数据特征对剪力墙结构生成式智能设计的影响

Influence of data features on the generative adversarial network-based intelligent design for shear wall structures

深度学习通过提取和学习剪力墙结构数据高维特征掌握智能设计技术,有效推动了建筑结构智能化设计的发展。数据驱动的结构智能设计通常受训练数据的数量和分布特征影响,但鲜有研究开展数据特征相关分析。该文通过收集设计习惯相似且地域相近的剪力墙建筑-结构设计图纸数据,开展了数据分布和数量特征与设计结果的相关性分析;提出混合数据训练联合特征分组训练的两阶段改进训练方法,有效提升了智能设计质量;采用与训练数据高度非同源的其他地区数据对研究结果进行了测试与检验。分析表明:数据低维特征的回归分析难以有效指导设计生成,而基于深度学习的高维特征学习则能有效掌握结构设计的潜在规律;随着训练数据量增加,智能设计效果也将得到平均20%以上的提升。通过典型案例研究证明:智能设计与工程师设计的平面设计相似性及整体结构力学性能相似性均较高,最大层间位移角的平均差异仅约8%。但是,当测试数据与训练数据特征差异过于显著时,智能设计质量将会受限。该研究通过对数据分布和数量特征与智能设计结果的相关性分析,为下一步开展基于深度学习的智能化结构设计研究提供了数据特征影响方面的参考。

[Objective] Deep learning methods have mastered sophisticated structural design skills via feature extraction and data-driven learning, thus significantly advancing intelligent building structural design. The performance of data-driven intelligent structural design typically depends on the quantity and distribution of training data. However, only a few studies have investigated the relationship between the learned data features and the intelligent design, thus limiting its performance. This study aims to study the influence of training data distribution and quantity on structural design and subsequently improve intelligent design quality. [Methods] Two typical datasets are generated by collecting the shear wall architectural-structural design drawing data from two distinct regions(Beijing and Shandong) with different design habits. Based on the Shandong dataset, the data distribution characteristics are studied, and the correlation between the design conditions, the structural planer, the vertical shape, and the equivalent ratio of the shear wall to the architectural plan area are investigated using linear regression analysis. Subsequently, a generative adversarial network-based intelligent design method for building structures is adopted to extract and learn high-dimensional data features. According to the quantity and the distribution characteristics of data, this study has proposed a data augmentation method and a two-stage training method, wherein all data is hybridized for training in the first stage, and the data grouped by different design conditions is used in the second stage, thereby improving the design quality of intelligent design models. In addition, the Beijing dataset, which is substantially different from the training data, is used to evaluate and validate the study results using various training methods and training data quantities. Finally, to validate the results and illustrate the performance of the intelligent structural design, typical cases with three different design conditions are used, namely, 7 degree seismic intensity and 27 m structural height, 7 degree seismic intensity and 54 m structural height, and 8 degree seismic intensity and 39 m structural height. [Results] The analysis results demonstrated the following. 1) The regression analysis based on low-dimensional data features could not guide the design generation properly for complicated structures, such as shear walls. In contrast, high-dimensional feature learning based on deep learning might effectively capture the potential design laws and optimize the design generation. 2) With the improvement of the quantity of training data and the training strategy, the quality of the intelligent design structure increased by an average above 20%;however, the quality of the intelligent design was compromised when the properties of the test and training data were considerably different(with a shear wall ratio difference of over 50%). 3) Moreover, the analysis results were evaluated using relevant case studies. Regarding the plane's design and the mechanical performance of the overall structure, the intelligent design and the engineer's design had a high degree of resemblance, and the maximum interstory drift ratio varied by up to 8% on average. [Conclusions] Consequently, by assessing the data distribution, design conditions, and quantitative properties, the generative adversarial network-based intelligent structural design may provide high-quality designs with suitable training datasets. Furthermore, this study provides a reference for future research on intelligent structural design based on deep learning and the influence of data features.