Revolutionizing Well Integrity: Temporal Deep Learning for Precise and Continuous Axial Groove Detection

Well integrity operations require consistent monitoring of pipeline conditions throughout the production cycle. Corrosion defects in wells, such as axial grooves, can have a catastrophic impact on both safety and operational costs, making early detection and thorough follow-up crucial. However, manually analysing data spanning thousands of meters within a well remains a labour- intensive process that relies heavily on technical expertise. A key limitation of conventional convolutional neural networks (CNNs) in addressing this challenge is splitting long borehole images into small parts, also called patches and treating them as static inputs, neglecting the continuous nature of the data. In deep well imaging, where axial grooves can span large sections of a well, maintaining continuity between predictions in each image patch is essential to preserve critical information about defect progression. Traditional CNNs struggle to effectively capture the temporal evolution of defects along the pipeline, causing discontinuity in groove detection and impairing the ability to trace defect evolution accurately. To address this limitation, we introduce a new approach that treats images in a spatiotemporal context, exploiting both the spatial features of the image and the temporality (sequence) of the image patches within a well. Our framework integrates novel techniques, such as temporal attention UNet (TAUNet), which introduces temporal blocks incorporating bidirectional convolutional long short-term memory (BConvLSTM) layers at strategic positions within the model architecture. This enables memory and feature sharing between neighboring images, allowing the model to better capture relationships and provide an improved representation of axial grooves present in challenging areas of the well. Additionally, we introduce temporal entropy loss as part of an overall hybrid loss function, ensuring that the model maintains consistent predictions across multiple frames by penalizing the model’s temporal aspect. A new postprocessing technique, temporal bridging, is also proposed to further refine the model prediction output by reducing false predictions and closing more gaps, ensuring greater temporal coherence across the entire wellbore extent. Apart from these proposed techniques, a custom statistical simulation and augmentation method has been introduced to further enrich the data input for model training, allowing the model to better generalize across a wider range of cases despite the initially limited dataset. Overall, these advancements enhanced the data representation through data augmentation, improved spatio-temporal coherence via TAUNet, refined continuity through post-processing, and the