Task — fault_detection_global

benchmark version 1.1.0 · held_out results v1.0.0 · splits v1.0/held_out, v1.1/multi_grid
updated 2026-09-29 13:18 UTC · git 7243f310ddbc

tier 1 · fault_detection · global view · 2 classes · all cells measured

What this task is. Decide, from one 50 ms waveform window, whether a short-circuit fault is present. This is the core protection decision: a relay that misses a fault leaves it burning (missed-fault rate), one that trips on a healthy window disconnects customers for nothing (false-alarm rate). Both error types are reported alongside balanced accuracy because their costs are asymmetric in practice. Global view: every cubicle in the grid at once (width differs per grid), the wide-area upper bound on observability.

How it's scored

Headline metric balanced_accuracy (↑ higher is better); full metric set: balanced_accuracy · macro_f1 · missed_fault_rate · false_alarm_rate · accuracy. Metrics are a frozen contract (src/evemtbench/evaluation/metrics.py); label derivation: src/evemtbench/tasks/labels.

Example window

example waveform window

event flt_1phg_incipient · grid double_line (adapt_grid/test split) · sample 2, window #46, cubicle 3 (max-RMS) · 50 ms / 480 samples @ 9.6 kHz · top: 3-phase current, bottom: 3-phase voltage · figure provenance in assets/manifest.json

Results (held_out)

fault_detection_global — every grid × baseline; click headers to sort
taskgridbaselineheadlinetestbenchmarkparamsfit/seedtrace
fault_detection_globalcigre_mvmajoritybalanced_accuracy ↑0.500 ± 0.000 (n=5)0.500 ± 0.000 (n=5)—0strace: output · eval-code · log · train-code · config · W&B
fault_detection_globalcigre_mvrandom_forestbalanced_accuracy ↑0.827 ± 0.003 (n=5)0.985 ± 0.001 (n=5)—3.0mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalcigre_mvmlpbalanced_accuracy ↑0.763 ± 0.001 (n=5)0.959 ± 0.000 (n=5)42.9M14.2mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalcigre_mvgrubalanced_accuracy ↑0.786 ± 0.027 (n=5)0.963 ± 0.006 (n=5)117k12.9mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalcigre_mvcnnbalanced_accuracy ↑0.799 ± 0.022 (n=5)0.961 ± 0.008 (n=5)124k12.6mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalcigre_mvresnetbalanced_accuracy ↑0.799 ± 0.037 (n=5)0.964 ± 0.015 (n=5)602k9.9mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaldouble_linemajoritybalanced_accuracy ↑0.500 ± 0.000 (n=5)0.500 ± 0.000 (n=5)—0strace: output · eval-code · log · train-code · config
fault_detection_globaldouble_linerandom_forestbalanced_accuracy ↑0.890 ± 0.001 (n=5)0.500 ± 0.000 (n=5)—40strace: output · eval-code · log · train-code · config · W&B
fault_detection_globaldouble_linemlpbalanced_accuracy ↑0.849 ± 0.001 (n=5)0.986 ± 0.000 (n=5)12.0M4.4mtrace: output · eval-code · log · train-code · config
fault_detection_globaldouble_linegrubalanced_accuracy ↑0.858 ± 0.021 (n=5)0.943 ± 0.114 (n=5)69k2.6mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaldouble_linecnnbalanced_accuracy ↑0.867 ± 0.017 (n=5)0.987 ± 0.001 (n=5)68k3.8mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaldouble_lineresnetbalanced_accuracy ↑0.863 ± 0.016 (n=5)0.987 ± 0.001 (n=5)530k8.2mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvmajoritybalanced_accuracy ↑0.500 ± 0.000 (n=5)0.500 ± 0.000 (n=5)—0strace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvrandom_forestbalanced_accuracy ↑0.866 ± 0.003 (n=5)0.500 ± 0.000 (n=5)—1.1mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvmlpbalanced_accuracy ↑0.848 ± 0.001 (n=5)0.985 ± 0.000 (n=5)26.7M8.4mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvgrubalanced_accuracy ↑0.857 ± 0.022 (n=5)0.985 ± 0.003 (n=5)92k5.7mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvcnnbalanced_accuracy ↑0.878 ± 0.010 (n=5)0.987 ± 0.003 (n=5)94k8.4mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globaltestgrid_110kvresnetbalanced_accuracy ↑0.864 ± 0.020 (n=5)0.987 ± 0.004 (n=5)564k9.5mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39majoritybalanced_accuracy ↑0.500 ± 0.000 (n=5)0.500 ± 0.000 (n=5)—0strace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39random_forestbalanced_accuracy ↑0.934 ± 0.001 (n=5)0.500 ± 0.000 (n=5)—6.9mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39mlpbalanced_accuracy ↑0.854 ± 0.000 (n=5)0.714 ± 0.072 (n=5)103.4M47.3mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39grubalanced_accuracy ↑0.854 ± 0.004 (n=5)0.848 ± 0.222 (n=5)211k35.8mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39cnnbalanced_accuracy ↑0.865 ± 0.010 (n=5)0.594 ± 0.267 (n=5)234k40.9mtrace: output · eval-code · log · train-code · config · W&B
fault_detection_globalieee39resnetbalanced_accuracy ↑0.870 ± 0.001 (n=5)0.500 ± 0.000 (n=5)744k48.6mtrace: output · eval-code · log · train-code · config · W&B

Metric definitions

headline & reported metrics for this view — the metric set is a frozen contract; each links to the exact function that computes it

metricdefinitionentry
balanced_accuracymean of per-class recall — the majority class cannot buy a good scoreevaluate()
macro_f1unweighted mean of per-class F1 (zero_division=0)evaluate()
missed_fault_rateFN / (TP+FN) — fraction of true faults the model missedevaluate()
false_alarm_rateFP / (FP+TN) — fraction of non-faults flaggedevaluate()
accuracyfraction of exact-match predictionsevaluate()

Reproduce one cell

PYTHONPATH=src python -m evemtbench.baselines.runner --task fault_detection_global --protocol held_out \
    --grid <grid> --baseline <baseline> --seeds 0 1 2 3 4