Proactive maintenance is in the data


Posted on: May 14, 2019 by Sarah Ewing

In the giant space of the Industrial Internet of Things (IIOT) detecting when equipment starts to break, before it breaks, is a cost-effective way to move from equipment downtime into proactive maintenance. Proactive maintenance is the act of preventing failures through harnessing data to engage equipment and processes with helpful measures before a point of no return.  This moves maintenance from reacting to a downtime event, to proactive responses, scheduled around the ebbs and follows of the process, increasing overall equipment effectiveness (OEE).

Rather than reacting to a failure, or creating a maintenance schedule, proactive maintenance can be used. Proactive maintenance occurs during natural off hours before the equipment breaks, preventing a stop in operations from ever even occurring and creating business optimization. With the power of machine learning, advanced analytics and big data this will be an implementable solution.

The Process

First, raw data is delivered, parsed and then loaded into a database with the aim of focusing on writing algorithms that check for a component heartbeat and anomalous behavior. A piece of equipment is said to have its own heartbeat when it’s preforming regularly within its expected functions.  The heartbeat sometimes becomes warped as the parts age, and thus is performing beyond its expected functions.

When the heartbeat of the equipment starts to differ from its regular functions, we want to generate an alarm to indicate that proactive maintenance needs to be performed because an anomalous behavior is impending. Which is just a nice way to say that the equipment is about to break. Key to proactive maintenance is creating an algorithm that identifies when an element, or other component, deters from regularity.

The Data

It is difficult to make generalizations about the regular heartbeat that applies to all the systems and components. For example, manufacturers release statements about the longevity of components, but this is not true for all components all the time. Taking that gamble on something as integral as a brake system could be disastrous, thus the heartbeat for each sensor, component and system needs to be explored individually across hundreds of components and timeframes.

Machine Learning

Many machine learning approaches are typically considered to do predictive maintenance, including logistic regression, support vector machines, nearest neighbor, clustering, outlier identification and decision trees. The most accurate solution produced is a custom algorithm that creates work orders or alarms for maintenance to be completed before a downtime occurs.

The custom algorithm specifically parses the component’s data into windows and looks at the rate of transition between the different states. It has been discovered that the length and severity of anomalous behavior correlates to the likelihood of a failure.  The stronger and longer the anomalous behavior is recorded, the more likely it is for an equipment failure to occur.

Unique Solutions

When algorithms automatically calibrate on historical data for each individual sensor it accommodates the different behaviors for each component, as some naturally have a faster heartbeat than others. This also reduces the number of false positives, as the ground truth cannot be generalized across components and systems. This automatic calibration also provides an easy to understand output that identifies when failures are most likely to occur and gives maintenance a heightened awareness when upkeep needs to take place. This allows for preventative actions to be taken during a natural downtime or off hours, so that production will not be affected. The custom alarm algorithm easily translates onto many different types of equipment.

In conclusion, predicting when maintenance needs to occur before an integral part fails is crucial.  And more important, it’s a luxury that hasn’t always been available until certain advances in technology. Predictive maintenance produces a safer and more productive workplace with less downtime and increased worker satisfaction, as well as allows for business optimization and maintenance to be performed during off hours before failure events ever occur.

Share this blog article


About Sarah Ewing

Senior Data Scientist and member of the Scientific Community
Sarah Ewing is a senior data scientist at Atos North America. She serves as a member of the Atos Scientific Community, a global network comprised of 150 of the top scientists, engineers and forward thinkers from across the Group. In her role, Ewing creates predictive maintenance algorithms on millisecond data for integral processes at a common household brand. She has experience in data science spanning various industries, including medical, nuclear, psychological and engineering. She has authored more than 13 peer-reviewed publications relating to data science and analytics. Prior to Atos, she worked for the federal government as a senior statistician creating models for the United States Nuclear Regulatory Commission (NRC). Ewing has a master’s degree in statistics from the University of Idaho. In her spare time, she serves on the board for Middleton Food Pantry, helping to eliminate hunger in the community by distributing food to those in need.

Follow or contact Sarah