IoT Spotlight: Predictive Maintenance and the Promised Land of Zero Unexpected Downtime

What is predictive maintenance?

If you drive a modern car, chances are you are already very familiar with predictive maintenance. All modern cars are equipped with 100 or more sensors to continuously monitor performance and predict part failures well before they fail, to give you a fair warning and enough time to schedule a service appointment.

The much dreaded “Service Engine Soon” light, one that often gives you nightmares, is actually your friend. One that has your back. It makes sure you are not stranded on the side of a freeway while driving to an important meeting, or while driving to the airport to catch a flight.

This concept similarly applies to manufacturing equipment or oil rigs. If we can predict a part failure well in advance, we can schedule maintenance/repair work for the part as per our convenience, while continuing to operate the equipment to avoid unexpected downtime. This will help reduce large repair expenses, as the part will be repaired or replaced well before it fails, stopping it from causing further damage to the equipment. Also, the productivity loss from sudden and unexpected downtime will be minimal. The repairs can be done at an optimal time to ensure minimum impact on productivity. Reduction in both maintenance cost and loss of productivity can have a significant impact on the bottom line of a business. Also, you would not need to schedule downtime just for the sake of inspection anymore, which results in productivity gains.

Predictive maintenance: A step forward from preventive maintenance

Preventive maintenance is the act of scheduling regular equipment maintenance based on the mean time to failure. It is similar to servicing your car every 10,000 miles or changing the oil every 3000 miles. Since you are not monitoring the equipment continuously, you can only hope regular maintenance is good enough to preempt any kind of  equipment failure. Basically, you are vulnerable to unexpected failures of your equipment.

Predictive maintenance is a much more sophisticated form of maintaining your equipment. Instead of only servicing the equipment at a regular frequency (i.e., preventive maintenance), you monitor the equipment continuously to detect any signs of possible failure. This strategy gives you much more control over the operations of your equipment. You can continue to schedule regular maintenance, since preventive maintenance is essentially a good practice that helps you to take care of your equipment, but at the same time you have the ability to check the health your equipment at any given time, predict when a part is going to fail, and avoid unplanned downtime.

How does predictive maintenance work?

Predictive maintenance requires data to be collected from the equipment sensors on a frequent basis, “normal behavior patterns” then need to be established with the help of data, and finally, the data is analyzed for possible patterns of failure using data models. The following steps can serve as rough guideline on how to set up predictive maintenance for your equipment:

Collect machine/sensor data:

There are many types of industrial sensors available in the market depending on the equipment. Most modern equipment comes with pre-built sensors. You will need to collect all sensor data in one central location where it can be analyzed. As part of the data collection process, you need to make sure that you have tested end-to-end network connectivity from the sensor to the backend server. This is a very important step, as the reliability of the end-to-end connection will determine whether you can trust your system to provide you with predictions about the health of your equipment.

Create models for failure patterns:

This is the brain of your predictive maintenance system. This is where all incoming events are analyzed to find failure signatures. Based on the historical data of failure, you can use multiple algorithms to create these failure patterns. Many of these failure patterns are correlated with multiple independent events, which signal impending failures. A failure pattern is often a combination of several factors that are hard to visualize, and hard to find via human inspection alone. However, sometimes one catastrophic event alone can signal impending failure, and therefore it is also important to have anomaly detection models in place, to raise an alert as soon as such an event takes place.

Machine learning:

You can use machine learning technologies such as Apache Mahout, Spark, MLlib, etc. to train your algorithms as new events and failures take place. This step ensures that you are not limited only to known anomalies and failure patterns; you can add new ones to your detection techniques as you collect more information about your equipment.

How can you get started?

The Stream Processing Quick Start Solution helps you get started with setting up real-time data workflows for event streams from sensors, statistical aggregations within stream processing engines, along with a search-based visualization interface to gain insights into outliers and trends. Installation and configuration of the MapR cluster is included within the scope of this Quick Start Solution.


Streaming Data Architecture:

New Designs Using Apache Kafka and MapR Streams




Download for free