意象驱动的产品造型智能设计方法  

Image-driven product modeling intelligent design methodology

作　　者：苏建宁[1,2] 鱼宝银 李雄 张志鹏 郭睿 SU Jianning;YU Baoyin;LI Xiong;ZHANG Zhipeng;GUO Rui(School of Design Art,Lanzhou University of Technology,Lanzhou 730050;School of Mechanical&Electrical Engineering,Lanzhou University of Technology,Lanzhou 730050;School of Bailie Mechanical Engineering,Lanzhou City University,Lanzhou 730070)

机构地区：[1]兰州理工大学设计艺术学院,甘肃兰州730050 [2]兰州理工大学机电工程学院,甘肃兰州730050 [3]兰州城市学院培黎机械工程学院,甘肃兰州730070

出　　处：《机械设计》2024年第8期115-120,共6页Journal of Machine Design

基　　金：国家自然科学基金(52165033)。

摘　　要：为更准确地挖掘用户情感偏好,采用人工智能技术辅助设计满足需求的产品方案,提出一种意象驱动的产品造型智能设计方法。对产品在线评论数据进行筛选,应用词频-逆文件频率技术(Term Frequency-Inverse Document Frequence,TF-IDF)提取描述产品造型的代表性感性意象词汇,经聚类分析后获得目标意象,并结合语义差分量表获得样本意象评价值;采用GoogLeNet卷积神经网络构建意象回归模型,预测其余样本评分,获得意象评价数据;针对不同意象进行样本构成元素解构,依据重要度排序设置训练提示词;最后通过微调稳定扩散模型(Stable Diffusion XL,SDXL),构建低秩适应(Low-Rank Adaption,LoRA)意象造型生成模型。以吹风机为例实现目标感性意象的创新设计方案生成,验证了方法的可行性与合理性,可有效指导产品的创新设计。To excavate users’emotional preferences more accurately,an image-driven intelligent design method of product modeling was proposed using artificial intelligence technology to assist designing product solutions to meet the requirements.The product online review data was filtered.Term Frequency-Inverse Document Frequence(TF-IDF)technique was applied to extract the representative perceptual image vocabulary that described product modeling.The target image was obtained after cluster analysis.The sample image evaluation value was obtained combined with the semantic difference scale.GoogLeNet convolutional neural network was used to construct the image regression model.The scores of other samples were predicted to obtain the image evaluation data.The constituent elements of the sample were deconstructed.The prompt words were trained according to the importance ranking setting.Finally,the Low-Rank Adaption(LoRA)image modeling generation model was constructed by finetuning the Stable Diffusion XL(SDXL).Taking the hairdryer as an example,the innovation design scheme of the target perceptual image was generated.The feasibility and rationality of the method were verified,that could effectively guide the product innovation design.

关 键 词：产品造型 感性意象 智能设计 词频-逆文件频率技术 GoogLeNet LoRA 

分 类 号：TB472[一般工业技术—工业设计]