Logging-while-drilling (LWD) dipole sonic tools have been introduced to the industry as a supplement to monopole and quadrupole measurement because they can provide shear slowness anisotropy, which is essential for formation characterization and well completion applications. Due to the presence of the collar, which acts as a strong waveguide, the recorded formation signal is significantly affected at low frequencies. Consequently, an automated interpretation of LWD-dipole sonic data remains a challenge. The traditional dispersive semblance-based method requires accurate estimates of parameters such as borehole size and/or mud slowness to avoid bias in the dispersion model used in the processing. Recently, a frequency-slowness domain inversion scheme has been developed that can invert for both the formation shear slowness and mud slowness by minimizing the guidance-mismatch cost function. However, this method uses an isotropic dispersion model and requires selecting narrow-band dispersion data in the low-frequency range with good-quality, which can limit the range of applicability of the method and also requires user input throughout the process.
We have previously developed a physics-driven machine learning-based method to enhance the interpretation of wireline dipole sonic data. However, the LWD scenario introduces additional complexity. This work extends the method to support the interpretation of LWD dipole sonic. An anisotropic root-finding mode-search algorithm is first used to generate extensive synthetic formation flexural dispersion curves that can match dispersion measurements in strong anisotropic formations in high-angle and horizontal wells, with a known tool model. Special care needs to be taken to pick the formation flexural mode from several co-existing modes arising from the strong coupling between tool and formation. After quality control and verification, this comprehensive synthetic dataset is used to train a neural network model. We then develop an inversion-based algorithm, taking advantage of this efficient neural network model and combining it with a clustering algorithm, to reliably label and extract the formation flexural mode, processed from either the modified Prony’s method, or a broadband dispersion analysis algorithm. The extraction around the formation flexural kick-in frequency is used for developing a quality control method. The strongest collar arrival, on the other hand, can be confidently removed due to the fundamental difference in its dispersion characteristics from the formation flexural mode.
This novel method can automatically and efficiently label the formation flexural mode and simultaneously invert it for formation shear slowness together with other relevant parameters such as mud slowness without user intervention. Since this method is built upon an anisotropic model, it can be applied to the full frequency range of the data spectrum without the traditional isotropic model assumption. Additionally, the regression analysis of the inverted mud slownesses can further provide physical constraint to reduce uncertainties in the inverted shear slowness. The algorithm has been tested on field data showing good performance. It makes edge deployment possible so that LWD telemetry can be optimized to transmit the processed data to the surface in real-time, which is essential to leverage the advantages of the conveyance method.