James Steidl - Fotolia
With more than 50 million members and patients nationwide, medical insurer and healthcare provider Highmark Health generates a lot of documents that need to be mailed -- invoices, explanation-of-benefits forms, patient letters and more. The nonstop flow of paper -- amounting to more than 1 billion pages and 128 million envelopes last year -- keeps the Pittsburgh-based organization's printing and mailing operations working 24/7 at two facilities in Pennsylvania. "We really can't afford to be down at all," said Paul Jorgensen, director of output services at Highmark's HM Health Solutions subsidiary.
The quest to avoid downtime led Jorgensen to sign on earlier this year as a beta user of new predictive maintenance software and services from Pitney Bowes Inc., which makes the production printers and mail inserters used by HM Health. Data analysts at Pitney Bowes are running automated algorithms in the Clarity Advisor software against streams of operational data from the mailing systems in an effort to detect signs of potential equipment failures. In one case, they found anomalies pointing to a looming problem with a vacuum pump seal in an inserter. "That allowed us to fix it on our own schedule instead of having the seal fail at two in the morning and that shift grind to a halt," Jorgensen explained.
Other companies are launching similar initiatives to more proactively maintain manufacturing machines and other types of industrial equipment to keep them running and prevent breakdowns. This once lofty goal is becoming more feasible thanks to a combination of internet of things (IoT) technologies, big data platforms and predictive analytics tools.
Predictive maintenance is "the Holy Grail" of the industrial IoT, said Heather Ashton, a manufacturing industry analyst at IDC. "Everyone is talking about it." But it's still an emerging capability, she added. In a survey of manufacturing organizations conducted by IDC last year, 65% of the 330 respondents said they had connected one-quarter or less of their products in the field to the IoT; none had linked more than half of installed products to a network. Looking three years down the road, only 12% said they expected to top the halfway mark on IoT connectivity by then [see "Slow to Make Connections"].
And a predictive project isn't a simple or inexpensive undertaking, according to early adopters. It can take "troves and troves of data," as Ashton put it, to make users fully confident about the findings of predictive maintenance algorithms and queries. All of that data needs to be collected, processed and analyzed in near real time, which requires a robust IT architecture -- often involving big data systems built around Hadoop, the Spark processing engine and related technologies. Organizations may also have to build up an analytics team with machine learning skills and make big changes to internal business processes.
Early on the predictive curve
"Frankly, we're very early days on this," said Grant Miller, a product management vice president at Pitney Bowes who's helping to lead its predictive maintenance efforts. The Stamford, Conn., company began beta-testing Clarity Advisor last December, along with companion applications for optimizing mail insertion workloads and dynamically scheduling jobs. It launched the application trio commercially in North America in March and plans to roll them out in Europe this quarter and the rest of the world by early next year.
To speed development and deployment, Pitney Bowes built the software on General Electric Co.'s Predix platform, a cloud architecture designed to support industrial IoT applications. Miller said data captured from mail inserters at customer sites is processed and stored in GE's Hadoop-based cloud environment. In creating the Clarity applications, Pitney Bowes also used data science methodologies and statistical analysis tools offered by GE as part of Predix.
Customers can use the predictive maintenance software to analyze the operational data themselves, but Miller thinks most are likely to rely on the analytics team at Pitney Bowes to do the work and alert them to possible equipment issues. For now, there's a delay of a couple seconds in transmitting data from mailing machines to the Predix cloud for analysis. "But we're eventually going to make it a true real-time [system]," he said. "For us, it has to be something people can interact with immediately" -- before inserters break down and mailing operations fall short of their service-level agreements on mail volumes and delivery times.
Pitney Bowes, which has a Clarity user base in the low double digits, expects to be collecting petabytes of data by next year as more customers sign on. The predictive analytics capabilities should get better as the algorithms are run against more and more data, Miller said; the same goes for benchmarking analysis that the company plans to do so customers can compare themselves against other organizations on mailing performance.
Predictive maintenance ups and downs
ThyssenKrupp Elevator AG also is in the early stages of launching predictive maintenance applications. The applications are designed to identify elevator components that need to be repaired or replaced before they cause unplanned service outages. Like Pitney Bowes, ThyssenKrupp is using a cloud platform to expedite the deployment -- in its case, Microsoft Azure, including the software vendor's HDInsight managed Hadoop service, Azure SQL Database relational data store and Azure Machine Learning analytics technology.
"In the beginning, we planned to build our own system," said Sascha Froemming, head of innovation and sustainability management at the Essen, Germany, elevator maker. "But very quickly we said, 'Come on, we're not an IT supplier.' " Still, ThyssenKrupp is designing a homegrown sensor unit to pull operational data out of elevator control panels and send it to the Azure cloud via the IoT -- an effort that Froemming said has posed software development challenges the company is still working to resolve.
A lack of internal analytics expertise was another hurdle confronting ThyssenKrupp. "We didn't have a couple hundred data scientists already working in the company," Froemming said. So the manufacturer has turned to Microsoft and third-party analytics service providers for help in creating predictive models while it sets up a full-fledged analytics team. It also has had to revamp field service, dispatch and training processes for its 20,000 elevator technicians and amend its customer service contracts to incorporate the predictive maintenance capabilities.
Nonetheless, ThyssenKrupp has an aggressive rollout schedule for the data collection units and predictive maintenance software. After initially testing out the technology on a couple hundred elevators in the U.S., Germany and Spain, the company planned to expand to as many as 2,000 in May. Within 18 months, it expects to have 150,000 elevators connected to the Azure platform, each transmitting data at least twice a day. "I don't want to call it a trial anymore," Froemming said, adding that the development team is "feeding Azure Machine Learning like hell" to harden its first two analytical models.
Already, the models have pinpointed some data anomalies that were confirmed as indicators of possible equipment failures upon visual examination of the affected elevators, according to Froemming. Going forward, company officials want to create separate predictive models for doors, motors and other elevator components to increase analytical accuracy. The ultimate goal of the initiative, he said, is converting field service from a focus on troubleshooting to "a more advisory mode," which could help tighten customer relationships and reduce service costs on both sides.
In addition to minimizing unplanned equipment downtime, predictive maintenance tools can enable both manufacturers and their customers to avoid "the cost of sending out a crew at a moment's notice to fix something," said Bill McBeath, an analyst at ChainLink Research. "Emergency fixes are much more expensive than scheduled ones."
A spectrum of predictive possibilities
But while McBeath sees a lot of interest in the predictive path among the companies he works with, he said real initiatives are still "in an experimental phase for the most part." He also noted that organizations often start with preventive monitoring -- setting thresholds on various equipment metrics with alerts generated if those thresholds are exceeded. "There's a spectrum of steps toward the more predictive stuff," culminating in the use of automated algorithms to analyze data in real time, McBeath explained.
Bill McBeathanalyst at ChainLink Research
Data storage vendor NetApp Inc. is another company progressing to the far end of the predictive maintenance spectrum. As part of a program called AutoSupport, NetApp has been collecting configuration and operational data from its storage systems for about 20 years -- first through weekly disk mail-ins by customers, then via automated feeds, also sent once a week. In 2012, the Sunnyvale, Calif., company deployed a Hadoop cluster based on Cloudera Inc.'s distribution of the open source framework to process the data. NetApp also upgraded the software that controled the storage devices to send daily snapshots of their settings and condition.
Two years later, NetApp added a second Hadoop system as a data lake for its data scientists, who analyze the collected information to identify configuration issues that could affect individual customers and potential problems with disks or other components. Now the company is working to turn the original intake cluster into a stream processing platform that will pull in data from storage equipment on a continual basis instead of waiting for the daily batch jobs. "We'll be able to process data as soon as it's created," said Shankar Pasupathy, a technical director at NetApp.
The data will then be fed immediately into the analytics data lake, Pasupathy added. He wouldn't disclose a scheduled launch date for the expanded architecture, but said development work is "well underway." The new setup will augment the core Hadoop platform with the Spark processing engine's Spark Streaming module and Kafka, an application-to-application messaging system. NetApp is also looking to add machine learning tools to the mix to further automate the predictive maintenance efforts. "We're hoping to do more real-time analytics in the future," Pasupathy said.
Paul Jorgensendirector of output services at HM Health Solutions
That kind of real-time window into the internal workings of industrial machinery is a big draw for an operations manager like HM Health's Jorgensen. And being able to tap the predictive maintenance software developed by Pitney Bowes to go beyond tracking output statistics and doing postmortem assessments of equipment failures in his mail inserters is a particularly logical capability for someone who has spent his entire career in the healthcare industry.
"I come from a mindset that prevention is a good thing," Jorgensen said. "It's like when I go to the doctor: I don't really want them to find anything, but I'm glad if they do."
Predictive maintenance work helps a utility keep its turbines running
Big data architectures required to drive IoT analytics initiatives
The predictive modeling process shouldn't be an academic exercise