Diagnosis, recovery ordering, and a learned detector answer different questions
What the case studies show
- Stage isolation works: swapping only evaluation moves correct decisions from 9 to 25 of 31
- Dynamics beat symptoms: trajectory similarity separates faults that share the same limit crossings
- Closing the loop works: procedure ETRs + alarm verification recover 28 of 31 under P/T or highest P
- Learned detection beats thresholds and holds under noise: 31/31 detect on every mode×intensity cell; F1 stays ≥ 0.92 (worst ar1/med), median latency ≤ 4 s
Limitations
- Recovery study uses knowledge-graph diagnosis only
- LSTM train / holdout share one synthetic generator
- Study C open-loop detection not yet gating full diagnosis
- Gate thresholds and verify window not tuned
- Robotics still a surrogate; no operator-in-the-loop yet