A Physics-Conditioned Multimodal Token-Stream Decoder and Contract-Grounded Workflow for Cross-Condition Aircraft Structural Health Monitoring

Abstract

Fatigue cracking around rivet holes is a major concern in aircraft skin panels, while practical structural health monitoring must handle heterogeneous sensors, unseen loading amplitudes, missing channels, and limited coupon data. This study develops SHM-GPT, a compact physics-conditioned multimodal token-stream decoder, together with a contract-grounded monitoring workflow. A 26-slot causal representation combines acoustic-emission, strain, Fibre Bragg Grating, and operating-condition information. Percentage-life progress is estimated with a continuous Huber regression head; early-onset classification and modality reconstruction are auxiliary training objectives, and a variance-floor penalty discourages constant predictions. Spectral constraints are applied to the condition encoder and query-key projections to limit sensitivity under load shifts. The 4.72-million-parameter model is evaluated on a 15-coupon working set drawn from an in-house 7050-aluminium centre-hole campaign. Across five scenarios, it obtains the lowest mean absolute error in four, and meets the stated spread and information diagnostics in three: boundary load extrapolation, loading-profile shift, and leave-one-specimen-out evaluation, with relative margins of 11.93%, 15.28%, and 3.01% over the strongest baseline, respectively. A small Temporal Convolutional Network is better in the two-coupon regime, and all learned methods fail the spread diagnostic when acoustic emission is removed. The deterministic contract checker matches the rule-defined verdict on 50 of 50 synthetic traces. The evidence therefore supports a shared multimodal decoder and an auditable decision-support prototype on this benchmark.