Inside the algorithm: how Eyer builds three dynamic baselines from your time-series data

Most anomaly detection approaches use a single reference point per metric: a threshold, a moving average, or a fixed normal range. Eyer uses three. Understanding why three baselines are necessary — and how they are built from your data — is the foundation of understanding what Eyer sees that static monitoring cannot.

Why one baseline is not enough

A single baseline per metric assumes that the metric has one normal state. In industrial operations, this is almost never true. A pump in morning startup behaves differently than a pump at steady-state load. A reactor in the first week of a campaign behaves differently than the same reactor approaching its cleaning cycle. A compressor running in winter at ambient temperature of minus ten degrees behaves differently than the same compressor in summer at plus thirty.

A single baseline cannot capture these variations. It is either too wide — capturing all operating states, which means it misses deviations that are significant within a specific state — or too narrow — calibrated for one state, which means it fires false positives during transitions and other states.

Three baselines capture the operating envelope of a metric in a way that a single baseline cannot: a lower baseline representing the typical minimum behaviour, a central baseline representing the typical steady-state behaviour, and an upper baseline representing the typical maximum behaviour — all dynamic, all learned from how your specific asset has actually behaved.

How the baselines are built

Eyer's baseline learning works directly from your historical time-series data — the same data your historian has been collecting. The algorithm analyses the distribution of values for each metric across your operational history, segmenting the data by operating context: load levels, time-of-day patterns, seasonal variation, asset state transitions.

From this segmented historical distribution, three baselines are derived: the lower baseline captures the 10th-percentile behaviour under each operating condition, the central baseline captures the median, and the upper baseline captures the 90th-percentile. The segmentation means that the baselines are dynamically appropriate for the current operating context — the baseline for a pump under high load is different from the baseline for the same pump under low load, and Eyer knows which is applicable at any given moment.

The baselines adapt continuously. As your operation changes — new equipment, process modifications, seasonal shifts — the baselines update to reflect the new normal. This means that Eyer's model of your operation is always current, without any manual intervention.

Correlation: the second layer

Baselines alone are not sufficient. An anomaly in a single signal might be a sensor fault, a transient fluctuation, or a genuine process event — and without correlating it against other signals, there is no way to know which.

Eyer maps the statistical correlation structure of your operation across every metric. This is not event-based correlation (which signals fired alarms at the same time?) but time-series correlation (which signals move together statistically, and with what time-lag relationship?). A bearing temperature rise that consistently precedes a pressure drop in a correlated downstream component tells a different story than the same temperature rise in isolation.

When a deviation from a learned baseline appears, Eyer immediately checks the correlation map: which other signals are currently deviating? Are they deviating in a direction and magnitude consistent with the historical correlation pattern? If so, the alert carries the full cascade context — what is primary, what is correlated, what the likely causal relationship is.

Try the algorithm on your data

The fastest way to understand what Eyer's algorithm sees in industrial time-series data is to run it on your data. The Fast Forward takes one week: your historical data in, a POV report out, showing the patterns your existing monitoring has never flagged.

Alternatively, explore the interactive demonstration at sim.eyer.ai, which shows how the three-baseline model and correlation analysis apply to real industrial data patterns.