Constructing a MOEA approach for product form Kansei design based on text mining and BPNN

Abstract

Kansei Engineering (KE) is a product design method that aims to develop products to meet users’ emotional preferences. However, traditional KE faces the problem that the acquisition of Kansei factors does not represent the real consumers demands based on manual and reports, and using traditional methods to calculate relationship between Kansei factors and specific design elements, which can lead to the omission of key information. To address these problems, this study adopts text mining and backward propagation neural networks (BPNN) to propose a product form design method from a multi-objective optimization perspective. Firstly, Term Frequency-Inverse Document Frequency (TF-IDF) and WordNet are used to extract key user Kansei requirements from online review texts to obtain more accurate Kansei knowledge. Secondly, the BPNN is used to establish the non-linear relationship between product Kansei factors and specific design elements, and a preference mapping prediction model is constructed. Finally, BPNN is transformed into an iterative prediction value of non-dominated sorting genetic algorithm-II (NSGA-II), and the model is solved through multi-objective evolutionary algorithm (MOEA) to obtain the Pareto optimal solution set that satisfies the user’s multiple emotional needs, and the fuzzy Delphi method is used to obtain the best product form design scheme that meets the user’s multiple emotional images. Using the example of electric bicycle form design could show that this proposed method can effectively complete multi-objective product solutions innovation design.

Keywords

Text mining Back propagation neural network (BPNN)Multi-objective evolutionary algorithm (MOEA)Non-dominated sorting genetic algorithm-II (NSGA-II)Kansei engineering

1 Introduction

In recent years, the progress of manufacturing technology has led to the emergence of a mature product market. It has become increasingly challenging for companies to distinguish themselves through innovation in usability and technology. In fact, the customers’ purchasing decisions are not solely driven by the function and usability of the product. Instead, they place higher value on products that offer an aesthetically pleasing appearance and comfortable texture [1–3]. Consumers have elevated expectations for new products based on the emotional aspects, such as an attractive design that appeals to their psychological needs [4–6].

KE is a branch of engineering that examines the interrelationships between human perception, behavior, and the environment. Its objective is to create more human-centered, comfortable, secure, and sustainable products and environments [7, 8], and convert human needs into specific design elements [9, 10], as depicted in Fig. 1. In recent times, KE has gained significant attention and is regarded as the most dependable and practical approach for addressing emotional requirements of users [11], which is extensively employed in automobiles [12], clothing, furniture [13], mobile phones [14], electric bicycles [15, 16] and other industries field.

Fig. 1

The production design based on Kansei Engineering.

One of the primary challenges in KE is to accurately identify the perceptual needs for users. In the traditional research phase of KE, the needs of user are typically obtained through two approaches. Firstly, a common method involves collecting perceptual adjectives from various sources, such as books, literature, newspapers, and magazines to understand the psychological perceptions for users. This information is processed through extensive questionnaires and statistical analysis to identify relevant perceptual terms. Although this traditional approach provides high-quality results, it is limited by its small scale and one-time nature, leading to limitations in terms of data size, updating efficiency, and data collection [17]. Secondly, many researchers obtain perceptual vocabulary through customer interviews and subjective questionnaires [18, 19]. However, this approach results in a certain degree of subjectivity and uncertainty in the research process due to the implicit and ambiguous nature of customer needs [20], which may mislead the design direction.

With the advancement of computer and internet technologies, online shopping has become a ubiquitous trend that has significantly impacted consumption patterns. Millions of customers now have the opportunity to select their favourite products on platforms such as Amazon and Taobao [21], and to post their own reviews. This review data constitutes a vast database of information connecting product attributes with user experiences, and provide precise and real-time user perceptions [22]. Thus, the text mining of review data can provide valuable insights into consumers’ experiences and expectations of products and services [23]. This technique involves the use of natural language processing (NLP), and computational linguistics to identify and extract important information from text [5]. Many unsupervised text mining algorithms, such as TF-IDF [2] and TextRank [24], could be used to effectively and automatically extract keywords from text. However, the extracted results may require manual secondary processing due to the irregularity and ambiguity of user-generated natural language descriptions. To address these issues, the lexicon WordNet model can be incorporated into word frequency statistics to analyze the semantic relationship of opinion words in phrases. WordNet provides various semantic relationships, such as synonyms and antonyms, to represent the relationship between words, which could used as a tool for understanding consumer emotions and preferences. The current application of WordNet focuses on NLP and sentiment analysis to identify and analyze the emotions and sentiments expressed in product reviews, so as to gain a deeper understanding of customer needs and preferences.

In addition, another challenge in KE is to establish connection between design elements and Kansei needs. In KE research, the mapping between perceptual factors and specific design features is often established through the linear quantitative approach [25, 26]. However, linear regression methods are limited in their ability to measure non-normal affective needs of users as they only assess linear relationships between variables, which has been limited use of predictive models with non-linear characteristics. Neural networks, as a non-linear artificial intelligence model, can simulate the human brain’s thinking process and handle complex non-linear relationships effectively [27]. The Back Propagation Neural Network (BPNN) is a commonly used method in artificial neural network (ANN) models and is well suited to model the mapping between perceptual factors and design elements. BPNN has the ability to adapt to the training data and continuously optimize its structure for different prediction tasks [28]. Compared to conventional linear regression algorithms, BPNN has better model prediction accuracy, faster convergence, good local approximation effect, so as to address the ambiguity and uncertainty. Hence, this study uses the BPNN model to establish the mapping relationship between the perceptual factors and design elements, which could effectively compensate for the information loss inherent in traditional linear methods.

Furthermore, consumers often have multiple Kansei needs, such as technological, modern, and minimalist, and it can be challenging to accurately characterize their emotional needs with conflicting complexities through the use of specific weight values or a single objective [29]. Hence, there is a crucial requirement for developing multi-objective evolutionary algorithms that can fulfill the tangible demands of users for products. The Non-Dominated Sorting Genetic Algorithm-II (NSGA-II) has been utilized by many researchers to study the evolutionary design of products with multiple image objectives. Guo et al. [30] integrating the radial basis function neural network with NSGA-II to achieve a multi-objective optimization of product color design based on two optimization objectives of color harmony and user sentiment preference. Shieh et al. [31] employed NSGA-II as a multi-objective evolutionary algorithm to derive a Pareto optimal solution. Therefore, in this study, we utilize BPNN to establish a mapping function between Kansei factors and specific design elements for production, which could enable us to obtain the product design combinations with the highest perceptual evaluation values through the application of the NSGA-II, thereby providing theoretical guidance for significantly enhancing customer satisfaction.

This objective of this study is to promote designers in efficiently identifying customer needs and developing product forms that fulfill consumer demands. To this end, we propose an integrated product development decision method, which combines text mining and multi-objective evolutionary algorithms. Specifically, we utilize Term Frequency-Inverse Document Frequency (TF-IDF) and WordNet to mine and analyze customer-focused Kansei appeals related to products from customer reviews, and extract high-frequency words using TF-IDF, and make judgments of users’ emotional needs based on the WordNet dictionary. A quantitative mapping relationship between user imagery and product modeling elements is constructed using Neural Networks (NN). The NSGA-II is then used to develop a Pareto product solution set that automatically generates solutions to meet the multidimensional demands of users. Finally, the Fuzzy Delphi method is employed to obtain the optimal solution design from the Pareto solution set, so as to explore the optimal product design solution parameters. Table 1 compares the similarities and differences between this study and previous studies.

Table 1

An overall comparison between this proposed approach and other studies

References	User need	Kansei features	Functional features	Product configuration
This paper	TF-IDF	WordNet		NSGA-II, BPNN
Nagamachi [32]		KE		RST
Guo et al. [33]	HCA, Delphi	KE,SD		BPNN
Bae and Kim [34]				ARM, DT
Wang [35]		KE	RST	CA, GRA
Wang et al. [36]		ResNet		DCGAN
Wang and Wang [37]			Fuzzy Kano, Fuzzy AHP
Wang [1]		KE	RST	FCRP, TRIZ
Wang and Zhou [15]	Kano	EGM		IGA, Evaluation time
Hsiao et al. [38]		AHP		QT-I, GA,
Wang [39]		SD,KJ, K-means	GRA	SVR, ANN
Shi et al. [40]		KE		RST, ARM
Chiu and Lin [2]	Text mining	SD		QTT-I
Jiao and Qu [22]	NLP
Hsiao and Huang [41]		KE		NN

QFD: quality function deployment, CA: conjoint analysis, FA: factor analysis, KE: Kansei engineering, EGM: evaluation gird method, NLP: natural language processing, RST: rough set theory, ARM: association rule mining, GA: genetic algorithm, SVR: support vector regression, TRIZ: the theory of inventive problem solving, DT: decision tree.

2 Literature review

2.1 Text mining

Feldman and Dagan [42] first introducing text mining as a technique for Knowledge Discovery from Text (KDT), especially for the typically large, noisy, and unstructured social media data. Since then, many researchers have utilized text mining to analyze product reviews. Zhu and Zhang [43] discussed the dependence of consumers on online reviews and analyzed the impact on their purchasing decisions. Christensen et al. [44] applied text mining and machine learning to identify and detect new product ideas from online communities. Jin et al. [23] proposed a methed that manually converted customer needs from online reviews into Engineering Characteristics (ECs).

In the realm of text mining, various scholars have utilized unsupervised algorithms, such as TF-IDF and TextRank, to extract keywords from text. In fact, the TF-IDF algorithm is biased towards high frequency words, which have to cause important semantic and emotional information to be overlooked. To address this issue, incorporating sentiment analysis into text analysis can provide a more comprehensive understanding of the contextual information of words. The lexicon-based approach to sentiment analysis is a efficient way, involving the creation of a lexicon that maps words to their semantic values, followed by an iterative process of finding new sentiment words within a large corpus, such as WordNet, HowNet, or Thesaurus, then the iterative process is executed so as to make the process of semantic analysis more efficient [45]. Wang et al. [5] proposed a text mining method that systematically identify Kansei vocabulary based on WordNet semantic thesaurus. Liu [46] develop the algorithm based on the HowNet sentiment dictionary to classify the sentiment orientations of product features.

Furthermore, the WordNet is a comprehensive English semantic thesaurus that encompasses nouns, verbs, adjectives, and adverbs that organises words with synonyms and antonyms. As a lexical database, WordNet provides a structure for categorizing and retrieving words and phrases based on their semantic relationships, which could enable designers to gain deeper understanding of semantics of Kansei words and thereby facilitates the systematic and efficient extraction analysis of Kansei words [47]. Recent studies have indicated that the application of WordNet in the examination of emotional attributes and consumer perceptions has provided valuable insight into customer preferences and needs [48, 49], so as to increase the efficiency of acquiring consumer emotional data. Hence, this research employs a combination method of TF-IDF and WordNet to obtain sensory vocabulary in an efficient and effective manner. The TF-IDF approach is utilized to identify objective high-frequency Kansei keywords in the online review text, while the WordNet method is applied for semantic sentiment analysis of the high-frequency vocabulary. Moreover, this proposed approach considers semantic features of the high-frequency vocabulary to uncover significant semantic information, allowing for rapid acquisition of key user demands in the product development process.

2.2 Neural networks

Gallant introduced the technique of Neural Networks of mathematical models in 1993 [27, 50], which could emulate human perceptual behavior, closely resembling the non-linear processes of the human brain to address the multifaceted intricate problems. It can process diverse types of data by simulating the structure and function of neurons in the human brain. The networks comprise interconnected artificial neurons that are capable of learning from the data, adapting their structure and parameters to the unique characteristics of the data [28]. Through this learning process, Neural Networks can form complex non-linear relationships between input and output data, and this ability allows Neural Networks to handle incomplete data, solve complex and ill-defined problems. A Neural Network comprises basic computational units called nodes, which are organized into three layers: the input layer, the output layer, and the hidden layer [50].

The three-layer feed-forward and Back-Propagation Neural Network (BPNN) model is a commonly used Artificial Neural Network (ANN) model [51]. This model operates by comparing the actual iteration output to the desired value and back-propagating the resulting error signal through the BPNN layer, thereby gradually adjusting the weights of individual network connections until the specified error criterion is met. The BPNN model have been proven to be effective tool for learning, storage, and prediction, which is widely used in various research areas. In KE, Wu [52] used ANN to establish a mapping relationship between morphological characteristics of electric motorbikes and users’ needs, and verified the validity of proposed model. Fang et al. [53] proposed a computational model using BPNN to simulate the mental function of colour aesthetic evaluation. Chen and Chang [54] constructed a linear regression model and BPNN research framework to determine the correlation for product form characteristics and the feasibility is verified. In view of BPNN could effectively handle subjective and imprecise emotional problems, this study employs BPNN model to establish the mapping relationship model between user Kansei need and product design elements, then the BPNN is trained and analyzed for bias, resulting in the evaluation automation of product form, which could provide a solid foundation for further automatic iteration and innovative design.

2.3 Non-dominated sorting genetic Algorithm

Evolutionary algorithms are a widely-used meta-heuristic optimization technique that follow the principles to find optimal solutions, which have been applied to product design optimization, specifically through the use of genetic algorithms. However, user needs for products are often diverse, and single-objective optimization (SOO) falls short in meeting user needs. To this end, some multi-objective optimization (MOO) algorithms based on the Pareto concept have been developed to address problems with multiple conflicting objectives. The unique advantage of MOO algorithms in solving multi-objective planning problems lies in their ability to retain multiple non-dominated solutions, rather than just one optimal solution [55]. The Non-dominated Sorting Genetic Algorithm-II (NSGA-II), introduced by Deb et al. [56], is a widely employed Multi-Objective Evolutionary Algorithm (MOEA) utilized for effectively addressing multi-objective optimization problems. The NSGA-II employs elitism and crowdedness comparison operators to ensure diversity by dividing the objective space into non-dominated layers, which contain relatively better solutions, and ensure that solutions from earlier layers are not dominated by those from later layers, so the NSGA-II is able to better preserve solution dispersion, converge more effectively on the obtained non-dominated frontier.

In production design fields, Yang [55] combined NSGA-II with Support Vector Regression (SVR) to generate optimal product design solutions. Jiang et al. [29] proposed a new engineering research method of a multi-objective optimization model of chaos-based NSGA-II. In this study, a BPNN and NSGA-II-based algorithm is proposed to obtain the better solutions, which is divided into three steps: (1) acquiring user needs through text mining, (2) establishing the relationship between user needs and design elements through BPNN, (3) establishing a morphological evolutionary design method for user multi-objective needs based on the NSGA-II algorithm to solve the user’s multi-objective need problem.

3 Proposed methodologies

This study presents a multi-objective evolutionary design method for product morphology by incorporating text mining and the BPNN model, and this proposed approach is depicted in Fig. 2. First, we apply the TF-IDF method to extract the key high-frequency Kansei word from online review. Subsequently, morphological analysis is utilized to extract and encode the morphological elements of products. Second, the relationship between user demand and product design elements is established based on the BPNN model. To take sample data and consumer evaluation demand data as input and output for training BPNN model, allowing for simulation of user predictions of product Kansei imagery evaluations that adapt to changes in product shape features. Finally, the BPNN model is applied to the multi-objective optimization to evaluate product solutions generated by the multi-objective evolutionary algorithm, and uses the NSGA-II algorithm to implement product iteration and innovation, so as to result in alternative product design solutions that meet user requirements. Moreover, using the FDM method to find the Pareto-optimal frontier product design solution that satisfies the user multi-objective needs.

Fig. 2

The research framework for this study.

3.1 Text mining

This study applies text mining to extract and identify Kansei need of consumers from online reviews. Firstly, by applying TF-IDF method to count the key perceptual information in user comments, and obtain preliminary high-frequency Kansei words, and classify them into categories according to their semantic similarity through KJ method [57]. Secondly, we use WordNet dictionary to analyze the meaning of text to obtain bipolar Kansei words, and use synonym and antonym retrieval to organize words to obtain a complete list. However, some words are ambiguous and have different semantic connotations. To this end, this study applies six extraction rules to solve the contradictions and conflicts in the retrieval process, so that the subgroups in each large group contain a word list with similar semantics. Finally, in order to select key words from the list network, sociometric status is applied to select the Kansei words with the highest value from each representative group as the group representative, so as to obtain the real emotional appeal of users. The acquisition process of perceptual needs is shown in Fig. 3.

Fig. 3

The complete process for the analysis of Kansei images need.

3.1.1 Text pre-processing

In order to ensure that the results of the text mining are representative of a broad range of consumers, it was necessary to collect as many customer reviews as possible. The first step was to browse a wide range of shopping websites to gather many representative samples of product in question. Products with a high number of reviews, typically exceeding 30, were considered valid candidates as they provide a greater reliability in the reviews. The focus of this study was on the reviewer Kansei perspective, and only the text content of the reviews was analyzed. User names, review dates, and rating information were not considered. The raw customer reviews were often interspersed with invalid and unclear information, so the data had to be cleaned to improve its quality. The text of the user comments was cleaned according to clean rules [2, 22, 46], and further word separation and lexical annotation was performed using the Jieba separation tool.

3.1.2 Applying TF-IDF to extract the Kansei word

Firstly, the key assertion in the study of user perceptual imagery is that it encompasses the emotional experience of the user. To gain insight into experience, the frequency of the user’s emotional disposition is analyzed. The significance of each term is determined based on the TF-IDF method, which is a widely utilized approach in the fields of information retrieval and text mining. The TF-IDF method assigns weight to each word in a document based on its frequency of occurrence within the document, and its rarity in the broader corpus of documents. The importance of a word increases as its frequency of occurrence within the document increases, and decreases as its frequency of occurrence across the corpus of documents decreases. The specific steps are shown [58, 59]:

For the set D containing M number of Chinese texts, we use the Jieba segmentation word tool to perform word segmentation for each text in the set D, and then use the TF-IDF algorithm to calculate its weight TF-IDF_i,j in the text, which could represent the weight of the word t_i in the text D_j(j = 1,2,3,...,M) and it could be calculated by formula (1): $TF - {IDF}_{i, j} = {TF}_{i, j} \times {IDF}_{i}$ (1)

n i,j is the number of occurrences of the word t i in the text D j, and the denominator is the sum of the number of occurrences of all words in the text D j. ${TF}_{i, j} = \frac{n_{i, j}}{\sum_{k} n_{k, j}}$ (2)

In formula (3),|D| is the total number of texts in the corpus, | { j : ti ∈ D_j } | could refer to the text number of contained the word t i(i.e., the number of texts with n_i,j ≠ 0). If the word is not in the corpus, it will result in a denominator of zero. Thus, in general, use the 1 + | { j : t_i ∈ D_j } | to address actual project. The high frequency of words in a specific text and the low text frequency of the words in the entire text collection can produce high-weight TF-IDF value. Therefore, the TF-IDF tends to filter common words and retain important words. ${IDF}_{i} = \lg \frac{| D |}{| {{j : t}_{i} \in D_{j}} |}$ (3)

3.1.3 Semantic analysis of Kansei words by WordNet

WordNet is a large semantic thesaurus for English commonly used in many text processing studies [60]. The WordNet consists of nouns, verbs, adjectives and adverbs, and it organizes words using synonyms and antonyms, where synonyms are related to each other by similar semantics. We divided the Kansei words in each group into two subgroups based on their semantics using KJ method, and used WordNet adjective lexicon to find semantically appropriate antonyms and synonyms for all Kansei words in all subgroups. After assigning the initial Kansei words to different subgroups, the antonyms of the subgroup Kansei words were retrieved from WordNet and added to the other subgroups belonging to the same group. Words are ambiguous and have different meanings and semantics. In other words, a word may belong to the same group, but to different subgroups. To resolve these contradictions, we apply rules 1 and 2 [5], as shown in Table 2.

Table 2
The conflict resolution rules of antonym

Rule 1 If two initial Kansei words belong to the same subgroup which are antonyms, then the words are assigned according to their initial subgroup.

Rule 2 If the retrieved word is the antonym of the initial word belonging to the same group but different subgroups, then the retrieved word is assigned to the minority group. If the frequency is the same, the retrieved word should be deleted.

We named the retrieved words as intermediate Kansei words. Then, we use the adjective thesaurus of WordNet to find synonyms for Kansei words from all the subgroups. Similarly, conflicts may also occur. Therefore, the resolution of the conflict for synonym retrieval is based on rules 3–6 [5](Table 3).

Table 3

The conflict resolution rules of synonym

Rule 3	If two initial or intermediate Kansei words belong to the same group, but are in different subgroups, and they are synonyms, then they are assigned based on their initial or intermediate subgroup.
Rule 4	If the retrieved word is a synonym of multiple initial words that belong to the same group but belong to different subgroups, the retrieved word is assigned to the multiple array. If the frequency is the same, the retrieved word should be deleted.
Rule 5	If the retrieved word is only a synonym of the initial word, and then the retrieved word is assigned to the subgroup of the initial word. In other words, when the retrieved word is a synonym of the initial word, ignore the synonym information of the intermediate word.
Rule 6	If the retrieved word is a synonym of the middle word in the same group but in a different subgroup, then the retrieved word is assigned to the multiple array. If the frequency is the same, the retrieved word should be deleted.

Therefore, based on the application of these six rules, the result of each subgroup obtained in the experiment contains a semantically similar word list. In Kansei attribute selection process, we attempt to select a representative word for each subgroup, and measure the sociometric status of all Kansei words collected from the sub-group, and the word with the highest value is selected as the representative item of the subgroup, and the social measurement status of a word is defined as follows [5]: $sociometric_status = \frac{2}{g - 1} \sum_{j = 1}^{g} X_{ij}$ (4)

Where i and j are individual words, x_ij is the edge values from word i to word j, g is the sum total of words in the network. If word i and word j are synonyms, the x_ij = 1, otherwise, x_ij = 0.

3.2 Developing the predictive models for affective responses based on BPNN

The aim of this study was to develop predictive models relating design variables to affective responses, so we construct a BPNN model in order to develop predictive model that is non-linear with mathematical equations. Therefore, we used a three-layer BPNN with a single hidden layer and an input layer with 28 nodes, i.e. the product form performance characteristics corresponding to the six form variables, for a total of 28 elements. 48 product design solution variables were used as inputs and multidimensional Kansei evaluation factors of user needs were used as output nodes.

The specifical implement process of the BPNN model need five key steps: input data encoding, network model construction, error calculation, training iteration, and result output. First, the design parameters of the electric bicycle product, such as frame shape, wheel form, seat style, etc., are encoded, and these design elements are converted into the input layers of the BPNN in a binary encoding manner. The rules are as follows: the number of digits in the code is equal to the number of types of design variables. In the code for each design variable, there is only one digit for “1” and the rest are “0". This input data relies on the labelled datasets of input variables (the code of key design variables) and output variables (assessed for stylistic demands of style, cozy, and bright through Likert scale).

Second, the network model is constructed, including input layer, hidden layer, and output layer, and these layer nodes are connected to the input layer through weights and nonlinear transformation functions to learn more complex relations between input and output.

The error calculation is the key phase of the BPNN model, the output of each node is processed through an activation function and passed to the next layer. Subsequently, the output values of these nodes become the final output of the neural network based on the corresponding predictions of the nodes in the output layer. Then, in the back propagation stage, the output of the network is compared with the actual desired output to obtain the error between the predicted and actual values.

Finally, the model is trained and the process of forward and backpropagation is iterated until the output of the network approximates the true output, so as to output prediction results of the style, cozy, and bright demand of electric bicycle.

In addition, in order to mitigate dimensional inaccuracies, normalization of the input and output data is indispensable. The desired outcome of the training function requires the output parameters of the style, cozy, and bright demand of electric bicycle styling to be in the range [0, 1], but the results obtained from Kansei evaluations through Likert scale do not fall within this interval. To this end, the fast linear transformation algorithm is employed to normalize the data to ensure that experimental results are within the appropriate range, which is defined as follows: $x_{i}^{'} = \frac{x_{i} - x_{\min}}{x_{\max} - x_{\min}} .$ (5)

The notation x_i and x’ _i represent the values original and after normalization, respectively, while x_max and x_min denote the maximum and minimum values in the vector, respectively. The hidden layer activation function of BPNN uses the log sigmoid transfer function [41]. $f (x) = \frac{1}{1 + e^{- x}}$ (6)

The RMSE was applied to test the standard deviation of predicted and measured values, so as to assess the performance of the model and the accuracy of this fitted model. $RMSE = \sqrt{\frac{\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2}}{n}}$ (7)

Where y_i is the evaluation value given by this research participant, ${\hat{y}}_{i}$ is the predicted value from SVR, n is the specific number of data. If there is no difference between two values, the RMSE is equal to 0. Otherwise, the value of RMSE will be large.

3.3 Deriving the Pareto optimal solution set by using the NSGA-II

Pareto optimal solutions are the situations in a multi-objective optimization problem where no one solution in a set of solutions can outperform the others in all objectives [61]. This means that each solution in the set may be better than the others in some objectives, but worse than the others in other objectives, which could be described as follows: (1) Multi-objective problem definition. The multi-objective problem needs to be defined explicitly by identifying the objective functions of the problem. These objective functions may be independent of each other and may be in conflict, so as to improve one objective will lead to deterioration of the other objectives. (2) Solutions space search. After the objective functions are defined, a solution space search needs to be performed. This can be achieved by using different optimization algorithms, and the goal is to find a set of potential solutions that cover the entire solution space. (3) Pareto optimal solution. a solution is considered Pareto optimal solution only if there exists at least one objective on which the solution is superior to other solutions and no other solution can outperform it on all objectives.

The NSGA-II algorithm is an optimization approach to construct the Pareto optimal solution set. The NSGA-II algorithm introduces the concept of congestion, which measures the attribute factors of congestion around an individual in the same non-dominated level. Individuals with larger congestion distances are preferentially selected, leading to an increased population diversity while avoiding the concentration of individuals in a local optimum, the procedure of the NSGA-II algorithm [56] is listed as follows:

Step 1: Constructing and locating consumers’ multi-objective image demand for products.

Step 2: Initializing the product population, and setting the number of population products based on product modeling parameters.

Step 3: Importing the BPNN-based evaluation and prediction model trained by both the product’s Kansei image and the modeling elements, and imaging prediction values of each sample are calculated as product’s fitness values.

Step 4: Performing the championship selection operation on the original population P(t) of size N according to the product’s multi-image modeling fitness evaluation value, and then the crossover and mutation operations is performed on the selected product according to the crossover probability and the mutation probability to produce offspring population Q(t) of size N, and then the combination of the retained product and the cross-mutated product is performed to generate R(t) with the size of 2 N. Then, individuals in R(t) are calculated by nondominated sorting.

Step 5: Sorting the products of the population according to the non-inferior classification strategy.

Step 6: Calculating the crowding degree of individuals in the same classification level.

Step 7: Filling the next generation population P(t + 1) according to the priority order of these levels with starting from the best nondominated level, then the second nondominated level, and so on, until P(t + 1) is filled. Regarding the final level, the solution with a larger crowded distance is preferentially filled with P(t + 1), so that the evolution proceeds in the direction of non-inferior solutions and uniform distribution.

Step 8: Determining whether to terminate. If the maximum evolution algebra termination condition is reached, the loop is ended and the optimized product modeling result is output; Otherwise, it returns to 3 step to continue the genetic operation.

Step 9: Output the best evolutionary individual.

3.4 Selecting the best solution by using fuzzy delphi method

The Delphi method is a qualitative technology used to collect opinions from distributed groups related to specific issues [62]. Then, the improved fuzzy Delphi method proposed by Ishikawa et al. [63] is a concept that combines fuzzy theory with the Delphi method, which could be used to solve the vagueness of the consensus for experts. Consequently, at this stage, the FDM [64] is mainly used to integrate the opinions of users and experts, and then the key project criterion with high consensus is screened out. Specifically, the most optimistic cognitive score can be represented by a triangular fuzzy number (TFN) as O = (L, M, U), where L, M and U represent the minimum, the geometric mean and the maximum, respectively. All of the experts’ most conservative cognition scores for each criterion are stated as a TFN for C = (l, m, u), where l, m and u represent the minimum, the geometric mean, and the maximum, respectively. Then, the double TFNs can be used to test the consistency of expert opinions, the relationship between these two TFNs will have two results: Firstly, if these two TFNs do not overlap, that is, L≥u, this indicates that there is consensus among experts. Therefore, the importance value G_k of the evaluation criterion K is calculated as G_k = (M + m)/2. Secondly, if two TFN would overlap, that is, L < u, when the interval value of L and u is between M and m, this indicates that there is consensus among experts, so the importance value G_k is calculated as G_k = (M + m)/2. However, if L < u, and the interval value of L and u is not between the values of M and m, then the experts’ opinions are not consistent, which need to reuse the questionnaire for data collection until the evaluation result reaches the consensus of experts.

The threshold setting can be determined according to different needs. If G_k is greater than the standard threshold, the k-th criterion is judged as a key evaluation item. If G_k is less than the standard threshold, this criterion is eliminated. Therefore, the final weight W _k of the key item k is calculated as follows: $W_{k} = \frac{G_{k}}{\sum_{k = 1}^{m} G_{k}}$ (8)

In this stage, we build a decision matrix C for n criteria and m solutions based on the optimized product modeling plan obtained by NSGA-II, which can be expressed as follows: $C = [\begin{matrix} c_{11} & c_{12} & \dots & c_{1 n} \\ c_{21} & c_{22} & \dots & c_{2 n} \\ ⋮ & ⋮ & c_{ij} & ⋮ \\ c_{m 1} & c_{m 2} & \dots & c_{mn} \end{matrix}]$ (9)

Therefore, the v_i performance value of the i-th design scheme can be calculated as follows: $v_{i} = \sum_{j = 1}^{n} C_{ij} W_{j},$ (10)

Where W_j is the normalized weight of the j-th criterion calculated by fuzzy Delphi Method. Finally, sort the design solutions accurately of each product performance v_i value.

4 An illustrative Case

4.1 Selecting samples

To identify the prevalent shape elements in electric bicycles, we selected models with a higher number of reviews from the two leading e-commerce platforms, Taobao and Jingdong, and obtained their corresponding morphological data. The models were differentiated based on their shape similarity, and a sample of 30 products was selected from Taobao and another 30 products from Jingdong based on the number of reviews, resulting in a total sample size of 60 products. To increase the efficiency of subsequent emotional cognition experiments, we established a panel of 12 experts (6 female and 6 male designers with an average age of 34.5 years and five years of product design experience) to select representative products from the sample. As a result, 48 representative electric bicycle product samples were selected, which are shown in Fig. 4.

Fig. 4

48 electric bicycle products.

4.2 Capturing user Kansei needs from online reviews

4.2.1 Filter initial Kansei vocabulary based on TF-IDF

The process of obtaining user comment information and extracting product-related adjectives begins with utilizing the crawler function of requests.get, so as to obtain the original corpus for the online review of e-bike products, some of the results of online reviews are displayed in Fig. 5.

Fig. 5

Online reviews of 48 e-bikes (partially).

The obtained information is then processed to five preprocessing rules, as outlined in section 3.1.1 section, so as to eliminate words or phrases that are irrelevant to the semantic meaning of the text. The text is then further processed through word segmentation, resulting in the extraction of text feature words, which are obtained based on formula (1). A total of 260 product-related adjectives being extracted in descending order of weights factors, and a focus group of six industrial designers, each with three years of work experience, selected these adjectives and merged them using the KJ simplification method. This resulted in the selection of 67 Kansei vocabularies related to product images, which served as representative Kansei images. These vocabularies were divided into 12 groups based on their semantics, and each group was further divided into two subgroups (positive and negative), the positive subgroup being referred to as the A group and the negative subgroup referred to as the B group. Furthermore, the results for each Kansei adjective are presented in Table 4.

Table 4

12 group Kansei words

No.	Positive	negative
1	handy	light	…	flexible	agile	bulky	thick	Stocky
2	comfortable	cosy				restrainted
3	Durable	dependable				unreliable
4	Exquisite	sophisticated	detailed	meticulous		rough	crude
5	Solid	sturdy	well-knit	stable		flimsy
6	Simple	brief	concise	natty	neat	complex
7	Cool	hi-tech	grim	dazzling	technology	classial
8	young	fashionable	cheerful			common
9	bright	Lustrous	luxurious	luminous		dim
10	novel	innovational	fresh			boring
11	Spacious	lofty	steady	stable	staid	narrow	frail
12	Streamlined	smooth	graceful	soft		Flat	…	straight

4.2.2 Semantic analysis of Kansei vocabulary based on WordNet

The WordNet is a large semantic thesaurus of English, which is composed of nouns, verbs, adjectives and adverbs. It uses synonyms and antonyms to organize words, and synonyms are related to each other through similar semantics. In order to obtain a complete list of Kansei words and extensively collect more related Kansei vocabulary, we use WordNet adjective thesaurus to find all the antonyms and synonyms of each Kansei words for all subgroups. After the initial Kansei words are assigned to different subgroups, the antonyms of the Kansei words of the subgroups are retrieved from the WordNet which could be added to other subgroups belonging to the same group. However, words are ambiguous, in order to solve the ambiguous situation of antonyms in the grouping process, rules 1 and 2 of 3.1.2.2 section are applied to resolve the conflict.

Furthermore, we use WordNet’s adjective thesaurus to find synonyms for all subgroups of all Kansei words. In order to resolve problems in synonym grouping process, the conflict resolution rules of synonym retrieval is carried out based on rules 3–6 of 3.1.2.2 section. As a result, each subgroup contains a list of words with similar semantic meaning. Moreover, in order to select the representative word in each subgroup, the sociometric status is measured of all Kansei words collected from each subgroup based on formula (4). Then, the word with the highest value is selected as the representative of subgroup. Therefore, 24 Kansei attributes are selected based on 24 subgroups, the experiment results are shown in Table 5.

Table 5
12 group representative Kansei words

1 dexterous 0.364 thick 0.857

2 cozy 2.000 restrainted 0.000

3 dependable 2.667 unreliable 2.000

4 sophisticated 1.000 rough 1.000

5 Solid 0.286 flimsy 0.500

6 simple 0.222 complex 0.667

7 technology 0.444 classial 0.000

8 stylish 0.800 common 0.500

9 bright 0.500 dim 2.000

10 novel 0.667 boring 2.000

11 broad 0.182 narrow 0.400

12 soft 0.182 stiff 0.400

In addition, in order to extract the key perceptual dimensions from the 12 groups of words, the semantic difference method based on a 7-point scale was applied to experimental measure. 184 subjects (92 females and 92 males, mean age 27.35 years, age range 21–28 years, with at least 3 years of designed educational background) were recruited to participate in the experiment, asked to complete the questionnaire, and use the sample mean statistic to calculate the strength of semantic preference. The results of the questionnaire were analysed using factor analysis and the test results revealed that the KMO = 0.742, thus could demonstrate the credibility and validity of this results. Next, the minimum eigenvalue of 1 was used as the cut-off value, thus dividing the 12 adjectives into three main factors, with three main factors accounting for 87.883% of the cumulative variance. Factor 1, factor 2 and factor 3 accounted for 37.778%, 34.509% and 15.596% of the variance respectively (Table 6). We used the adjective with the largest loading factor, resulting in three factors, namely Stylish-Common, Cozy-Restrainted and Bright-Dim, and thus these three factors were used as target criteria for the multi-latitude emotional user needs in this study.

Table 6

Factor loadings of 12 adjectives using 3 factors

Pairwise adjectives		Factor 1	Factor 2	Factor 3
Stylish	Common	0.953	– 0.074	0.178
Sophisticated	Rough	0.952	– 0.101	0.145
Novel	Boring	0.936	– 0.126	0.152
Technology	Classial	0.903	– 0.232	0.195
Dexterous	Thick	0.753	– 0.395	0.397
Cozy	Restrainted	– 0.111	0.944	– 0.018
Dependable	Unreliable	– 0.231	0.916	0.08
Broad	Narrow	– 0.202	0.912	– 0.287
Solid	Flimsy	– 0.093	0.888	0.018
Soft	Stiff	– 0.06	0.671	0.666
Bright	Dim	0.434	0.076	0.776
Simple	Complex	0.392	– 0.305	0.682
Eigenvalue		4.533	4.141	1.871
Percentage of variance		37.778	34.509	15.596
Cumulative percentage		37.778	72.287	87.883

4.3 Using BPNN to construct mapping relationship model for affective responses

4.3.1 Morphological analysis of product form elements

The morphometric analysis was performed on 48 representative e-bike samples to extract shape elements. A focus group of 10 subjects was convened to integrate similar observations using the KJ simplification method. As a result, six distinctive shape elements were identified from the sample set and associated with specific shape types. These shape elements exhibit a range of morphological types, from type 1 to type 6, each with its own unique characteristics. For instance, the ‘wheel (C₂)’ element was found to have four different morphological types, including ‘Y-type (C₂₁)’, ‘pentagram-shaped (C₂₂)’, ’ conventional (C₂₃)’ and ‘rotating (C₂₄)’, as shown in Table 7. Thence, the coding of the product form is further specified based on the shape element.

Table 7
Deconstruction of electric bicycle modeling features

Six elements Category Code Quantitative indicators

C₁ Front structure C₁₁ 1 Square pane type

C₁₂ 2 Y type

C₁₃ 3 Motorcycle type

C₁₄ 4 T tube

C₁₅ 5 Rounded panel type

C₂ Wheel hub C₂₁ 1 Y type

C₂₂ 2 Pentagram type

C₂₃ 3 Conventional type

C₂₄ 4 Rotation shape

C₃ Waist line C₃₁ 1 Triangle shape

C₃₂ 2 U shape

C₃₃ 3 Straight tube shape

C₃₄ 4 S shape

C₃₅ 5 Multi broken line shape

C₃₆ 6 Straight and curve hybrid

C₄ Cushion C₄₁ 1 Egg shape

C₄₂ 2 Flat shape

C₄₃ 3 Geometric line

C₄₄ 4 Triangle shape

C₅ Headlight C₅₁ 1 Round

C₅₂ 2 Square

C₅₃ 3 Hexagon

C₅₄ 4 Flat long oval

C₅₅ 5 Small bulb

C₅₆ 6 Tortuous

C₆ Fender C₆₁ 1 Hat shape

C₆₂ 2 Thin slice type

C₆₃ 3 Sharp type

Six elements	Category	Code	Quantitative indicators
C₁ Front structure	C₁₁	1	Square pane type
	C₁₂	2	Y type
	C₁₃	3	Motorcycle type
	C₁₄	4	T tube
	C₁₅	5	Rounded panel type
C₂ Wheel hub	C₂₁	1	Y type
	C₂₂	2	Pentagram type
	C₂₃	3	Conventional type
	C₂₄	4	Rotation shape
C₃ Waist line	C₃₁	1	Triangle shape
	C₃₂	2	U shape
	C₃₃	3	Straight tube shape
	C₃₄	4	S shape
	C₃₅	5	Multi broken line shape
	C₃₆	6	Straight and curve hybrid
C₄ Cushion	C₄₁	1	Egg shape
	C₄₂	2	Flat shape
	C₄₃	3	Geometric line
	C₄₄	4	Triangle shape
C₅ Headlight	C₅₁	1	Round
	C₅₂	2	Square
	C₅₃	3	Hexagon
	C₅₄	4	Flat long oval
	C₅₅	5	Small bulb
	C₅₆	6	Tortuous
C₆ Fender	C₆₁	1	Hat shape
	C₆₂	2	Thin slice type
	C₆₃	3	Sharp type

4.3.2 Building the BPNN model

This study applies a three-layer neural network with a single hidden layer to establish this non-linear relationship. The input layer has 28 nodes, i.e. 6 morphological variables, for a total of 28 levels. The input data are 48 product design modelling elements, which are coded in depth by a modelling element deconstruction table for the samples to achieve the input, and three nodes in the output layer for the Kansei imagery obtained from the results of the user evaluation for three Kansei terms. The number of neurons in the hidden layer is set as the arithmetic mean of the number of input neurons and the number of output neurons [41], Hence, the number of neurons in the hidden layer is determined as (28 + 3)/2 = 15.5, which is set to 16 neurons in this study, thus the specific framework of BPNN used in this study is shown in Fig. 6 below.

Fig. 6

Construction of the BPNN model.

In order to be able to use the design variables as input parameters, we need to encode specific design elements of the e-bike. Each sample is coded with the same number of bits as the design level and each design element is coded with just one number being 1 and the rest being 0. Also, since the output of the training function needs to be in the range [0,1], so we use the fast linear transformation algorithm of equation (5) to normalise the data so that the evaluation results are scaled to the appropriate range. Finally, we input the normalised results as output parameters into a BPNN model of the e-bike design to complete the model train process.

Referring to previous production design research, the training data sets are 12 [27], 14 [53], 35 [33], and 40 [41]. Generally speaking, the more training data sets, the training performance of the BPNN model will be improved, so as to enhance the generalization ability of the model and reduce the risk of over-fitting. Hence, we defined the first 42 samples as the training dataset and the last 6 samples as the test dataset. We created a neural network (NN) using Newff and established relationships between the 28 design elements (C₁-C₆) and the 3 high-level Kansei words (Y1-Y3) by applying BPNN. After training the network on 42 products, we evaluated the performance of the model from the test set as input parameters. We assessed the difference between the predicted and measured values by using the root mean squared error (RMSE). If there was no difference, the RMSE was 0. We compared the model predictions with the actual values of samples to determine the RMSE, and the model could be used for prediction and inference in real problems only if the RMSE was small. The RMSE is 0.212, 0.295 and 0.122 for Cozy-Restrainted, Stylish-Common, Bright-Dim, respectively. Based on the Hsiao’s study [65], this error is acceptable, thus the validity of the present model was verified.

Table 8

The predictive performance of BPNN model

Kansei words	Cozy-Restrainted		Stylish-Common		Bright-Dim
Production Scheme	Target values	Predicted values	Target values	Predicted values	Target values	Predicted values
NO.43	4.816	4.942	5.421	5.325	5.126	5.139
NO.44	5.737	5.661	4.421	4.800	4.947	5.016
NO.45	4.926	5.153	5.053	5.046	5.105	5.144
NO.46	4.553	4.942	5.263	5.325	5.263	5.139
NO.47	4.789	4.802	4.932	5.130	4.737	4.990
NO.48	4.842	5.055	4.474	5.045	4.968	5.027
RMSE		0.212		0.295		0.122

4.4 Deriving the pareto optimal solutions based on NSGA-II

4.4.1 Establish the multi-image form evolutionary design model

This research utilizes an elite non-dominated sorting genetic algorithm (NSGA-II) to address the problem of multi-objective requirement optimization. Each production samples are represented as 28-bit binary attribute chromosomes. The optimization process of NSGA-II begins by generating an initial population of chromosomes within a range of shape features through the orthogonal method, followed by calculation of adaptation values for each chromosome through the BPNN prediction model. Selected chromosomes undergo crossover and mutation operations to form offspring populations. In this study, the adjustment parameters for NSGA-II are defined as follows: 1000 generations, population size is set to 80, a crossover rate of 0.8, and a mutation rate of 0.3. When the predetermined number of 1000 generations is reached, the product form search terminates, resulting in 80 Pareto solutions, as presented in Table 9 and the Pareto optimal frontier is illustrated in Fig. 7.

Table 9
80 Pareto solutions obtained by using NSGA-II

Sample number Design variables Affective responses

C₁₁-C₆₃ Cozy-Restrainted Stylish-Common Bright-Dim

1 0001000100000010010000010100 3.011 3.868 4.837

2 0001000100100000010100000001 4.587 3.858 3.500

3 0001000100100000010100000010 4.260 3.742 3.600

4 0001000100100000010100000100 4.608 3.399 3.581

⋮ ⋮ ⋮ ⋮ ⋮

80 0010000100000100010000010100 3.148 4.290 4.871

Sample number	Design variables	Affective responses
1	0001000100000010010000010100	3.011	3.868	4.837
2	0001000100100000010100000001	4.587	3.858	3.500
3	0001000100100000010100000010	4.260	3.742	3.600
4	0001000100100000010100000100	4.608	3.399	3.581
⋮	⋮	⋮	⋮	⋮
80	0010000100000100010000010100	3.148	4.290	4.871

Fig. 7

Pareto optimal front in objective space for three affective responses.

It should be noted that as the experimental results of the algorithm are generated randomly, the solutions obtained will be slightly different for each run. Figure 7 shows the set of non-inferior solutions distributed in a relatively dispersed manner, which could indicate that each set of solutions can be used as a result of product design optimisation. Furthermore, the crowded distance of multiple images of products not only solves the problem of products in the same category not being able to be further sorted, but also maintains a uniform distribution of products. Therefore, this proposed multi-objective optimisation method we have developed for product form to guide the iteration of product design solutions is effective and feasible.

4.4.2 Use the fuzzy delphi to select best solution

In this study, we used Fuzzy Delphi Method to determine the best design solution. According to the operation steps of FDM, the gray interval of each item feature and item weight value G^j are generated (as Table 10). In order to screen out key items, the threshold G^j = 4.0 was set in this research. By comparing the predicted value and threshold value in turn, we find the value of three criterion are greater than the threshold value G^j, so kept as the key items. Then, we apply the normalized method to convert the absolute value G_j into the relative value W_j. Therefore, the final weight of Cozy-Restrainted, Stylish-Common, Bright-Dim is 0.380, 0.352, 0.268, respectively (Table 10).

Table 10
Weights of three-dimensional emotion items

Criteria Gray interval Global weight

(L,u) (m,M) G^j W_j

Cozy-Restrainted (5, 7) (4.097, 7.260) 5.678 0.380

Stylish-Common (5, 6) (3.915, 6.621) 5.268 0.352

Bright-Dim (3, 5) (2.650, 5.362) 4.006 0.268

Criteria	Gray interval	Global weight
Cozy-Restrainted	(5, 7)	(4.097, 7.260)	5.678	0.380
Stylish-Common	(5, 6)	(3.915, 6.621)	5.268	0.352
Bright-Dim	(3, 5)	(2.650, 5.362)	4.006	0.268

80 Pareto optimal solutions are sorted according to the performance values, the rank result is shown in Table 11. Solution number 62 ranks 1st, its performance value is 4.415. Solution number 49 ranks 2st, its performance value is 4.399, the last ranking is solution number 11, and its performance value is 3.349. Consequently, we could conclude that the optimal product form design solution is solution number 62.

Table 11

Ranking of Pareto set for design solutions

Rank	Sample	Design variables	Affective responses			Performance
	number	C₁₁-C₆₃	Cozy-Restrainted	Stylish-Common	Bright-Dim	value
1	62	0100000100100000010000001001	5.115	4.183	3.728	4.415
2	49	0001000100100000010000001001	4.998	4.238	3.759	4.399
3	42	0100000100100000010000001010	4.607	4.673	3.600	4.360
4	48	0001000010100000010100000100	5.117	3.970	3.750	4.347
⋮	⋮	⋮	⋮	⋮	⋮	⋮
80	11	0010010000000100001001000001	3.344	2.976	3.845	3.349

4.4.3 Verification result

To validate the results of this study, the optimal design scenario was selected and 124 university students majoring in industrial design were recruited to participate in a product form design evaluation. A satisfaction survey was administered to gather data, and the results showed a satisfaction score is 4.613 for the product design solution, which could indicate that it effectively satisfied the aesthetic needs of users, thereby validating this proposed method. Designers can utilize this method to derive Pareto optimal design solutions that align with consumers’ multidimensional emotional responses, which is beneficial for product redesign and the development of new products.

5 Discussion

This study employed a research approach that blends TF-IDF and WordNet in order to precisely understand customers’ psychological requirements for a product. The frequency of words was calculated using TF-IDF, which extracted essential perceptual words, and unsupervised dictionary semantic analysis was utilized to identify users’ perceptual demands. Subsequently, sociometric analysis was employed to analyze the relationship network of the user’s perceptual needs and quantify the impact factors, thereby revealing the real Kansei needs of the users, which is the cozy, stylish and bright.

In order to construct the non-linear relationship between user Kansei imagery and product styling elements, the morphological analysis was utilized to extract the e-bike styling attractiveness factors, and this resulted in the creation of a database comprising of 6 styling evaluation items and 28 specific styling element evaluation items. Subsequently, a non-linear quantitative mapping relationship was established based on the BPNN. The styling elements were coded as input layer parameters of consumer evaluations of the product were set as the output parameters. The mapping model was trained using the data and its performance was validated using the RMSE, which indicated that the accuracy of the model was reasonable. Hence, the obtained fitted functions serve as the predictive reference for the subsequent stage of multi-objective product form optimization design. Furthermore, this proposed method provides a scientific and systematic approach to quantifying the non-linear relationship between product Kansei imagery and product styling elements. Compared to traditional parametric statistics and linear regression methods, the BPNN has the advantage of having strong learning capabilities, flexibility and non-linear data pattern recognition, while being able to overcome local minimum and overfitting learning problems, resulting in more satisfactory decision results.

Furthermore, a novel approach is proposed to optimize the design of a product by considering the multidimensional emotional imagery perceived by the user, the NSGA-II is utilized to iteratively update the product modelling approach, with the mapping model trained by BPNN serving as its fitness function. The product design solutions that meet the user’s multidimensional aspirations are obtained through the iteration and evolution to form the Pareto solution set. The NSGA-II algorithm is enhanced with non-inferiority classification, crowding distance and pruning operations, resulting in a more horizontally and uniformly distributed product multi-image modelling process. The resulting set of Pareto non-inferiority solutions can meet the consumers’ expectations for product form optimization, thus providing a satisfactory design outcome. In addition, the optimal solution from the Pareto set is selected through the application of the fuzzy Delphi method, so as to lead to a more rational approach to derive the optimal design solution for the production.

In conclusion, the method proposed in this paper aims to accurately capture users’ real emotional preferences in online evaluations of product internet to reflect their actual demands. This is achieved by applying TF-IDF to mine users’ emotional preference factors and using unsupervised dictionary WordNet semantic analysis to identify users’ emotional semantic factors. The emotional demands are then further clarified in combination with social relationship measurement. Furthermore, a combination method of BPNN and NSGA-II is proposed to drive product innovation and evolution. This approach overcomes the limitations of previous studies that relied on single image product shape evolution design and automatically generates optimal parameters for product innovation design. This study uses electric bicycles as an example to demonstrate the effectiveness of the proposed integration method, with the results showing that Scheme 62 ranks first in product design, and it is found that optimum parameter for C₁, C₂, C₃, C₄, C₅, and C₆ are of type 2, 3, 2, 3, 6, and 3 respectively (Table 12), and it visual presentation of production is shown in Fig. 8. Thence, adhering to this design rule in the development of electric bicycles will align with users’ psychological demands and enhance user satisfaction.

Table 12
The optimal parameters design

Front structure Wheel hub Waist line Cushion Headlight Fender

Fig. 8

The visual presentation of design solution.

6 Conclusions and future works

This study proposes a user demand-driven design method that integrates text analysis with BPNN to product design. Firstly, this study analyzes the key perceptual appeals of the product through the use of TF-IDF method and WordNet and quantifies the shape elements of the product through morphological analysis. Subsequently, BPNN is utilized to establish the relationship between user needs and product design elements, and the NSGA-II is adopted to identify a set of Pareto optimal solutions that cater to multi-dimensional user demands. Additionally, FDM is used to rank the design schemes in the Pareto optimal solution set, thereby enabling the analysis of the optimal product design. The method proposed in this study offers several benefits, such as reducing product development costs, shortening the development cycle, and allowing enterprises to respond to user needs in a timely manner and main contributions of this study are summarized as follows.

This study extract product Kansei adjectives from online review accurately and efficiently use TF-IDF to clarify user multidimensional Kansei needs based on the WordNet and social relationship measures. This unsupervised text mining method facilitates the acquisition of real user needs.

A BPNN-based mapping relationship between Kansei imagery and product styling elements is established and incorporated into the product evolution process, so as to allow the NSGA-II model to output product styling solutions that satisfies users’ multidimensional emotional needs.

The Delphi method is applied to further calculate the comprehensive product expressiveness in Pareto optimal solution set to select the optimal product solution.

The results enables designers to quickly and effectively obtain a design solution that meets the emotional imagery needs of their target user, thus improving the success rate of their designs.

In addition, the limitations of this study are as follows: First, in view of the mapping relationship between product image and design elements is effectively constructed through BPNN, although the RMSE value of the test results is acceptable. However, if the prediction accuracy needs more accurate, other prediction models can be used. Second, online review is effective, and user needs are also dynamic over time, so the latest review of products should be focused on, while the relatively long-term comments should be weakened. In addition, this study did not consider the brand factor of the product. Therefore, in the future, the brand culture of the product should also be explored as a priority, so as to complete the product innovation development.

Footnotes

Acknowledgments

The authors thanks for the anonymous reviewers for their help to improve this work.

References

Wang

C.-H.

, Combining rough set theory with fuzzy cognitive pairwise rating to construct a novel framework for developing multi-functional tablets, Journal of Engineering Design 29 (2018), 430–448.

Chiu

M.-C.

and Lin

K.-Z.

, Utilizing text mining and Kansei Engineering to support data-driven design automation at conceptual design stage, Advanced Engineering Informatics 38 (2018), 826–839.

Wang

and Yang

, Combining GRA with a Fuzzy QFD Model for the New Product Design and Development of Wickerwork Lamps, Sustainability 15 (2023), 4208.

Wang

W.M.

, Wang

J.W.

, Li

, Tian

Z.G.

and Tsui

, Multiple affective attribute classification of online customer product reviews: A heuristic deep learning method for supporting Kansei engineering, Engineering Applications of Artificial Intelligence 85 (2019), 33–45.

Wang

W.M.

, Li

, Tian

Z.G.

, Wang

J.W.

and Cheng

M.N.

, Extracting and summarizing affective features and responses from online product descriptions and reviews: A Kansei text mining approach, Engineering Applications of Artificial Intelligence 73 (2018), 149–162.

Vilares

, Alonso

M.A.

and Gómez-Rodríguez

, Supervised sentiment analysis in multilingual environments, Information Processing & Management 53 (2017), 595–607.

Nagamachi

, Kansei engineering as a powerful consumer-oriented technology for product development, Applied Ergonomics 33 (2002), 289–294.

Nagamachi

, Kansei Engineering: A new ergonomic consumer-oriented technology for product development, International Journal of Industrial Ergonomics 15 (1995), 3–11.

Hsiao

Y.-H.

, Chen

M.-C.

and Liao

W.-C.

, Logistics service design for cross-border E-commerce using Kansei engineering with text-mining-based online content analysis, Telematics and Informatics 34 (2017), 284–302.

10.

Wang

, A Novel Approach of Integrating Natural Language Processing Techniques with Fuzzy TOPSIS for Product Evaluation, Symmetry 14 (2022).

11.

Lin

Y.-C.

, Lai

H.-H.

and Yeh

C.-H.

, Consumer-oriented product form design based on fuzzy logic: A case study of mobile phones, International Journal of Industrial Ergonomics 37 (2007), 531–543.

12.

and Zhu

, Extracting knowledge for product form design by using multiobjective optimisation and rough sets, Journal of Advanced Mechanical Design, Systems, and Manufacturing 14 (2020), 1–16.

13.

Akgül

, Özmen

, Sinanoğlu

and Aydoğan

E. Kizilkaya

, Rough Kansei Mining Model for Market-Oriented Product Design, Mathematical Problems in Engineering 2020 (2020), 1–11.

14.

Wang

C.-H.

, Incorporating the concept of systematic innovation into quality function deployment for developing multi-functional smart phones, Computers & Industrial Engineering 107 (2017), 367–375.

15.

Wang

and Zhou

, A method for product form design of integrating interactive genetic algorithm with the interval hesitation time and user satisfaction, International Journal of Industrial Ergonomics 76 (2020).

16.

Wang

and Zhou

, Integrating rough set theory with customer satisfaction to construct a novel approach for mining product design rules, Journal of Intelligent & Fuzzy Systems 41 (2021), 331–353.

17.

, Tian

Z.G.

, Wang

J.W.

, Wang

W.M.

and Huang

G.Q.

, Dynamic mapping of design elements and affective responses: A machine learning based method for affective design, Journal of Engineering Design 29 (2018), 358–380.

18.

, Jiao

R.J.

, Yang

, Helander

, Khalid

H.M.

and Opperud

, An analytical Kano model for customer need analysis, Design Studies 30 (2009), 87–110.

19.

and Wang

, A continuous fuzzy Kano’s model for customer requirements analysis in product development, Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture 226 (2011), 535–546.

20.

Guo

, Qu

Q.-X.

, Nagamachi

and Duffy

V.G.

, A proposal of the event-related potential method to effectively identify kansei words for assessing product design features in kansei engineering research, International Journal of Industrial Ergonomics 76 (2020), 1–10.

21.

Jin

, Ji

and Gu

, Identifying comparative customer requirements from product online reviews for competitor analysis, Engineering Applications of Artificial Intelligence 49 (2016), 61–73.

22.

Jiao

and Qu

Q.-X.

, A proposal for Kansei knowledge extraction method based on natural language processing technology and online product reviews, Computers in Industry 108 (2019), 1–11.

23.

Jin

, Ji

, Liu

and Lim

S.C. Johnson

, Translating online customer opinions into engineering characteristics in QFD: A probabilistic language analysis approach, Engineering Applications of Artificial Intelligence 41 (2015), 115–127.

24.

Anh

K.Q.

, Nagai

and Le

, Minh, Extracting user requirements from online reviews for product design: A supportive framework for designers, Journal of Intelligent & Fuzzy Systems 37 (2019), 7441–7451.

25.

Y.-C.

, Lo

C.-H.

and Chen

C.-C.

, Influence of Personality Traits on Consumer Preferences: The Case of Office Chair Selection by Attractiveness, Sustainability 10 (2018).

26.

Zhang

B.Y.

and Li

, Design for environmental protection: Measuring the appeal factors of green product for consumers, Ekoloji 28 (2019), 1699–1707.

27.

Yeh

Y.-E.

, Prediction of Optimized Color Design for Sports Shoes Using an Artificial Neural Network and Genetic Algorithm, Applied Sciences 10 (2020).

28.

Shieh

M.-D.

and Yeh

Y.-E.

, Developing a design support system for the exterior form of running shoes using partial least squares and neural networks, Computers & Industrial Engineering 65 (2013), 704–718.

29.

Jiang

, Kwong

C.K.

, Liu

and Ip

W.H.

, A methodology of integrating affective design with defining engineering specifications for product design, International Journal of Production Research 53 (2014), 2472–2488.

30.

Guo

, Li

, Nagamachi

, Hu

and Li

, Research on color optimization of tricolor product considering color harmony and users’ emotion, Color Research & Application 45 (2019), 156–171.

31.

Shieh

M.-D.

, Li

and Yang

C.-C.

, Product Form Design Model Based on Multiobjective Optimization and Multicriteria Decision-Making, Mathematical Problems in Engineering 2017 (2017), 1–15.

32.

Nagamachi

, Kansei Engineering and rough sets model, in: Rough Sets and Current Trends in Computing Springer, 2006, pp. 27–37.

33.

Guo

, Liu

W.L.

, Liu

F.T.

, Wang

and Wang

T.B.

, Emotional design method of product presented in multi-dimensional variables based on Kansei Engineering, Journal of Engineering Design 25 (2014), 194–212.

34.

Bae

J.K.

and Kim

, Product development with data mining techniques: A case on design of digital camera, Expert Systems with Applications 38 (2011), 9274–9280.

35.

Wang

C.-H.

, Integrating Kansei engineering with conjoint analysis to fulfil market segmentation and product customisation for digital cameras, International Journal of Production Research 53 (2015), 2427–2438.

36.

Wang

, Ma

and Yang

, Creativity and Sustainable Design of Wickerwork Handicraft Patterns Based on Artificial Intelligence, Sustainability 15 (2023), 1574.

37.

Wang

C.-H.

and Wang

, Combining fuzzy AHP and fuzzy Kano to optimize product varieties for smart cameras: A zero-one integer programming perspective, Applied Soft Computing 22 (2014), 410–416.

38.

Hsiao

S.-W.

, Chiu

F.-Y.

and Lu

S.-H.

, Product-form design model based on genetic algorithms, International Journal of Industrial Ergonomics 40 (2010), 237–246.

39.

Wang

K.-C.

, A hybrid Kansei engineering design expert system based on grey system theory and support vector regression, Expert Systems with Applications 38 (2011), 8738–8750.

40.

Shi

, Sun

and Xu

, Employing rough sets and association rule mining in KANSEI knowledge extraction, Information Sciences 196 (2012), 118–128.

41.

Hsiao

S.-W.

and Huang

H.C.

, A neural network based approach for product form design, Design Studies 23 (2002), 67–84.

42.

Feldman

and Dagan

, Knowledge discovery in textual databases (KDT), in: First International Conference on Knowledge Discovery and Data Mining (KDD-95), AAAI Press, Montreal, Canada, (1995), pp. 112–117.

43.

Zhu

and Zhang

, Impact of Online Consumer Reviews on Sales: The Moderating Role of Product and Consumer Characteristics, Journal of Marketing 74 (2010), 133–148.

44.

Christensen

, Nørskov

, Frederiksen

and Scholderer

, In Search of New Product Ideas: Identifying Ideas in Online Communities by Machine Learning and Text Mining, Creativity and Innovation Management 26 (2017), 17–30.

45.

Moreo

, Romero

, Castro

J.L.

and Zurita

J.M.

, Lexicon-based Comments-oriented News Sentiment Analyzer system, Expert Systems with Applications 39 (2012), 9166–9180.

46.

Liu

, Bi

J.-W.

and Fan

Z.-P.

, Ranking products through online reviews: A method based on sentiment analysis technique and intuitionistic fuzzy set theory, Information Fusion 36 (2017), 149–161.

47.

Madel

, Cahigas

and Tri

, Prasetyo, Kansei Engineering-based Model and Online Content Assessment in Evaluating Service Design of Lazada Express, in: 2020 The 6th International Conference on Industrial and Business Engineerin, (2020), pp. 49–55.

48.

Wang

W.M.

, Li

, Liu

, Tian

Z.G.

and Tsui

, Mining of affective responses and affective intentions of products from unstructured text, Journal of Engineering Design 29 (2018), 404–429.

49.

Tumewu

T.W.

and Dharma

, The Exterior Form Design of Men Formal Shoe with Kansei Engineering Through Structure of Meaning, in: Joint International Conference-APCHI-Ergofuture-PEI-IAIFI 2014, Udayana University Press, 2014.

50.

, Hui

C.-L.

and Choi

T.-M.

, An empirical study of intelligent expert systems on forecasting of fashion color trend, Expert Systems with Applications 39 (2012), 4383–4389.

51.

Russell and Stuart

, Artificial Intelligence: A Modern Approach, The People’s Posts and Telecommunications Press(Posts & Telecom Press), 2002.

52.

, Product form evolutionary design system construction based on neural network model and multi-objective optimization, Journal of Intelligent & Fuzzy Systems 39 (2020), 7977–7991.

53.

Fang

, Muramatsu

and Matsui

, A computational model simulating the mental function of multicolor aesthetic evaluation, Color Research & Application 42 (2017), 216–235.

54.

Chen

H.-Y.

and Chang

H.-C.

, Consumers’ perception-oriented product form design using multiple regression analysis and backpropagation neural network, Artificial Intelligence for Engineering Design, Analysis and Manufacturing 30 (2015), 64–77.

55.

Yang

C.-C.

, Constructing a hybrid Kansei engineering system based on multiple affective responses: Application to product form design, Computers & Industrial Engineering 60 (2011), 760–768.

56.

Deb

, Pratap

, Agarwal

and Meyarivan

, A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation 6 (2002), 182–197.

57.

Scupin

, The KJ Method: A Technique for Analyzing Data Derived from Japanese Ethnology, Human Organization 56 (1997), 233–237.

58.

Zhang

, Yoshida

and Tang

, A comparative study of TF-IDF, LSI and multi-words for text classification, Expert Systems with Applications 38 (2011), 2758–2765.

59.

Wang

, Sun

, Zhou

and Gao

, Construction of a novel production develop decision model based on text mined, in: Design, User Experience, and Usability: UX Research and Design: 10th International Conference, DUXU 2021, Held as Part of the 23rd HCI International Conference, HCII 2021, Virtual Event, July 24–29, 2021, Proceedings, Part I, Springer, 2021, pp. 128–143.

60.

Fellbaum

, Miller

, WordNet: An electronic lexical database, MIT Press, 1998.

61.

Shieh

M.-D.

, Li

and Yang

C.-C.

, Comparison of multi-objective evolutionary algorithms in hybrid Kansei engineering system for product form design, Advanced Engineering Informatics 36 (2018), 31–42.

62.

Singh

P.K.

and Sarkar

, A framework based on fuzzy Delphi and DEMATEL for sustainable product development: A case of Indian automotive industry, Journal of Cleaner Production 246 (2020).

63.

Ishikawa

, Amagasa

, Shiga

, Tomizawa

, Tatsuta

and Mieno

, The max-min Delphi method and fuzzy Delphi method via fuzzy integration, Fuzzy Sets and Systems 55 (1993), 241–253.

64.

Lin

, Application of fuzzy Delphi method (FDM) and fuzzy analytic hierarchy process (FAHP) to criteria weights for fashion design scheme evaluation, International Journal of Clothing Science and Technology 25 (2013), 171–183.

65.

Hsiao

S.-W.

and Tsai

H.-C.

, Use of gray system theory in product-color planning, Color Research & Application 29 (2004), 222–231.

Rule 1	If two initial Kansei words belong to the same subgroup which are antonyms, then the words are assigned according to their initial subgroup.
Rule 2	If the retrieved word is the antonym of the initial word belonging to the same group but different subgroups, then the retrieved word is assigned to the minority group. If the frequency is the same, the retrieved word should be deleted.

1	dexterous	0.364	thick	0.857
2	cozy	2.000	restrainted	0.000
3	dependable	2.667	unreliable	2.000
4	sophisticated	1.000	rough	1.000
5	Solid	0.286	flimsy	0.500
6	simple	0.222	complex	0.667
7	technology	0.444	classial	0.000
8	stylish	0.800	common	0.500
9	bright	0.500	dim	2.000
10	novel	0.667	boring	2.000
11	broad	0.182	narrow	0.400
12	soft	0.182	stiff	0.400

Constructing a MOEA approach for product form Kansei design based on text mining and BPNN

Abstract

Keywords

1 Introduction

2.1 Text mining

2.2 Neural networks

2.3 Non-dominated sorting genetic Algorithm

3 Proposed methodologies

3.1.2 Applying TF-IDF to extract the Kansei word

3.4 Selecting the best solution by using fuzzy delphi method

4.1 Selecting samples

4.2.1 Filter initial Kansei vocabulary based on TF-IDF

4.3.1 Morphological analysis of product form elements

4.4.1 Establish the multi-image form evolutionary design model

Table 10 Weights of three-dimensional emotion items Criteria Gray interval Global weight (L,u) (m,M) G j W j Cozy-Restrainted (5, 7) (4.097, 7.260) 5.678 0.380 Stylish-Common (5, 6) (3.915, 6.621) 5.268 0.352 Bright-Dim (3, 5) (2.650, 5.362) 4.006 0.268

5 Discussion

Table 12 The optimal parameters design Front structure Wheel hub Waist line Cushion Headlight Fender

Footnotes

Acknowledgments

References

Table 10
Weights of three-dimensional emotion items

Criteria Gray interval Global weight

(L,u) (m,M) G^j W_j

Cozy-Restrainted (5, 7) (4.097, 7.260) 5.678 0.380

Stylish-Common (5, 6) (3.915, 6.621) 5.268 0.352

Bright-Dim (3, 5) (2.650, 5.362) 4.006 0.268

Table 12
The optimal parameters design

Front structure Wheel hub Waist line Cushion Headlight Fender