by Andy Oram
November 15, 2015
This article originally appeared on the International Manufacturing Technology Show site.
There’s a rush in manufacturing and industry toward the use of sensors, analytics, and timely alerts to maximize the availability of your machines and tools. This movement, called predictive maintenance or predictive monitoring, gives your staff an unprecedented look into the innards of the infrastructure on which your manufacturing and your customers depend.
Is a gear getting out of line and crippling the productivity of a million-dollar lathe? Is some key support developing hidden cracks that will cause it to collapse? Sensors and analytics can reveal such problems, and alerts can prompt your staff to do something to rescue the situation. A recent article discusses techniques such as thermal imaging, vibration analysis, and ultrasound to find wear and tear. Predictive analytics may even help child protection agencies find at-risk families—although that approach was also likened to 1984’s Big Brother.
The difference between traditional preventative maintenance (where you bring a machine in for repair at regular intervals) and preventative maintenance is like the difference between going to the doctor for your annual check-up and going to the emergency room when you feel chest pain. Both preventative maintenance and predictive maintenance are valuable, but you can save a lot of money and avoid risk with predictive maintenance.
A recent posting of mine laid out the value of predictive maintenance and the practices that have to be in place within your organization to make such maintenance productive. This article will delve a little deeper and look at some current trends.
A presentation on September 30 at the Strata + Hadoop World conference in New York City showed how to create a predictive maintenance software system and keep it up to date with new lessons from the field. Speaker Yan Zhang pointed out that the knowledge underlying predictive maintenance starts in the lab. One can run engines under a number of different conditions (adjusting temperature, humidity, dust levels, etc.) and find out when they fail. Even more powerfully, one may discover changes in the engine that predict failure.
But this is only the starting point for efficient maintenance. Once your engines and your predictive system are in the field, you can keep monitoring them for new information that will help you make better judgments concerning when to repair or replace an engine.
Data comes from the machines themselves, from operators, and from the environment. Typical data of value includes failure history, repair history, machine conditions (real-time data coming from machines), fixed features of each machine, and operating conditions (such as whether machine is in a building with air conditioning, or who is driving a car).
An evolving system uses machine learning: a kind of programming that looks over historical data and finds relationships that don’t immediately spring to a human’s attention. These systems generally require you to divide the data you know about your system into two sets: a training set and a test set. It’s a bit like going over some math problems in front of a class, and then assigning homework for the class to do that evening.
Machine learning draws inferences from the training data, through standard statistical techniques such as regression analysis and clustering (finding which items are similar). But quirks in the training data can lead to incorrect inferences, so the test data shows you whether you’ve really got the expertise you think you have. As you collect data out in the field, you continue to train and test with it.
Zhang also presented principles tools for choosing when to send out an alert, giving staff time to respond, and how to display statistics on maintenance schedules through useful visualizations. She also pointed out that the data we commonly collect now—used for anomaly detection and control—may not be the most useful data for predictive maintenance.
General Electric, a leader in the use of predictive maintenance, has recently spun out a general platform called Predix for companies who offer predictive maintenance tools. A comprehensive solution offering cloud storage and a developer toolbox, Predix supports software on the devices, tools for transmitting data to the cloud, and analytics in the cloud.
An executive overview lays out the context for Predix, and an in-depth white paper (no longer available) discusses how developers can exploit it. This development could herald an onslaught of platforms and tools to bring predictive maintenance to every sector that could benefit from it.
This work is licensed under a Creative Commons Attribution 4.0 International License.
Author’s home page
Other articles in chronological order
Index to other articles