ML on strain data: promising, but physics still does the heavy lifting

I have been testing ML on long strain gauge records to flag early fatigue issues and loose joints. The models are great at obvious anomalies, but the score is often not actionable. We saw bursts of false alarms at maneuver transitions, and one model sailed past an actual early crack because the signature was sequence driven and quiet, not a spike.

What helped was not a bigger net, but better features and guardrails. Rainflow cycle pairs, R ratio and mean stress, dwell time statistics, modal band energy, and temperature context improved transfer from lab coupons to flight. Adding physics constraints like monotonicity with Delta K and using calibrated uncertainty beat raw end to end training. A model that says "inspect soon, 60% confidence" is useful. One that is sure and wrong is not.

For those doing similar work, what features or constraints made your models stick in the real world? Anyone using Gaussian processes or conformal prediction for bounds? And how are you validating with the tiny amount of real crack data we actually have?