With the latest advances in well logging technology, data can now be obtained dynamically during drilling (LWD) with quality comparable to that obtained under controlled conditions, using wireline equipment. When data is successfully acquired during drilling, a significant amount of time and resources are spared by the operator. However, of all the tools typically found in the bottom-hole assembly, the bore-hole image tends to be the most sensitive to the noisy environment encountered during drilling. Considering that wireline acquisitions occur after this period, it should be possible to decide whether to also run the image tool in the wireline assembly, based on the quality of the LWD data collected. Unfortunately, due to the large size of the data, image logs demand a significant time to be processed and analyzed by a professional specialist, typically exceeding the time available between the operations. Furthermore, the evaluation of the images by different professionals is subject to inconsistent interpretations for different operations. Therefore, the objective of the work is to automate and standardize the quality assessment of images acquired via LWD, to support the decision of whether to run a wireline image log.
To achieve this goal, it is necessary to efficiently analyze multiple criteria that would ordinarily be dependent on the speed and subjectivity of human interpretation. The chosen solution involves the combination of convolutional neural networks with feed forward networks, to extract and automatically classify features from the image. Using LWD image logs from eleven different wells and their respective quality annotation logs, provided by a group of professional image specialists, as well as the application of data augmentation techniques, a total of 14,000 dependent and independent variables were generated for the training and validation processes. Additionally, a blind test was conducted to confirm the robustness and reliability of the model.
The automation of image log quality assessment enables real-time decisions to be made. Consequently, unnecessary expenses on redundant acquisitions are avoided and critical resources, such as wireline image tools, are preserved. Additionally, the developed model can prevent the operator from leaving a location without having acquired proper image information, which could ultimately lead to flaws in the reservoir development and in the completion of the well. Finally, the assessment's reliance on a trained model ensures consistency in evaluations across different wells.
A plugin compatible with different petrophysical software has been developed and incorporated into operational routines. This has enabled real-time assessment to be performed in scenarios where it would not have been previously possible. This has consistently led to optimized wireline acquisition designs in terms of costs, resource allocation and well log quality.
