estkit

Benchmark explorer

Steady-state state-of-charge RMSE, in percent of capacity, for every cell-level estimator across the twelve fault scenarios. Darker is worse. Switch plants, sort by any column, hover a cell for the number. Lower is better everywhere.

better worse

div in the source data marks a run that diverged in at least one seed; blank cells are scenarios not defined on that plant.

Composite ranking & cost

Composite ranking
Geometric-mean RMSE over the twelve scenarios (bar) with the worst scenario (tick), by family.
ECM vs SPM
Nominal accuracy on the two plants is uncorrelated: the converged Kalman filters are twenty times worse on the electrochemical plant.

Pack, attitude and state-of-health

Pack
Twelve-cell module: averaging inside the filter halves the per-cell error.
Attitude
Attitude estimators: the accelerometer is not a gravity reference in a banked turn.
SOH
120-mission ageing: joint/dual filters track capacity to the current-sensor accuracy.

The reference stack

A 0.5 µs EKF that adapts its measurement noise and bounds its memory, with a gated capacity state and an independent Coulomb counter beside it on a calibrated current sensor, has no failure scenario in this benchmark on either plant. The full argument — eight design rules and a reference architecture — is in the report's recommendations chapter.

Read the recommendations →