【作者】 王勇；

【导师】 宋方昊；

【作者基本信息】 山东大学 ， 工业设计， 2025， 博士

【摘要】 人工智能生成内容技术拓展了产品设计创意的来源,但也带来了方案的美感评价难题。同时,美感在工业设计中扮演着重要的角色,产品美感是用户情感需求中的重要内容,也是企业发展不可或缺的竞争利器。评价具有美感的产品通常要依赖产品美感评价方法。传统产品美感评价方法主要基于人的评分数据或图像像素数据来研究,而人产生的多种生理数据常常被忽略。生理数据为实现客观、快速且准确的产品美感评价提供了重要依据。若要利用生理数据为产品设计方案评价提供可靠的理论依据与实践支持,补充使用生理数据进行产品美感评价的可行性和准确率相关研究尤为迫切和必要。为了解决使用生理数据进行产品美感评价的可行性和准确率问题,并提供产品美感评价研究框架和方法参考,本文开展了产品美感评价中生理数据与美感质量的相关性和评价模型研究。首先通过采集实验参与者在产品美感评价过程中的眼动追踪(眼动)、肌电图(肌电)、皮肤电(皮电)和近红外脑成像(脑氧)数据,其次通过统计科学的数据分析方法分析生理数据与美感质量的相关性等问题,然后引入情感计算领域的多模态数据融合方法为构建产品美感评价模型提供更多的互补信息,最后通过人工智能领域的机器学习算法来构建产品美感评价模型。本文的主要研究工作与结论如下:(1)针对相关性分析和评价模型研究的数据需求问题,开展了产品美感评价中的生理数据采集实验。进行了实验材料搜集与图像处理、问卷调查和统计分析,以及实验材料选取等工作。招募了实验参与者,并规划了详细的实验流程。开展了5个生理数据采集实验,包括:眼动、肌电与脑氧数据采集实验,皮电和眼动数据采集实验,以及多模态生理数据采集实验。上述实验为后续研究提供了客观且充足的数据支撑。(2)针对使用生理数据进行产品美感评价的可行性问题,开展了生理数据与美感质量的相关性分析研究。研究了数据预处理和特征提取,诊断性分析和正态分布检验,以及相关性分析等问题。结果显示,与产品美感质量显著相关的生理特征包括:眼动数据中的注视时间、注视点数、注视次数和访问时间;肌电数据中的积分肌电值、均方根值、最大值、最小值和标准差;脑氧数据中的颞叶区(S3-D1)、顶叶区(S8-D5/S8-D8)、额叶区(S4-D3/S7-D8/S16-D10)、枕叶区(S18-D15)和前额叶区(S19-D11);皮电和眼动数据中的平均瞳孔直径、平均值和半衰期等9项特征;以及多模态生理数据中的平均瞳孔直径、积分肌电值和半衰期等22项特征。研究还发现,具有显著相关性的生理特征中,眼动、肌电与脑氧特征中不存在冗余特征,而皮电特征的最大值、最小值和基准值是冗余特征。结果表明,直接使用传统的生理特征数据求和或均值统计无法准确评价产品美感质量,但是生理特征数据可用于构建评价模型。相关性分析研究筛选出用于产品美感评价的关键生理特征,为后续的评价模型研究提供了理论依据。(3)针对使用生理数据进行产品美感评价的准确率问题,开展了基于生理数据的产品美感评价模型研究。进行了机器学习算法的选型与数据集划分,研究了产品美感评价模型训练和构建问题,探讨了评价模型的准确率验证等问题。结果显示,在多元逻辑回归算法下,基于眼动数据与基于脑氧数据构建的产品美感评价模型准确率高于基于肌电数据构建的模型。在多层感知机神经网络算法下,基于皮电和眼动双模态生理数据的产品美感评价模型准确率优于单模态数据构建的模型。此外,在全连接神经网络算法下,基于多模态生理数据的产品美感评价模型准确率高于单模态或双模态生理数据构建的模型。进一步比较发现,在多模态生理数据集下,使用全连接神经网络算法构建的产品美感评价模型准确率最高,而使用深度神经网络、K近邻和支持向量机算法构建的模型准确率相对较低。结果表明,不同数据类型和算法构建的产品美感评价模型准确率存在差异,采用多模态生理数据和全连接神经网络算法能够提供一种客观和高准确率的产品美感评价方法。(4)针对产品美感评价系统开发中设计方案缺乏问题,开展了产品美感评价系统设计实践。根据研究成果和用户需求分析结果设计了产品美感评价系统的整体框架、设计规范和功能模块等内容,并对其可用性和用户体验进行了评估。结果显示,系统可用性量表平均得分高于行业标准。同时,大部分测试参与者认为产品美感评价系统在可用性、易用性和美观性方面表现较好。结果表明,产品美感评价系统的设计方案具有较高的可用性,并能够提供较好的用户体验,为评价系统开发提供了界面设计方案。本文的研究完善了产品美感评价理论,扩展了多模态生理数据和机器学习算法在美感评价领域的应用边界,为产品设计方案决策和人工智能海量的生成式设计方案评价提供了科学的美感评价依据和技术支撑。更多还原

【Abstract】 Artificial intelligence generated content(AIGC)technology expands the source of product design ideas,but it also poses the challenge of evaluating the aesthetics of the solution.Meanwhile,aesthetics plays an important role in industrial design,and product aesthetics is an important content in the emotional needs of users,as well as an indispensable competitive tool for enterprise development.Evaluating aesthetically pleasing products usually relies on product aesthetics evaluation methods.Traditional product aesthetics evaluation methods are mainly based on human rating data or image pixel data,while a variety of physiological data generated by humans are often ignored.Physiological data provides an important basis for realizing objective,fast and accurate product aesthetics evaluation.If physiological data are to be utilized to provide a reliable theoretical basis and practical support for the evaluation of product design solutions,it is particularly urgent and necessary to supplement the research on the feasibility and accuracy of using physiological data for product aesthetics evaluation.In order to address the feasibility and accuracy of using physiological data for product aesthetics evaluation and to provide a research framework and methodological reference for product aesthetics evaluation,this paper carries out a research on the correlation and the evaluation model between physiological data and aesthetics quality in product aesthetics evaluation.Firstly,by collecting eye movement tracking(EMT),electromyography(EMG),electrodermal(EDA)and functional near-infrared spectroscopy(fNIRS)data from experimental participants in the process of product aesthetics evaluation,secondly,by analysing the correlation between the physiological data and the aesthetic quality through the data analysis method of statistical science,and other issues,and then by introducing the multimodal data fusion method in the field of affective computing to provide more complementary information,and finally construct the product aesthetics evaluation model through machine learning algorithms in the field of artificial intelligence.The main research work and conclusions of this paper are as follows:(1)Aiming at the data demand problem of correlation analysis and evaluation model research,physiological data collection experiments in product aesthetics evaluation were carried out.Experimental material collection and image processing,questionnaire survey and statistical analysis,and experimental material selection were carried out.Experimental participants were recruited and a detailed experimental process was planned.Five physiological data acquisition experiments were carried out,including:EMT,EMG and fNIRS data acquisition experiments,EDA and EMT data acquisition experiments,and multimodal physiological data acquisition experiments.The above experiments provided objective and sufficient data support for the subsequent studies.(2)To address the issue of the feasibility of using physiological data for product aesthetics evaluation,a correlation analysis study of physiological data and aesthetics quality was conducted.Data preprocessing and feature extraction,diagnostic analysis and normal distribution test,and correlation analysis were studied.The results showed that the physiological features significantly correlated with the aesthetics quality of the product included:gaze duration,number of gaze points,number of gaze times,and access time in the EMT data:integral EMG value,root mean square value,maximum value,minimum value,and standard deviation in the EMG data;and temporal lobe area(S3-D1),parietal lobe area(S8-D5/S8-D8),frontal lobe area(S4-D3/S7-D8/S16-D10),occipital area(S18-D15).and prefrontal area(S19-D11)in the fINR data;9 features such as mean pupil diameter,mean value and half-life in EDA and EMT data;and 22 features such as mean pupil diameter,integral EMG value,and half-life in multimodal physiological data.It was also found that among the physiological features with significant correlations,there were no redundant features in the EMT,EMG and fNIRS features,whereas the maximal.minimal and baseline values of the EDA features were redundant features.The results showed that the product aesthetics quality could not be accurately evaluated by directly using the traditional summation or mean statistics of physiological feature data,but the physiological feature data could be used to construct an evaluation model.The correlation analysis study screened out the key physiological features used for product aesthetics evaluation,which provided a theoretical basis for the subsequent evaluation model study.(3)Aiming at the accuracy of using physiological data for product aesthetics evaluation,research on product aesthetics evaluation models based on physiological data was carried out.The selection of machine learning algorithms and dataset division were carried out,the training and construction of product aesthetics evaluation models were studied,and issues such as the verification of the accuracy of the evaluation models were explored.The results show that under the multivariate logistic regression(MLR)algorithm,the accuracy of the product aesthetics evaluation model constructed on the basis of EMT data and fNIRS data was higher than the model constructed on the basis of EMG data.Under the multilayer perceptron(MLP)neural network algorithm,the accuracy of the product aesthetics evaluation model constructed based on EDA and EMT bimodal physiological data was better than the model constructed based on unimodal data.In addition,under the fully connected neural network(FCNN)algorithm,the accuracy of the product aesthetics evaluation model based on multimodal physiological data was higher than that of the model constructed from unimodal or bimodal physiological data.Further comparisons revealed that the product aesthetics evaluation model constructed using the FCNN algorithm had the highest accuracy rate under the multimodal physiological dataset,whereas the models constructed using the deep neural network(DNN),K-nearest neighbor(KNN),and support vector machine(SVM)algorithms had relatively low accuracy rates.The results show that there are differences in the accuracy of product aesthetics evaluation models constructed with different data types and algorithms,and that the use of multimodal physiological data and FCNN algorithms can provide an objective and highly accurate method for product aesthetics evaluation.(4)Aiming at the lack of design solutions in the development of the product aesthetics evaluation system(PAES),the design practice of the PAES was carried out.The overall framework,design specifications and functional modules of the PAES were designed based on the research results and the results of the user requirements analysis,and its usability and user experience were evaluated.The results showed that the average score of the system usability scale was higher than the industry standard.Meanwhile,most of the test participants thought that the PAES performed better in terms of usability,ease of use and aesthetics.The results indicate that the design solution of the PAES has high usability and can provide a better user experience,providing an interface design solution for evaluation system development.The research in this paper improves the theory of product aesthetics evaluation,extends the application boundaries of multimodal physiological data and machine learning algorithms in the field of aesthetics evaluation,and provides scientific aesthetics evaluation bases and technical support for the decision-making of product design solutions and the evaluation of massive generative design solutions by artificial intelligence.更多还原