Support-Vector-Machine Phase Classification of Downhole Leak Flows Based on Acoustic Signals
- Atchyuta Ramayya Venna (Halliburton) | Yi Yang Ang (Halliburton) | Nam Nguyen (Halliburton) | Yinghui Lu (Halliburton) | Darren Walters (Halliburton)
- Document ID
- Society of Petrophysicists and Well-Log Analysts
- Publication Date
- December 2018
- Document Type
- Journal Paper
- 841 - 848
- 2018. Society of Petrophysicists & Well Log Analysts
- 6 in the last 30 days
- 96 since 2007
- Show more detail
Acoustic hydrophones have been used for detection and localization of leak flows in wellbores. In addition to the location of detected leak flows, fluid phases of the flows are also valuable information to extract from the measured acoustic signals.
This work introduces a support-vector-machine (SVM) approach to classify four different flow phase scenarios: liquid-to-liquid (L2L), gas-to-gas (G2G), gas-to-liquid (G2L), and liquid-to-gas (L2G). The proposed algorithm consists of three steps: spectrum estimation, principal component analysis (PCA), and SVM classification. First, the frequency spectrum is computed from the time-domain acoustic signals. Second, PCA is used to extract features from the spectrum profile, during which principal components of the spectrum that contain most of the information are extracted. Third, the extracted features from Step 2 are used as inputs to the SVM classifier, which labels the input feature values with one of the previously mentioned scenarios.
Experimental data were used to train and test the SVM classifier. For each of the four flow phase scenarios, multiple leak types were simulated to account for variations of actual leak flows downhole. For each leak type, various leak flow rates were produced using different orifice sizes and differential pressures across the orifices. An acoustic leak-detection tool was used to record the acoustic signal of the leak sources. Field data were used to validate the SVM classifier, achieving an accuracy of 91%. In addition, it was determined that the combination of spectrum estimation, PCA, and SVM outperformed the algorithm that used only spectrum estimation and SVM.
|File Size||7 MB||Number of Pages||8|