A Loose Screw Moves In More Ways Than a Screw Should

Researchers at the Karlsruhe Institute of Technology built a robot that takes broken machines apart. It unscrews, it mills, it manipulates, and the point is salvage: recover the parts that still work before the rest becomes e-waste. They presented it at ICRA this year in Vienna, and IEEE Spectrum wrote it up last week.

The arm is not the interesting part. Robot arms that unscrew things have existed for a long time.

The interesting part is how it figures out what is wrong. And it is described in one throwaway line about finding a loose screw:

the system moves the screw, it will notice that it can move in more ways than a screw should move.

I have been building condition monitoring for two years and I have never had a definition of failure that clean.

What the Robot Actually Does

The procedure has four steps, and the third one is the one worth stealing.

First, it loads a CAD model of the device and forms a mathematical guess about what is probably damaged. Second, it starts taking things apart, nudging each component as it goes. Third, it compares how the part actually moved against how that part is supposed to be able to move. Every component in an assembly has a permitted set of motions, called its degrees of freedom. A hinge pin rotates on one axis. A screw turns and advances along one axis. A press-fit bearing race should not move at all.

Fourth, when the observed motion exceeds the permitted motion, the robot revises its guess about what is broken and changes its plan for the rest of the teardown.

So the fault detector is not a camera looking for cracks and it is not a model trained on pictures of broken gearboxes. It is a comparison between how much freedom a part has and how much freedom it is allowed.

Why This Reorganized My Own Work

Run that definition backward through everything I monitor and it holds.

A bearing is a constraint on where a shaft is allowed to sit. When it wears, the shaft gets radial freedom it is not supposed to have. It wobbles a little, it rubs where it should glide, friction goes up, and the motor pulls more current to turn the same load. The current rise is not the failure. The current rise is the echo of the failure, three weeks later, arriving at my smart plug.

A motor mount is a constraint on the whole assembly's position. When a bolt backs out, the motor gets freedom, the shaft alignment goes off, and again: more current, longer cycles.

A coupling is a constraint on relative rotation between two shafts. It fails by permitting slip.

Every fault my system has ever caught fits the pattern, and I did not know that until a German disassembly robot said it out loud. What I have been selling, described honestly, is this: your machine is a stack of parts held in place by other parts, and every mechanical failure starts as something getting loose. I watch the electrical signature of loose.

That is a better sentence than the one I have been using, which was some version of "we track changes in how hard your equipment is working." True, and completely inert.

The Same Idea Tells Me What I Cannot See

Here is why the definition earns its keep, and it is not the sales line. It is the filter.

If a fault is a constraint that stopped constraining, then the question I should ask about any piece of equipment a customer points at is: does this thing fail by getting loose, and does getting loose change something electrical I can already measure?

Sometimes the answer is a clean yes. Anything that spins, mostly. Pumps, fans, compressors, blower motors, circulators. There is a rotating mass held in place by bearings, the bearings are the constraint, and the motor's own current draw reports on how hard it is fighting. That failure mode is loud in a signal I can get for fifteen dollars.

Sometimes the answer is no, and I need to say so:

The bottom half of that table is work I should decline, and I have not always been crisp about it. There is a real temptation, when someone is ready to pay you, to say the system will probably pick something up. It might. But "probably picks something up" is how monitoring gets a reputation for being decoration.

So the intake question is now one line: what part of this machine is supposed to be held still, and what happens to you when it comes loose? If they can answer it, I can help and I will quote it. If the thing they are actually afraid of is fouling or leaking or drift, I will tell them current monitoring is the wrong instrument and point them somewhere else. A fast honest no is worth more to both of us than a slow maybe.

The Second Thing the Robot Does, Which Is Better Than What I Do

Now the part that stung.

The KIT system starts with a guess from the CAD model. It does not trust it. It takes a physical action, partly to make progress and partly to gather evidence. It looks at what happened. It revises. Then it acts again.

It never commits on the strength of one observation, because one observation is not enough and the system is built as though it knows that.

My alerting, on the other hand, watches a number, and when the number crosses a line, it emails a human. That is a system that commits on one observation, at three in the morning, to somebody's phone.

I have written before that half of vibration alerts are wrong. This is a large part of why. It is not that the threshold is set badly. It is that a threshold crossing is evidence, and I have been treating it as a verdict.

What I Am Changing This Week

Two stages instead of one.

Stage one: the hypothesis. A threshold crossing no longer sends anything. It writes a record that says what it thinks is happening and what that would imply about the near future. If this is a worn bearing, then the next several cycles run long and current stays elevated. That prediction is the whole point; it is what makes the guess falsifiable.

Stage two: the check. The system watches the following cycles against that prediction. If the pattern holds, a human hears about it, and the alert arrives with its evidence attached rather than as a bare number. If the pattern does not hold, the hypothesis closes quietly and nobody's night gets interrupted by a shop vac.

Same sensors. Same thresholds. Same hardware in the basement. The only thing that changes is that the system is allowed to be uncertain for an hour before it bothers anyone.

And I can find out what it is worth before I ship it, which is the good part. I have two years of history from my own pump. I can run the two-stage logic backward over that record and count exactly how many of the alerts I actually sent would have died at stage one. I do not know that number yet. I will publish it either way, including if it turns out to be small.

The Part I Keep Coming Back To

A research team in Germany built a robot to pull salvageable parts out of dead machines, and in solving that, they had to give the machine a way to tell working from broken using nothing but motion. What they landed on was: compare how much freedom this part has to how much it is allowed.

That is not a robotics insight. It is a maintenance insight that happened to be discovered by roboticists, because they were the ones who needed to write it down precisely enough for a machine to execute.

I do not have a robot arm and I am not going to build one. But I have a definition of failure I did not have three days ago, a filter for which jobs to take, and a reason my alerts have been firing on thin evidence. That is a good week for reading somebody else's conference paper.