针对输电铁塔结构参数不确定性及监测信息不完备导致的响应预测难题, 提出了一种基于图神经网络(graph neural network, GNN)的深度学习模型,实现了不完全信息条件下的结构力学响应预测。该模型通过融合网络结构优化与随机标签掩盖算法, 有效缓解了数据缺失的影响。为适应工程不完全信息场景,设计了双阶段(训练-推理)验证框架:训练阶段通过模拟信息缺失学习数据间内在联系,推理阶段通过知识迁移实现未知响应预测。通过消融实验对比分析不同改进模块对模型性能的影响, 并进一步研究了RPTTGN-II(response prediction of transmission towers based on GNN under incomplete information)模型在不同未知响应参数占比下的预测性能, 验证该模型的准确性和有效性。研究结果表明, 该模型在处理位移预测任务时具有明显优势, 未知位移参数占比高达50%时预测精度在90%以上, 而在处理轴力预测任务时, 尽管模型预测精度有所下降, 但在一定的信息缺失情况下, 其结果仍可提供有效参考。
英文摘要:
To address the challenge of response prediction for transmission towers under structural parameter uncertainty and incomplete monitoring information, a graph neural network (GNN)-based deep learning model is proposed, achieving structural mechanics response prediction under incomplete information conditions. By integrating network architecture optimization with a random label masking algorithm, the impact of data missingness has been effectively mitigated. To adapt to engineering scenarios with incomplete information, a two-stage (training-inference) validation framework is designed: during training, data intrinsic correlations are learned through simulated information missingness, while during inference, unknown responses can be predicted through knowledge transfer. Ablation experiments are conducted to compare the effects of different improvement modules on performance of model. Furthermore, the prediction performance of RPTTGNN-II (response prediction of transmission towers based on GNN under incomplete information)under varying proportions of unknown response parameters is investigated, validating its accuracy and effectiveness in practical engineering applications. Research results demonstrate that the model exhibits significant advantages in handling displacement prediction tasks, maintaining over 90% accuracy when the proportion of unknown displacement parameters is as high as 50%. For axial force prediction tasks, although the prediction accuracy performance of model is constrained, the model still provides valuable references under certain information missing conditions.