The fear that “The machines are coming for you!” has been around since before the Luddites, and we can still find it everywhere from blockbuster movies about sentient killer robots to mainstream news reports. A recent The New York Times article asks, “What Exactly Are the Dangers Posed by A.I.?” before warning us that a rapid rise in disinformation is only the start of the problems with artificial intelligence (AI). The final fear, we’re warned: loss of control and the destruction of humanity.
But for facility and maintenance professionals, it’s always made more sense to say, “The machines are coming to you, not for you.” Modern maintenance management is all about working with large data sets; often to track key performance indicators, but increasingly to use predictive maintenance (PdM), where, for example, you feed an asset’s digital twin a steady stream of real-time data so it can warn you about future failures long before they happen.
And that’s just the start of the possible applications for AI in maintenance.
What is the future of AI in maintenance?
The future applications of AI in maintenance likely lay in the technology’s two main abilities: distilling large data sets into discrete actionable insights and creating natural-sounding language from simple prompts.
Basically, maintenance teams are going to use AI to make big numbers smaller, more digestible and short phrases into longer, easier-to-understand instructions and follow-ups.
What is artificial intelligence (AI)?
Before looking at its future in maintenance, it’s important to have a solid sense of what AI is and where it’s heading.
A basic definition of AI is that it’s an intelligent entity created by humans that can perform tasks without explicit instructions. Also, it can think and act rationally and humanly. Because of how often humans are irrational and inhuman, it’s possible to define AI as the best possible version of ourselves. But that’s a bit philosophical.
Back down at a more practical level, a good way to understand AI is to look at some of the common ways researchers divide it into different types.
Strong and weak AI
So, you can classify AI as strong or weak. Strong AI has a wider scope, human-level intelligence, and can cluster and associate data. Weak AI has a relatively narrower scope, is only good at specific tasks, and can use a combination of supervised and unsupervised learning to process data. A good example of weak AI is a simple computer chess game.
Artificial narrow intelligence (ANI), artificial general intelligence (AGI), and artificial super intelligence (ASI)
ANI is the only type of AI we already have. It’s able to do one task well, but only that one task. An example could be an AI that can guess what products you might also be interested in by looking at past purchases. Another would be an AI that can predict the weather.
For facility and maintenance managers, an ANI would be able to predict future occupancy rates by looking at historical data or tell you the best time to change out the air filers on a piece of equipment based on sensor data fed into a digital twin.
AGI is currently only theoretical. Instead of being good at one task, an AGI would excel at a range of competencies, including:
- Language processing
- Image processing
- Computational functioning
- Reasoning
Basically, it would be like a regular person, able to pay attention to a conversation while driving a car and trying, in the back of their head, to estimate how much longer the ride is going to take based on the amount of traffic.
ASI is the system in those sci-fi movies that can beat humans at all levels of cognition and production. Not only can it outwrite Shakespeare, but it can also emotionally connect to literature better than we can.
And that’s just the beginning, because one of the many things it will be better at than us is developing new AIs. If we can create something that’s twice as smart as we are, what can that AI go on to create by itself?
In classic science fiction and horror, we create some version of Frankenstein’s monster, which we quickly lose control of, unleashing chaos. But imagine those monsters heading back into the lab to work on their own experiments, their own attempts to create life.
What is predictive maintenance and what role does AI play?
Like with all trendy terms, the definition of PdM depends on who you ask. It also depends on when you ask them, because the technology is quickly evolving. But a broad definition would include the fact that PdM leverages both historical and real-time data from assets and equipment to warn you of future failures so you can find and fix small issues before they have a chance to create costly downtime.
PdM involves a lot of data from many different sources. According to a report on using machine learning in manufacturing, “ML algorithms are fed OT data (from the production floor: sensors, PLCs, historians, SCADA), IT data (contextual data: ERP, quality, MES, etc.), and manufacturing process information describing the synchronicity between the machines, and the rate of production flow.”
But what’s the role of AI?
PdM is like a carnival fortune teller for your assets, where the crystal ball is all the data, and the AI is the person peering into it. Once you train the AI on what the data from the assets should look like, the software can find trends and then anomalies.
If you removed the AI from the process, you’d have condition-based maintenance, where the software warns you as soon as the asset moves outside of what you told it are the acceptable operational ranges. You don’t even need fancy software for that, though. A simple physical thermostat with two metals bent into one coil can tell you when things are getting too hot inside an engine, for example.
What is a chatbot and what is ChatGPT?
A chatbot is a piece of software that simulates conversation with a user. In the past, they were nothing more than basic, interactive FAQs, hard coded with a select set of responses matched to a limited number of questions.
Modern chatbots now leverage a type of AI called natural language understanding to figure out a user’s intentions and goals, which helps the AI provide the best possible answers.
Over time, the chatbot can also employ machine learning and deep learning to expand and refine its base of questions and answers, improving its ability to predict how best to help future users.
ChatGPT is a popular chatbot from OpenAI. The GPT in the name stands for generative pre-trained transformer, which is a form of large language model. For this type of AI, the language model holds a neural network with billions of parameters. The software trains itself using self-supervised learning or semi-supervised training.
For the average facility or maintenance manager, the important parts here are that ChatGPT has so far proven so successful at such a wide range of language tasks that research in this area of AI has shifted away from older paradigms. And ChatGPT’s impact is set to keep spreading. In fact, since January of 2023, it’s been the fastest growing consumer application in history.
What are some current and future use cases for language-based AI in maintenance?
Outside of predictive maintenance, where AI works with large data sets from sensors and digital twins to tell you the best times and tasks for maintenance, the future of AI in maintenance will likely leverage the new technology’s capacity to generate natural-sounding language.
Sorting maintenance requests with a simple chatbot
Traditionally, maintenance managers have struggled to get accurate information on needed repairs from outside the department. Your average maintenance technician might be able to write up a good work order simply because they understand what to include, but what about the average office manager or production manager?
In the future, by upgrading a static request portal to an interactive chatbot, facility and maintenance managers are much more likely to get the right data from the right people.
Right from the start, the chatbot will know what questions to ask, following up when it needs more information. It will also act as an active gatekeeper, making sure real emergencies go directly to the team the same way current customer service chatbots work on, for example, a bank’s website. Using plain language, the user explains the sort of help they need.
If the AI knows the information already exists on the site, it provides it directly or provides directions to it. But as soon as a user mentions a stolen or lost credit card, it connects them directly with a real person.
Improving communication with AI writing assistants
Most maintenance managers are looking for technicians who can read instructions, follow diagrams, and work their way through checklists. But those skills don’t directly translate into being able to write instructions, sketch out diagrams, and create checklists. They’re separate skill sets.
The result is that it’s generally easy to get information into your techs’ heads but more difficult to get it out. Work orders go out with clear instructions and neat lists of associated parts and materials, but the follow-up messages and close-out notes might come back riddled with typos, spelling mistakes, and unclear sentences.
Even if a technician were a latter-day Bill Shakespeare when sitting comfortably at a desk in an air-conditioned office, being clear and concise is more challenging out in the field trying to type out messages on a mobile device.
With what’s being marketed as “generative AI,” everyone in the maintenance department gets a real-time communication assistant that helps them with everything from spelling and clarity to tone and style. Your tech can use prompts to explain to the AI the message they need to send. Because the AI already knows the department’s preferred tone and style, it can make a good first draft, which the tech can then fine-tune.
Or, when the tech writes the first draft themselves, the AI can check it for them, offering corrections for problems with spelling and grammar as well as suggestions on tone and style, including word choice and sentence structure.
What are the current limitations for AI in maintenance?
For AI and predictive maintenance, the limitations are related to cost. According to a recent McKinsey report on why implementing predictive maintenance can be expensive, organizations looking to scale their systems “require substantial investment in [research and development] along with deep industry knowledge, access to relevant data, and practical operational experience.”
You need to invest in the sensors that go on the assets to collect real-time streams of data. The digital models you feed that data also take time and money to build and maintain. Most technicians won’t have experience working with the new tech, so you have to set up additional training.
The solution is the same as for any other project involving a lot of money and moving parts. Organizations need to go into predictive maintenance with patience and persistence, a clear set of goals and a strong commitment to change management.
For language-based AI, the limitations are more related to accuracy than cost. When you ask an AI to write something, you can’t assume it’s going to limit itself to facts. Writing for FastCompany, Harry McCracken, a real person, explains the problems AIs have with the truth. “If a rogue software engineer set out to poison our shared corpus of knowledge by generating convincing-sounding misinformation in bulk, the end result might look something like this.
It’s prone to botching the chronological order of events, conflating multiple people with similar backgrounds, and—like an unprepared student—lobbing vague pronouncements that don’t require it to know anything about the topic at hand.”
An example of AI inaccuracy is the lawyer who accidentally cited cases that never existed. He used an AI to help him generate legal briefs, which he then submitted to the court, but the paperwork was full of what developers call “AI hallucinations,” where the software appears to be guessing, making up answers that are not true and do not match the data it’s been given.
Even though it’s a new problem, the solution is the same as when implementing any technology into existing workflows. Everyone needs to be aware of the inherent limitations of the tools. If someone on the team uses AI to help them write or check anything, they need to go back and double-check the AI’s output for accuracy.
Summary
Although there is a long, colorful history both in popular culture and mainstream news of exploring the threats of artificial intelligence, for facility and maintenance managers, the technology has always promised positives, including the ability to predict and prevent failures long before they happen. But that’s only the start of the possibilities of AI in maintenance. “Artificial intelligence” covers a wide variety of software that excels at what are traditionally human capabilities, including language processing and reasoning.
In maintenance, we can leverage these for predictive maintenance, where large amounts of data are fed into an AI so it can look for anomalies that suggest future problems. The main issue, though, is the technology tends to be expensive to set up and then run. Organizations need to start with a clear set of goals and a commitment to change management.
Maintenance managers can also use generative AI. Here, chatbots can act as work order request portal gatekeepers, ensuring the maintenance department gets the information it needs from outside sources. AI can also help technicians craft clear comments and emails, even when out in the field working under tough conditions.
