Shear slowness is commonly computed from well log dipole flexural mode data using slowness-timecoherence (STC) processing. Flexural dispersion is handled by restricting the signal's frequency content to the low frequencies that travel close to the formation's shear velocity, or by altering the phase relationships within the waveforms prior to STC processing in accordance with observed dispersion characteristics. Restricting the frequency range may not eliminate the need for a residual dispersion correction, however, and in noisy environments the dispersion curves needed for modifying phase relationships may be of poor quality. Formation and borehole properties have a significant influence on observed frequency content, and selection of bandwidth for the optimal balance between noise and size of the residual dispersion correction adds to overall processing time. Inversion addresses these difficulties by computing shear slowness directly from observed dispersion characteristics, but the process needs to be fast and tolerant to noise to be effective in commercial applications. In order to make the inversion efficient the iterative steps which compare observed and forward modeled dispersion curves are replaced with a fast neural net trained on a large number of pre-modeled curves generated with known formation and borehole properties. Automated mode frequency detection constrains the bandwidth over which dispersion curves are matched, accounting for potentially high levels of noise seen, for example, in horizontal wells. Results from 127,000 modelled and field data points show improved accuracy and precision relative to STC processing.
展开▼
