State duration
ReliabilityBench treats a row as a change effective at its timestamp. The last row is an open state with no duration. An entry into the exact normalized label RUNNING is counted as one start observation; an exit from RUNNING is counted separately.
Time at load
The load analyzer separates exactly 0% as idle, uses ten positive bands through 100%, and retains values above 100% in a review band. It does not interpolate irregular timestamps or convert between ratings.
Threshold excursion duration
Consecutive intervals in the same outside zone form one event. Values exactly equal to a limit are inside. This left-hold convention assumes the current sample represents the interval that follows; the final sample is not assigned time.
Rate of change
A rise event requires rate greater than the positive rise limit. A fall event requires rate less than the negative fall magnitude. The time base changes reporting units only; it does not smooth or resample the series.
Validation boundaries
- Timestamps must parse and increase strictly; duplicates and reversed rows are rejected.
- Values and limits must be finite; NaN and Infinity are rejected.
- Load readings must be 0–200% and equally spaced; the interval must be positive and no more than 1,440 minutes.
- Threshold lower limit must be less than the upper limit.
- Rate rise/fall magnitudes and time base must be positive.
- Each pasted dataset is bounded to 240 rows or readings for responsive local-browser use.
Decision boundary
These methods summarize declared source data. They do not establish the correct operating envelope, infer failure causality, calculate degradation or remaining life, quantify sensor uncertainty, replace historian quality flags, or make a safety/compliance decision.