Graham School News

Failure Prediction Models Shine at MScA’s Summer Capstone Showcase

Summer 2017 Analytics Capstone Showcase winners Arun Bodapati, Cyrus Safaie, and Srini Vokkarane

Calling the presentations great and the participation excellent, Sema Barlas, Director of the Masters of Science in Analytics program at the Graham School, congratulated the MScA students who took part in the Summer Capstone Showcase Event on Saturday, August 19. With the $5,000 best-in-show prize awarded to the winning team of Arun Bodapati, Cyrus Safaie, and Srini Vokkarane for their presentation entitled “A Machine Learning Approach to Predicting Fuel Injector Failure,” honorable mention was also given to Akin Akyüz for his project “Predicting Failure of Industrial Machines” and John Balcarcel, Nathan Jensen, and Taylor Marx for “Social Media Tells: Predicting Future Spending Through Text Analysis.”

“I would like to thank all the students and their supervisors who worked very hard to achieve such great outcomes,” said Barlas, who, along with Francisco Azeredo, Brian Clark, and Anish Gera, was also on the jury conferring the prize. “It was quite difficult for the jury to select the winner, but in the end the jury felt that the winning team distinguished itself through the rigor and quality with which they developed their project. Not only did they skillfully justify their decisions, they offered a framework for solving similar problems.”

Working with their faculty advisor, Francisco Azaredo, Vice President of Analytics at Northern Trust as well as a lecturer in the MScA program, the winning team employed machine learning techniques to develop a framework to predict the rolling probability of a unique fuel injector failure mode identified by Navistar. Affecting a family of medium duty legacy engines first designed about 20 years ago, the team faced the challenge of finding a solution to the fuel injector failure without recourse to the rich data streams stemming from the sensors embedded on engines designed today.

“Given the type of data we had,” the team said, “we knew we wouldn’t be able to arrive at a solution with 80 or 90 percent predictability. We had to develop proxies and build innovative variables from the data while exploring a wide range of algorithms. After seeing that none of them worked too well, we were able to bring multiple models together to develop a metalearner, but in the end we ended up going with a simpler model for its improved interpretability.”

While focused on providing a solution to the problem at hand, the team also worked with Navistar to position their solution to the problem in terms of a framework so that internal teams at Navistar would have a tool to use when confronting similar problems in the future.

“The idea was to look at it from a big-picture perspective,” the team said. “If you work at Navistar and you’re given a problem like this to solve in a month, you won’t have time to investigate the problem fully when developing a solution. The advantage to working on a problem like this for nine months or even a year is that you’ll be able to explore all the different options available. By solving the problem with a framework solution, we’ve given the internal teams at Navistar a helpful way to approach all failure-related problems in the future.”

Also focusing his Capstone on a problem involving failure prediction, Akin Akyüz worked with faculty advisor Yuri Balasanov, Founder and President of Research Software International, Inc., and lecturer in the MScA program, with the goal of using streaming sensor data from a large number of industrial machines to predict when breakages will occur. With a dataset provided by Uptake, a predictive analytics company identifying patterns and insights that aid growth opportunities, Alkin sought to understand the trends and patterns within the data in order to develop classification models for predicting various failure modes and the Remaining Useful Life of the machines.

“Working on a real world problem is always challenging and the datasets are never clean like the ones that are used for assignments,” Alkin said. “There is no perfect solution for a problem like this. I was lucky to have access to this dataset because of the partnership between the MScA program and Uptake. I learned a lot from this experience and am very grateful to Yuri Balasanov and Sema Barlas for their constant support.”