(715c) Identify Fracture-Hit Events in Low-Frequency DAS Strain Rate with Convolutional Neural Networks
AIChE Annual Meeting
2021
2021 Annual Meeting
Fuels and Petrochemicals Division
Unconventional Oil and Natural Gas: Science & Technology Advancement II
Tuesday, November 16, 2021 - 4:00pm to 4:15pm
Event Classification. CNN applications in image-based real-time fiber optic sensing event detection. A propagating fracture exhibits a characteristic cone-shaped extension zone surrounded by compression at the fracture tip (Tang and Zhu, 2021). We adopt a fracture propagation model to simulate fracture geometry and stress field around the fracture. With this approach, we generate the strain rate responses for various completion scenarios. The responses are passed to the CNN models for event classification and localization.
When installed outside of the casing of a monitoring well, the fiber optic cable is considered mechanically coupled with the casing and the formation. When a hydraulic fracture perpendicular to the monitoring well approaches the well, it causes axial deformation of the fiber. Low-frequency DAS can detect such deformation because the phase change in the backscattered light is linearly correlated with the axial displacement. Hence the displacement can be used to approximate the axial strain rate. (Jin and Roy, 2017).
Convolutional neural network (CNN) is the state-of-the-art technology used in computer vision. CNN models are compelling in computer vision tasks because the trainable convolutional filters can learn patterns in a visual field. They can be trained to recognize objects in an image. For training, the input is labeled images in RGB (red, green, blue) channels or in grayscale, and a trained model predicts the intended target class or location of the object. This study uses strain rate images to train models that can identify and locate fracture-hit events.
Many image classification applications utilize well-known pre-trained. In fact, we find a simple CNN architecture sufficient. The model consists of two convolutional layers followed by two fully connected flat layers and a sigmoid output for binary classification. Each convolutional layer includes convolution, batch normalization, ReLU activation, and max pooling.
Event Localization. In addition to identifying whether an image contains a fracture-hit event, we also want to know the location of the event. This is considered an object localization problem, practically a regression problem with the task of identifying where in the image the object of interest is in terms of time and depth. The location is typically characterized by a 2D coordinate, and width and height that define a bounding box around the object. In this application, we are only concerned with the 2D coordinate, hence the model has two outputs: location in X-axis (time) and in Y-axis (depth). The architecture of the localization model is the same the classification model trained to identify the presence of the event, except that the sigmoid output is replaced with two linear activations (essentially no activations). Correspondingly, we use mean squared error as the loss function since the training data are generated with well-defined distributions.
Results and Observation. The developed approach was trained and validated by datasets. The results are with satisfaction.
This study demonstrated the efficiency and accuracy of using convolutional neural network models to classify and localize fracture-hit patterns in simulated strain rate data. Both classification and localization models achieved near-perfect predictions. The results prove the feasibility of using CNN for real-time event detection from low-frequency DAS data. Because of CNN's flexibility and trainable nature, its applications extend beyond identifying strain rate patterns from low-frequency DAS data. This functionality can be developed as software packages for fiber optic sensing deployment, creating additional value for this already versatile technology. Edge detection is another plausible technique for locating fracture-hits, but it relies on assumptions about the image. The cone-shaped pattern must be smoothly and sharply defined. These assumptions often do not hold for field DAS data. The comparison of edge detection and CNN further supports the need for using CNN for image-based real-time event detection applications from fiber optic sensing data.