Pressure fluctuates a little more than usual. At the same time a pump runs slightly longer, and cycle time grows by fractions of a second. Every single value stays inside its limits. No alarm. Two hours later the machine is down.

Afterwards there is usually an explanation: a seal beginning to leak, a worn part, a changed material. At the moment it happened, nobody knew. That raises an uncomfortable question: Must a system understand the cause before it can warn in a useful way?

Cause and pattern are two different things.

A cause explains why something happens. A pattern describes what occurs together — and what followed. Classic diagnosis looks for the cause. That requires knowledge of components, physics, and failure modes, which has to be modeled at length for every machine type.

A learning system can take another path. It observes which signal combinations recurred in comparable situations and what followed them. It does not need a model of the machine for that — only enough reliable experience with this exact machine.

Why alarms do not do this.

Alarms react to limits on individual signals. That is useful — but it means they usually fire only once the critical state has already arrived. Early change often stays inside every single limit. It becomes visible only in the interplay of several signals.

That is why stoppages can occur that no alarm predicts: not because nobody was looking, but because the information is not in any single value.

What a pattern can say — and what it cannot.

A pattern can support a statement like: “The last time these signals occurred together like this, a stoppage followed.” It cannot support the statement: “The seal is defective.”

That difference matters. A pattern is experience, not a diagnosis. It shows a relationship between observations, not the mechanism behind them. A system that invents a cause from a correlation creates trust it has not earned. A system that stays with the correlation can say honestly what it knows.

The boundary belongs to the concept.

Anyone who warns without a causal model has to impose rules:

No invented diagnosis. The system does not name a component or a location it cannot know.

Signals instead of guesses. It shows which signals stand out together — and leaves the physical interpretation to people.

References only with evidence. It may point to an earlier stoppage only when the current situation can actually be matched to it.

Know what it does not know. If comparable experience is missing, uncertainty stays visible instead of forcing an answer.

What does this mean for Copass?

Copass does not know technical causes. We do not treat that as a gap, but as a deliberate boundary. Copass should recognize relationships between signals, compare them with the experience of this one machine, and report what followed in a similar situation — early, before an alarm fires. The person remains the one who finds the physical cause and decides.

Not everything we describe here is fully in production today. It describes the direction in which we investigate and develop Copass — without selling Copass as a finished predictive-maintenance system.

The decisive question is therefore not whether a machine knows why something happens. It is: Does it know what followed last time — and can it show that?

Related research: When Is a Deviation Really a Deviation? — why early hints also gain meaning only through the right context.