Good luck with that algorithm – Machine diagnostics

It is important that certain diagnostic responsibilities are not just assigned to AI tools without considering the need for human intervention and experience, warns Annemie Willer, manager of WearCheck’s ARC (asset reliability care) division.

“We keep hearing worrying claims from industry stakeholders and customers’, says Annemarie, ‘that if you throw enough data from vibration, oil, thermography, process sensors, ultrasound, and AE (acoustic emission) into an AI system, it’ll somehow converge into a perfect picture of machine health, complete with the exact corrective action to take.

I don’t buy it

“It’s a nice idea. In fact, it sounds like the future. But I don’t buy it,’’ she says.
“Importantly, this is not because I’m anti-technology, quite the opposite, in fact. I’ve worked in diagnostics long enough to see the value of every tool we have. But I’ve also been around long enough to know this: machines don’t behave according to theory. And AI doesn’t understand that.

‘Convergence’

“For example, I keep encountering the myth of “convergence”: the idea that all condition monitoring technologies can fuse into one holistic truth, which assumes that machines behave in predictable, repeatable ways. Only problem? They don’t.

“You can install ten pumps from the same OEM, running under the same process conditions, in the same plant, with the same lube, and still, they won’t age the same. One might run clean for six years. Another might seize up in eight months. And no amount of sensor data is going to tell you why – not reliably.

Machines are not cloned

“Why? Because machines are not clones. They’re flawed. They are manufactured to tolerance, not perfection. Machined surfaces differ microscopically, and assembly is never identical. And once you add human hands, production targets, rushed shutdowns, and midnight shift decisions into the mix — good luck feeding that into an algorithm!
“It is important to take the real-world situation into account when assessing an asset. AI relies on data, but data only captures what the sensors see, not what the human maintainer did when nobody was looking. It does not record the subtle looseness that a technician ‘felt’ but didn’t log. It does not register the fact that someone topped up the wrong grease, or skipped torque checks, or ran a fan uncoupled for three minutes at startup.

Root cause

“No historian records that. And without that real-world information, AI is flying blind on the stuff that actually causes most failures.
“I believe that every condition monitoring technology has its place, and its limits. For example, vibration monitoring tells us about mechanical behaviour; oil analysis identifies lubricant condition and contamination; thermography picks up heat and load imbalance, AE and ultrasound testing give early warnings of friction, turbulence, or sparking; and process data provides the operating context but not the root cause of failure.

Team effort, not a solo act

“These monitoring techniques and their test results don’t converge neatly. They weren’t designed to. One doesn’t combine them to get a better truth, rather, they should be compared to demonstrate different perspectives. That’s what makes condition monitoring powerful: it’s a team effort, not a solo act.
“AI is useful” insists Annemarie, “just not the way that the vendors keep claiming. It can spot changes over time. It can rank the risks, it can filter out noise and highlight anomalies, all of this is valuable.

And when you get there?

“Importantly, however, AI cannot know the history of every shaft and housing. It cannot understand why a lube change worked for one gearbox and not the next.
It cannot interpret subtle mechanical behaviour that only a human technician would notice, and it cannot predict how different people on different shifts handle the same piece of equipment.

“In other words, AI can help one find where to look, but not what to do when you get there.

Machines have personalities

“I have always told our customers that machines are messy, and that this is not a problem, it is merely the reality. Here’s the truth: machines have personalities. Not literally, of course, but in how they wear, respond, and behave under pressure. And a lot of that has nothing to do with engineering design or process control. It has to do with maintenance history, human touch, and physical realities that no AI-powered model, however sophisticated, can learn.

“The idea that AI will converge all technologies into one correct decision ignores this complexity. It reduces the craft of diagnostics to a logic problem, when in truth, it’s part science, part art, and always tied to context.”

Annemarie concludes, “Let AI support us. Let it help us scale, see patterns, and work smarter. But let’s stop pretending it can replace understanding or diagnose machines like a seasoned engineer can. Because machines don’t live in the cloud. They live in the real world. And in the real world, convergence isn’t the goal. Clarity is.”