SAP for Mill Products Blogs
Strengthen your foundation of SAP knowledge with practical tips and insights for optimizing your production processes and achieving operational efficiency.
Showing results for 
Search instead for 
Did you mean: 
0 Kudos

How to approach predictions and machine learning according to master Yoda

“Size matters not. Look at me."

Image by Ciro Duran from Caracas, Venezuela (Master Yoda) [CC BY 2.0], via Wikimedia Commons

Let’s get one thing right from the start. The internet of things, and machine learning are not just for the Fortune 500. Many innovative SMBs have embarked on this journey. One of my favorite stories is a specialty paper manufacturer from the black forest, the heartland of innovative small companies in Germany. They have taken a very pragmatic, exploratory approach – starting with a few use cases and question, connected a few data sources (less than 100 input variables from many different sources including QM, MES, ERP, Energy, and many process control parameters).

They have plugged this into their existing manufacturing landscape, added a SAP HANA database and used predictive analytics capabilities to start their learning journey towards predictive quality.

Find more information about their approach linked here.

“You will find only what you bring in.”

This seems obvious, like a different way of saying “garbage in, garbage out”. But actually, this is deep Yoda wisdom, and a tricky business case. When you start with a classical QM and process control data set, you may surely find correlations and patterns. What has proven to be very powerful in several IoT projects at mill products companies is to go broader and look beyond the operational technology (OT)-data set:
Include CRM complaints or warranty claims. Include ERP plant maintenance information. Include QM data from ERP, the lab, or even your supplier (if they are willing to share). Blend this with streaming sensor data, historian information, image/camera data.

I do admit that this requires some serious data preparation, and comes at a cost.
Also, not all information may be available. If you do not sample every piece of raw material you use (e.g. for cost reasons), you may not have the crucial input variable to predict quality. This comes down to the question how much value you generate (higher quality, less scrap, lower cost of quality, faster grade chances) compared to the cost of acquiring and analyzing data.

“Many of the truths that we cling to depend on our point of view.”

I recently had a very interesting discussion with a paper & packaging company about corporate culture and how to talk about quality. This integrated company had an internal supply chain across multiple intermediate products. To predict customer value (or quality defects that impact the customer), it would have seemed logical to analyze the entire value chain across departmental boundaries. This would have required, in the first place, to share quality (and defectst) with the next department. Or to put it differently: to “externally” admit that you do not always produce perfect quantity.

The cultural change to admit & share this even across company borders is even more daunting. But there are companies I know of who do exactly this – actually in both directions. They share quality information about the pulp they produce (to help their customers optimize their downstream process), and in return connect to the machines of the customers to support them when they run into problems, thus advise how to run optimally.

Bad quality may not actually be so bad. Quality in metal and paper is not a black and white situation. We deal first of all with ranges and tolerances, and one roll may still be ok for one customer, but not for another. Also, we rarely have the same quality every where on a roll. Most areas of the paper roll may have the intended quality but could be a few spots and holes as well. Instead of scrapping the complete roll, it may be smarter to share the quality details along the supply chain to cut up the roll in such a way that all holes and spots are avoided. Thus the total quantity of available material is slightly reduced, but we do not need to scrap the roll. Plus we can deliver on the agreed time.

“Mudhole? Slimy? My home this is!!

Empathy is important. Data science and prediction models will not find trust by decision makers and your process & quality engineers if they are not “interpretable”. Your deep learning algorithm is a black box. A very good one. But difficult to convey and trust. The interpretability of model is an important aspect, even more when regulators or society are affected. Or when the decision maker does not trust your predictions.

These are slippery grounds. On one hand, we would like to use the best prediction model available. On the other hand, if we do not win trust and support from the stakeholders, we will get nowhere.

My point of view: You will have to built trust first, later you can move on to uncharted land.

“You must unlearn what you have learned”.

When hearing of AI being used for asset maintenance and service, the first thought that comes to mind is a predictive maintenance use case. Mining giant Vale is actually using AI for image recognition & classification to avoid costly process errors in spare parts requisitioning.

Finally, there are many possible ares to apply machine learning. "Not everything is appropriate to be automated. Most companies cannot rip and replace all their processes with machines, but are looking to strike a balance between what tasks an agent accomplishes and what can be handed off to intelligent solutions." [Customer service automation in 3 steps with AI and machine learning. ]

"Secret, shall I tell you? Grand Master of Jedi Order am I. Won this job in a raffle I did, think you?"

Finally, this your  learning journey. It may be helpful to get a Yedi master to learn some of the darker secrets of data science. It may be equally helpful to attend a course at openSAP:

Enterprise Machine Learning in a Nutshell

Enterprise Deep Learning with Tensor Flow

Data Science in Action: Building a Predictive Churn Model