Definition
No existing KPI measures autonomous control loop oscillation specifically.
The Measurement Gap
The decision PODI supports is whether the autonomous controller is tuned correctly for the process dynamics and whether the control authority envelope needs to be narrowed or the control loop needs damping. The measurement gap exists because aggregate SPC statistics (Cpk, range charts) can look stable while the underlying process oscillates rapidly within the control limits. The AI makes an adjustment, the process responds, the AI counter-adjusts, and the cycle repeats at a frequency that SPC averaging masks. The invisible failure mode is control energy waste and micro-variation injection. The oscillation consumes control authority without improving the process, and the rapid parameter swings inject variation into the product that manifests as subtle quality inconsistency not visible in standard SPC charts. In multi-parameter systems, oscillation on one controlled parameter can induce sympathetic variation on physically coupled but uncontrolled parameters, contributing to the drift that MDQDI measures.
Formula
For each AI-controlled parameter, count the number of direction reversals per unit time and weight by magnitude. The formula is: PODI_p = (Reversals_p / T) times (Avg_reversal_magnitude_p / Range_p).
Variables: Reversals_p is the count of direction reversals for parameter p in the measurement window, where a reversal is a sign change in the parameter's first derivative. T is the measurement window duration in hours. Avg_reversal_magnitude_p is the average absolute change per reversal for parameter p in parameter units. Range_p is the normal operating range of parameter p (max setpoint minus min setpoint) in parameter units.
Sign convention: 0 equals no oscillation (stable control). Higher values indicate more oscillation. The result is dimensionless, enabling comparison across parameters and across time.
Edge cases: normal hunting around setpoint exists in every controller, so establish a noise floor where reversals below a minimum magnitude threshold (such as 2% of operating range) are excluded. Legitimate step changes during setpoint transitions should be excluded by removing the settling window (such as the first 5 minutes after a setpoint change). For multi-parameter oscillation where two parameters oscillate in anti-phase, track cross-parameter correlation as a supplementary signal.
Proposed formula, pending validation pilots.