Using Predictive Analytics in a Smart Factory - Perficient Blogs
Blog
  • Topics
  • Industries
  • Partners

Explore

Topics

Industries

Partners

Using Predictive Analytics in a Smart Factory

So you want a smart factory?

Something that drives greater value by combining business processes and physical machines. Update your manufacturing processes from legacy automation to a fully connected stream of data. The solution translates a specific business problem into a data problem that can be solved with predictive analytics.

Many manufacturing systems operate at less than full capacity (Blanchard).  The smart factory uses connected Internet of Things (IoT) sensors to help systems make better decisions (Burke et al.). Watching real-time (R/T) sensors, with help of a predictive analytics module, can determine the best time for machines to be repaired, adjusted, or replaced (Li et al.).

The legacy state is a costly machine that needs downtime for inspection and general maintenance. Regardless if it is an best time to do so. Sometimes this machine fails when it is needed most. And, maintenance means cost of labor and materials (Blanchard). This is reacting by scheduling regular maintenance. Or at best, proactively addressing defects. However, the next step in getting better is to predict machine failures.

Predictive Maintenance (PdM) is the fixing of problems before failure occurs through the analytical comparison of known machine thresholds and recorded physical measurements (Vibralign).

Here’s the good news. Although shown to have great promise, PdM is still rare (Zenisek et al.; Yamato et al.). So, that’s what creates the competitive advantage. The PdM market is expected to continue growing at a significant rate (Columbus) to $12.7 billion by 2025 (Smith).

Disadvantages of PdM include the initial capital expenditure and training personnel (Vibralign). However, machine monitoring is getting cheaper with new sensors and analysis software (Mariappan et al.; Sciban). The newest software is greatly simplified. So, the associated training burden can be diminished.

Where to Start?

Always start with what you want to achieve (CRISP-DM). For example, something simple is “remaining useful life.” RUL is the available time until a set of sensor measurements collectively reach a given threshold (Yamato et al.). Therefore, PdM needs knowledge of error patterns in the data that happen at failure (Ravi et al.). Secondly, rank the subsystems. Focus can be given to the subsystems of higher value (Thoppil et al.).

So, simply put the data modelers with the machine techs. Yes, have a meeting. They will define the error patterns and subsystem rank. They will identify the model training data. Then, build some models. This is a laptop task. Yes, get good info without capital investment. The PdM model can even include costing to help you with the ROI.

Have a work out meeting. Build a model. Run it side by side for awhile. Check it out! Digital transformation does not need to be hard.

 

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.