Automobile modeling determination method, device and equipment based on self-adaptive multi-criterion evaluation

Abstract

The invention provides an automobile modeling determination method, device and equipment based on adaptive multi-criterion evaluation, and the method comprises the steps: collecting the multi-modal data of an automobile modeling scheme, carrying out the feature extraction, enabling all modal features to be aligned and fused into a unified feature expression through a cross-modal fusion model, and carrying out the recognition of the multi-modal data. And performing multi-criterion evaluation on the fusion feature based on the initial criterion weight to obtain a sub-item score and a comprehensive score, and sorting the candidate schemes according to the sub-item score and the comprehensive score. And then feedback data of the target user group to the sorted schemes is obtained, and the response degree of each evaluation criterion to the user feedback is calculated according to the fusion feature representation and the feedback data, so that the weight configuration of each criterion is dynamically adjusted. The adjusted weight is used for re-executing evaluation and sorting, so that the criterion weight can be automatically optimized according to actual feedback of a specific user group, the evaluation model is converted from a static fixed weight to a dynamic adaptive weight, and the finally output optimal modeling scheme can be matched with differentiated aesthetic preferences of target market users.

Description

Technical Field

[0001] This invention relates to the field of automotive styling design, and in particular to a method, apparatus, and device for determining automotive styling based on adaptive multi-criteria evaluation. Background Technology

[0002] Automotive styling design is one of the core aspects of automotive product development, and the quality of the styling design directly affects the product's market competitiveness. With the development of generative artificial intelligence technology, designers can quickly generate a large number of candidate styling designs. However, how to efficiently select the optimal design from this massive pool of options that simultaneously satisfies aesthetic appeal, functional performance, and market acceptance has become a key bottleneck restricting design efficiency.

[0003] Existing automotive styling evaluation methods typically employ multi-criteria decision models, comprehensively considering multiple evaluation criteria such as aesthetics, function, and market. However, the weighting of each criterion in these methods is usually static, either pre-set by experts or determined in a single calculation using methods such as the analytic hierarchy process (AHP). This static weighting approach faces a core problem in practical applications: target user groups in different regions and market segments have significantly different aesthetic preferences for automotive styling. For example, European market users may prefer a sporty and dynamic styling, while Asian market users may favor a more stable and sophisticated style; younger and more mature user groups also have different focuses regarding styling elements.

[0004] Evaluation models using fixed weights cannot adapt to these differentiated user preferences: when the evaluation model sets weights based on European market preferences, the optimal solution it selects may not meet the aesthetic expectations of users in the Asian market; and vice versa. This forces companies to repeatedly adjust the weights of evaluation criteria and re-evaluate when designing for different markets, resulting in long design cycles, low iteration efficiency, and a lack of objective basis for weight adjustment, still relying on the subjective judgment of the design team.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention discloses a method, apparatus, and device for determining automobile styling based on adaptive multi-criteria evaluation, aiming to solve the technical problem that existing automobile styling evaluation methods have statically fixed criterion weights and cannot adapt to the aesthetic preferences of different user groups.

[0007] The first embodiment of the present invention provides a method for determining vehicle styling based on adaptive multi-criteria evaluation, comprising: Multimodal data of the vehicle styling scheme to be evaluated are collected, and feature extraction is performed on the multimodal data to obtain the feature vector of each mode; A cross-modal fusion model is used to align and fuse the feature vectors of each modality to generate a unified fused feature representation. Based on the preset initial criterion weights, the fusion feature representation is evaluated using a multi-criterion evaluation model, the sub-scores and comprehensive scores of each evaluation criterion are calculated, and multiple candidate styling schemes are ranked according to the comprehensive scores. Obtain feedback data from the target user group on the sorted candidate design schemes, calculate the responsiveness of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. Based on the adjusted criterion weights, the multi-criterion evaluation and scheme ranking are re-executed until the preset convergence conditions are met, and then the optimal modeling scheme and the corresponding criterion weight configuration are output.

[0008] Preferably, the multimodal data includes visual modal data, semantic modal data, and numerical modal data; The feature extraction process of the visual modality data includes: inputting the shape image into a pre-trained convolutional neural network, extracting the high-level semantic features of the image, and mapping them into a fixed-dimensional visual feature vector through a fully connected layer; The feature extraction process of the semantic modality data includes: inputting the design description text into the pre-trained language model, and taking the pooling output of the last hidden state as the semantic feature vector; The feature extraction process of the numerical modal data includes: normalizing the aerodynamic coefficients and ergonomic parameters, and then concatenating them into a numerical feature vector.

[0009] Preferably, the step of aligning and fusing the feature vectors of each modality using a cross-modal fusion model to generate a unified fused feature representation specifically involves: The feature vectors of each modality are mapped to a common feature space of the same dimension through a linear projection layer. Modality type encoding is added to the mapped feature vectors of each modality, and they are concatenated along the sequence dimension. The concatenated feature sequence is input into a multi-layer Transformer encoder, and cross-modal information interaction is achieved through a self-attention mechanism. The output of the Transformer encoder is pooled to obtain the fused feature representation.

[0010] Preferably, the evaluation criteria include aesthetic criteria, functional criteria, and market criteria; the evaluation process of the multi-criteria evaluation model includes: The fusion feature representations are input into the scoring subnetworks corresponding to each evaluation criterion to obtain the sub-scores of each criterion. The sub-scores of each criterion are weighted and summed using the current criterion weights to obtain the comprehensive score of each candidate design scheme. The multiple candidate design schemes are sorted in descending order according to the comprehensive score.

[0011] Preferably, the scoring process of the scoring subnetwork corresponding to the aesthetic criteria includes: The fused feature representation is input into the aesthetic scoring subnetwork, which sequentially includes a feature dimensionality reduction layer, a non-linear activation layer, and a scoring output layer. The scoring output layer outputs scores for the proportion coordination of the shape, symmetry, and dynamism. The scores for proportional harmony, symmetry, and dynamism are weighted and summed to obtain the individual scores for the aesthetic criteria.

[0012] Preferably, the scoring process of the scoring subnetwork corresponding to the functional criteria includes: Extract function-related feature components from the fused feature representation; The feature components are concatenated with the aerodynamic coefficients and ergonomic parameters in the numerical modal data and then input into the functional scoring subnetwork. The functional scoring subnetwork outputs the sub-scores of the functional criteria.

[0013] Preferably, feedback data from the target user group on the current design scheme is obtained, specifically as follows: Obtain the rating data and text evaluation data submitted by the target user based on the sorted candidate design schemes; Statistical analysis is performed on the rating data to obtain a group preference score. Sentiment analysis and keyword extraction are performed on the text evaluation data to obtain a feedback feature vector. The group preference score and the feedback feature vector are combined to form the feedback data.

[0014] Preferably, the dynamic adjustment of the weights of each evaluation criterion specifically involves: The evaluation deviation value is calculated based on the current comprehensive score and user feedback data. The fusion feature representation and the evaluation deviation value are concatenated and input into the weight optimization network. The weight optimization network outputs the weight adjustment amount of each evaluation criterion. The current criterion weight is added to the weight adjustment amount and normalized to obtain the adjusted criterion weight.

[0015] The second embodiment of the present invention provides a vehicle styling determination device based on adaptive multi-criteria evaluation, comprising: A multimodal data processing unit is used to collect multimodal data of the car styling scheme to be evaluated, and to extract features from the multimodal data to obtain feature vectors for each mode. The fusion unit is used to align and fuse the feature vectors of each modality using a cross-modal fusion model to generate a unified fused feature representation. The multi-criteria evaluation unit is used to evaluate the fusion feature representation based on preset initial criterion weights using a multi-criteria evaluation model, calculate the sub-scores and comprehensive scores of each evaluation criterion, and rank multiple candidate styling schemes according to the comprehensive scores. The weight adjustment unit is used to obtain feedback data from the target user group on the current design scheme, calculate the degree of response of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. The scheme output unit is used to re-execute multi-criteria evaluation and scheme ranking based on the adjusted criterion weights until the preset convergence conditions are met, and then output the optimal modeling scheme and the corresponding criterion weight configuration.

[0016] The third embodiment of the present invention provides a vehicle styling determination device based on adaptive multi-criteria evaluation, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a vehicle styling determination method based on adaptive multi-criteria evaluation as described in any of the above embodiments.

[0017] This invention provides a method, apparatus, and device for determining automotive styling based on adaptive multi-criteria evaluation. The method involves collecting multimodal data of automotive styling schemes and extracting features. A cross-modal fusion model aligns and fuses the features of each modality into a unified feature representation. Based on initial criterion weights, the fused features are evaluated using multiple criteria to obtain individual scores and a comprehensive score, which are then used to rank candidate schemes. Feedback data from the target user group on the ranked schemes is then obtained. The responsiveness of each evaluation criterion to user feedback is calculated based on the fused feature representation and feedback data, thereby dynamically adjusting the weight configuration of each criterion. The adjusted weights are used to re-execute the evaluation and ranking, enabling the criterion weights to automatically optimize based on the actual feedback from a specific user group. This transforms the evaluation model from static fixed weights to dynamic adaptive weights, ensuring that the final optimal styling scheme matches the differentiated aesthetic preferences of the target market users. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for determining vehicle styling based on adaptive multi-criteria evaluation provided in the first embodiment of the present invention. Figure 2This is a schematic diagram of a vehicle styling determination device based on adaptive multi-criteria evaluation provided in the second embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] This invention discloses a method, apparatus, and device for determining automobile styling based on adaptive multi-criteria evaluation, aiming to solve the technical problem that existing automobile styling evaluation methods have statically fixed criterion weights and cannot adapt to the aesthetic preferences of different user groups.

[0022] Please see Figure 1 The first embodiment of the present invention provides a method for determining automobile styling based on adaptive multi-criteria evaluation, which can be executed by an automobile styling determination device (hereinafter referred to as the determination device or system), and in particular, by one or more processors within the determination device or system, to at least implement the following steps: S101, Collect multimodal data of the car styling scheme to be evaluated, and extract features from the multimodal data to obtain feature vectors for each mode; In this embodiment, the determining device or system can be a terminal with data processing capabilities, such as a desktop computer, a laptop computer, or a server workstation. The auxiliary device can be equipped with a corresponding operating system and application software, and the functions required in this embodiment can be realized through the combination of the operating system and application software.

[0023] In this embodiment, three types of multimodal data are collected: visual modal data, semantic modal data, and numerical modal data. Visual modal data mainly includes rendered images of the vehicle's styling or multi-view screenshots of the 3D model, used to capture the appearance and morphological features of the design. Semantic modal data mainly includes design descriptions, design concept descriptions, or user evaluations of the styling scheme, used to express the design intent and semantic information of the design. Numerical modal data mainly includes aerodynamic coefficients and ergonomic parameters obtained through simulation analysis or actual measurements, used to quantify the functional performance indicators of the styling.

[0024] After completing multimodal data acquisition, the system extracts features from each modality to obtain corresponding feature vectors. For visual modality data, the system inputs the model image into a pre-trained convolutional neural network for processing. This convolutional neural network can be a deep network model pre-trained on a large-scale image dataset, such as ResNet, VGG, or EfficientNet. Through multi-layer convolutional operations, the network extracts high-level semantic features from the image. These high-level semantic features effectively represent visual elements such as the contour lines, surface transitions, and proportional relationships of the model. The extracted high-level semantic features are then mapped to fixed-dimensional visual feature vectors via fully connected layers, for example, 512-dimensional or 1024-dimensional vector representations. For semantic modality data, the system inputs the design description text into a pre-trained language model for processing. This pre-trained language model can be BERT, RoBERTa, or other Transformer-based language models. After word segmentation and encoding, the text is input into the language model. The system takes the pooled output of the last hidden state as the semantic feature vector. This semantic feature vector encodes semantic information in the text regarding style, design concept, and emotional tendency. For numerical modal data, the system first normalizes the aerodynamic coefficients and ergonomic parameters. Specifically, min-max normalization or Z-score normalization methods can be used to scale the parameter values ​​to a uniform numerical range to eliminate the influence of differences in the dimensions and numerical scales of different parameters. After normalization, the parameters are concatenated into a numerical feature vector in a preset order. Through the above feature extraction process, the system obtains visual feature vectors, semantic feature vectors, and numerical feature vectors. These three feature vectors represent the vehicle styling scheme from three dimensions: visual appearance, semantic description, and functional performance, respectively.

[0025] S102, A cross-modal fusion model is used to align and fuse the feature vectors of each modality to generate a unified fused feature representation; After extracting features from each modality, the system employs a cross-modal fusion model to align and fuse visual, semantic, and numerical feature vectors to generate a unified fused feature representation. Since the feature vectors from each modality originate from different feature extraction networks, their vector dimensions and feature distributions differ, necessitating feature alignment first. Specifically, the system maps the feature vectors of each modality to a common feature space of the same dimension through their respective linear projection layers. For example, it maps the originally different-dimensional visual, semantic, and numerical feature vectors to a unified 256-dimensional or 512-dimensional vector representation, making the features of different modalities comparable within the same vector space.

[0026] After dimensional alignment, the system adds a modality type code to each mapped modal feature vector. This modality type code is a learnable embedding vector used to identify the modality category to which the feature vector belongs, enabling the fusion model to distinguish feature information from different sources. The modality type code can be added using vector addition, that is, adding the modality type code vector to the corresponding modality feature vector. The system then concatenates the modality feature vectors with added modality type codes along the sequence dimension to form a feature sequence containing multiple modal features. The length of this feature sequence is equal to the number of modalities involved in the fusion.

[0027] The system inputs the concatenated feature sequence into a multi-layer Transformer encoder for cross-modal fusion processing. The Transformer encoder consists of multiple stacked encoding layers, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer. In the multi-head self-attention sub-layer, each modal feature vector in the feature sequence is used as a query vector to perform attention calculations with all other modal feature vectors. This self-attention mechanism enables cross-modal information interaction and fusion, allowing the feature representation of each modality to dynamically aggregate relevant information from other modalities. Through iterative processing by the multi-layer Transformer encoder, information exchange and fusion are fully achieved among the features of each modality. Finally, the system performs pooling on the output of the Transformer encoder to obtain the fused feature representation. Pooling can be performed using methods such as average pooling or taking the output vector at a specific position, converting the sequence-based output into a single vector of fixed dimensions. This fused feature representation comprehensively encodes information from multiple dimensions of the automotive styling scheme, including visual appearance, semantic description, and functional performance, providing a unified input feature for subsequent multi-criteria evaluation.

[0028] S103, based on the preset initial criterion weights, the fusion feature representation is evaluated using a multi-criterion evaluation model, the sub-scores and comprehensive scores of each evaluation criterion are calculated, and the multiple candidate styling schemes are ranked according to the comprehensive scores; After obtaining the fused feature representation, the system evaluates the fused feature representation using a multi-criteria evaluation model based on preset initial criterion weights to calculate the sub-scores and overall score for each evaluation criterion. In this embodiment, the evaluation criterion system includes three dimensions: aesthetic criteria, functional criteria, and market criteria, which are used to evaluate the performance of the automotive styling scheme in terms of visual appeal, functional performance, and market acceptance, respectively. The initial criterion weights can be preset by domain experts based on experience; for example, the weight of the aesthetic criterion can be set to 0.4, the weight of the functional criterion to 0.3, and the weight of the market criterion to 0.3, with the sum of the weights of all criteria being 1.

[0029] The multi-criteria evaluation model includes scoring sub-networks corresponding to each evaluation criterion. The system inputs the fused feature representation into each scoring sub-network to obtain the sub-scores for the corresponding criterion. For the aesthetic criteria, the system inputs the fused feature representation into the aesthetic scoring sub-network for processing. This aesthetic scoring sub-network adopts a multilayer perceptron structure, which sequentially includes a feature dimensionality reduction layer, a nonlinear activation layer, and a scoring output layer. The fused feature representation first passes through the feature dimensionality reduction layer to map high-dimensional features to a lower-dimensional hidden space. Then, it passes through the nonlinear activation layer to introduce nonlinear expressive power. Common activation functions such as ReLU or GELU can be used. Finally, the scoring output layer outputs the scoring results for multiple aesthetic sub-items, specifically including styling proportion harmony score, symmetry score, and dynamic sense score. The styling proportion harmony score is used to evaluate whether the proportional relationship between different parts of the car body is harmonious. The symmetry score is used to evaluate the degree of left-right symmetry and the processing effect of symmetrical elements. The dynamic sense score is used to evaluate the sense of movement and speed conveyed by the styling. The system performs a weighted sum of the scores of the above three sub-items according to preset sub-item weights to obtain the sub-scores for the aesthetic criteria.

[0030] For functional criteria, the system first extracts feature components related to functional performance from the fused feature representation. This extraction process can be achieved through linear projection or attention mechanisms to filter feature dimensions that are highly correlated with aerodynamic performance and ergonomics. Subsequently, the system concatenates the extracted feature components with aerodynamic coefficients and ergonomic parameters from numerical modal data to form an input vector for functional evaluation. This concatenation operation allows the functional scoring process to simultaneously utilize implicit functional representations learned from the fused features and explicit functional parameters obtained through simulation or measurement. The concatenated input vector is fed into a functional scoring sub-network, which outputs sub-scores for the functional criteria, used to quantify the overall performance of the design scheme in terms of aerodynamic efficiency and human-computer interaction convenience.

[0031] For market criteria, the system inputs the fused feature representation into the market scoring sub-network to evaluate the market acceptance potential of the styling scheme. The market scoring sub-network first maps the fused feature representation to the market preference feature space through a feature mapping layer, then extracts deep features related to market performance through multiple hidden layers, and finally outputs the scoring results for multiple market sub-items through the scoring output layer, specifically including user preference matching score and competitor differentiation score. The user preference matching score evaluates the degree of matching between the styling scheme and the historical preference data of the target user group. This score is calculated based on the similarity measure between the fused feature representation and a pre-built user preference feature library, which is trained from historical market research data and user evaluation data. The competitor differentiation score evaluates the degree of differentiation of the styling scheme relative to existing competing models in the market. This score is obtained by calculating the distance between the fused feature representation and each competitor feature in the competitor styling feature library; a larger distance indicates a higher degree of differentiation. The system weights and sums the user preference matching score and the competitor differentiation score according to preset sub-item weights to obtain the sub-item scores for the market criteria.

[0032] After obtaining the scores for each evaluation criterion, the system uses the current criterion weights to perform a weighted sum of the scores for the aesthetic, functional, and market criteria sub-categories, calculating the overall score for the candidate design. The system repeats this evaluation process for all candidate designs to be evaluated, obtaining the overall score for each design. Then, based on the overall scores, the multiple candidate designs are sorted in descending order, with the design with the highest overall score ranked first, and so on. The ranking results intuitively reflect the overall merits of each candidate design under the current criterion weight configuration.

[0033] S104, obtain feedback data from the target user group on the sorted candidate design schemes, calculate the degree of response of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. After completing the initial multi-criteria evaluation and ranking the candidate styling schemes, the system enters the user feedback collection phase to obtain feedback data from the target user group on the ranked candidate styling schemes. The system displays the ranked candidate styling schemes to the target user group through online questionnaire platforms, on-site survey terminals at auto shows, or mobile applications. The selection of the target user group is determined based on the target market positioning; for example, for sports sedans, users aged 25 to 40 with a high level of interest in vehicle performance could be selected. The system acquires rating data and text evaluation data submitted by target users based on the ranked candidate styling schemes. The rating data consists of quantitative scores given by users to each styling scheme, using rating scales such as five-point or ten-point scales. The text evaluation data consists of users' subjective evaluation opinions on the styling schemes expressed in natural language.

[0034] After acquiring the raw feedback data, the system processes the rating data and text evaluation data separately to generate structured feedback data. For the rating data, the system performs statistical analysis to obtain a group preference score. Specific processing includes removing outlier ratings, calculating the average score and standard deviation of each design scheme, and performing stratified statistical analysis of ratings from different user subgroups. The final result is a group preference score vector that reflects the overall preference tendency of the target user group. For the text evaluation data, the system first performs text preprocessing, including word segmentation, removal of stop words and punctuation, etc. Then, it performs sentiment analysis on the preprocessed text to determine the sentiment polarity and intensity of user evaluations. Sentiment analysis can be implemented using a sentiment classifier based on a pre-trained language model, outputting positive, negative, or neutral sentiment labels and corresponding confidence scores. Simultaneously, the system extracts keywords from the text evaluation data, identifying frequently occurring design-related words in user evaluations, such as "streamlined," "stable," "radical," and "avant-garde." These keywords reflect users' focus and preference direction regarding design features. The system converts sentiment analysis results and extracted keywords into feedback feature vectors using vector encoding. The dimension of these feedback feature vectors matches the dimension of the fused feature representation for subsequent calculations. Finally, the system combines the group preference score with the feedback feature vectors to form a complete feedback data representation.

[0035] After receiving feedback data, the system dynamically adjusts the weights of each evaluation criterion based on the fused feature representation and the feedback data. First, the system calculates the evaluation deviation value based on the current comprehensive score and user feedback data. This evaluation deviation value quantifies the gap between the output of the current multi-criterion evaluation model and the user's actual preferences. Specifically, the system compares the comprehensive score ranking of each candidate scheme calculated based on the current criterion weights with the preference ranking reflected in the user feedback data, calculating the degree of difference between the two rankings. This difference can be quantified using a ranking correlation coefficient or a ranking loss function, yielding an evaluation deviation value in scalar or vector form. Subsequently, the system concatenates the fused feature representation and the evaluation deviation value along the feature dimension to form the input vector of the weight optimization network. This concatenation operation allows the weight optimization process to simultaneously consider the feature information of the styling schemes and the deviation status of the current evaluation model.

[0036] The concatenated input vector is fed into a weight optimization network for processing. This network employs a multilayer perceptron structure, containing an input layer, several hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the number of neurons in the output layer equals the number of evaluation criteria. The training objective of the weight optimization network is to learn the mapping relationship from fused features and evaluation bias to the optimal weight adjustment. The training process is supervised learning based on historical modeling evaluation data and user feedback data. The weight optimization network outputs the weight adjustment for each evaluation criterion. This adjustment is a signed numerical value; a positive value indicates that the weight of the corresponding criterion should be increased, and a negative value indicates that the weight of the corresponding criterion should be decreased. The system adds the current criterion weight to the weight adjustment output by the weight optimization network to obtain the updated criterion weights. The updated criterion weights are then normalized to ensure that all criterion weights are non-negative and that the sum of the weights equals 1. Normalization can be implemented using the Softmax function or linear normalization. After this processing, the system obtains adjusted criterion weights, which, compared to the initial weights, better reflect the actual preference characteristics of the target user group.

[0037] S105, based on the adjusted criterion weights, re-execute the multi-criterion evaluation and scheme ranking until the preset convergence conditions are met, and then output the optimal modeling scheme and the corresponding criterion weight configuration.

[0038] After adaptively adjusting the criterion weights, the system re-executes multi-criterion evaluation and scheme ranking based on the adjusted criterion weights. Specifically, the system replaces the original criterion weight configuration with the adjusted criterion weights, keeping the scoring sub-networks corresponding to each evaluation criterion unchanged, and recalculates the multi-criterion evaluation of the fusion feature representations of all candidate styling schemes. The system inputs the fusion feature representations of each candidate scheme into the aesthetic scoring sub-network, functional scoring sub-network, and market scoring sub-network, respectively, to obtain the sub-scores for each criterion. Then, it uses the adjusted criterion weights to perform a weighted summation of the sub-scores to obtain the updated comprehensive score for each candidate scheme. Since the criterion weights have been adjusted based on user feedback, the updated comprehensive score better reflects the preferences of the target user group compared to the initial evaluation results. The system re-ranks all candidate styling schemes in descending order based on the updated comprehensive score, generating a new scheme ranking result.

[0039] After the re-ranking is completed, the system re-enters the user feedback collection phase, displaying the re-ranked candidate designs to the target user group and collecting a new round of scoring and textual evaluation data. The system processes this new round of feedback data using the aforementioned method, calculates the evaluation deviation between the current evaluation result and the user feedback, and calculates new weight adjustments through a weight optimization network to update the criterion weights. This process of multi-criterion evaluation, user feedback collection, and adaptive weight adjustment is repeated cyclically, forming a closed-loop iterative optimization process.

[0040] After each iteration, the system determines whether the current state meets the preset convergence conditions. In this embodiment, the convergence conditions include any one of the following three scenarios. The first scenario is the weight stability condition, where the system detects the change in the weights of each criterion in several consecutive iterations. If the change in the weights of each criterion is less than a first preset threshold in a preset number of consecutive iterations, the criterion weights are considered to have stabilized and the convergence condition is met. For example, convergence can be determined when the change in the weights is less than 0.01 in three consecutive iterations. The second scenario is the deviation convergence condition, where the system calculates the deviation between the current comprehensive score ranking and the preference ranking reflected in user feedback. If the deviation is less than a second preset threshold, it indicates that the output of the current evaluation model is highly consistent with the user's actual preferences and the convergence condition is met. For example, convergence can be determined when the ranking correlation coefficient is greater than 0.95. The third scenario is the iteration count condition, where the system records the number of iterations executed. If the number of iterations reaches the preset maximum number of iterations, the iteration process is forcibly terminated to avoid falling into an infinite loop due to failure to converge. For example, the maximum number of iterations can be set to 10.

[0041] When any of the above convergence conditions is met, the system terminates the iterative optimization process and enters the output stage. The system extracts the candidate design with the highest comprehensive score from the final ranking results as the optimal design, and simultaneously records the criterion weight configuration after iterative convergence. This criterion weight configuration has been adaptively adjusted through multiple rounds of user feedback, accurately reflecting the relative preferences of the target user group in the three evaluation dimensions of aesthetics, functionality, and market. It can serve as a reference weight for subsequent design or evaluation targeting this market.

[0042] The system outputs the optimal styling scheme and its corresponding criterion weight configuration, and can generate an evaluation report for the design team's reference as needed. The evaluation report generation process includes the following: the system extracts the sub-scores of each evaluation criterion corresponding to the optimal styling scheme, presenting the distribution of each criterion score in a radar chart format, allowing designers to intuitively understand the scheme's performance across various evaluation dimensions; the system records the numerical changes of the criterion weights during each iteration, plotting a weight convergence curve in a line graph format, showing the evolution of the weights from the initial configuration to the final configuration; the system calculates the product of each criterion's sub-score and its corresponding final weight, marking the criterion with the largest product value as the key influencing criterion, indicating that this criterion contributes the most to the final comprehensive score and is the main factor influencing the target user group's preference judgment. Through the above outputs, the design team not only obtains the optimal styling scheme for the target market but also obtains personalized criterion weight configurations validated by user feedback and a visualized evaluation analysis report, providing data support for subsequent styling optimization and design decisions.

[0043] Please see Figure 2 The second embodiment of the present invention provides a vehicle styling determination device based on adaptive multi-criteria evaluation, comprising: A multimodal data processing unit is used to collect multimodal data of the car styling scheme to be evaluated, and to extract features from the multimodal data to obtain feature vectors for each mode. The fusion unit is used to align and fuse the feature vectors of each modality using a cross-modal fusion model to generate a unified fused feature representation. The multi-criteria evaluation unit is used to evaluate the fusion feature representation based on preset initial criterion weights using a multi-criteria evaluation model, calculate the sub-scores and comprehensive scores of each evaluation criterion, and rank multiple candidate styling schemes according to the comprehensive scores. The weight adjustment unit is used to obtain feedback data from the target user group on the current design scheme, calculate the degree of response of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. The scheme output unit is used to re-execute multi-criteria evaluation and scheme ranking based on the adjusted criterion weights until the preset convergence conditions are met, and then output the optimal modeling scheme and the corresponding criterion weight configuration.

[0044] The third embodiment of the present invention provides a vehicle styling determination device based on adaptive multi-criteria evaluation, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a vehicle styling determination method based on adaptive multi-criteria evaluation as described in any of the above embodiments.

[0045] This invention provides a method, apparatus, and device for determining automotive styling based on adaptive multi-criteria evaluation. The method involves collecting multimodal data of automotive styling schemes and extracting features. A cross-modal fusion model aligns and fuses the features of each modality into a unified feature representation. Based on initial criterion weights, the fused features are evaluated using multiple criteria to obtain individual scores and a comprehensive score, which are then used to rank candidate schemes. Feedback data from the target user group on the ranked schemes is then obtained. The responsiveness of each evaluation criterion to user feedback is calculated based on the fused feature representation and feedback data, thereby dynamically adjusting the weight configuration of each criterion. The adjusted weights are used to re-execute the evaluation and ranking, enabling the criterion weights to automatically optimize based on the actual feedback from a specific user group. This transforms the evaluation model from static fixed weights to dynamic adaptive weights, ensuring that the final optimal styling scheme matches the differentiated aesthetic preferences of the target market users.

[0046] Exemplary examples show that the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in implementing an adaptive multi-criteria evaluation-based vehicle styling determination device. For example, the apparatus described in the second embodiment of the present invention.

[0047] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the aforementioned adaptive multi-criteria evaluation-based vehicle styling determination method, connecting various parts of the method through various interfaces and circuits.

[0048] The memory can be used to store the computer program and / or modules. The processor, by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory, implements various functions of an adaptive multi-criteria evaluation-based vehicle styling determination method. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0049] If the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0050] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0051] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for determining vehicle styling based on adaptive multi-criteria evaluation, characterized in that, include: Multimodal data of the vehicle styling scheme to be evaluated are collected, and feature extraction is performed on the multimodal data to obtain the feature vector of each mode; A cross-modal fusion model is used to align and fuse the feature vectors of each modality to generate a unified fused feature representation. Based on the preset initial criterion weights, the fusion feature representation is evaluated using a multi-criterion evaluation model, the sub-scores and comprehensive scores of each evaluation criterion are calculated, and multiple candidate styling schemes are ranked according to the comprehensive scores. Obtain feedback data from the target user group on the sorted candidate design schemes, calculate the responsiveness of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. Based on the adjusted criterion weights, the multi-criterion evaluation and scheme ranking are re-executed until the preset convergence conditions are met, and then the optimal modeling scheme and the corresponding criterion weight configuration are output.

2. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 1, characterized in that, The multimodal data includes visual modal data, semantic modal data, and numerical modal data; The feature extraction process of the visual modality data includes: inputting the shape image into a pre-trained convolutional neural network, extracting the high-level semantic features of the image, and mapping them into a fixed-dimensional visual feature vector through a fully connected layer; The feature extraction process of the semantic modality data includes: inputting the design description text into the pre-trained language model, and taking the pooling output of the last hidden state as the semantic feature vector; The feature extraction process of the numerical modal data includes: normalizing the aerodynamic coefficients and ergonomic parameters, and then concatenating them into a numerical feature vector.

3. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 1, characterized in that, The cross-modal fusion model is used to align and fuse the feature vectors of each modality to generate a unified fused feature representation, specifically as follows: The feature vectors of each modality are mapped to a common feature space of the same dimension through a linear projection layer. Modality type encoding is added to the mapped feature vectors of each modality, and they are concatenated along the sequence dimension. The concatenated feature sequence is input into a multi-layer Transformer encoder, and cross-modal information interaction is achieved through a self-attention mechanism. The output of the Transformer encoder is pooled to obtain the fused feature representation.

4. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 1, characterized in that, The evaluation criteria include aesthetic criteria, functional criteria, and market criteria; The evaluation process of the multi-criteria evaluation model includes: The fusion feature representations are input into the scoring subnetworks corresponding to each evaluation criterion to obtain the sub-scores of each criterion. The sub-scores of each criterion are weighted and summed using the current criterion weights to obtain the comprehensive score of each candidate design scheme. The multiple candidate design schemes are sorted in descending order according to the comprehensive score.

5. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 4, characterized in that, The scoring process of the scoring subnetwork corresponding to the aesthetic criteria includes: The fused feature representation is input into the aesthetic scoring subnetwork, which sequentially includes a feature dimensionality reduction layer, a non-linear activation layer, and a scoring output layer. The scoring output layer outputs scores for the proportion coordination of the shape, symmetry, and dynamism. The scores for proportional harmony, symmetry, and dynamism are weighted and summed to obtain the individual scores for the aesthetic criteria.

6. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 4, characterized in that, The scoring process of the scoring subnetwork corresponding to the functional criteria includes: Extract function-related feature components from the fused feature representation; The feature components are concatenated with the aerodynamic coefficients and ergonomic parameters in the numerical modal data and then input into the functional scoring subnetwork. The functional scoring subnetwork outputs the sub-scores of the functional criteria.

7. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 1, characterized in that, Obtain feedback data from the target user group regarding the current design scheme, specifically: Obtain the rating data and text evaluation data submitted by the target user based on the sorted candidate design schemes; Statistical analysis is performed on the rating data to obtain a group preference score. Sentiment analysis and keyword extraction are performed on the text evaluation data to obtain a feedback feature vector. The group preference score and the feedback feature vector are combined to form the feedback data.

8. The method for determining vehicle styling based on adaptive multi-criteria evaluation according to claim 1, characterized in that, The dynamic adjustment of the weights of each evaluation criterion is specifically as follows: The evaluation deviation value is calculated based on the current comprehensive score and user feedback data. The fusion feature representation and the evaluation deviation value are concatenated and input into the weight optimization network. The weight optimization network outputs the weight adjustment amount of each evaluation criterion. The current criterion weight is added to the weight adjustment amount and normalized to obtain the adjusted criterion weight.

9. A vehicle styling determination device based on adaptive multi-criteria evaluation, characterized in that, include: A multimodal data processing unit is used to collect multimodal data of the car styling scheme to be evaluated, and to extract features from the multimodal data to obtain feature vectors for each mode. The fusion unit is used to align and fuse the feature vectors of each modality using a cross-modal fusion model to generate a unified fused feature representation. The multi-criteria evaluation unit is used to evaluate the fusion feature representation based on preset initial criterion weights using a multi-criteria evaluation model, calculate the sub-scores and comprehensive scores of each evaluation criterion, and rank multiple candidate styling schemes according to the comprehensive scores. The weight adjustment unit is used to obtain feedback data from the target user group on the current design scheme, calculate the degree of response of each evaluation criterion to user feedback based on the fusion feature representation and the feedback data, and dynamically adjust the weight of each evaluation criterion. The scheme output unit is used to re-execute multi-criteria evaluation and scheme ranking based on the adjusted criterion weights until the preset convergence conditions are met, and then output the optimal modeling scheme and the corresponding criterion weight configuration.

10. A vehicle styling determination device based on adaptive multi-criteria evaluation, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that can be executed by the processor to implement a vehicle styling determination method based on adaptive multi-criteria evaluation as described in any one of claims 1 to 7.

Images