Shared-first risk-aware open-set domain adaptation for catalogue-absent structural state identification

Abstract Bridge structural health monitoring (SHM) models are normally trained with a fixed catalogue of structural states, yet the bridge continues to change after the model is deployed. Traffic, support restraint, ageing, and local cracking can produce target responses for which no source label exists. In that situation, closed-set domain adaptation is risky because alignment may pull an unfamiliar response into a known class. We address this problem with a shared-first riskaware open-set domain adaptation (OSDA) procedure. The input is an 18-channel acceleration window represented by local Transformer patches and whole-window statistics. Stage 1 calibrates the C00-C06 source anchors. Stage 2 combines shared-class confidence, source-private probability, and prototype distance to rank target risk, and aligns only the low-risk samples. Stage3 queries a small part of the reserved high-risk pool and learns C07-C09 from those labels. We evaluate the procedure on the Abaqus MEDIUM10 and LOW10 protocols using anchor calibration, risk separation, unknown enrichment, alignment-pool contamination, matched method comparisons, and same-pool Stage 3 ablations. Across these tests, the main benefit is a cleaner shared alignment pool and a reserved set that remains available for expert review. The evidence is simulation based; real-bridge and long-term field validation are left for future work.