skip to Main Content

Having trouble modeling your IoT data? Create the data you need!

The applications of IoT data are many, as you can see here.  The most touted application is predictive maintenance: foretelling when an asset such as a machine or component might encounter a fault, or fail outright. 

 

Sometimes we are not concerned about something as drastic as a failure.  Perhaps our machinery is not operating at peak performance.  Manufacturers measure themselves with a metric called OEE – Overall Equipment Effectiveness, of which performance is a key element.

 

IoT data can help tune the machinery to achieve better performance, but there is a catch.  In reality, the data from machines can be challenging to work with toward this objective.  First, as discussed in another blog post, there are underlying data generation issues that make things messy.  Second, sometimes the range of data generated does not cover the “space” of potential outcomes which may shed light on the optimal machine parameters and settings. 

 

While statistical models can be built to understand the relationship between settings and performance, they typically are able to provide insights only within the domain of data they are derived from.  If there is little, or even no information about certain ranges of parameters or settings, the models will have a difficult or impossible task of definitively providing confident insights and recommendations.

, Having trouble modeling your IoT data?  Create the data you need!

 

This is where Design of Experiments comes in.  DOE is a scientific approach to generating data for modeling.  It does so in an optimal way.  For example, one’s inclination may be to “just run the machine with a bunch of set points and let’s see what comes out.”  Well, there can be an infinite number of combinations of set points, and that approach still might not systematically cover the range of outcomes that merit exploration. 

 

DOE is mathematically optimal in the sense that it can take all of the factors (settings) and their various levels, and generate “runs” to produce the minimum amount of outcome data that should contribute to a statistically valid model.

 

The technical details of DOE are beyond the scope of this post, but of course you can always start with Wikipedia to learn more.  From a business standpoint, check out this HBR article written by our chairman, Tom Davenport.

 

DOE has applicably in many areas beyond IoT.  It is an often overlooked, yet mature and well-established tool that you should have in your toolbox.

 

 

 

Back To Top