Archive | 2019
Machine learning for the automated detection of diagnosis-revealing features on leaky flexural wave imager data
Abstract
Cement evaluation in cased oil and gas wells is important to ascertain well integrity before putting the well on production. It is conducted through the use of acoustic measurements, in particular an advanced ultrasonic tool with two modalities: pulse-echo and pitch-catch with the latter relying on the excitation and detection of the steel casing-thickness flexural mode. When the annular fill behind the casing is comprised of a solid with a compressional (P) wave velocity that intersects the dispersive flexural mode phase velocity curve within the signal frequency bandwidth, phase matching to a headwave at the casing-solid occurs and we observe an additional contribution that follows closely the arrival of the casing flexural mode: a feature we refer to as a clinging P. We propose to use this occurrence exhibited in the raw waveforms to our advantage, as a basis for a quantitative diagnosis of the solid behind the casing. To leverage this effect in a non-assisted robust (online) diagnosis, we use a machine learning based workflow that automates the detection of clinging P arrivals in the flexural wave data. The automated workflow highlights axial and azimuthal regions, where the annular fill behind casing is confirmed to be a solid with additionally a tight range for its compressional wave velocity - therefore rendering the diagnosis of the tool measurement more quantitative and more autonomous than delivered with existing inversion schemes.