A SYNTACTIC APPROACH AND VLSI ARCHITECTURES FOR SEISMIC SIGNAL CLASSIFICATION
Abstract
Syntactic pattern recognition has been applied to seismic classification in this study. Its performance is better than many existing statistical approaches. VLSI architectures for syntactic seismic recognition are also proposed which take advantage of parallel processing and pipelining so that a constant time complexity is attainable when processing large amount of data. Application of syntactic pattern recognition to damage assessment is also proposed and demonstrated on a set of experimental data. Seismic waveforms are represented by strings of primitives, i.e., sentences, in this study. String-to-string similarity measures based on both distance and likelihood concepts are discussed along with the symmetric property and the hierarchy. A fixed-length segmentation is used in the experiment. Encouraging results comparable to those of the best statistical approaches are obtained with only two very simple features, namely, zero-crossing count and log energy. Primitives are automatically selected using a hierarchical clustering procedure and two decision criteria. Nearest-neighbor decision rule and finite-state error-correcting parsers are used for classification. For error-correcting parsing, finite-state grammars are first inferred from the training samples. These two approaches have same performance in the experiment, whereas the nearest-neighbor rule is faster in speed. Attributed grammar and its parsing are also proposed for seismic recognition, which could reduce the complexity and increase the descriptive flexibility of the pattern grammars. VLSI architectures are proposed for fast recognition of seismic waveforms. Three systolic arrays perform the feature selection, primitive recognition and string distance computation. These individual units can be used in other similar applications. Although this study is on seismic classification, it can be extended or modified to tackle other signal recognition problems.
Degree
Ph.D.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.