Oumayma Jerbi | BETAVIB

The Layer That Does Work

None of this means artificial intelligence has no place in condition monitoring. It means the technology has been pointed at the wrong layer.

Consider how a competent analyst actually works. They do not stare at raw amplitude values and intuit a conclusion. They calculate. Bearing fault frequencies come from bearing geometry and shaft speed. Gear mesh frequency comes from tooth count. The analyst matches measured peaks against those calculated targets, weighs the pattern of evidence, rules out the alternatives and writes it up.

Every one of those steps except the last is deterministic physics, and deterministic physics is exactly what software has always been good at. A well-built diagnostic engine therefore has three distinct layers, and keeping them separate is the entire design.

1. The physics layer resolves fault frequencies from machine kinematics and measured running speed, then matches spectral peaks against them. Working in orders—multiples of shaft speed—rather than in hertz is what makes this portable. A rule written for a bearing outer race defect applies to every machine in the plant, because it references the fault frequency symbolically rather than numerically.
2. The statistics layer handles what changes over time. Robust methods matter here more than sophisticated ones. Vibration trends contain transients from process changes, poor sensor mounting and simple operator error; a conventional least-squares trend line is dragged badly off course by a single bad reading, while a median-based slope estimate ignores it entirely. This layer requires no labeled faults at all. It learns only what normal looks like for that specific asset, from that asset’s own history.
3. The language layer is where a language model finally earns its place—and it earns it well. Given a structured set of findings from the layers below, a model can produce a clear, readable report in the plant’s working language, at the technical register the audience needs. It can answer plain-language questions across a fleet. It can help an analyst turn a hard-won diagnostic heuristic into a formal rule.
4. The critical constraint is that the model never sees the spectrum. It receives only the conclusions and the evidence supporting them. It cannot invent a frequency it was never given.

Evidence, Not Scores

There is a second reason to prefer this structure, and in practice, it matters more than accuracy.

A diagnosis must survive contact with a maintenance planner. “Anomaly score 0.87” does not survive that meeting. What survives is “outer race fault frequency present at 109 hertz with six harmonics, no shaft-speed sidebands, envelope band energy up five-fold in 90 days, all four motor points stable.”

That second statement does more than justify itself. It tells the planner what part to order. The absence of sidebands is not a detail; a defect on a stationary outer race sits permanently in the load zone and produces harmonics without modulation, while a defect on the rotating inner race passes through the load zone once per revolution and generates shaft-speed sidebands. That single distinction separates two different repairs with different urgencies, and it is only available to a system that reasons about physics rather than pattern-matching on amplitude.

Explainability in this domain is not a regulatory nicety. It is the mechanism by which a diagnosis becomes an action.