So you've boarded the plane and settled into your seat, ready for your trip to a holiday paradise for a well-earned vacation. But then there's that dreaded announcement from the cockpit, where the pilot announces that she's "got a light on the dashboard", and needs to send in a maintenance crew to check it out. Before you know it, you've been sitting on the tarmac for an hour, you're probably going to miss your connection, and the trip is already off to a bad start.
Maintenance delays are major inconvenience for passengers, and they're a serious issue for airlines as well: according to Airlines for America, maintenance glitches cost US airlines 1.4 billion dollars in 2014. So airlines are naturally looking for ways to cut down on these costs, and minimize unexpected maintenance issues. This is a perfect opportunity to use predictive analytics: modern aircraft generate a wealth of data -- a 787 generates as much as a terabyte of sensor data per flight -- and airlines have extensive records of flight delays and their causes. So it makes sense to look the sensor data from flights that had unexpected maintenance issues, and see if you can find patterns in the data that indicate a likelihood of a maintenance problem, so yiou can fix any such issues before they become a delay.
This PowerBI dashboard demonstrates exactly this idea. Click on "View Report" and then the "Fleet Summary" tab at the bottom to get an overview of all aircraft. The fleet is shown as icons (with tail numbers) on the left; those with likely maintenance issues are flagged in red. (Information on aircraft locations and schedules is also provided, to help the Operations team find a suitable replacement aircraft, if necessary.) You can click on any of the aircraft to review the specific issues that are causing it to be flagged, as shown for Boeing 777 1ABCK below:
Anomaly alerts are generated from machine learning models, trained on historic sensor data and maintenance incidents, and then flagged using telemetry from aircraft in flight. The anomaly in this particular case is caused by a sudden increase in pressure in Sensor 17:
Many technology components come together to make this application possible. You can find the details in this blog post, and if you want to spin up the Cortana Analytuics application yourself, the Predictive Maintenance for Aerospace template provides a step-by-step guide. The underlying predictive models that detect anomalies and generate maintenance alerts are implemented using a combination of Azure ML machine learning algorithms, and the R language. In the video below, Microsoft data scientists Xinwei Xue and James Ren describe how they used Microsoft R Services to build and compare several models (including Decision Forests, Boosted Decision Trees, Poisson Regression and Neural Networks) to predict maintenance events from the sensor data, and then use those models to generate real-time alerts using SQL Server 2016:
So next time you're on a flight that actually takes off on time, think about the bullet you may have dodged thanks to predictive maintenace!
Cortana Analytics Gallery: Predictive Maintenance for Aerospace