Teaching machines to speak drilling

Maana’s Jeff Dalgliesh discusses a recent project with Chevron aimed at training a machine to understand how drillers describe problems encountered during operations.

Image from iStock.

Typically, when people think of Intelligent Machines they think of algorithms that monitor equipment sensors and alert people when a problem is suspected in the operations. However, there is another set of use cases that is beginning to emerge in the industry where Intelligent Machines are being used to understand what people are talking about and use this understanding to assist in making operational decisions.

An example of this was presented at SPE Intelligent Energy in Aberdeen, last year, by Maana, a technology company based in Palo Alto, California, and Chevron. “Natural Language Processing Techniques on Oil and Gas Drilling Data” set out how Maana and Chevron trained a machine to understand how drillers describe problems they encountered in operations. This enables well planning engineers to get a better understanding of potential risks associated with drilling a well by seeing how often a problem happened in the past.

For simplicity, let’s say an average well might take about 30 days to drill and on average the driller may make eight comments each day describing the operations and problems they encountered. This equates to about 240 comments per well. A company such as Chevron has more than 100,000 wells worldwide, which represents 24 million comments. Training a machine to read all 24 million comments and classify each comment into a problem type would allow a well planning engineer to get a better understanding of the frequency of a certain type of problem. If a person took five seconds to read and classify each of the 24 million comments, it would take about three years and nine months – non-stop. A properly trained algorithm could do this in a few minutes.

Here is an example of a typical comment describing a drilling operation where the driller encountered the pipe being stuck in the hole:

PIPE STUCK WHILE ROTATING 10in OFF BTM. LOST PUMP PRESS AND GAINED STGROKES WHILE JARRING ON PIPE. MIX & PUMP 100BBLS DFE1310 WHILE JARRING ON PIPE. SPOT IN PLACE AT 2400HRS. JAR PIPE FREE 3MINS. AFTER SPOT WAS IN PLACE. PULL 12STDS. PUMP 30BBLS TO CLEAR SPOT OUT OF DP. PUMP SLUG. POOH

There is a lot of jargon and technical terminology in the comment. This jargon and technical language provides a lot of clues as to what the driller is talking about. Maana uses natural language processing algorithms to take these clues and build statistical language models to classify the comment into a problem type.

A challenge when training natural language understanding algorithms for the oil and gas industry based on comments is identifying when the author is describing an actual problem that happened or when they are describing a Health, Safety and Environment (HSE) meeting on the rig where they are training to respond to the problem. For example, our first pass of the natural language understanding algorithm identified a lot of kicks on the rig based on the comments, when we dug deeper a huge majority of these comments were the HSE stand up meeting where the crew was reviewing well control emergency procedures.

Using Intelligent Machines to mine the vast amounts of unstructured text in organizations unlocks deeper understanding of operations so operators can make more informed decisions for future operations. In a world where people and Intelligent Machines coexist in operations, being able to understand each other is an important element to designing our technology systems of the future. Combining machine learning techniques, knowledge models and state-of-the-art information processing techniques can help organizations navigate the transformation to a human-guided machine-assisted future.

Jeff Dalgliesh is Maana’s Oil and Gas Specialist, working with clients to apply machine learning and artificial intelligence techniques. Prior to Maana, Dalgleish worked for Chevron for 18 years most recently as drilling and completions technology manager and previously was the drilling and completions technology architect, both at the Chevron Engineering Technology Co. He holds a BSc in Computer Science from University of British Colombia in Canada.