Kansei Design Optimization of Torque Tool Inspection Cabinets Using XGBoost Prediction Models

Abstract

In the context of the aesthetic economy and the rapid development of digital intelligence, product design is increasingly required to address not only functional performance but also users’ emotional needs. However, due to the ambiguity and subjectivity of perceptual requirements, it remains difficult to accurately translate user emotions into specific design solutions. To address this challenge, this study proposes an integrated Kansei Engineering–machine learning framework for optimizing product design. First, user perceptual data are collected through questionnaires and interviews, and key perceptual imagery words are extracted using the Latent Dirichlet Allocation (LDA) model and factor analysis. Then, product design elements are systematically decomposed, and their relative importance is determined using the fuzzy analytic hierarchy process (FAHP). Based on this, a mapping relationship between perceptual imagery and design elements is established. Subsequently, the XGBoost model is employed to predict and optimize design element combinations. The optimized design schemes are further generated using AIGC technology and validated through eye-tracking experiments and subjective evaluations.The results show that the proposed method achieves high predictive accuracy (R² = 0.87) and significantly improves the emotional expression of product design. This study contributes to the integration of Kansei Engineering and machine learning by providing a data-driven approach for emotional design optimization, offering theoretical, practical, and strategic guidance for intelligent product design in industrial contexts.

1. Introduction

Against the backdrop of the rapid development of the digital and intelligent era, consumer behavior has gradually entered the era of the aesthetic economy [1], which emphasizes the transformation of consumption from functional utility to aesthetic and experiential value. In this context, product design is no longer limited to functional realization but increasingly emphasizes emotional experience and user value [2]. However, traditional product design is largely dominated by designers and is often influenced by their professional experience and personal preferences, which makes it difficult to fully satisfy the diverse needs of users.

User-centered design has been widely recognized as an effective approach to improving customer satisfaction, and concepts such as “user-centered” and “user-driven” have attracted increasing attention [3]. Nevertheless, user needs are inherently intangible, ambiguous, and semantically diverse, which makes it challenging to accurately identify and translate them into design solutions [4]. As a result, designers often face difficulties in achieving both efficiency and accuracy during the design process.

Meanwhile, artificial intelligence technologies, with their strong capabilities in data processing and pattern recognition, have been increasingly integrated into product design to improve efficiency and support decision-making [5]. However, how to effectively combine AI technologies with user-centered design to accurately capture and utilize user emotional needs remains a critical research issue.

Although Kansei Engineering has been widely applied to analyze users’ emotional perceptions in product design [6], existing studies can generally be categorized into semantic-oriented approaches, quantitative decision-making methods, and data-driven predictive modelling techniques [7].

Moreover, these approaches are often applied independently, and most studies rely primarily on subjective evaluations with limited incorporation of objective validation methods. As a result, there is a lack of an integrated framework that systematically links emotional semantic extraction, quantitative weighting of design elements, predictive modelling, and objective validation for design optimization.

To this end, this study aims to explore users’ emotional needs with the assistance of AI technologies and data-driven design methods [8], improve the specificity and accuracy of emotional expression, and enhance the effectiveness of design models.

Accordingly, this study seeks to explore how user emotional perceptions can be quantitatively modelled and effectively translated into optimized design solutions. To further clarify the research focus, this study addresses the following research questions:

RQ1: How can user Kansei perceptions be effectively extracted and structured into representative perceptual dimensions?

RQ2: How can multi-dimensional design elements be systematically identified and quantitatively weighted under uncertainty?

RQ3: How can the mapping relationship between design elements and user emotional perceptions be accurately modelled and predicted using machine learning techniques?

RQ4: How can the predicted design solutions be efficiently generated and their emotional communication effectiveness objectively validated?

Taking the industrial design project of a torque-setting tool inspection system as a case study, this research integrates Kansei Engineering and AI-based methods to address the challenges of emotional need extraction and its application in design. Specifically, a quantitative modelling approach is adopted to construct the relationship between design elements and user emotional responses, thereby providing efficient, accurate, and systematic support for design decision-making. The proposed framework not only contributes to the methodological development of emotion-driven design but also offers empirical evidence for improving the effectiveness of product design in the context of intelligent manufacturing.

2. Research Methodology

This study adopts a quantitative modelling approach to systematically investigate the relationship between product design elements and users’ Kansei perceptions [9]. In this context, the research method defines the overall analytical paradigm, while the research methodology specifies the concrete procedures and techniques applied. Quantitative modelling is particularly suitable for this research because it enables the transformation of subjective emotional responses into measurable variables and supports the establishment of relationships between design features and perceptual outcomes. Compared with purely qualitative approaches, this method provides stronger analytical robustness and interpretability.

It is important to distinguish between the research method and the research methodology. In this study, the research method refers to the overall analytical paradigm, namely quantitative modelling, which provides the theoretical foundation for the study. In contrast, the research methodology refers to the specific techniques and procedures employed to implement this approach. The methodology in this research integrates multiple techniques. Latent Dirichlet Allocation (LDA) is used for extracting Kansei image semantics [10]. Factor analysis is applied for dimensionality reduction. The fuzzy analytic hierarchy process (FAHP) is used for determining the weights of design elements [11], and machine learning models (including XGBoost and SVR) are used for predictive modelling [12]. In addition, eye-tracking experiments are used to validate the perceptual effectiveness of the design schemes.

The overall research framework follows a structured multi-stage process. First, Kansei image data are collected and processed using LDA to identify representative perceptual dimensions. Second, cabinet design elements are identified and deconstructed into quantifiable variables. Third, FAHP is employed to evaluate and weight the relative importance of these design elements based on expert judgments. Fourth, statistical analysis and machine learning models are applied to establish the mapping relationship between design elements and Kansei perceptions, with model performance evaluated through comparative analysis. Finally, design solutions are generated and visualized using AIGC techniques and validated through eye-tracking experiments to assess their perceptual effectiveness. This integrated methodology combines perceptual analysis, multi-criteria decision-making, and data-driven modelling within a unified framework, enhancing both the interpretability and practical applicability of Kansei-based product design.

3. Research on Cabinet Design Based on Kansei Engineering

The research on Kansei Engineering began in the late 1970s and early 1980s. It is regarded as a consumer-oriented product development methodology that effectively captures users’ emotional needs. Its core idea is to transform these emotional needs into specific design elements and integrate them into the product development process [13]. The identification of emotional needs, as the primary step in Kansei Engineering, currently mainly quantifies user needs by distinguishing multiple emotional adjectives [14].

Manual collection and data-driven approaches are the two primary methods for extracting users’ perceptual needs [15]. As shown in Figure 1, manual collection analyzes user reports, literature, and physiological data through methods such as interviews and questionnaires. This method is easy to operate and can guide in-depth thinking, which is conducive to obtaining accurate needs. However, there are problems of data divergence and uneven quality, which can be optimized by standardizing the pre-process (such as unifying interview outlines and introducing emotion dictionaries). Data mining uses technologies such as front-end data embedding and web crawlers to intelligently and efficiently capture available data from real-time online content. After preprocessing and standardization, perceptual image vocabularies can be extracted. A large number of studies have used it as an effective support for product design [16], providing assistance for accurately grasping users’ perceptual needs.

After obtaining a large amount of product requirement information, due to the irregularity and complex composition of the data, data cleaning work needs to be carried out, covering aspects such as text screening, word segmentation, and word frequency statistics. Given the professional nature of product requirement data, general tools are often inadequate. To preprocess the collected textual data, Chinese word segmentation is first required. In addition, it has been widely applied in natural language processing tasks such as sentiment analysis and text mining [17]. Jieba is a widely used Chinese word segmentation tool that achieves high efficiency through prefix dictionaries and probabilistic models. It supports custom dictionaries and optimizes ambiguous segmentation through Hidden Markov Models, thereby improving segmentation quality. After completing data preprocessing, clustering and dimensionality reduction analyses are conducted on the perceptual vocabulary [18]. Based on data similarity, the chaotic data is automatically divided into multiple classes with high internal similarity, which helps researchers quickly understand the distribution characteristics of the data [19] and reduces the cognitive burden. Common text clustering analysis methods are shown in

Table 1. Statistical table of text clustering analysis methods.

Clustering Method	Application Scenarios	Advantages	Disadvantages
K-Means Clustering	Structured text, large-scale data	Fast speed	Susceptible to initial points
Hierarchical Clustering	Small-scale data, hierarchical analysis	Strong interpretability	High computational complexity
DBSCAN	Data with much noise, unknown number of categories	Can discover outliers	Depends on hyperparameters
LDA	Topic modeling, text classification	Suitable for semantic analysis	Requires parameter optimization

.

After completing data preprocessing, clustering and dimensionality reduction analyses are conducted on the perceptual vocabulary. Based on data similarity, the original data are automatically grouped into categories with high internal similarity, which helps researchers understand the distribution characteristics of the data and reduces cognitive burden. Among various methods, the Latent Dirichlet Allocation (LDA) topic model performs well in text analysis and semantic extraction [20]. It identifies latent semantic structures by representing documents as mixtures of topics and topics as distributions over words. Furthermore, factor analysis is introduced to achieve dimensionality reduction by transforming observed variables into a smaller number of latent factors. This process reduces data dimensionality and facilitates the identification of key perceptual image features.

Meanwhile, the integration of generative AI technology brings new breakthroughs in the mining of perceptual needs [21]. Research shows that combining generative models such as GPT-4 can enhance the efficiency and depth of emotional need extraction. For example, AI-assisted generation of design sketches and optimization of the semantic association of topic words promote the integration of “data-driven design” and “human–machine collaboration” [22].

Through this series of analysis and processing, it is possible to deeply mine users’ emotional and psychological needs, systematically construct the experience related to users’ emotional cognition, extract perceptual image vocabulary, and provide strong data support for personalized design and product optimization.

3.1. Extraction of Kansei Images of Cabinets

In the “Industrial Design Project of a Certain Torque Tool Detection System”, to explore the users’ perceptual needs for equipment products, the research directly collected users’ subjective feedback through interviews and questionnaires, and supplemented the samples with the help of literature surveys and AI to form a multi-dimensional corpus. In the data preprocessing stage, the data quality was improved by cleaning noisy data and unifying the format (such as letter case and digital standardization). The Jieba word-segmentation tool was used for word segmentation-processing, and efficient Chinese word-segmentation was achieved in combination with the Python3.13.1 ecosystem: the pandas and jieba libraries were installed to build the environment. After inputting the corpus, a general stop-word list and the NTUSD sentiment dictionary of National Taiwan University were set, and finally the word-segmentation results were output. An example of word-segmentation of the original data is shown in

Table 2. Word segmentation example table.

Raw Data	Word Segmentation Results
It is hoped that the appearance of the equipment is simple and elegant, with harmonious color matching, creating a professional and reliable impression. Meanwhile, the tactile sensation of the product should convey a sense of warmth or high-tech, enhancing the user experience, lowering the operational threshold for users, and adapting to multiple application scenarios through modular design.	Equipment, appearance, simple, elegant, color, matching, harmonious, professional, reliable, impression, product, tactile sensation, convey, warmth, high-tech, enhance, experience, modular design, scenario, equipment, robust, structure, industrial styling, tough, reliable, quality, detail optimization, increase, technology, precision, good, compressive resistance, durability, extended periods, high-load conditions, stable operation, affect, performance
The equipment should express its tough and reliable quality through a robust structure and industrial styling, while increasing the sense of technology and precision through detail optimization. Additionally, it should possess good compressive resistance and durability, enabling stable operation over extended periods under high-load conditions without affecting performance.	Equipment, appearance, simple, elegant, color, matching, harmonious, professional, reliable, impression, product, tactile sensation, convey, warmth, high-tech, enhance, experience, modular design, scenario, equipment, robust, structure, industrial styling, tough, reliable, quality, detail optimization, increase, technology, precision, good, compressive resistance, durability, extended periods, high-load conditions, stable operation, affect, performance

.

When conducting cluster analysis on segmented text based on the LDA topic model, first install libraries such as os, pandas, scikit-learn, and pyLDAvis. Extract text features and train the LDA model, and identify topics through keywords. Subsequently, use pyLDAvis to visualize the topic distribution. Combine the topic spacing graph (as shown in Figure 2) and the perplexity curve (a significant inflection point appears when the number of topics is 4, as shown in Figure 3) to determine the optimal number of topics.

The first category of perceptual image words focuses on the depiction of visual effects and aesthetic atmosphere; the second category pays more attention to the creation of texture and environmental atmosphere; the third category of words conveys the perception and experience of quality and details; the fourth category emphasizes the practicality and long-term stability of the product; the fifth category describes unique styles and personality characteristics. When the number of LDA topics K = 5, the topic bubbles only slightly overlap, and the classification of perceptual image words is clearer. At the same time, combined with the inflection point characteristics of the perplexity curve (a significant inflection point appears when K = 4, but the model has a better discrimination of perceptual needs when K = 5), finally determine K = 5 to balance topic interpretability and model stability. Set the number of iterations to 10 (passes = 10) to ensure that the model fully converges.

To reveal text similarity and thematic distribution after word segmentation—and to improve the efficiency of downstream tasks—a cluster analysis of the text is required. The Latent Dirichlet Allocation (LDA) topic model is widely employed for text clustering due to its flexibility and stability; in the analysis of product reviews, it can rapidly identify core user concerns and requirements while generating interpretable keywords for visual analysis, which aligns with real-world scenarios. Furthermore, Python offers a rich suite of Natural Language Processing (NLP) libraries conducive to efficient LDA implementation; thus, this phase was conducted within a Python environment. By analyzing the inter-topic distance map and the model perplexity curve in conjunction with the clustering results, five categories of user perceptual needs were preliminary identified, as shown in

Table 3. Perceptual demand classification table.

Category	Sensory Image Vocabulary
First-type	Grand, Modern, Concise, Comfortable
Second-type	Detailed, Technological, Textured, Futuristic, Delicate
Third-type	High-end, Industrial Style, Professional, Exquisite
Fourth-type	Durable, Steady, Long-lasting, Reliable
Fifth-type	Smooth, Characteristic, Affable

.

The paper uses the HIT-Bert pre-trained model specially developed for Chinese to vectorize perceptual imagery words. This model is deeply optimized for Chinese semantics and grammar. After being trained on a large amount of corpora, the extracted word vectors can be well transferred to downstream tasks such as text analysis.

To ensure that the word-vector data is applicable to the factor analysis model, KMO and Bartlett’s sphericity test need to be carried out. The closer the KMO value is to 1, the more suitable it is for factor analysis. When the p-value of Bartlett’s sphericity test is less than 0.05, the correlation between variables is significant, which is suitable for analysis. By analyzing the perceptual imagery words of five types of themes with SPSS 23.0, the results show that the KMO values of all five types of themes are greater than 0.7, and the p-values are all less than 0.001. This indicates that the words of each theme have sufficient correlation with factor analysis, and the correlation between variables is significant, making them suitable for factor analysis. As shown in

Table 4. KMO and Bartlett’s test.

Theme	KMO	Approx. ${\mathit{\chi }}^2$	df	p-Value
First-type Theme	0.877	603.650	6	<0.001
Second-type Theme	0.912	885.913	10	<0.001
Third-type Theme	0.853	663.249	6	<0.001
Fourth-type Theme	0.870	644.098	6	<0.001
Fifth-type Theme	0.787	425.132	3	<0.001

, all KMO values exceed 0.7 and the p-values of Bartlett’s test are below 0.001, indicating that the data are suitable for factor analysis.

Factor analysis was performed on various theme image vocabulary using the principal component analysis method. The total variance explanation describes the proportion of the variance explained by each factor to the total variance. The number of factors was determined by selecting the number of initial eigenvalues greater than one. Factor loadings were used to judge the correlation and contribution degree between variables and factors, and the component matrix shows the loadings of variables on factors. The word-vector data were input into SPSS for processing and analysis. For the first-type theme, the number of components with initial eigenvalues greater than one was one, which could explain 93.749% of the variance, and “concise” was selected as the representative vocabulary. Similarly, for the second type, “technological sense” was selected; for the third type, “professional” was selected; for the fourth type, “reliable” was selected; for the fifth type, “characteristic” was selected. Finally, vocabulary that meets users’ perceptual needs was obtained, providing reliable data for establishing a mapping relationship between perceptual images and design elements.

3.2. Identification and Deconstruction of Cabinet Design Elements

In an industrial design project of a certain torque tool detection system, according to the user’s requirements, the preliminary functional requirements of the product can be summarized from the proposed technical agreement as being simple and beautiful, ergonomic, and sustainable. For the first-level design element identification, the functional requirements are the primary content for design discussion and confirmation. After summarizing the sub-requirements of the preliminary functional requirements, three first-level design elements, namely design features, color configuration, and material texture, are summarized using the function-form comparison method and tabulated to provide structural guidance for the design, as shown in

Table 5. Function-form comparison table.

Clustering Method	Application Scenarios	Advantages	Disadvantages
K-Means Clustering	Structured text, large-scale data	Fast speed	Susceptible to initial points
Hierarchical Clustering	Small-scale data, hierarchical analysis	Strong interpretability	High computational complexity
DBSCAN	Data with much noise, unknown number of categories	Can discover outliers	Depends on hyperparameters
LDA	Topic modeling, text classification	Suitable for semantic analysis	Requires parameter optimization

.

Next, the deconstruction of secondary design elements is carried out (see in

Table 6. Deconstruction matrix of form comparison method.

Deconstruction Scheme	Design Features				Color Configuration		Material Texture
	Morphological Features	Layout Features	Structural Features	Connection Features	Primary Color	Secondary Color	Material	Surface Treatment
Deconstruction 1	Linear	Symmetrical	Large Rounded	Corners	Welding High-Lightness Gray + Low-Lightness Gray	Yellow	Cast Iron	Polishing
Deconstruction 2	Curvilinear	Asymmetrical	Small Rounded	Corners Screw Fixing	Low-Lightness Gray + High-Lightness Gray	Orange-Yellow	Aluminum Alloy Chrome	Chrome Plating
Deconstruction 3	Combined Right	Angle	Gluing	High-Lightness Gray	Orange	Aviation Aluminum	Wire Drawing
Deconstruction 4					Low-Lightness Gray	Aviation Blue
Deconstruction 5					Sky Blue

). There are significant differences in design characteristics between equipment products and consumer products. The design characteristics of this project revolve around the internal power and mechanical structure. The layout balance needs to be considered, with straight lines as the main element combined with curves. In terms of color matching, black and white are mainly used to convey a sense of reliability, and auxiliary and bright colors are used to guide operations. The materials are within the scope of metal materials. Such identification and deconstruction lay the foundation for creating high-quality products.

3.3. Analysis and Evaluation of Kansei Images and Design Elements

3.3.1. Fuzzy Analytic Hierarchy Process

To provide accurate guidance for the subsequent prediction model, it is necessary to evaluate the relative importance of design elements while considering the uncertainty and vagueness inherent in human judgments. The traditional Analytic Hierarchy Process (AHP) has been widely used for multi-criteria decision-making [23]. Within the quantitative modelling framework adopted in this study, the Fuzzy Analytic Hierarchy Process (FAHP) is employed as a specific research methodology to handle ambiguity in expert judgments. As an extension of AHP, the Fuzzy Analytic Hierarchy Process (FAHP) integrates fuzzy logic with hierarchical analysis, thereby improving the robustness and reliability of evaluation results [24]. Due to its systematic structure and enhanced capability to handle ambiguity, FAHP has been widely applied in product design and optimization [25].

The overall FAHP procedure in this study consists of four main steps, including pairwise comparison construction, fuzzy aggregation, defuzzification, weight calculation, and consistency verification. A panel of domain experts with relevant experience in industrial design and product development was invited to perform the pairwise comparisons. This approach allows uncertainty in human judgment to be explicitly modelled, thereby enhancing the robustness of the evaluation results [26].

The general steps in the research process of the Fuzzy Analytic Hierarchy Process are as follows:

A panel of experts was invited to participate in the evaluation process. All experts had relevant backgrounds in industrial design or product development and more than five years of professional experience in equipment design. The panel was selected based on domain relevance and practical expertise, ensuring the reliability and validity of the evaluation results.

• Conduct pairwise comparisons of the core design elements at the same level. Experts express their judgments using linguistic variables, which are then transformed into corresponding fuzzy numbers. These linguistic variables are represented by triangular fuzzy numbers (TFNs), defined as $(l,m,u)$.
• Based on expert judgments, the fuzzy pairwise comparison matrix is constructed as:

$$\begin{array}{c}\hfill S_k=\left[\begin{array}{cccc}{\tilde{b}}_{11k}& {\tilde{b}}_{12k}& \cdots & {\tilde{b}}_{1mk}\\ {\tilde{b}}_{21k}& {\tilde{b}}_{22k}& \cdots & {\tilde{b}}_{2mk}\\ ⋮& ⋮& \ddots & ⋮\\ {\tilde{b}}_{m1k}& {\tilde{b}}_{m2k}& \cdots & {\tilde{b}}_{mmk}\end{array}\right]\end{array}$$

(1)

Among them, ${\tilde{b}}_{ij}$ represents the fuzzy emotional preference of expert between attributes i and j. Thereafter, Equations (2)–(4) are used to aggregate the evaluation data of experts.

$$\begin{array}{c}\hfill {\tilde{b}}_{ij}=(L_{ij},M_{ij},U_{ij}),{\tilde{b}}_{ji}^{-1}=\left({\displaystyle \frac{1}{U_{ij}}},{\displaystyle \frac{1}{M_{ij}}},{\displaystyle \frac{1}{L_{ij}}}\right)\end{array}$$

(2)

$$\begin{array}{c}\hfill L_j=\underset{i}{min}\left(l_{ij}\right),M_j={\displaystyle \frac{1}{n}}\sum _{i=1}^n{m}_{ij},U_j=\underset{i}{max}\left(u_{ij}\right)\end{array}$$

(3)

$$\begin{array}{c}\hfill b_{ij}={\displaystyle \frac{L_{ij}+M_{ij}+U_{ij}}{3}}\end{array}$$

(4)

Among them, $b_{ij}$ represents the aggregated fuzzy number, and $b_{ij}$ represents the accurate value after defuzzification.

3.

Based on the defuzzified comparison matrix, the weights are calculated using the eigenvector method and then normalized to obtain the final weight vector.

$$\begin{array}{c}\hfill \mathbf{A}\mathbf{W}={\lambda }_{max}\mathbf{W}\end{array}$$

(5)

$$\begin{array}{c}\hfill \mathbf{A}=\left[\begin{array}{cccc}{b}_{11}& b_{12}& \cdots & b_{1m}\\ b_{21}& b_{22}& \cdots & b_{2m}\\ ⋮& ⋮& \ddots & ⋮\\ b_{m1}& b_{m2}& \cdots & b_{mm}\end{array}\right]\end{array}$$

(6)

Among them, $\mathbf{A}$ is a matrix of size $m\times m$ containing m attributes. Here, ${\lambda }_{max}$ is the maximum eigenvalue of matrix $\mathbf{A}$, and $\mathbf{W}$ represents its corresponding eigen-vector. In this paper, this eigen-vector is regarded as the importance weight.

4.

Verify the consistency of the matrix. The consistency check is performed after defuzzification to ensure the rationality of expert judgments. According to the transitivity principle, If $C_1$ is superior to $C_2$ and $C_2$ is superior to $C_3$, then $C_1$ must be superior to $C_3$. The following consistency index (C.I.) and consistency ratio (C.R.) are used to determine the consistency of the decision-making quality.

$$\begin{array}{c}\hfill \mathrm{C}.\mathrm{I}.={\displaystyle \frac{{\lambda }_{max}-n}{n-1}}\end{array}$$

(7)

$$\begin{array}{c}\hfill \mathrm{C}.\mathrm{R}.={\displaystyle \frac{\mathrm{C}.\mathrm{I}.}{\mathrm{R}.\mathrm{I}.}}\end{array}$$

(8)

Among them, C.I. is the inconsistency index (the closer it is to zero, the better the consistency), and R.I. is the random index. The random indices for different orders are shown in

Table 7. Random consistency index table.

Order	2	3	4	5	6	7
R.I.	0	0.58	0.90	1.12	1.24	1.32

. When $\mathrm{C}.\mathrm{R}.⩽0.1$, it can be considered that the decision-making process is acceptable; otherwise, the pairwise comparison matrix should be re-evaluated and adjusted to improve consistency.

To enhance robustness, uncertainty in expert judgments was incorporated using triangular fuzzy numbers, and aggregation was applied to reduce individual bias. A sensitivity analysis was conducted by slightly varying input judgments, showing that the ranking of key design elements remains stable.

In addition, a bootstrap resampling method was employed to estimate confidence intervals for the weights [27]. The expert judgments were resampled with replacement, and the FAHP procedure was repeated for each sample to obtain a distribution of weights. Based on this distribution, mean values were calculated to represent the central tendency of the weights. Percentile-based confidence intervals (95%) were further derived to quantify the uncertainty associated with the estimated weights. This approach provides a statistical measure of uncertainty and enhances the robustness and reliability of the evaluation results.

3.3.2. Analysis and Evaluation

To ensure the scientific and fair evaluation of design elements and provide data support for constructing a reasonable and effective mapping relationship model, the Fuzzy Analytic Hierarchy Process (FAHP) is used to determine the weights of evaluation indicators at each level. In this paper, five experts with rich experience in the fields of industrial and product design are invited to evaluate the relative importance of indicators in a pairwise comparison manner. Each expert forms four groups of judgment matrices. For matrices of the same category, the fuzzy hierarchy matrix is constructed using Formulas (1)–(4), forming the fuzzy hierarchy judgment matrix (

Table 8. Pair-wise judgment matrix of fuzzy hierarchy criterion for layer.

Criterion Layer	Design Feature	Color Configuration	Material Texture
Design Feature	1	6.9	5.7
Color Configuration	0.15	1	1.2
Material Texture	0.2	0.83	1

) of the criterion-layer indicators and the fuzzy hierarchy judgment matrix (

Table 9. Fuzzy hierarchy judgment matrix of scheme layer (design features).

Alternative Layer	Morphological Feature	Layout Feature	Structural Feature	Connection Feature
Morphological Feature	1	4.75	3.5	2.6
Layout Feature	0.21	1	2.24	1.5
Structural Feature	0.29	0.45	1	1.75
Connection Feature	0.39	0.67	0.57	1

,

Table 10. Fuzzy hierarchy judgment matrix of scheme layer (color configuration).

Alternative Layer	Main Color	Auxiliary Color
Main Color	1	6.33
Auxiliary Color	0.16	1

and

Table 11. Fuzzy hierarchy judgment matrix of scheme layer (material texture).

Alternative Layer	Material Surface	Treatment Process
Material	1	3.95
Surface Treatment Process	0.26	1

) of the scheme-layer indicators.

The importance of evaluation indicators is determined through progressive pairwise comparisons. For example, at the criterion level, design features are compared with material textures, and then the importance of each internal element is further compared in detail. To ensure the scientific and objective nature of the evaluation, by integrating expert judgments and mathematical logic, it is necessary to conduct consistency verification on the fuzzy hierarchical judgment matrices at all levels. The relevant values of the judgment matrices at each level are calculated using Formulas (5)–(8), and the results are shown in

Table 12. Consistency test.

Hierarchy	Maximum Eigenvalue (${\mathit{\lambda }}_{max}$)	C.I.Value	R.I.Value	C.R.Value	Consistency Check Result
Criterion Layer	3.076	0.038	0.580	0.066	Passed
Alternative Layer (Design Features)	4.250	0.083	0.900	0.0926	Passed
Alternative Layer (Color Configuration)	2.006	0.006	0.000	0.000	Passed
Alternative Layer (Material Texture)	2.013	0.013	0.000	0.000	Passed

. The indicator judgment matrix values of design features, color configuration, and material textures are all lower than 0.1, indicating good consistency and ensuring the rationality of the judgment matrices. Although the CR value of the alternative layer (design features) is relatively close to the threshold (0.1), it still falls within the acceptable range. To further ensure robustness, sensitivity analysis and bootstrap-based confidence interval estimation were conducted.

The weights of all evaluation indicators for the evaluation of equipment-type design elements are shown in

Table 13. Evaluation weights of design elements for equipment-type products.

Evaluation Level	Scheme Category	Scheme Name	Hierarchical Weight (%)	Total Weight (%)
Criterion layer		Design features	75.2
		Color configuration	12.5
		Material texture	12.3
Scheme layer	Design features	Morphological features	53.7	$40.38\pm 2.10$
		Layout features	19.2	$14.44\pm 1.25$
		Structural features	14.3	$10.75\pm 1.05$
		Connection features	12.8	$9.63\pm 0.95$
	Color configuration	Main color	86.3	$10.79\pm 0.85$
		Auxiliary color	13.7	$1.71\pm 0.25$
	Material texture	Material	79.6	$9.79\pm 0.80$
		Surface treatment process	20.4	$2.51\pm 0.30$

. Among them, the total weight is calculated by multiplying the weight of this level by the weight of the corresponding criterion-level indicator. The determination of these weights combines subjective experience and objective logic, providing a data benchmark for the design process.

Generally speaking, through a multi-step and multi-dimensional weight distribution in a progressive manner, on the premise of ensuring fairness and transparency, the importance of each evaluation index is clarified, the key indicators that need to be concerned in the design of equipment-type products are revealed, a unified and standardized data benchmark is provided for subsequent research, and a solid foundation is laid for in-depth analysis and the construction of prediction models. This process enhances the transparency and reproducibility of the evaluation results.

4. Prediction Model and Design Evaluation Based on Kansei Images

This study adopts quantitative modelling as the overall research method, while specific methodologies, including LDA, FAHP, SVR, and XGBoost, are employed as analytical tools to implement the research framework.

The overall research process is structured as follows. First, user perceptual needs are extracted and key design elements are identified based on Kansei image analysis. Second, the relationships between perceptual images and design elements are established, and the importance weights of design elements are determined using FAHP. Third, multiple prediction models are constructed to capture the relationship between design elements and perceptual evaluation values. Finally, the optimal design scheme is generated and validated through experimental methods.

Based on the in-depth analysis of “a certain torque tool detection system” in the previous stage, this study has accurately captured the users’ perceptual needs and refined the key design elements. In this section, based on the mapping relationship between perceptual images and design elements, multiple algorithm prediction models are established for comparative analysis, and the model with the best performance is selected for the design of “a certain torque tool detection system”.

This section will discuss the application of two algorithms, SVR and XGBoost, in predicting the perceptual image evaluation values. SVR is based on the support vector machine and is suitable for the prediction of continuous variables in nonlinear problems and small sample data [12]. XGBoost is based on the gradient boosting framework and can be quickly trained and provides excellent prediction ability [21]. These algorithms have been widely applied in data-driven design and predictive modelling tasks [27]. To ensure that these models can accurately predict the perceptual image scores, data preprocessing is required. Data preprocessing includes data cleaning, feature engineering, and data standardization. The specific process is as follows: first, collect the original data and handle missing values and outliers; then, conduct feature engineering to extract the position features related to the design element sequence; next, normalize or standardize the data according to the model requirements. SVR requires normalization, while XGBoost directly inputs the feature set; finally, divide the data set. Through preprocessing, standard input data is provided for the model to ensure the effectiveness of training.

4.1. Prediction Model XGBoost

The core idea of XGBoost (Extreme Gradient Boosting) is based on the Gradient Boosting framework. Gradient Boosting is an ensemble learning method that makes predictions by combining multiple weak learners into a strong one. The objective function of XGBoost consists of two parts: the Loss Function, which measures the difference between the predicted values and the actual values of the current model, and the Regularization Term, which controls the complexity of the tree to prevent overfitting. The formula for its objective function is:

$$\begin{array}{c}\hfill \mathcal{L}\left(\theta \right)=\sum _{i=1}^{N}L(y_i,{\widehat{y}}_i)+\sum _{k=1}^T\Omega \left(f_k\right)\end{array}$$

(9)

where $\mathcal{L}\left(\theta \right)$ is the overall objective function, L is the loss function, $y_i$ is the actual value, ${\widehat{y}}_i$ is the predicted value, $\Omega \left(f_k\right)$ is the regularization term representing the complexity of the k-th tree, N is the number of samples, T is the number of trees, and $f_k$ is the k-th tree. In regression problems, a common loss function is:

$$\begin{array}{c}\hfill L(y_i,{\widehat{y}}_i)={\displaystyle \frac{1}{2}}{(y_i-{\widehat{y}}_i)}^2\end{array}$$

(10)

The form of the regularization term is usually:

$$\begin{array}{c}\hfill \Omega \left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda \sum _{j=1}^T{w}_j^2\end{array}$$

(11)

where $\gamma$ controls the number of leaves, $\lambda$ is the L2 regularization coefficient, T is the number of leaves in the k-th tree, and $w_j$ is the weight of the j-th leaf. To improve the training speed, XGBoost adopts an efficient optimization method, which is to perform a second-order approximation of the objective function through Taylor expansion. Assume that the output of the current model is ${\widehat{y}}_t$. Each iteration will train a new tree and update the output of the current model. The second-order approximation form of the Taylor expansion of the objective function is:

$$\begin{array}{c}\hfill \mathcal{L}\left(t\right)\approx \sum _{i=1}^N\left[g_{i}f\left(x_i\right)+{\displaystyle \frac{1}{2}}{h}_{i}f{\left(x_i\right)}^2\right]+\Omega \left(f\right)\end{array}$$

(12)

where $g_i={\displaystyle \frac{𝜕L(y_i,{\widehat{y}}_t)}{𝜕{\widehat{y}}_t}}$ is the first-order derivative (gradient), $h_i={\displaystyle \frac{{𝜕}^{2}L(y_i,{\widehat{y}}_t)}{𝜕{\widehat{y}}_t^2}}$ is the second-order derivative (Hessian), and $f\left(x\right)$ is the correction value of the prediction.

4.2. Selection of Prediction Model

This section discusses the performance and parameter tuning process of SVR and XGboost models in predicting the Kansei image scores of different design element sequences. The training adopts the five-fold cross-validation method, and Mean Squared Error Mean Squared Error ((MSE), RootRMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²) are selected as the evaluation indicators of the models. Among them, MSE and RMSE are used to measure the magnitude of errors, MAE calculates the average of the absolute values of prediction errors, MAPE represents the error percentage, and R² is used to measure the model’s ability to fit the data—the closer it is to one, the better the model performance.

Parameter Tuning of SVR Model: The main hyperparameters include regularization parameter (C), insensitive loss function (epsilon), kernel function type (kernel), and kernel coefficient (gamma). The parameter adjustment steps are as follows: first determine the kernel function type, then adjust the values of C, epsilon, and gamma parameters respectively. After multiple adjustments, the optimal parameter combination is shown in

Table 14. SVR model parameters.

Parameter Name	Parameter Value
Training Time	2.69 s
Data Shuffling	Yes
Cross-Validation	5
C	1
Epsilon	0.5
Kernel	rfb
Gamma	scale

, and the model evaluation results are shown in

Table 15. SVR model evaluation results.

	MSE	RMSE	MAE	MAPE	R2
Training Set	0.057	0.408	0.384	5.976%	0.772
Cross-Validation Set	0.070	0.446	0.506	6.849%	0.744
Test Set	0.062	0.593	0.490	6.250%	0.761

. The R² values on the training set, cross-validation set, and test set are 0.772, 0.744, and 0.761 respectively, indicating good prediction accuracy.

Parameter Tuning of XGboost Model: The main hyperparameters include the number of base learners $\left(n_{e}stimators\right)$, maximum tree depth $\left(ma{x}_{d}epth\right)$, etc. The tuning process first adjusts parameters such as the number of base learners to control model complexity, then reduces overfitting risk through sample sampling ratio, and finally adjusts regularization parameters. The combination optimal parameter is shown in

Table 16. XGboost model parameters.

Parameter Name	Parameter Value
Training time	3.25 m
Data shuffling	Yes
Cross-validation	5
Number of base learners	50
Maximum depth of the tree	3
Learning rate	0.05
Minimum sample weight of leaf nodes	2
L1 regularization	0.1
L2 regularization	1
Sample sampling ratio	0.9
Feature sampling ratio used for each tree	0.8

, and the model evaluation results are shown in

Table 17. XGboost model evaluation results.

	MSE	RMSE	MAE	MAPE	R2
Training Set	0.0263	0.1622	0.1184	2.3753%	0.949
Cross-Validation Set	0.0385	0.2879	0.1740	2.7920%	0.885
Test Set	0.0449	0.1869	0.2231	2.8729%	0.870

. The R² values on the training set, cross-validation set, and test set are 0.949, 0.885, and 0.870 respectively, indicating high accuracy and generalization ability.

Through the aforementioned evaluation process, it can be concluded that the XGboost model performs better in most performance evaluation indicators. Meanwhile, the test data prediction graphs of the two models shown in Figure 4 (with the abscissa representing sample serial numbers and the ordinate representing actual scores and predicted scores) intuitively demonstrate the high accuracy of the XGboost model in predicting Kansei image scores. The matching degree between actual values and predicted values highlights its trend of outperforming the SVR model in various evaluation indicators, reflecting the excellent performance of the XGboost model in handling nonlinear data.

The predicted sequences with the highest and lowest scores under different Kansei images, along with their corresponding design languages and score values, are clearly defined. The experimental results support the aforementioned analysis conclusions, proving the excellent performance of the XGboost algorithm. This emphasizes the importance of selecting appropriate algorithms and provides guidance for subsequent design practices.

4.3. Comprehensive Analysis of Prediction Results

In the section, the XGboost model was used to predict the scores of different combinations of design element sequences under five categories of perceptual images respectively. The prediction results are shown in

Table 18. Prediction situations under various perceptual images.

Perceptual Image	Sequence with the Highest Predicted Score	Score	Sequence with the Lowest Predicted Score	Score
Concise Sense	[1-2-2-3-3-1-3-4]	5.72	[2-1-2-4-1-5-1-1]	2.48
Technological	[3-1-3-4-2-4-3-5]	5.85	[2-2-1-2-4-3-1-1]	2.88
Professional	[1-1-2-2-4-4-1-4]	5.87	[2-2-1-3-1-5-2-2]	2.52
Reliable	[1-1-3-1-4-4-1-5]	5.66	[2-2-1-3-3-3-2-2]	2.41
Characteristic	[2-2-1-4-3-5-3-3]	5.72	[1-1-2-2-4-2-2-1]	2.25

.

In summary, the experimental results predicted by the XGBoost model support the previous analysis conclusions, further proving the excellent performance of the XGBoost algorithm. It also emphasizes the importance of choosing an appropriate algorithm for model training from the side. The above experimental results provide more precise guidance for the subsequent design practice, making the purpose of the design activities more clear.

5. Detection, Optimization, and Verification of the Kansei Design Offets Cabin with Torque-Setting Tools

5.1. Kansei Design of Cabinets Based on AIGC

5.1.1. Application Design of AIGC Technology

Based on the aforementioned research prediction results, the perceptual images are transformed into specific design elements. After extracting the design language, it is input into Stable Diffusion, and creation is carried out using computer-recognizable language to gain inspiration. The paper selects the large model $realvisxlV{50}_{v}40LightningBakedvae$ and the Lora model $Slee{k}_{P}roduc{t}_{D}esign$. As shown in the Figure 5, Figure 6, Figure 7, Figure 8 and Figure 9, design schemes are generated for different perceptual images:

“Simple”: Focus on minimalist industrial equipment, with a light-gray main tone, accented with bright yellow, regular shapes, simple lines, and an emphasis on the integration of form and function.

“Technological sense”: Centering around industrial-grade equipment, a combination of dark and light gray colors is used, with aviation blue as accents. A left–right symmetrical layout is employed to showcase a stable, rigorous, and technological style.

“Professional”: Dark gray is blended with aviation blue. Regular geometric shapes, a symmetrical structure, and an integrated layout are adopted, where rounded corners coexist with straight lines to create a professional atmosphere.

“Reliable”: Predominantly using the cold tone of dark gray, with a small amount of aviation blue incorporated to relieve the sense of oppression and enhance the emotional connection with users.

“Distinctive”: With light gray as the main color, locally accented with sky-blue, irregular contour lines, and a combination of straight and curved lines are used to show its characteristics.

By converting the prediction results into a language understandable by large-scale models to generate creative design solutions, the accurate conveyance of design intentions and emotional appeals is ensured.

5.1.2. Optimization of Cabinet Design Schemes

An initial creative design solution is produced based on AIGC, but there is a gap between it and a manufacturable product. Based on this, combined with the requirements of “a certain torque tool detection system”, the solution is transformed considering the structural rationality. The total height of the cabinet is determined to be 850 mm, the height of the handle is 807 mm, and the diameter is 40 mm, taking both cost and comfort into account. The cabinet interfaces are classified and integrated and clear labels are provided to improve the accuracy and safety of operations. Rear-side wiring is adopted to avoid line interference, and the operation area and wiring area are separated for easy operation and maintenance, as shown in Figure 10.

5.2. Optimization and Verification of Design Schemes Based on Eye-Tracking Experiments

1.

experiment design and Setup

To ensure the accuracy and reliability of the verification experiment data, the crucial experimental step of participant screening was controlled and adjusted. Ultimately, 11 participants passed the strict screening and participated in the verification experiment, including five males and six females aged between 21 and 26. After completing the participant screening and conducting the experiment, an in-depth analysis was carried out on the experimental sample data. The experimental sample data of this study passed the Shapiro–Wilk normality test and the Levene’s test for homogeneity of variances, meeting the statistical requirements for analysis of variance, and the effect size ${\eta }^2$ of the core index was close to one. However, small- and medium-sized samples still have statistical and representativeness defects. In subsequent research, the sample size needs to be increased to further enhance the reproducibility and stability of the experimental results. The eye-tracking verification experiment was conducted in a bright and quiet classroom. As shown in Figure 11, adjustable seats were selected to eliminate interference and ensure the validity and scientific nature of the data.

The tools and materials used in the verification experiment are essential for ensuring the reliability of the results. In terms of data collection, the Tobii Pro Fusion eye tracker was employed. This device enables the non-invasive acquisition of eye movement data, including fixation duration and fixation frequency, which are widely recognized indicators of visual attention and cognitive processing [22].

These metrics are also commonly used to analyze user preferences and decision-making processes in visual tasks [23]. By analyzing these parameters, participants’ preference tendencies can be effectively inferred, providing quantitative support for evaluating the perceptual communication effectiveness of the design scheme [24].

The system can accurately capture subtle eye movements and transmit data in real time, thereby ensuring the efficiency and stability of the experimental process. To further support experimental design, data collection, and analysis, ErgoLab software was employed. Its standardized procedures enable precise control over stimulus presentation sequences, ensuring consistency across participants and improving the repeatability of the experiment. In addition, the software facilitates the recording of multiple types of data and provides various visualization tools, thereby enhancing the accuracy and scientific rigor of data analysis.

The experimental stimuli consisted of feasible design schemes representing different perceptual images. The stimulus images were generated based on three-dimensional simulations of real products, with variables strictly controlled under a unified perspective. To minimize order effects and external interference, the presentation sequence of images was randomized. Furthermore, paired bipolar perceptual adjectives were used to guide participant evaluation tasks, providing reliable support for the collection of high-quality experimental data. Eye-tracking has been widely applied in design evaluation studies to investigate user perception and interaction with product features.

2.

Experimental procedure

The verification experiment was programmed using ErgoLab software. The procedure is shown in Figure 12, and the stages and their durations are presented in

Table 19. Experimental flow table of the observation task of pictures with different perceptual images.

Experimental Stage	Experimental Content	Duration
Experimental preparation	Experiment introduction	2 min
	Eye tracker calibration	3 min
Pre-training	Random observation task	1 min
Formal experiment	5 groups of observation tasks	3 min
	5 groups of scoring tasks	3 min

.

Since the emotional image evaluation tasks that include both positive and negative poles can better reflect data differences, both the picture-stimulus observation and subjective evaluation tasks were guided by negative–positive word pairs to obtain real feelings and reliable data. The experiment lasted about 12 min. First, the researchers introduced the procedure and precautions. The participants read the instructions and calibrated the eye-tracker, which took about 5 min. Then, there was a pre-training session where the participants observed random pictures to get familiar with the content, lasting about 1 min. Finally, the formal experiment was conducted, including five sets of observation tasks and five sets of subjective evaluation tasks, which took about 6 min. Data were collected from multiple dimensions to support the research.

3.

Experimental data analysis

For the research on the emotional images of “concise”, “technological sense”, “professional”, “reliable”, and “characteristic”, an index selection strategy was adopted to focus on the analysis of professional eye-movement indicators such as the total fixation duration and the number of fixations in the Areas of Interest (AOI). Eye-movement data were presented using characteristic visualization tools such as heat maps and trajectory maps to reliably analyze how different design schemes attracted the participants’ attention.

Taking the analysis of “concise” as an example, the eye-movement trajectory and heat map are shown in Figure 13 and Figure 14. As can be seen from the figures, Design Scheme 1 could attract the participants’ attention for the longest time, followed by Design Scheme 4, then Design Scheme 5. Design Scheme 2 attracted the participants’ attention for the shortest time but had a relatively large number of fixations.

As shown in Figure 15, it presents the comparison results between the eye-movement data and evaluation values of design schemes in the “concise” perceptual image. The total duration of AOI fixation and the subjective evaluation data show a good consistent change. However, there is a certain difference in the change of the No. 2 design scheme when compared with the number of AOI fixations. For the design scheme that can best convey the feeling of “conciseness”, the data indicators of the participants consistently point to the No. 1 design scheme.

In order to comprehensively understand the prediction effect of the prediction model and the emotional communication effect of design solutions, this paper imported both subjective and objective data into SPSS for Spearman correlation analysis. As an objective physiological indicator, eye-movement data can capture the attention distribution of participants during the stimulus presentation process. Subjective evaluation data can reflect the subjective feelings and preferences of participants towards design solutions under different perceptual images, while the model prediction values are objective prediction results obtained from training with existing data. After combining and analyzing the three-dimensional data of eye-movement data, subjective evaluation, and prediction values, the paper concludes that there is a high degree of consistency among them, which indicates the reliability and validity of the research method.

Spearman consistency analysis is suitable for data with non-normal distribution or non-linear relationships. It measures the monotonic correlation between variables by calculating the rank correlation coefficient and can effectively resist the influence of outliers. Therefore, it is often used to analyze ordinal data such as subjective scores and perceptual images to evaluate the consistency trend between variables. Eye-movement data reflects attention distribution, subjective evaluation reflects preferences, and model prediction values are objective prediction results. We take the “simple” image as an example for analysis. In the “simple” perceptual image, the heat map of the consistency between subjective and objective data is shown in Figure 16.

The Spearman correlation coefficients between the total AOI fixation duration and the number of AOI fixations, predicted values, and evaluation values are 0.80, 0.72, and 0.84 respectively, indicating good consistency. This verifies that the prediction model and the emotional communication effect of the design scheme are excellent. The correlation coefficients between the number of AOI fixations and the predicted values and the evaluation values are only 0.25 and 0.41 respectively, showing weak consistency. It is greatly affected by other factors, making it difficult to verify the effect. However, the correlation coefficient between the predicted values and the evaluation values reaches 0.97, indicating strong consistency, which shows that the prediction model performs excellently under this image and the emotional communication effect of the design scheme is good. Similarly, analyze “sense of technology”, “professional”, “reliable”, and “characteristic”.

6. Conclusions

6.1. Summary of Findings

This paper focuses on the product design methods and processes based on Kansei Engineering in the digital and intelligent era. Firstly, the theoretical foundation of intelligent design is reviewed. The clustering models are compared, the methods for extracting design elements are briefly described, the application potential of machine learning and AI is explored, and the process of eye-tracking experiments and data analysis methods are summarized to provide scientific guidance for design activities.

Next, the Kansei images are clustered based on the LDA model. The design elements are deconstructed, the main images are extracted, and the weight indicators of the elements are determined to provide a benchmark for subsequent research.

Then, the mapping relationship between the Kansei images and the design elements is established to reveal the influence of design elements on users’ Kansei perception, providing data support for building a prediction model.

Subsequently, the SVR and XGBoost machine learning models are constructed. The XGBoost model with better performance is selected to predict the optimal combination of design elements under different Kansei images. The AIGC technology is introduced to complete the transformation of creative design schemes, realizing the intelligent innovation of the design process. Finally, the model and the scheme are verified through eye-tracking experiments and subjective evaluation tasks to verify the feasibility of the prediction model and the emotional communication effect of the design scheme. A complete closed-loop of “theory–model–practice–verification” is formed to explore a new approach to product design based on Kansei Engineering in the digital and intelligent era. It provides theoretical and practical references and paradigms for promoting future intelligent design and helps the product design transform from “experience-driven” to “data-driven”.

6.2. Implications

Beyond the empirical results, this study provides several implications from theoretical, practical, and strategic perspectives. From a theoretical perspective, this study extends Kansei Engineering by establishing a quantitative linkage between emotional perception and design elements. By combining LDA-based semantic extraction, FAHP-based weight calculation, and XGBoost-based prediction, the research moves beyond traditional descriptive analysis toward a predictive modelling framework. This integrated approach addresses the limitation that existing studies often treat emotional analysis and design optimization separately. This directly addresses the identified research gap regarding the lack of integration between semantic analysis, weighting methods, and predictive modelling in Kansei-driven design.

From a practical perspective, the proposed method provides a clear workflow for designers and decision-makers to translate user emotional needs into specific design decisions. The FAHP results help identify key design elements with higher perceptual influence, while the XGBoost model enables the prediction of user responses to different design combinations (R² = 0.87), supporting more accurate scheme selection. In addition, the use of AIGC facilitates rapid visualization of optimized design solutions, and the eye-tracking experiment provides objective validation of users’ attention and emotional response, enhancing the reliability of the design evaluation.

From a strategic perspective, the study demonstrates the feasibility of applying data-driven emotional design methods to industrial equipment, such as torque tool inspection cabinets, which are traditionally dominated by functional considerations. The integration of AI technologies and Kansei Engineering provides a reference for promoting intelligent and user-centered design in industrial contexts, supporting the transformation toward more competitive and experience-oriented product development.