Oil and Gas Needs Automated Machine Learning

demo

What Does Automated Machine Learning Do for Oil and Gas?

  • Predictive maintenance helps avert asset failures and unscheduled downtime
  • Manually building and maintaining predictive models takes time, money, and data science talent companies can’t afford

  • Automated machine learning allows non-technical users to rapidly create flexible, scalable predictive models

In an industry like oil and gas, where asset failures and unscheduled downtime can be catastrophic to the bottom line, few things are as critical as having an efficient and effective system in place to monitor and repair assets. It’s for this reason that predictive maintenance is becoming an increasingly popular approach in oil and gas operations. But to make predictive maintenance work at scale, oil and gas companies must look to automated machine learning.

Predictive maintenance is the process of analyzing patterns in machine sensor data to predict impending asset failures before they occur. These predictions allow operators to prepare for failures by either performing the needed maintenance on the asset while it is not in use or bringing in a replacement asset beforehand.

This method of asset monitoring is a genuine cost-saver. Research by the Electric Power Research Institute has found that the annual cost of scheduled maintenance, performed at regular, predetermined intervals, is $24 per horsepower. Reactive maintenance, which performs maintenance on assets after they have run to failure, costs $17 per horsepower per year—though this is before considering the additional costs of the safety hazards or operational damaged that may be incurred by asset failure. Predictive maintenance was found to be by far the most cost-effective method, costing only $9 per horsepower annually and including no hidden costs or dangers.

Automated machine learning enables predictive maintenance for oil and gas

The Trouble with Traditional Predictive Maintenance

Unfortunately, many oil and gas operators are aware of the benefits of predictive maintenance, but still find it beyond their reach. Predictive maintenance has traditionally required teams of data scientists and technicians to constantly monitor and analyze incoming sensor data from all assets. This is hardly a scalable approach, even before factoring in the increasing scarcity of data scientists, who are in high demand but short supply. In 2016, it was reported that there were 6,500 total people listing themselves as data scientists on LinkedIn, but 6,600 job listings for data scientists in San Francisco alone1. Data scientists are a hot commodity, and the companies that are lucky enough to have scooped some up can’t afford to waste them on routine data collection.

Furthermore, these manually created predictive models take a great deal of time to create, perform poorly in extreme or unusual conditions, and can’t adapt to changing conditions—alter one variable of the asset being monitored, and the model is instantly rendered useless.

Automated Machine Learning for Better Predictions

The missing piece of this puzzle is automated machine learning (autoML). Simply put, automated machine learning is artificial intelligence (AI) that builds AI. AutoML technology uses machine learning techniques to create powerful, flexible predictive models in a matter of weeks or even days. These models can continually learn and refine themselves over time, and require minimal human involvement and even less data science or programming expertise to create, implement, and use.

Automated machine learning is particularly well-suited for the oil and gas industry, as its models perform well even in tasks with sparse data. Traditional predictive models require data from a minimum of fifteen failures to function properly, but the vast majority of drilling operators will elect to perform preventative maintenance on assets before they fail, and therefore don’t have much in the way of failure data to train a model. AutoML doesn’t need those fifteen failures, as it is able to find trends and build and train reliable models in data sets with few or no failures at all. These models are still far more accurate and robust as compared to their traditional counterparts.

Case Study: Automated Machine Learning for an E&P Operator

In one case study, a major exploration and production (E&P) operator used autoML to quickly build a solution capable of predicting workover needs and production in oil fields. This E&P operator was looking to minimize costs by using an automated, reliable solution to monitor the state of all the wells in a field. Specifically, it wanted to do so by assisting process engineers in identifying wells in need of maintenance, providing those engineers with insights on the potential return of individual wells to help prioritize maintenance work, and predicting failures far enough in advance to minimize unscheduled downtime.

An E&P operator used automated machine learning to predict maintenance needs

Using an autoML product, the operator was able to build new models in a matter of days with a fairly limited data set of well static properties, such as location; reservoir characteristics; and monthly oil, water, and gas production. These models had dramatic results, allowing the operator’s process engineers to predict workover, rod change, and cleaning operations needs in 12 out of 17 wells across the field with 70 to 80% accuracy. The models were able to predict asset failures and other maintenance needs three to six months in advance. Finally, the models developed with autoML revealed insights on production potential for individual wells, allowing engineers to focus their efforts on the most promising wells.

All in all, the E&P operator was able to dramatically increase its insight into its wells and their maintenance needs, at a far lower cost and in far less time than would have been possible with a traditional approach.

This is why automated machine learning is the future of predictive maintenance: because it makes predictive maintenance accessible to operators in a way it never has been before and yields better results than any other approach.

1“The State of Data Engineering.” Stitch, Inc.

Originally published April 2018.

Latest blogs

SparkCognition is committed to compliance with applicable privacy laws, including GDPR, and we provide related assurances in our contractual commitments. Click here to review our Cookie & Privacy Policy.