The ever-increasing acquisition of real-time borehole images in the oil and gas industry has unleashed an unprecedented potential for the geomechanical and geological interpretation of hydrocarbon reservoirs. Real-time decision making, whether it is for well placement or wellbore stability analysis, is at our fingertips provided efficient automation techniques are implemented to interpret the geological features from the images.

The present work describes fitting customized light neural network (NN) specifically designed to extract stress-induced features from high-resolution ultrasonic images. Thanks to its small size and low memory requirements, it is directly embedded at the heart of a recently deployed logging-while-drilling (LWD) dual physics imager and run downhole in real-time during acquisition. A synthetic description of the extracted features is then streamed to the surface together with a compressed version of one of the borehole images (usually the apparent resistivity image, rich in geological information).

This technology brick successfully addresses one of the most drastic limitations in modern LWD acquisition: the limited amount of information transmitted to the surface in real-time through mud pulse telemetry (pressure waves travelling through the mud), reserving the high-resolution images to recorded mode only. Overcoming this problem is of prime importance to spot in real-time, prevent or mitigate drilling instability issues, one of the major causes of nonproductive time on the rig.

This new deep-learning methodology builds on expert labeling of stress-induced features (such as breakouts (BO) or induced fractures (IF), followed by rigorous learning on both real and synthetic data. The model detects breakouts downhole with >93% accuracy. It uses a fine-tuned, pre-trained state-of-the-art NN, on which quantization and layer-based pruning of the weights is applied. The model is fed with annotated data from many geological environments, acquired from field tests the new LWD tool, public domain data, and other US imagers. Because drilling-induced features do not necessarily occur over long intervals, a new method was designed to artificially augment the dataset with realistic annotated images, resulting in a significant improvement of the model performance. A post-processing step, guided by geomechanic considerations and proceeding on a larger scale than the downhole network, is deployed at the surface to considerably decrease false positives and increases the robustness of the model.

Beyond its current application to wellbore instability prediction, the present method could be generalized at low cost to other real-time applications, other feature or image types, provided a sufficiently large database customized to the desired application is available for training.

This content is only available via PDF.
You can access this article if you purchase or spend a download.