Recorded Well Data Enriches the Testing of Automation Systems by Using a Deep Neural Network Approach

Abstract

Abstract One technique useful in the testing and development of drilling automation system is to use synthetic data. A good drilling time series simulator can enrich a dataset for testing, and enable the inference of the states of drilling in real time. However, conventional simulators do not generate the "warts" of real data (noise, gaps, etc.). The proposed solution is a model that learns from real data, characterizes the different drilling responses and conditions the data with a deep neural network (DNN) approach, and generate realistic drilling time series dataset. To simulate a drilling-time series dataset, a DNN can model physical properties of the formation, rig, and sensors, and generates data with realistic curve patterns when it is trained on actual measurements, e.g. block position, hook load, standpipe pressure, and surface torque. The neural network has multiple convolutional, recurrent, and fully-connected layers. The model, trained with wellsite recorded data, captures the spatio-temporal distributions among data channels, and then uses a windowed input to predict the next data points, which are then fed back into the network to generate the simulated data sequence recursively. An actual sensor drilling-time series dataset containing various channels are input into the DNN. The networks contain eight convolutional layers with three max-pooling layers, three recurrent layers, and four fully-connected layers. The time window used in the input contains 512 samples for each channel, while the output is 1 sample for each channel. After training the network for 200 epochs, the network can successfully simulate time series data recursively. The simulated time series preserve the features of the original training data, while maintaining the data distribution of multiple channels. For example, the network shows a consistent "inslips" pattern in the hook load channel when the block position moves quickly from bottom to top. Currently the simulation is autonomous based on the training data, and does not take input as controls, which is our future steps. The proposed DNN model is a low-cost, robust model that simulate drilling-time series datasets containing complex spatio-temporal patterns. Our proposed algorithm is the first known simulator of drilling time series datasets that models the nontrivial physics laws and properties, including formation, rig, and sensors, and generates data containing realistic curve patterns with a deep neural network approach. The simulator greatly helps the inference component of automation systems with the enrichment of datasets that are available for testing.