A Method for the Front-End Design of Electric SUVs Integrating Kansei Engineering and the Seagull Optimization Algorithm

Abstract

With the rapid expansion of the Electric Sport Utility Vehicle (ESUV) market, capturing consumer aesthetic preferences and emotional needs through front-end styling has become a key issue in automotive design. However, traditional Kansei Engineering (KE) approaches suffer from limited timeliness, subjectivity, and low predictive accuracy when extracting affective vocabulary and modeling the nonlinear relationship between product form and Kansei imagery. To address these challenges, this study proposes an improved KE-based ESUV styling framework that integrates data mining, machine learning, and generative AI. First, real consumer reviews and front-end styling samples are collected via Python-based web scraping. Next, the Biterm Topic Model (BTM) and Analytic Hierarchy Process (AHP) are used to extract representative Kansei vocabulary. Subsequently, the Back Propagation Neural Network (BPNN) and Support Vector Regression (SVR) models are constructed and optimized using the Seagull Optimization Algorithm (SOA) and Particle Swarm Optimization (PSO). Experimental results show that SOA-BPNN achieves superior predictive accuracy. Finally, Stable Diffusion is applied to generate ESUV design schemes, and the optimal model is employed to evaluate their Kansei imagery. The proposed framework offers a systematic and data-driven approach for predicting consumer affective responses in the conceptual styling stage, effectively addressing the limitations of conventional experience-based design. Thus, this study offers both methodological innovation and practical guidance for integrating affective modeling into ESUV styling design.

1. Introduction

In the context of global energy transition and environmental imperatives, electric vehicles (EVs) have become a pivotal driver of sustainable mobility and a key contributor to carbon neutrality. Among them, sport utility vehicles (SUVs) are particularly favored for their spacious interiors, versatility, and comfort [1]. Studies show that the market share of electric SUVs (ESUVs) in China rose from 45% to 55% between 2019 and 2021 [2], and projections suggest that ESUVs will surpass internal combustion engine (ICE) SUVs in the European Union by 2025, reaching nearly six million units by 2030 [3]. More importantly, ICE SUVs consume more fuel and emit more due to their weight and drag [4], whereas electric propulsion meets performance and space needs while improving energy efficiency. As a result, ESUVs represent significant strategic value and growth potential in the evolving automotive market. With consumer decision-making increasingly driven by emotional and experiential factors—and core product attributes becoming more homogeneous—automotive styling, especially front-end design, has become a key point of differentiation and emotional branding [5]. This visual focal point not only forms consumers’ first impressions but also conveys brand identity and cultural value, thereby influencing emotional resonance and purchasing behavior. Within this context, shape imagery—defined as the symbolic and aesthetic meaning embedded in form—plays a vital role in fostering emotional connection [6]. Furthermore, electrification reduces the need for large front grilles, granting designers greater creative freedom in ESUV front-end design [7]. Thus, accurately capturing and predicting consumers’ emotional preferences and integrating them into front-end styling is of theoretical and practical significance for advancing emotion-driven EV design.

Kansei Engineering (KE) [8] is an established method for analyzing the relationship between user emotion and product form, and has been widely applied in automotive styling [9,10]. It typically involves three stages: feature decomposition, affective information extraction, and predictive model construction [11]. However, traditional KE studies often rely on interviews, surveys, focus groups, or basic text analysis, which struggle to capture diverse, multilayered affective needs and suffer from limited objectivity and timeliness. With the rise of e-commerce and big data, large volumes of user-generated content (UGC) have become an important source for collecting authentic emotional feedback [12]. Scholars have employed Python-based Scrapy and natural language processing (NLP) techniques to mine online reviews and construct affective corpora, as demonstrated by Lai et al. [13], who combined Scrapy with Word2Vec (a classic NLP technique) to identify Kansei vocabulary related to EV exterior design, thereby facilitating emotional need analysis and shape imagery prediction. However, conventional text mining techniques often fail to reveal latent topics and subtle emotional expressions in large-scale short-text UGC due to sparsity, noise, and fragmentation [14]. To address these issues, unsupervised topic modeling has been widely adopted to extract semantic structures from unstructured text [15]. While Latent Dirichlet Allocation (LDA) [16] is frequently applied in KE-related studies [17,18], its performance declines in short-text scenarios due to sparse word co-occurrence [19]. As an alternative, the Biterm Topic Model (BTM) [20] models biterms across the corpus, improving topic extraction from short texts. For instance, Pan et al. [21] used BTM to analyze public perceptions in heritage district reviews, while Zhang et al. [22] applied BTM to identify latent user demands in intelligent product–service systems. These studies confirm BTM’s effectiveness for affective information extraction in design contexts. In predictive modeling, KE traditionally employs Multiple Linear Regression (MLR) [23] and Quantification Theory Type I (QTT-I) [24]. For example, Liu and Yang [25] used MLR to map product features to affective responses, while Xue et al. [26] adopted QTT-I to quantify user perception of design attributes. However, affective cognition is often nonlinear, subjective, and dynamic, limiting the effectiveness of linear models [27]. In response, recent studies have explored machine learning methods such as Back Propagation Neural Networks (BPNNs) [28] and Support Vector Regression (SVR) [29]. For example, Zhu et al. [30] employed BPNN to incorporate affective and sustainable design parameters for product optimization and satisfaction prediction, whereas Yang and Shieh [31] utilized SVR to estimate consumer affective responses to styling attributes. Although these models handle nonlinearities well, their performance is sensitive to initialization and hyperparameter selection [32,33]. Thus, optimization techniques such as evolutionary algorithms [34], random search [35], and exhaustive grid search [36] have been used to improve generalization. The research conducted by Lin et al. [37], Liu et al. [38], and Yang et al. [39] serves as a prime example of this.

However, as problem dimensionality and complexity increase, traditional optimization techniques often struggle with highly nonlinear and dynamically evolving shape imagery prediction tasks, leading to premature convergence, limited global search, and high computational costs. To address these limitations, this study adopts Swarm Intelligence Algorithms [40] to optimize BPNN and SVR, leveraging their robustness and adaptability in complex search spaces. Among them, the Seagull Optimization Algorithm (SOA) [41], proposed by Gaurav and Vijay in 2019, simulates seagull soaring and diving behaviors through spiral flight and stochastic migration, enabling efficient global search in high-dimensional, multimodal spaces. SOA has shown strong performance in domains such as power systems [42], environmental modeling [43], and engineering safety [44], and demonstrates better parameter adaptability than Whale Optimization Algorithm (WOA) and Grey Wolf Optimizer (GWO), particularly under sparse-sample conditions. Although underexplored in product design, its potential for high-precision nonlinear modeling is evident. Nonetheless, the inherent limitations of single-algorithm strategies, such as restricted search diversity and susceptibility to local optima, can compromise optimization effectiveness. To address this, this study also employs Particle Swarm Optimization (PSO) [45], which simulates collective learning behaviors by dynamically updating particles’ positions based on personal and global best experience. Due to its simple structure and high computational efficiency, PSO has been widely applied to nonlinear prediction tasks. For example, Fu et al. [46] successfully used PSO to optimize SVR parameters in Ming-style furniture design, significantly enhancing model accuracy and applicability. Based on these considerations, this study constructs four predictive models, namely SOA-BPNN, SOA-SVR, PSO-BPNN, and PSO-SVR, with their respective predictive performance systematically evaluated and analyzed based on error comparison methods. The objective is to identify the optimal associative model for the precise prediction of ESUV front-end shape imagery.

After determining the optimal predictive model, verifying its practical applicability is crucial. Traditional manual modeling for product styling is time-consuming, highly prone to subjective biases, and lacks reproducibility. With the rise of Artificial Intelligence-Generated Content (AIGC), design processes increasingly shift toward intelligent automation, providing an efficient framework for emotional expression and generative creativity [47]. Text-to-Image (T2I) technology, in particular, synthesizes images from natural language descriptions, enabling multidimensional evaluation of shape imagery prediction models and reducing the workload and biases associated with manual design [48]. To demonstrate this feasibility, the present study employs Stable Diffusion (SD), a state-of-the-art T2I tool, to generate ESUV front-end styling proposals and validate the predictive model’s effectiveness in Kansei Engineering-based applications.

In summary, this study proposes a KE-based ESUV front-end styling method by progressively integrating multiple analytical techniques. It begins by collecting frontal-view ESUV images from public platforms to build a styling dataset, from which representative samples are selected and their design features deconstructed as input for predictive modeling. Building on this, the BTM is combined with the AHP to extract Kansei vocabulary as model output, capturing consumers’ affective preferences. To further enhance model performance, SOA and PSO are employed to optimize the parameters of BPNN and SVR, respectively. Finally, the four models are evaluated through error comparison to identify the optimal configuration, which is subsequently validated via SD-generated styling proposals. The main contributions are threefold: (1) introducing BTM-based affective analysis into KE to improve the extraction of latent emotional needs from short, fragmented user texts; (2) enhancing the robustness and accuracy of shape imagery prediction by optimizing BPNN and SVR with SOA and PSO; and (3) validating the model via SD-generated proposals to reduce subjectivity and improve the rigor and reproducibility of the design process. Together, these contributions reinforce the reliability of KE-based predictive modeling and offer practical guidance for intelligent, emotion-driven design.

The remainder of this paper is organized as follows: Section 2 presents the methodology and models; Section 3 outlines the ESUV front-end design experiments; Section 4 concludes the study.

2. Materials and Methods

To address the limitations of conventional KE in capturing complex emotional responses, this study proposes a multimodal framework that integrates big data mining, optimization algorithms, and artificial intelligence to model the Kansei imagery of ESUV front-end design. As shown in Figure 1, the methodology unfolds in three stages.

2.1. Affective Data Acquisition Based on BTM

As a probabilistic topic modeling approach, BTM focuses on the co-occurrence relationships of global biterms. By directly modeling all biterms in the text and utilizing Gibbs sampling to infer topic distributions, BTM generates word distributions for each topic, thereby accurately uncovering the latent thematic structures within the text. The training model structure is illustrated in Figure 2.

In the figure: α, β represent prior parameters; θ denotes the latent topic distribution of all biterms in the corpus; K represents the optimal number of topics; φ denotes the word distribution of topics in the corpus; z represents the latent topics in the corpus; w_i, w_j denote two different words forming a biterm; |B| represents the total number of biterms in the corpus.

Since product imagery styling involves highly subjective and diverse affective vocabulary, the model does not preset a fixed number of topics. Instead, Gibbs sampling dynamically adjusts the topic number based on data characteristics during training. Prior to sampling, the hyperparameters are empirically set as α = 50/k and β = 0.01, where k denotes the candidate topic number. Multiple values of k are tested, and Gibbs sampling is applied iteratively to train each. The process updates biterm-topic assignments via conditional probability, gradually converging to the optimal topic distribution. The corresponding formula is as follows:

$$P(z\mid z_{-b},B,\alpha ,\beta )\propto (n_z+\alpha )\frac{(n_{w_i|z}+\beta )(n_{w_j|z}+\beta )}{{({\displaystyle \sum _w}{n}_{w|z}+V\beta )}^2}$$

(1)

where V represents the total number of unique words in the corpus; n_z denotes the number of biterms currently assigned to topic z; z_−b represents the topic distribution of all biterms in the corpus, excluding biterm b; B denotes the set of all possible biterms in the text.

To enhance model interpretability and practical applicability, the optimal number of topics for BTM in this study is determined by evaluating candidate topic numbers using a combination of Perplexity and Coherence metrics. Perplexity measures the model’s ability to generalize to new data, where lower values indicate better generalization performance. Coherence, on the other hand, assesses the semantic consistency of biterms within a topic, with higher scores reflecting stronger associations and higher-quality topics. The formulas for Coherence and Perplexity are as follows:

$$Perplexity=\exp (-\frac{{\displaystyle \sum _{d=1}^M}\log p(w_d)}{{\displaystyle \sum _{d=1}^M}{N}_d})$$

(2)

$$Coherence=\sum _{n=2}^T\sum _{j=1}^{n-1}\log \frac{D(w_n^z,w_j^z)+1}{D(w_j^z)}$$

(3)

where M represents the number of texts to be modeled; N_d denotes the number of biterms contained in the d-th text; p_(wd) represents the probability of word w appearing in the d-th text; T refers to the number of high-frequency words in a topic; D($w_n^z$, $w_j^z$) represents the co-occurrence frequency of biterms in the text.

2.2. Construction of Nonlinear Predictive Models

2.2.1. Back Propagation Neural Network

First, to prevent overfitting, the dataset is divided into a training set and a test set, which are used for model learning and generalization validation, respectively. Additionally, since data collected through questionnaires often exhibit different scales and variances, it is necessary to normalize the data before constructing the BPNN model, following neural network computation rules. In detail, normalization aims to rescale data with different magnitudes into a fixed range, typically [0, 1] or [−1, 1]. This process, achieved through linear transformation, reduces scale differences among features, ensuring that each variable contributes equally to the training process and preventing any feature from dominating or being overshadowed. The normalization formula is as follows:

$$x^{\prime }={\displaystyle \frac{2(x-x_{\min })}{x_{\max }-x_{\min }}}-1$$

(4)

where x represents the original data; x_min and x_max denote the minimum and maximum values of feature x, respectively.

In terms of network architecture, the BPNN model constructed in this study consists of an input layer, a hidden layer, and an output layer. To be specific, the number of nodes in the input layer is determined based on the number of decomposed product styling feature categories, ensuring that all design elements are fully represented. The number of nodes in the output layer corresponds to the number of affective vocabulary terms, with each node representing a specific user’s affective preference. The number of hidden layer nodes (r) is initially determined using an empirical formula, which is given as follows:

$$r=\sqrt{l+k}+\mathsf{\partial }$$

(5)

where n represents the number of nodes in the input layer; l represents the number of nodes in the output layer; $\mathsf{\partial }$ is a constant, typically set within the range [1, 10].

In nonlinear shape imagery prediction, activation function selection is critical to balancing model expressiveness and computational efficiency. This study applies the Tan-Sigmoid function in the input layer and the Purelin linear function in the hidden layer, with Trainlm as the training algorithm. This architecture enables the input layer to perform a nonlinear transformation, while the hidden layer propagates information linearly to reduce computational complexity. Thus, the configuration ensures a balance between nonlinear learning capacity and efficiency.

2.2.2. Support Vector Regression

In the process of establishing the complex nonlinear relationship between ESUV front-end styling features and affective demands, SVR constructs a tolerance-based nonlinear mapping, enabling it to capture subtle correlations between variables in high-dimensional feature spaces. This ensures both prediction accuracy and model robustness.

First, the regression function for the SVR model needs to be constructed, which is formulated as follows:

$$f(x)={\omega }^{T}x+b$$

(6)

where ω represents the weight vector; x denotes the nonlinear mapping function; b represents the bias term; T denotes the transpose operation, applied to vectors or matrices. This expression defines the linear decision-making structure in function space via the formulation of the weight vector and bias term. However, the key innovation of SVR lies in introducing the error tolerance margin (ε) and incorporating weighted prediction errors into the objective function. This enhances not only prediction accuracy but also generalization. Moreover, the ε-insensitive mechanism filters out minor fluctuations, improving the model’s adaptability to larger sample spaces.

More importantly, the core objective of the SVR model is to establish an optimization process that balances model complexity and error tolerance. In this process, SVR achieves global data distribution optimization by minimizing a weighted combination of the squared weight norm and the sum of deviations, formulated as follows:

$$\underset{\omega ,b}{\min }\frac{1}{2}\parallel \omega {\parallel }^2+C\sum _{i=1}^n{l}_{\epsilon }(f(x_i),y_i)$$

(7)

where C represents the regularization parameter, which controls the trade-off between model complexity and error tolerance; n denotes the total number of data samples; x_i represents the feature vector of the i-th styling sample; y_i denotes the affective evaluation score of the i-th sample; lε(f(x_i),y_i) is the ε-insensitive loss function, defined as follows:

$$l_{\epsilon }(f(x_i),y_i)=\left\{\begin{array}{cc}0,& |f(x_i)-y_i|<\epsilon \\ |f(x_i)-y_i|-\epsilon ,& |f(x_i)-y_i|\ge \epsilon \end{array}\right.$$

(8)

where ε represents the maximum acceptable deviation.

In the optimization process, to allow the model greater flexibility in handling training samples, particularly in the presence of outliers, slack variables ${\xi }_i$ and $\widehat{{\xi }_i}$ are introduced into SVR to account for errors. The reformulated optimization problem is expressed as follows:

$$\begin{array}{c}\underset{\omega ,b,{\xi }_i,\widehat{{\xi }_i}}{\min }\frac{1}{2}\parallel \omega {\parallel }^2+C{\displaystyle \sum _{i=1}^n}({\xi }_i+\widehat{{\xi }_i})\\ \begin{array}{c}\mathrm{s}.\mathrm{t}.\left\{\begin{array}{l}f(x_i)-y_i\le \epsilon +{\xi }_i\hfill \\ y_i-f(x_i)\le \epsilon +\widehat{{\xi }_i}\hfill \\ {\xi }_i\ge 0,\widehat{{\xi }_i}\ge 0,i=1,2,\cdots ,n\hfill \end{array}\right.\end{array}\end{array}$$

(9)

where ${\xi }_i$ and $\widehat{{\xi }_i}$ represent non-negative slack variables, introduced to handle deviations between predicted values and actual values.

Furthermore, to simplify computation and enhance solution efficiency, SVR reformulates the original optimization problem into its dual optimization form using the principles of the Lagrange multiplier. By introducing Lagrange multipliers ${\alpha }_i$ and ${\widehat{\alpha }}_i$, the dual optimization problem is expressed as follows:

$$\begin{array}{c}\max [{\displaystyle \sum _{i=1}^n}{y}_i({\widehat{\alpha }}_i-{\alpha }_i)-\epsilon ({\widehat{\alpha }}_i+{\alpha }_i)-\frac{1}{2}{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}({\widehat{\alpha }}_i-{\alpha }_i)({\widehat{\alpha }}_j-{\alpha }_j)K(x_i,x_j)]\\ s.t.\left\{\begin{array}{l}{\displaystyle \sum _{i=1}^n}({\widehat{\alpha }}_i-{\alpha }_i)=0\\ 0\le {\alpha }_i,{\widehat{\alpha }}_i\le C\end{array}\right..\end{array}$$

(10)

where K(x_i,x_j) represents the kernel function, which is used to map data into a high-dimensional feature space.

In the high-dimensional feature space, the introduction of the kernel function enables the transformation of the nonlinear problem into a convex quadratic programming problem. The final regression model is expressed as follows:

$$f(x)=\sum _{i=1}^n({\widehat{\alpha }}_i-{\alpha }_i)K(x_i,x)+b$$

(11)

where The kernel function K(x_i,x_j) is typically computed using the Radial Basis Function (RBF), which is defined as follows:

$$K(x_i,x_j)=\exp \left(-g\parallel x_i-x_j{\parallel }^2\right)$$

(12)

where g represents the kernel function parameter, which controls the influence range in the input space.

2.3. Introduction of Optimization Algorithms

2.3.1. Seagull Optimization Algorithm

The core principle of the SOA lies in its dynamic balancing mechanism, which enables the algorithm to efficiently search for optimal solutions in multidimensional nonlinear optimization problems. Specifically, SOA consists of two primary behaviors: migration and foraging. The migration behavior focuses on global exploration, while the foraging behavior emphasizes local exploitation, ensuring a comprehensive optimization process for the target problem. The operational steps of the SOA algorithm are as follows:

(1)

Migration behavior

This behavior can be regarded as the global search process, where seagulls move from one position to another, avoiding collisions while converging toward the optimal solution. During this process, three conditions must be satisfied:

The first condition is that, to prevent becoming trapped in local optima, seagulls utilize an additional behavioral variable A to avoid collisions with other seagulls. The corresponding formula is as follows:

$$C_s=P_s(t)+A$$

(13)

where C_s represents the position update of seagull s in a collision-free scenario; P_s(t) denotes the original position of seagull s at iteration t.

$$A=fc\cdot (1-{\displaystyle \frac{t}{T_{\max }}})$$

(14)

where fc controls the variation of the behavioral variable A; T_max represents the maximum number of iterations.

The second condition is that, after avoiding collisions, seagulls move toward the global optimal solution, simulating the process of searching for the optimal solution in a multidimensional parameter space. The corresponding formula is as follows:

$$M_s=P_s(t)+B\cdot [P_{bs}(t)-P_s(t)]$$

(15)

where M_s represents the position update of seagull s moving toward the optimal solution; P_bs(t) and P_s(t) denote the positions of the best solution (bs) and individual seagull s, respectively, at iteration t; B is a random parameter that balances global exploration and local exploitation, defined as follows:

$$B=2\cdot rand-1$$

(16)

where rand represents a random number uniformly distributed in the range [0, 1].

The third condition is that, during the global search process, seagulls gradually approach the optimal solution. The position update formula is as follows:

$$D_s=C_s+M_s$$

(17)

where D_s represents the overall displacement of seagull s, which is composed of both the collision-avoidance position update and the displacement toward the optimal solution.

(2)

Foraging Behavior

This behavior can be regarded as the local exploitation phase of SOA. During this process, seagulls perform spiral movements at varying angles and velocities to simulate foraging, thereby refining parameter adjustments to further enhance model performance. The local exploitation process is defined as follows:

$$x(t+1)=x(t)+r\cdot \cos (k\cdot u)\cdot e^{-v^2}$$

(18)

where r represents the spiral diameter of the seagull’s attack; k is a random variable, where k ∈ [0, 2π]; u and v are hyperparameters defining the spiral pattern, typically set to 1. The position update formula for the seagull is given as follows:

$$x_{\mathrm{new}}=x(t+1)$$

(19)

2.3.2. Particle Swarm Optimization Algorithm

PSO optimizes model parameters by enabling particles to collaborate and evolve, ultimately converging toward the global optimal solution. During the iterative search process, each particle represents a potential solution and continuously adjusts its velocity and position within the search space, guided by both its own experience and the global best position. The algorithm operates through two key mechanisms. First, it facilitates global exploration by integrating individual particle search results with swarm-wide search dynamics, using the current best solution to direct the global search process. Second, it ensures progressive convergence to the optimal solution by iteratively refining particle positions, combining local learning with collective intelligence to enhance optimization efficiency. The velocity update and position update are formulated as follows:

$$v_i(t+1)=\omega v_i(t)+c_1{r}_1[p_i(t)-x_i(t)]+c_2{r}_2[g(t)-x_i(t)]$$

(20)

$$x_i(t+1)=x_i(t)+v_i(t+1)$$

(21)

where v_i(t) represents the velocity of particle i at iteration t; x_i(t) represents the position of particle i at iteration t; ω denotes the inertia weight, which controls velocity continuity; c₁ and c₂ are learning factors, typically ranging from [0, 2]; r₁ and r₂ are random numbers uniformly distributed within [0, 1], introduced to enhance search diversity by incorporating stochasticity; p_i(t) represents the historical best position of particle I; g(t) represents the global best position of the swarm.

2.4. Prediction Model Optimization

To enhance predictive accuracy and stability, both BPNN and SVR models are optimized via the SOA and PSO. In BPNN, the network architecture (e.g., number of hidden neurons) and key hyperparameters (learning rate, momentum, initial weights, thresholds) are encoded as high-dimensional vectors—treated as seagull or particle “positions”—with the BPNN’s test-set loss function serving as the fitness metric. Through iterative updates governed by SOA’s migration–foraging dynamics or PSO’s velocity–position mechanisms, the global best solution is continuously refined, thereby reducing training errors and improving generalization. In SVR, the penalty coefficient (C) and kernel parameter (g) are similarly represented as two-dimensional vectors, and each candidate combination is evaluated using the SVR’s validation-set loss to guide the search. By leveraging global exploration, both SOA and PSO efficiently traverse the parameter space to avoid local optima, significantly outpacing manual or grid-based tuning in larger-scale optimization tasks. This approach ensures that the selected hyperparameter configurations strike an optimal balance between accuracy and computational cost, thereby substantially enhancing performance in complex nonlinear regression scenarios. The respective optimization workflows for BPNN and SVR using SOA and PSO are illustrated in Figure 3.

2.5. Application of the Stable Diffusion Model

To rapidly generate ESUV images, this study employs the text-to-image generation capability of the SD model for the subsequent Kansei imagery prediction experiments. As a typical latent diffusion model, SD efficiently performs text-to-image synthesis by integrating a pretrained text encoder and a Variational Autoencoder (VAE) to operate within the latent space. The generation process is shown in Figure 4.

3. Results

3.1. Data Crawling of ESUV Front-End Styling Samples and Kansei Vocabulary

The present study selects Autohome (https://www.autohome.com.cn, accessed on 17 April 2025), one of China’s most influential automotive business platforms, as the primary data source for Kansei imagery analysis. This platform not only hosts a large and active user base but also encourages consumers to provide authentic and diverse opinions on various vehicle models. Given these characteristics, it serves as an ideal source for collecting a high-quality initial dataset that is highly representative of consumer perceptions. Specifically, this study focuses on 26 ESUV models that hold significant market shares and high brand recognition among mainstream consumer groups, including Model Y (Tesla Inc., Palo Alto, CA, USA), AITO M9 (Seres Group Co., Ltd., Chongqing, China), and Li Auto L7 (Li Auto Inc., Beijing, China), etc.. The research extracts online review texts related to these models using a Python 3.9-based web crawling program, which automates the collection of user comments from targeted webpages by employing “Electric SUV” as the primary keyword. To enhance data acquisition efficiency and accuracy, the web crawler integrates processes such as page parsing, data cleaning, and duplicate comment filtering. As a result, 34,697 valid consumer reviews were successfully retrieved. Simultaneously, 156 ESUV front-end styling sample images were collected. In the subsequent sample preprocessing phase, to reduce information redundancy and eliminate highly overlapping brand-specific design features, a focus group manually filtered the dataset based on front-end styling similarity. This process ultimately refined the dataset to 140 ESUV front-end images, forming a structured sample repository, as illustrated in Figure 5.

3.2. Selection of Representative Samples and Decomposition of Styling Features

3.2.1. Selection of Representative Samples

To objectively and systematically identify representative ESUV front-end samples, the 140 initial samples were classified using the KJ method. Given that color perception preferences and environmental factors might affect the accuracy and stability of classification results, preliminary background removal and grayscale processing were performed on all samples using Adobe Photoshop 2024 (v25.12). This preprocessing step aimed to minimize non-styling influences during observation and evaluation. Following this, a focus group was formed, consisting of 37 participants, including graduate and doctoral students specializing in vehicle engineering and automotive design, professors, and automotive designers with over five years of styling experience. These experts conducted detailed visual assessments of the preprocessed samples and applied the KJ method based on their subjective perceptions. Ultimately, 44 representative samples were identified to encapsulate the diverse characteristics of ESUV front-end styling. Furthermore, to enable SD questionnaire participants to focus exclusively on front-end styling features, while minimizing biases related to viewing angles, branding, and color perception, and to mitigate copyright concerns associated with using original vehicle images, additional vector-based stylization was performed using Adobe Illustrator 2024. The results are shown in Figure 6.

3.2.2. Decomposition of ESUV Front-End Styling

In this study, to obtain the input-layer data required for the predictive model, the morphological features and design elements of ESUV front-end styling were systematically analyzed and deconstructed using the Morphological Analysis method. Through a comprehensive review of the relevant literature, market research, and expert insights from the focus group, the key elements that significantly influence the overall visual perception of the ESUV front end were identified and screened. Ultimately, from a range of potential styling factors, seven representative design items were selected based on their strongest correlation with the overall front-end styling, as illustrated in Figure 7. These include side mirrors, headlights, fog lights, windshield, lower grille, hood, and front-end contour.

Based on the preliminary definition of these seven design items, the specific morphological elements contained in the 44 representative samples were further decomposed by the focus group to fully reveal the differentiated characteristics of each ESUV front-end design in terms of points, lines, surfaces, and their interrelated mappings. Through iterative refinement and evaluation, 30 element categories were finalized, as presented in

Table 1. ESUV front-end styling design items and element categories.

Design Items	Element Categories
	Serial Number
Side mirrors
	X1	X2			X3		X4			X5
Headlights
	X6	X7			X8		X9			X10
Fog lights
	X11	X12			X13		X14			X15
Windshield
	X16			X17				X18
Lower grille
	X19		X20			X21			X22
Hood
	X23		X24			X25			X26
Front-end contour
	X27		X28			X29			X30

. Within this classification framework, the seven design items and their corresponding 30 element categories effectively encompass all major design variations observed in the 44 representative ESUV front-end samples.

3.3. Kansei Vocabulary Mining Based on BTM

Before applying BTM for topic modeling on ESUV-related online reviews, the code implementation is conducted in Jupyter Notebook (v6.5.4) within Python 3.9, enabling a structured preprocessing workflow. This process includes data cleaning, tokenization, and stop-word removal, ensuring that the final dataset is refined, machine-readable, and optimized for subsequent topic modeling and Kansei vocabulary extraction.

3.3.1. Determining the Optimal Number of Topics

To determine the optimal number of topics, this study applies Equation (1) for text clustering and systematically evaluates the relationship between model performance and topic numbers using Equations (2) and (3). To enable objective comparison, Perplexity and Coherence values are normalized to a common scale. The trends across different topic numbers are visualized in a line chart, as shown in Figure 8.

As shown in Figure 8, Perplexity decreases markedly with more topics, indicating improved model fit. However, when the topic number is below 4, Coherence remains low, suggesting weak semantic association and limited topic interpretability. When the number exceeds 6, Perplexity continues to decline, but topic distinction becomes excessive, raising the risk of over-segmentation. Considering both metrics, the optimal range lies between 4 and 6. Within this range, Coherence peaks at K = 4, indicating the best balance between model fit and semantic coherence. Therefore, 4 is selected as the optimal number of BTM topics.

3.3.2. Topic Evaluation and Selection of Representative Kansei Vocabulary

After determining the optimal number of BTM topics, this study analyzes the latent semantic features of each topic. A refined vocabulary selection and contextual comparison are conducted to identify core imagery words reflecting consumer affective preferences. Specifically, high-probability candidate words are reviewed to eliminate irrelevant or redundant terms, retaining those with clear distinctiveness and strong interpretability in the context of ESUV styling. Word probability, contextual relevance, and usage in the original corpus are then evaluated. Finally, the top 12 feature words per topic are selected, and their probabilities are recorded. The resulting Topic–Feature Word Probability Distribution is shown in

Table 2. Topic-feature word probability distribution for ESUV front-end styling.

Topic 1		Topic 2		Topic 3		Topic 4
Feature Word	Probability Value	Feature Word	Probability Value	Feature Word	Probability Value	Feature Word	Probability Value
Technological	0.031	Premium	0.029	Stable	0.021	Dynamic	0.022
Streamlined	0.027	Powerful	0.029	Ecofriendly	0.019	Technological	0.022
Futuristic	0.022	Expressive	0.027	Premium	0.018	Premium	0.019
Dynamic	0.021	Ecofriendly	0.025	Reserved	0.017	Cute	0.016
Luxury	0.021	Minimalist	0.025	Streamlined	0.017	Aesthetic	0.016
Smart	0.018	Streamlined	0.019	High-end	0.014	Streamlined	0.015
Cool	0.016	Individualistic	0.018	Innovative	0.013	Smooth	0.014
Geometric	0.016	Smooth	0.015	Balanced	0.013	Geometric	0.014
Grand	0.012	Reserved	0.011	Elegant	0.011	Energyefficient	0.009
Minimalist	0.011	Avant-garde	0.011	Agile	0.008	Rounded	0.008
Aesthetic	0.009	Novel	0.010	Classic	0.007	Elegant	0.008
Lightweight	0.006	Low-key	0.007	Graceful	0.005	Fresh	0.006

.

As shown in Table 2, some feature words appear across multiple topics. To better identify the key focus areas of user comments for each topic, low-probability words and those with no substantial relevance to the topic analysis were filtered out. For instance, the word “Streamlined” appears in all four topics, but based on its probability distribution across topics (0.027, 0.019, 0.017, 0.015), it is most frequently associated with Topic 1. Therefore, it is retained in Topic 1 while being removed from the other topics to ensure clearer topic differentiation. Additionally, words such as “Agile” and “Fresh”, which have weaker relevance to ESUV exterior styling, were also eliminated to enhance the accuracy and interpretability of topic characteristics. The final selection of Kansei feature words under each topic was rigorously reviewed and refined by a focus group comprising 8 Ph.D. students in automotive design and 13 experienced automotive designers with over five years of styling expertise. The results are presented in

Table 3. Kansei feature words for ESUV front-end styling.

Topic	Kansei Feature Words
	Serial Number
Topic 1	Technological	Streamlined		Futuristic		Dynamic		Cool
	Y11	Y12		Y13		Y14		Y15
Topic 2	Premium		Powerful		Minimalist		Avant-garde
	Y21		Y22		Y23		Y24
Topic 3	High-end	Reserved		Stable		Elegant		Classic
	Y31	Y32		Y33		Y34		Y35
Topic 4	Rounded		Aesthetic		Cute		Geometric
	Y41		Y42		Y43		Y44

.

To prevent the decline in questionnaire reliability and validity due to a large-scale Semantic Differential (SD) scale, and to maintain participant engagement, the importance of the Kansei feature words in Table 3 was quantified and weighted using the Analytic Hierarchy Process (AHP). This approach ensures that only the most representative Kansei imagery vocabulary is selected for each topic. More precisely, the four identified topics were established as the criterion layer, while the Kansei feature words under each topic were incorporated into their respective sub-criterion layers, thereby constructing a hierarchical analytical structure. Subsequently, a focus group of 21 experts, as previously described, was invited to conduct pairwise comparisons of elements at each hierarchical level. The relative importance of each element was assessed based on Saaty’s scale, and the values were assigned within a judgment matrix.

Finally, this work further applied the geometric mean method to perform comprehensive weighting of Kansei feature words within each topic. Based on these computed weights, a consistency test was conducted on the judgment matrix. When the Consistency Ratio (CR) < 0.1, it indicates that the internal consistency of the judgment matrix is acceptable, meaning that the consistency test is passed. The computed weights (W) and corresponding consistency test results are presented in

Table 4. AHP calculation results.

Topic	Kansei Feature Words	W	CR
Topic 1	Y11	0.4048	0.098
	Y12	0.1520
	Y13	0.1205
	Y14	0.1132
	Y15	0.2095
Topic 2	Y21	0.5481	0.052
	Y22	0.2371
	Y23	0.1566
	Y24	0.0582
Topic 3	Y31	0.4086	0.053
	Y32	0.2697
	Y33	0.1519
	Y34	0.0951
	Y35	0.0746
Topic 4	Y41	0.4658	0.012
	Y42	0.2771
	Y43	0.0960
	Y44	0.1611

.

Based on the above analysis, the feature words with the highest weight values in each topic (Y₁₃, Y₂₁, Y₃₁, and Y₄₁) were selected as the Kansei imagery vocabulary for subsequent experiments. To ensure a more precise and structured evaluation, each selected Kansei word was paired with its antonym, forming a set of oppositional scales that effectively capture the perceptual differences in ESUV front-end styling. Through this process, four representative Kansei word pairs were finalized: Futuristic/Traditional, Premium/Ordinary, High-end/Low-end, and Rounded/Rigid. These contrasting pairs serve as the core semantic dimensions for the upcoming experimental phases, enabling a systematic and quantifiable assessment of how consumers perceive ESUV front-end styling.

3.4. Kansei Evaluation Experiment

To objectively capture consumer emotional perceptions and psychological differences regarding ESUV front-end styling, this investigation adopted the principles of the Semantic Differential method. A questionnaire survey was designed based on the 44 representative ESUV front-end styling samples, structured around the four pairs of Kansei imagery vocabulary. Participants were invited to evaluate each representative sample across different emotional dimensions using a 7-point Likert scale to measure their subjective perceptions. A total of 519 responses were collected, and after removing 34 invalid responses, 485 valid questionnaires were retained for further analysis. To further explore the latent relationships between front-end styling design elements and consumer Kansei evaluations, the styling design elements were encoded and integrated with the emotional assessment data obtained from the survey. Subsequently, four nonlinear predictive models—SOA-BPNN, SOA-SVR, PSO-BPNN, and PSO-SVR—were constructed to systematically analyze and predict Kansei responses based on front-end styling characteristics.

3.5. Construction of the Kansei Imagery Prediction Model

3.5.1. Construction of SOA-BPNN and PSO-BPNN Models for Kansei Imagery Prediction

To systematically represent ESUV front-end styling features, this research constructs an association model using BPNN. The input layer consists of 30 neurons, corresponding to the 30 design element categories, while the output layer contains 4 neurons, mapping to the 4 pairs of representative Kansei imagery vocabulary. The number of hidden layer neurons was initially estimated using Equation (5). Subsequently, with all other hyperparameters fixed, the model was repeatedly trained in MATLAB R2022b, using Minimum Mean Square Error as the performance evaluation metric. Experimental results indicated that when the hidden layer contained 9 neurons, the prediction error reached its minimum value. Therefore, the hidden layer size was set to 9. To prevent overfitting and ensure training stability and generalization capability, 40 out of the 44 ESUV front-end samples were selected for training, while the remaining 4 samples were reserved for validation. Additionally, to maintain consistency across feature dimensions and accelerate network convergence, the input features were normalized prior to training following Equation (4).

During model optimization, this study integrates SOA to perform global optimization of BPNN’s initial weights and biases using Equations (13)–(19). Given that SOA hyperparameters are highly sensitive to model performance, multiple adjustments and experiments were conducted based on previous research insights, evaluating the model’s behavior under different parameter configurations. To ensure fast convergence, achieve a globally optimal solution, and minimize prediction error, the set of hyperparameters yielding the lowest prediction error was selected for SOA-BPNN model training. To be specific, the core SOA parameters were set as follows: Population size (N) = 40, Maximum number of iterations (k_max) = 50, Training error threshold = 0.00001. Additionally, the learning rate was fixed at 0.01, and the maximum training iterations were set to 1000, balancing convergence speed and global search capability. During the training process, convergence curve analysis revealed that at the 50th iteration, the Root Mean Square Error reached its minimum value, indicating that the SOA-BPNN model achieved the expected convergence accuracy. Furthermore, when the best confirmation performance reached 4.33 × 10⁻² at epoch 3, the iteration process was terminated, finalizing the construction of the ESUV front-end styling and Kansei imagery association model, as illustrated in Figure 9.

To further assess the effectiveness of SOA in improving Kansei imagery prediction accuracy, this study also introduces PSO to optimize BPNN. To ensure a balance between search diversity and computational efficiency, hyperparameter tuning was performed multiple times based on previous research insights and Equations (20) and (21). The configuration with the lowest prediction error was ultimately selected for PSO-BPNN model training. Specifically, the selected PSO parameters were: Population size = 30, Maximum iterations = 50. To effectively navigate the high-dimensional weight-bias search space, ensuring wide exploration in early iterations and precise convergence in later stages, a linearly decreasing inertia weight strategy from 0.9 to 0.4 was adopted. Further experimentation revealed that when the Maximum Inertia Weight was set to 0.8 and the Minimum Inertia Weight was 0.3, the model achieved the lowest prediction error. Additionally, the Learning Factor 1 and Learning Factor 2 were both set to 2, ensuring a balanced fusion of local and global optima. Regarding BPNN training settings, in this research, the model’s maximum network training steps were set to 1000, the learning rate was fixed at 0.01, and the target training error was 0.0001. After 50 iterations, the average fitness value reached its optimal state. The best validation performance was recorded as 8.95 × 10⁻² at epoch 2, at which point iteration stopped, successfully completing the construction of the ESUV front-end styling Kansei evaluation prediction model, as illustrated in Figure 10.

Next, to visually compare the differences between consumer Kansei evaluation values and the predicted values from SOA-BPNN and PSO-BPNN models, the styling feature encodings of four validation samples were used as input data for both models. The prediction results are illustrated in Figure 11.

As shown in the figure, SOA-BPNN produces predictions closer to actual consumer Kansei evaluations than PSO-BPNN, indicating that the SOA-BPNN model exhibits superior prediction performance in mapping ESUV front-end styling to Kansei imagery.

3.5.2. Construction of SOA-SVR and PSO-SVR Models for Kansei Imagery Prediction

Given the multidimensional nature of nonlinear regression in machine learning, it is essential to compare alternative techniques for Kansei imagery prediction. In this study, alongside BPNN, an SVR model was constructed based on Equations (6)–(11). SVR is known for its strong generalization and robustness, especially with small sample sizes. However, its performance is highly sensitive to the penalty coefficient (C) and kernel parameter (g), which must be pre-defined. These parameters directly influence prediction accuracy, generalization, and learning efficiency. To address this, SOA and PSO are introduced to iteratively optimize SVR parameters, aiming to build a more accurate model for ESUV front-end Kansei imagery prediction. Specifically, the design elements of 40 ESUV front-end samples were encoded as input variables, and four corresponding emotional evaluation values per sample were used as model outputs for training.

In the SOA optimization process, based on previous research and multiple rounds of manual parameter tuning, the seagull population size was set to 50, and the maximum number of iterations (k_max) was set to 150. The RBF kernel, as specified in Equation (12), was used for model training. To enhance prediction performance and reduce errors caused by imbalanced data distribution, the 40 samples were divided using the K-fold cross-validation method, resulting in 32 training samples and 8 validation samples. Subsequently, SOA’s iterative updating was employed to globally optimize the SVR parameters until the optimal parameter set was found for the final SVR model configuration. The results showed that when the optimal training parameters (c, g) were (0.347, 0.065), the model’s prediction performance was at its best, thus completing the construction of the SOA-SVR model. Additionally, to compare the performance of different optimization algorithms in SVR modeling, PSO was also used to optimize the SVR model. Based on prior knowledge from related research and manual parameter tuning, the particle swarm size was set to 40, and the maximum number of iterations was set to 100. The same RBF kernel function and K-fold cross-validation strategy were used to ensure consistency with the previous training and test sets. Under these conditions, the optimal parameters (c, g) computed by PSO were (0.392, 0.058), leading to the construction of the PSO-SVR model.

To evaluate the prediction performance of the SOA-SVR and PSO-SVR models, the 8 validation samples were input into both models, and the predicted Kansei imagery values from each model were compared with the actual evaluation values. The results are shown in Figure 12. As observed in the figure, both models accurately fitted the trend of the actual emotional ratings to a large extent, indicating that both SOA-SVR and PSO-SVR models were capable of effectively capturing the consumer’s Kansei imagery responses.

Additionally, to facilitate a comprehensive comparison of prediction errors with other models in subsequent analyses, 4 additional test samples were input into both the SOA-SVR and PSO-SVR models. The corresponding emotional prediction values are presented in

Table 5. Kansei imagery prediction values for SOA-SVR and PSO-SVR models.

Kansei Imagery Vocabulary	Test Sample	Prediction Values
		SOA-SVR	PSO-SVR
Futuristic/Traditional	Sample 1	3.498	4.214
	Sample 2	3.595	3.498
	Sample 3	3.501	3.778
	Sample 4	5.482	5.476
Premium/Ordinary	Sample 1	3.508	3.546
	Sample 2	3.753	3.628
	Sample 3	4.310	3.359
	Sample 4	4.529	4.628
High-end/Low-end	Sample 1	3.508	5.228
	Sample 2	3.287	3.852
	Sample 3	3.500	3.598
	Sample 4	2.589	3.117
Rounded/Rigid	Sample 1	3.997	4.221
	Sample 2	3.998	3.558
	Sample 3	3.345	3.971
	Sample 4	5.452	4.626

.

3.6. Comprehensive Comparison of Prediction Errors

To quantitatively evaluate the accuracy of the ESUV front-end styling emotional prediction models and to visually present the prediction accuracy of each optimization model, the model performance is comprehensively assessed using the Average Error Rate (AER). In concrete terms, we selected 4 test samples and input their real emotional evaluation values into each prediction model to obtain the corresponding predicted values. Then, the AEV and RER for each test sample were calculated using the corresponding formula The results are presented in

Table 6. Absolute error and relative error rate for prediction models.

Test Sample	Prediction Models	Error Comparison Method	Output Variable
			Y13	Y21	Y31	Y41
Sample-1	SOA-BPNN	AEV	0.175	0.392	0.309	0.175
		RER	4.68%	9.92%	6.48%	4.68%
	SOA-SVR	AEV	3.498	3.508	5.204	3.997
		RER	0.248	0.441	0.433	0.251
	PSO-BPNN	AEV	0.986	0.522	0.538	0.986
		RER	26.32%	13.21%	11.28%	26.32%
	PSO-SVR	AEV	0.468	0.403	0.457	0.475
		RER	12.50%	10.21%	9.57%	12.69%
Sample-2	SOA-BPNN	AEV	0.172	0.186	0.176	0.172
		RER	4.50%	4.75%	5.01%	4.50%
	SOA-SVR	AEV	0.223	0.158	0.230	0.180
		RER	5.84%	4.04%	6.54%	4.72%
	PSO-BPNN	AEV	0.196	0.164	0.446	0.196
		RER	5.13%	4.18%	12.69%	5.13%
	PSO-SVR	AEV	0.320	0.283	0.335	0.260
		RER	8.38%	7.24%	9.53%	6.80%
Sample-3	SOA-BPNN	AEV	0.016	0.132	0.205	0.016
		RER	0.43%	3.42%	6.40%	0.43%
	SOA-SVR	AEV	0.160	0.454	0.288	0.316
		RER	4.37%	11.78%	8.97%	8.63%
	PSO-BPNN	AEV	0.121	1.122	0.728	0.121
		RER	3.30%	29.10%	22.67%	3.30%
	PSO-SVR	AEV	0.117	0.497	0.386	0.310
		RER	3.20%	12.89%	12.02%	8.47%
Sample-4	SOA-BPNN	AEV	0.325	0.288	0.044	0.325
		RER	6.40%	6.92%	1.60%	6.40%
	SOA-SVR	AEV	0.406	0.368	0.174	0.376
		RER	7.99%	8.84%	6.29%	7.40%
	PSO-BPNN	AEV	0.732	0.555	0.537	0.732
		RER	14.42%	13.35%	19.43%	14.42%
	PSO-SVR	AEV	0.400	0.467	0.354	0.450
		RER	7.87%	11.22%	12.82%	8.87%

.

Finally, after calculating the arithmetic average of the Relative Error Rates (RER) for all test samples, the Average Error Rate (AER) evaluation metric was obtained, as shown in

Table 7. Calculation results of the average error rate.

Prediction Models	AER
SOA-BPNN	4.00%	6.25%	4.87%	4.00%
PSO-BPNN	12.29%	14.96%	16.52%	12.29%
SOA-SVR	6.20%	8.96%	7.72%	6.87%
PSO-SVR	7.99%	10.39%	10.99%	9.21%

. By comparing the AER results of the models, it was found that the SOA-BPNN model exhibited the lowest average error rate across the four pairs of Kansei imagery vocabulary, with values of 4.00%, 6.25%, 4.87%, and 4.00%, respectively. This indicates that the SOA-BPNN model demonstrated the best performance in predicting the Kansei imagery of product forms. Further analysis revealed that, from an optimization algorithm perspective, SOA showed higher efficiency and accuracy compared to PSO in improving the performance of Kansei imagery prediction models. In terms of overall prediction performance, the hierarchy of model effectiveness can be described as: SOA-BPNN > SOA-SVR > PSO-SVR > PSO-BPNN. This result suggests that, compared to other models, SOA-BPNN has significant advantages in terms of accuracy and robustness.

Furthermore, to further validate the reliability of the model accuracy comparison, the current research introduces MAE, RMSE, and R² as evaluation metrics to quantitatively assess the performance of the four Kansei imagery prediction models (SOA-BPNN, SOA-SVR, PSO-BPNN, and PSO-SVR). The smaller the MAE and RMSE values, the smaller the error between the predicted values and true values, indicating better prediction accuracy. Additionally, the closer the R² is to 1, the better the model’s goodness of fit. Based on Equation, the MRE, RMSE, and R² values for the four models are as follows: the MRE values are 3.11%, 4.05%, 4.15%, and 5.27%, respectively; the RMSE values are 0.221, 0.248, 0.324, and 0.395; and the R² values are 0.983, 0.956, 0.934, and 0.849. These results indicate that the SOA-optimized BPNN model demonstrates significant advantages in Kansei imagery prediction accuracy. Therefore, this model has high reliability and practical value when applied to ESUV front-end styling design.

3.7. Application of the Best Model

Based on the previous comparison of prediction errors across models, it is evident that the SOA-BPNN model performs best in Kansei imagery prediction. To avoid unnecessary iterations and subjective aesthetic interference during the front-end design development of the ESUV front-end styling, the present study applies the best-performing model to the Kansei imagery prediction phase in the early design process. Building on this foundation, the SD model is further introduced to assist in meeting the emotional needs within the styling imagery of the ESUV front-end. Ultimately, using Stable Diffusion technology, four ESUV front-end conceptual design schemes were rapidly generated, as shown in Figure 13.

Subsequently, the SOA-BPNN model was employed to perform Kansei imagery evaluation on each scheme, and real-time feedback and adjustments were made based on the evaluation results. At this stage, the design projects and element categories from Table 1 were used to encode the four design schemes, and this data was then imported into the model. After running the model, the Kansei imagery prediction values for the four schemes under the four sets of Kansei imagery vocabulary were obtained, as shown in

Table 8. Kansei imagery prediction for ESUV front-end styling design schemes.

Design Schemes	Kansei Imagery Prediction Values
	Y13	Y21	Y31	Y41
Scheme A	5.14	4.96	3.42	3.81
Scheme B	3.06	3.61	2.86	3.43
Scheme C	4.13	4.02	3.85	3.02
Scheme D	2.93	3.64	4.17	3.95

.

According to the prediction results in Table 8, the four ESUV front-end styling design schemes exhibit significant differences across various Kansei imagery dimensions. Scheme A scores the highest in the “futuristic” and “high-end” styling imagery categories, indicating that it evokes stronger consumer associations with forward-thinking and premium quality in its appearance. Scheme B, on the other hand, performs more prominently in the “low-end” imagery category, likely due to its simpler overall design language and the lack of advanced technological elements, which makes it less likely to convey a sense of sophistication or modernity. Scheme C, with its rounded surface transitions and fewer sharp angles, best showcases the “rounded” stylistic feature. Notably, only Scheme D performs exceptionally well in the “traditional” Kansei imagery. However, in the “futuristic” imagery, both Scheme B and Scheme D underperform, suggesting that further refinement of critical components such as the headlights and side mirrors is needed to better align with consumers’ expectations for advanced and technological designs. These results demonstrate that, through the precise evaluation of the relationship between ESUV front-end styling elements and Kansei imagery, the model provides rational and objective data support for automotive exterior design. This, in turn, helps the design team make targeted adjustments and improvements to emotional elements during the early concept development phase. Building on this foundation, it is possible not only to more accurately grasp consumers’ potential preferences in styling perception but also to significantly improve the accuracy of design decisions and the overall efficiency of product development. Furthermore, by combining Stable Diffusion with the optimal prediction model, this study successfully establishes a complete feedback loop from rapid concept generation to Kansei imagery evaluation and scheme optimization, greatly enhancing both the scientific rigor and timeliness of ESUV front-end styling design.

4. Discussion and Conclusions

Addressing the common challenges in existing research on Kansei Engineering in automotive front-end styling design, such as insufficient timeliness in emotion vocabulary extraction, strong subjective dependency, and relatively low accuracy of nonlinear prediction models, this study proposes an improved Kansei Engineering-based ESUV front-end styling design method. First, through Python web scraping technology, 34,697 real consumer reviews and 156 ESUV front-end samples were collected from platforms such as Autohome and Yiche. BTM was then applied to conduct an in-depth analysis of the preprocessed corpus, uncovering 48 emotion words across 4 themes. Subsequently, using the AHP method, the weight ranking clearly identified 4 representative sets of Kansei imagery vocabulary. Next, BPNN and SVR were used to construct the Kansei imagery prediction models, and SOA was employed for global optimization to significantly enhance the prediction accuracy of the models. Furthermore, PSO was introduced to make comparative improvements to these models in order to assess the relative advantages of SOA. Through average error rate analysis of the prediction results for 4 validation samples and 4 sets of vocabulary, experimental results demonstrated that the SOA-BPNN model achieved the highest Kansei imagery prediction accuracy, and thus, it was applied to the early design phase of ESUV front-end styling. In conclusion, this work developed a novel and systematic emotion-driven design framework for ESUV front-end styling, providing designers with a clearer development pathway, and addressing the scientific and rationality challenges often faced when traditional research and development are driven by subjective experience. Additionally, the application of Stable Diffusion (SD) for generative design showcases the integration of cutting-edge artificial intelligence (AI) tools into the design process. By enabling real-time creation and evaluation of multiple design concepts, SD provides automotive designers with a powerful tool to iterate and validate emotional resonance in design proposals quickly. This approach enhances the efficiency and scientific rigor of the design process, making it possible to align aesthetic outcomes more closely with consumer emotional preferences while reducing subjective bias.

To further improve this work, future research could enhance the emotional vocabulary extraction process by exploring more advanced techniques, such as deep learning, to reduce manual intervention. Additionally, incorporating temporal models like Long Short-Term Memory (LSTM) networks would allow for better tracking of evolving consumer emotional preferences. Lastly, expanding the model to consider additional design elements, such as color, material, and texture, would offer a more comprehensive framework for emotion-driven design. In conclusion, this study introduces a data-driven framework integrating Kansei Engineering, machine learning, and generative design, significantly advancing emotion-driven automotive design. The findings highlight the potential of combining these technologies to better align product design with consumer emotional preferences. The proposed framework offers both theoretical insights and practical tools for automotive designers, enhancing the scientific rigor and emotional resonance of ESUV front-end styling design.