Under the assumption of an acoustic signal corrupted with additive Gaussian, white, uncorrelated noise the LSE is theoretically the minimum variance, unbiased estimate of the slowness vector. The infrasonic noise field present at most arrays is known to violate the assumption of white, uncorrelated noise. The following work characterizes the noise field at two infrasound arrays operated by the University of Alaska Fairbanks, The power distribution and coherence of the noise fields was determined from atmospheric pressure measurements collected from 2003-2006. The estimated power distribution and coherence of the noise field were not the white, uncorrelated noise field assumed in the analytic derivation of the LSE of the slowness vector. The performance of the LSE of azimuth and trace velocity with the empirically derived noise field was numerically compared to its performance under the standard noise assumptions. The effect of violating the correlation assumption was also investigated. The inclusion of clutter in the noise field introduced a dependence to the performance of the LSE on the relative signal amplitude. If the signal-to-clutter ratio was above 10 dB, the parameter estimates made with the correlated noise field were comparable to the estimates made with uncorrelated noise. From the results of these numerical studies, it was determined that the assumption of Gaussian, white, uncorrelated noise had little effect on the performance of the LSE at signal-to-noise ratios greater than 10 dB, but tended to over estimate the performance of the LSE at lower signal-to-noise ratios.