Speaking at the 2018 Nvidia GPU Technology Conference, in the “AI at the edge” track, Schneider Electric’s Matthieu Boujonnier showed how Schneider is applying deep learning to analyze artificial lift with Dynacard records. Labelled Dynacard records can be treated as image data for classification with a convolutional neural net, just as pictures of cats are classified on the internet! The problem is that it is hard to obtain a decent set of labelled Dynacard data, there is more usually only a smallish set of typical pump responses labelled by an expert.
Schneider’s approach is to ‘augment’ the data set by combining cards showing similar effects. This is said to make for a simpler model and less ‘overfitting’. The dataset can be further augmented using an autoencoder and a latent space methodology. Another approach is to extract ‘new’ features from images such as gradients. Different statistical models can be combined in an ensemble model which is said to increase the odds of success. Schneider’s ‘Realift’ solution rod pump controller, bundled with data acquisition and cleansing apps, can now be deployed ‘at the edge’ i.e. in the field for on-the-spot diagnostics and beam pump optimization.
Steve Dominguez (CGG) compared current interpreter guided interpretation techniques such as deployed in CGG’s InsightEarth* 3D interpretation solution with a data-driven deep learning approach. CGG is currently working on accelerating compute intense processes for automatic fault extraction, geobody extraction and noise reduction by applying simpler and faster neural network approaches. Dominguez has retrained the LeNet system to recognize faults and gets 70-80% accuracy in fault identification.
* Developed by the Geoscience Interpretation Visualization Consortium (GIVC) in a 12 plus year R&D effort.
© Oil IT Journal - all rights reserved.