AI in Manufacturing: What's to know

AI, Machine Learning, Deep Learning... Do you know how to explain these buzzwords and how to apply them to Industrial use cases?

What is Artificial Intelligence (AI)? 

AI refers to the application of technologies designed to equip machinery and systems with the capability to simulate cognitive functions similar to human intelligence, such as understanding and responding to instructions, analyzing data, offering solutions, and more. 

A very important side note is that the AI system was never programmed to do a specific task, it learns to do it. Let’s take a robot as an example.

A traditional robot, e.g. a welding robot used in the Automotive industry could be programmed to perform a very specific set of movements. Every spot is defined and the robot is moving from one x,y,z coordinate to the next one. A small unforeseen deviation could require the process to stop, typically indicated by a red signal light. An operator needs to step in, rectify whatever went wrong and restart the machine. 

• The big advantage is that we can read (and thus understand) the programming code,

• The big disadvantage is that every change & every situation needs to be programmed.

A robot that is solely using AI to perform the same task would need to learn by itself, it would try to weld thousands, even millions of metal sheets until the weld is exactly where it needs to be (luckily, we can use computer simulations to learn). Compare this to how a baby learns to walk or how it tries to solve its first wooden jigsaw. A baby does not understand spoken instructions, he/she hasn’t downloaded an historical database from previous babies trying to do the same. They try the activity over and over again, until they get feedback from you: “Yes!! Great job!”. This feedback is stored into their memory and they try again, await new feedback and do it once more.

• The big advantage is that (given enough time), the robot could be doing things which are extremely hard to program.

• The (extremely hard to solve) disadvantage is that we cannot read the code, we can’t exactly know what the system is doing. It might be funny when Siri have a completely wrong understanding of your spoken instruction, but it’s less funny if the AI algorithm is mixing chemical components which might blow up.

Obviously, these are two extreme examples. More and more smart robots are a combination of both: eg they apply AI powered Vision systems to identify a small deviation in the metal sheet. In the video below we show robots playing football, without being programmed to do so (the only thing they know is that making a goal is what they are supposed to do)  

AI is not just 1 technology, it’s a Collection

While AI is often perceived as an individual system, it's actually a suite of technologies implemented into a system to enable it to reason, learn, and act to resolve complex tasks. There are multiple technologies available or under development, however, for industrial applications we believe that you should at least know about the following ones: 

• Machine Learning (ML) aims to teach a machine how to perform a specific task and provide accurate results by identifying patterns in an existing (historical) dataset. If the prediction you are trying to engineer is not represented in the available data: stop & go home.
  We’ve found this brilliant summary of Machine Learning in the Book “Designing Machine Learning Systems” by Chip Huyen:  Machine learning is an approach to learn complex patterns from existing data and use these patterns to make predictions on unseen data.

• Deep learning is a more advanced application of Machine Learning and uses neural networks to mimic a human brain.

It’s not black and white, but all other things you hear about are typically applications of those two major technologies. For example:

• Natural language processing (NLP) and Speech Recognition is a field of artificial intelligence that enables computers to understand, interpret, and generate human language by using algorithms and linguistic rules to analyze and process large amounts of textual data. It uses both Machine Learning and Deep Learning. But that is not enough, additional rules about how Human Language works need to be added as well. It is now incredibly popular thanks to OpenAI’s ChatGPT, but already being researched for decades.

• Computer vision typically applies Deep Learning to detect patterns in images.

• Expert systems also apply both Machine Learning and Deep Learning 

And so one…

Visual representation of different technologies: AI as part of a very broad range of mathematical/statistical methods. Machine Learning (ML) as an application of AI. Deep Learning can be seen as a subset of ML.

First things first: fixing the Elephant in the Room 

Let’s start with the elephant in the room: we need data. Let’s circle back to the definition of Machine Learning from the previous paragraph: “Machine learning is an approach to learn complex patterns from existing data and use these patterns to make predictions on unseen data”. 

Unfortunately, the “existing data” statement is a big problem in manufacturing for multiple reasons:

• In many cases the right data is just not there

  • Because we do not measure it (eg: if we want to predict the quality of a product, we need to have reliable data from the past to start our modeling process)

  • Because it’s on paper 

• In other cases, the data we have is useless 

  • Because it gets downsampled (eg: if we want to analyze vibration measurements, we might need high frequency data and not hourly averages)

  • Because it is noisy, the sensor is not calibrated or other quality issues exist

• More often than not the data has not enough variation 

  • Because we all try to run our processes as stable as possible, the resulting data sets often have a very low level of variability. 

• Finally, in almost all cases, the data is hiding in a wide variety of subsystems

  • Making it very difficult to access it using new (cloud native) machine learning platforms 

  • Making it even more difficult to combine several data sources with each other

Good news: fixing “the data problem” is possible. But it’s way more difficult than most people think or most vendors will make you believe. 

Where can AI be applied in Manufacturing? 

Good things always come in 3, so meet the clusters in which AI can be applied given a Manufacturing/Industrial context: 

• Design & Engineer: Design, draw & build new assets and processes (“greenfield”) or update existing ones (“brownfield”).

• Operate & Maintain: Once the asset has been built, we need to operate & maintain it in a reliable, safe and predictable way.

• Plan & Optimize: Optimize the supply chain, decide which raw materials will be blended, reduce energy consumption… just a few examples of what is possible. 

Here are some final examples of applications in the cluster “Operate & Maintain”:

• Predictive Maintenance (already a hot topic for many years, but still a long way to go!)

• Quality Control (estimate product quality using machine learning)

• Operator & Technician Guidance (think expert systems, NLP…)

Collaborative Robots / Autonomous vehicles (combination of all goodies including computer vision)