Abstract
Predictive maintenance produces probabilities, anomalies and remaining-life estimates. Maintenance decisions require something else: the mechanism, the uncertainty around that mechanism, and the verification that discriminates between competing explanations. This paper argues that prediction should be treated as evidence entering a diagnostic process rather than as a diagnosis in itself, and that some of the most important diagnostic information is not present in the steady-state signal at all.
The pump that ran backwards for six years
We got a call-out to an oil site. Their pump was not producing enough pressure.
We drove out and looked everything over, and the site was right. It was not putting out the pressure you would expect from a pump that size. That much was correct, measurable, and completely useless on its own.
I asked them to shut it down. As it slowed, I watched the shaft, and I could see it turning the wrong direction.
I told the operator his pump was going in reverse. He said no it isn't. I said yeah, it is, I'm telling you. There's the arrow on the casing. Give it a bump and you'll see. They bumped it. Wrong rotation, confirmed against the arrow.
The operator said nobody had worked on that pump. He went away, checked, came back, and said nobody had worked on that pump in six years.
Well. Then it has been going backwards for six years.
He asked how that is even possible. It is a centrifugal pump. It still pumps in reverse, because the impeller slings fluid outward regardless of which way it turns. The volute and the vanes are optimized for one direction, so reverse just gives you degraded performance. My read on the six years is that the impeller clearance was tight, either set that way or by luck, and the tight clearance compensated. It gave them the pressure they wanted despite the rotation being wrong. Over six years the clearance wore, and the compensation wore with it, until the output finally dropped below what they needed.
We set the clearance back and rewired it to turn the right way. As far as I know it ran fine for at least three years after.
What that story does to predictive maintenance
Sit with the timeline, because there are three separate lessons in it and each one lands somewhere different.
The complaint was not the fault appearing. It was the compensation running out. The fault was present and constant from day one. What changed over six years was a second, unrelated parameter that had been quietly hiding it. Any system watching for a change would have flagged the wrong event, at the wrong time, six years late, and would have been looking at wear when the answer was wiring.
Six years of wrong was the baseline. Anything trained on that asset's history learned reverse rotation as normal operation. A model asked to detect deviation from normal would have detected the deviation from a normal that was itself the defect. The plant had a complete, accurate, well-behaved dataset describing a pump running backwards, and nothing in it was labelled wrong.
The diagnostic information was not in the steady-state signal. At full speed a shaft is a blur and rotation direction is invisible. It shows up during the coast-down, for a few seconds, to somebody who happens to be standing there looking at it.
That last one is the part I would put in front of anybody selling continuous monitoring. Transients are diagnostic windows. Startup, shutdown, coast-down and changeover are where a machine tells you things it will not tell you while it is running, and they are exactly the periods most monitoring systems discard as non-representative.
The seductive number
A system reports an 82 percent probability of failure within fourteen days. The number looks precise, urgent and actionable.
The planner still has to know what is failing, what physical mechanism is suspected, whether the signal comes from operating conditions or instrumentation or actual deterioration, what inspection would discriminate between those, and what job should be prepared.
Prediction answers what may happen. Diagnosis asks why. Planning asks what to do. Three different questions, and a probability answers only the first one.
Nothing in an 82 percent would have told anyone to watch the shaft on the way down.
Fourteen days is not an invented horizon. In September 2026 TDK SensEI took its edgeRX platform to IMTS with the line that it can "predict maintenance needs up to 14 days in advance." No asset class, no failure mode, no false-positive rate, and nothing about what anybody is supposed to do on day one of the fourteen. That is a vendor claim and it cannot be checked as stated. It is also a clean example of the pattern: a horizon offered where the mechanism should be.
From isolated signals to evidence-grounded reasoning
IBM's Condition Insight Agent matters here because it combines work-order language, operational behaviour and engineering failure semantics rather than treating a single prediction as the final answer, and it applies a verification loop to suppress unsupported conclusions. Multimodal maintenance research points the same way, combining sensor data, images and text to produce grounded guidance.
Vendors are starting to build it. At IMTS 2026, deviceWISE demonstrated a system on live robotic cells that pairs what cameras see with live machine data and process state, runs root-cause diagnostics at the edge, and generates a recovery procedure for the specific fault. It is a vendor-controlled demonstration with no published performance figures, so it proves nothing about field results. The architecture is the point. The anomaly is only the entry, and the diagnosis comes from a second, independent line of evidence laid against it.
The useful model is that a predictive output becomes one item in an evidence package. The system then compares competing mechanisms and recommends the next verification step, instead of jumping from anomaly to repair.
For the reverse pump, an evidence-grounded system would not have needed to know about rotation. It would only have needed to say: output is below curve, the mechanisms consistent with that are wear, clearance, cavitation, obstruction, speed and rotation, and here is the cheapest observation that separates them. Watching the coast-down is free.
The cost of confusing the two
When prediction is treated as diagnosis, plants replace good components, create unnecessary downtime, and write incorrect failure histories.
That last one compounds, and it is the part worth being frightened of. A false repair is recorded as a repair. Later it reads as confirmation that the original diagnosis was right, and bad inference becomes institutional memory.
The mechanism does not even need AI to get started. We had a vibrating conveyor cracking itself apart, stress fractures everywhere, replacement weeks out. I got my welders in to weld the hell out of whatever they could, as an explicit stopgap to limp it to the next down day. Next down day I scheduled the welders again, and the plant manager was angry, because I had told him it was fixed.
I had not. I told him it was OK to run until the next down day. It was nowhere near fixed and it was not going to be, it needed replacing, and we were chasing our own tails hoping it did not let go before the new one arrived. We had the same argument the down day after that.
He was not being stupid. He defaulted to "running equals done" because the system let him. The distinction between a holding action and a repair lived in a verbal aside instead of in a recorded completion state. One line on the work order, TEMPORARY, recurring until replacement, would have ended the argument before it started.
Now put a language model on top of that record. It reads a closed work order, a completed weld, and a machine that ran. It has no way to know the difference between a repair and a holding action, because nobody wrote one down. It will summarize that asset's history fluently and it will be wrong, and its wrongness will be indistinguishable from its correctness in tone.
A decision-oriented output
A useful maintenance AI output should separate observed condition, predicted risk, suspected mechanisms, supporting evidence, contradicting evidence, alternative explanations, verification required, and work readiness.
The system should be entirely comfortable saying that an asset is at elevated risk while the repair scope is not ready. Those are different findings and collapsing them is how a plant ends up sending a crew to do the wrong job urgently.
Delay-aware maintenance research adds a point that the reverse pump makes concrete. When evidence arrives asynchronously, the system has to distinguish the current estimated condition from the age of the information used to estimate it. A reading from a sensor installed six years ago, describing a machine nobody has opened since, is not current information. It is a very old claim being repeated confidently.
Conclusion
Predictive maintenance is most valuable when it creates time to reason. The point of an early signal is not a more impressive alarm. It is an earlier and better decision.
That requires treating prediction as the beginning of a diagnostic process rather than the end of one, keeping the mechanism question open until an observation closes it, and understanding that the observation which closes it may be free, may take four seconds, and may only be available while the machine is slowing down.
Six years is a long time for a pump to run backwards. What made it possible was not a lack of data. There was plenty of data, all of it accurate, describing a machine that was wrong in a way nothing was watching for.
References
- IBM Research (O’Donncha et al.), “Evidence-Driven Reasoning for Industrial Maintenance Using Heterogeneous Data,” ACL 2026. https://research.ibm.com/publications/evidencedriven-reasoning-for-industrial-maintenance-using-heterogeneous-data
- Hsu et al., “Cache-augmented multimodal generative AI for energy-aware predictive maintenance,” Computers & Industrial Engineering, 2026. https://www.sciencedirect.com/science/article/pii/S0360835226003517
- Fang & Wang, “Delay-aware predictive maintenance control for multi-equipment systems using approximate Bayesian neural reliability networks,” Electronic Research Archive, 2026. https://www.aimspress.com/article/doi/10.3934/era.2026294
- Li, Feng, Chen, Tsai, Sun, Das, El Maghraoui, Lin & Patel, “PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools,” 2026. https://arxiv.org/abs/2604.01532
- deviceWISE (Telit Cinterion), “deviceWISE to Demonstrate Agentic AI and Automated Fault Detection & Recovery on Live Robotic Lines at IMTS 2026,” PR Newswire, September 2026. https://www.prnewswire.com/news-releases/devicewise-a-telit-cinterion-company-to-demonstrate-agentic-ai-and-automated-fault-detection--recovery-on-live-robotic-lines-at-imts-2026-302875313.html
- TDK SensEI, “IMTS 2026.” https://sensei.tdk.com/event/imts-2026/