Research proposal · v1.6 preview · English supplement
Not an official page of Kyushu University or I²CNER. The Japanese proposal is the full document; this English page is a supplement. Proposed host: I²CNER, Kyushu University (under discussion). The research stay is not confirmed.
Benchmark v1.5.0 (illustrative)
Illustrative (例示). All numbers come from a public, reproducible benchmark (v1.5.0; Python = JavaScript = expected_output.json). They do not describe any real station, alarm setting or product.
Trade-off
FA = false alarms on no-leak test fills (unnecessary shutdowns). Leak columns = correct diagnosis (sign included), 2% / 5% of the hold flow, δ = 0. TTD = median time to detection, detected cases only.
| Detector | FA declared ±0.3 s (max) | FA +0.5 s | FA drift | FA meter lag 1 s | FA meter-bias step | 2% leak | 5% leak | TTD 2% |
|---|---|---|---|---|---|---|---|---|
| A — inventory-corrected balance | 1% | 81.5% | 57% | 100% | 0% | 0/400 | 2/400 | — |
| B2 — fixed timing allowance u = 0.4 s | 0.5% | 99.75% | 35.75% | 100% | 13% | 36/400 | 307/400 | 56.5 s |
| B3 — state-dependent allowance u_t | 0.75% | 99% | 31.25% | 100% | 3.5% | 1/400 | 252/400 | 153.0 s |
| B4 — two-stage: B2 (low threshold) → CUSUM | 0.25% | 40% | 26.75% | 1.5% | 59.5% | 223/400 | 387/400 | 5.3 s |
| B5 — GLR: leak + admissible offset | 0% | 8.25% | 13.25% | 100% | 24.25% | 64/400 | 320/400 | 41.7 s |
| C — one-sided CUSUM pair (1 s blocks) | 0.25% | 37% | 28% | 0.75% | 82% | 322/400 | 386/400 | 4.4 s |
What it says (and does not say)
- The state-dependent allowance (B3) did not help: 2% leak −8.75 pp (95% CI −16 to −2.25) versus B2.
- Two-stage detection (B4) recovers small-leak detection while keeping false alarms ≤ 0.25% inside the declared range, but a meter-bias step raises alarms in 59.5% of fills.
- Calibration matters more than the detector: calibrated at δ = 0 only, A detects 280/400 of 2% leaks but shuts down 100% of no-leak fills at +0.2 s; calibrated over the declared range, A detects 0/400.
- Failures are detector failures, not proven indistinguishability.
- Value of added measurements (200 fills): an independent time reference (±0.02 s) lets A detect 122/200 of 2% leaks (C 182/200).
Reproduce
Code (MIT): mass_balance_clock.py · mass_balance_clock.mjs · expected_output.json · README
SHA-256 (frozen):
653e4ed91a06d6cc2ff777b629641a591427a8b58a921785abb5dde41bb7ba88 mass_balance_clock.py
c5f43ad410a677cc914dd42ae50245fce1884879ffab38ef478f0a8120047aa3 mass_balance_clock.mjs
c2642719411660360de6113e17f84b462ab109589c9a38d62df4fff9748242a7 expected_output.json
e7cf1981c6f08b95a3b1bc753d9879db889bbc32626ccfcdffad21b6fc10f444 README.mdInteractive tables and ROC (Japanese): #/benchmark