Data-driven product family design method and system based on artificial intelligence

Abstract

The invention provides a data-driven product family design method and system based on artificial intelligence. The method comprises the following steps: respectively extracting appearance dominant DNA features and recessive DNA features of a to-be-designed product; product family appearance design is carried out in combination with dominant DNA features and recessive DNA features, pictures are preliminarily generated through text cues, and the cues are iterated until pictures meeting design requirements are generated; screening the pictures meeting the design requirements to obtain a plurality of pictures with highest scores, determining the appearance design of the to-be-designed product in combination with design elements in the dominant DNA feature pictures, and migrating all dominant feature heritable design elements in the appearance design of the to-be-designed product to the to-be-designed product to form a product family appearance design; according to the method, design objectivity and data driving are achieved, recessive DNA keywords are obtained by combining semi-structured interviews and perceptual intention investigation, fusion of emotional appeals and morphological characteristics is achieved, and a generated design scheme is made to conform to brand tonality and have innovativeness.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial design and human-computer interaction, and particularly relates to a data-driven product family design method and system based on artificial intelligence. BACKGROUND

[0002] In the fields of consumer electronics, instruments and medical devices, etc. with fierce market competition, enterprises often maintain brand recognition and user loyalty by building product families. A product family refers to a series of products with common core technology, function or brand genes, and the appearance design thereof needs to meet the functional differences of various product models on the basis of unified style. Traditionally, product family modeling relies on the experience and aesthetic judgment of designers, and lacks a unified analysis framework, resulting in inconsistent visual styles and a decrease in user brand recognition. The shortened development cycle of new products also makes it difficult for manual design to quickly iterate to meet market demand.

[0003] In order to systematically manage such visual language, the design field often borrows the concept of "design DNA", which includes the core design principles and characteristic elements of a product, and can be divided into two categories: explicit DNA, referring to perceptible and objective visual elements such as form, line, color, etc.; and implicit DNA, referring to intangible and subjective semantic attributes such as "simplicity", "stability" or "technological sense" perceived by users, etc.

[0004] However, the traditional product DNA definition and implementation process highly depends on the personal experience and aesthetic intuition of a few senior designers. This highly subjective method has many problems: poor consistency: the design results are greatly influenced by the personal style of the designer, and in large organizations or team changes, it is difficult to ensure the unity and continuity of the visual language of the product family; difficult to inherit and expand: design knowledge is mostly implicit, making it difficult to be clearly recorded, inherited and applied on a large scale, resulting in a gap in style when developing new products; low efficiency and lack of objective basis: the process of converting abstract brand value (implicit DNA) into specific form features (explicit DNA) is not transparent and is mainly driven by intuition, lacking data support and systematic framework, resulting in a long design iteration cycle and difficulty in assessing the degree of fit between the final solution and the brand strategy.

[0005] Therefore, there is an urgent need in the prior art for a systematic and data-driven product family design method that can overcome subjectivity and improve design consistency and efficiency. SUMMARY

[0006] In order to solve the problems of strong subjectivity, poor consistency and low efficiency of product family design process in the prior art, the present application provides a data-driven product family design method and system based on artificial intelligence, which integrates the objective analysis ability of computer vision and the creative exploration ability of generative artificial intelligence, establishes a systematic, quantifiable and replicable process for accurately defining and deploying the design DNA of products, thereby efficiently creating a product family with high brand recognition and consistency.

[0007] In order to achieve the above-mentioned purpose, in the first aspect, the present application provides a data-driven product family design method based on artificial intelligence, comprising the following steps: Respectively extracting the explicit DNA features and implicit DNA features of the appearance of the product to be designed; Combining the explicit DNA features and the implicit DNA features to design the appearance of the product family, generating pictures through text prompts, iterating the prompts until pictures meeting the design requirements are generated, screening the pictures meeting the design requirements to obtain a number of pictures with the highest scores, determining the appearance design of the product to be designed in combination with the design elements in the explicit DNA feature pictures, and migrating the design elements of each explicit feature in the appearance design of the product to be designed to the product to be designed to form the appearance design of the product family.

[0008] Further, the extraction of the explicit DNA features of the appearance of the product to be designed comprises the following steps: Firstly, the test machine is analyzed for competitive products from the aspects of function and appearance, and the appearance features of the test machine are preliminarily extracted; Secondly, the pictures of the test machine and similar instruments are collected and preprocessed, and the preprocessing includes adjusting the image size, removing the background and low-pixel photos; Thirdly, the test set, the training set and the validation set are divided, the top layer of the ResNet50 model is adjusted, and the correct rate of the adjusted ResNet50 model for recognition and classification is counted; Fourthly, the test machine picture samples are selected for test machine feature recognition, the corresponding heat map is obtained, the frequency of the appearance features in all test pictures is counted, and the three highest frequencies are selected as the explicit DNA features of the test machine.

[0009] Further, the pictures of the surrounding products of the product to be designed are obtained, and the obtained pictures are preprocessed to obtain picture data sets with uniform format, uniform pixels and uniform naming rules; Training and saving the ResNet50 model: the top layer of the ResNet50 model is replaced with a new layer, including a global average pooling layer and two fully connected layers, the last layer uses a softmax activation function, outputs the probability of 6 categories, the original ResNet50 layer is set as untrainable, and the features in the pre-trained weight are retained; Using the trained ResNet50 model to screen the heat map generated by the product to be designed, the three appearance features with the highest proportion are obtained as the test machine appearance explicit DNA features.

[0010] Further, the extraction of the implicit DNA features of the product to be designed includes: Through interviews with enterprise employees, the strategic goals and long-term vision of the enterprise, the product line and market strategy of the enterprise are obtained, the characteristics and functional requirements of each main product of the enterprise are collected, and a preliminary understanding of the overall features of the product to be designed is formed, and the interview records are sorted to obtain a word cloud diagram; Combined with the results of the enterprise user interview, the implicit key words with strong correlation to brand information are extracted; Based on expert opinions and research, the perceptual intention vocabulary of the test machine is obtained, the enterprise implicit key words are benchmarked to the intuitive perceptual intention vocabulary, and an antonym is matched for each intention vocabulary to obtain a pair of perceptual intention vocabulary. According to the scoring results of the perceptual intention questionnaire, statistical analysis is performed on the intention vocabulary, and the highest correlation word is screened out to obtain the implicit DNA features of the product to be designed.

[0011] Further, based on the professional field, core competitive advantage, key characteristics and technical specifications of the product, and brand identification content, an interview outline is developed; Based on the interview results, key words are extracted, the frequency of use of the words is quantitatively analyzed, meaningless words are removed, and a word cloud diagram is generated.

[0012] Further, according to the scoring results of the perceptual intention questionnaire, statistical analysis is performed on the intention vocabulary, and the highest correlation word is screened out to obtain the implicit DNA features of the product to be designed. Select multiple representative product pictures, all details of the product pictures are clear and visible, set the same perceptual vocabulary evaluation questions for each product picture in the questionnaire, and let the participants evaluate each picture using the same standard; the questionnaire uses a Likert seven-point scale, the positive and negative represent the positive and negative of the perceptual vocabulary, and the numerical value represents the degree; Invite test machine industry personnel, industrial design related personnel and ordinary users to evaluate the questionnaire respectively, and the participants score each picture according to the positive and negative words listed in the questionnaire; A stratified analysis method was used to calculate the mean and variance of the ratings for each image for each group. The mean was used to measure the overall evaluation trend, and the variance was used to assess the dispersion of the ratings. The average sample mean of the ratings for each image was then calculated. Finally, a weighted average was used to further calculate the words with the largest absolute value of the reinforced average sample mean of the emotional intention words.

[0013] Furthermore, the dominant and recessive DNA characteristics of the product to be designed are used as keywords in the text-based image prompts. Combined with usage scenarios, rendering, and material keywords, images are initially generated based on the keywords. Based on the initially generated images, the prompt words are adjusted, and the images are generated iteratively using the prompt words. The images are generated multiple times until the number and quality of the generated images meet the requirements, and then the images are classified according to dominant DNA characteristics. The images corresponding to the prompts in each dominant DNA trait category undergo a first round of screening, with four of the most valuable images retained for each prompt. A second round of screening is then conducted by design personnel, retaining the most valuable images for each prompt. The screening is based on a Likert five-point scale, meaning images with an average user score higher than a set value are retained. A rating questionnaire was created for the images corresponding to each category. Relevant experts and users were invited to evaluate and select the images. The questionnaire contained clear images of each dominant DNA feature scheme. Respondents rated the schemes from the perspective of different recessive DNA features. Based on the questionnaire results, the images with the highest scores that reflected the recessive DNA feature elements were selected. Based on the final selected images, design elements are extracted from each dominant DNA feature image to determine the appearance design of the product to be designed.

[0014] Secondly, the present invention provides a data-driven product family design system based on artificial intelligence, including a feature extraction module and a design module; The feature extraction module is used to extract the dominant and recessive DNA features of the appearance of the product to be designed, respectively. The design module is used to combine dominant and recessive DNA features to design the appearance of product families. It initially generates images based on text prompts, iterates through the prompts until images that meet the design requirements are generated, filters the images that meet the design requirements to obtain the top-scoring images, determines the appearance design of the product to be designed by combining the design elements in the dominant DNA feature images, and transfers the heritable design elements of each dominant feature in the appearance design of the product to be designed to the product to be designed, thus forming the appearance design of the product family.

[0015] Furthermore, the feature extraction module includes a dominant DNA feature extraction unit and a recessive feature extraction unit; The dominant DNA feature extraction unit is used to analyze from both functional and appearance perspectives, initially extracting the appearance features of the test machine; collecting images of the test machine and similar-looking instruments and preprocessing them, including adjusting image size, removing backgrounds, and removing low-resolution photos; dividing the test set, training set, and validation set, adjusting the top layer of the ResNet50 model, and statistically analyzing the classification accuracy of the adjusted ResNet50 model; screening test machine image samples for feature recognition, obtaining corresponding heatmaps, statistically analyzing the frequency of appearance features in all test images, and selecting the three with the highest frequency as the dominant DNA features of the test machine; The latent DNA feature extraction unit is used to combine enterprise user interview results to extract latent keywords that are strongly related to brand information. Based on expert opinions and research, the unit obtains the emotional intention words of the test machine, matches the enterprise's latent keywords with intuitive emotional intention words, and pairs each intention word with an antonym to obtain emotional intention word pairs. Based on the scoring results of the emotional intention survey questionnaire, statistical analysis was conducted on the intention words, and the words with the highest relevance were selected to obtain the implicit DNA characteristics of the product to be designed.

[0016] Furthermore, the design module includes an analysis unit that employs stratified analysis to calculate the mean and variance of each person's rating for each image based on the questionnaire results. The mean is used to measure the overall evaluation trend, and the variance is used to assess the dispersion of the ratings. Then, the average sample mean of the ratings for each image is calculated. Finally, a weighted average is used to further calculate the words with the largest absolute value of the reinforced average sample mean of the emotional intention words.

[0017] Compared with existing technologies, this invention has at least the following beneficial effects: This invention achieves objectivity and data-driven design, employing deep residual networks such as ResNet50 to extract dominant DNA features of products, accurately identifying key appearance elements and possessing strong generalization ability; the residual structure alleviates degradation problems through skip connections, making the training of deep networks more stable. Combining implicit DNA keywords obtained from semi-structured interviews and emotional intent surveys, it achieves the fusion of emotional appeal and morphological characteristics, making the generated design scheme both consistent with brand tone and innovative. This method can systematically construct a coherent and unique visual language and ensure its consistent application throughout the entire product family, thereby improving product consistency and recognizability; verification experiments show that product families designed using this method achieve a 100% accuracy rate in brand classification tasks, higher than traditional designs, demonstrating significant effectiveness. In terms of execution efficiency, this invention utilizes generative AI to explore thousands of design possibilities in a short time, greatly expanding the breadth of creative ideas, and quickly converges to high-value solutions through a structured screening process, thereby improving design efficiency and the breadth of innovation, and shortening the cycle from concept to final product.

[0018] This invention provides an end-to-end, standardized operating framework that achieves systematization and replicability, making design knowledge explicit and process-oriented, facilitating replication and inheritance in different projects and teams. It clearly defines the roles of AI and human designers in the design process—AI is responsible for objective analysis and divergent exploration, while humans are responsible for strategic setting, semantic interpretation, and creative synthesis—forming a complementary and highly efficient optimized human-machine collaboration model. Attached Figure Description

[0019] Figure 1 A schematic diagram of the product family appearance design methodology framework assisted by AIGC.

[0020] Figure 2 A schematic diagram illustrating the framework of a product family appearance design method that combines MJ and SD tools with DNA features.

[0021] Figure 3 A schematic diagram of the design framework for the product family.

[0022] Figure 4 This is a schematic diagram of the hot zone of the feature recognition algorithm.

[0023] Figure 5 The results of the hot zone feature identification operation.

[0024] Figure 6 Word cloud for interview transcripts.

[0025] Figure 7 Word cloud of the interview transcript after removing irrelevant words.

[0026] Figure 8 This is a diagram of the shape segmentation scheme.

[0027] Figure 9 A design scheme for decorative lines.

[0028] Figure 10 This is a diagram of a heat dissipation mesh design.

[0029] Figure 11 These are the results of the second round of screening.

[0030] Figure 12 The heat dissipation mesh is converted into a hand-drawn illustration.

[0031] Figure 13 This is the final reference design drawing.

[0032] Figure 14 The results of the exploration and refinement of the FG-1's appearance.

[0033] Figure 15 The results of the exploration and refinement of the appearance of ZS-3 are shown in the image.

[0034] Figure 16 The design results of the front face scheme exploration are shown in the figure.

[0035] Figure 17 The results of the exploration of fuselage design scheme are shown in the figure.

[0036] Figure 18 The result diagram of the test machine's appearance design scheme.

[0037] Figure 19 A schematic diagram illustrating the process of exploring the appearance design of the test machine.

[0038] Figure 20 The product family's appearance design includes (a) a representative design drawing of the test machine, (b) a representative design drawing of the slot-type instrument chassis, and (c) a representative design drawing of the desktop test machine. Detailed Implementation

[0039] 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, not all, of the embodiments of the present invention. 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.

[0040] Taking a test machine product designed by a certain radio frequency test company as an example, the method described in this invention will be explained in detail.

[0041] Product DNA comprises both dominant and latent characteristics. Dominant DNA characteristics include the product's appearance and design, which are the most immediate features consumers encounter upon first contact. These include the product's shape, color, materials used, and basic design principles. Latent DNA characteristics originate from the company's philosophy, cultural heritage, and brand essence, and are gradually understood by consumers through use and experience, such as product quality and user experience. Figure 3 As shown, dominant DNA features are extracted and screened from competitor products of the testing machine to guide the appearance design of the product family. Analysis of corporate culture and brand information yields preliminary implicit DNA features, which can directly guide the product's functionality and user experience design. However, these implicit DNA features need to be transformed when designing the product family's appearance by converting them into emotionally resonant vocabulary to guide the design of the testing machine.

[0042] Product family design encompasses aspects such as appearance, functionality, and user experience. This invention primarily introduces the appearance design of product families, which includes two aspects: first, the determination of dominant and recessive DNA characteristics; and second, the application of these characteristics. For example... Figure 1As shown, the process is divided into three modules. Module one determines the dominant DNA features by analyzing the appearance and functions of competing products to obtain the appearance characteristics of the test machine. An AI image recognition algorithm based on the ResNet50 model is used for deep learning training and recognition of images. After data analysis and filtering, the dominant DNA features of the test machine's appearance are determined. Module two determines the recessive DNA features by acquiring data from user interviews, enterprise implicit keywords, and emotional intention surveys to extract the product's recessive DNA features. Module three integrates the results of the first two modules, using the MJ (Midjourney) and SD (StableDiffusion) intelligent image generation tools to generate images. The generated images are then filtered to obtain images with dominant DNA features for reference. The appearance design of the test machine is then combined with these dominant DNA feature images, and the style is transferred to target products, such as general-purpose chassis and desktop test machines, as examples to achieve the product family appearance design.

[0043] Extracting dominant DNA features of a test device's appearance using AI image recognition algorithms: For extracting dominant DNA features of a product's appearance, dominant features are typically identified from three aspects: shape, color, and material. Since color and material are largely determined at the initial design stage, only the product's shape needs to be evaluated. Traditional methods include: eye-tracking experiments, which record subjects' gaze points and gaze trajectories on test device images to infer the location and importance of dominant DNA features in the images. However, this method involves expensive eye-tracking equipment, a complex and time-consuming experimental process, and a small sample size. Expert surveys, which invite experts in the field to label and evaluate the dominant DNA features of the test device's appearance, suffer from inconsistencies in labeling the same image by different experts, a small sample size, and are typically time-consuming and inefficient.

[0044] Based on this, this invention uses an AI image recognition algorithm to extract dominant DNA features of the test machine's appearance. The AI ​​image recognition algorithm can automatically identify the outline, color, shape, and pattern of objects in an image, and then classify the image. By collecting a large number of images, a deep learning-based convolutional neural network automatically extracts features from the collected images, inputs the extracted features into a neural network, and adjusts the weight parameters so that the model can predict the category of the input image. The image recognition algorithm can quickly analyze large amounts of image data and process large amounts of sample data, improving efficiency, reducing bias caused by subjective human judgment, and is unaffected by factors such as human emotion and fatigue, thus improving the stability of feature extraction.

[0045] Taking the design of a testbed as an example, a ResNet50 model is used with adjustments to the top layer. The ResNet50 model employs a residual learning framework, directly passing input information to subsequent layers through skip connections. This allows the model to extract rich features from images, making it particularly suitable for processing complex image data. This application uses the ResNet50 model as the basic architecture, replacing the top fully connected layer with a fully connected layer used in testbed classification tasks, which improves the classification accuracy for instrument-related images such as testbeds. Through transfer learning on the ResNet50 model, instrument-related images such as testbeds are identified and classified. To intuitively see which features the model uses for classification, Grad-weighted Class Activation Mapping (Grad-CAM) is employed to explain the model's behavior during predictions for specific categories. When making predictions, Grad-CAM generates heatmaps to highlight the most important regions in the model's visual input, indicating that the appearance features corresponding to the most important regions play a crucial role. Overlaying a heatmap onto the original image allows for a more intuitive identification of image features. Specifically, this involves using functions from the Open Source Computer Vision Library (OpenCV) to convert the standardized heatmap into a pseudo-color image, and then overlaying it onto the original image by adjusting the transparency, ultimately obtaining a visualization of the model's features.

[0046] The extraction of dominant DNA characteristics from the testing machine specifically includes the following steps: The first step is to conduct a competitive analysis of the test machine, analyzing it from both functional and appearance perspectives. The main purpose of the functional analysis is to accurately identify each functional component of the test machine, while the appearance analysis identifies the appearance features of each test machine and initially extracts its appearance characteristics.

[0047] The second step is to collect a large number of images of test machines and instruments with similar appearances and preprocess them. Preprocessing includes adjusting image size, removing backgrounds, and removing photos with too low resolution.

[0048] The third step is to divide the model into test, training, and validation sets. Then, write code to train the model, adjust the top layers of the ResNet50 model, and calculate the accuracy of the new model's classification.

[0049] The fourth step involves selecting test machine image samples for feature recognition to obtain the corresponding heat map. The frequency of appearance features in all test images is statistically analyzed, and the three most frequent features are selected as the dominant DNA features of the test machine.

[0050] Extracting the latent DNA characteristics of the product to be designed: A company's latent DNA characteristics refer to those features that are not easily observed directly but profoundly influence corporate culture and product image. These are usually closely related to the company's brand information and are not easily extracted directly through simple surveys or data analysis. To accurately obtain the latent DNA characteristics of the product to be designed, a method based on user interviews and emotional intention questionnaires is used. The steps of this method are as follows: First, enterprise user interviews: By interviewing company employees, we gain a deeper understanding of the company's strategic goals and long-term vision, as well as its product lines and market strategies. We collect information on the characteristics and functional requirements of the company's main products to form a preliminary understanding of the overall features of the product to be designed. We organize the interview records to obtain word clouds, which helps in the extraction of brand information, implicit keywords, and the final determination of implicit DNA characteristics.

[0051] Second, extraction of implicit keywords: Based on the results of enterprise user interviews, implicit keywords with strong relevance to brand information are extracted. These keywords represent the core connotation and value of the enterprise brand.

[0052] Third, the transformation from implicit keywords to emotive keywords: Based on expert opinions and research on the emotive keywords of the test machine, the company's implicit keywords are matched with intuitive emotive keywords. Emotive keywords better convey the overall feeling of the product and are also easier to guide the appearance design of the test machine. For example, implicit keywords such as "innovation," "technology," and "R&D" are transformed into emotive keywords related to "technology," and each emotive keyword is paired with an antonym to obtain emotive keyword pairs.

[0053] Fourth, the design and selection of the emotional intention survey questionnaire: Emotional intention vocabulary is obtained through surveys of corporate culture and brand information. For the appearance of the test machine, a questionnaire rating system is still needed for selection. The questionnaire includes having respondents rate the relevance of intentional word pairs in images of the test machine to select the words that best represent the company's test machine product. Based on the questionnaire rating results, statistical analysis is performed on the intentional words, and the words with the highest relevance are selected as the implicit DNA characteristics of the product to be designed.

[0054] Product family appearance design using MJ and SD tools combined with dominant and recessive DNA features: Based on the results of Module 1 and Module 2, a product family appearance design method combining dominant and recessive DNA features with the AI ​​image generation tools Midjourney and Stable Diffusion is used.

[0055] Midjourney provides designers with diverse sources of inspiration, quickly generating high-quality images related to product design through simple text descriptions. Designers can easily generate different types of images based on prompts. Stable Diffusion can transform user-input text descriptions into a rich variety of images. Specifically: First, the dominant and recessive DNA characteristics of the determined test machine are used as keywords in the MJ and SD text-to-image prompts. Combined with usage scenarios, rendering, and material keywords, the prompt content is improved and images are initially generated.

[0056] Second, based on the initially generated images, the prompt words are adjusted to improve image quality and more accurately represent dominant and recessive DNA characteristics. The process of generating images using prompt words is iterated continuously, selecting prompt words with ideal image effects and generating images multiple times until the quantity and quality of generated images meet the requirements. The images are then categorized according to dominant DNA characteristics.

[0057] Third, the images corresponding to the prompts in each dominant DNA trait category are screened in the first round, with four of the most valuable images retained for each prompt. Then, the design team conducts a second round of screening, retaining the most valuable images for each prompt, based on a Likert five-point scale, i.e., images with an average user score of 3.5 or higher are retained.

[0058] Fourth, the rating questionnaire and final selection: A rating questionnaire was created for the images corresponding to each category, and relevant experts and users were invited to evaluate and select the best options. The questionnaire included clear images of each dominant DNA trait scheme, and respondents could rate the schemes from the perspective of different recessive DNA traits. The questionnaire results were analyzed, and the images that best represented the recessive DNA trait elements were selected.

[0059] Fifth, test machine design and modeling rendering: Based on the final selected images, design elements are extracted from each dominant DNA feature image to determine the design scheme of the test machine.

[0060] Sixth, the heritable design elements of the dominant features in the test machine design scheme are transferred to general chassis and desktop test machines to form an extended product family appearance design.

[0061] refer to Figure 2 This invention utilizes AI image generation technology and screening strategies to ensure the efficiency of the design process and fully leverage the determined dominant and recessive DNA characteristics, providing enterprises with inheritable product family appearance design solutions.

[0062] AI Image Recognition Algorithm Based on ResNet50 Model Extraction of Dominant DNA Features: Taking the test machine as an example, we will implement the specific steps of Module 1 to extract the dominant DNA features of the test machine's appearance.

[0063] The purpose of competitive analysis is to comprehensively understand the direct competitors and their products in the market. By deeply analyzing competitors' test equipment products, we can identify the functional and appearance features of test equipment on the market. Further analysis allows us to initially extract factors influencing product recognizability, namely, the appearance features of the test equipment. To gain a deeper understanding of the competitive landscape of the test equipment market, we selected five well-known domestic and international chip test equipment manufacturers and three international manufacturers for detailed analysis.

[0064] Functional Analysis of the Test Machine: The main purpose of functional analysis is to identify the various functional components of the test machine, which serve as its external features. A product can be viewed as a technical system integrating various main functions. The operation of the main functions requires the cooperation of their sub-functions, and the operation of the sub-functions requires specific components. Therefore, it is necessary to determine which main functions the product should possess, analyze the corresponding sub-functions through these main functions, and then deduce the components required by the product. The main functional components of the test machine include dust covers, engineering carts, interface panels, heat dissipation mesh, caster bases, buttons, and handles.

[0065] To improve the model's accuracy in recognizing instrument-related images, a large number of such images need to be collected. These instruments share similarities with testing machines in appearance and functional components. For example, many testing devices are equipped with sophisticated input / output ports, control panels, and heat dissipation meshes—common components of testing machines. As shown in Table 1, the final instrument categories selected, excluding testing machines, are automated test equipment, RF analyzers, wafer probe stations, PCB testing equipment, and oscilloscopes. Abbreviations are used for easier naming of the images later, and the similarities between each category and the appearance of testing machines have been analyzed.

[0066] Table 1 Information on Five Types of Instruments

[0067] Find images in these six categories through image datasets (such as ImageNet, Kaggle, Open Images Dataset), search engines (such as Google, Bing, Baidu), professional websites or resource libraries (such as Getty Images, Unsplash, Pixabay, QianTu.com), and the official websites of the corresponding product companies.

[0068] The main purpose of image preprocessing is to improve the quality of image data, thereby enhancing the effectiveness of subsequent model training and the accuracy of predictions. Image preprocessing is performed according to the following steps: (1) Manually screen the images, removing those with too low resolution or that are too blurry. Adjust all images to a uniform resolution and size, scaling all images to 256×256 pixels.

[0069] (2) Simplify the background of the image. Cut out images with cluttered backgrounds. This will make it easier for the model to recognize image features.

[0070] (3) Adjust the contrast and brightness of the image to make the key features in the image more prominent, so that the algorithm can recognize them.

[0071] (4) Adjust the image format and naming. Each image should be in "jpg" format, and the naming rule should be "abbreviation-xxx". For example, the image of the ATE automatic testing equipment should be named "ATE-001.jpg". At the same time, ensure the accuracy of the model prediction and that the number of images is the same.

[0072] After the above steps, a total of 612 images were collected, including 102 images each of the Automated Test Equipment (ATE), RF Analyzer (FXY), Wafer Probe Station (JY), PCB Test Equipment (PCB), Oscilloscope (SBQ), and Chip Tester (ceshiji).

[0073] Training the model requires splitting the original image data to ensure good generalization ability during training and testing. The `sklearn.model_selection.train_test_split` function in Python is used to split the images, ensuring random and reasonable allocation to the training, validation, and test sets. 70% are used for training, 15% for validation, and 15% for model testing. Images are then copied from the original folder to the corresponding training, validation, and test set folders using `shutil.copy` for easier subsequent processing.

[0074] Model Architecture: The ResNet50 model contains 50 deep layers and uses a residual learning framework to train the deep network. The model utilizes weights pre-trained on the ImageNet dataset, which helps improve classification accuracy, especially under limited data conditions.

[0075] Training strategy: Replace the top layers of the ResNet50 model with new layers to adapt to the specific classification task. This includes a global average pooling layer and two fully connected layers, with the last layer using a softmax activation function to output the probabilities of the six classes. Except for the newly added layers, the original ResNet50 layers are set to be untrainable, preserving the features from the pre-trained weights.

[0076] The model is trained on the training set, with each iteration using a subset of all available data (32 samples per batch), for a total of 10 iterations to optimize the model parameters. Simultaneously, the model performance is evaluated on a separate validation set to monitor for overfitting during training. The model is saved as final_model_resnet50.h5.

[0077] In the dataset split, 15% of the test set was not used during model training. These images were used to evaluate the model's classification accuracy. The test set contained 15 images for each category. The category with the highest output probability was selected as the final predicted category, and the accuracy was calculated by counting the number of correct predictions. Statistical analysis showed that the average classification accuracy exceeded 50%, and the accuracy of random guessing was approximately 16.67% (1 / 6), thus meeting expectations.

[0078] like Figure 4 To identify dominant DNA features in the test images, Grad-CAM technology was used to generate heatmaps. A pre-trained `final_model_resnet50` model, trained on the ImageNet dataset, was used to identify six different instruments. Each test image was resized to 224×224 pixels and processed using ResNet50's preprocessing functions. The heatmaps reflect the regions in the test images that the model considers most relevant to the predicted category; these regions are then considered to contain key features.

[0079] Feature statistics and filtering: such as Figure 5 As shown, from the test images on the testing machine, images with small angular deviations, similar appearances, and low quality are removed. Thirty high-quality test images are selected for testing, resulting in 30 corresponding heat map images.

[0080] As shown in Table 2, the test images do not contain all appearance features. The frequency of each appearance feature in the 30 test images is counted. For example, 27 out of the 30 test images contain heat dissipation mesh, so the frequency of the heat dissipation mesh feature is 27. Based on the heat map generated by Grad-CAM, the frequency of test machine features covered by the heat map is identified and recorded. For example, 27 out of the 30 test images contain heat dissipation mesh, and in 20 of these heat map images, the heat dissipation mesh is covered; therefore, the heat map feature frequency of the heat dissipation mesh feature is 20. The proportion of each appearance feature is calculated. For example, the proportion of heat dissipation mesh is 74.07%, meaning that in 20 out of the 27 test machine images containing heat dissipation mesh, the heat map images cover the heat dissipation mesh, accounting for 74.07%. This indicates that the algorithm model selects heat dissipation mesh as an important appearance feature for classification with nearly 74% probability. The three appearance features with the highest proportions are selected as the dominant DNA features of the test machine appearance.

[0081] In summary, the explicit visual DNA features of the test unit are its body shape segmentation, decorative lines, and heat dissipation mesh. These visual features are the key areas of focus when classifying models, serving as the basis for classification. Therefore, they represent the most critical visual characteristics of the test unit and have significant guiding significance for the appearance design of the product family.

[0082] Table 2. Statistics and Screening of Appearance Features

[0083] Extracting latent DNA features of the product to be designed: When extracting latent DNA features of the product to be designed, user interviews, hidden keywords and emotional intention questionnaires are combined to systematically mine the latent DNA features of the product to be designed.

[0084] Corporate Strategy Analysis: Through user interviews, we gain a deeper understanding of the company's strategic goals and long-term vision, its development history, core mission, and main strategic directions in the RF testing field. We obtain key features and technical specifications of the company's RF test equipment products, gaining a thorough understanding of their main functions, performance advantages, and technological innovations. According to David Aaker's brand identity theory, brand identity comprises nine aspects: corporate spirit, corporate philosophy, corporate culture, product characteristics, quality, user experience, brand history, brand personality, and brand visual identity (VIS). However, not every company possesses all of these identity elements; additional content can be selected or supplemented based on the company's specific circumstances. Obtaining brand information through user interviews will help reveal the company's core values ​​and cultural philosophy.

[0085] An interview outline was developed, as shown in Table 3, to comprehensively gather brand information from the company in order to extract latent DNA characteristics and guide the appearance design of subsequent product families.

[0086] Table 3 Interview Outline

[0087] Interviews were conducted with five employees of the company, and the interview records were organized and analyzed using Nvivo software. The interview data was converted into a matrix format using matrix coding for statistical analysis and comparison, followed by quantitative analysis. The final result was presented in the form of a word cloud. Figure 6 As shown, word frequency analysis can reveal frequently mentioned keywords in interviews, helping to identify the focus of the interview content. Simultaneously, by quantitatively analyzing the frequency of word usage in interviews, the importance of different topics can be objectively measured. The extraction of implicit keywords provides data support, making the generation process of implicit keywords more scientific.

[0088] The word cloud after removal is as follows:Figure 7 As shown in the image, many words in the word cloud, such as "market," "product," "company," and "enterprise," although frequently used, are not helpful in generating implicit keywords for businesses and therefore need to be removed. Additionally, frequently used verbal words, such as "we," "certainly," "like this," and "therefore," are also of little reference value for generating implicit keywords and should also be removed.

[0089] Organizing the interview transcripts will help extract brand information and hidden keywords for Mingjian, and ultimately determine its hidden DNA characteristics.

[0090] Extracting Hidden Keywords: Based on research and interviews with enterprise users, and combined with David Aaker's brand identity theory, the Mingjian brand information includes 10 aspects: corporate spirit, corporate positioning, corporate culture, corporate mission, corporate vision, values, flagship leading products, core design principles, surface treatment, and development direction. As shown in Table 4, the left column represents the brand identity content, and the right column corresponds to the explanatory information.

[0091] Table 4. Information on a Certain Brand

[0092] As shown in Table 5, by analyzing a brand's information and combining it with the word cloud generated from the interview results, we obtained implicit keywords that are strongly related to the company's brand information.

[0093] Table 5 Latent Keywords

[0094] Implicit keywords driven by brand connotation often fail to intuitively and appropriately express people's perceptions and feelings about a product. Therefore, it is necessary to further transform implicit keywords into target words. As shown in Table 6, target words describing semiconductor testing products were obtained through literature review, online collection, magazine articles, and consumer descriptions. These words need to meet users' desired perceptions of the tester's appearance and are divided into four categories: appearance, user experience, quality, and emotional.

[0095] Table 6. Collection of Intended Keywords for Semiconductor Testing Products

[0096] As shown in Table 7, words with repetitive meanings in terms of emotional intention were removed and compared with implicit keywords.

[0097] Table 7. Implicit Keywords and Corresponding Sentimental Keywords

[0098] As shown in Table 8, words that clearly do not match the description of the appearance in the benchmarked emotional connotation words were removed, resulting in 7 words that can reflect the implicit keywords of the sword: rugged, steady, cutting-edge, technological, precise, lively, and simple. Each connotation word was paired with an antonym, resulting in 7 word pairs.

[0099] Table 8. Emotional Intentional Vocabulary Pairs for Semiconductor Testing Products

[0100] Because the source of emotional intention vocabulary is the implicit keywords of semiconductor testing product benchmark companies, which are continuously filtered, the screening process is influenced by personal bias and subjective judgment. Therefore, it cannot be directly used as the implicit DNA characteristics of the testing machine. Questionnaires provide a standardized way to collect information, which can reduce the influence of personal bias and subjective judgment. Therefore, an emotional intention questionnaire can be created to further screen the emotional intention vocabulary of the testing machine as implicit DNA characteristics. Three representative images of the testing machine were selected, ensuring that all details of the images were clearly visible to avoid the impact of poor image quality on participants' evaluations. The questionnaire included the same emotional vocabulary evaluation questions for each image, allowing participants to evaluate each image using the same set of standards. The questions used a seven-point Likert scale, with positive and negative indicating the positive or negative of the emotional vocabulary, and the numerical value representing its degree. Test machine industry personnel, industrial design personnel, and general users were invited to evaluate the questionnaire. Participants rated each image based on the positive and negative vocabulary listed in the questionnaire (rating range from -3 to 3). The closer the score for positive words is to 3, the better the image reflects the effect of positive words; conversely, the closer the score for negative words is to -3, the better the image reflects the effect of negative words. A total of 108 questionnaires were collected, including data from 17 people in the testing machine industry, 20 people in industrial design, and 71 ordinary users. Before data analysis, the collected data was first screened and cleaned to ensure its validity and accuracy. Clearly invalid questionnaires, such as those completed in too short a time or with all scores of 0, were removed. After these steps, 81 valid questionnaires were selected, including 17 from testing machine industry personnel, 18 from industrial design personnel, and 46 from ordinary users for further analysis.

[0101] As shown in Table 9, to analyze the differences in evaluations of each image by participants with different identities, a stratified analysis method was used for data analysis. This involved calculating the mean and variance of the ratings for each image for each participant group. The mean was used to measure the overall evaluation trend, and the variance was used to assess the dispersion of the ratings. The sample mean of the average ratings for the three images was then calculated.

[0102] Table 9. Average Sample Means of Three Types of Personnel

[0103] As shown in Table 10, since different groups have varying degrees of importance in product evaluation, a weighted average is needed for further calculation. Testing industry personnel are more familiar with the product's technical details and have a higher weight (40%); industrial design personnel have deeper insights into design aesthetics and have a lower weight (35%); feedback from ordinary users is also crucial, but has a slightly lower weight than that of technical and design professionals (25%). Using a weighted average method can balance the evaluations of each group, avoiding the extreme influence of one group's evaluation on the overall result. Furthermore, by using a weighted average, the different perspectives of each group can be comprehensively considered, resulting in a more comprehensive and objective evaluation result.

[0104] Table 10 Weighted Average Sample Mean

[0105] The calculation results show that the three emotional connotations of "stable," "technological," and "simple" have the largest absolute values ​​of reinforced average sample mean. Therefore, the implicit DNA characteristics of the company's testing machine products are determined to be "simple," "stable," and "technological."

[0106] The following details how to use MJ and SD tools to combine the dominant and recessive DNA characteristics of a product for product family appearance design. Through steps such as generating images from text prompts, iterating through the prompts, conducting two rounds of image selection, using a scoring questionnaire, determining the appearance design scheme on a test machine, and style transfer, the product family appearance is obtained, and the extensibility of the product family appearance design is verified.

[0107] Image Generation Approach: MJ is an AI-powered image generation tool based on text. It generates images based on user-input descriptive text prompts, which include specific design requirements, desired visual effects, and overall design style. The system analyzes these prompts and generates a set of images for the user to choose from and further optimize. SD is similar to MJ, both generating images based on prompts, but SD's image-to-image generation function is more powerful.

[0108] The dominant DNA characteristics are organism shape segmentation, decorative lines, and heat dissipation mesh, while the recessive DNA characteristics are "simple," "stable," and "technological." Therefore, different shape segmentation, decorative line, and heat dissipation mesh schemes can be created. Through continuous iteration and optimization, several design schemes for reference can be generated. As shown in Table 12, since SD and MJ image outputs require relatively precise cue word control, the DNA characteristic factors can be converted into keywords for generating images.

[0109] Table 11 Generated Image Keywords

[0110] As shown in Table 12, once you have the keywords for generating images, you can try to generate prompts, such as generating a scheme for shape segmentation with a simple style.

[0111] Table 12 Midjourney prompts

[0112] Therefore, the final prompt text generated is: imagine prompt: A minimalist chip testing machine design reference, showcasing a segmented layout. The design should feature clean lines, a modern aesthetic, and clearly defined sections for different components such as the control panel, chip slot, and interfaceports. The color scheme should be neutral with metallic and matte finishes. The machine should be displayed in a simple, uncluttered environment to emphasize its sleek design. --v 6 --ar 1:1 --quality 2 --style minimalistic.

[0113] Table 13 shows the suggested keywords for Stable Diffusion (SD). In addition to positive keywords, SD also includes negative keywords, primarily to avoid low-quality, non-compliant responses. Furthermore, SD's suggested keywords do not require detailed descriptions; only the keywords are needed.

[0114] By incorporating DNA-related keywords such as "stylist segmentation," "simple," and "stable" into positive prompts, different types of positive prompts can be generated. Adjusting negative prompts can make image generation more stable and the resulting images more informative.

[0115] Table 13 Stable Diffusion Tips

[0116] Set the parameters of SD, input positive and negative prompt words into SD, use realisticVisionV60B1_v51VAE.safetensors [15012c538f] for the large model, and perform graph-to-graph iteration on the reference images.

[0117] Image Generation Iteration and Initial Screening: During the modification of the MJ and SD prompts, some prompts generated images that were disorganized and offered little reference value for the design. Therefore, iterative processing of both the prompts and images was necessary. This involved generating multiple prompts, selecting suitable images for iteration, and then choosing those images as the reference images for each prompt. Thus, each prompt generated multiple reference images. For example, for the shape segmentation and decorative line schemes, there were multiple prompts and corresponding reference images for each prompt. To ensure the reliability and scientific rigor of the subsequent screening, the above two methods selected six prompts with high-quality output and high reference value, and for each prompt, four more representative and valuable images were selected. For the heat dissipation mesh scheme, since it only involved pattern design, it was relatively simple, and six images were directly selected.

[0118] A design segmentation scheme refers to a product's appearance using geometric segmentation and structural design to present different visual effects and functional zones. This involves only the visual and structural design of the outer shell, such as outlines, dividing lines, panel divisions, and protruding shapes. When selecting a design segmentation scheme, factors such as geometric aesthetics, the rationality of functional zoning, manufacturability, structural strength, and brand consistency are comprehensively considered. Geometric aesthetics require the design to have visual balance and harmony. The rationality of functional zoning ensures that dividing lines and panel divisions are scientifically sound and easy for users to operate and maintain. Manufacturability requires the design to be easy to produce and assemble, avoiding complex dividing lines and shapes to reduce production difficulty and costs. Furthermore, the exterior color is uniformly white to avoid color affecting the selection process.

[0119] like Figure 8As shown, the final selection of the styling segmentation scheme yielded 6 prompts, with 4 images selected for each prompt. Each image was named according to the prompt number and image number. Decorative line schemes refer to decorative lines added to the product's appearance. These lines not only enhance visual appeal but also strengthen the overall design style. When selecting decorative line schemes, the lines should enhance the product's aesthetics, harmonizing with the overall design style and improving the brand image. Simultaneously, these lines should not only serve a decorative function but also enhance the product's functionality to some extent. Designs easily implemented using existing processes should be chosen, ensuring suitability for the materials used and avoiding increased production complexity and costs. Emphasis should be placed on design uniqueness to ensure product market recognition. Since the color and material of the testing machine have been determined, to avoid color influence on the selection, blue was uniformly used as the color scheme for the decorative lines to ensure that the evaluation focuses on the line design itself. Figure 9 As shown, the decorative line scheme ultimately yielded 6 key words, with 4 images selected for each key word. Each image was named according to the key word's English number and the image's number. For example... Figure 10 As shown, by adjusting the prompts, heat dissipation mesh designs were obtained. When selecting mesh patterns, it was ensured that the mesh design could facilitate airflow and heat dissipation. The mesh pattern should be consistent with the overall design style of the product to enhance visual appeal. Patterns easily implemented using existing processes were chosen to avoid increasing production costs and complexity. Six mesh pattern images were selected. Several industrial design students were invited as participants, as their design background helped them better understand the selection criteria for the testing machine. As shown in Table 14, the selection criteria for the shape segmentation scheme and decorative line scheme were explained to the participants. Each student independently evaluated and selected the image they believed best met the above criteria.

[0120] Table 14 Screening Criteria

[0121] The number of people who selected the image corresponding to each prompt word was counted, and the image with the most selections was selected. This list of images that best met the selection criteria for each prompt word was then compiled and categorized. For example... Figure 11 As shown, the reference images that emphasize the front design are grouped into one category, while the images that emphasize the body design are grouped into another category. These images will be used as samples for the subsequent style preference rating questionnaire.

[0122] The latent DNA trait scoring questionnaire systematically collects subjective evaluations from experts and users on different design schemes to determine the design scheme that best reflects the latent DNA traits (simple, stable, and technological) of the product to be designed.

[0123] After two rounds of image screening, images with styling and decorative line schemes were obtained for design reference. However, because some of the images of the heat dissipation mesh were hand-drawn and some were actual object images, and the image quality of different actual object images varied, in order to reduce the impact of these factors on testers when evaluating the mesh pattern, such as... Figure 12 As shown, the heat dissipation mesh holes are uniformly converted into hand-drawn diagrams.

[0124] The questionnaire was still targeted at individuals in the testing industry, industrial design professionals, and general users. A total of 58 questionnaires were collected, including 23 from testing industry professionals, 18 from industrial design professionals, and 17 from general users. After removing questionnaires submitted too recently, 22 questionnaires remained from testing industry professionals, 16 from industrial design professionals, and 12 from general users. The mean score for each group was calculated separately, and then a weighted average was used for further calculation. The weighting percentages remained the same: testing industry professionals (40%), industrial design professionals (35%), and general users (25%). SPSS was used to calculate the mean scores for different latent DNA features for each image from different individuals, and then the weighted average score for each image was calculated for the three groups. As shown in Table 15, the styling segmentation scheme and decorative line scheme were categorized by front-face and body, and the mean scores for the latent DNA features corresponding to each image were used to calculate the mean of the three latent DNA features.

[0125] Table 15 Sample mean scores for each image style

[0126] In the design segmentation and decorative line feature scheme, the image with the highest average score from the sample is selected for both the front face and the body. For example... Figure 13 As shown, in the styling segmentation scheme, the front panel design is FG-1, and the body design is FG-4. In the decorative line scheme, the front panel design is ZS-3, and the body design is ZS-6. Since the ventilation mesh is only a pattern design, the design with the highest average value, SE-2, was chosen.

[0127] Two images each of the shape segmentation scheme and decorative line scheme, and one image of the heat dissipation mesh pattern were selected through a screening strategy. These images can be used as design references. By integrating the advantages and features of these schemes, we can explore the appearance design of the test machine and the final product family series design. Figure 14 As shown, the FG-1's geometric division is simple and clear, with clean lines; therefore, it was extracted as the basis for the heat dissipation mesh panel division in the testing machine method. Since the heat dissipation air intake and exhaust positions are located at the front and back of the testing machine respectively, the layout of the heat dissipation mesh panel is also positioned at the front and back of the testing machine. (Reference) Figure 15The decorative lines of the ZS-3 in Figures (a) and (b) cover the front, thus primarily providing a reference for the front design of the test unit, extracted through hand-drawn sketches of the front features. (Reference) Figure 16 In section (a), the front panel layout of the test machine was obtained by combining the solutions refined from FG-1 and ZS-3. For example... Figure 16 As shown in (b) and (c), the SE-2 heat dissipation mesh pattern was placed on the heat dissipation mesh panel, thus obtaining the front design scheme of the test machine. Figure 17 As shown in (a), (b), and (c), the design scheme of the test aircraft body was obtained by extracting the line design of the ZS-6 reference drawing and the outline of the FG-4. Figure 18 As shown, the overall appearance design sketch was obtained by combining the fuselage design with the front design. Handles and casters were added during modeling, resulting in the final appearance design of the test unit. The overall process of exploring the test unit's appearance is as follows: Figure 19 As shown.

[0128] The CMF (Color, Material, Texture) design of the test machine is a crucial part of the product development process. A well-designed CMF not only enhances the product's appearance and feel but also improves the user experience and brand recognition. Based on the colors provided by the company, black and white were chosen as the primary colors, with blue as the accent color, resulting in two primary colors and pure white. Using different primary colors, four color schemes for the test machine were derived: Scheme 1 uses white as the primary body color, blue as the accent color, and white as the logo color; Scheme 2 uses white as the primary body color, gray as the accent color, and white as the logo color; Scheme 3 uses gray as the primary body color, white as the accent color, and blue as the logo color; Scheme 4 uses gray as the primary body color, blue as the accent color, and white as the logo color. Considering the renderings of the four schemes, a primary color consistent with the brand image was chosen: white or light gray for a minimalist style, dark gray or black for a more sophisticated style, and metallic or cool tones for a technological style. Therefore, Scheme 3 was selected as the color scheme for the test machine.

[0129] Choose materials with good functionality, such as wear resistance, corrosion resistance, and water resistance. For example, the outer shell can be made of high-strength plastic or metal to ensure the product's durability. Select comfortable materials, such as soft rubber or silicone, for areas frequently touched by users to enhance the user's tactile experience and overall feel. Choose appropriate surface treatment processes, such as matte finishing, polishing, painting, or electroplating, to ensure the surface texture and visual appeal. Metal shells can be polished to enhance a technological feel; plastic shells can be matte to improve tactile feel. Through surface treatment technologies, improve the product's texture and feel, and enhance the user experience.

[0130] After completing the design of the test unit, its style was then transferred to two other products in a specific brand series: a chassis and a desktop test unit. This approach ultimately created a unified product family design. Style transfer involves not only unifying visual elements but also enhancing the overall design philosophy and brand recognition.

[0131] Identify the core elements of the test machine's exterior design, including color, shape segmentation, decorative lines, and heat dissipation mesh. These elements will serve as the basis for style transfer. The test machine's exterior design is as follows: Figure 20 Figure (a) shows the exterior design reference for the slot-type instrumentation chassis of the RF millimeter-wave switch matrix and RF power amplifier. Figure 20 Figure (b) shows the exterior design of the desktop test machine. Figure 20 Figure (c) shows a desktop testing machine used to verify the extensibility of the product family's appearance design. Six competing companies were selected, with three products chosen from each company. To avoid using logos for classification, an AI-powered logo removal tool was used to eliminate the brand logos from each product. Twelve participants, unfamiliar with the products, divided them into eight groups of three images each, based on their shape features and decorative lines. The success rate of correct grouping for each brand was calculated. The original products were correctly classified less than 40% of the time, while the new product family's appearance design was correctly classified 100% of the time. This demonstrates, to some extent, the extensibility of the product family's appearance design, meaning that companies can maintain the brand image and design style of the product family's appearance design when launching new products. It also illustrates the feasibility of using AI to assist in product family appearance design.

Claims

1. A data-driven product family design method based on artificial intelligence, characterized in that, Includes the following steps: The dominant and recessive DNA features of the product to be designed were extracted separately. Product family appearance design is carried out by combining dominant and recessive DNA features. Initial images are generated through text prompts, and the prompts are iterated until images that meet the design requirements are generated. The images that meet the design requirements are selected to obtain the top-scoring images. The design elements in the dominant DNA feature images are combined to determine the appearance design of the product to be designed. The heritable design elements of each dominant feature in the appearance design of the product to be designed are transferred to the product to be designed, forming the product family appearance design.

2. The data-driven product family design method based on artificial intelligence according to claim 1, characterized in that, Extracting the dominant DNA characteristics of the appearance of the product to be designed specifically includes the following steps: The first step is to conduct a competitive analysis of the test machine, analyzing it from the perspectives of function and appearance, and initially extracting the appearance features of the test machine; The second step is to collect images of the test machine and similar-looking instruments and preprocess them. Preprocessing includes adjusting image size, removing backgrounds, and removing photos with too low a resolution. The third step is to divide the test set, training set and validation set, adjust the top layer of the ResNet50 model, and calculate the accuracy of the classification of the adjusted ResNet50 model. The fourth step is to select test machine image samples for test machine feature recognition, obtain the corresponding heat map, count the frequency of appearance features in all test images, and select the three with the highest frequency as the dominant DNA features of the test machine.

3. The data-driven product family design method based on artificial intelligence according to claim 2, characterized in that, Obtain images of surrounding products for the product to be designed, and preprocess the obtained images to obtain an image dataset with uniform format, uniform pixel count, and uniform naming rules; Train and save the ResNet50 model: Replace the top layer of the ResNet50 model with a new layer, which includes a global average pooling layer and two fully connected layers. The last layer uses the softmax activation function and outputs the probabilities of the 6 classes. The original ResNet50 layer is set to be untrainable, and the features in the pre-trained weights are retained. The heat map generated by the trained ResNet50 model was used to filter the three appearance features with the highest proportion as the dominant DNA features of the test machine appearance.

4. The data-driven product family design method based on artificial intelligence according to claim 1, characterized in that, Extracting recessive DNA features from the product to be designed includes: By interviewing company employees, we obtained the company's strategic goals and long-term vision, product lines and market strategies, collected the characteristics and functional requirements of the company's major products, formed a preliminary understanding of the overall characteristics of the product to be designed, and compiled the interview records to obtain a word cloud. Based on the results of enterprise user interviews, we extracted implicit keywords that are highly relevant to brand information; Based on expert opinions and research, the intuitive and emotional vocabulary of the test machine was obtained. The implicit keywords of the enterprise were matched with intuitive and emotional vocabulary, and an antonym was matched for each vocabulary to obtain intuitive and emotional vocabulary pairs. Based on the scoring results of the emotional intention survey questionnaire, statistical analysis was conducted on the intention words, and the words with the highest relevance were selected to obtain the implicit DNA characteristics of the product to be designed.

5. The data-driven product family design method based on artificial intelligence according to claim 4, characterized in that, Develop an interview outline based on the professional field, core competitive advantages, key characteristics and technical specifications of the product, and brand identity content. Based on the interview results, keywords were extracted, the frequency of word usage was quantitatively analyzed, meaningless words were removed, and a word cloud was generated.

6. The data-driven product family design method based on artificial intelligence according to claim 4, characterized in that, Based on the scoring results of the emotional intention survey questionnaire, statistical analysis was conducted on the intentional vocabulary, and the words with the highest relevance were selected. The implicit DNA characteristics of the product to be designed include: Multiple representative product images were selected, with all details clearly visible. The questionnaire included the same evaluative questions using emotional vocabulary for each product image, allowing participants to evaluate each image using the same set of standards. The questionnaire used a seven-point Likert scale, with positive and negative numbers representing the positive or negative aspects of the emotional vocabulary, and the numerical value representing its degree. Test machine industry professionals, industrial design professionals, and ordinary users were invited to evaluate the questionnaire. Participants rated each picture based on the positive and negative words listed in the questionnaire. A stratified analysis method was used to calculate the mean and variance of the ratings for each image for each group. The mean was used to measure the overall evaluation trend, and the variance was used to assess the dispersion of the ratings. The average sample mean of the ratings for each image was then calculated. Finally, a weighted average was used to further calculate the words with the largest absolute value of the reinforced average sample mean of the emotional intention words.

7. The data-driven product family design method based on artificial intelligence according to claim 1, characterized in that, The dominant and recessive DNA characteristics of the product to be designed are used as keywords in the text-based image prompts. Combined with usage scenarios, rendering, and material keywords, images are initially generated based on the keywords. Based on the initially generated images, the prompt words are adjusted, and the images are generated iteratively using the prompt words. The images are generated multiple times until the number and quality of the generated images meet the requirements, and then the images are classified according to dominant DNA characteristics. The images corresponding to the prompts in each dominant DNA trait category are screened in the first round, and four of the most valuable images are retained for each prompt. Then, the design staff conducts a second round of screening, retaining the most valuable images for each prompt, based on the Likert five-point scale, that is, retaining images whose average user score is higher than the set value. A rating questionnaire was created for the images corresponding to each category. Relevant experts and users were invited to evaluate and select the images. The questionnaire contained clear images of each dominant DNA feature scheme. Respondents rated the schemes from the perspective of different recessive DNA features. Based on the questionnaire results, the images with the highest scores that reflected the recessive DNA feature elements were selected. Based on the final selected images, design elements are extracted from each dominant DNA feature image to determine the appearance design of the product to be designed.

8. A data-driven product family design system based on artificial intelligence, characterized in that, Includes a feature extraction module and a design module; The feature extraction module is used to extract the dominant and recessive DNA features of the appearance of the product to be designed, respectively. The design module is used to combine dominant and recessive DNA features to design the appearance of product families. It initially generates images based on text prompts, iterates through the prompts until images that meet the design requirements are generated, filters the images that meet the design requirements to obtain the top-scoring images, determines the appearance design of the product to be designed by combining the design elements in the dominant DNA feature images, and transfers the heritable design elements of each dominant feature in the appearance design of the product to be designed to the product to be designed, thus forming the appearance design of the product family.

9. The data-driven product family design system based on artificial intelligence according to claim 8, characterized in that, The feature extraction module includes a dominant DNA feature extraction unit and a recessive feature extraction unit; The dominant DNA feature extraction unit is used to analyze from both functional and appearance perspectives, and to initially extract the appearance features of the testing machine; Images of the test machine and similar instruments were collected and preprocessed, including image resizing, background removal, and removal of low-resolution photos. The ResNet50 model was divided into test, training, and validation sets, and its top layer was adjusted. The accuracy of the adjusted ResNet50 model in classification was statistically analyzed. Test machine image samples were selected for feature recognition, resulting in corresponding heatmaps. The frequency of appearance features in all test images was statistically analyzed, and the three most frequent features were selected as the dominant DNA features of the test machine. The latent DNA feature extraction unit is used to combine enterprise user interview results to extract latent keywords that are strongly related to brand information. Based on expert opinions and research, the unit obtains the emotional intention words of the test machine, matches the enterprise's latent keywords with intuitive emotional intention words, and pairs each intention word with an antonym to obtain emotional intention word pairs. Based on the scoring results of the emotional intention survey questionnaire, statistical analysis was conducted on the intention words, and the words with the highest relevance were selected to obtain the implicit DNA characteristics of the product to be designed.

10. The data-driven product family design system based on artificial intelligence according to claim 8, characterized in that, The design module includes an analysis unit that uses stratified analysis to calculate the mean and variance of each person's rating for each picture based on the questionnaire results. The mean is used to measure the overall evaluation trend, and the variance is used to assess the dispersion of the ratings. Then, the average sample mean of the rating for each picture is calculated. Finally, a weighted average is used to further calculate the words with the largest absolute value of the reinforced average sample mean of the emotional intention words.

Images