How AI Helped a Leading Energy Corporation Solve an Impossible Problem – Saving Million$
Oil drilling is big business: The total revenue of the United States oil and gas industry was roughly $181 billion in 2018. And when things go wrong, the losses are equally impressive.
According to one recent report, offshore oil and gas operators lose $49 million annually, on average, due to unplanned downtime. Some lose as much as $88 million. A common cause of such unplanned downtime is when drill bits fail.
What’s a Drill Bit?
Oil and gas companies use drill bits to create wellbores in the ground to extract crude oil and natural gas. These bits have to make their way through layers of rock, which generates a lot of heat and pressure. They constantly rotate and vibrate as they grind through the rock, and over time, they weaken and break. When they do, drilling comes to a halt.
Drilling is a complex operation that involves dozens of companies working together. Costs for one drilling project can easily exceed $4 million and take three weeks or more to complete — and that’s without any drilling complications. But when drills have to be removed from the ground and repaired or replaced while machines — and workers — sit idle, schedules slip and budgets get blown out of the water. The wells are thousands of feet deep vertically and horizontally, and crews have to pull the equipment back to the surface to replace a broken drill bit. As each leased rig costs $50,000-100,000 to operate daily, any downtime is a significant loss.
Why Do Drill Bits Fail?
Drill bits can fail for myriad reasons including incorrect or inappropriate operating conditions, too much force on the bit, insufficient fluid to keep the bit clean and more. Sensors are often used to monitor drilling conditions, to collect performance data in the hopes of understanding how to best use and protect the bits and improve drilling efficiency.
But actionable insights from this sensor data are often unavailable, as down-hole sensors — which can provide the best performance data — are difficult and expensive to deploy. Additionally, many of these sensors don’t provide real-time data, so results can only be analyzed after drilling, making it difficult if not impossible to predict impending bit failures and stop them before they occur.
Predictive Modeling Detects and Prevents Possible Bit Failures
Predictive models that can correlate data collected at that surface with down-hole data enable early warning if a bit is close to failing. At the surface, sensors can provide information about the weight on the bit, how fast the bit is turning, torque and resistance and amount of vibration — all factors that can contribute to bit damage. A Fortune 500 oil and gas company was told it was not possible to make such predictions to help eliminate these drill bit failures.
Then 2predict stepped up to help, and here’s how the problem was solved:
First, we set to work using statistical learning techniques to analyze the sensor data for insights and reconstruct a story of drilling operations.
Next, we created a model that would correlate what happens at the surface with what happens down- hole. The model was based on “sequential neural networks,” which is well-suited to time series data.
After creating the model, we trained it using a set of synthetic data to build confidence that the architecture and methods were appropriate for the dataset size and dimensionality. The synthetic models consisted of classification tasks, during which a specific signal signature was associated with certain events – mostly highs and lows for the overall drilling activity.
We then de-noised the original data, using a combination of Savitzky-Golay smoothing and discrete wavelet transform, which dramatically improved the predictive performance of the algorithms.
Finally, we built regression models to correlate surface and down-hole data using the Tensorflow library. The models were then trained on servers with GPUs and delivered to the customer.
By using statistical models, the energy company can now predict impending drill bit failures with a high level of accuracy, which helps them improve drilling penetration and completion rate times. The ability to predict potential bit failures and stop drilling before they happen substantially reduces unscheduled and costly downtime. Based on past failure rates across 30 of the company’s wells, they expect to avoid $600,000 of losses. Reducing such failures by just 50% across 200 wells could save about $60 million.
To learn more about the process, the model that was developed, and what benefits the oil and gas company has realized, read the full case study.