Industry 4.0 or Lean 4.0?

Cape Town is sometimes called the “Silicon Valley of South Africa”. Even so, it was a pleasant surprise to learn that Atlantis Foundries had become a leader of the foundry industry not only in Africa but worldwide, using artificial intelligence (AI) and machine learning to achieve zero defects in the parts it delivered to its customers in Europe and the United States.

In this interview, Pieter du Plessis, the CEO of Atlantis Foundries at the time of this initiative, tells the story (which you can read or watch below) of how they used AI to solve a strategic business problem. Considering all the hype that Industry 4.0 has received, there are some valuable lessons here from someone who actually made AI work in the real world.

There is a lot to learn from what Pieter shares with us, but I would like to highlight one aspect in particular. It has become fashionable to use the term "beyond lean” in the rush to market Industry 4.0 technologies. Certainly, these new technologies are game changers and cannot be ignored, but I think referring to them as "going beyond lean” is the result of a fundamental misunderstanding of what lean thinking is.

Before introducing AI, Atlantis Foundries invested in continuous improvement for many years to improve its defect rate and reach the industry standard. This means that, by the time they began experimenting with AI, there was already a problem-solving culture in the organization on which they could build. You will hear Pieter refer to AI as another “tool” on the improvement journey. In other words, just as we use the five whys as a tool to find root causes, there is now this very powerful tool called AI that can be used in situations where humans cannot cope with the amount of information that needs to be processed.

Like any other problem-solving tool, AI can be used incorrectly or inappropriately if you do not apply lean to the problem at hand. Asking what problem we need to solve has to be the first step, which will inform our problem solving and make it effective. Atlantis Foundries thought very carefully about this question, and it guided the company as it figured out how to exploit AI.

This story shows us how important lean thinking is to the success of new technologies such as machine learning and AI.

Anton Grütter: Pieter, can you introduce Atlantis Foundries to us?

Pieter du Plessis: Atlantis Foundries is an offshoot of Atlantis Diesel Engines, an engine plant bought in the late 1970s. Atlanta Diesel Engines went bankrupt, but Atlanta Foundries was bought by Daimler/Mercedes Benz to continue casting cylinder blocks for trucks. Daimler already had a foundry in Mannheim, Germany, but with the country's strict environmental laws it could not expand its capacity. So it bought Atlantis Foundries in 1999 to cast these parts. Atlantis Foundries went on to export all of its engines to customers in the United Kingdom, the USA, South Korea and Germany.

AG: Can you explain to us why quality was so important for Atlantis Foundries?

PDP: Being so far from the customers we serve means that, if you have any quality problem detected in the machining operations at the customer's plant, the entire pipeline could be contaminated. The cost of searching for quality defects in the pipeline – which can take six to eight weeks – is enormous. That is why delivering the right quality the first time is so important to us. Our quality costs are significant.

AG: How did you initially try to solve this quality problem?

PDP: With the classic quality tools, such as statistical process control, control plans, process sheets and so on. Anything traditional companies would do to improve quality. However, we reached a point where we could not advance any further, because the feedback loop was too long, and the process was very complicated and full of variables. Finding a pattern in those variables was difficult. A benchmark foundry making the same parts as us typically had between 4 and 8% scrap, and we were at around 6% – with an additional 1.5% at the customer's site. We were in the middle of the pack.

AG: What made you decide to experiment with machine learning?

PDP: We wanted to predict which part would turn out defective at the customer, in the USA or in Germany, so that we would not even ship it. Since the defect is a function of the process, there must be a pattern in the process that can be recognized. That way, you know that if the process is a certain way, it will produce a defective part. We are talking about defects that are hidden and cannot be seen physically, nor found with ultrasonic methods, because of their size and volume. We wanted to predict, based on what the process was at that moment, whether a part would be scrap or not. In the end, we had a 70% success rate. We would say “throw these 10 parts away” and we knew that seven of them would have become scrap. We knew we would be throwing three good parts away, but the associated cost was insignificant compared with what was being spent in the USA.

AG: When did you make that decision?

PDP: We started with machine learning in March 2017, giving “Data Prophet” [machine learning specialists in Cape Town] 18 months of process data and quality results (internal and external) that we had accumulated.

We did not need to generate new data. We just took the data we had every day but had never used. The hardest task is matching the data with the part you make: we normally used time series – checking the process every half hour, every 40 minutes or once a day –, but our production rate was one part every minute. We used to make 700 parts a day, an enormous number if you think about it. Of the 180,000 truck engines we had already built, at first we had perfectly matching data for only 7,000. So we used that data to build the first model. As you feed in more data every week, the algorithm learns more. You can also start doing your tests to correspond to the exact parts, which is quite critical for the process: when we make part number 1 or number 20, we know which parameters we should check, instead of checking randomly at regular intervals. Essentially, our testing methods changed.

AG: How does this work in practice?

PDP: There are several types of artificial intelligence and machine learning. There is supervised and unsupervised learning. With unsupervised learning, the answer is not known; you give the machine the raw data and it finds the patterns there. With supervised learning, we tell the machine whether the result was a good or a bad part and provide the process parameters for each one; the algorithm uses brute force, calculating all the iterations until it arrives at an answer (however, we had 185 variables in the process, so you can imagine the complexity).

AG: Were there problems with people? Did they accept this?

PDP: We had problems. There were many who did not believe in it, including me. I was very skeptical at first, because I did not understand how the specialists could just ask for data without understanding our process. Some of the engineers feared for their jobs and thought AI would replace them. I underestimated this, because I thought intelligent, well-educated people would not be afraid of AI. It took a while for people to understand it and use it, but the results were spectacular.

Using the model, we wanted to predict which part would become scrap in the USA and, in the second phase, learn what we had to do so as not to make parts that would become scrap. The model that predicted scrap went on to prescribe the process. It took process parameters and calculated the region for all of them (min 11), which meant that our process limits were squeezed significantly, and some of them were adjusted. The combination of these two things allowed us, in the second half of 2017, to achieve zero defects at three plants for three months in a row. That was unprecedented.

I spoke with some people recently, and this trend continues – even if zero defects is not achievable all the time, because the process parameters become very tight over time. And that is why people will always be important in this AI thing: they will have to analyze that report and decide what they need to do on the machine so that it achieves the desired result. You cannot tell a machine: “You can't do that”. Machines are not that smart. The notion that AI will replace all human beings is a fallacy. Maybe two centuries from now. Certainly much later than many people think.

AG: What advice do you have for people interested in Industry 4.0?

PDP: I think many people are having a hard time understanding what Industry 4.0 really means. They say: “We have data, let's do something with it”. To me, that approach is wrong. They should ask themselves what the biggest problem in their business is and how this additional tool can help them reach their goal. At Atlantis Foundries, for example, we wanted to avoid scrap in the USA and Germany. Once we were able to predict which part would be defective, we wanted to find out what we had to do to prevent such parts from being made. You need a clear objective. Just doing Industry 4.0 for its own sake is a waste of time.

AG: There is more and more talk these days about going “beyond lean” when discussing technologies such as artificial intelligence. What do you think about that?

PDP: In my view, lean seeks to simplify processes, reduce waste and make work repeatable and efficient. To me, AI is another tool that lets you do that. It is not a silver bullet; it is just another tool. As I became more involved with it, I quickly realized how far we are from machines doing everything on their own without the need to involve people. We are incredibly far from that. The human aspect is important, because part of the process is human influence. If an operator does not follow the rules, you can have all the artificial intelligence in the world, but you will still have a problem. So, to me, AI is not beyond lean (in English)... it is another tool for lean.