Scholarlink

<p>Wave motion in variable sea states affects SOV operation times, particularly SOV-RM. A hybrid framework, further extended to an Adaptive Transformer–Transfer Learning (AdaTF–TL) approach, is proposed. By integrating an adaptive transformer architecture with transfer learning, the framework aims to extend the feasible construction windows for SOV operations. Temporal Similarity Quantization (TSQ) quantifies inter- and intra-sequence distribution shifts in SOV-RM caused by wave frequency and direction. Intra-sequence heterogeneity is mitigated through Dynamic Weighted Distance (DWD), which adjusts feature weights via adaptive attention. For inter-sequence heterogeneity across sea states, a Dynamic Unfreezing Strategy (DUS) transfers pre-trained model parameters for conditional probability alignment. Validation under sea states 3–6, using Pierson–Moscowitz and Longuet–Higgins spectra, demonstrates that the Adaptive Transformer–Transfer Learning (AdaTF–TL) framework achieves a prediction accuracy of 99.18%, outperforming four comparative models, including AdaRNN. Its robustness is further confirmed through RT-LAB–based simulations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Expanding the Construction Window Period: Heterogeneity Analysis and Compensation Prediction of Service Operation Vessel (SOV) Roll Motion (SOV-RM) Using Adaptive Transformer and Transfer Learning

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Qin Zhang,
  • Feng Zhou,
  • Bang-ping Gu,
  • Xiong Hu

摘要

Wave motion in variable sea states affects SOV operation times, particularly SOV-RM. A hybrid framework, further extended to an Adaptive Transformer–Transfer Learning (AdaTF–TL) approach, is proposed. By integrating an adaptive transformer architecture with transfer learning, the framework aims to extend the feasible construction windows for SOV operations. Temporal Similarity Quantization (TSQ) quantifies inter- and intra-sequence distribution shifts in SOV-RM caused by wave frequency and direction. Intra-sequence heterogeneity is mitigated through Dynamic Weighted Distance (DWD), which adjusts feature weights via adaptive attention. For inter-sequence heterogeneity across sea states, a Dynamic Unfreezing Strategy (DUS) transfers pre-trained model parameters for conditional probability alignment. Validation under sea states 3–6, using Pierson–Moscowitz and Longuet–Higgins spectra, demonstrates that the Adaptive Transformer–Transfer Learning (AdaTF–TL) framework achieves a prediction accuracy of 99.18%, outperforming four comparative models, including AdaRNN. Its robustness is further confirmed through RT-LAB–based simulations.