In today’s video in the Machine Learning myth buster series, I’m covering the following myths:
- Machine Learning only makes sense when we go “all-in” and cover an entire production line or factory
- My production line is not prepared for Machine Learning
Myth 1: "Machine Learning only makes sense if I go all-in. That means, when I cover the entire production line."
The question here is: “How do I get started and where should I look first?” In an industrial manufacturing, where we have complex production lines, even small pilots can already add value and get us a long way on our smart factory journey. But what makes a pilot project a good one?
Any good Machine Learning pilot project is characterized by the following three aspects:
- A pilot should always be compared to and increase the status quo
- After a successful pilot that covers only a part of a production line, value-add increases significantly when we add other process steps
- Beyond the pilot, we can then at some point analyze the entire process holistically
This means that, while the end goal is gaining actionable insights over the entire production line with quality predictions, predictive and prescriptive maintenance, it is very simple and highly valuable to run Machine Learning pilots first on select parts of the production line.
For instance, if we choose our pilot well, we can uncover relationships in our process that we might have suspected in the past but were never able to test out or prove, or even relationships that we had no idea existed in the first place.
Myth 2: "My production line is not prepared for ML. I need to retrofit a lot of sensors first."
Every fairly automated process inherently has a lot of sensor data. Many machines and process steps are already connected to an intranet or remotely accessible, e.g. to access MES or ERP systems or for remote maintenance. In consequence, sensor data is being generated, often in a standardized format. This does not necessarily mean all this data is stored or even transmitted somewhere outside of the machine (yet). But often times, we can get this data into permanent storage with reasonable effort.
Ultimately, data needs to be obtained somehow. In terms of pilots, we can use a small dataset first - even if we might have to export it via a USB stick in the first run.
And yes, in some cases, we ultimately might have to retrofit some sensors. But that brings us back to the goals of a pilot project: a well-chosen pilot helps avoiding the situation of having to retrofit immediately. Only when we decide to go forward with integrating additional process steps, we get to this point.
It is not hard to get started with machine learning in industrial manufacturing. A feasible pilot use case exists in almost all fairly automated production sites. So, in order to get started, a simple pilot is better than a large, potentially overengineered project.
If you’d like to find out more about how to pilot Machine Learning in your production, attend our “Leading the Factory of the Future” Masterclass, where you can see how Siemens, Porsche, and BMW set up their smart factory and how they’re profiting from it.