The first three monographs in this series — FDI, FTC, PHM — covered the engineering of fault-handling as the textbooks present it. Clean models, clean problems, clean algorithms. This one covers what happens when you try to put any of it into a real production system.
It turns out the gap between textbook and fleet is enormous. The model you trained on one engine doesn’t work on a different one. Your beautiful neural network won’t fit in 64 KB. The regulator wants explanations for every decision. The labels you assumed don’t exist. The system you’re diagnosing isn’t even one system — it’s fifty computers on a CAN bus, only some of which talk to you.
This monograph is the fourth and final post in the diagnostics series, and it’s a survey of the open problems that the field is actively working on:
- Fault Detection and Isolation
- Fault-Tolerant Control
- Prognostics and Health Management
- Frontiers in Diagnostics (this post)
What it covers
Eight chapters. About forty-five minutes to read.
§1 — The gap between the textbook and the fleet. The recurring observation that motivates everything else in the post.
§2 — Domain adaptation. Training here, deploying there. Why the model doesn’t transfer and what to do about it. CORAL, DANN, the small set of practical techniques.
§3 — Physics-informed neural networks. Physics as regularizer. The hybrid approach that’s eating both pure-physics and pure-data approaches.
§4 — Explainable AI. The certification problem. Why “the model said so” isn’t an answer that survives an aerospace audit, and what counts as a decision-grade explanation.
§5 — Edge deployment. The 64 KB constraint. What survives when you have to fit a diagnostic model on a microcontroller alongside the control software it’s monitoring.
§6 — Digital twins at fleet scale. When you have ten thousand identical units in the field, the per-unit model can be richer than any one engineer could maintain. The architecture and economics of fleet-scale telemetry.
§7 — Anomaly detection without labels. The reality: failures are rare, labels are non-existent, and you have to find faults using only the structure of normal data. SVDD, one-class methods, deep autoencoders.
§8 — Distributed FDI. When the “system” is fifty computers on a CAN bus. Consensus algorithms, redundant diagnosis, the new constraints that come with distribution.
Read it
← Back to Autonomy