Various industries have started to take advantage of advanced technologies, such as computer vision, robotics to automate operations, increase production, reduce cost, enhance safety and improve asset integrity. Often considered slow to change, our oil and gas industry has started to implement these technologies and capture the related benefits. Implementation of these advanced technologies is often referred as digitalization.
To prevent environmental catastrophes and substantial losses, every oil and gas operator has to satisfy the general visual inspection (GVI) requirement for subsea asset integrity. In practice, field operations installations are either hard to get to or may present hazardous working conditions for humans, therefore remotely operated vehicles (ROV) and to an extent autonomous underwater vehicles (AUV) are used to provide automated anomaly detection on large-scale subsea infield and pipeline inspection.
To achieve this level of digitalization, state-of-the-art computer vision techniques with deep learning are applied to real-time camera imagery onboard the AUV or the ROV supporting vessel. Nowadays, deep learning methods based on convolutional neural networks ensure state-of-the-art performance in different areas of computer vision like image classification, activity recognition, and object detection. Object detection enables the identification of objects in the presence of occlusion.
Object detection can also be useful for the analysis of subsea video data, where sensing limitations in terms of resolution, range, and noise can influence the ability to accurately identify and track objects in a complex scene. Object detection is typically approached as a combination classification and regression problem where the algorithm will learn the characteristics of all the structures of interest (classes) using a set of labeled images.
While they may look incredible in practice and in many cases surpass human performance, state-of-the-art object detection techniques relying on deep neural networks require a great amount of manually labeled data by experts. Obtaining relevant labeled data is critical to evaluating and building the object detection computer algorithms. This labelled data can then be used for training and testing purposes to implement object detection. Like in the medical field, the bottleneck resides in establishing the ground truth for this data (labels). Observations are domain specific, sensitive and therefore require expert annotators. Moreover, labeling is both expensive and time consuming. Luckily with every problem comes solutions!
To tackle that problem there has been a push on the research front to optimize the cost of labeling while ensuring top performance of object detectors. This is known as active learning, it enables us to cope with the problem of high data availability versus limited annotation budget. It consists in incrementally selecting samples to be annotated based on their informativeness drastically reducing the cost of labeling. With such technology in place one can provide automated anomaly detection on real large-scale subsea infield and pipeline inspection data acquired by ROVs.
With active learning in place, one can produce efficient workflows for computer vision from ingesting the data, labeling the data, and training for machine learning for automated feature detection onboard the robots. Very promising efficiencies can be gained through the machine learning workflow and through automated anomaly detection. Furthermore, it has shown to remove human error during the acquisition process and shown a faster turnaround from data acquisition to insight.
Author information: From academic research on 3D surface reconstruction and meshing at INRIA (Sophia-Antipolis, France) to artificial intelligence research at Schlumberger-Doll Research (Cambridge, MA, USA), through 3D multi-Z subsurface salt meshing at Schlumberger (Stavanger, Norway). When not programming you'll find him designing, prototyping, reading, thinking or planning the next experimental project. Nader was part of the Schlumberger robotics team and is currently team leader of DELFI data ecosystem, core services.