Work

Attention-BiLSTM-LMA: An Attention-Pooled Bidirectional LSTM-Conditioned Laplace Moving-Average Model for Non-Gaussian Hull Girder Stress Responses in Container Ships

Year: 2026
Type: article
Abstract: Wave-induced hull girder stresses in container ships can exhibit non-Gaussian response characteristics that influence fatigue damage and extreme-response assessment. However, classical extreme-value and fatigue assessment procedures rely on Gaussian assumptions and may severely underestimate design loads. The Laplace-driven moving-average (LMA) process offers an analytically tractable non-Gaussian alternative, but its kernel and shape parameters are typically obtained by moment matching a single measured sea-state record. This study proposes Attention-BiLSTM-LMA, which uses an attention-pooled bidirectional long short-term memory (BiLSTM) encoder to process short stress-history windows. The encoded features are combined with a multilayer perceptron embedding of sea-state covariates and passed to a Gaussian probabilistic output head to predict LMA skewness and kurtosis parameters. The model is calibrated using real measurements and validated using full-scale hull girder stress measurements from a 2800 TEU container ship. The rainflow damage for the reference voyage agreed with the published result to within approximately 0.3%. The Gaussian narrow-band estimate was 14.2% higher than the rainflow result. Winterstein’s crossing-intensity formula for Hermite-transformed Gaussian processes is used to provide a closed-form, directly comparable treatment of extremes and fatigue for three models: Gaussian, classical LMA, and Attention-BiLSTM-LMA. The Gaussian, classical LMA, and Attention-BiLSTM-LMA models yield 100-year return-period stress estimates of 253, 359, and 282 MPa, respectively, indicating substantial differences in the extrapolated upper tail. The Gaussian narrow-band fatigue estimate exceeds the measured rainflow damage by 18.1%, whereas the error decreases to 8.3% when spectrum-matched Gaussian stress histories are evaluated using rainflow counting. Both non-Gaussian models yield fatigue damage estimates closer to those obtained by rainflow counting of measured stresses, with errors of +1.6% for classical LMA and −0.8% for Attention-BiLSTM-LMA. On held-out voyages, the skewness predicted by the BiLSTM output head correlates with classical per-window estimates at r = 0.63, while real-data recalibration substantially reduces the mean bias in kurtosis predictions. These findings indicate that the proposed model provides fatigue damage estimates comparable to those of classical per-window LMA fitting without manual refitting. Nevertheless, the lower extreme-stress estimate relative to classical LMA indicates potential non-conservatism for ultimate-limit-state (ULS) design, despite the model’s close agreement with measured fatigue damage.
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7237200739
FWCI:
Cites: 34
Cited by:
Related to: 10