User Preference-Based Method for Characterizing Automotive Wheel Hub Styles

Abstract

User preferences serve as a crucial dimension in describing the product style and its psychological experience. To assist designers in rapidly producing wheel designs that align with user preferences and enhance design efficiency by improving design transparency, this research establishes a user preference-based Automotive Wheel Hub Style Characterization Model (AWSCM). The model primarily consists of a style space, a design element space, and a matching relationship space. The style space is used to capture user preferences and product styles, the design element space for sample clustering and design element extraction, and the matching relationship space for calculating and analyzing the mathematical relationships between product styles and design elements. The model explores and reveals the logical and mathematical quantitative relationships among user preferences, product styles, and design elements. The results of controlled experiments indicate that the AWSCM has significant positive impacts on the design time and design quality. Particularly for non-professional designers, the AWSCM can help reduce the design time and enhance design comprehension, thereby improving design efficiency. This suggests that the AWSCM is beneficial for aiding designers in quickly generating reasonable designs in their design practice.

1. Introduction

User preferences depict individuals’ perceptions and understanding of objects, embodying a fusion of subjective emotions and objective imagery. In product design, they describe the psychological perception and evaluation of users regarding product style, and have become a significant reference in product design processes [1,2,3]. Existing studies confirm that user preferences profoundly impact consumer behaviour. For instance, Ryan Elder’s [4] and Aradhna Krishna’s research describes the formation of mental preference, its multimodal properties, and its role in consumer behaviour. Wu Jinnan’s team [5] shows that product presentation online significantly affects user preferences. These preferences also influence the willingness to purchase wearable devices through perceived social risk and positive emotions. Some scholars and designers commonly believe that uncovering and satisfying user preferences is crucial in determining whether a product will be accepted by the market and consumers [6,7,8]. User preferences are elicited by human subjective cognition and the external design elements of a product. Given the collective nature of user cognition and the diversity of design elements, satisfying user preferences necessitates prior research into the characteristics of these preferences. This involves delving into and analyzing the preferred product styles of the target users and the design elements of the target product. Currently, quantitative analysis is a key method for analyzing and uncovering the traits of objects. This method enhances people’s comprehension and use of these traits. The literature [9,10,11,12,13,14] shows the characterization of various styles. These include players’ playing style, online consumption style, 3D shape style, singing style, driving style, and packaging design style. The studies adopt different observational perspectives. They demonstrate the application of a quantitative analysis across multiple domains. Therefore, the use of quantitative analyses of the differences in people’s preferences for product styles and the relevance of design elements to product styles is feasible and necessary.

Meanwhile, over the past three years, we located only a few articles. Lee Hyo [15] conducted experiments on automotive wheel models for research purposes. These experiments aim to assist designers to ensure that wheel designs meet the engineering objectives. Li [16] applied big data techniques to analyse design factors. These factors impact the structure of agricultural vehicle wheel hubs. The analysis aims to enhance the performance design of axle hubs. Both of them focus on the effects of the hub style on structural performance and modes. Consequently, a characterization study of hub style from users’ preferences and the product styling design perspective is necessary.

The current collection of research focuses on product style evaluations and design solution recommendations and evaluations. In terms of the style assessment, Leblebici-Başar [17] and his team investigated the roles of styles and emotion in transforming concepts into product forms. Su [18] and colleagues employed Term Frequency-Inverse Document Frequency (TF-IDF) to extract the target style requirements from online reviews. Liu [19] et al. utilized a semantic quantitative matching technique to mine the salient styles of car appearance. Pu [20] et al. combined text mining and neural network techniques to extract the styles of high-speed train styling. Wang [21] et al. proposed a product perceptual style evaluation method that considers attribute associations. Belfi [22] and colleagues quantified the relationship between emotional potency, emotional arousal, and the vividness of evocative styles with aesthetic attractiveness using linear mixed-effects modeling. The results indicate a close relationship between emotional potency, style vividness, and the aesthetic appeal of music. Song [23] systematically designed a CMF for automotive interiors. The design is based on perceptual engineering and eye-tracking techniques. Song mathematically characterized the correlation between CMF design elements and three design styles for automotive interiors. Zhang [24] describes a multimodal measurement product styling design method. This method combines electroencephalography (EEG) and eye tracking. It is used to characterize different product styling styles. This is achieved by assessing the degree of similarity in EEG and eye tracking data. Shi [25] proposed a method for evaluating the product style design quality. This method is based on optimized multiple-attribute group decision-making under uncertainty. They employed the TODIM and EDAS techniques to assist in design decision-making.

In terms of design solution recommendations and evaluations, Xu [26] and his team developed a cultural and creative product design evaluation method based on a computer-perceived style system. They established a deep connection between style legends and sub-legends through data analysis and visualization. Su [27] et al. constructed a product styling perceptual style evaluation model using a convolutional attention-parallel neural network. Zhao [28] and others used StackGAN for intelligent conversion from product styles to product form design schemes. Liu [29] and others applied an interactive genetic algorithm generative adversarial network for intelligent design from color styles to color schemes. Ding [30] and others trained a color scheme generator with a generative adversarial network and combined it with a BP neural network to train the style evaluator. They collectively built a product color intelligent design system. Additionally, some studies [31,32,33] also used intelligent algorithms to implement product solution recommendations and style assessments. A comparison of all the above-mentioned relevant studies is summarized in

Table 1. The comparison of all the above-mentioned relevant studies.

Methods	Related Research	Advantages	Disadvantages
Text/Words	The literature [17,18,19,20,21,26]	The process of information acquisition is simple.	Text and word information is prone to ambiguity and multiple interpretations, which can readily cause misunderstandings.
Physiological data	The literature [23,24,25]	The information is abundant and precise.	Physiological data information possesses a degree of specialization that can be challenging for designers to grasp and comprehend. The process of acquiring information is intricate.
Artificial Neural Networks	The literature [22,27,28,29,30,31,32]	It can quickly generate a large number of design schemes.	Neural networks possess black box characteristics and a lack of information transparency, allowing only for the acquisition of design outcomes.

.

By comparing the aforementioned studies, we observe some deficiencies. When conveying and representing user-preferred product styles through text formats or physiological data like EEG, designers struggle to quickly comprehend and apply these professional data, hindering the rapid creation of designs that cater to user preferences. Furthermore, the use of artificial algorithm-based product solutions in the recommendation and evaluation process introduces a level of opacity and black-boxing, making it challenging for designers to grasp the underlying design logic of the output results, and thus failing to enhance their understanding of product styles favored by users. To assist designers in the improvement of design comprehension and in swiftly generating wheel designs that align with user preferences, our goal is to develop a user preference-based Automotive Wheel Hub Style Characteristic Model (AWSCM) for automotive wheel hub design. This model aims to explore and uncover the latent quantitative relationships between user preferences, product styles, and design elements, thereby enhancing design transparency and efficiency.

2. Materials and Methods

2.1. Product Style Characterization Model Based on User Preferences

Characterization, also known as mental characterization or knowledge characterization, is a key concept in cognitive psychology. It deals with how information or knowledge is presented and recorded in mental activities [34,35]. This concept has been widely integrated into various research fields. Zou [36] constructed a design feature hierarchical relationship model and a design feature attribute dependency network model. These models abstractly express conceptual product topology to process and express numerous design information accurately. They achieve a mathematical description of conceptual product design information based on multicolour sets. Zhang [37] used event-related potential technology to monitor users’ EEG information. He characterized users’ perceptual style thinking through changes in EEG signals. Devamanyu Hazarika [38] and his team proposed a method to apply modal invariant space and specific modal space. This method realizes the modal characterization of heterogeneous signals and performs a multimodal sentiment analysis of user-generated videos. Wang [39] and others developed a multipoint multi-objective decision-making model. This model characterizes driving styles by analyzing drivers’ visual field characteristics and decision-making willingness. These studies show that characterization can be understood as the reproduction of external things. It transforms complex, vague, and obscure information carriers into explicit, clear, and easy-to-understand forms. In product style characterization, the goal is to transform ‘style’ from textual forms to multimodal forms. These forms include image symbols, audio beats, and design parameters. The transformation aims to convert ‘style’ from textual polysemy and ambiguity into concrete and clear design information. Therefore, this paper proposes a wheel style characterization model based on user preferences. The model covers the design element space, style space, and relationship space, as shown in Figure 1.

Style space describes the outcomes of the style perception and selection of target products by designers and the public. This process is driven by cognitive psychology and influenced by social culture, educational background, and personal experience. The space takes users’ subjective evaluation information as input. It relies on perceptual engineering theory and related algorithms to analyze and reason about the evaluation information. The goal is to screen style needs and judge the importance of different styles. The construction of this space involves two layers:

The first layer involves mining and extracting the style demand. It collects and refines key style terms using data crawler technology and the KJ clustering method.

The second layer differentiates the importance degree of key styles. It uses the semantic difference method to construct the semantic decision matrix. The entropy weight method calculates the importance degree of wheel style.

The design element space refers to taking the research object as the input and using morphological similarity as the defining criterion. It employs computer capability and designer experience to cluster samples of the research object. The output is representative key morphological symbols. The construction of the space is divided into three layers:

The first layer is the sample collection of the research object, manually gathered from various channels.

The second layer involves data cleansing and clustering processing. It screens the most representative samples from each cluster to build the product sample space.

The third layer uses a morphological analysis to deconstruct wheel samples, extract key morphological symbols, and build a design factor library for design innovation and optimization.

The relationship space records and expresses the transmission relationship and characteristics of ‘style’. It is the output result of computer technology and cognitive analysis. The space takes ‘style’ as the input and outputs material information. It constructs the quantitative characterization information chain of style and design elements, and visualizes the characterization path of style. The construction of this space includes three levels:

The first level is the semantic evaluation stage, which evaluates style elements from the user’s perspective.

The second level is the morphology coding stage, which encodes samples using the design factor library and transforms design elements into computer-recognizable sequences.

The third level is the mapping stage, which establishes the map between semantic evaluation values and morphology coding. It calculates the contribution degree between style and each design factor, establishing a precise characterization chain from style to design elements.

2.2. Data Material

2.2.1. Sample Data Collection and Pre-Processing

The market offers a wide variety of car wheels with varying quality. To ensure the applicability and timeliness of the samples, the wheel hub samples were collected within the range of 2013 to 2023. Samples were collected through online sources (like professional wheel hub websites and wheel hub magazine websites; all links are listed in Supplementary File SIV), resulting in a total of 408 wheel hub images. Due to the large number and variety of sample images, the study must preprocess the wheel images systematically. The goal is to obtain a uniform image format and avoid the influence of redundant factors. The preprocessing is detailed in Figure 2. Firstly, the perspective angles of the wheel pictures are standardized. The study area is the front side of the wheel, as styling features are primarily focused on the front view. Redundant pictures that do not conform to the front view angle are eliminated. Next, the background and brand logo in the pictures are removed. This step is necessary to eliminate their influence on feature extraction and user evaluation. Then, the size of the wheel images is standardized to avoid the influence of image size on visual perception. All images are resized to 512 px × 512 px, with a resolution of 72 dpi and a bit depth of 8. Finally, the images are converted to grayscale to eliminate the influence of color differences on the classification. After processing, 255 wheel sample images are selected for the study. All samples are displayed in Supplementary File SI.

2.2.2. Sample Data Clustering and Filtering

The product sample space represents a condensed and refined version of product samples, which is primarily used for clustering and extracting features from these samples. Efficiently clustering samples allows for the more effective extraction of typical design elements. Thus, the way samples are clustered is crucial for the subsequent extraction of design elements. Common image sample clustering methods include manual clustering, k-means clustering, and clustering using deep learning techniques [40,41,42,43]. Considering the unique features of wheel hub images, this study uses a VAE-DBSAN as the sample clustering model, which is a combination of Variational Autoencoder (VAE) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN). This approach is applied to perform unsupervised learning clustering and filtering of the tire sample database, aiming for rapid and effective clustering. VAE [44] is an unsupervised generative model that integrates deep learning with probabilistic modelling. Its architecture comprises two parts: an encoder and a decoder. The encoder maps observations x to hidden variables z in the latent space. This parametric mapping follows a Gaussian or uniform distribution. The decoder, on the other hand, maps the latent variable z back to the observed data x. VAE enhances the robustness of the decoding process to noise by introducing noise into the encoded results. DBSCAN [45] is a density-based spatial data clustering algorithm. It eliminates the need to predefine the number of clusters. DBSCAN can handle noisy points and discover clusters of arbitrary shapes. It identifies core, boundary, and noise objects using the Eps neighbourhood and MinPts parameters. A region in the clustering space must contain a number of objects above a set threshold for clustering to occur. Compared to the classical K-means approach, DBSCAN effectively fits the distribution of the overall samples. It also ensures the distinction and differentiation between clustered communities, demonstrating a better anti-interference ability. Compared to deep learning classification, DBSCAN has lower requirements for computational hardware and environments, making it easier to operate and implement.

The sample clustering model consists of two stages: feature extraction and clustering. In the feature extraction stage, VAE is used to map high-dimensional data to the latent space and learn their feature characterization. In this study, VAE consists of an encoder and a decoder. The encoder is composed of two convolutional layers, each with a convolution kernel size of 4, a stride of 2, and a ReLU activation function. The decoder consists of two deconvolutional layers, each with a convolution kernel size of 4, a stride of 2, and activation functions of ReLU and Sigmoid, respectively. The loss function is comprised of two parts: Reconstruction Loss and Kullback–Leibler Divergence (KL Divergence). In the clustering stage, the feature vectors are input into DBSCAN for the clustering analysis of the overall samples. Eps radius and MinPts are two key parameters in DBSCAN. To optimize the clustering results, we introduce a function called optimize_dbscan_params, which employs a grid search to iterate through various combinations of the Eps and MinPts parameters. The function evaluates each combination’s clustering performance using the silhouette coefficient, with the goal of identifying the set of parameters that maximizes this coefficient, thereby yielding the best clustering outcome. The final optimal DBSCAN parameters are Eps = 0.9 and MinPts = 5. The detailed complete code is provided in Supplementary File SII. This experiment was conducted on a computer. The computer is equipped with an Intel Core i5 13600KF processor and an RTX3060Ti discrete graphics card. The experiment was run in the Python-3.9.7 environment. The clustering results are presented in Figure 3 and Table S1 of Supplementary File SIII.

In addition, to verify the reliability of the VAE-DBSCAN clustering results proposed in this paper, the clustering effects of methods such as manual clustering, K-means clustering, and ResNet-50+HC clustering are compared. The evaluation indicators include the number of clusters, style consensus, and clustering time. Among them, style consensus refers to the consistency of sample styles within a cluster, calculated as the ratio of the total number of samples minus the questionable samples to the total number of samples (questionable samples are those that the expert panel considers do not belong to the cluster). The results show that VAE-DBSCAN outperforms manual clustering and K-means clustering in terms of style consensus and clustering time, and it performs better than ResNet-50+HC in terms of clustering time. Therefore, it can be concluded that the clustering results of VAE-DBSCAN are reliable and can achieve rapid and effective sample clustering. The comparison of the results from the four methods are presented in

Table 2. Comparison of the results from the four clustering methods.

Method	Cluster Quantity	Style Consensus (%)	Computational Cost (min)
Manual Categorization	9	76.5	180
K-means (Raw Pixels)	7	63.1	97
ResNet-50+HC	8	79.8	34
VAE-DBSCAN	7	81.2	13

.

2.2.3. Design Element Deconstruction

In current design practice, designers in enterprises play a central role. They focus closely on user needs and design goals. They rely on their own subjective design experience and creative thinking to create designs. The logic of the style characterization of design and its results must align with the cognitive logic and judgement of designers. The cognitive characterization of design simulates the cognitive logic and thinking mode of designers. It requires analyzing and characterizing the intrinsic connection between styles and styling design elements. This analysis occurs across the dimensions of design intention, drawing behavior, and functional structure.

Firstly, this study involves 10 senior corporate designers. They deconstruct and analyze design elements in wheel styling. The front view of the wheel reveals two major components: the rim and the spokes. The rim, the peripheral round part of the wheel, mounts the tire and supports the vehicle’s weight. Its styling features are relatively singular and less variable. The spokes, which support structure connecting the axle and the rim, vary in shape and quantity by vehicle model. They are the focal point of wheel styling design. This study summarizes wheel design elements into five categories: spoke layout, spoke shape, spoke number, spoke width, and center shape. Each category is further subdivided, with a detailed elaboration and explanation provided for each subcategory.

(1)

Spoke layout classification involves the arrangement pattern of the spokes. It shapes the overall shape and basic direction of the spoke profile. This is a key factor in determining the visual impact of the spoke profile. The study systematically summarizes and classifies spoke profiles from 255 wheel samples. The spoke layout is divided into four main categories: point line, petal, rotating, and netted. The point line type refers to spokes radiating outward in a straight line from the center point. The petal type describes spokes spreading out like flower petals. The rotating type is characterized by twisted spokes forming a rotating visual effect. The netted type involves spokes expanding outward and intertwining to form a mesh structure. The classification of spoke layouts is shown in Figure 4.

(2)

The spoke shape classification is a key aspect of wheel design. The spokes are the core component of the wheel shape. Their design characteristics are repetitive and regular, arrayed around the center. Consequently, the study of spoke profiles focuses on the morphological characteristics of a single spoke. This study analyzed and summarized the spoke characteristics of 255 previously collected wheel samples. The spoke profiles were classified into six basic forms: V-shaped, polygon-shaped, Y-shaped, rectangles, irregularly shaped, and curve-shaped. The classification of spoke shapes is presented in Figure 5.

(3)

The classification of the number of spokes is based on the wheel design. A single spoke serves as the basic unit, arranged in a circular array. The number of spokes directly impacts the visual sense of sparseness and density, influencing the viewer’s sensory experience. This study details the count of spokes in the 255 previously collected wheel samples. The samples are classified into five categories: one piece, three piece, four piece, five piece, and seven piece. The one piece category refers to a single spoke in the shape of a disc. The three piece category includes designs with a number of spokes divisible by three. The four piece category consists of designs with a number of spokes divisible by four. The five piece category represents designs with a number of spokes divisible by five. The seven piece category includes designs with a number of spokes divisible by seven. The classification of the number of spokes is illustrated in Figure 6.

(4)

The spoke width classification is based on the ratio of the width to height (w/h) of a single spoke. A smaller ratio indicates a finer spoke width. This study evaluates and counts the spoke widths of 255 wheel samples from previously collected images. The spoke widths are classified into three categories: string, block, and surface. The string category is defined by a spoke width ratio w/h less than 1/4. The block category refers to a spoke width ratio between 1/4 and 1/2. The surface category is defined by a spoke width ratio greater than 1/2. The category of spoke width is shown in Figure 7.

(5)

The classification of the center shape pertains to the shape of the area formed by the hub center hole and its connection to the spokes. This area is the focal point of visual perception. An in-depth analysis and summary of the center area characteristics are conducted using images from 255 previously collected wheel samples. The center shapes are classified into five categories: circle, pentagon, hexagon, octagon, and oval. The category of center shapes is demonstrated in Figure 8.

2.2.4. Design Element Space

This study relies on the designer’s experience and cognitive judgement. It decomposes wheel styling elements and summarizes them into five categories, totaling 23 design elements. Clarifying and unifying the attribution relationship between each design factor and its associated design elements is necessary for memory and record-keeping. Therefore, all design elements and elements are marked with an identifier. The operation is as follows: if the wheel design element set is A, it contains several element families, denoted as Ai. Each family includes multiple design elements, which can be marked as Aij. For example, the first element family in the first design factor is A11, the second is A12, and so on. This method not only organizes and differentiates wheel styling design features comprehensively but also provides a graphical reference for subsequent design work. It efficiently demonstrates the distribution of design elements. The final wheel design factor space is presented in

Table 3. Wheel design factor space.

Design Elements	Design Factor
Spoke Layout (A1)	Point line (A11), Petal (A12), Rotating (A13), Netted (A14)
Spoke shape (A2)	V-shaped (A21), Polygon-shaped (A22), Y-shaped (A23), Rectangles (A24), Irregularly shaped (A25), Curve-shaped (A26)
Number of spokes (A3)	One piece (A31), Three piece (A32), Four piece (A33), Five piece (A34), Seven piece (A35)
Spoke width (A4)	String (A41), Block (A42), Surface (A43)
Center formation (A5)	Circle (A51), Pentagon (A52), Hexagon (A53), Octagon (A54), Oval (A55)

.

2.2.5. Online Review Data Mining and Processing

The style space maps users’ direct feelings about product style preferences and expectations intuitively. Users typically express their style needs using texts or terms rich in emotional meaning, known as ‘style terms’. So, the study in this section aims to explore users’ style needs for automotive wheel styling by a combination of data mining and subjective evaluation, which is divided into two phases: data mining and processing, and style need extraction. In the data mining and processing phase, the crawler tool ‘Octopus Gatherer’ is first used to collect comments on car styles and wheel styles from the internet. To reduce noise interference, the collected data are initially screened and cleaned. This process removes useless spam comments and duplicate content to construct a clean initial corpus. The initial corpus then undergoes a segmentation operation. The Jieba segmentation tool is used to split the comment text into independent word units. The Baidu stop word list filters out deactivated terms from the comments. The adjectives in the dataset are counted and filtered by word frequency. To ensure the validity of the results, adjectives with a word frequency of at least 3 times are selected. This study initially filters and numbers 126 style terms, as detailed in Table S2 of Supplementary File SIII.

2.2.6. Style Space

In the stage of style requirement extraction, the 126 style terms extracted from the initial corpus cover a wide range. There may be semantic overlap and identity among these terms, which is not conducive to a targeted study of style requirements. Consequently, it is necessary to downscale these 126 style terms. Firstly, this study invited five designers, five postgraduates with a language background, and five ordinary users to classify the style terms using the KJ (Kawakita Jiroh Method). They reduced the 126 style terms to 26 terms. The results are presented in Table S3 of Supplementary File SIII.

Subsequently, to enhance the representativeness of the style terms and seek to reduce the cognitive burden on users during the post-evaluation process, a questionnaire conducted a secondary screen of the initially screened 26 style terms. In the web questionnaire, the 26 style terms are randomly ranked. Participants were instructed to select the 8 terms that best align with their personal preferences and expectations. The questionnaire targets automotive designers, car salespeople, and actual users. A total of 247 valid questionnaires were distributed and collected. Finally, the collected data were analyzed using SPSS-28.0.1.1 software. The frequency distribution of style terms was obtained, as shown in Figure 9. The analysis reveals that 8 style terms, including V2 (geometric), V4 (fairing), V5 (elegant), V7 (lush), V10 (steady), V12 (firm), V9 (line), and V18 (dynamic), have a word frequency of more than 50%. These 8 style terms are identified as the final style requirements.

2.2.7. Style Semantic Evaluation

In this section of our study, we aim to explore the implicit relationship between design elements and style values through semantic evaluation and Quantification Theory Type I, thereby constructing a chain of information that can represent style and design elements. Firstly, the study explores users’ perceptual perception of wheel styling. If people are required to browse and evaluate 255 wheel hubs, it would greatly increase their cognitive burden. At the same time, the number of clusters in each group is different, and the amount of sample information needs to be balanced across different clusters to avoid data redundancy. Therefore, reducing the dimensionality of the number of wheel hubs to be evaluated is necessary. After in-depth discussions and exchanges among the expert panel, a decision is made to select samples. Ten samples are chosen from each cluster. These samples are distributed equidistant from the cluster centre. Each sample interval is one-tenth of the distance from the cluster centre. This method ensures an adequate number of samples. It effectively covers the characteristics of the samples while preventing excessive similarity between them. Consequently, 70 wheel samples are selected as representative samples. These samples are displayed in Figure 10.

It uses 70 representative sample pictures and 8 style terms. These were screened in the previous section to compile a semantic evaluation questionnaire. The questionnaire adopts a 1~5 level semantic evaluation method. Subjects are asked to score each sample’s 8 style terms based on their actual feelings. Higher scores indicate a higher degree of style recognition.

It is believed that males are more enthusiastic about understanding and paying attention to cars [46,47]. To ensure the authenticity and validity of the evaluation results, we invited 50 subjects to participate in this study. The evaluation took place in a 4S shop. Of the 50 subjects, 35 were males and 15 were females. All subjects planned to buy a car, were aged between 25 and 40 years, had an education level of high school or above, and had a visual acuity >5.0. Subsequently, the evaluation results were collated by sample number and analyzed for reliability and validity, and the results are shown in Table S4 of Supplementary File SIII. The Cronbach α coefficients were all >0.85, indicating high reliability. The KMO values were all >0.75, and the p-values were all <0.05, indicating good validity. This suggests that the questionnaire content is valid and the evaluation results are reliable. Finally, the evaluation values were homogenized and used as the results of wheel style evaluation, as shown in

Table 4. Evaluation matrix for wheel styles.

Number	Value
	Geometric	Fairing	Elegant	Lush	Steady	Firm	Linear	Sporty
No. 1	3.54	3.36	3.64	3.44	3.32	3.52	3.54	3.38
No. 2	3.4	3.38	3.52	3.16	3.54	3.3	3.18	3.34
No. 3	3.48	3.46	3.16	3.32	3.52	3.36	3.32	3.46
No. 4	3.32	3.46	3.26	3.22	3.3	3.36	3.34	2.98
No. 5	3.36	3.24	3.4	3.38	3.2	3	3.04	3.38
No. 6	3.4	3.36	3	3.5	3.28	3.46	3.56	3.24
No. 7	3.62	3.62	3.32	3.64	3.48	3.5	3.5	3.54
No. 8	3.96	3.82	3.74	3.98	3.58	3.68	3.94	3.5
No. 9	3.74	3.36	3.46	3.64	3.52	3.7	3.7	3.38
No. 10	3.8	3.36	3.4	3.68	3.48	3.68	3.84	3.58
…	…	…	…	…	…	…	…	…
No. 67	3.5	3.36	2.98	3.46	3.38	3.42	3.38	3.54
No. 69	3.66	3.48	3.48	3.56	3.5	3.52	3.46	3.42
No. 70	3.58	3.24	3.36	3.36	3.48	3.42	3.36	3.58

.

2.2.8. Sample Morphology Code

To explore and characterize the mapping relationship between style and design elements using computer algorithms, we must help the computer identify styling elements in wheel sample pictures, which means converting the abstract and vague symbolic language of the pictures into a computer-understandable language. Accordingly, the study of this section manually encodes wheel samples morphologically. It encodes and marks the design element items within the 70 samples using a binary system, where the presence of the Aijth design element is counted as ‘1’ and its absence as ‘0’. This encoding allows each sample to be represented as a 23-bit string, for example, ‘0100/010000/01000/010/010/01000’. The morphology codes for wheel samples are listed in Table S5 in Supplementary File SIII.

3. Results

3.1. Results of Style Weighting

Given the inherent differences in style ontology, users often have diverse style feelings and needs for the target object in reality. Their recognition of and demand for different styles also show individual differences. To deeply explore users’ style preferences, quantitatively differentiating the importance of style is necessary. Based on the evaluation data in Table S4, the study of this section constructs an initial evaluation matrix for wheel style. The matrix rows represent 70 different samples, and the columns correspond to 8 evaluation dimensions. Additionally, the entropy weighting method serves as an objective method for assigning weights, as it calculates the weights of each evaluation index and provides a quantitative basis for a comprehensive multi-indicator evaluation system [48,49]. It calculates the weights of each evaluation index. So, we applied the entropy weight method to calculate the entropy weights between different styles in this study. This quantifies the importance of differences between them. Through this process, we obtained the coefficient of differentiation and weight value for each style. The detailed results are shown in

Table 5. Results of style weighting.

Weights	W1	W2	W3	W4	W5	W6	W7	W8
Terms	geometric	fairing	elegant	lush	steady	firm	linear	dynamic
Score	0.072	0.219	0.095	0.183	0.134	0.104	0.124	0.070

.

The data in Table S5 indicate that the ‘fairing’ style ranks first, with a weight value of 0.219. This suggests that the ‘fairing’ style has the highest recognition among users. The ‘dynamic’ style ranks last, with a weight value of 0.070, indicating that it is the least recognized style. The recognition rankings for the eight style are as follows: V4 (fairing) > V7 (lush) > V10 (steady) > V9 (linear) > V12 (firm) > V5 (elegant) > V2 (geometric) > V18 (dynamic). Consequently, in the subsequent design process, designers should prioritize incorporating top-ranked style, such as ‘fairing’, into the design concept.

3.2. Quantitative Analysis of the Results of Design Elements with Style

This study aims to depict the contribution degree relationship between each design element and wheel style accurately by Quantification Theory Type I. Quantification Theory Type I is suitable for research where the independent variable is qualitative and the dependent variable is quantitative. It constructs a numerical model through multiple regression analysis. This model transforms complex qualitative and quantitative problems into quantifiable data, revealing intrinsic connections and laws [50,51]. In the study of this section, the morphological coding of 70 samples serves as the independent variable. The semantic evaluation value of each style is the dependent variable. Then, we conduct in-depth analyses of the eight styles in turn.

(1)

The research results include the quantitative mapping relationship analysis of ‘geometric’. The PCE (partial correlation coefficients) for each design element and the design elements’ scores are presented in

Table 6. Results of quantitative ‘geometric’ calculations.

Dependent Variable: Geometric
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.281	5	A11	−0.338
			A12	−0.201
			A13	−0.274
			A14	−0.652
A2	0.732	1	A21	0.17
			A22	0.16
			A23	0.027
			A24	0.13
			A25	0.234
			A26	0.208
A3	0.544	3	A31	−0.317
			A32	−0.102
			A33	−0.224
			A34	−0.26
			A35	−0.157
A4	0.457	4	A41	0.089
			A42	0.067
			A43	−0.033
A5	0.689	2	A51	0.693
			A52	0.653
			A53	0.369
			A54	0.955
			A55	0.46

.

The data from Table 6 show that A2 ‘spoke shape’ (PCE = 0.732) and A5 ‘center formation’ (PCE = 0.689) have significant scores. This indicates that these design elements significantly influence the feelings of ‘geometric’. The in-depth analysis reveals that ‘spoke shape’ has the most outstanding performance in design factor A25 ‘irregularly shaped’ (with a score of 0.234). This suggests that ‘irregularly shaped’ has a significant impact on the feelings of ‘geometric’. There is a strong positive correlation between ‘irregularly shaped’ and ‘geometric’. Designers should focus on and prioritise these key elements in the creative process to maximize their positive impacts. Conversely, the PCE of A1 ‘spoke layout’ is only 0.281. This indicates a relatively weak connection with the feelings of ‘geometric’. In design practice, the attention given to these weakly related design elements should be reduced accordingly.

(2)

For ‘fairing’, the quantitative mapping relationship analysis shows the PCE of each design element and the design factor score results in

Table 7. Results of quantitative ‘fairing’ calculations.

Dependent Variable: Fairing
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.467	3	A11	−0.183
			A12	−0.064
			A13	0.12
			A14	0.021
A2	0.768	1	A21	0.047
			A22	0.032
			A23	−0.054
			A24	0.099
			A25	0.117
			A26	0.478
A3	0.288	5	A31	−0.037
			A32	0.237
			A33	0.003
			A34	0.179
			A35	0.155
A4	0.591	2	A41	0.144
			A42	−0.063
			A43	−0.08
A5	0.45	4	A51	0.441
			A52	0.327
			A53	0.242
			A54	0.468
			A55	0.353

.

The data in Table 7 indicate that the design elements A2 ‘spoke shape’ (PCE = 0.768) and A3 ‘spoke width’ (PCE = 0.591) have high scores. This reveals their significant influences on the feelings of ‘fairing’. For ‘spoke width’, the design factor A31 ‘string’ (with a score of 0.144) is outstanding. In the case of ‘spoke shape’, A26 ‘curve-shaped’ (score of 0.468) also has a high score. These data suggest that ‘curve-shaped’ spokes and the ‘string’ spoke width are strongly related to the feelings of ‘fairing’. Designers should focus on these elements and prioritize their positive contributions. In contrast, A4 ‘number of spokes’ (PCE = 0.288) has a lower partial correlation coefficient. This indicates a weaker correlation with the feelings of ‘fairing’. Consequently, the importance of this element in the design process should be reduced accordingly.

(3)

For ‘elegant’, the quantitative mapping relationship analysis, PCE of each design element, and design factor score results are shown in

Table 8. Results of quantitative ‘elegant’ calculations.

Dependent Variable: Elegant
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.632	1	A11	0.624
			A12	0.762
			A13	0.653
			A14	0.558
A2	0.543	2	A21	0.258
			A22	0.201
			A23	0.118
			A24	0.125
			A25	0.205
			A26	0.035
A3	0.456	3	A31	0.671
			A32	−0.12
			A33	−0.079
			A34	−0.121
			A35	−0.168
A4	0.369	4	A41	−0.199
			A42	−0.117
			A43	−0.342
A5	0.297	5	A51	−0.214
			A52	0.105
			A53	−0.138
			A54	−0.165
			A55	−0.197

.

The data in Table 8 show that the design elements A1 ‘spoke layout’ (PCE = 0.632) and A2 ‘spoke shape’ (PCE = 0.543) have high scores. This suggests that they significantly influence the feelings of ‘elegance’. In the ‘spoke layout’ category, the design factor A12 ‘petal’ has the highest score (0.762). For ‘spoke shape’, A21 ‘V-shape’ is the most prominent, with a score of 0.258. These data indicate that ‘petal’ and ‘V-shaped’ have strong positive correlations with the feelings of ‘elegant’. Designers should prioritize these design elements in their creations. In contrast, A5 ‘center formation’ has a low partial correlation coefficient (PCE = 0.297). This indicates a relatively weak correlation with the semantic meaning of ‘elegant’. Consequently, attention to this element in design practice could be reduced.

(4)

For ‘luxury’, the quantitative mapping relationship analysis, PCE of each design element, and design factor score results are shown in

Table 9. Results of quantitative ‘lush’ calculations.

Dependent Variable: Lush
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.269	4	A11	−0.47
			A12	−0.134
			A13	−0.336
			A14	−0.16
A2	0.741	2	A21	0.642
			A22	0.609
			A23	0.292
			A24	0.868
			A25	0.363
			A26	0.381
A3	0.402	3	A31	−0.378
			A32	−0.198
			A33	−0.362
			A34	−0.499
			A35	−0.222
A4	0.877	1	A41	0.559
			A42	0.474
			A43	0.374
A5	0.162	5	A51	0.026
			A52	0.132
			A53	−0.008
			A54	0.241
			A55	0.36

.

The data in Table 9 indicate that the design elements A4 ‘spoke width’ (PCE = 0.877) and A2 ‘spoke shape’ (PCE = 0.741) have significant scores. This suggests strong associations with the feelings of ‘lush’ compared to other elements. In-depth analyses reveal that for ‘spoke width’, the design factor A41 ‘string’ has a higher score (0.559). In the ‘spoke shape’ category, A24 ‘rectangle’ also has a high score (0.868). These data show that the ‘string’ spoke width and ‘rectangle’ spoke shape have strong positive correlations with the feelings of ‘lush’. Designers should therefore prioritize these elements in their creations. In contrast, A1 ‘spoke layout’ (PCE = 0.269) and A5 ‘center formation’ (PCE = 0.162) have lower partial correlation coefficients. This indicates weaker relationships with the feelings of ‘lush’. Consequently, the focus on these elements could be reduced during the design process.

(5)

The quantitative mapping relationship between design factors and ‘steady’ is analyzed. The PCE of each design element and the design factor score results are shown in

Table 10. Results of quantitative ‘steady’ calculations.

Dependent Variable: Steady
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.721	1	A11	0.635
			A12	0.623
			A13	0.605
			A14	−0.192
A2	0.379	4	A21	0.123
			A22	0.105
			A23	−0.01
			A24	0.069
			A25	0.101
			A26	0.014
A3	0.644	2	A31	0.147
			A32	0.253
			A33	0.381
			A34	0.471
			A35	0.293
A4	0.481	3	A41	0.228
			A42	0.252
			A43	0.281
A5	−0.186	5	A51	0.101
			A52	−0.231
			A53	−0.687
			A54	−0.192
			A55	−0.226

.

The data in Table 10 show that the design elements A1 ‘spoke layout’ (PCE = 0.721) and A3 ‘number of spokes’ (PCE = 0.644) have high partial correlation coefficients. This indicates a strong correlation with the feelings of ‘steady’. Further exploration reveals that within the ‘spoke layout’ category, the design factor A11 ‘point line’ has a high positive score (0.635). Similarly, in the ‘number of spokes’ category, the design factor A34 ‘five’ also has a high positive score (0.471). These data suggest that the ‘point line’ layout and a spoke count of five have significant positive correlations with the feelings of ‘steady’. Designers should focus on these elements and prioritize their development in design. In contrast, A5 ‘center formation’ (PCE = −0.186) has a low and negative coefficient. This suggests a weak or potentially negative correlation with the feelings of ‘steady’. In design practice, attention to this element should be reduced accordingly.

(6)

The quantitative mapping relationship analysis of ‘firm’ is presented. The PCE of each design element and the design factor score results are shown in

Table 11. Results of quantitative ‘firm’ calculations.

Dependent Variable: Firm
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.349	4	A11	−0.195
			A12	−0.142
			A13	−0.127
			A14	−0.047
A2	0.579	3	A21	0.237
			A22	0.228
			A23	0.212
			A24	1.185
			A25	0.241
			A26	0.154
A3	0.614	2	A31	0.004
			A32	0.263
			A33	−0.047
			A34	0.352
			A35	0.401
A4	0.683	1	A41	−0.272
			A42	0.486
			A43	0.449
A5	0.092	5	A51	0.144
			A52	0.096
			A53	−0.052
			A54	−0.025
			A55	0.047

.

The data in Table 11 indicate that the design elements A3 ‘spoke width’ (PCE = 0.683) and A4 ‘number of spokes’ (PCE = 0.614) have high partial correlation coefficients. This suggests significant correlations with the feelings of ‘firm’. Within the ‘spoke width’ category, the design factor A42 ‘block’ has the highest score (0.486). In the ‘number of spokes’ category, the design factor A45 ‘seven piece’ also has the highest score (0.401). These data indicate that the ‘block’ spoke width and a spoke count of seven have more significant positive correlations with the feelings of ‘sturdy’. Designers should focus on these elements and prioritize their development to enhance their positive impacts. In contrast, A5 ‘center formation’ (PCE = 0.092) has a lower partial correlation coefficient. This suggests a weaker correlation with the feelings of ‘sturdy’. This element may receive less attention in the design process.

(7)

The quantitative mapping relationship analysis for ‘linear’ is presented. The PCE of each design element and the design factor score results are shown in

Table 12. Results of quantitative ‘linear’ calculations.

Dependent Variable: Linear
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.526	2	A11	1.148
			A12	0.397
			A13	0.502
			A14	−0.143
A2	0.367	4	A21	0.202
			A22	0.201
			A23	0.000
			A24	0.055
			A25	0.11
			A26	0.241
A3	0.442	3	A31	0.259
			A32	0.356
			A33	−0.257
			A34	0.269
			A35	0.348
A4	0.584	1	A41	0.475
			A42	−0.384
			A43	−0.714
A5	−0.129	5	A51	−0.114
			A52	−0.226
			A53	−0.314
			A54	−0.217
			A55	−0.143

.

The data in Table 12 reveal that design elements A3 ‘spoke width’ (PCE = 0.584) and ‘spoke layout’ (PCE = 0.526) have higher partial correlation coefficients. This indicates closer relationships with the feelings of ‘linear’. In the ‘spoke width’ category, the design factor A31 ‘string’ has the highest score (0.475). For ‘spoke layout’, the design factor A11 ‘point line’ exhibits a very high positive value (score of 1.148). These findings suggest that the ‘string’ spoke width and ‘point line’ layout have the most significant positive correlations with the feelings of ‘linear’. Designers should prioritize these elements in their designs. In contrast, A5 ‘center formation’ has a negative partial correlation coefficient (−0.129). This indicates a weak or slightly negative correlation with the feelings of ‘linear’. In design practice, the focus on this element could be reduced accordingly.

(8)

The quantitative mapping relationship analysis for ‘dynamic’ is presented. The PCE of each design element and the design factor score results are shown in

Table 13. Results of quantitative ‘dynamic’ calculations.

Dependent Variable: Dynamic
Design Elements	PCE (Partial Correlation Coefficient)	Rank	Design Factor	Score
A1	0.562	2	A11	0.358
			A12	0.425
			A13	0.582
			A14	0.055
A2	0.781	1	A21	0.701
			A22	0.158
			A23	0.415
			A24	0.179
			A25	0.758
			A26	0.419
A3	0.289	4	A31	−0.517
			A32	0.213
			A33	0.19
			A34	0.187
			A35	−0.436
A4	0.436	3	A41	0.473
			A42	0.361
			A43	0.181
A5	0.058	5	A51	0.047
			A52	0.058
			A53	−0.057
			A54	0.079
			A55	0.181

.

The data in Table 13 show that the design elements A2 ‘spoke shape’ (PCE = 0.781) and A1 ‘spoke layout’ (PCE = 0.526) have significant partial correlation coefficients. This indicates strong correlations with the feelings of ‘dynamic’. In the ‘spoke shape’ category, the design factor A25 ‘irregularly shaped’ has the highest score (0.758). For ‘spoke layout’, the design factor A13 ‘rotating’ exhibits a high positive value (score of 0.582). These findings suggest that ‘irregularly shaped’ spoke shapes and ‘rotating’ layouts have particularly strong positive correlations with ‘dynamic’ semantics. Designers should prioritize these elements in their creations. In contrast, A5 ‘center formation’ has a low partial correlation coefficient (PCE = 0.058), indicating its lesser relevance to ‘dynamic’ semantics. In the design process, the focus on this element could be reduced accordingly.

In summary, this study explores the cognitive characterization path of design. Through steps such as refining, pictorial processing, and evaluation, it quantifies the correspondence between style and design elements. The study constructs a characterization chain containing eight styles and 23 design factors. This chain structure, presented visually in numerical and graphical forms (Figure 11), reveals complex interactions between style and design factors. This method not only prioritizes design elements but also provides objective data support for design decision-making. It allows designers to understand and select appropriate design elements more intuitively. Consequently, the design process is optimized and decision-making efficiency is improved.

4. Experimental Analysis

To verify whether the Automotive Wheel Hub Style Characterization Model (AWSCM) established in this study achieves the purpose of deepening designers’ understanding of style and design efficiency, this section of this research aims to evaluate the practical application effects of AWSCM in actual design scenarios through empirical research. To complete the experimental validation, we are using a controlled experiment method to assess the impacts of the model on designers, design efficiency, and design quality. The experimental steps are as follows.

4.1. Participant Selection and Experimental Design

For this experiment, 20 designers were recruited to participate in an empirical research project. These individuals were currently active in the automotive design industry or related disciplines, with 10 being professional designers from automotive firms, each possessing at least three years of industry experience. The remaining 10 participants were students pursuing a degree in automotive design, equipped with relevant knowledge but lacking practical project experience. Consequently, the participants were categorized into two groups: Group A, comprising the professional designers, and Group B, consisting of the design students. This research employed a controlled experimental design, featuring an experimental group and a control group. In the control group, designers were utilizing traditional methods and personal experience to create wheel hub designs adhering to a specified style. Conversely, the experimental group has access to data from the Automotive Wheel Style and Configuration Model (AWSCM) as a reference for their designs. Participants from both groups were randomly assigned to either the experimental or control group, ensuring a fair distribution and minimizing potential biases. This arrangement resulted in four groups, two experimental (A′ and B′) and two control (A″ and B″) groups, with each group comprising five designers.

4.2. Experimental Procedure

The experiment was structured in three phases: task allocation, execution of the design process, and data collection and analysis. In the initial phase, all groups were given the same style design task after reviewing the experimental guidelines and requirements. This task involved creating sketches and three-view drawings of a wheel hub within a three-hour time limit. During the second phase, participants concentrated on the design process. The control group used their personal experience and traditional design methods, while the experimental group referred to the results of the Automotive Wheel Style and Configuration Model (AWSCM) for their designs. In the final phase, following the completion of the design tasks by each group, the design time for each participant was documented. Then, all design proposals were subjected to a quality assessment, where 20 users were invited to score each design on a 100-point scale. A statistical analysis of these evaluation results was then performed to assess the impact of AWSCM on both design efficiency and effectiveness.

4.3. Experimental Results and Analysis

The results of this experiment aim to evaluate the impact of AWSCM on design efficiency and design quality in order to determine whether the model can deepen designers’ understanding of style and improve design efficiency.

(1)

Design time. In terms of design time, the average time for experimental group A′ was 57 min, the average time for experimental group B′ was 76 min, and the overall average time for the experimental groups was 66.5 min. The average time for control group A″ was 83 min, the average time for control group B″ was 123 min, and the overall average time for the control groups was 103 min. There was a 26 min difference between experimental group A′ and control group A″ and a 47 min difference between experimental group B′ and control group B″; both differences were significant (p < 0.01). Overall, there was a 36.5 min difference between the experimental groups and the control groups, which was statistically significant (p < 0.05). Additionally, through cross-comparison, it can be found that there was a 6 min difference between experimental group B′ and control group A″, which was statistically significant (p < 0.05). Based on the above data, it can be considered that AWSCM has a significant positive impact on reducing the design time and improving design efficiency.

(2)

Design quality. In terms of design quality, the average score for experimental group A′ was 85, the average score for experimental group B′ was 78, and the overall average score for the experimental groups was 81.5. The average score for control group A″ was 82, the average score for control group B″ was 65, and the overall average score for the control groups was 73.5. There was a 3-point difference between experimental group A′ and control group A″ and a 13-point difference between experimental group B′ and control group B″; both differences were significant (p < 0.05). Overall, there was an 8-point difference between the experimental groups and the control groups, which was statistically significant (p < 0.05). Additionally, through cross-comparison, it can be found that there was a 4-point difference between experimental group B′ and control group A″, which was statistically significant (p < 0.05). Based on the above data, it can be considered that AWSCM has a significant positive impact on improving the design quality of non-professional participants.

(3)

Design efficiency. In terms of design efficiency, by comparing the statistical results of design time and design quality, for professional participants, there was a 26 min time difference and a 3-point score difference between experimental group A′ and control group A″, suggesting that AWSCM can improve design efficiency by reducing the design time. For non-professional participants, there was a 47 min time difference and a 13-point score difference between experimental group B′ and control group B″, suggesting that AWSCM can improve design efficiency by reducing the design time and enhancing design understanding. Through cross-comparison, it can be further suggested that AWSCM can help non-professional participants improve design efficiency by reducing the design time and enhancing design understanding, as there was a 6 min time difference and a 4-point score difference between experimental group B′ and control group A′.

5. Discussion

This study elucidates the interplay between user preferences, product style, and design elements by developing a conceptual model for automotive hubs. It converts the subtle and indistinct aspects of product style into tangible visual representations and quantifiable data. Reflecting on the research methodology and outcomes, the discussion of the research’s innovations and shortcomings is as follows:

(1)

To enhance design efficiency, this research proposes a VAE-DBSCAN clustering approach in the sample clustering phase, which demonstrates superiority in both clustering performance and time efficiency for wheel samples compared to methods like manual clustering and K-means. This improvement accelerates the process of wheel sample clustering and the extraction of design element features.

(2)

The study presents a dual-chain representation framework (AWSCM) capable of quantifying user preference weights (using an entropy-based semantic analysis) and the contribution of design elements (through Quantification Theory Type I), creating a transparent link between abstract styles and concrete design elements. This framework assists designers in comprehending the interconnections between user preferences, product style, and design elements, enabling them to rapidly produce designs that cater to user preferences.

(3)

The AWSCM proposed in this paper is a transparent-first approach that replaces the reliance on black-box algorithm recommendations and subjective experience with designer-centric symbolic logic and intuitive data, thereby enabling rapid and traceable design decision-making and enhancing design efficiency. Compared to research results based on text or physiological data (such as EEG), this paper adopts a symbolic representation for design elements. These symbolic elements are easier for both the general public and designers to recognize, understand, and utilize in decision-making. Compared to product design research based on intelligent algorithms, the model proposed in this paper is more transparent and is more likely to assist designers in understanding the correlation between design elements and styles. This understanding can contribute to the improvement of design thinking. However, in terms of design speed, this method may not be as fast as artificial intelligence algorithms.

(4)

The characterization results presented in this paper are based on a theoretical analysis with a limited number of samples. To enhance the universality and feasibility of these results, it is essential to increase the quantity of user evaluation data. This expansion will facilitate further validation and analysis of the model. Additionally, design involves the recombination of multiple elements. Future research will aim to explore and quantify the combined effects of different design elements on style perception. For instance, it will investigate whether the simultaneous presentation of design factor A51 and design factor A11 influences their individual contributions to the ‘geometric’ style.

(5)

When user preferences for an object are not singular but a composite of multiple intertwined styles, the design solution needs to balance different style requirements. The combination of style weights and quantitative analysis results (the product of style weights and bias correlation coefficients of design elements) reveals the quantitative relationship between each design item and style requirements. For a different number of styles, the quantitative results should be flexible. The style weights and analysis results can be recalculated based on the actual number of styles, and the combination can be recalculated accordingly.

6. Conclusions

The form of wheels is the main visual carrier of product style. Different styles can emerge from the aesthetic combination of various design elements. This study proposes an Automotive Wheel Hub Style Characterization Model (AWSCM) based on user preferences that analyzes and characterizes the relationships between user preferences, style, and design elements. On one hand, this research constructs a representation chain from user preferences to product style. This includes stages such as mining, filtering, and evaluation, utilizing the entropy weight method to calculate the importance of each style in user preferences. On the other hand, the study decomposes and classifies the design factors within the hub through morphological analysis, guided by the designer’s thinking and functional structure. This establishes a mathematical relationship between ‘style’ and ‘design elements’, constructing a feature chain between ‘style’ and ‘design elements’. The combination of these two chains builds a top-down automotive hub style representation chain from user preferences to style and then to design elements. This effectively transforms the implicit and ambiguous ‘style’ from textual information into intuitive and clear design elements, improving transparency in design and thereby assisting designers in deepening understanding and enhancing design efficiency. Future research will investigate the comprehensive impacts of different design elements on style perception, aiming to provide a more detailed reference for design.