As skilled tradespeople retire, manufacturers are feeling the squeeze between their customers’ reshoring initiatives and the challenge of finding skilled labor to fulfill projects.
But a new wave of generative AI that has made headlines across the world could bring relief.
Managing skills has always been a challenge for technical manufacturers. When new employees join, it can take years for them to learn not only how to operate machines, but how to deal with the exceptions and edge cases that inevitably come up with complex modern equipment.
With average tenures declining, however, employees often never become an expert before they move on. According to the BLS, average manufacturing tenure has declined 14% in 10 years. And while 55-64 year-olds have an average tenure of 10 years, those in the 25-34 bracket stay for less than three years.
“There’s a difference in generations,” said Derek Moeller, founder of CognitionWorks, a company that develops AI technical assistants for manufacturers’ employees.
“When someone spends 20 years at one place, they learn the machines. If they’re only there for a year – it’s really hard to become productive.”
Manufacturers respond by increasing their efforts to document processes and procedures.
Process engineers create PowerPoints and process documents to standardize how things should be done. That helps. But as any manufacturer knows, there can be a big gap between writing the document and having it be used on the floor.
“We found that team members on the floor all know how to use phones, but they’re not always that conversant with traditional computers, which is where the information is,” said Stina McConkey, owner of McConkey Plastics near Tacoma, WA. “So there’s this gap between the process engineer and the technician.”
“We tried to fix it by printing everything out and putting it in binders. But there’s no way to search it, and it’s tough to find nuggets of relevant information in a stack of paper, especially when there’s other things happening on the floor.”
McConkey partnered with CognitionWorks to develop a solution: an AI ‘brain’ that employees could talk to, and which had every scrap of process knowledge a manufacturer has. The system uses a ‘large language model’ on the back-end, which is a system with over 100 billion neurons.
“We just threw everything at it,” said McConkey. “Word documents, maintenance logs, databases, PowerPoints, we even scanned our old manuals. The system broke everything down and absorbed it. It learned it all.”
CognitionWorks developed a texting app for phones and tablets, and integrated it with McConkey’s Slack messaging system. “A shift lead just texts the AI assistant. If a machine has a problem, they ask it a question, like they’re talking to a friend. It’s all natural language. ‘Hey, I need help on this injection molding machine – it’s suddenly not building pressure, what do I do?’” said Moeller.
The manufacturer’s ‘AI Brain’ then gives them a list of helpful tips, step-by-step instructions and sources, synthesized from many different documents. This is made possible by the advent of Generative Pre-trained Transformers, popularly known from the ChatGPT app from OpenAI that now has over 100 million users, which is uniquely able to analyze large quantities of documents and summarize them for users.
The AI can search through source documents from many different sources and types, and then compile a straightforward list of things for the tech to try.
Managers like it too. Connecting to a production record database allows a manager to simply ask, “how did production go last night?” and the model will provide a concise summary of the major issues that came up, much like tools such as ChatGPT can summarize an essay. Using database connectors, the AI assistant can also perform more complex queries, such as “how many work orders on Machine 12 are still open?” or “what was the average uptime on cell 3 over the last 3 days?” all in natural language.
Avoiding Language AI Pitfalls
The problem with existing large language models available on the Internet, such as OpenAI’s ChatGPT or Google’s newly released Bard, is that they’re only trained on what’s publicly available. That’s a lot: billions of documents available on the web. But there’s also a lot that isn’t there.
For most manufacturers, that means they’re of limited use in their public release versions. Partly that’s because manufacturers have their own proprietary information. But even when they have standard equipment they’re using, such as an injection molding machine, manuals for those machines are often not publicly available – and therefore, not part of the training set for these language models.
That means ChatGPT works remarkably well for anything you might find on Wikipedia. But ask the ChatGPT website a detailed question about how to operate a specific injection molding machine or CNC cell, and you’ll pull blanks – or, even worse, it might make things up. There’s no way for it to know about content that wasn’t in its training set.
Language models will only be useful to companies that have already worked to document their processes.
“If a company doesn’t have basic documentation and procedures down, they’re going to be disappointed by what AI language models can do,” said Moeller. “But these language models are going to have a big impact for companies that put the effort into documentation, and just need a better way for their employees to access it conversationally. It’s going to create an even bigger performance gap between operations that document process and those that don’t.”
That’s why implementation of natural language AI takes careful consideration of what data is available, and then developing systems to access that data. The interfaces become tools that an AI can choose to use.
“It’s kind of wild to watch,” said Moeller. When we set these systems up, you see the AI actually talk to itself, like it’s thinking out loud: ‘the user asked about this machine procedure, what tool should I use?’ and then it looks at the tools you’ve given it, like machine manuals, production databases, maintenance logs and so on, and it chooses the one it wants. Then it uses the tool to get the information.You see it look at the information returned, and it doesn’t just regurgitate it to the user. It reads it, and says, ‘okay, I got a few different documents from different tools. I should synthesize and summarize these for the user,’ and then it writes it up in a neatly numbered series of steps.”
Bridging Language Barriers
Large language models have some unexpected benefits. One is in the name: language. During training, large language models have absorbed billions of documents in every major language on the planet. That means that these neural networks are excellent at translation, because they can understand the semantic meaning of concepts across languages.
For manufacturers who face language barriers on their shop floor, or international operations that operate with many different languages, this can deliver surprising benefits.
“You have a Spanish-speaking processing lead on the floor, working on an issue. All your documentation is in English. Even if he could access it all, it wouldn’t be of any value to him. Now, he can text your corporate AI brain, in Spanish, it understands his request, searches your English documentation, synthesizes it into a Spanish summary, and returns it to him in a tidy set of steps to take. When we first tried it, we didn’t even expect it to work – it just did it. The capabilities of these neural networks are sometimes surprising.”
WIthout having to invest hundreds or thousands of hours into translating procedure documents, large language models allow companies to create one set of documentation that can be used across a workforce using a diverse set of languages.
How it works: ChatGPT versus Integrated AI
So far, most peoples’ experience with generative AI is through OpenAI’s ChatGPT service. Using this, you can make requests or paste information to query it. But the area of explosive use for industry will be the use of language models to query and summarize internal data and information. When a user is trying to ‘chat’ with large amounts of data, there has to be an intermediary layer.
There are two practical methods today for a manufacturer to train a large language model AI on your data. In either way, relevant data or documents are first fetched, and then provided to an AI as context to answer your question.
1. Vector Embeddings
The first method uses a technique the industry calls ‘vectorized embeddings.’ This technique (best used on text data, such as procedures or reference manuals) translates all of the text in your knowledge repositories into sets of numerical matrices, which are the ‘vectors’ in the name. Though it looks like nothing but long strings of numbers to us, to a computer it has underlying meaning – as if you could assign a number to the actual meaning behind a word. That allows the computer to then quantitatively compare meanings.
That’s very different from traditional keyword search, which only finds the exact word the user types. In traditional keyword search, looking for the word ‘dog’ won’t find ‘golden retriever,’ but vectorized search will because it knows the underlying meaning of each phrase. That’s a lot more like how a human brain works, and it means much more natural conversation with an AI assistant.
So once your data has been turned into numbers this way, you can ask it all sorts of questions, and by turning your question into a set of numbers, it can find the data that’s the closest in meaning. It can do that over vast sums of information: it’ll find the closest match to your query over 100 million words in fractions of a second. Vectorized embeddings are ideal for unstructured documents in PDFs, PowerPoints, Word documents, Sharepoints or even Teams chats.
They’re perfect for answering questions such as ‘what do I do when the A01 light is flashing red on the Line 3 servo board?’
2. Generated SQL
The second method is for the AI to query your databases. Remember how language models know lots of different languages? It turns out SQL is just another language they know, like French or English. So by providing a little information on your databases, they can turn your natural language query – ‘How many hours of uptime did we get on machine 12 over the last four days?’ – into a SQL database query that fetches your answer from live data and provides it to your chat window. This method is ideal for structured information.
These two basic approaches can be used on different databases or stores of information, but ultimately, they are all accessible by the same conversational chatbot. That means you can text a single interface that has a commanding knowledge of your company’s operations and procedures. They can be turned on or off for different users, as well – so, for example, a technician can have access to the portion of the AI that knows how to search for a repair procedure, but not the method that searches for sales data.
Ultimately, the latest advances in AI dating from just a few months ago will mean a sea-change in how we interact with our company’s data and procedures. In a world with deficits of skilled labor, smart manufacturers are already evaluating how they can use these novel intelligent agents to bring decades of knowledge to junior workers, a critical constraint to growing domestic manufacturing.