Force Spectrum Audits for AI-Based Structural Monitoring of a Full-Scale Aircraft
Vibration-based structural health monitoring can mistake a change in excitation for a change in structural condition. We audit measured shaker forces before evaluating a simple AI anomaly detector on six white noise records from the open BAE Systems Hawk T1A full structure dataset: the healthy aircraft and a controlled 254.3 g mass addition at one starboard wing location, each at nominal excitation levels 1, 3, and 5. Force power spectral density (PSD) was estimated over 5-256 Hz for five load cells and ten repeats per record. Five record medians followed the expected level progression. The mass-added level 1 record did not: its median force PSD was 0.1778 N2/Hz, 8.89 times the published level 1 value of 0.02 N2/Hz and close to the measured level 3 spectra, with band-limited RMS of 6.68 N versus 2.25 N for healthy level 1. We exclude this pair from matched-level damage claims. At the two force-matched levels, a Ledoit-Wolf Mahalanobis novelty detector trained without target-level healthy calibration was unstable. With five healthy target-level calibration repeats, MIMO frequency response features gave zero false alarms and ten of ten detections at both levels; raw acceleration features gave one false alarm in five healthy repeats at level 3 and none at level 5. These exploratory results come from one aircraft and one test record per state and level. The principal finding is methodological: audit measured input spectra before using nominal excitation labels to benchmark AI-based structural monitoring.
Authors
- Aaron Acheneje
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24424246
- Primary Topic
- Structural Health Monitoring Techniques
- Type
- preprint