Treating the cause and not the symptom

In 2023, you might think that AI/ML has solved every problem under the sun.

But as with most things in life, we need to cut through the hype and look at things differently – in this case treating the cause of an issue, and not the symptom.

?

We had a prospect all too recently who was very proud to have solved a use case using via AI/ML.

In this specific case, they wanted to address the issue of predictive maintenance of punching tools. The goal was, among other things, to minimize setup times and downtimes.

But – our prospect’s management team was disappointed in the results delivered by this AI/ML “miracle cure”.

So what had gone wrong?

• The customer wanted to know when to replace a tool early before it became blunted.
• They had used image recognition to record the result of the punched holes and compared them with a reference image.
• The problem here was that if the AI decided that the hole was of poor quality, it was too late as the tool was already blunt. Even worse, because the machine would now be making considerably more noise, an experienced worked could have detected the issue themselves.

Therefore, the management was also of the opinion that AI/ML is not for them, because for all the investment, it is only telling them something they already know.

As for the approach to solve the problem, we were a bit puzzled. However, this was not due to the Data Scientist, but to the problem definition, which was probably a bit too technical. They had forgotten to give the Data Scientist some assistance.

The solution, and it's not the only one, was quite simple and also easy to implement.

What happens to a punching machine when it has to work with a blunt tool, or let's say a tool that is no longer sharp?

It has to increase the force to create the hole, or in other words, the power consumption has to be increased to make this possible.

The solution was simple: measure the power consumption of the machine and thus detect an anomaly, or just the need for more power.

We have been able to show in a simple example within 2 days that AI/ML can already predict the change of the tool 24 hours in advance.

Now, during a shift change, the tool can be changed and therefore there is no more downtime, quality was increased, scrap was reduced and the customer was happy.

The takeaway from this use case is very simple, had the Data Scientist been given support in the form of an experienced machine operator, the solution would have been found relatively quickly.

The Data Scientist knew from his training the subject of image recognition for quality detection. If you have a hammer everything is a nail.

Here we are again with the all-familiar: Why do AI/ML projects fail?

Artificial Intelligence (AI) and Machine Learning (ML) are enabling managers to increase their decision-making capabilities like never before.

The areas of application for AI are becoming more varied. At the same time, technologies are rapidly evolving. This makes it increasingly difficult for companies to identify the application areas that are critical to their value creation. Key challenges are in particular the determination of the concrete cost-benefit ratio.

My conclusion here is:

The only long-term solution? Deal with the issue yourself and learn to recognize and understand the causes. If you can't do that, then find an expert who will take enough time for you.

We need to focus on trying to solve the treatment of the cause and not the symptom.

Therefore, data democratization is one of the key recipes for success. Involve specialists and empower them to use data science.

Since we have had to see these very often now. My question to you, do you also encounter such cases. Or do you possibly even have other examples.

I look forward to your feedback.