Plant managers check out time-series data for close-up analysis

In many plants, sensors function as a kind of automotive “check-engine” light. The information they impart lets the plant manager know it's time for preventative maintenance, or that a key piece of equipment is about to fail.

The problem lies in the amount of information that sensors impart. First up is how to manage it. Sensors can take millisecond measurements, and storing and making sense of those numbers becomes overwhelming.

Time-series information—literally a list of measurements and the time they were taken—can be archived in what are called standard relational databases. But users often have trouble making sense of those long lists in any meaningful way, says Stephen Johnston, director of server software at Logical Information Machines (LIM).

To envision a relational database, think of Microsoft Excel—or a number of boxes arranged in rows and columns. These databases definitely have their strong points. They can be used for far-ranging applications and can account for many discrete numbers. But their strength also is their weakness, says Johnston. Information contained within a relational database isn't easily broken out or analyzed, and relational databases that contain huge amounts of information can perform slowly.

Johnston says that due to limited storage options, many companies still store time-series information in relational databases. “In general [they are used] because they're so generic, and can be applied to so many problems,” he says. “But if it's time-series you're dealing with, you can more efficiently store it.”

LIM's Historis is an alternative specially designed to optimize storage of time-series data using appropriate storage, retrieval, and manipulation algorithms. When Johnston says this system can upload more than 250,000 readings per second and extract more than 220,000 readings per second for analysis, the need for optimized data storage becomes clear. Once taken, the readings can be analyzed with LIM's analysis software or with a manufacturer's custom written in-house application.

LIM's query tools use an English-like language to return specific numbers. But why would sensor readings need to be taken as often as one millisecond? Some very finely tuned or important machines must be followed minutely—almost in real time—so that operators immediately know of abnormal operating patterns, Johnston says.

For example, a power plant turbine contains 500 sensors, and each is measuring operating aspects such as temperature or vibration or pressure, Johnston says. Each sensor might record a value every 500 milliseconds.

The plant typically inputs that information into data storage for months and years. Over time, operators may analyze data to get feedback on typical operating parameters and everyday patterns. Any discrepancies in those patterns might show a machine or instrument is on its way to a breakdown. A slight deviation in operating parameters alerts the manufacturer to schedule maintenance.

“A query might show me what happens five days after a certain condition repeats three days in a row,” Johnston says. “You might find out the instrument failed 100 percent of the time five days later.”

Analysis also helps operators plan downtime.

“If they see they'll have to do maintenance on a machine, they can mitigate any damage caused by taking the machine down,” Johnston says. “They can see that the machine is primarily active from midnight through 9 a.m., so they know not to schedule downtime during this window.”

Numbers like those returned by database analysis also can lead to lower maintenance costs.

“Rather than blindly scheduling maintenance on a set schedule—maybe they replace a part once a year—their data analysis shows a part doesn't need to be replaced as often as it has been,” Johnston concludes.