Using AI to Predict Machine Failures or Replacements
The profitability of oil and natural gas development activity depends on both the prices realized by producers and the cost and productivity of newly developed wells. Prices, costs, and new well productivity have all experienced significant changes over the past decade. A fleet of rigs costs an estimated to $54M per year for scheduled maintenance. At the risk of stating the obvious, preventing costly mechanical repairs, and in turn operating at a more efficient level, provides significant competitive advantage. McKinsey estimates predictive maintenance can save manufacturers more than $240B by 2025. In addition, the injuries that can result from mechanical failures cost more than just money.
The Challenge and the Solution
An ounce of prevention is worth a pound of cure. – Benjamin Franklin
With the growing proliferation of sensors, we now have real-time monitoring and access to more information than ever before. The challenge lies in managing the growing volume and velocity of data, as well as the speed required to act on it, before disaster—a rupture, for example—occurs. Traditionally, only 1% of the information gathered from an oil rig’s approximately 30,000 separate data points is leveraged. To act preemptively requires processing the data, extracting a signal, and creating a predictive model. This must happen quickly, before the underlying phenomena has changed. With machine learning automation, companies can automate the mechanical aspects of predictive modeling workflow to extract the signal and quickly build a good model. The predictive model can ultimately reduce downtime, increase asset availability, and reduce total ownership cost. McKinsey estimates such predictive models can reduce maintenance cost by 10-40%, reduce downtime by 50%, and, by extending machine life, lower equipment and capital investment by 3-5%.
Mark is an expert operations manager for a major oil refinery. He has been with the company for over 15 years, originally starting as a machine operator. He can recite every machine’s blueprint and specification from memory. He understands that planning maintenance based on a periodic schedule is causing unnecessary cost and increased downtime of his machines so he decides to start using a predictive maintenance approach to ensure optimal maintenance planning.
In this example, you will be using data from a Kaggle competition Predictive Maintenance dataset (originally available from Kaggle.com, /ludobenistant/predictive-maintenance ).
For information about getting access to that dataset or a similar dataset, see the discussion here. (Note that the Kaggle competition Predictive Maintenance dataset used for this Use Case is provided under the under the Creative Commons Attribution-ShareAlike License).