SAIC has predictive analytics system ready for market

Engineering giant Scientific Applications International (SAIC) in McLean, Va, Friday told us it completed work on software that predicts failures in distribution and transmission systems days, weeks or months before they occur. Called Distribution Monitoring System, it is ready for release after nearly a year of research, said project lead Paul Halpin in an exclusive interview.

“Making predictions like this is a very, very hard problem and we've done some really smart things to solve it,” said Halpin, a mathematician and mechanical engineer.

The software consists of four pieces:

• A complex-event processing engine from San Diego's Zementis (SGT, Oct-26), that evaluates masses of data -- two or more years' worth -- against rules laying out the relationship between specific, fairly obvious events in the life of a device and the likelihood such a device will fail. For example, a rule might state that “if a transformer experiences a voltage spike above a certain defined operational level, which lasted longer than a specified duration and the voltage spread was greater than a defined range, then this is an event likely to cause a failure,” Halpin said. The number, severity and timing of voltage spikes and drops in a transformer are events of potential significance, he added, as are many other operational conditions such as secondary net-neutral faults.
  Millions of records containing event data can be run daily or in real time against those rules. The engine identifies equipment that is likely to fail within the next several days.
  
  
• A knowledge database that correlates faults and failures -- in effect, learning to spot the problems that cause failures.
• A neural network -- an artificial-intelligence module that can deduce from ongoing patterns of even minor equipment aberrations whether failure will occur at some point and can say, with a stated probability, when the failure will occur. The aberrations that the neural network detects are far smaller than those dealt with by the complex-event processing engine and the predicted failure times are farther out: days, weeks or months away. All neural networks tend deliver better predictions the longer they run, Halpin noted.
• A web-based user interface and workflow manager using open standards and Java. Through licensed use of Google Earth, the system can show on a map any transformer it predicts will fail and “if there are two transformers on a pole, it can show you which one will have the problem,” Halpin said. The workflow manager directs email alerts to the designated operator and tracks the repairs when made. If alerts are ignored, the system tells someone at a higher level. Distribution Monitoring System accepts data from a variety of sources. A firm in the Southeastern US, that Halpin said declined to be named, is using it with a backhaul data feed arriving via BPL. But it will also use data from SCADA and data histories, Halpin said.
  The software can be used in transmission systems by writing rules specific to transmission equipment such as intelligent relays and SCADA systems, though it is currently in use only on a distribution system, he added.
  
  “We designed it to be agnostic to whatever the data is -- the goal was to make it usable with either distribution or transmission,” Halpin said.
  
  SAIC plans to market the software worldwide alongside hardware and consulting services but as for how much it will cost, Halpin couldn't even report a ballpark, he told us.
  

Is SAIC on right track?

Experts in the field of predictive analytics told us the software could have potential.

“If they have the right data set, it sounds like a great application of the technology,” said Eric Siegel, PhD, president of consultancy Prediction Impact in San Francisco and a former computer science professor at Columbia University.

The software could even be used to make failure predictions in locales other than where it gathered the data, though accuracy might be a bit lower, said Karl Rexer, president of consultancy Rexer Analytics.

“The accuracy of software like this is going to basically depend on how big the data sets are,” Rexer added. “Data gathered over a year or more sounds good.”