A dual-pathway transformer framework for corrosion-fatigue crack growth rate prediction from raw acoustic emission signals

Real-time prediction of crack growth rate (CGR) in marine structures under corrosion-fatigue conditions is critical for preventing catastrophic failures. Traditional acoustic emission (AE) monitoring relies heavily on mechanical parameters such as fatigue cycle count ( N ) and stress intensity factor ( Δ K ), as well as threshold-based AE features such as counts and energy, which are often unreliable or unavailable in real-time scenarios. Therefore, we propose a novel parallel transformer framework to predict the CGR directly from raw AE signals, eliminating the dependency on mechanical priors and threshold-based AE parameters. This study utilizes in situ AE signals acquired at 5 MHz sampling rate during corrosion-fatigue propagation in ship steel defence metallurgical research (DMR) 249A. The experiments were conducted under three distinct load ratios of 0 . 1 , 0 . 25 , and 0 . 5 , simulating the variable loading conditions of sea-state IV. The B-spline functional measurement filter is applied to the high-dimensional raw AE waveforms to mitigate the computational challenges arising from high dimensions of raw AE data. This preprocessing step reduces the input dimension of data by approximately 60%, while preserving the crucial signal topology. The resulting B-spline coefficients are integrated with separately computed cumulative damage state features and processed by a dual-stream parallel transformer, capturing both historical damage trends and instantaneous fracture dynamics. The proposed framework attained an R 2 of 0 . 87 , 0 . 87 , and 0.90 for load ratios of 0 . 1 , 0 . 25 , and 0 . 5 , respectively, validated against experimental fatigue machine logs. Field emission scanning electron microscopy analysis verified that captured transients originated from crack propagation rather than ambient noise. Furthermore, the proposed model demonstrated its generalization capability by rapidly calibrating on new stress ratios using only 30% of target data. The computational efficiency and robust generalization of the proposed framework confirm the model’s suitability for autonomous and continuous structural health monitoring of naval assets.