Why the failure window closes before any threshold fires

The most actionable intelligence about a developing failure exists in the hours before any alarm fires. That window is the only time you can actually prevent the failure. Once the threshold is crossed, you are in response mode. The question is whether you have a system that can see inside the window.

The structure of an industrial failure

Industrial failures rarely happen without warning. What they happen without is visible warning — warning that can be read by a system configured to watch for threshold crossings.

The typical failure pattern looks like this: a primary signal begins to drift from its normal behaviour, incrementally, over hours or days. The drift is subtle enough that no single data point falls outside any configured threshold. Correlated signals begin to show corresponding shifts — mechanical, thermal, chemical signals that normally move together start moving in a new relationship. Eventually, one or more signals crosses a threshold and alarms fire. By then, the failure is already in progress.

The early detection window: a case

In a pump system, four alarms fired in a six-hour period. The maintenance team investigated each alarm in sequence, treating them as unrelated events. The root cause — a developing fault in a related component — was not identified until the asset failed.

The historical data, analysed by Eyer, told a different story. The primary deviation from learned baseline behaviour appeared five hours and twenty-five minutes before the failure. The correlated pattern — the sequence of signals that Eyer reads as a single event rather than four separate alarms — was visible and interpretable in the data from the moment it began.

The window was five hours and twenty-five minutes. It was completely invisible to the threshold-based system.

Why static thresholds cannot see inside the window

Static threshold alarms are binary: a value is either above or below the line. They have no memory — each measurement is evaluated independently, with no reference to how that signal has behaved historically. They have no context — a deviation that would be significant on a Monday morning is evaluated the same way as an identical reading during a scheduled maintenance cycle.

The early detection window contains exactly the kind of information that threshold-based systems cannot process: gradual deviations from historical normal, correlated across multiple signals, in a context where the magnitude of the individual deviations is not yet alarming but the pattern is.

What it takes to see inside the window

Reading the early detection window requires three things: a historical baseline of how each signal normally behaves, a correlation model that maps statistical relationships across signals, and the ability to evaluate current behaviour against both — continuously, autonomously, without manual configuration.

This is what Eyer builds from your historical data. Three dynamic baselines per metric, not set by an operator but learned from how your specific assets have actually behaved over their operational history. A correlation matrix that captures the statistical relationships between signals as they exist in your operation, not in a generic industry model.

When current behaviour deviates from that learned normal — when the correlation pattern shifts in a way that is statistically significant but has not yet crossed any threshold — Eyer identifies it. That is the early detection window. It has been in your data all along.