PATTERN RECOGNITION FOR A CLASS OF WARPED WAVEFORMS WITH APPLICATION TO WELL LOG SIGNATURE RECOGNITION.

摘要

Scope and Method of Study. The well log signature recognition problem considered in this study is essentially a pattern recognition problem involving waveform shapes. A signature class can be defined as a set of waveforms, all of which are warped versions of some original signature. A digitized signature S(n) is given, along with a digitized log section Y(n). A subsection of Y(n) is thought to belong to this signature class; the problem is to systematically search log Y(n) and identify the correct subsection. Several possible solutions to this problem have been investigated; all start by sliding a search window across the log section to select candidate subsections. Each candidate must then be compared with the given signature. The key question is how to determine a matching figure of merit for two sequences when warping is involved. The possible solutions to this question addressed in this study can be divided into two major categories: (1) methods based on dynamic programming, and (2) methods based on nonlinear prewarping filters. The second category can be roughly divided into two subcategories: (a) direct template matching, and (b) statistical pattern recognition techniques. A method of artificially creating a training set for the signature class has been explored. Proposed signature search algorithms have been evaluated by generating artificial random signature recognition problems. In addition to synthetic data, real data has been investigated.;Finding and Conclusions. The dynamic programming method has been determined to be a viable approach to the signature recognition problem. This method outperforms direct template matching in terms of the success rate, but is significantly slower. Direct template matching based on prewarping filters shows great promise because of its speed. A hybrid technique combining dynamic programming with direct template matching appears to have special promise since it outperforms dynamic programming operating alone both in terms of the success rate and speed. The results based on statistical pattern recognition are disappointing. However, the method of artificially creating a signature class training set has been shown to be useful in automatically selecting the paramaters for prewarping filters.