廖文杰 陸新征 黃羽立 趙鵬舉 費(fèi)一凡 鄭哲
摘要:建筑結(jié)構(gòu)的智能化方案設(shè)計(jì)是智能建造的重要內(nèi)容。既有研究提出了基于深度神經(jīng)網(wǎng)絡(luò)的剪力墻結(jié)構(gòu)生成式設(shè)計(jì)方法框架、智能設(shè)計(jì)算法、設(shè)計(jì)性能評(píng)價(jià)方法等,完成了從數(shù)據(jù)驅(qū)動(dòng)到物理增強(qiáng)的智能化設(shè)計(jì)方法的發(fā)展,但目前尚未有研究針對(duì)不同設(shè)計(jì)條件下數(shù)據(jù)驅(qū)動(dòng)和物理增強(qiáng)方法的設(shè)計(jì)能力進(jìn)行詳細(xì)對(duì)比,且基于計(jì)算機(jī)視覺與基于力學(xué)性能的評(píng)價(jià)方法尚未有明確的關(guān)系,難以有效保證計(jì)算機(jī)視覺評(píng)價(jià)方法的合理性?;谏疃壬墒剿惴▽?duì)比和算例分析,開展數(shù)據(jù)驅(qū)動(dòng)和物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)方法的詳細(xì)對(duì)比,并進(jìn)一步驗(yàn)證基于計(jì)算機(jī)視覺評(píng)價(jià)與基于力學(xué)分析評(píng)價(jià)方法的正相關(guān)性。結(jié)果表明:數(shù)據(jù)驅(qū)動(dòng)的方法易受到數(shù)據(jù)質(zhì)量與數(shù)量的約束,而物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)的方法設(shè)計(jì)性能更加穩(wěn)定,基本擺脫數(shù)據(jù)質(zhì)量和數(shù)量的約束;基于計(jì)算機(jī)視覺綜合評(píng)價(jià)指標(biāo)SCV的合理性閾值為0.5,對(duì)應(yīng)力學(xué)性能差異約為10%。
關(guān)鍵詞:智能化結(jié)構(gòu)設(shè)計(jì);生成對(duì)抗網(wǎng)絡(luò);數(shù)據(jù)驅(qū)動(dòng);物理增強(qiáng);設(shè)計(jì)評(píng)價(jià)
中圖分類號(hào):TU318? ? ?文獻(xiàn)標(biāo)志碼:A? ? ?文章編號(hào):2096-6717(2024)01-0082-11
Intelligent generative structural design methods for shear wall buildings: From data-driven to physics-enhanced
LIAO Wenjie, LU Xinzheng, HUANG Yuli, ZHAO Pengju, FEI Yifan, ZHENG Zhe
(Key Laboratory of Civil Engineering Safety and Durability of China Education Ministry, Department of Civil Engineering, Tsinghua University, Beijing 400045, P. R. China)
Abstract: Intelligent structural design in the scheme phase is an essential component of intelligent construction. Existing studies have proposed the deep neural network-based framework of intelligent generative structural design, intelligent design algorithms, and design performance evaluation methods for shear wall structures, which have developed intelligent structural design methods from data-driven to physics-enhanced data-driven. However, little detailed design performance comparison of data-driven and physics-enhanced methods under different design conditions is conducted. Furthermore, the relationship between the computer vision-based and mechanical analysis-based evaluation methods are still unclear, resulting in difficulties in effectively guaranteeing the rationality of the computer vision-based evaluation methods. Hence, in this study, the comparative analysis of data-driven and physics-enhanced intelligent design methods is conducted by algorithm comparison and case studies; and the consistent relationship between computer vision-based and mechanical analysis-based evaluation methods is validated. The comparison results reveal that data-driven methods are more prone to be limited by the quality and quantity of training data. In contrast, the physics-enhanced data-driven design method is more robust under different design conditions and is little affected by the data-caused limitation. Moreover, the rationality threshold of the computer vision-based evaluation index (SCV) is 0.5, corresponding to a difference in the mechanical performance of approximately 10%.
Keywords: intelligent structural design; generative adversarial networks; data-driven; physics-enhanced; design evaluation
人工智能的應(yīng)用為土木工程領(lǐng)域的規(guī)劃、設(shè)計(jì)、建造、維護(hù)和防災(zāi)技術(shù)帶來了重大變革與重塑[1-4],智能建造已成為土木領(lǐng)域的重點(diǎn)發(fā)展方向。建筑結(jié)構(gòu)智能化設(shè)計(jì)則是智能建造的重要內(nèi)容,不僅可以減少大量繁瑣的人工設(shè)計(jì)流程,還能為設(shè)計(jì)人員提供更加多樣的初始設(shè)計(jì)選擇,最終達(dá)到更好的設(shè)計(jì)效果[5-14]。建筑結(jié)構(gòu)設(shè)計(jì)主要包括方案設(shè)計(jì)、優(yōu)化設(shè)計(jì)及施工圖設(shè)計(jì)階段,其中方案設(shè)計(jì)對(duì)后續(xù)設(shè)計(jì)影響關(guān)鍵,且對(duì)設(shè)計(jì)知識(shí)與經(jīng)驗(yàn)的需求較高,并對(duì)設(shè)計(jì)速度有較高的要求。
智能化方案設(shè)計(jì)通常采用進(jìn)化算法等對(duì)結(jié)構(gòu)設(shè)計(jì)方案進(jìn)行搜索和設(shè)計(jì)[15-17],但是,進(jìn)化算法難以有效學(xué)習(xí)結(jié)構(gòu)設(shè)計(jì)數(shù)據(jù)與經(jīng)驗(yàn),迭代優(yōu)化效率有限,難以滿足快速完成結(jié)構(gòu)方案設(shè)計(jì)的需求。相比而言,具備大數(shù)據(jù)特征提取與學(xué)習(xí)能力的深度生成算法快速進(jìn)步,例如,基于卷積的生成神經(jīng)網(wǎng)絡(luò)、圖(Graph)生成神經(jīng)網(wǎng)絡(luò)與生成對(duì)抗網(wǎng)絡(luò)等,為建筑和結(jié)構(gòu)智能化設(shè)計(jì)提供了新的可能性[18-27]。
生成對(duì)抗網(wǎng)絡(luò)(Generative Adversarial Networks, GAN)可通過對(duì)既有設(shè)計(jì)圖像和文本的學(xué)習(xí),掌握設(shè)計(jì)數(shù)據(jù)中潛在的設(shè)計(jì)規(guī)律,實(shí)現(xiàn)新設(shè)計(jì)圖像的生成。剪力墻結(jié)構(gòu)是一種廣泛應(yīng)用的典型住宅建筑結(jié)構(gòu)形式[28],具備設(shè)計(jì)需求較大、結(jié)構(gòu)相對(duì)規(guī)則、圖紙表達(dá)有效等特點(diǎn),可作為智能設(shè)計(jì)的重要研究對(duì)象[12-14, 25-26]。筆者基于生成對(duì)抗網(wǎng)絡(luò)開展了從數(shù)據(jù)驅(qū)動(dòng)到物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)的剪力墻結(jié)構(gòu)智能化生成式設(shè)計(jì)方法研究,基于生成對(duì)抗網(wǎng)絡(luò)的剪力墻結(jié)構(gòu)智能化設(shè)計(jì)方法包括StructGAN[12]、txtimg2img[13](后續(xù)改稱為StructGAN-TXT)、StructGAN-PHY[14]。但是,目前對(duì)于數(shù)據(jù)驅(qū)動(dòng)及物理增強(qiáng)智能設(shè)計(jì)方法的對(duì)比研究相對(duì)缺乏,不同方法中,數(shù)據(jù)特征分析與數(shù)據(jù)集創(chuàng)建、智能設(shè)計(jì)算法開發(fā)、測(cè)試評(píng)估方法以及典型案例分析的研究相對(duì)缺乏且不明確。筆者針對(duì)提出的3種智能化結(jié)構(gòu)設(shè)計(jì)方法開展詳細(xì)的對(duì)比分析研究,明確不同方法的性能及適用范圍,并對(duì)相關(guān)評(píng)價(jià)指標(biāo)進(jìn)行合理性分析,為未來的智能化設(shè)計(jì)方法研究提供對(duì)比模型和評(píng)價(jià)方法及指標(biāo)的參考。
1 剪力墻結(jié)構(gòu)智能化設(shè)計(jì)方法
基于生成對(duì)抗網(wǎng)絡(luò)的剪力墻結(jié)構(gòu)智能化設(shè)計(jì)方法如圖1所示,主要包括語義化模塊、設(shè)計(jì)模塊、結(jié)構(gòu)建模模塊[12-14]。目前,該方法主要應(yīng)用于基于像素圖表達(dá)的建筑和結(jié)構(gòu)設(shè)計(jì)中,對(duì)于可采用像素圖表達(dá)的不同結(jié)構(gòu)類型的智能設(shè)計(jì),具備一定的可拓展性和通用性。
1)語義化模塊是指將復(fù)雜的原始CAD建筑圖,通過對(duì)關(guān)鍵元素的提取,并采用顏色填充生成為包含建筑設(shè)計(jì)初步特征的語義化圖紙。語義化過程可有效排除復(fù)雜信息的干擾,有效提升深度生成網(wǎng)絡(luò)的設(shè)計(jì)效果,更詳細(xì)的語義化方法詳見既有研究[12]。該過程可通過人工提取數(shù)據(jù)特征,精度高但效率較低,也可進(jìn)一步研發(fā)基于BIM(Building Information Modeling)和CAD的數(shù)字化自動(dòng)提取方法,效率高且精度可控。
2)設(shè)計(jì)模塊是指采用深度生成網(wǎng)絡(luò),根據(jù)輸入的語義化建筑圖智能地生成語義化結(jié)構(gòu)設(shè)計(jì)圖。深度生成網(wǎng)絡(luò)是生成對(duì)抗網(wǎng)絡(luò)模型的生成器,在采用生成網(wǎng)絡(luò)進(jìn)行設(shè)計(jì)前,需要采用構(gòu)建的建筑-結(jié)構(gòu)圖紙數(shù)據(jù)集對(duì)生成對(duì)抗網(wǎng)絡(luò)進(jìn)行訓(xùn)練,使生成網(wǎng)絡(luò)基本掌握剪力墻結(jié)構(gòu)設(shè)計(jì)能力。
3)建模模塊是指基于生成的結(jié)構(gòu)設(shè)計(jì)像素圖建立對(duì)應(yīng)的結(jié)構(gòu)計(jì)算模型。通過像素圖到矢量圖的自動(dòng)轉(zhuǎn)化算法提取剪力墻構(gòu)件的矢量坐標(biāo),并基于結(jié)構(gòu)計(jì)算軟件的應(yīng)用程序接口(Application Programming Interface,API)建立對(duì)應(yīng)的結(jié)構(gòu)計(jì)算模型,完成整體結(jié)構(gòu)的方案設(shè)計(jì)。
用于訓(xùn)練和測(cè)試的數(shù)據(jù)集構(gòu)建方法如圖2所示,即將CAD圖紙數(shù)據(jù)進(jìn)行語義化處理。在該數(shù)據(jù)集中,7度設(shè)防且結(jié)構(gòu)高度50 m以下(簡稱:7d-H1)數(shù)據(jù)63組、7度設(shè)防且結(jié)構(gòu)高度50~150 m(7d-H2)數(shù)據(jù)80組、8度設(shè)防(8d)數(shù)據(jù)81組[12]。
2 數(shù)據(jù)驅(qū)動(dòng)和物理增強(qiáng)設(shè)計(jì)方法對(duì)比
StructGAN采用的核心算法為圖像合成生成對(duì)抗網(wǎng)絡(luò)[12],StructGAN-TXT采用的核心算法為文本-圖像特征融合生成對(duì)抗網(wǎng)絡(luò)[13],二者均為數(shù)據(jù)驅(qū)動(dòng)方法;StructGAN-PHY采用的核心算法為物理增強(qiáng)生成對(duì)抗網(wǎng)絡(luò)[14],是一種物理增強(qiáng)方法。3種方法的核心差異在于設(shè)計(jì)模塊,分別為3種不同的條件生成對(duì)抗網(wǎng)絡(luò)模型,如圖3所示。圖3(a)所示為StructGAN模型,由圖像生成器與圖像真假判別器構(gòu)成,基于建筑圖輸入生成結(jié)構(gòu)設(shè)計(jì)圖紙。其對(duì)應(yīng)數(shù)據(jù)集的輸入為建筑設(shè)計(jì)圖(圖2(b)),標(biāo)簽為結(jié)構(gòu)設(shè)計(jì)圖(圖2(c))。訓(xùn)練過程中,圖像生成器不斷生成結(jié)構(gòu)設(shè)計(jì)圖,判別器不斷判斷生成圖的真假,判別器判斷正確則提升生成器,判斷錯(cuò)誤則提升判別器;生成器與判別器對(duì)抗訓(xùn)練不斷提升,直到二者性能均難以提升達(dá)到納什均衡,完成訓(xùn)練獲得結(jié)構(gòu)設(shè)計(jì)能力,詳細(xì)訓(xùn)練過程可見相關(guān)研究源代碼[12,29]。訓(xùn)練時(shí),7d-H1、7d-H2、8d數(shù)據(jù)集分別訓(xùn)練3個(gè)StructGAN設(shè)計(jì)模型[12]。
圖3(b)所示為StructGAN-TXT模型,基于建筑圖和對(duì)應(yīng)結(jié)構(gòu)設(shè)計(jì)條件同時(shí)輸入,生成滿足建筑圖和設(shè)計(jì)條件約束的結(jié)構(gòu)設(shè)計(jì)圖紙。相比于StructGAN,StructGAN-TXT增加了文本輸入,其對(duì)應(yīng)數(shù)據(jù)集的輸入為建筑設(shè)計(jì)圖(圖2(b))和抗震設(shè)防烈度及結(jié)構(gòu)高度文本,標(biāo)簽為結(jié)構(gòu)設(shè)計(jì)圖(圖2(c))。因此,其圖像生成器需要同時(shí)對(duì)圖像和文本特征進(jìn)行提取和融合,進(jìn)而基于文本-圖像融合特征,生成結(jié)構(gòu)設(shè)計(jì)圖紙;判別器同樣需要對(duì)真假圖像真假文本進(jìn)行判別;生成器和判別器對(duì)抗訓(xùn)練至性能穩(wěn)定,進(jìn)而具備結(jié)構(gòu)設(shè)計(jì)生成能力。訓(xùn)練時(shí),7d-H1、7d-H2、8d所有圖像數(shù)據(jù)混合,并對(duì)每個(gè)圖像給予對(duì)應(yīng)文本,創(chuàng)建數(shù)據(jù)集訓(xùn)練一個(gè)StructGAN-TXT設(shè)計(jì)模型[13]。
圖3(b)所示為StructGAN-PHY模型,由圖像生成器、判別器、物理評(píng)估器構(gòu)成,基于建筑圖輸入生成器,生成滿足建筑圖和相應(yīng)設(shè)計(jì)條件的結(jié)構(gòu)設(shè)計(jì)圖紙。其對(duì)應(yīng)數(shù)據(jù)集的輸入為建筑設(shè)計(jì)圖(圖2(b)),標(biāo)簽為結(jié)構(gòu)設(shè)計(jì)圖(圖2(c)),輸入的抗震設(shè)防烈度和結(jié)構(gòu)高度并不會(huì)輸入生成器中,而是輔助物理性能評(píng)估器進(jìn)行力學(xué)性能評(píng)估。與StructGAN和StructGAN-TXT相比,其網(wǎng)絡(luò)架構(gòu)中新增了物理性能評(píng)估器。物理性能評(píng)估器為基于深度判別神經(jīng)網(wǎng)絡(luò)的代理模型,通過對(duì)結(jié)構(gòu)平面圖紙及其對(duì)應(yīng)力學(xué)性能數(shù)據(jù)集的訓(xùn)練,物理性能掌握對(duì)結(jié)構(gòu)設(shè)計(jì)對(duì)應(yīng)力學(xué)性能的預(yù)測(cè),進(jìn)而有效指導(dǎo)生成器的結(jié)構(gòu)設(shè)計(jì)。訓(xùn)練時(shí),7d-H1、7d-H2、8d所有圖像數(shù)據(jù)混合,創(chuàng)建多個(gè)設(shè)計(jì)條件組合,設(shè)防烈度包括6度(0.05g)、7度(0.10g)、7度(0.15g)、8度(0.20g)、8度(0.30g),結(jié)構(gòu)高度包括“<40 m”“40~60 m”“60~80 m”“80~100 m”“>100 m”,采用物理增強(qiáng)提升StructGAN-PHY設(shè)計(jì)性能,得到不同設(shè)計(jì)條件對(duì)應(yīng)的智能設(shè)計(jì)模型。相比而言,當(dāng)StructGAN-TXT和StructGAN沒有特定設(shè)計(jì)條件下的數(shù)據(jù)時(shí)(例如6度50 m),便無法開展對(duì)應(yīng)的設(shè)計(jì)模型訓(xùn)練[14]。所以StructGAN-PHY是在數(shù)據(jù)驅(qū)動(dòng)學(xué)習(xí)的基礎(chǔ)上,由物理性能評(píng)估器進(jìn)行增強(qiáng)的智能設(shè)計(jì)方法,能有效地保證設(shè)計(jì)結(jié)果盡可能滿足設(shè)計(jì)規(guī)范要求,避免由于數(shù)據(jù)數(shù)量不足或質(zhì)量不佳導(dǎo)致的設(shè)計(jì)效果降低。
整體而言,StructGAN與StructGAN-TXT為數(shù)據(jù)驅(qū)動(dòng)的方法,其設(shè)計(jì)能力與訓(xùn)練數(shù)據(jù)的數(shù)量和質(zhì)量密切相關(guān),而StructGAN-PHY則是物理增強(qiáng)的方法,能通過GAN對(duì)結(jié)構(gòu)力學(xué)性能學(xué)習(xí),有效克服數(shù)據(jù)數(shù)量或質(zhì)量的影響,并保證設(shè)計(jì)結(jié)果力學(xué)性能的穩(wěn)定性。
3 典型案例分析與對(duì)比
采用4個(gè)實(shí)際工程案例進(jìn)行設(shè)計(jì),并開展性能分析、對(duì)比。需指出,由于設(shè)計(jì)資料的知識(shí)產(chǎn)權(quán)限制,將隱去工程項(xiàng)目的真實(shí)信息,以及真實(shí)的圖紙和對(duì)應(yīng)的結(jié)構(gòu)設(shè)計(jì)模型,僅展示研究提出方法的設(shè)計(jì)結(jié)果。典型案例命名為“案例1-7d83m”,其中,“1”代表案例編號(hào)、“7d”代表設(shè)防烈度為7度(0.1g)、“83m”代表結(jié)構(gòu)高度83 m。
3.1 評(píng)價(jià)方法
圖4(a)為基于圖像相似性的評(píng)價(jià)方法,將生成設(shè)計(jì)的剪力墻布置與工程師的設(shè)計(jì)逐像素對(duì)比以及逐構(gòu)件輪廓對(duì)比,通過式(1)~式(4)所示的交并比計(jì)算相似性,以SCV作為綜合相似性指標(biāo)[12-14]。圖4(b)為基于結(jié)構(gòu)力學(xué)模型進(jìn)行力學(xué)性能分析的評(píng)價(jià),其中像素圖剪力墻構(gòu)件矢量提取方法見Lu等[14]的研究。
式中:SCV為生成設(shè)計(jì)與目標(biāo)設(shè)計(jì)的綜合相似性指標(biāo),該指標(biāo)越大,代表相似性越高;RSWratio為剪力墻占比率,即圖像中剪力墻像素面積(Aswall)占總墻體面積(Aswall+Ainwall)的比例;ηSWratio為剪力墻占比率一致性指標(biāo),該指標(biāo)越大,表明生成設(shè)計(jì)剪力墻率(R_generate^SWratio)與目標(biāo)設(shè)計(jì)剪力墻率(R_target^SWratio)越一致;SSIoU為剪力墻輪廓交并比一致性指標(biāo),即生成與目標(biāo)設(shè)計(jì)的剪力墻輪廓交集面積(Ainter)與并集面積(Aunion)的比例;SWIoU為剪力墻像素一致性指標(biāo),k是總的像素類別(類別0是背景,類別1是剪力墻,類別2是填充墻,類別3是窗戶,類別4是戶外門洞);pii是生成正確的像素點(diǎn)數(shù)量,pij和pji則是生成錯(cuò)誤的像素點(diǎn)數(shù)量。
需指出,平面設(shè)計(jì)結(jié)果的評(píng)價(jià)采用圖4(a)所示的基于計(jì)算機(jī)視覺的評(píng)價(jià)方法,快速評(píng)價(jià)生成設(shè)計(jì)結(jié)果與工程師設(shè)計(jì)結(jié)果的一致性。整體結(jié)構(gòu)設(shè)計(jì)評(píng)價(jià)采用圖4(b)所示的基于結(jié)構(gòu)力學(xué)性能的評(píng)價(jià)方法。整體結(jié)構(gòu)模型構(gòu)建就以工程師設(shè)計(jì)模型的參數(shù)為基準(zhǔn),剪力墻布置由像素到矢量的自動(dòng)化程序轉(zhuǎn)化,剪力墻厚度和材料參數(shù)以及沿結(jié)構(gòu)高度的變化與工程師設(shè)計(jì)一致,梁構(gòu)件設(shè)計(jì)則根據(jù)墻體布置進(jìn)行適應(yīng)性調(diào)整?;谏鲜鲈O(shè)計(jì)和評(píng)價(jià)方法,保證3種方法對(duì)比條件盡可能一致,更加有效體現(xiàn)差異。
3.2 案例1-7d83m對(duì)比分析
1) 案例1-7d83m基本信息
結(jié)構(gòu)總高度83 m,28層,平面尺寸為:49.4 m×18.7 m。抗震設(shè)防烈度7度(0.1g),特征周期0.55 s。工程師完成的建筑和結(jié)構(gòu)平面設(shè)計(jì)CAD及其對(duì)應(yīng)的語義化圖紙如圖5所示。
2)案例1-7d83m平面設(shè)計(jì)結(jié)果對(duì)比
在該案例的平面設(shè)計(jì)中,StructGAN、-TXT、-PHY三種方法的設(shè)計(jì)結(jié)果以及對(duì)應(yīng)的工程師設(shè)計(jì)結(jié)果如圖6所示。同時(shí),基于SCV指標(biāo)評(píng)價(jià)的智能設(shè)計(jì)與工程師設(shè)計(jì)的相似性結(jié)果也如圖6所示。在該案例中,StructGAN和StructGAN-PHY兩種方法的設(shè)計(jì)結(jié)果較好,與工程師設(shè)計(jì)相似性高(即量化SCV指標(biāo)較高,均為0.47),而StructGAN-TXT的設(shè)計(jì)則差別較大(量化SCV指標(biāo)僅為0.39)。具體分析,StructGAN和StructGAN-PHY生成設(shè)計(jì)的ηSWratio指標(biāo)接近1,意味著在該案例中二者生成的剪力墻總像素?cái)?shù)量與工程師設(shè)計(jì)的剪力墻像素量基本一致,而StructGAN-TXT生成設(shè)計(jì)的ηSWratio指標(biāo)則偏低,代表剪力墻總量與工程差異較大;3個(gè)案例的SWIoU和SSIoU指標(biāo)則較為接近(均接近0.5),表明剪力墻布置位置與工程師布置位置差異較小。
3)案例1-7d83m整體結(jié)構(gòu)性能對(duì)比
整體結(jié)構(gòu)案例分析的模型如圖7所示,不同設(shè)計(jì)結(jié)果的剪力墻布置不同,但建模方法與工程師設(shè)計(jì)均一致。
開展結(jié)構(gòu)動(dòng)力特性分析以及基于振型分解反應(yīng)譜法的結(jié)構(gòu)力學(xué)響應(yīng)分析,對(duì)應(yīng)的結(jié)構(gòu)動(dòng)力特性和層間位移角的對(duì)比結(jié)果如表1和圖8所示。StructGAN-TXT的設(shè)計(jì)結(jié)果與工程師設(shè)計(jì)差異較大,生成的剪力墻數(shù)量偏多,結(jié)構(gòu)剛度偏大,導(dǎo)致動(dòng)力特性差異20%左右,層間變形的最大差異達(dá)到了30%;而StructGAN和StructGAN-PHY的設(shè)計(jì)與工程師差異則較小,最大性能差異約10%。
3.3 案例2-8d96m、案例3-7d77m、案例4-7d41m對(duì)比分析
案例1-7d83m的分析表明,StructGAN-PHY與StructGAN的設(shè)計(jì)效果較好,為進(jìn)一步研究不同智能設(shè)計(jì)方法的通用性和泛化性,進(jìn)一步開展了案例2-8d96m、案例3-7d77m、案例4-7d41m的對(duì)比分析。案例2-8d96m的抗震設(shè)防烈度8度(0.2g)、結(jié)構(gòu)總高度96 m(30層);案例3-7d77m的抗震設(shè)防烈度7度(0.1g)、結(jié)構(gòu)總高度77 m(26層);案例4-7d41m的抗震設(shè)防烈度7度(0.1g)、結(jié)構(gòu)總高度41 m(14層)。
1)案例2-8d96m平面與整體設(shè)計(jì)結(jié)果對(duì)比
在該案例的平面設(shè)計(jì)中,StructGAN、-TXT、-PHY三種方法的設(shè)計(jì)結(jié)果以及對(duì)應(yīng)的工程師設(shè)計(jì)結(jié)果如圖9所示。在該案例中,直觀視覺判斷3種方法的設(shè)計(jì)均與工程師設(shè)計(jì)非常接近,且量化的SCV指標(biāo)較高(>0.5)。其中,ηSWratio均較接近1,意味著生成設(shè)計(jì)的總剪力墻數(shù)量與工程設(shè)計(jì)的基本一致;且SWIoU和SSIoU均大于0.5,意味著剪力墻布置的位置一致性較高。
進(jìn)一步開展整體結(jié)構(gòu)設(shè)計(jì)的對(duì)比分析,結(jié)果如表2所示。3種設(shè)計(jì)與工程師設(shè)計(jì)的結(jié)果差異均較小,動(dòng)力特性差異在3%左右,最大層間變形在15%左右,且整體變形模式非常接近。其中,StructGAN-TXT的差異最小(動(dòng)力特性差異1.00%,最大層間變形差異3.79%)。
2)案例3-7d77m平面設(shè)計(jì)結(jié)果對(duì)比
該案例的3種方法設(shè)計(jì)結(jié)果以及對(duì)應(yīng)SCV評(píng)價(jià)結(jié)果如圖10所示??梢钥吹?,StructGAN和StructGAN-PHY兩種方法的設(shè)計(jì)結(jié)果較好,與工程師設(shè)計(jì)相似性高,而StructGAN-TXT的設(shè)計(jì)則差別較大。主要原因在于StructGAN-TXT設(shè)計(jì)的剪力墻墻體過多,ηSWratio指標(biāo)偏小,僅0.65。
進(jìn)一步分析整體結(jié)構(gòu)案例對(duì)應(yīng)的結(jié)構(gòu)動(dòng)力特性和層間位移角的對(duì)比結(jié)果如表3所示。與平面設(shè)計(jì)結(jié)果的對(duì)比類似,StructGAN-TXT的設(shè)計(jì)結(jié)果與工程師設(shè)計(jì)結(jié)果差異較大,剪力墻布置偏多,導(dǎo)致性能差異20%左右,而StructGAN和StructGAN-PHY的設(shè)計(jì)與工程師差異則為10%左右。
3)案例4-7d41m平面設(shè)計(jì)結(jié)果對(duì)比
該案例的3種方法設(shè)計(jì)結(jié)果以及對(duì)應(yīng)SCV評(píng)價(jià)結(jié)果如圖11所示??梢钥闯?,StructGAN-TXT和StructGAN-PHY兩種方法的設(shè)計(jì)結(jié)果較好,與工程師設(shè)計(jì)相似性高,而StructGAN的設(shè)計(jì)則差別較大。主要原因在于StructGAN設(shè)計(jì)的剪力墻墻體偏少,ηSWratio、SWIoU、SSIoU指標(biāo)均偏小。
整體結(jié)構(gòu)案例分析對(duì)應(yīng)的結(jié)構(gòu)動(dòng)力特性和層間位移角的對(duì)比結(jié)果如表4所示。與平面設(shè)計(jì)結(jié)果的對(duì)比類似,StructGAN的設(shè)計(jì)結(jié)果與工程師差異較大,其剪力墻布置偏少,動(dòng)力特性差異20%左右,層間變形的最大差異達(dá)到了40%;而StructGAN-TXT和StructGAN-PHY的設(shè)計(jì)與工程師差異同樣偏大,動(dòng)力特性差異10%左右,層間變形最大差異達(dá)到了20%左右。主要原因是7度41 m設(shè)計(jì)條件下對(duì)應(yīng)的剪力墻需求較少,剪力墻布置的較小差異都容易引起較大的結(jié)構(gòu)整體特性不同。
3.4 基于計(jì)算機(jī)視覺與力學(xué)分析的評(píng)價(jià)分析
通過對(duì)4個(gè)典型案例的平面設(shè)計(jì)結(jié)果和整體結(jié)構(gòu)分析和對(duì)比,分析結(jié)果匯總至表5中??梢钥闯觯?)在不同案例中,StructGAN-PHY是3種方法中最有效的,較少受到數(shù)據(jù)質(zhì)量的限制,能更準(zhǔn)確地匹配對(duì)應(yīng)設(shè)計(jì)需求,且設(shè)計(jì)結(jié)果與工程師設(shè)計(jì)最接近;2)StructGAN和StructGAN-TXT的設(shè)計(jì)能力均受到訓(xùn)練數(shù)據(jù)的制約,僅有部分案例效果較好,意味著如果待設(shè)計(jì)的建筑與訓(xùn)練數(shù)據(jù)的特征域較為接近,則對(duì)應(yīng)的設(shè)計(jì)結(jié)果較良好,反之則設(shè)計(jì)質(zhì)量不佳;3)結(jié)構(gòu)抗震性能需求較高的案例,設(shè)計(jì)結(jié)果通常較好,原因在于所需布置的剪力墻較多,設(shè)計(jì)變化較少;對(duì)于抗震性能需求較少的案例,對(duì)應(yīng)布置的剪力墻較少,設(shè)計(jì)變化則較豐富,人工智能不能保證找到最合適的結(jié)果。既有研究認(rèn)為當(dāng)生成設(shè)計(jì)與目標(biāo)設(shè)計(jì)的交并比大于0.5則相對(duì)較優(yōu)秀[12-14],但目前尚未明確基于計(jì)算機(jī)視覺的平面設(shè)計(jì)評(píng)價(jià)指標(biāo)與結(jié)構(gòu)力學(xué)性能評(píng)價(jià)指標(biāo)之間的關(guān)系。將所有案例分析結(jié)果繪于圖12中,可以看出:1)當(dāng)SCV大于0.5時(shí),基本可以保證智能設(shè)計(jì)的結(jié)構(gòu)力學(xué)性能與工程師設(shè)計(jì)的結(jié)構(gòu)動(dòng)力特性差異在5%以內(nèi),且層間位移角在10%以內(nèi);2)對(duì)于地震作用比較大的情況,例如8度(0.2g)設(shè)防、96 m案例中,即使SCV相差僅0.02且動(dòng)力特性差異在5%以內(nèi),層間位移角的差異仍舊會(huì)超過10%,主要原因在于較大地震作用會(huì)導(dǎo)致更大的結(jié)構(gòu)變形,相對(duì)差異也變得更大。
因此,可以認(rèn)為基于計(jì)算機(jī)視覺的SCV指標(biāo)與結(jié)構(gòu)力學(xué)性能評(píng)價(jià)的指標(biāo)具有高度的正相關(guān)性,即SCV越高則結(jié)構(gòu)力學(xué)性能越優(yōu),更接近專家的優(yōu)化設(shè)計(jì)。SCV=0.5可作為基于視覺評(píng)價(jià)的合理性閾值,其對(duì)應(yīng)的結(jié)構(gòu)動(dòng)力特性差異約為5%,層間位移角差異約為10%。
4 結(jié)論
針對(duì)建筑結(jié)構(gòu)智能化方案設(shè)計(jì)方法進(jìn)行了對(duì)比分析研究。從數(shù)據(jù)驅(qū)動(dòng)到物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)方法,對(duì)比了核心算法構(gòu)建和設(shè)計(jì)性能,并通過典型案例對(duì)比了相同設(shè)計(jì)條件下不同算法的實(shí)際表現(xiàn),主要結(jié)論如下:
1)數(shù)據(jù)驅(qū)動(dòng)與物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)是建筑結(jié)構(gòu)智能化設(shè)計(jì)的兩個(gè)階段,數(shù)據(jù)驅(qū)動(dòng)方法搭建了智能化設(shè)計(jì)的數(shù)據(jù)、算法、評(píng)價(jià)和應(yīng)用方法的基礎(chǔ),物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)方法則進(jìn)一步提升了算法性能。
2)數(shù)據(jù)驅(qū)動(dòng)方法(StructGAN和StructGAN-TXT)通常受限于特定設(shè)計(jì)條件下的數(shù)據(jù)質(zhì)量與數(shù)量,而物理增強(qiáng)數(shù)據(jù)驅(qū)動(dòng)(StructGAN-PHY)則可以有效降低對(duì)結(jié)構(gòu)設(shè)計(jì)數(shù)據(jù)的依賴性。
3)對(duì)于抗震設(shè)防需求較低的設(shè)計(jì),剪力墻布置的需求較少,其布置的位置更加靈活,導(dǎo)致智能化設(shè)計(jì)與工程師設(shè)計(jì)存在一定差異,未來將進(jìn)一步提升相關(guān)設(shè)計(jì)條件下的智能化設(shè)計(jì)能力。
4)確定了基于計(jì)算機(jī)視覺與力學(xué)分析評(píng)價(jià)方法的正相關(guān)性,且SCV=0.5可作為基于視覺評(píng)價(jià)的合理性閾值,可供未來智能化設(shè)計(jì)評(píng)價(jià)方法使用。
5)主要針對(duì)剪力墻結(jié)構(gòu)開展了智能設(shè)計(jì)相關(guān)研究,框架、框架-剪力墻等多種結(jié)構(gòu)類型的智能設(shè)計(jì)方法研究有待進(jìn)一步從結(jié)構(gòu)表達(dá)、智能算法和評(píng)估方法等方面開展。
采用的4個(gè)典型案例分析已在GitHub中開源(https://github.com/wenjie-liao/StructGAN-PHY/blob/main/StructGAN-TXT-PHY.zip)。
參考文獻(xiàn)
[1]? 鮑躍全, 李惠. 人工智能時(shí)代的土木工程[J]. 土木工程學(xué)報(bào), 2019, 52(5): 1-11.
BAO Y Q, LI H. Artificial intelligence for civil engineering [J]. China Civil Engineering Journal, 2019, 52(5): 1-11. (in Chinese)
[2]? 勾紅葉, 楊彪, 華輝, 等. 橋梁信息化及智能橋梁2019年度研究進(jìn)展[J]. 土木與環(huán)境工程學(xué)報(bào)(中英文), 2020, 42(5): 14-27.
GOU H Y, YANG B, HUA H, et al. State-of-the-art review of bridge informatization and intelligent bridge in 2019 [J]. Journal of Civil and Environmental Engineering, 2020, 42(5): 14-27. (in Chinese)
[3]? 趙天祺, 勾紅葉, 陳萱穎, 等. 橋梁信息化及智能橋梁2020年度研究進(jìn)展[J]. 土木與環(huán)境工程學(xué)報(bào)(中英文), 2021, 43(Sup1): 268-279.
ZHAO T Q, GOU H Y, CHEN X Y, et al. State-of-the-art review of bridge informatization and intelligent bridge in 2020 [J]. Journal of Civil and Environmental Engineering, 2021, 43(Sup1): 268-279. (in Chinese)
[4]? 劉紅波, 張帆, 陳志華, 等.人工智能在土木工程領(lǐng)域的應(yīng)用研究現(xiàn)狀及展望[J]. 土木與環(huán)境工程學(xué)報(bào)(中英文), 2024, 46(1): 14-32.
LIU H B, ZHANG F, CHEN Z H, et al. Applied research status and prospects of artificial intelligence in civil engineering field [J]. Journal of Civil and Environmental Engineering. 2024, 46(1): 14-32.(in Chinese)
[5]? BROWN N C, MUELLER C T. Design for structural and energy performance of long span buildings using geometric multi-objective optimization [J]. Energy and Buildings, 2016, 127: 748-761.
[6]? ZHANG Y, MUELLER C. Shear wall layout optimization for conceptual design of tall buildings [J]. Engineering Structures, 2017, 140: 225-240.
[7]? TAFRAOUT S, BOURAHLA N, BOURAHLA Y, et al. Automatic structural design of RC wall-slab buildings using a genetic algorithm with application in BIM environment [J]. Automation in Construction, 2019, 106: 102901.
[8]? 何政, 來瀟. 參數(shù)化結(jié)構(gòu)設(shè)計(jì)基本原理、方法及應(yīng)用[M]. 北京: 中國建筑工業(yè)出版社, 2019.
HE Z, LAI X. Parametric structural design: Theory, approach and applications [M]. Beijing: China Architecture & Building Press, 2019. (in Chinese)
[9]? 杜修力, 劉占省, 趙研. 智能建造概論[M]. 北京: 中國建筑工業(yè)出版社, 2021.
DU X L, LIU Z S, ZHAO Y. Introduction to intelligent construction [M]. Beijing: China Architecture & Building Press, 2021. (in Chinese)
[10]? 周緒紅, 劉界鵬, 馮亮, 等. 建筑智能建造技術(shù)初探及其應(yīng)用 [M]. 北京: 中國建筑工業(yè)出版社, 2021.
ZHOU X H, LIU J P, FENG L, et al. A preliminary exploration and application of building intelligent construction technology [M]. Beijing: China Architecture & Building Press, 2021. (in Chinese)
[11]? SUN H, BURTON H V, HUANG H L. Machine learning applications for building structural design and performance assessment: State-of-the-art review [J]. Journal of Building Engineering, 2021, 33: 101816.
[12]? LIAO W J, LU X Z, HUANG Y L, et al. Automated structural design of shear wall residential buildings using generative adversarial networks [J]. Automation in Construction, 2021, 132: 103931.
[13]? 廖文杰, 黃羽立, 鄭哲, 等. 基于生成對(duì)抗網(wǎng)絡(luò)融合文本圖像數(shù)據(jù)的剪力墻結(jié)構(gòu)生成式設(shè)計(jì)方法[C]//第30屆全國結(jié)構(gòu)工程學(xué)術(shù)會(huì)議論文集(第Ⅲ冊(cè)), 2021: 381-387.
LIAO W J, HUANG Y L, ZHENG Z, et al. Building structural generative design using generative adversarial networks-based fused-image-text to image translation method [C]//Proceedings of the 30th National Conference on Structural Engineering (NO.Ⅲ), 2021: 381-387. (in Chinese)
[14]? LU X Z, LIAO W J, ZHANG Y, et al. Intelligent structural design of shear wall residence using physics-enhanced generative adversarial networks [J]. Earthquake Engineering & Structural Dynamics, 2022, 51(7): 1657-1676.
[15]? HUANG X, XIE Y M. Evolutionary topology optimization of continuum structures [M]. Chichester, UK: John Wiley & Sons, Ltd, 2010.
[16]? SHEA K, AISH R, GOURTOVAIA M. Towards integrated performance-driven generative design tools [J]. Automation in Construction, 2005, 14(2): 253-264.
[17]? 趙憲忠, SHEA Kristina. 空間桿系結(jié)構(gòu)的智能生成與設(shè)計(jì)[J]. 建筑結(jié)構(gòu)學(xué)報(bào), 2010, 31(9): 63-69.
ZHAO X Z, KRISTINA S. Intelligent generation and design of spatial truss structures [J]. Journal of Building Structures, 2010, 31(9): 63-69. (in Chinese)
[18]? OH S, JUNG Y, KIM S, et al. Deep generative design: Integration of topology optimization and generative models [J]. Journal of Mechanical Design, 2019, 141(11): 111405.
[19]? SHU D L, CUNNINGHAM J, STUMP G, et al. 3D design using generative adversarial networks and physics-based validation [J]. Journal of Mechanical Design, 2020, 142(7): 071701.
[20]? HUANG W X, ZHENG H. Architectural drawings recognition and generation through machine learning [C]//Proceedings of the 38th Annual Conference of the Association for Computer Aided Design in Architecture (ACADIA), ACADIA Proceedings. October 18-20, 2018. Mexico City, Mexico. ACADIA, 2018.
[21]? CHAILLOU S. ArchiGAN: A generative stack for apartment building design [EB/OL]. [2019-7-17]. https://developer.nvidia.com/blog/archigan-generative-stack-apartment-building-design/.
[22]? NAUATA N, CHANG K H, CHENG C Y, et al. House-GAN: Relational generative adversarial networks for graph-constrained house layout generation [M]// Computer Vision, ECCV 2020. Cham: Springer, 2020: 162-177.
[23]? NAUATA N, HOSSEINI S, CHANG K H, et al. House-GAN++: Generative adversarial layout refinement networks [EB/OL]. 2021: arXiv: 2103.02574. https://arxiv.org/abs/2103.02574.
[24]? CHANG K H, CHENG C Y. Learning to simulate and design for structural engineering [EB/OL]. 2020: arXiv: 2003.09103. https://arxiv.org/abs/2003.09103.
[25]? PIZARRO P N, MASSONE L M, ROJAS F R, et al. Use of convolutional networks in the conceptual structural design of shear wall buildings layout [J]. Engineering Structures, 2021, 239: 112311.
[26]? PIZARRO P N, MASSONE L M. Structural design of reinforced concrete buildings based on deep neural networks [J]. Engineering Structures, 2021, 241: 112377.
[27]? QIAN W L, XU Y, LI H. A self-sparse generative adversarial network for autonomous early-stage design of architectural sketches [J]. Computer-Aided Civil and Infrastructure Engineering, 2022, 37(5): 612-628.
[28]? 田晴, 趙崇錦, 王忠凱, 等. 新疆某超高層住宅樓動(dòng)力彈塑性分析[J]. 土木建筑與環(huán)境工程, 2016, 38(Sup 2): 30-34.
TIAN Q, ZHAO C J, WANG Z K, et al. Elastic and plastic dynamic analysis of a super high-rise residential building in Xinjiang [J]. Journal of Civil, Architectural & Environmental Engineering, 2016, 38(Sup 2): 30-34. (in Chinese)
[29]? WANG T C, LIU M Y, ZHU J Y, et al. High-resolution image synthesis and semantic manipulation with conditional GANs [C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA. IEEE, 2018: 8798-8807.
(編輯? 胡英奎)
收稿日期:2022?04?25
基金項(xiàng)目:國家重點(diǎn)研發(fā)計(jì)劃(2019YFE0112800);騰訊基金會(huì)(科學(xué)探索獎(jiǎng));清華大學(xué)“水木學(xué)者”計(jì)劃項(xiàng)目(2022SM005)
作者簡介:廖文杰(1995- ),男,助理研究員,主要從事建筑結(jié)構(gòu)智能化設(shè)計(jì)研究,E-mail:liaowj17@tsinghua.org.cn。
通信作者:陸新征(通信作者),男,教授,博士生導(dǎo)師,E-mail:luxz@tsinghua.edu.cn。
Received: 2022?04?25
Foundation items: National Key R & D Program of China (No. 2019YFE0112800); Tencent Foundation (XPLORER PRIZE); Shuimu Tsinghua Scholar Program (2022SM005)
Author brief: LIAO Wenjie (1995- ), assistant researcher, main research interest: intelligent structural design of buildings, E-mail: liaowj17@tsinghua.org.cn.
corresponding author:LU Xinzheng (corresponding author), professor, doctorial supervisor, E-mail: luxz@tsinghua.edu.cn.