A Product Design Strategy Generation Method Based on Environmental Feature Extraction

Abstract

This invention discloses a product design strategy generation method based on environmental feature extraction, belonging to the field of intelligent product design technology. The steps are as follows: constructing a multimodal dataset containing ambient sound, scene images, and odor data; extracting root mean square energy, zero-crossing rate, spectral centroid, and spectral flux features from the ambient sound; extracting explicit visual features from the scene images and implicit style features through an improved Swin-Transformer model; extracting intensity, pleasantness, and familiarity features from the odor data; performing semantic mapping through Mamdani fuzzy inference to generate a design guidance parameter vector, including sensory stimulation, morphological simplicity, emotional affinity, technological feel, interaction rhythm, stability, sense of security, and environmental integration; and generating a product design strategy based on this vector. This invention automatically transforms environmental perception into a product design strategy, improving the objectivity and efficiency of design decisions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent product design technology, specifically relating to a product design strategy generation method based on environmental feature extraction. Background Technology

[0002] In product design, design elements such as product form, color, materials, and interaction methods typically need to be coordinated with their usage environment to create a unified and coherent overall experience. However, existing product design methodologies often rely on designers' subjective experience, aesthetic judgments, or user research results obtained through questionnaires, interviews, and batch review analysis. Environmental factors are often treated as background conditions for qualitative reference rather than as quantifiable and calculable input information for design decisions. For example, in actual design processes, designers often manually adjust the product's shape or color scheme based on subjective judgments of the environment, such as "quiet or noisy," "bright or dim," or "comfortable or stimulating." This lack of systematic collection and objective modeling of environmental information leads to significant differences in design results from different designers under the same environmental conditions, resulting in low design consistency and reusability.

[0003] While some existing multimodal environmental perception technologies incorporate multi-source information such as ambient sound and images for fusion analysis, their applications are primarily focused on areas like emotion recognition, scene classification, or user state analysis. They emphasize outputting environmental or state-based judgments rather than directly supporting decision-making in the product design process. Even when multimodal data fusion is involved, the related technologies often remain at the recognition or classification level, lacking a systematic approach to further transform environmental perception results into design elements applicable to product styling, color, materials, or interaction methods.

[0004] Meanwhile, in existing design assistance systems, the relationship between environmental features and product design elements largely relies on manual experience rules, lacking a unified feature extraction process and semantic mapping mechanism. This makes it difficult to automatically generate and reuse product design strategies under different environmental conditions, thus hindering the development of intelligent and large-scale application of product design.

[0005] Therefore, existing technologies lack an effective technical solution that can comprehensively utilize multimodal environmental information such as ambient sound, scene images, and odors to uniformly extract and semantically map environmental features, and further generate product design strategies that are coordinated with the overall environment. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention proposes a product design strategy generation method based on environmental feature extraction. The method is rationally designed, overcomes the shortcomings of existing technologies, and achieves good results.

[0007] A product design strategy generation method based on environmental feature extraction includes the following steps: S1: Construct a multimodal environmental perception dataset, which includes environmental sound data, scene image data, and odor data; S2: Extract features from the ambient sound data to obtain ambient sound features; S3: Perform feature extraction on the scene image data to obtain scene image features including explicit visual features and implicit style features extracted based on the improved Swin-Transformer model; S4: Extract features from the odor data to obtain odor features; S5: Perform multimodal environmental feature design semantic mapping on the environmental sound features, scene image features, and odor features to generate a design guidance parameter vector; S6: Generate a product design strategy based on the design guidance parameter vector.

[0008] Furthermore, the construction of the multimodal environment-aware dataset in step S1 includes: S1.1: Construct a subset of ambient sound data based on the ESC-50 dataset; S1.2: Construct a subset of scene image data based on the SUN397 dataset; S1.3: Construct a subset of odor data based on the Keller2016 odor perception dataset.

[0009] Furthermore, the environmental sound characteristics mentioned in step S2 include: root mean square energy. Zero crossing rate Spectral centroid and spectral flux .

[0010] Furthermore, the explicit visual features mentioned in step S3 include the main color tone, color scheme, brightness value, texture roughness, contrast, saturation, and illumination uniformity. The dominant color tone is extracted from the image pixels using the K-means clustering algorithm. First, the image is converted to the RGB color space. Then, the pixels are reconstructed, and the K-means clustering algorithm is used to divide the pixels into multiple categories. The color with the highest frequency is selected as the dominant color tone of the image. The expression is: ; in, The main color tone for the output. This is a color space conversion operation. The set of pixels for the entire image; The color scheme is determined by converting the image to the HLS color space and calculating the average hue, expressed as follows: ; in, This refers to the output color scheme type. It is a warm color scheme. It is a cool color scheme. Neutral color tone The average hue angle in the HLS color space, normalized to [0,1]; The brightness value is calculated by converting the image from the RGB color space to the HSV color space, and the expression is: ; in, This is the brightness value. For the first The luminance component value of each pixel in the HSV color space This represents the number of pixels in the image; based on the calculation results, the image's brightness values ​​are divided into five levels. The texture roughness is calculated using Shannon entropy, and the expression is: ; in, For texture roughness, This represents the probability distribution of pixels with a gray value of 𝑘 in the image; and is divided into three texture roughness levels: smooth, medium, and rough based on the calculation results. The contrast ratio is calculated as the ratio of the grayscale standard deviation to the mean, expressed as: ; in, For contrast, The standard deviation of the image grayscale. The mean value of the image grayscale. It is a very small value; the contrast value is divided into five levels to characterize the brightness and darkness of the image; The saturation is calculated by converting the image from the RGB color space to the HSV color space, and the expression is: ; in, For overall image saturation, For the first The saturation of each pixel is divided into three levels of color saturation. The uniformity of illumination is calculated using the standard deviation of the luminance component, expressed as: ; in, For uniform illumination, The luminance component in the HSV color space The standard deviation; divided into three levels of light uniformity.

[0011] Furthermore, the construction and training of the improved Swin-Transformer model described in step S3 includes the following steps: S3.1: Data annotation, assigning style labels to each sample in the subset of scene image data; S3.2: Randomly divide the labeled dataset into training and validation sets in an 8:2 ratio; S3.3: Model training, using the following improvement strategies: (1) Dynamic layered unfreezing and differentiated learning rate strategy: Freeze the parameters of the first 50% of the network layers of the Swin-Transformer, unfreeze the parameters of the last 50% of the network layers for training, and allocate a learning rate to the backbone network that is less than that to the classification head. (2) ECA channel attention module embedding strategy: ECA channel attention module is embedded after the feature extraction layer of Swin-Transformer to enhance style-related channel features; (3) Lightweight data augmentation strategy guided by style features: retain random horizontal flipping and color dithering, reduce the adjustment range of color brightness, contrast and saturation from 0.3 to 0.2, and reduce the hue coefficient from 0.05 to 0.03; (4) Adaptive learning rate scheduling strategy: A combination of 5 rounds of linear preheating and cosine annealing restart is adopted. In the first 5 rounds, the learning rate is linearly increased from 0.001 to the set value. After that, cosine annealing is restarted every 20 rounds, and the learning rate is gradually reduced to 1e-6. (5) TTA Enhanced Inference Strategy: During the inference phase, the test image is horizontally flipped for enhancement, and the inference results of the original image and the flipped image are averaged. S3.4: Implicit style features assign style labels to each image using the classification head of the improved Swin-Transformer model after training.

[0012] Furthermore, the extraction of odor features in step S4 includes the following steps: S4.1: Based on the Keller 2016 odor perception dataset, the subjects' ratings for each odor sample on three dimensions: intensity, pleasantness, and familiarity are calculated. Intensity reflects the strength of the odor, pleasantness reflects the degree of pleasantness or unpleasantness of the odor, and familiarity reflects whether the odor is familiar to the subject. The mean scores of each dimension are calculated as the mean intensity, mean pleasantness, and mean familiarity of the odor. The formula for calculating the mean of the intensity dimension is as follows: ; in, For the number of valid data points, For the first Intensity ratings for each subject; S4.2: Count the number of valid subjects for each odor sample under each sub-odor category label, select the sub-label corresponding to the maximum number of valid subjects as the primary label, and record the maximum value and its corresponding label name; S4.3: Output the odor feature data for each odor sample, including: unique odor identifier, number of valid subjects, mean intensity, mean pleasantness, mean familiarity, maximum value of sub-odor category labels and corresponding label names.

[0013] Furthermore, the design semantic mapping of the multimodal environment features in step S5 includes the following steps: S5.1: Normalize the environmental sound characteristics to obtain normalized characteristics. The minimum-maximum normalization method with quantile truncation is used: ; ; in, These are the original eigenvalues ​​to be normalized. The median value after quantile truncation. For quantiles, It is a very small positive number; The environmental sound semantic variables are constructed as follows: Ambient noise level variables , used to characterize the intensity of environmental sound energy; Environmental noise active variables , The weighting coefficients for the zero-crossing rate feature. It is a sigmoid nonlinear mapping function used to characterize the rhythm and variation characteristics of environmental sounds; Sharp variables of ambient sound , used to characterize the proportion of high-frequency components; Stable variables of ambient sound ; right The language values ​​of low, medium, and high were used for blurring respectively; S5.2: Construct the basic visual variables as follows: The explicit visual features in the scene image are subjected to ordered numerical encoding and normalization, and the discrete values ​​at each level are converted into discrete values. Map to the [0,1] interval using the following formula: ; in, This refers to the maximum number of levels for this feature, specifically the maximum number of levels for brightness and contrast. The maximum number of grades for texture roughness, saturation level, and illumination uniformity. Let the mapped brightness level variable be... Contrast level variable is Saturation level variable is The light uniformity level variable is Texture roughness level variable is ; Hue temperature variable Based on the color scheme: Warm is set to 1, Neutral to 0.5, and Cool to 0; Based on explicit visual attributes and style conditions, construct visual semantic variables: Visual stimulus variables Where, 𝐶 is the normalized contrast value and S is the normalized saturation value; Visual morphological variables ,in These represent the complexity scores of texture roughness, contrast, and style tags in implicit style features, respectively. Visual technology variables ,in These represent the technical scores for style tags in the tonal temperature variable, contrast, and implicit style features, respectively. Visual fusion variables ,in These represent the blending scores of style tags in hue temperature, brightness, lighting uniformity, and implicit style features, respectively. The scoring of style tags in implicit style features is determined by expert experience; Establish membership functions for Low, Mid, and High linguistic values ​​for the basic visual variables and visual semantic variables, respectively; S5.3: Construct olfactory semantic variables as follows: olfactory stimulus variables Olfactory pleasure variables olfactory affinity variables By respectively through the mean intensity Mean of pleasure Average familiarity Obtained by quantile truncation and normalization; Odor category variable The maximum value of each sub-label and its corresponding label name Characterization; right The three language values ​​of Low, Mid, and High were used for blurring respectively. Dedicated rules are used for triggering; S5.4: Employ Mamdani-type fuzzy inference, using the fuzzified results of environmental sound semantic variables, basic visual variables, visual semantic variables, and olfactory semantic variables as antecedents, to design guiding parameter vectors. As the consequent, reasoning is performed according to a preset rule base; in, For stimulation level, For simplicity, For emotional affinity, For a sense of technology, For the rhythm of interaction, For stability, For a sense of security, For degree of integration; The antecedent of a rule is determined by the rule definition using either an AND or OR operation, i.e., min or max. Rule implication uses min, and rule aggregation uses max. For output For each dimension, establish three language values: Low, Mid, and High, and define corresponding membership functions; S5.5: For each rule Activation strength is calculated based on the antecedent. : ; in, , , These are the membership functions for auditory, visual, and olfactory semantic variables, respectively. Will The function is applied to the consequent output membership function, and multiple rule results on the same output dimension are aggregated using max aggregation to obtain the output fuzzy set. ; For each output dimension, the centroid method is used to defuzzify and obtain a definite value: ; in, The determined value after deblurring. To output continuously valued variables in the universe of discourse, The membership function value of the aggregated output fuzzy set at the value u. To avoid extremely small positive numbers with a denominator of 0; The defined values ​​of all output dimensions together constitute the design guiding parameter vector. .

[0014] Furthermore, the generation of the product design strategy in step S6 includes: Based on stimulation level The values ​​of these parameters are used to output adjustment commands for the contrast coefficient and saturation coefficient. Based on simplicity The value of is used to output the adjustment instructions for the number of product structure levels and interface information density parameters; Based on emotional affinity The value of is used to output the adjustment commands for surface roughness and surface curvature radius parameters; Based on the sense of technology The value of the parameter determines the output material type parameter adjustment command; According to the rhythm of interaction The value of is used to output instructions for adjusting the interactive feedback frequency and response delay parameters; According to stability The values ​​of are used to output adjustment commands for the bottom support width parameter and the center of gravity height parameter; Based on sense of security The value of the parameter determines the output of the accidental touch protection level parameter and the confirmation step strength parameter adjustment command. According to the degree of integration The value of is used to output the color difference parameter adjustment command between the main color of the product and the main color of the environment; The instructions are used as input to the computer-aided design system to automatically generate or adjust the digital model of the product.

[0015] The beneficial technical effects of this invention are as follows: This invention introduces multimodal environmental information such as ambient sound, scene images, and odors to systematically collect and process environmental features. It transforms the environmental perception process, which originally relied on the subjective experience of designers, into a quantifiable and calculable feature input, thereby improving the objectivity and consistency of product design decisions.

[0016] This invention, by constructing a unified environmental feature extraction process and design semantic mapping mechanism, realizes the conversion of multimodal environmental perception results into design guidance parameters, avoiding the limitations of existing technologies where environmental information only stays at the identification or classification level. This enables environmental features to directly participate in the product design strategy generation process, improving the matching degree between design results and the usage environment.

[0017] By comprehensively analyzing environmental sound characteristics, explicit visual attributes, implicit visual style characteristics, and odor perception characteristics, this invention can fully characterize the product's usage environment from multiple perception dimensions, which is beneficial to improving the overall coordination and experience quality of the product in terms of shape, color, material, and interaction methods.

[0018] This invention employs fuzzy reasoning to semantically map multimodal environmental features, which can effectively handle the uncertainty and fuzziness in environmental perception information, making the generation process of design guidance parameters more stable and continuous, avoiding drastic fluctuations in design strategies due to minor changes in environmental features, thereby improving the reliability of design strategy generation.

[0019] The design guidance parameter vector generated by this invention can be used as a unified input to drive the generation of multi-dimensional product design strategies, enabling design strategies under different environmental conditions to be automatically generated and reused, reducing the workload of manual analysis and repeated adjustments, improving product design efficiency, and reducing design costs.

[0020] The overall process structure of this invention is clear, and the feature extraction method and reasoning rules have good versatility and scalability, making them applicable to design scenarios of different types of products. They are easy to promote and apply in intelligent product design systems and have high practical value. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of a product design strategy generation method based on environmental feature extraction in this invention.

[0022] Figure 2 This is a diagram showing the environmental sound feature extraction results in an embodiment of the present invention.

[0023] Figure 3 This is a diagram showing the scene image feature extraction results in an embodiment of the present invention.

[0024] Figure 4 This is a diagram showing the odor feature extraction results in an embodiment of the present invention.

[0025] Figure 5 This is a diagram of the improved Swing-Transformer network structure in this invention.

[0026] Figure 6 To improve the confusion matrix of the previous Swin-Transformer network training results.

[0027] Figure 7 This is the confusion matrix of the improved Swin-Transformer network training results.

[0028] Figure 8 This is a structural diagram of multimodal semantic mapping and fuzzy reasoning in this invention.

[0029] Figure 9 This is an image of the product's usage environment in an embodiment of the present invention.

[0030] Figure 10 This is a diagram illustrating the application effect of product design in an embodiment of the present invention. Detailed Implementation

[0031] The specific embodiments of the present invention will be further described below with reference to specific examples: A product design strategy generation method based on environmental feature extraction, such as Figure 1 As shown, it includes the following steps: S1: Construct a multimodal environmental perception dataset, which includes environmental sound data, scene image data, and odor data; The construction of a multimodal environment-aware dataset includes: S1.1: An ambient sound data subset was constructed based on the ESC-50 dataset. This subset includes all 50 categories of ambient sound samples from the ESC-50 dataset, used to comprehensively characterize the auditory characteristics of different usage environments. The dataset can be found at: https: / / github.com / karolpiczak / ESC-50; S1.2: Construct a subset of scene image data based on the SUN397 dataset; select scene categories corresponding to typical product usage environments from the SUN397 dataset, including beach, botanical garden, campsite, park, pasture, and picnic area, to represent the spatial structure, color features, and overall style characteristics of the environment at the visual level. The dataset is available at: https: / / opendatalab.org.cn / OpenDataLab / SUN397; S1.3: An odor data subset was constructed based on the Keller2016 odor perception dataset. This subset contains odor perception data corresponding to all odor stimuli in the Keller2016 dataset, used to characterize olfactory perception characteristics under different environmental conditions. The dataset can be found at: https: / / github.com / pyrfume / pyrfume-data / tree / main / keller_2016.

[0032] S2: Extract features from the ambient sound data to obtain ambient sound features; Ambient sound characteristics include: root mean square energy Zero crossing rate Spectral centroid and spectral flux In this embodiment, the extracted environmental sound feature results are as follows: Figure 2 As shown.

[0033] Regarding root mean square energy For ambient sound signals The audio signal is processed to calculate the root mean square (RMS) value for each segment, which characterizes the energy intensity of the ambient sound. The RMS calculation formula is as follows: ; in, This represents the total number of sampling points for the audio signal. For audio signals at time The calculated values ​​are normalized to obtain numerical characteristics reflecting the ambient sound intensity.

[0034] Regarding the zero-crossing rate The zero-crossing rate is calculated in an audio signal, representing the number of times the signal changes from positive to negative or vice versa within a certain time window. The formula is as follows: ; in, Indicates signal At any moment The sign function, if but ,like but .

[0035] For the Spectral Centroid The ambient sound signal is subjected to a Fast Fourier Transform (FFT) to calculate the centroid value of the spectrum, thus reflecting the center of the frequency distribution of the audio signal. The calculation formula is as follows: ; in, The total number of frequencies in the spectrum. For the spectrum at frequency The amplitude value on; For Spectral Flux The spectral flux value is obtained by calculating the spectral changes of consecutive frames, thus measuring the rate at which the spectral content of a signal changes over time. The calculation formula is as follows: ; in, Indicates at time and frequency The spectral amplitude at that location.

[0036] S3: Perform feature extraction on the scene image data to obtain scene image features including explicit visual features and implicit style features extracted based on the improved Swin-Transformer model; in this embodiment, the extracted explicit visual features and implicit style features are as follows: Figure 3 As shown.

[0037] Explicit visual features include interpretable visual characteristics such as dominant color tone, color scheme, brightness value, texture roughness, contrast, saturation, and illumination uniformity, which are used to describe the explicit visual representation of the environment. Dominant color tone: Color clustering analysis is performed on the scene image. The dominant color tone is extracted from the image pixels using the K-means clustering algorithm. First, the image is converted to the RGB color space. Then, the pixels are reconstructed, and the K-means clustering algorithm is used to divide the pixels into multiple categories. The color with the highest frequency is selected as the dominant color tone of the image. The expression is: ; in, The main color tone for the output. This is a color space conversion operation. The set of pixels for the entire image; Color scheme: based on the average of the hue angles It is determined by converting the image to the HLS color space and calculating the average hue, expressed as: ; in, This refers to the output color scheme type. It is a warm color scheme. It is a cool color scheme. Neutral color tone The average hue angle in the HLS color space, normalized to [0,1]; Brightness value: Calculated by converting the image from RGB color space to HSV color space, the expression is: ; in, This is the brightness value. For the first The luminance component value of each pixel in the HSV color space The number of pixels in the image; based on the calculation results, the image brightness value is divided into five levels: "extremely low brightness", "low brightness", "medium brightness", "high brightness", and "extremely high brightness". Texture roughness: By converting the image to grayscale, the Shannon entropy of the image is calculated based on the grayscale distribution. It is used to characterize the overall texture complexity of the image, and its expression is: ; in, For texture roughness, This represents the probability distribution of pixels with a gray value of 𝑘 in the image; and is divided into three texture roughness levels: smooth, medium, and rough based on the calculation results. Contrast ratio: After converting an image to grayscale, the ratio of the standard deviation of grayscale values ​​to the mean is calculated. The expression is: ; in, For contrast, The standard deviation of the image grayscale. The mean value of the image grayscale. It is a very small value used to avoid division by zero errors; the contrast value is divided into five levels, from "very low contrast" to "very high contrast", to characterize the brightness and darkness of the image; Saturation: Convert the image from RGB color space to HSV color space and extract the saturation component. And calculate the average saturation of all pixels in the image as the overall saturation index of the image, the expression is: ; in, For overall image saturation, For the first The saturation of each pixel is divided into three levels of color saturation. Illumination uniformity: Based on the luminance component in the HSV color space The standard deviation of the pixel values ​​is calculated to reflect the dispersion of the illumination distribution, and the illumination uniformity index is obtained by inverse normalization, expressed as: ; in, For uniform illumination, The luminance component in the HSV color space The standard deviation; divided into three levels of light uniformity.

[0038] An improved Swin-Transformer model is proposed to extract implicit style features from scene images. This implicit style feature extraction method automatically learns and extracts style information from the image, particularly implicit features such as texture, color scheme, and composition style, through deep learning processing. The construction and training of the improved Swin-Transformer model includes the following steps: S3.1: Data labeling. Style labels are assigned to each sample in the subset of scene image data. In this embodiment, style labels include rugged, retro, minimalist, tranquil, fresh, and fantastical. The label assignment method is as follows: Rough: Characterized by strong contrast, complex textures, and irregular tones.

[0039] Vintage: The color tone is warm, giving it a vintage effect, and the texture is relatively soft.

[0040] Minimalist: Clear and simple color scheme, few and regular elements, highlighting a sense of simplicity.

[0041] Serene: The tones are soft, the image is calm, and there is almost no noise or extreme variation.

[0042] Fresh: Bright and refreshing, often associated with natural scenes.

[0043] Fantasy: characterized by its dreamlike or fantastical quality, vibrant colors, or surreal features.

[0044] S3.2: Randomly divide the labeled dataset into training and validation sets in an 8:2 ratio; S3.3: Model training, such as Figure 5 As shown, the traditional Swin-Transformer model includes an input encoding layer, a backbone network, normalization and pooling layers, and a classification head. The backbone network consists of four sequentially connected Swin Block feature extraction layers; the number of channels C in the four feature extraction layers are 96, 192, 384, and 768, respectively. This model adopts the following improvement strategy: (1) Dynamic layered unfreezing and differentiated learning rate strategy: Freeze the parameters of the first 50% of the network layers of the Swin-Transformer, unfreeze the parameters of the last 50% of the network layers for training, and allocate a learning rate to the backbone network that is less than that to the classification head. (2) ECA (Efficient Channel Attention) module embedding strategy: ECA channel attention modules are embedded after each feature extraction layer of Swin-Transformer. The input and output dimensions dim of the four ECA channel attention modules are 96, 192, 384 and 768, respectively. The adaptive channel-level attention mechanism is used to enhance the style-related channel features. (3) Lightweight data augmentation strategy guided by style features: In order to avoid the interference of traditional data augmentation methods (such as rotation and blur) on style features, this invention designs a lightweight data augmentation method guided by style features, which retains random horizontal flipping and slight color jitter, and reduces the amplitude of color jitter, thereby avoiding unnecessary style interference, reducing the adjustment range of color brightness, contrast and saturation from 0.3 to 0.2, and the hue coefficient from 0.05 to 0.03; (4) Adaptive learning rate scheduling strategy (learning rate warm-up and cosine annealing restart): A combination of 5 rounds of linear warm-up and cosine annealing restart is used to gradually increase the learning rate in the early stage of training and gradually decrease it in the later stage. The learning rate is linearly increased from 0.001 to the set value in the first 5 rounds, and then cosine annealing is restarted every 20 rounds, and the learning rate is gradually reduced to 1e-6; this strategy ensures stable convergence during training and improves the training efficiency of the model; (5) TTA (Test Time Augmentation) Enhanced Inference Strategy: During the inference phase, the test image is horizontally flipped for enhancement, and the inference results of the original image and the flipped image are averaged to improve the stability and accuracy of the model inference. like Figure 6 and Figure 7 As shown, the improved model has increased the classification accuracy of each category, and improved the stability and reliability of style recognition.

[0045] S3.4: Implicit style features assign style labels to each image using the classification head of the improved Swin-Transformer model after training.

[0046] S4: Perform feature extraction on the odor data to obtain odor features; in this embodiment, the extracted odor feature results are as follows: Figure 4 As shown.

[0047] The extraction of odor characteristics includes the following steps: S4.1: Based on the Keller 2016 odor perception dataset, the subjects' ratings for each odor sample were calculated on three dimensions: intensity, pleasantness, and familiarity. Intensity reflects the strength of the odor, pleasantness reflects the degree of pleasantness or unpleasantness of the odor, and familiarity reflects whether the odor is familiar to the subject. The mean scores for each dimension were calculated as the mean intensity, mean pleasantness, and mean familiarity of the odor. The formula for calculating the mean of the intensity dimension is as follows: ; in, For the number of valid data points, For the first Intensity ratings for each subject; The calculation methods for the Valence and Familiarity dimensions are the same as above.

[0048] S4.2: Sub-scent category labels include: edible, bread-like, sweet, and fresh, denoted as EDIBLE, BAKERY, SWEET, and FRESH, respectively. Count the number of valid participants for each scent sample under each sub-scent category label. Select the sub-label corresponding to the maximum number of valid participants as the primary label, and record this maximum value and its corresponding label name, for example: ; in, This represents the valid count of each subcategorized odor category label in the odor sample.

[0049] S4.3: Output the odor characteristic data for each odor sample, including: Odor_ID: A unique identifier for the odor (from the Stimulus column of the original dataset).

[0050] Valid_Sample_Count: Number of valid participants (excluding invalid data).

[0051] Intensity_Mean: The average intensity of the odor.

[0052] Valence_Mean: Mean of pleasantness of the odor.

[0053] Familiarity_Mean: The average level of familiarity with the scent.

[0054] Sub_Label_Max_Value: Maximum value of the sub-label (number of valid values).

[0055] Sub_Label_Max_Name: The label name corresponding to the maximum value.

[0056] S5: As Figure 8 As shown, the environmental sound features, scene image features, and odor features are subjected to multimodal environmental feature design semantic mapping to generate a design guidance parameter vector, and intermediate decision variables for product design strategy generation are output through fuzzy inference. Let the multimodal input feature vector be: ; in, As environmental sound characteristics, As a visual feature, It is an olfactory characteristic.

[0057] The output is a vector of design guiding parameters: ; These parameters are used to characterize the comprehensive guiding role of environmental conditions on product design in terms of sensory stimulation level, form simplicity and complexity, emotional affinity, technical attribute presentation, interaction rhythm and method, usage stability, safety control requirements, and degree of harmony with the environment. The value range of each design semantic parameter is normalized to [0,1].

[0058] The design of semantic mapping for multimodal environment features includes the following steps: S5.1: Normalize the environmental sound characteristics to obtain normalized characteristics. The minimum-maximum normalization method with quantile truncation is used: ; ; in, These are the original eigenvalues ​​to be normalized. The median value after quantile truncation. For quantiles, It is a very small positive number; Construct environmental sound semantic variables, converting acoustic features into auditory semantic variables, as follows: Ambient noise level variables , used to characterize the intensity of environmental sound energy; Environmental noise active variables , The weighting coefficients for the zero-crossing rate feature. It is a sigmoid nonlinear mapping function used to characterize the rhythm and variation characteristics of environmental sounds; Sharp variables of ambient sound , used to characterize the proportion of high-frequency components; Stable variables of ambient sound ; right The language values ​​are fuzzed using three levels: low, medium, and high. Triangular or trapezoidal membership functions are used, with the 20%, 50%, and 80% quantiles preferred as nodes.

[0059] S5.2: Construct the basic visual variables as follows: Explicit visual attributes and style tags are extracted from raw visual data, including color scheme, brightness, texture, contrast, saturation, illumination uniformity, and image style tags. The explicit visual features in the scene image are then subjected to ordered numerical encoding and normalization, converting the discrete values ​​at each level... Map to the [0,1] interval using the following formula: ; in, This represents the maximum number of levels for this feature. Maximum number of brightness and contrast levels The maximum number of grades for texture roughness, saturation level, and illumination uniformity. ;Specifically: Brightness level variable , ; Contrast level variable , ; Saturation level variable , ; Illumination uniformity level variable , ; Texture roughness level variable , ; Hue temperature variable Based on the color scheme: Warm is set to 1, Neutral to 0.5, and Cool to 0; Based on explicit visual attributes and style conditions, construct visual semantic variables: Visual stimulus variables Where, 𝐶 is the normalized contrast value and S is the normalized saturation value; Visual morphological variables ,in These represent the complexity scores of texture roughness, contrast, and style tags in implicit style features, respectively. Visual technology variables ,in These represent the technical scores for style tags in the tonal temperature variable, contrast, and implicit style features, respectively. Visual fusion variables ,in These represent the blending scores of style tags in hue temperature, brightness, lighting uniformity, and implicit style features, respectively. The scoring of style tags in implicit style features is determined by expert experience; Membership functions for three levels of linguistic values—Low, Mid, and High (or Warm / Neutral / Cool)—are established for the basic visual variables and visual semantic variables, respectively, for subsequent inference. S5.3: Construct olfactory semantic variables as follows: olfactory stimulus variables Olfactory pleasure variables olfactory affinity variables By respectively through the mean intensity Mean of pleasure Average familiarity Obtained by quantile truncation and normalization; Odor category variable The maximum value of each sub-label and its corresponding label name Characterization (e.g., EDIBLE, SWEET, BAKERY, FRESH, etc.); right The three language values ​​of Low, Mid, and High were used for blurring respectively. Discrete rules are used for triggering (different categories trigger different affinity / safety / fusion inference enhancements); S5.4: Employs Mamdani-type fuzzy inference, using the fuzzified results of environmental sound semantic variables, basic visual variables, visual semantic variables, and olfactory semantic variables, along with style tags, as antecedents, to design a guiding parameter vector. As the consequent, reasoning is performed according to a preset rule base; in, For stimulation level, For simplicity, For emotional affinity, For a sense of technology, For the rhythm of interaction, For stability, For a sense of security, For degree of integration; The antecedent of a rule is determined by the rule definition using either an AND or OR operation, i.e., min or max. Rule implication uses min, and rule aggregation uses max. For output In each dimension, establish three language values: Low, Mid, and High, and define corresponding membership functions (equidistant method). Typical rule set: ① Stimulation ; R1: If For High or If it is High, then For High.

[0060] R2: If If it is High, then For High.

[0061] R3: If For High and If it is Low, then For High.

[0062] ② Simplicity ; R4: If For a high-end style that is minimalist / tranquil, then... For High.

[0063] R5: If For High and If it is High, then It is Low.

[0064] ③ Emotional affinity ; R6: If For High and If it is High, then For High.

[0065] R7: If If it is Warm and the style is retro / tranquil, then Increase (Low→Mid) / (Mid→High).

[0066] R8: If For EDIBLE / BAKERY / SWEET and If it is Mid / High, then For High.

[0067] ④ Technological feel ; R9: If For High or ( For Cool and) If it is High, then For High.

[0068] R10: If the style = fantasy and If it is High, then promote.

[0069] ⑤ Interactive rhythm ; R11: If For High or High volatility means For High.

[0070] R12: If the style = fantasy / rugged and... For High For High.

[0071] ⑥ Stability ; R13: If For a high-end style = tranquil / minimalist If it is Mid / High, then For High.

[0072] R14: If For Low and If it is High, then It is Low.

[0073] ⑦ Sense of security ; R15: If For High and If it is Low, then For High (requires stronger security and verification mechanisms).

[0074] R16: If For High and If it is Low, then For High (unpleasant and stimulating environments, controllability and accidental touch prevention strategies need to be improved).

[0075] ⑧ Integration ; R17: If the style = fresh / tranquil and... Neutral / Cool and If it is High, then For High.

[0076] R18: If For FRESH and For High, and If it is Neutral / Cool, then For High.

[0077] R19: If If it's "High" and the style is fantasy / rugged, then... It is Low.

[0078] S5.5: For each rule Activation strength is calculated based on the antecedent. : ; in, , , These are the membership functions for auditory, visual, and olfactory semantic variables, respectively. Will The result is applied to the consequent output membership function (with min truncation) and then aggregated by max for multiple rule results of the same output dimension to obtain the output fuzzy set. ; For each output dimension, the centroid method is used to defuzzify and obtain a definite value: ; in, The determined value after deblurring. To output continuously valued variables in the universe of discourse, The membership function value of the aggregated output fuzzy set at the value u. To avoid extremely small positive numbers with a denominator of 0; The defined values ​​of all output dimensions together constitute the design guiding parameter vector. As the "environment-design guidance parameter mapping result", it is passed to S6 for product design strategy generation.

[0079] S6: Generate a product design strategy based on the design guidance parameter vector. The product design strategy is used to input the computer-aided design system to automatically generate a digital model of the product.

[0080] The generation of product design strategy vectors includes: Based on stimulation level The output commands adjust the contrast and saturation coefficients based on their values; where the contrast coefficient... Saturation coefficient ;when When, it is defined as low contrast; when When defined as medium contrast; when High contrast is defined as a value with a saturation coefficient. The corresponding relationship of the value range is as follows: when When, it is defined as low saturation; when When, it is defined as medium saturation; when When the value is high, it is defined as high saturation.

[0081] Based on simplicity The value of is used to output adjustment instructions for the number of product structure layers and interface information density parameters; the number of product structure layers Interface information density ;when When defined as a low-level structure; when When, it is defined as a medium-level structural hierarchy; when At this point, it is defined as a high-level structural hierarchy. Interface information density. The corresponding relationship of the value range is as follows: when When, it is defined as low information density; when When, it is defined as medium information density; when High information density is defined as such.

[0082] Based on emotional affinity The values ​​are used to output adjustment commands for surface roughness and surface curvature radius parameters; product surface roughness parameters Surface curvature radius parameter Product surface roughness parameters Surface curvature radius parameter Product surface roughness parameters The corresponding relationship of the value range is as follows: when When, it is defined as low roughness; when When, it is defined as medium roughness; when When this is defined as high roughness, the surface curvature radius parameter is used. The corresponding relationship of the value range is as follows: when When, it is defined as the small radius of curvature; when When, it is defined as the radius of curvature of medium; when When , it is defined as the radius of large curvature.

[0083] Based on the sense of technology The values ​​are used to output adjustment commands for material type parameters, edge sharpness parameters, and cool tone ratio parameters; the proportion parameter of technological materials is also included. Edge sharpness parameter Cool color proportion parameters ; Proportion of technological materials The corresponding relationship of the value range is as follows: when When, it is defined as the proportion of low-tech materials; when When, it is defined as the proportion of medium-tech materials; when At that time, it was defined as the proportion of high-tech materials. Edge sharpness parameter The corresponding relationship of the value range is as follows: when When defined as low edge sharpness; when When defined as medium edge sharpness; when When defined as high edge sharpness. Cool tone proportion parameter. The corresponding relationship of the value range is as follows: when When, it is defined as a low proportion of cool tones; when When, it is defined as a medium cool color proportion; when When defined as the proportion of cool and sophisticated colors.

[0084] According to the rhythm of interaction The value of is used to output the adjustment instructions for the interactive feedback frequency and response delay parameters; the interactive feedback frequency parameter Response delay parameter Interactive feedback frequency parameter The corresponding relationship of the value range is as follows: when When, it is defined as a low feedback frequency; when When, it is defined as a medium feedback frequency; when At this time, it is defined as a high feedback frequency. Response delay parameter. The corresponding relationship of the value range is as follows: when When, it is defined as low response latency; when When, it is defined as a medium response latency; when When this occurs, it is defined as high response latency.

[0085] According to stability The value of is used to output adjustment commands for the bottom support width parameter and the center of gravity height parameter; bottom support width parameter Center of gravity height parameter Bottom support width parameter The corresponding relationship of the value range is as follows: when When, it is defined as a narrow support structure; when When, it is defined as a medium-support structure; when When defined as a wide-support structure, the center of gravity height parameter is used. The corresponding relationship of the value range is as follows: when When, it is defined as a low center of gravity; when When, it is defined as the middle center of gravity; when At this time, it is defined as the high center of gravity.

[0086] Based on sense of security The value of the parameter determines the output of the accidental touch protection level parameter and the confirmation step strength parameter adjustment command; accidental touch protection level parameter Confirm the strength parameters of the steps Accidental contact protection level parameters The corresponding relationship of the value range is as follows: when When defined as low accidental touch protection level; when When, it is defined as a medium level of accidental touch protection; when At this point, it is defined as a high level of accidental touch protection. Confirm the strength parameters of the procedure. The corresponding relationship of the value range is as follows: when When, it is defined as a weak confirmation mechanism; when When, it is defined as a medium confirmation mechanism; when When this is defined as a strong confirmation mechanism.

[0087] According to the degree of integration The value of is used to output the color difference parameter adjustment command between the main color of the product and the main color of the environment; color difference parameter Comprehensive color difference parameters The corresponding relationship of the value range is as follows: when When, it is defined as high environmental integration; when When, it is defined as medium environmental integration; when When this occurs, it is defined as low environmental integration.

[0088] The instructions are used as input to the AI ​​design model to automatically generate or adjust the digital model of the product.

[0089] like Figure 9 As shown, the image is an environmental image of a park. The sound data example is 1-101296-A-19.wav from the ESC-50 dataset (Natural Recreation and Relaxing Environmental Sounds 1-101296-A-19). The olfactory data example is the 10th data point from the Keller 2016 Odor Perception Dataset. The environmental feature extraction results are shown in Table 1. Table 1. Environmental Feature Extraction Results ; 1) Construction of environmental sound semantic variables; The above results are then normalized using quantile truncation and Min-Max normalization: ; ; ; ; Therefore, loudness variable Mid, activity variable For Low-Mid, sharpness variable For Low, stability variable For High.

[0090] 2) Construction of visual semantic variables; Medium brightness → 0.5; Low contrast → 0.2; Medium saturation → 0.5; Texture coarsness → 0.7; Uniform illumination → 0.7; Hue: Neutral → 0.5; Visual stimulus variables Mid-Low, visual morphological variables Mid, visual technology variable For Low, visual fusion variable For High.

[0091] 3) Construction of odor semantic variables; Strength 26.36 → Mid; Pleasure 43.07 → High; Familiarize yourself with 30.25 → Mid; olfactory stimulation For Mid, olfactory pleasure For High, olfactory affinity It is rated High, and its scent category is Sweet.

[0092] 4) According to fuzzy rules, the following reasoning is obtained: Stimulation level: Mid; Simplicity: Mid; Emotional intimacy: High; Technical sophistication: Low; Interaction rhythm: Low; Stability: High; Sense of security: Mid; Blending degree: High; Based on the level of stimulation: ; ; This indicates that the product uses medium contrast and moderate color saturation.

[0093] Based on simplicity: ; ; This indicates that the product structure has fewer layers and the information is more concise.

[0094] Based on emotional affinity: ; ; This indicates that the product surface is relatively smooth and the overall design features a large curvature and rounded shape.

[0095] Based on technical sense: ; ; ; This indicates that the product reduces the proportion of technological materials and increases the proportion of natural materials, with a softer design for the edges.

[0096] Based on the interaction rhythm: ; ; This indicates that the product's interaction feedback frequency is low and the response rhythm is relatively stable.

[0097] Based on stability, we obtain: ; ; This indicates that the product adopts a wider bottom support structure and a lower center of gravity design.

[0098] Based on a sense of security: ; ; This indicates that the product has a moderately high level of protection against accidental touches and a strong confirmation mechanism.

[0099] Based on the degree of fusion: ; This indicates that the main color of the product maintains a small color difference with the main color of the environment, and the overall product has a high degree of environmental integration.

[0100] like Figure 10 The image shown is a rendering of the final outdoor speaker product produced according to the above design strategy.

[0101] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A product design strategy generation method based on environmental feature extraction, characterized in that, Includes the following steps: S1: Construct a multimodal environmental perception dataset, which includes environmental sound data, scene image data, and odor data; S2: Extract features from the ambient sound data to obtain ambient sound features; S3: Perform feature extraction on the scene image data to obtain scene image features including explicit visual features and implicit style features extracted based on the improved Swin-Transformer model; S4: Extract features from the odor data to obtain odor features; S5: Perform multimodal environmental feature design semantic mapping on the environmental sound features, scene image features, and odor features to generate a design guidance parameter vector; S6: Generate a product design strategy based on the design guidance parameter vector.

2. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, Step S1 involves constructing a multimodal environment-aware dataset, including: S1.1: Construct a subset of ambient sound data based on the ESC-50 dataset; S1.2: Construct a subset of scene image data based on the SUN397 dataset; S1.3: Construct a subset of odor data based on the Keller2016 odor perception dataset.

3. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, The environmental sound characteristics mentioned in step S2 include: root mean square energy. Zero crossing rate Spectral centroid and spectral flux .

4. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, The explicit visual features mentioned in step S3 include the main color tone, color scheme, brightness value, texture roughness, contrast, saturation, and illumination uniformity. The dominant color tone is extracted from the image pixels using the K-means clustering algorithm. First, the image is converted to the RGB color space. Then, the pixels are reconstructed, and the K-means clustering algorithm is used to divide the pixels into multiple categories. The color with the highest frequency is selected as the dominant color tone of the image. The expression is: ； in, The main color tone for the output. This is a color space conversion operation. The set of pixels for the entire image; The color scheme is determined by converting the image to the HLS color space and calculating the average hue, expressed as follows: ； in, This refers to the output color scheme type. It is a warm color scheme. It is a cool color scheme. Neutral color tone The average hue angle in the HLS color space, normalized to [0,1]; The brightness value is calculated by converting the image from the RGB color space to the HSV color space, and the expression is: ； in, This is the brightness value. For the first The luminance component value of each pixel in the HSV color space This represents the number of pixels in the image; based on the calculation results, the image's brightness values ​​are divided into five levels. The texture roughness is calculated using Shannon entropy, and the expression is: ； in, For texture roughness, This represents the probability distribution of pixels with a gray value of 𝑘 in the image; and is divided into three texture roughness levels: smooth, medium, and rough based on the calculation results. The contrast ratio is calculated as the ratio of the grayscale standard deviation to the mean, expressed as: ； in, For contrast, The standard deviation of the image grayscale. The mean value of the image grayscale. It is a very small value; the contrast value is divided into five levels to characterize the brightness and darkness of the image; The saturation is calculated by converting the image from the RGB color space to the HSV color space, and the expression is: ； in, For overall image saturation, For the first The saturation of each pixel is divided into three levels of color saturation. The uniformity of illumination is calculated using the standard deviation of the luminance component, expressed as: ； in, For uniform illumination, The luminance component in the HSV color space The standard deviation; divided into three levels of light uniformity.

5. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, The construction and training of the improved Swin-Transformer model described in step S3 includes the following steps: S3.1: Data annotation, assigning style labels to each sample in the subset of scene image data; S3.2: Randomly divide the labeled dataset into training and validation sets in an 8:2 ratio; S3.3: Model training, using the following improvement strategies: (1) Dynamic layered unfreezing and differentiated learning rate strategy: Freeze the parameters of the first 50% of the network layers of the Swin-Transformer, unfreeze the parameters of the last 50% of the network layers for training, and allocate a learning rate to the backbone network that is less than that to the classification head. (2) ECA channel attention module embedding strategy: ECA channel attention module is embedded after the feature extraction layer of Swin-Transformer to enhance style-related channel features; (3) Lightweight data augmentation strategy guided by style features: retain random horizontal flipping and color dithering, reduce the adjustment range of color brightness, contrast and saturation from 0.3 to 0.2, and reduce the hue coefficient from 0.05 to 0.03; (4) Adaptive learning rate scheduling strategy: A combination of 5 rounds of linear preheating and cosine annealing restart is adopted. In the first 5 rounds, the learning rate is linearly increased from 0.001 to the set value. After that, cosine annealing is restarted every 20 rounds, and the learning rate is gradually reduced to 1e-6. (5) TTA Enhanced Inference Strategy: During the inference phase, the test image is horizontally flipped for enhancement, and the inference results of the original image and the flipped image are averaged. S3.4: Implicit style features assign style labels to each image using the classification head of the improved Swin-Transformer model after training.

6. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, The extraction of odor features in step S4 includes the following steps: S4.1: Based on the Keller 2016 odor perception dataset, the subjects' ratings for each odor sample on three dimensions: intensity, pleasantness, and familiarity are calculated. Intensity reflects the strength of the odor, pleasantness reflects the degree of pleasantness or unpleasantness of the odor, and familiarity reflects whether the odor is familiar to the subject. The mean scores of each dimension are calculated as the mean intensity, mean pleasantness, and mean familiarity of the odor. The formula for calculating the mean of the intensity dimension is as follows: ； in, For the number of valid data points, For the first Intensity ratings for each subject; S4.2: Count the number of valid subjects for each odor sample under each sub-odor category label, select the sub-label corresponding to the maximum number of valid subjects as the primary label, and record the maximum value and its corresponding label name; S4.3: Output the odor feature data for each odor sample, including: unique odor identifier, number of valid subjects, mean intensity, mean pleasantness, mean familiarity, maximum value of sub-odor category labels and corresponding label names.

7. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, The design semantic mapping of the multimodal environment features in step S5 includes the following steps: S5.1: Normalize the environmental sound characteristics to obtain normalized characteristics. The minimum-maximum normalization method with quantile truncation is used: ； ； in, These are the original eigenvalues ​​to be normalized. The median value after quantile truncation. For quantiles, It is a very small positive number; The environmental sound semantic variables are constructed as follows: Ambient noise level variables , used to characterize the intensity of environmental sound energy; Environmental noise active variables , The weighting coefficients for the zero-crossing rate feature. It is a sigmoid nonlinear mapping function used to characterize the rhythm and variation characteristics of environmental sounds; Sharp variables of ambient sound , used to characterize the proportion of high-frequency components; Stable variables of ambient sound ; right The language values ​​of low, medium, and high were used for blurring respectively; S5.2: Construct the basic visual variables as follows: The explicit visual features in the scene image are subjected to ordered numerical encoding and normalization, and the discrete values ​​at each level are converted into discrete values. Map to the [0,1] interval using the following formula: ； in, This refers to the maximum number of levels for this feature, specifically the maximum number of levels for brightness and contrast. The maximum number of grades for texture roughness, saturation level, and illumination uniformity. ; Let the mapped brightness level variable be... Contrast level variable is The saturation level variable is The light uniformity level variable is Texture roughness level variable is ; Hue temperature variable Based on the color scheme: Warm is set to 1, Neutral to 0.5, and Cool to 0; Based on explicit visual attributes and style conditions, construct visual semantic variables: Visual stimulus variables Where, 𝐶 is the normalized contrast value and S is the normalized saturation value; Visual morphological variables ,in These represent the complexity scores of texture roughness, contrast, and style tags in implicit style features, respectively. Visual technology variables ,in These represent the technical scores for style tags in the tonal temperature variable, contrast, and implicit style features, respectively. Visual fusion variables ,in These represent the blending scores of style tags in hue temperature, brightness, lighting uniformity, and implicit style features, respectively. The scoring of style tags in implicit style features is determined by expert experience; Establish membership functions for Low, Mid, and High linguistic values ​​for the basic visual variables and visual semantic variables, respectively; S5.3: Construct olfactory semantic variables as follows: olfactory stimulus variables Olfactory pleasure variables olfactory affinity variables By respectively through the mean intensity Mean of pleasure Average familiarity Obtained by quantile truncation and normalization; Odor category variable The maximum value of each sub-label and its corresponding label name Characterization; right The three language values ​​of Low, Mid, and High were used for blurring respectively. Dedicated rules are used for triggering; S5.4: Employ Mamdani-type fuzzy inference, using the fuzzified results of environmental sound semantic variables, basic visual variables, visual semantic variables, and olfactory semantic variables as antecedents, to design guiding parameter vectors. As the consequent, reasoning is performed according to a preset rule base; in, For stimulation level, For simplicity, For emotional affinity, For a sense of technology, For the rhythm of interaction, For stability, For a sense of security, For degree of integration; The antecedent of a rule is determined by the rule definition using either an AND or OR operation, i.e., min or max. Rule implication uses min, and rule aggregation uses max. For output For each dimension, establish three language values: Low, Mid, and High, and define corresponding membership functions; S5.5: For each rule Activation strength is calculated based on the antecedent. : ； in, , , These are the membership functions for auditory, visual, and olfactory semantic variables, respectively. Will The function is applied to the consequent output membership function, and multiple rule results on the same output dimension are aggregated using max aggregation to obtain the output fuzzy set. ; For each output dimension, the centroid method is used to defuzzify and obtain a definite value: ； in, The determined value after deblurring. To output continuously valued variables in the universe of discourse, The membership function value of the aggregated output fuzzy set at the value u. To avoid extremely small positive numbers with a denominator of 0; The defined values ​​of all output dimensions together constitute the design guiding parameter vector. .

8. The product design strategy generation method based on environmental feature extraction according to claim 1, characterized in that, Step S6, generating the product design strategy, includes: Based on stimulation level The values ​​of these parameters are used to output adjustment commands for the contrast coefficient and saturation coefficient. Based on simplicity The value of is used to output the adjustment instructions for the number of product structure levels and interface information density parameters; Based on emotional affinity The value of is used to output the adjustment commands for surface roughness and surface curvature radius parameters; Based on the sense of technology The value of the parameter determines the output material type parameter adjustment command; According to the rhythm of interaction The value of is used to output instructions for adjusting the interactive feedback frequency and response delay parameters; According to stability The values ​​of are used to output adjustment commands for the bottom support width parameter and the center of gravity height parameter; Based on sense of security The value of the parameter determines the output of the accidental touch protection level parameter and the confirmation step strength parameter adjustment command. According to the degree of integration The value of is used to output the color difference parameter adjustment command between the main color of the product and the main color of the environment; The instructions are used as input to the computer-aided design system to automatically generate or adjust the digital model of the product.

Images