Machine Learning & The Last 5%

Nathan Lai

Nathan Lai

Sales Engineer

I have a confession: I’m not really into cars. I know that might sound strange coming from someone who’s spent over a decade manufacturing them, but I didn’t become a mechanical engineer to work in automotive. I was just lucky enough to get my first job out of school at Linamar, and that laid the foundation for the rest of my career.

A decade is not a long time in the auto industry, especially given its history, but I’d argue that coming out of the Great Recession, the electrification of the car, and the beginnings of Mobility as a Service, the automotive landscape is rapidly changing.

We could very well see more change in the next 10 years than we have in the previous 50. (Don’t tell me your company hasn’t modified it’s product portfolio to accommodate vehicle electrification, whether it’s in R&D or straight up acquisitions.) Everyone knows the industry is changing; maybe we aren’t sure exactly when and where we will end up, but we know it’s inevitable. 

Industry 4.0 is the preferred term to describe the current paradigm shift in manufacturing, incited by a proliferation of disruptive technologies, particularly artificial intelligence (AI), additive manufacturing (AM), and the Internet of Things (IoT). The fourth industrial revolution is happening in automotive, but it’s a unique industry, and that means the pitfalls and possibilities of this transition are unique as well. 

What's So Special About Automotive Manufacturing?

One of the very first things I learned out of school was that manufacturing is a kind of know-how you learn from doing, not something you learn from a classroom. Sure, I learned about CCT curves for different steels in class, but it’s a whole other thing to work on induction hardener parameters to heat up steel to 900C in 10s and then quench it down to case-harden a part on the floor. Southern Ontario has some great automotive manufacturing firms with substantial know-how—it’s knowledge and experience that’s developed in-house and passed on internally. 

As much as there might be university courses on lean manufacturing or auto manufacturing, most of your knowledge comes from being on the plant floor. I’ve probably learned more from working with shop floor technicians than anyone, and I came to trust their experience over what I learned in books. (I’ll get the theory to match once when the shipment’s protected.) I truly appreciate all the folks who let me hold a light for them while they worked so I could learn the intricacies of all the operations that go into making a quality machined part.

Automotive manufacturing is defined by its pace, something everyone in the industry seems to understand: we all know what it’s like to get that phone call on Friday at 4pm. It’s not for everyone and those who stay in it understand the expectations. Surprisingly, it’s actually a small community. You end up running into the same people all over the place, especially if you’re in a niche market. 

When I was working on hydraulic products, I got to know the people that I worked with in our supply base but then went on to my product’s DRE’s at an OEM. People changed companies, but I’d still be working with them, just maybe at a different place. There’s something about the high-volume pace that keeps us in it, love it or hate it.

What really sets automotive apart from other manufacturing sectors, though, is the unique combination of aggressive productivity margins and demanding product requirements. That’s the challenge of automotive that most people don’t see. It’s one thing to build something that works 85-90% of the time, but that’s nowhere near enough for automotive. 

When I managed engine/transmission pump designs, the expectation was always that the product would work for at least 5 years without the user even knowing it existed. In casual conversations, I felt a sense of pride if someone had no idea what I made because that meant it was never a problem. 

Of course, you need more than just robust designs when you’re making millions of components and each one has the exact same expectation of quality. Your approach to quality is different at high volumes because you just can’t check every feature on every single part, sub-assembly, or assembly; you don’t have that luxury. 

I’ve worked on assembly lines that crank out a part every seven seconds that had over 30 components in it. Even if the quality rate was 90%, that’s still over 300 assemblies that need to be reworked or scrapped every shift. If you’ve found yourself in that situation, I hope you have the floor space and you’ve factored in the teardown labour into your piece price.

Product design may be more technical, but process design and operations management is  another beast. The logistics of solving problems is typically your biggest hurdle, as opposed to technical challenges. As my old engineering manager used to say, “It’s not easy to fix a plane when it’s in the air.” 

Automotive operation patterns are 24 hours: you can’t just stop it at any time, and if you shut the customer down you might as well put the keys to the plant in the mailbox. You’d better be as close to 100% certainty before you take the line down, or you’re going to hear it from the production team. And that’s why the more data you have, and the better tools you have to analyze it, the more informed and impactful your decisions can be.

Manufacturing Data: Better to Have It & Not Need It...

When I first started on assembly lines, we were collecting and aggregating data, but nowhere near the level of today. The pallets were tracked via RFID scanners and we only stored the critical parameters on a computer at the end of the line. Compare that to the modern assembly lines I’ve worked on, where almost every sub-component has its own 2D matrix with birth history, tabulating all the parameters through the operations it’s been processed through. All of which is at least written to an SQL database so part and assembly data can be traced all through production right up to and including the EOL test. Depending on how well that’s been integrated, you’ll have traceability data all the way through, that’s a lot of data.

If I’m making 2,000 assemblies per shift, and I’m collecting data from gauging stations, press force graphs, leak test data, screwgun data, etc. obviously I’ll be generating a lot of data points, and, in light of diminishing IT costs, I might as well collect and store it.

In principle, it’s great to have all this information, but the problem in practice is that a lot of manufacturers aren’t doing much with it. Basically, they’re using it to solve problems in the same way they would have in the past—having a database just means you don’t have to go measure the part (but you probably will anyways). 

And if you want to look for trends, you can go back and see, for example, that you’re having a flare up today on a clearance check station, and you can evaluate that trend in comparison to others, but you’re still taking the same reactive approach to solving the same problems.

Enter Machine Learning

The fact is, the best way to leverage the huge volumes of manufacturing data isn’t the traditional way. We can’t just look at historically or reactively anymore; we need to be proactive and use that data to build predictive models. Once you start aggregating all your production and testing data, you can start applying data science and machine learning. And when I say all of it, I truly mean all of it. There are insights to be gained not just from average depth measurements but from collecting each individual LVDT measurement too. 

Measuring torque or pressure response after solenoid actuation? Record the time-series traces, not just the summary stats.

The human mind has always been very good at making qualitative assessment—it’s second nature to look at two plots and comment on why they’re different—but we have never been good at consuming large amounts of data. Contrast that to computers that have largely been unable to make qualitative assessment but are great at consuming large amounts of data. 

All that changes with machine learning and AI: computers now have models that can make qualitative assessments by consuming large amounts of data. That could fundamentally change how we build quality and product control into new assembly lines, but first we need to get past the skepticism that data science and machine learning tend to evoke in manufacturers.

I’ve seen two kinds of skepticism in automotive when it comes to data science and machine learning. In the first, you have automakers who see ‘data science’ as the latest in a long line of buzzwords, because they don’t really understand how it’s supposed to work. I recently talked to one engineer who’s been in the industry for 25 years, and even though he was curious, he was still skeptical because it’s all new to him. The other kind of skepticism comes from manufacturers who have embraced data science but ask, “What can Acerta data scientists do with my data that I can’t?”

I think in both cases, that skepticism comes from experience. We’ve all had problems that lead to a high-priced consultant being brought in after the 2nd week of daily 4pm conference calls to “help us through our problems” (which typically means complicated and out of context studies). We’ve all gone down the DFSS study rabbit hole.

I don’t know about you, but I’ve never come out of a meeting with one of those consultants thinking “Problem solved, let’s pack it up!” And the trouble is, a lot of people in automotive paint data scientists with the same brush. 

Quite frankly, I did too, at first.

New Tools in Automotive Engineering

Data science and machine learning have been described as new tools in the engineer’s toolkit. That’s partly because of how novel they are. Some of the most knowledgeable engineers in automotive have been making their products for 20 or 30 years. In that time, data science emerged, developed in academia, and started finding real-life applications. It may be relatively new, but it shouldn’t be underestimated.

More importantly, like any tool, machine learning is ideal for some applications and impractical in others. Here’s an example: I was having difficulty with the launch of a new line for an electric pump assembly because we were getting some low pressure failures off the end of line test. An A-B-A study identified a problem with one of the wire connections—the brazing from the shunt wire to the choke. This problem doesn’t need machine learning.

In contrast, issues like NVH or EOL performance trace problems that are likely a function of several components and multiple features upstream cannot easily be discovered through an A-B-A study or DOE. These are prime candidates for machine learning techniques. 

Traditional root causing methods start having problems when it’s difficult for a human being to take all the relevant factors into account simultaneously; problems where there are many degrees of freedom and a daunting amount of feature interaction. That’s where machine learning can really shine. With this capability in mind, automakers could set up assembly lines knowing what data to collect and even integrate machine learning in the development phase.

But I think that’s just the beginning.

Way back before my time, automakers assembled components using a selective fit methodology: workers would measure the cylinder bores on an engine block, for example, and then select the piston head that gave them the right running fit. That approach has fallen by the wayside for most applications because you can’t take that much time in mass production.

However, with all the data that’s being collected, we could use machine learning to efficiently bring that approach back to the high-volume manufacturing environment. An axle assembly with certain carrier and cover dimensions could be flagged to avoid matching it with an incoming pinion and ring gear set with a particular transmission error.

You might be making parts at 95%, but if you want to get that last five percent, you’re going to need something you don’t already have, and that’s machine learning. Obviously, we’re talking about an ambitious target, but I’ve never had a General Manager tell me that we’ll get to relax if we get to 85% OEE or 95% FTT.

From a broader perspective, as the industry shifts to a Mobility as a Service model, OEMs will need to be more conscious of vehicle reliability than ever. No one would sign up for a mobility service with a streak of issues, and if we’re getting into autonomous vehicles, there’s no driver to diagnose them. Skepticism notwithstanding, automotive needs data science and machine learning at every stage in the product life cycle.

That’s what brought me to Acerta: while others are just starting to dabble in machine learning to see what data science can do for them, we’re helping automakers use those tools in real life production and on-road applications today. Compared to my colleagues, I still can’t say I’m really into cars, but I am into making them, and machine learning is the key to making them better.

For More Information...

Check out other blog posts or read our case studies.

Leave a Comment

Your email address will not be published. Required fields are marked *

How can we Help?

Contact Info

  • 30 Duke St. West, Suite 605, Kitchener, Ontario, Canada, N2H 3W5
  • +1 (519) 341-6080
  • info@acerta.ca
Share on facebook
Share on twitter
Share on linkedin
Share on email
Enjoying the Blog? Sign up for Acerta's Monthly Newsletter!

Welcome Back!

Stay on the cutting edge of automotive Intelligence

Subscribe to Our Monthly Newsletter for Industry Trends, Best Data Practices & Industrial Machine Learning Applications