Product design method and system of differentiable hierarchical generative adversarial network

Abstract

The application discloses a product design method and system of a differentiable hierarchical generative adversarial network, acquires user multi-source input, and constructs a user preference feature vector by using a gate fusion network; product design specifications are coded into a continuous differentiable symbolic distance function to construct a rule constraint; a hierarchical generative adversarial network containing a basic shape generator, a style adjustment generator and a high-resolution generator is further constructed, and a multi-condition discriminator with a cross-attention mechanism is designed; a composite loss function containing an adversarial loss, a conditional consistency loss and a rule differentiable loss is used for joint training. After generating a candidate scheme, reinforcement learning is used for local optimization, and online injection fine-tuning of user feedback is supported. The above scheme effectively improves the process compliance and personalized matching degree of generative design, and can be widely applied to the field of intelligent design.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer-aided design technology, and is a product design method and system based on generative adversarial networks, particularly a technical solution that embeds product engineering rules into the generative adversarial network training process in a differentiable manner. Background Technology

[0002] Generative Adversarial Networks (GANs) have been widely applied in fields such as image generation, style transfer, and product design due to their powerful ability to fit data distributions. In product design scenarios, researchers have attempted to use conditional GANs to generate corresponding product design schemes based on user-input style, color, and other labels.

[0003] However, existing technologies have revealed the following inherent flaws in practical applications: First, existing GAN models learn only from data distributions. While the images they generate may be visually realistic, they often violate basic engineering design standards, such as structural strength constraints for furniture products, heat dissipation space requirements for electronic products, and size standards for clothing. These standards are difficult to learn from image data alone, making the generated solutions unsuitable for direct use in downstream computer-aided design and manufacturing processes.

[0004] Second, when users' personalized needs, such as "ultra-narrow bezels," conflict with product physical rules, such as "bezel width must not be less than 5mm to ensure structural strength," existing models lack a mechanism to balance the two, either sacrificing user experience or outputting non-compliant solutions.

[0005] Third, traditional GANs and their variants tend to converge to a few patterns during training, resulting in highly homogeneous design schemes that fail to meet users' expectations for diverse choices.

[0006] To address the aforementioned issues, some existing techniques attempt to add rule validation filters in the post-processing stage or train a separate rule classifier to filter out violations. However, these methods are all "post-hoc remedies," inefficient, and unable to guide the generator to fundamentally learn to avoid violations. Other techniques attempt to use rules as additional label inputs to conditional GANs, but discrete labels cannot represent the continuous boundaries of rules and are difficult to handle complex geometric constraints. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, a product design method and system using differentiable hierarchical generative adversarial networks is provided. Product design rules are encoded as differentiable signed distance functions and used as loss terms in the generator's backpropagation. This allows the generator to learn the rule boundaries from the training phase, ultimately outputting a technical solution that simultaneously satisfies user preferences and engineering specifications.

[0008] To solve one of the aforementioned technical problems, the following technical solution is adopted: This invention first acquires the user's natural language description and the user's behavioral interaction sequence during use. A BERT model is used to semantically encode the natural language, obtaining a first semantic vector; a temporal convolutional network is used to perform pattern mining on the behavioral sequence, obtaining a second behavioral vector. To dynamically fuse these two heterogeneous features, this invention designs a gated fusion network. This network outputs two dynamic weights based on the statistical characteristics of the input vectors, and the two vectors are weighted and summed to obtain the final user preference feature vector. Compared to simple concatenation or fixed-weight summation, the gated fusion mechanism can adaptively adjust the proportion of semantic and behavioral information according to the uncertainty of the input data, improving the accuracy of preference representation.

[0009] Each product design rule is encoded as a continuously differentiable signed distance function (SDF). For a rule involving geometric dimensions, such as "handrail width not less than 15cm," its equivalent constraint can be written as: Define the signed distance function. When the design parameters satisfy the rule, the SDF value is negative or zero; when the rule is violated, the SDF value is positive, and the magnitude of the positive value directly quantifies the severity of the violation. For more complex rules, such as proportional constraints, multiple SDF functions can be combined. Ultimately, the overall rule-differentiable loss is defined as the sum of the positive parts of the SDF of each rule at all sampled geometric points. Because the SDF function is differentiable in design, the gradient of this loss can propagate back along the computation graph to each layer of the generator's parameters, thereby guiding the generator to actively adjust geometric parameters to avoid rule violations.

[0010] In some implementations, a three-tiered cascaded generator is constructed to simultaneously control both the macroscopic structure and microscopic details of the product: Basic Shape Generator: Employing a deep convolutional generative adversarial network (GAN) structure, this generator takes product category identifiers as input and outputs a low-resolution basic shape tensor. It is responsible for learning the basic outline and proportions of the product.

[0011] Style Adjustment Generator: Cascaded after the basic shape generator, it employs an improved StyleGAN structure. Its core consists of multiple adaptive instance normalization layers. This generator receives the basic shape tensor and the user preference feature vector, and modulates the user preference style information onto the basic shape through the adaptive instance normalization layers. Specifically, the transformation parameters of the adaptive instance normalization layers in this invention are not directly obtained from the fully connected preference feature vector, but are generated by an independent hypernetwork. This design increases the generator's ability to express complex preference combinations.

[0012] High-resolution generator: Cascaded after the style adjustment generator, a progressive generative adversarial network growth structure is used to gradually upsample the stylized design tensor to the final high resolution, outputting a high-fidelity product design image that can be directly reviewed by users.

[0013] The present invention also provides a multi-condition discriminator, comprising: The authenticity discrimination branch, similar to the standard GAN discriminator, is used to distinguish between generated images and real product images.

[0014] The conditional consistency discriminant branch is a multimodal discriminator. Its inputs include the feature map of the generated image, the semantic embedding of the user preference feature vector, and the constraint embedding of the rule base. Through an internally configured cross-attention module, this branch outputs a score between 0 and 1 to measure how well the generated image matches the given conditions.

[0015] The total loss function of this invention is a weighted sum of adversarial loss, conditional consistency loss, and rule differentiability loss. The training process consists of three stages: pre-training, joint training, and fine-tuning. The differentiable rule loss guides the generator to actively approach the rule boundaries while optimizing user preferences.

[0016] After generating multiple candidate design schemes, the system filters them based on three dimensions: preference matching degree, rule compliance, and design diversity. For the selected candidate schemes, a reinforcement learning agent based on proximal policy optimization is launched. Using the comprehensive score as a reward, the agent performs a local policy search in the latent space of the high-resolution generator and outputs the optimal scheme after fine optimization.

[0017] More importantly, there is an online feedback injection mechanism. When a user provides suggestions for modifying the recommended solution, the system parses the user's modification into a low-dimensional local adjustment vector. This vector is directly injected into the current forward propagation process of the high-resolution generator, enabling real-time design fine-tuning without retraining the model.

[0018] Compared to existing technologies, this invention is the first to embed product engineering rules into the training process of a generative adversarial network in the form of a continuous, differentiable signed distance function. Compared to post-processing filters or rule classifiers, the differentiable loss of the rules in this invention directly participates in gradient backpropagation, enabling the generator to fundamentally learn the rule boundaries and significantly improving the process compliance rate.

[0019] By decoupling the basic shape generator, style adjustment generator, and high-resolution generator at different levels, the macro structure, style attributes, and micro details of the product are modeled separately. This effectively alleviates the problem of mutual interference between user preference constraints and rule constraints in traditional GANs, enabling the two to achieve synergistic optimization at their respective levels.

[0020] A gated fusion network is employed to dynamically fuse semantic and behavioral vectors. Compared to static weight concatenation, this method adaptively adjusts the information weighting based on input uncertainty, improving the robustness and accuracy of user preference representation. In the style adjustment generator, the parameters of the adaptive instance normalization layer are generated by a hypernetwork, rather than a simple fully connected mapping. This enhances the model's ability to express complex, multimodal user preferences and increases the diversity of generative designs. User interaction feedback is parsed into local adjustment vectors in real time and injected into the generator's forward process, enabling online design fine-tuning without retraining. This results in fast response times and a superior user experience. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the product design method for differentiable hierarchical generative adversarial networks in this embodiment of the invention; Figure 2 This is a schematic diagram of the structure of the stacked generative adversarial network (H-GAN) in an embodiment of the present invention; Figure 3 This is a schematic diagram of the internal structure of the multi-condition discriminator in an embodiment of the present invention; Figure 4 This is a schematic diagram of the gating fusion network and user preference feature encoding process in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the online feedback injection mechanism in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the embodiments. The embodiments are only intended to provide a clearer understanding of the technical features, objectives and effects of the present invention.

[0023] refer to Figure 1-5 This invention provides a product design method for differentiable hierarchical generative adversarial networks, comprising the following steps: Acquire multi-source input data, which includes user natural language descriptions, user behavior interaction sequences, and product category identifiers; The natural language description and user behavior interaction sequence are encoded and dynamically fused through a gated fusion network to generate a user preference feature vector. Encode each rule in the preset product design rule base into a continuously differentiable symbolic distance function; Constructing a hierarchical generative adversarial network, which includes: A basic shape generator is used to generate a basic shape tensor based on the product category identifier; A style adjustment generator, cascaded after the base shape generator, includes an adaptive instance normalization layer for receiving the base shape tensor and the user preference feature vector, and outputting a stylized design tensor. A high-resolution generator, cascaded after the style adjustment generator, is used to upsample the stylized design tensor to the target resolution to generate the final product design image. A multi-condition discriminator, including a truth-based discriminator branch and a condition consistency discriminator branch; The hierarchical generative adversarial network is trained using a composite loss function, which includes: an adversarial loss calculated by the realism discriminant branch, a conditional consistency loss calculated by the conditional consistency discriminant branch, and a rule-differentiable loss calculated by substituting the geometric point coordinates in the basic shape tensor or the stylized design tensor into the symbolic distance function. The trained hierarchical generative adversarial network is used to generate product design schemes that conform to the user preference feature vectors and satisfy the product design rule base.

[0024] The calculation method for the rule-differentiable loss is as follows: For the k-th rule in the product design rule base, its symbolic distance function is denoted as: The geometric point set P is obtained by sampling from the basic shape tensor or stylized design tensor. but ;in, This indicates the degree of violation of rule k at that point. The loss is differentiable, and its gradient is directly backpropagated to the basic shape generator and style adjustment generator.

[0025] The process of generating user preference feature vectors includes: The natural language description is encoded using the BERT model to obtain the first semantic vector; The behavioral interaction sequence is encoded using a temporal convolutional network to obtain a second behavioral vector; The first semantic vector and the second behavior vector are concatenated and then input into a gated fusion network. The gated fusion network outputs two dynamic weights, which are used to perform a weighted summation of the first semantic vector and the second behavior vector to obtain the user preference feature vector.

[0026] In the style adjustment generator, the adaptive instance normalization layer's transformation parameters γ,β are not directly obtained from the fully connected user preference feature vector, but are generated by an independent supernetwork that takes the user preference feature vector as input and outputs the γ,β parameters.

[0027] Among them, the conditional consistency discriminant branch is a multimodal discriminant branch, and its input includes: The first feature map extracted from the final product design image by a convolutional network; The global semantic embedding obtained by linear mapping of the user preference feature vector; The product design rule base is embedded with rule constraints obtained by the encoder. The conditional consistency discrimination branch has a cross-attention module, which uses the first feature map as a query, the global semantic embedding and the rule constraint embedding as keys and values, and outputs a conditional consistency score.

[0028] It also includes: performing multi-dimensional screening on the generated candidate design schemes to select the top-ranked candidate schemes; using a near-end policy optimization algorithm, with the weighted sum of the multi-dimensional scores as the reward function, performing a local policy search on the latent space of the high-resolution generator, and iteratively generating optimized design schemes; receiving user interaction feedback on the optimized design schemes, parsing the interaction feedback into a local adjustment vector, and directly injecting the local adjustment vector into the current forward propagation process of the high-resolution generator to achieve online design fine-tuning.

[0029] This invention also discloses a product design system based on rule-differentiable hierarchical generative adversarial networks, comprising: The input parsing module receives the user's natural language description and behavioral interaction sequence, and outputs a user preference feature vector. The rule encoding module stores the product design rule base and encodes each rule into a differentiable signed distance function; A hierarchical generative adversarial network module, which embeds a basic shape generator, a style adjustment generator, a high-resolution generator, and a multi-condition discriminator trained according to the above method, is used to generate candidate product design images based on the user preference feature vector. The feedback optimization module is used to receive user interaction feedback and perform online design fine-tuning of the above methods.

[0030] Example: Personalized Design of Smart Sofa This example uses a smart sofa as the design object to explain in detail the implementation process of the present invention.

[0031] Step 1: Multi-source input acquisition The user inputs a natural language description through the front-end interface: "Design a modern minimalist smart sofa, three-seater, made of gray technical fabric, with electric reclining function and USB charging port, with a budget of around 6,000 yuan. I like a low backrest and wide armrests, and I hope the armrests are slightly rounded." The system also reads the user's historical behavior data: In the past 30 days, the user has browsed 23 sofas on the e-commerce platform. The average time spent on the sofas labeled "modern minimalist" is 45 seconds, the click rate of the "tech fabric" label is 80%, and the sofa with "wide armrests" design has been saved twice.

[0032] Step 2: Construction of User Preference Feature Vectors (see...) Figure 4 ) A pre-trained BERT-base-chinese model is used to encode the natural language description, and the 768-dimensional output at the [CLS] position is taken as the first semantic vector. A temporal convolutional network containing four layers of dilated convolutions is used to process the user behavior sequence, and outputs a 256-dimensional second behavior vector.

[0033] The gated fusion network consists of two fully connected layers: the first layer maps the concatenated 1024-dimensional vector to 128 dimensions, and the second layer outputs 2-dimensional logits, which are then processed by Softmax to obtain weights α and β. The final user preference feature vector is the product of α and the first semantic vector plus the product of β and the second behavioral vector.

[0034] Step 3: Encoding the symbolic distance function of the rule base For smart sofa products, the following three core rules apply: armrest width is no less than 15cm; the angle between the backrest and the seat cushion is between 100° and 120°.

[0035] Sufficient space is required under the seat cushion to accommodate the electric mechanism. During training, a set of geometric points is sampled from the design parameters output by the generator, and the rule-based loss is calculated.

[0036] Step 4: Construction and training of hierarchical generative adversarial networks (see...) Figure 2 , Figure 3 ) A three-level generator is constructed: the basic shape generator outputs a basic shape image of 128×128×3; the style adjustment generator combines it with the user preference feature vector to output a stylized design image of 512×512×3; the high resolution generator progressively upsamples to 1024×1024 and outputs the final product design image.

[0037] In the multi-condition discriminator, the authenticity discrimination branch adopts the PatchGAN structure, while the conditional consistency discrimination branch fuses image features, user preference embeddings, and rule constraint embeddings through a cross-attention module to output a consistency score. Loss weights are set, and the Adam optimizer is used for training.

[0038] Step 5: Candidate Solution Selection and Reinforcement Learning Optimization Input user preference feature vectors into a trained hierarchical generative adversarial network (GAN), generating 20 candidate solutions in a single run. Calculate the preference matching degree, rule compliance, and diversity of each solution, and select the 5 solutions with the highest overall scores. Use a proximal policy optimization algorithm to refine these 5 solutions and output the optimal solution.

[0039] Step 6: Online feedback injection (see...) Figure 5 ) The optimal recommended solution is presented to the user. The user provides feedback: "The curve at the front of the armrest could be extended a little further, and the USB port should ideally be moved to the side." The system analyzes this feedback, generates a local adjustment vector, and directly injects it into the current forward propagation process of the high-resolution generator. Within 2 seconds, a modified solution is generated for the user's confirmation.

[0040] This invention fundamentally solves the technical problems of poor practicality and prominent contradiction between personalization and compliance in AI-generated design by encoding engineering rules into differentiable symbolic distance functions and embedding them into the training process of hierarchical generative adversarial networks.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A product design method for differentiable hierarchical generative adversarial networks, characterized in that, include: Acquire multi-source input data, which includes user natural language descriptions, user behavior interaction sequences, and product category identifiers; The natural language description and user behavior interaction sequence are encoded and dynamically fused through a gated fusion network to generate a user preference feature vector. Encode each rule in the preset product design rule base into a continuously differentiable symbolic distance function; Constructing a hierarchical generative adversarial network, which includes: A basic shape generator is used to generate a basic shape tensor based on the product category identifier; A style adjustment generator, cascaded after the base shape generator, includes an adaptive instance normalization layer for receiving the base shape tensor and the user preference feature vector, and outputting a stylized design tensor. A high-resolution generator, cascaded after the style adjustment generator, is used to upsample the stylized design tensor to the target resolution to generate the final product design image. A multi-condition discriminator, including a truth-based discriminator branch and a condition consistency discriminator branch; The hierarchical generative adversarial network is trained using a composite loss function, which includes: an adversarial loss calculated by the realism discriminant branch, a conditional consistency loss calculated by the conditional consistency discriminant branch, and a rule-differentiable loss calculated by substituting the geometric point coordinates in the basic shape tensor or the stylized design tensor into the symbolic distance function. The trained hierarchical generative adversarial network is used to generate product design schemes that conform to the user preference feature vectors and satisfy the product design rule base.

2. The product design method according to claim 1, characterized in that, The calculation method for the differentiable loss of the rule is as follows: For the k-th rule in the product design rule base, its symbolic distance function is denoted as: If a geometric point set P is obtained by sampling from the basic shape tensor or stylized design tensor, then... ;in, This indicates the degree of violation of rule k at that point. The loss is differentiable, and its gradient is directly backpropagated to the basic shape generator and style adjustment generator.

3. The product design method according to claim 1, characterized in that, The process of generating the user preference feature vector includes: The natural language description is encoded using the BERT model to obtain the first semantic vector; The behavioral interaction sequence is encoded using a temporal convolutional network to obtain a second behavioral vector; The first semantic vector and the second behavior vector are concatenated and then input into a gated fusion network. The gated fusion network outputs two dynamic weights, which are used to perform a weighted summation of the first semantic vector and the second behavior vector to obtain the user preference feature vector.

4. The product design method according to claim 1, characterized in that, The adaptive instance normalization layer in the style adjustment generator does not directly obtain the transformation parameters γ,β from the fully connected user preference feature vector. Instead, it is generated by an independent supernetwork that takes the user preference feature vector as input and outputs the γ,β parameters.

5. The product design method according to claim 1, characterized in that, The conditional consistency discriminant branch is a multimodal discriminant branch, and its inputs include: The first feature map extracted from the final product design image by a convolutional network; The global semantic embedding obtained by linear mapping of the user preference feature vector; The product design rule base is embedded with rule constraints obtained by the encoder. The conditional consistency discrimination branch has a cross-attention module, which uses the first feature map as a query, the global semantic embedding and the rule constraint embedding as keys and values, and outputs a conditional consistency score.

6. The product design method according to claim 1, characterized in that, Also includes: The generated candidate design schemes are screened from multiple dimensions to select the top-ranked candidate schemes. The near-end policy optimization algorithm is adopted, and the weighted sum of the multi-dimensional scores is used as the reward function to perform local policy search on the latent space of the high-resolution generator and iteratively generate the optimized design scheme. The system receives user feedback on the optimized design, parses the feedback into a local adjustment vector, and directly injects the local adjustment vector into the current forward propagation process of the high-resolution generator to achieve online design fine-tuning.

7. A product design system based on rule-differentiable hierarchical generative adversarial networks, characterized in that, include: The input parsing module receives the user's natural language description and behavioral interaction sequence, and outputs a user preference feature vector. The rule encoding module stores the product design rule base and encodes each rule into a differentiable signed distance function; A hierarchical generative adversarial network module, which embeds a basic shape generator, a style adjustment generator, a high-resolution generator, and a multi-condition discriminator trained according to any one of claims 1 to 5, is used to generate candidate product design images based on the user preference feature vector; The feedback optimization module is used to receive user interaction feedback and perform online design fine-tuning of the method according to claim 6.

Images