Machine learning algorithms for analysis of oil, gas and water well acoustic datasets

This thesis presents a comprehensive analytical and experimental study of machine learning algorithms and their application to real world problems. The data in this study is collected from distributed acoustic sensors attached to active oil, gas and water wells. The Terabyte size dataset of sounds recorded at every point from each well is in the form of acoustic amplitude as a two dimensional function of both distance along the pipeline and time. The new methods developed in this thesis offer an efficient, robust and accurate design for extracting the most informative features of the acoustic dataset and provide a deeper understanding of the features. The key achievements in the first part are the accurate estimations of speed and direction of fluid flow at each point along the pipeline, as well as the identification of flow regimes and predicting the fluid type in multiphase flow. Four methods have been developed using a combination of signal processing, image processing and machine learning algorithms to estimate the speed of sound at each point along the wells. Doppler shift is employed to compute speed of flow using the estimated speed of sound. To identify the fluid type in a multiphase flow regime, an Artificial Neural Network and a Convolutional Neural Network have been implemented. The second part of the thesis was devoted to analysing the acoustic datasets to identify the degree of opening of Inflow Control Valves (ICV), the combination of ICVs conditions and the location of ICVs in the well automatically. Two feature extraction techniques are developed to extract the global and local ICV acoustic features. The first one is a combination of Fourier Transform and PCA and the second one is Wavelet Time Scattering. Both feature sets are tested and validated using Neural Network classifier.