Long-tailed acoustic-emission-based structural condition recognition in marine steel structures via cross-state residual injection

Marine steel structures endure cyclic loading throughout their service life. Acoustic emission (AE) is an effective passive non-destructive testing technique. It captures transient elastic waves released during damage evolution in real time and supports online damage monitoring. AE waveforms are complex, and the structural conditions to be recognized are diverse. Deep learning has therefore become essential for automated structural condition recognition. However, components remain intact for most of their service life, leaving fracture signals extremely scarce. The resulting long-tailed distribution biases model decision boundaries toward the intact class, causing fracture miss rates that compromise safety assessment. To address this, we propose Intact-to-Fracture Feature Fusion (I2F). Within a frozen backbone’s feature space, I2F computes the feature residual between intact donors and fracture recipients through energy-guided adaptive channel selection. It then injects the residual into high-energy channels at a controllable intensity. This process expands fracture-class distribution coverage during training without adding inference overhead. We validate I2F on an AE dataset from cyclic loading fatigue experiments on thick plates and thin-web beams, covering the full life-cycle from cumulative degradation to final fracture. I2F outperforms all comparison methods under both ResNet18 and MobileNetV2, achieving a Macro F1 of 0.952 and a Macro FNR of 0.041, with robust noise tolerance. These results highlight the potential of I2F as a promising strategy for reliable AE-based non-destructive testing under long-tailed conditions.