Application of Computer Vision in Diagnosing Water Production Mechanisms in Oil Wells

Abstract

Abstract Diagnostic plots, introduced by K.S. Chan, are widely used to determine excessive water production mechanisms. In this paper, we introduce a computer vision model that is capable of segmenting and identifying multiple Chan signatures per plot, for the sake of surveillance and early screening, given that wells could exhibit diverse mechanisms throughout their lifecycle. As deep learning demands a vast amount of information, we start our workflow by building a dataset of 10,000 publicly available oil wells that have experienced varying water production mechanisms and annotating them. Next, we perform pre-processing and remove anomalies from production data, which could be misleading in analysis. Then, we visualize Chan plots as images, split the dataset, carry out augmentation, and have the data ready to be used as input for a CNN (Convolutional Neural Network) layer. Eventually, we utilize YOLO, a one-stage object detector, tune hyper-parameters and evaluate the model performance using mAP (mean average precision). The collected data from fields in Alaska and North Dakota represent oil wells that have been producing for decades. When working with some wells that possess noisy production data, we identified challenge, bias, and tedium in human interpretation of Chan plots. Subsequently, we observed the inevitability of cleaning well production data prior to constructing the plots, and thoroughly revealed its effect on enhancing the potentiality to get a satisfactory score. In addition, we concluded that following a simple approach of active learning, a technique that allows the user to analyze mistakes of prediction and label the data incrementally in order to achieve a greater score with fewer training labels, accomplished a significant boost in model performance especially with under-represented classes. The newly proposed model employs automatic feature extraction, expresses data in much more detail and is confirmed to be robust as it successfully predicted multiple mechanisms of excessive water production, with confidence scores higher than 80%, in wells that exhibit different production conditions such as horizontal trajectories, artificial lift, water flooding, stimulation, and other well intervention events. In this work, we introduce a novel computer-vision model, which combines image processing and deep learning techniques to identify multiple water production signatures that a well can undergo, and eliminate the subjectivity of human interpretation. This approach has the potential to be effective, as a part of workflow automation, in expeditious surveillance of large oilfields.