ML//applied ML//operating regime

An operating regime is one of the distinct modes a machine or a plant runs in (start-up, partial load, full load, cleaning, idle), each with its own normal values and its own relations between variables, and identifying it is the step that lets thresholds and models judge a reading against the right normal. A pump drawing 40 A is healthy at full flow and suspicious at minimum flow; a single limit for both is either blind or noisy.

An operating regime is one of the distinct modes a machine or a plant runs in (start-up, partial load, full load, cleaning, idle), each with its own normal values and its own relations between variables, and identifying it is the step that lets thresholds and models judge a reading against the right normal. A pump drawing 40 A is healthy at full flow and suspicious at minimum flow; a single limit for both is either blind or noisy.

When the regime is written down somewhere (a recipe number in the PLC, a speed setpoint, a batch phase in the MES), it is read from there and used as a column. When it is not, operating regime detection finds it in the data with clustering, which in a plant is the most profitable use of clustering: grouping the work cycles of 200 pumps by vibration and consumption reveals four modes nobody had labelled. K-means is fast and enough when the modes are compact clouds; a Gaussian mixture or spectral clustering handles elongated or curved ones.

Thresholds are normalized per regime. Each regime gets its own mean, spread and limits, or its own PCA monitoring model, because that method assumes a single regime and fills its QQQ statistic with false alarms when the plant switches modes. A residual model is treated the same way: fitted per regime, or given the regime as an input.

Transitions behave like a regime of their own. Start-ups and grade changes produce values that no steady regime explains, and a detector that ignores them alarms at every start; the usual remedies are a blanking time after each switch or a separate model for transients.

A clustering always returns as many groups as it was asked for, whether they exist or not. Four clusters do not prove four regimes, so they are checked against the physics or the operators before anyone writes thresholds on them.

Regime comes before the detector. A contextual anomaly (80 °C at no load) is visible only once the regime is known, which is why anomaly detection in a plant starts by splitting the data this way.