Prediction of fracture density using genetic algorithm support vector machine based on acoustic logging data

Existing methods of well-logging interpretation cannot be applied accurately for the exploration and evaluation of carbonate reservoirs because of the fracture development. Based on the fracture density obtained by core analysis in a carbonate reservoir located in the Ordos Basin, in northwest China, three types of fracture density (low fracture density, medium fracture density, and high fracture density) of the target formation were identified. We investigated the effect of fractures on acoustic logging signals in the time and frequency domains by the Hilbert-Huang transform (HHT) and extracted 11 features in the time domain and nine features in the frequency domain. Then, we reduced the features in the time and frequency domain to three principal components by principal component analysis. Finally, a new prediction model of genetic algorithm-support vector machine method based on HHT of acoustic logging data was reported to predict the fracture density. The results indicate that the fracture density has a greater effect on the attenuation of intrinsic mode function 2 (IMF2) and IMF3 components for three different types of formation by empirical-mode decomposition analysis. The energy of the Stoneley wave and S-wave has higher sensitivity than the P-wave. Compared with the time domain, the distribution in the high-frequency domain has a greater correlation with fracture density by the Hilbert spectrum and marginal spectrum. The correlation coefficients between the fracture density and nine features in the frequency domain (R-2 = 0.5 - 0.7) are better than the coefficients with 11 features in the time domain (R-2 = 0.3 - 0.5). The core analysis and interpretation of resistivity image logging support the validity and effectiveness of our model. The prediction accuracy using the features in the frequency domain can reach to 82%-90%, which is much higher than using the features in the time domain with accuracy of 52%-59%. The application with more information of original acoustic logging data in our model not only avoid the error in velocity picking but also point the direction for the future prediction.