Every now and then, usually in conferences or in small talks, I’m approached by manufacturing professionals who seem genuinely interested at what we’re doing at Plataine. After all, IIoT is not just an industry in tremendous growth, it’s also a trending buzzword. But when the word ‘AI’, enters the discussion things become even more interesting. “Oh, we already use AI,” people say.
I’m always happy to indulge in conversations about AI use cases, so I immediately ask if they’ll be ok with telling me a bit about their system, and how exactly it’s being used.
Then, as the discussion continues, it becomes clear that they refer to BI, reports and dashboards and not AI in the deep sense that I refer to.
Don’t confuse data analysis and dashboards with AI: The case of Navigation Apps
Did you ever use WAZE or Google maps, the GPS navigation app on your smartphone? Of course you have.
Imagine that the app would simply collect data, analyze it and create a nice dashboard with hundreds of alerts and marks on the app’s main screen? You will get a screen showing tiny icons for cars, roadblocks, traffic jams and whatnot.
But that’s it. Nothing more:
The above is a map of Paris, now if you need to drive from the Eiffel tower to the CDG airport, can you use the above to calculate an optimal route? While you’re driving? I thought not.
That’s because the route is missing.
WAZE’s and other apps’ value proposition does not revolve around collecting data from the road (as much as it is the app’s main enabler). It’s about solving a problem and guiding users in real-time.
The missing element:
AI is used to make predictions based on present and historical data, and make instant calculations for an optimal route. Presenting data on a nice report or dashboard DOES NOT equal AI.
The AI continuum
“There’s smart and there’s really smart … “
To put it differently, the real power of what we call AI is not found simply in presenting or smartly organizing data, or even in alerts that are merely based on data.
Let’s think of AI as a continuum, built on top of the data surface level. You can use AI to be responsive (let’s call it level 1): for example, as soon as the Waze system knows there’s a police car in your route, it can pop up an alert. Or you can use AI to make smart predictions (level 5) instead of waiting until something happens before you can. For example, if the system knows that in this road, during these hours, and only when it’s not raining, a police car is usually spotted, it alerts users of the possibility without having to wait for some user to mark it on the map first and only then react.
The AI continuum can be compared to an intelligence scale that lets us, humans, move from making conclusions based on data to optimizing our environment
The Key is context
Differences between AI levels are often analyzed through the lens of context, since context is a key concept when talking about AI, in any domain. In fact, it’s impossible to discuss intelligence levels outside of context.
Even a seemingly trivial question like “what should I wear today?” is completely context-dependent, and although it’s answered almost automatically, billions of times a day, all around the world, it calls for a very sophisticated process of contextualizing: What is the weather like today? Where am I going? Is it a work day or a day off? is there a dress code? What do I want other people to think of me? And most importantly: which of these questions are my top priority right now?
Pure data can only take you so far. At some point, you will need to run calculation power and add a sort of algorithmic thinking, considering both your experience (historical data) and variety of parameters (the context). That’s AI.
When you think about it deeply, you see that the answer for such a “simple” question is a lot more than a list of all options currently in my closet – which is more or less the equivalent of data presented on a dashboard.
Obviously, when it comes to manufacturing, the nature of the questions and decisions we’re dealing with is quite different than your everyday life as a consumer. Still, our AI’s competence is measured by its ability to offer optimized solutions while weighing all components of the relevant context. And that’s exactly what systems like Plataine’s IIoT are built to do.
Let’s explore five specific use cases.
1 – Prediction of work order latency
Let’s assume that you have a work order that has to meet its deadline. And you need the system to identify if you are at risk meeting this deadline.
As with most other aspects of AI solutions, we set the stage by observing what (simple) MES or similar systems are capable of. Well, they do possess some capability of assessing work order latency, especially when the context is not complex. Basically, a simple “prediction” of how long it should take to get a production cycle done. In some very simple situations, you don’t need more. The problem is that systems can be very inaccurate or identify the risk too late if they don’t take into account a lot of important information, such as historical data.
Let’s take as an example aerospace parts manufacturing; in many sites, 10-30% of work orders are late on a regular basis, information on bottlenecks at stations, availability of tools, and material shelf life isn’t a part of the calculation, and the condition of stations and machines is rarely acknowledged. Yes, there are missing parts to this context.
In this example, if you use a typical MES system that indicates the order progress and alerts if operations are running late you would not be getting any prediction on latency. When we used the Plataine system, we learned that if we kept our plan as is, we will be at risk due to the following issues:
- A couple of the workstations assigned to this project are almost always significantly late, considering historical data.
- The material we need to use is about to expire and we will, in fact, need to cut it first, thereby changing the order, or else we will need to reproduce the part again.
- A tool we were planning to use is away for service and maintenance.
The system alerted us and also provided us with specific actions to take in order to still meet the deadline: We needed to fortify a couple of workstations, cut a specific material first and assign another appropriate tool (while the one that was originally assigned is in service).
AI systems can dramatically improve the prediction of work order latency, and serve as a crucial enabler in maximizing on-time delivery of products and minimizing the financial damage resulting from inaccurate predictions. Context-dependent recommendations (actionable insights) allow manufacturers to make optimal and proactive decisions based on accurate predictions and prevent issues before they occur.
2 – Alerts on part misplacement’s
One of the most common problems in discrete manufacturing (owning huge production sites) is the issue of misplaced parts and tools.
When a part or a tool is misplaced, it can easily trigger a whirring chain of reactions, causing losses through delays and rework (you need to reproduce the same part and after a while when you do find the original it’s too late and you usually scrap it). Well, multiple systems know how to track parts when needed. You don’t need an AI for that… but AI solutions that know how to identify patterns can also identify a deviation and alert of a misplacement as it occurs, saving us the immense effort of searching for a missing part or tool later on. So should I teach the system the different routes of all my parts, components, kits and tools? Definitely not, for this we have AI and more specifically machine learning! By processing historical data using pattern recognition and other algorithms, the Plataine system “learns” the trail per each type of object and can fire an automatic alert every time it deviates from its normal course. Again, context is key. Moving parts are not simply data to be shown on a dashboard. They’re an intricate language of behaviors that can be understood, analyzed, predicted and optimized.
3 – Quality decisions
You fry an egg. Sometimes the pan is perfect. Sometimes it’s ruined. After a while, you realize it’s the pan that sets the standard. You learn from past mistakes and go on to avoid some pans just to prevent issues.
Root causes of quality issues in manufacturing are oftentimes very difficult to identify. There are so many elements in play on the shop floor, it’s almost impossible to tell them apart and pinpoint the single procedure, part, or condition that stands in the way of a better product. Intensive context analysis using AI-based algorithms, which run on production data and search for any telling patterns, is more productive than any other approach in dealing with such complex problems.
On the whole, finding differences in patterns and being able to compare them statistically is something that AI is very good at. We, as humans, are constantly running such testings, unconsciously and intuitively, but we have clear limits. When the amounts of data units are enormous, and the processes are mechanic (i.e., unrelated to our natural environment), and when there’s little time to reach a decision, our human intuition is not relevant anymore.
Moreover, we usually aren’t talking about some static one-time solution to a single specific problem. We’re dealing with an ongoing quality optimization process that is based on multiple historical data. Being able to monitor quality decisions in real time and under dynamic conditions is a great use case of industrial AI systems.
4 – Identifying, predicting and preventing future bottlenecks
Bottlenecks are like the weather, in that famous quote by Mark Twain: everybody talks about them, but nobody does anything to solve them.
Yep, it seems that bottlenecks are the end result of some convoluted law of physics (probably involving some quantum mechanics, too). They’re hard to predict, and yet they’re always there, somewhere, day in, day out.
However, AI-based systems can recognize their patterns when trained on historical production data. Then, in real time, when a certain pattern associated with a bottleneck is identified, the system can send an alert concerning the anticipated delay, so measures could be taken in order to prevent or prepare for it.
Again, context serves an important role: bottlenecks are essentially a part of the big picture. For instance, what does the usual trend look like? Can these workstations function well under pressure? What is their capacity limit? How many units can we usually cut out of this raw material? What usually happens with demand during this time of year? In what conditions does tool-maintenance cause a bottleneck? When does it go well?
5 – Optimization in cut planning
My last example involves cut planning, a supposedly simple process: you take a piece of raw material (let’s think of a very expensive carbon-fiber), and you cut parts from it. You need it for the production. Initially, it seems like a rather simple geometrical problem. However, it’s one of the more interesting problems to think about, when dealing with optimization.
First, as mentioned, we need to optimize the way we cut the parts, so the smallest possible amount of material is lost. But, we must also consider our work orders – some parts might be more urgent than others, and all could refer to the same potential piece of available material.
The example below shows the traditional way of cutting a specific material into different parts. Each type is cut separately.
here, the same cutting is done based on smart calculations. All parts are mixed and cut from the same sheet, resulting in almost 20% material saved:
But wait, there’s more!
Context is relevant here as well. If we only consider the geometric aspects of the problem, then yes, it’s still not that extremely complicated.
But processes on the shop floor are never ‘just about’ something.
There are other parameters to consider: What’s the material’s expiry date (we should cut X material before Y), what’s the fastest way to the cut parts, how to avoid redundant machine setups, and will it be more efficient to first cut parts from group B and only then from group A (even though geometrically it might make more sense to start with A), simply because we have an available station that is ready to work on these parts now? Which part are needed ahead of others…we must now cut too many parts before it’s relevant.
And so on. So many contextual parameters, so many open optimization questions, all related to what seemed like a relatively simple process.
I hope you can now see the differences between having a data system that includes fancy reports and dashboards, and a system that is based on AI algorithms. AI systems should be smarter than you in a way that allows them to help you navigate through multiple complicated factory floor processes, optimizing your decision making, giving you recommendations and actionable insights in real-time, every time.