A dynamic constraint evolution product appearance customization method and system

Abstract

This invention belongs to the field of computer-aided design and intelligent optimization, specifically relating to a method and system for product appearance customization with dynamic constraint evolution. The method includes: constructing a visual semantic mapping mechanism to obtain user feature requirement vectors; automatically intercepting physically infeasible solutions through an embodied constraint matrix based on a knowledge graph; and during the evolution process, monitoring the user's manual unlocking behavior of constraint interception schemes and dynamically adjusting the penalty weight of corresponding constraints according to a preset update formula to achieve adaptive evolution of constraint rules. Users only need simple operations such as browsing and clicking to complete the personalized customization of the appearance of highly constrained products. This invention can significantly reduce the interaction load, improve customization efficiency, and ensure the engineering feasibility of the design scheme.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided design and intelligent optimization, specifically relating to a product appearance customization method and system with dynamic constraint evolution, which is particularly suitable for product appearance personalization customization scenarios with hard engineering constraints, such as robots and mining equipment. Background Technology

[0002] In the context of "Industry 5.0," mass customization has become the mainstream production mode. Interactive Genetic Algorithms (IGAs), as a human-machine co-evolutionary algorithm that incorporates subjective human evaluation into the optimization loop, can involve users in the design process, achieving "human-machine co-creation." Currently, IGAs have been applied in multiple design fields: for example, in the conceptual design of automotive control panels, some studies have introduced hesitation mechanisms to reduce evaluation noise; in the design of train head shapes, some scholars have used fuzzy fitness functions to alleviate the user's scoring burden. However, in the customization of product appearance with high constraints, such as robots and mining equipment, existing IGA methods still have the following prominent problems: (1) Cognitive noise and interaction load issues: In the early stages of customization, users' needs are vague, and there are a large number of random and systematic errors (i.e., "cognitive noise") in subjective evaluations that deviate from true preferences. Existing IGA methods usually require users to give precise numerical ratings to multiple individuals in each generation of the population, which leads to severe cognitive fatigue in users in the early stages of evolution, resulting in heavy interaction load and low efficiency. Some studies have attempted to alleviate this problem through fuzzy semantic evaluation or hesitation modeling, but have failed to fundamentally simplify the interaction pattern.

[0003] (2) Lack of Embedded Constraints: The appearance design of highly constrained products must strictly meet physical hard constraints such as kinematics, dynamics, and spatial layout (i.e., "embodied constraints"). Existing patents such as CN114298308B introduce penalty functions to check spatial size constraints in furniture customization, but such methods are mostly static rule matching, which is difficult to handle the complex coupling constraints of form-function-environment in products such as robots. Many solutions generated by traditional IGA are physically infeasible (such as appearance parts interfering with motion mechanisms, occluding sensor fields of view), resulting in a low ratio of effective solutions, and a lot of user evaluation effort is wasted on screening infeasible solutions.

[0004] (3) Static constraints and inability to adapt to design intent: The constraint filtering rules in the existing IGA method are mostly one-time presets and fixed throughout the process. When users repeatedly try to bypass a certain type of non-critical constraint in the actual design process (indicating that the constraint can be negotiated in a specific context), the system still strictly intercepts it, cannot perceive and adapt to the user's true design intent, limits the design exploration space, and leads to slow evolutionary convergence or getting stuck in local optima.

[0005] (4) Insufficient evolutionary guidance: The evolutionary process of traditional IGA lacks phased guidance. Whether the evolution is in the initial exploration stage or the later convergence stage, users face the same evaluation interface and operation logic, making it difficult to focus on the appropriate design goals at the right time (such as focusing on feasibility in the early stage and aesthetic preferences in the later stage), resulting in low evolutionary efficiency.

[0006] The paper by Yufeng Zhang et al., "An interactive genetic algorithms system customizes robot appearance via cognitive noise filtering under embodied constraints" (Scientific Reports, February 2026), while introducing the concept of "embodied constraints," treats constraints statically—the constraint matrix remains fixed after system initialization and cannot adaptively adjust based on actual user behavior. The paper does not disclose a feedback mechanism for dynamically adjusting the constraint matrix based on user unlocking behavior, nor does it address weight aggregation methods in multi-user collaborative scenarios. When users show preference for solutions deemed "infeasible" by the system, existing systems can only mechanically block them, failing to learn user unlocking behavior and correspondingly soften the constraints.

[0007] While the patent document CN114298308B, titled "A Furniture Personalization Customization System and Method Based on Interactive Genetic Algorithm," employs IGA for scheme optimization, it primarily addresses the layout of internal storage space within wardrobes, offering limited optimization of user cognitive load and failing to address product appearance design. Its penalty function is merely used to guide the scheme's direction; it does not construct a multi-dimensional embodied constraint matrix covering form, function, and environment, nor does it achieve dynamic evolution of constraints.

[0008] Although the master's thesis "Research on 3D Product Appearance Design Based on Improved Interactive Genetic Algorithm" (2024) focuses on the problem of "evaluation noise", its improvement is concentrated on the global convergence performance at the algorithm level, and does not reduce noise from the source of demand acquisition - that is, it lacks an effective semantic guidance mechanism to help users accurately express their preferences.

[0009] In summary, existing technologies in the field of product appearance customization suffer from several shortcomings, including static constraint processing, ineffective suppression of user cognitive noise, lack of multi-user collaboration mechanisms, fragmented application scenarios, and lack of closed-loop feedback. Therefore, there is an urgent need for an improved interactive genetic algorithm that can organically embed embodied constraints into the IGA evolution process, dynamically and adaptively adjust the constraint strength based on user interaction feedback, and achieve low-load interaction. Summary of the Invention

[0010] The purpose of this invention is to provide a product appearance customization method and system with dynamic constraint evolution to address the problems existing in the prior art. Through an integrated architecture of "visual semantic mapping - knowledge graph constraint verification - dynamic constraint evolution", it solves the problems of static constraints being unable to adapt to user intentions and high interaction load in existing interactive customization methods.

[0011] The technical solution of this invention is: A method for product appearance customization with dynamic constraint evolution includes the following steps: S1. Demand Acquisition and Semantic Mapping: Based on a preset visual semantic mapping library of "text-image-feature encoding", the system receives the current user's selection of preset intention semantics, generates and displays representative samples corresponding to the intention semantics; after the user confirms that the sample meets expectations, the system decodes the sample to generate a feature demand vector containing the user's preferences. S2. Embodied Constraint Matrix Construction and Solution Space Filtering: Constructing a Parametric Relationship Matrix F based on a Pre-defined Knowledge Graph K The matrix defines the compatibility constraints between various design parameters, including morphological constraints, functional constraints, and environmental constraints; the feature requirement vector generated in step S1 is validated using the relation matrix, and physically infeasible design schemes are automatically intercepted to generate a valid requirement set; S3. Demand-guided population initialization: The effective demand set is used as the preference-guided individuals and mixed with randomly generated individuals to form the initial population of the interactive genetic algorithm (IGA). S4. Interactive Evaluation and Dynamic Constraint Evolution: Receive at least one user's interaction, and based on the comprehensive fitness function: Make a selection, among which Rate users To constrain the degree of violation, Weighted ratings for multiple users This represents the constraint weight; in the early stages of evolution... Higher weights are used to filter feasible solutions in the later stages of evolution. Higher weighting is used to optimize user preferences; At the same time, an adaptive mutation probability function is adopted:

[0012] in, Represents the probability of individual variation. Represents the basic mutation probability. The decay coefficient is represented by t, which is the generation number; a high mutation rate is applied to individuals with high constraint violation. During the evolution process, the system monitors users' manual unlocking behavior when they are unable to enter the next generation of individuals due to constraint blocking; based on the unlocking behavior records, the parameter relationship matrix is ​​dynamically adjusted according to a preset update formula. The penalty weight for the corresponding constraint type is adjusted to reduce the penalty for constraint types that are frequently unlocked by the user in subsequent evolutions. S5. Elite Retention and Termination Judgment: Receives the user's lock command for a satisfactory individual and directly retains that individual to the next generation; when the user confirms the final satisfactory solution or the preset maximum number of evolution generations is reached, the process terminates and the customized solution is output.

[0013] Through the steps described above, the core of this invention lies in introducing a dynamic constraint weight adjustment mechanism based on user unlocking feedback. When a constraint is repeatedly unlocked manually by the user (indicating that the constraint is not absolutely critical in the user's view), the system adaptively reduces the penalty weight of that constraint, thereby relaxing the restriction in subsequent evolutions and expanding the effective solution space. Simultaneously, through a phased weighting strategy, infeasible solutions are strictly filtered in the early stages of evolution, while fine-tuning is performed later based on user preferences, achieving a dynamic balance between constraint rigidity and design flexibility.

[0014] Specifically, the feature demand vector mentioned in S1 is a set of personalized demands formed by the user through multiple rounds of cross-selection of representative samples corresponding to different semantic intentions. .

[0015] Specifically, the relation matrix mentioned in step S2 It is an M×N matrix, where M is the number of classification items and N is the number of design objectives; the matrix representation is as follows:

[0016] Where the row vector is:

[0017] Matrix elements Defined as:

[0018] For each column Summation yields If all If the requirement set is valid, then the requirement set is valid; if any Make If the requirement set is invalid, automatic interception will be triggered.

[0019] Specifically, the functional constraint conflict detection described in S2 is implemented through a spatial occupancy matrix: a binary occupancy matrix is ​​established for different functional components on the same parameterized discrete grid. and ; Calculate the overlap matrix ,in For the Hadamard product; calculate the sum of the elements of the overlapping matrix:

[0020] in This refers to the grid size; if If a spatial conflict exists, the design is deemed invalid.

[0021] Specifically, the weight allocation strategy for the comprehensive fitness function described in step S4 is as follows:

[0022] in, The term is used to increase the fitness penalty for infeasible solutions; in the early stages of evolution, the focus is on feasibility, and in the later stages of evolution, the focus is on user preference.

[0023] Specifically, the preset update formula for dynamically adjusting the penalty weight in S4 is: For constraint types Its penalty weight The update rules are as follows:

[0024] in, For the first Substitute weight, For the updated weights, The learning rate has a range of values. , This represents the cumulative number of times constraint type K has been manually unlocked by the first role in the current session. The total number of times automatic interception is triggered for constraint type K. To prevent extremely small positive numbers from being divided by zero; the updated Directly substituted into the constraint violation degree It participates in fitness calculation.

[0025] Specifically, in S4, when multiple users participate in the interaction, the multiple users are pre-assigned to at least one first role and at least one second role, wherein the first role has hard-locking authority over functional constraint gene loci, and the second role has soft preference weight adjustment authority over styling gene loci; in each generation of evolution, the preferences of multiple users are weighted and fused through a preset social selection function to generate a unified fitness evaluation, which participates in the calculation of the comprehensive fitness function.

[0026] Specifically, the visual semantic mapping, validity verification, and population evolution process are all completed through simple interactive operations such as browsing and clicking by the user, without requiring the user to perform numerical scoring or parameter configuration.

[0027] Specifically, the product is a unibody robot or mining equipment, the morphological constraints include geometric size constraints and material property constraints, the functional constraints include kinematic feasibility constraints, and the environmental constraints include task scenario adaptability constraints.

[0028] This invention also provides a product appearance customization system with dynamic constraint evolution, comprising: The requirement acquisition and semantic mapping module is used to receive the current user's selection of the preset intention semantic based on the preset "text-image-feature encoding" visual semantic mapping library, generate and display representative samples corresponding to the intention semantic, and decode the sample to generate a feature requirement vector containing the user's preferences after the user confirms that the sample meets the expectations. The embodied constraint matrix construction and solution space filtering module is used to construct parameter relationship matrices based on a preset knowledge graph. The relationship matrix is ​​used to verify the validity of the feature requirement vector, automatically intercepting physically infeasible design schemes and generating a set of valid requirements. The population initialization module is used to mix the effective demand set as preference-guided individuals with randomly generated individuals to form the initial population of the interactive genetic algorithm (IGA). The interactive evaluation and dynamic constraint evolution module receives interactive operations from at least one user, selects individuals based on a comprehensive fitness function, performs mutation operations using an adaptive mutation probability function, and dynamically adjusts the parameter relationship matrix F based on the user's manual unlocking behavior when an individual is blocked from entering the next generation due to constraint. K The penalty weight corresponding to the constraint type; The Elite Retention and Termination Judgment Module is used to receive the user's lock command for a satisfactory individual, directly retain the individual to the next generation, and terminate the process when the user confirms the final satisfactory solution or the preset maximum number of evolution generations is reached, and output the customized solution.

[0029] This application improves the interactive genetic algorithm compared to existing technologies (especially the closest prior art—the paper by Yufeng Zhang et al. published in Scientific Reports in February 2026), and has the following significant technical effects: 1. Constraint handling has evolved from static interception to dynamic adaptation. The closest existing technology constructs an "embodied constraint chromosome specification" to intercept physically infeasible design schemes, but its constraint matrix remains fixed after system initialization and cannot be adaptively adjusted according to the user's actual preferences. US20070112698A1 also relies on a pre-set penalty function to handle physical constraints and does not involve a dynamic evolution mechanism for constraints.

[0030] This application monitors users' manual unlocking behavior when they are unable to enter the next generation of individuals due to constraint blocking, and dynamically adjusts the parameter relationship matrix (F) according to a preset update formula. K The penalty weights for the corresponding constraint types in the model reduce the penalties for constraint types that are frequently unlocked by users in subsequent evolutions. This mechanism enables the system to achieve a dynamic balance between "physical feasibility" and "user satisfaction," overcoming the fundamental flaw of static constraint processing in existing technologies.

[0031] 2. Dynamic constraint weight evolution improves search efficiency and design flexibility: By monitoring user unlocking behavior of constraint interception schemes, the system automatically distinguishes between hard and soft constraints and dynamically relaxes the penalty weight of non-critical constraints. This mechanism effectively expands the search space while ensuring engineering feasibility, avoiding convergence stagnation caused by excessive interception under traditional static constraints.

[0032] 3. Multi-user collaborative evaluation fills the gap in group customization. Existing product appearance customization systems are mostly based on a single-user independent evaluation mode. Although CN110717545B introduces interactive evaluation, its evaluation subject is a single user; CN111639990A belongs to the clothing recommendation field and does not involve multi-user collaborative evaluation scenarios. The closest existing technology is also only aimed at single-user scenarios.

[0033] This application introduces multi-user rating weights into the comprehensive fitness function. Each user rating is a weighted aggregate of multiple user ratings, and the weight coefficient of each user is dynamically adjusted based on the consistency of their historical evaluations. This mechanism enables this application to be effectively applied to multi-user collaborative customization scenarios such as home-shared robots and team office equipment, filling a gap in existing technologies for group customization.

[0034] 4. Semantic mapping has been upgraded from "noise reduction" to "precise guidance". The closest existing technology constructs a "visual-semantic mapping mechanism" to reduce initial cognitive noise through "text-image-symbol" association. However, its mapping relationship is relatively fixed, and no dynamic evaluation mechanism for semantic contribution value has been established.

[0035] In step S1, this application establishes a many-to-many mapping relationship of "text-image-feature encoding" and assigns a semantic contribution value to each mapping relationship (claim 6), enabling the system to more accurately capture user preferences and reduce cognitive noise from the source of demand acquisition. Although CN114298308B uses an interactive genetic algorithm for furniture customization, it mainly solves the problem of internal storage space layout in wardrobes and lacks a general visual semantic mapping framework.

[0036] 5. Adaptive mutation probability enables synergistic optimization of evolutionary efficiency and diversity. Existing interactive genetic algorithms generally suffer from slow convergence speed and a tendency to get stuck in local convergence. CN114298308B introduces a penalty function to guide the direction of scheme generation, obtaining optimization results with fewer evaluations, but it does not achieve adaptive adjustment of mutation probability.

[0037] This application employs an adaptive mutation probability function, applying a high mutation rate to individuals with high constraint violation rates to increase population diversity; simultaneously, the decay coefficient is dynamically adjusted with each generation, gradually reducing the sensitivity of the mutation probability to constraint violation rates in later stages of evolution. This mechanism achieves synergistic optimization between maintaining population diversity and accelerating convergence speed.

[0038] 6. Improved search efficiency and design flexibility: The dynamic constraint weight evolution mechanism monitors user behavior regarding constraint unlocking schemes, automatically distinguishing between hard and soft constraints and dynamically relaxing the penalty weights of non-critical constraints. This mechanism effectively expands the search space while ensuring engineering feasibility, avoiding convergence stagnation caused by excessive interception under traditional static constraints.

[0039] 7. Construct a closed-loop feedback architecture of demand-constraint-evolution. The biggest systemic flaw in existing technologies lies in the lack of an organic closed loop between user demand expression, physical constraint verification, and solution evolution optimization. Most solutions either only solve the semantic mapping problem, only optimize the genetic algorithm itself, or only focus on constraint filtering.

[0040] This application constructs a closed-loop system from requirement acquisition to solution output through a complete process: "S1 semantic mapping to obtain requirements → S2 knowledge graph constraint verification → S3 preference-guided initialization → S4 dynamic constraint evolution → S5 elite retention". In particular, the feedback mechanism of S4 to S2 (dynamically adjusting the constraint matrix according to user unlocking behavior) enables the constraint system to have online learning capabilities, which is a differentiated feature not disclosed in existing technologies.

[0041] 8. Simplified interaction significantly reduces user cognitive load: The evaluation mode of "scoring individuals one by one" in the traditional IGA is changed to simple operations such as browsing, clicking, selecting, and unlocking. Users do not need to understand the principle of genetic algorithm, nor do they need to input values ​​or configure parameters to complete the adaptive evaluation.

[0042] 9. Improved overall efficiency and user experience The closest existing technology claims a 35.2% reduction in user reviews, a 30.4% reduction in user burden, and a 15% reduction in iteration cycle. Building upon these advancements, this application employs a combination of mechanisms, including dynamic constraint evolution (reducing iterations of ineffective solutions), multi-user collaboration (sharing the burden of individual user reviews), and precise semantic mapping (reducing cognitive noise), to further improve customization efficiency, user satisfaction, and solution quality. Compared to CN114298308B, which relies solely on a penalty function to guide solution generation, and the master's thesis, which focuses on global convergence improvements at the algorithmic level, this application represents a systematic innovation across the entire process from requirement gathering and constraint handling to evolutionary optimization. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the structure of the method of the present invention; Figure 2 This is a schematic diagram of the user requirement set acquisition process in this invention; Figure 3 This is a schematic diagram of the intention semantic mapping process in this invention. Figure 4 This is a schematic diagram of the feature encoding generation map and parameter selection matrix in this invention. Detailed Implementation

[0044] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0045] Example 1:

[0046] This embodiment uses the method provided by the present invention to take the appearance customization of a GS-type industrial robot as an example. The user is a designer with basic engineering knowledge who performs appearance customization, such as... Figure 1 As shown, the specific steps are as follows.

[0047] S1: Requirements Acquisition and Semantic Mapping, specifically building a visual semantic mapping library and noise requirement encoding. First, construct a set of decomposable parameters for the product design objective G, for example: Each target is further subdivided into multiple feature categories, forming a parameter selection matrix.

[0048] Then, the system pre-builds a mapping library of "image semantics - representative samples - feature encoding", and image semantics. A schematic diagram of the mapping process is shown below. Figure 3 As shown, when a user selects imagery such as "technological feel," the interface displays the sample image with the highest semantic content.

[0049] After user confirmation, the system will decode the subjective semantics as follows: Feature demand vector Through multiple rounds of cross-selection with multiple samples, a personalized demand matrix is ​​formed. The process of obtaining the chromosome code, which is used for subsequent optimization, is illustrated in the diagram below. Figure 2 As shown.

[0050] S2: Constructing the embodied constraint matrix and filtering the solution space, with schematic diagrams of the feature encoding generation graph and parameter selection matrix as shown below. Figure 4 As shown First, construct a parametric relationship matrix that defines the compatibility between all feature classification items and the design goal. The matrix is A dimension, an element is defined by the following formula:

[0051] This matrix pre-integrates morphological constraints (such as spatial interference), functional constraints (such as the conflict between the installation positions of the gripper and the fire monitor), and environmental constraints.

[0052] Once the user-selected set of requirements is formed, the system sums the values ​​in each column j: If all If there exists a j such that the requirement set is valid; If this occurs, automatic interception will be triggered, and the conflicting parameter pair will be highlighted on the interface.

[0053] Furthermore, for typical functional conflicts, a spatial occupancy matrix is ​​used for precise detection. For example, a space occupancy matrix is ​​established on the robot's torso mounting panel. The discrete grid G. The occupancy matrix of the dexterous hand actuator is: The occupancy matrix of the fire monitor mounting base is as follows: Calculate the overlap matrix: as well as when If a spatial conflict is detected, the demand set is marked as an "invalid individual" and is directly prevented from entering the initial population.

[0054] S3: Demand-Driven Population Initialization The valid demand set confirmed in S2 is used as the preference-guided individuals, which, together with some randomly generated individuals, constitute the initial population. In this embodiment, the population size is 7, with 2-3 preference-guided individuals and the rest being random individuals, to balance convergence speed and diversity.

[0055] S4: Interactive Evaluation and Dynamically Constrained Evolution The system receives interaction input from at least one user, selects an individual to display, and calculates the fitness function. ,in Rate users To constrain the degree of violation, Weighted ratings for multiple users This represents the constraint weight; in the early stages of evolution... Higher weights are used to filter feasible solutions in the later stages of evolution. Higher weights are assigned to optimize user preferences.

[0056] Among them, the degree of constraint violation Constraint types Weighted violation composition: To constrain The original amount of violation, The penalty weight is dynamically updated, with an initial value of 1.0.

[0057] The weight allocation strategy adopts a phased mechanism: At the same time, an adaptive mutation probability is adopted: This embodiment takes , Individuals with higher constraint violation rates have a greater probability of mutation, thereby increasing population diversity.

[0058] Dynamic constraint weight updates are a significant improvement in this invention. During the evolutionary process, when a certain type of constraint... (For example, "interference between leg decorative shell and walking mechanism") When a designer's preferred shape is repeatedly blocked, the engineer can manually unlock the blocked individual and select it for the next generation. The system records the total number of blocks for this constraint type. and number of unlocks At the end of each generation, the weights are updated according to the following formula: Among them, learning rate , For example, if a certain type of constraint is intercepted 8 times ( Engineer unlocked 5 times ( If the weight increases, then its weight increases. The punitive effect is correspondingly weakened. The updated weights are immediately applied to the fitness evaluation of the next generation.

[0059] This dynamic mechanism works synergistically with the phased weighting strategy: high in the early stage... This ensures that core constraints are not easily violated, but when engineers repeatedly unlock a certain type of constraint, the system identifies it as a non-critical constraint and automatically relaxes it to prevent the feasible solution space from becoming too narrow; later... The dynamic adjustment range has also become more gradual, thus expanding the design space.

[0060] Step S5: Elite Retention and Termination The system receives a user's command to lock a satisfactory individual, which then directly advances to the next generation. The process terminates when the user confirms the final design or when the 20th generation is reached, outputting the 3D model and manufacturing parameters.

[0061] Example 2:

[0062] This embodiment uses a multi-role collaborative customization robot as an example to illustrate the method provided by the present invention. The users are an "engineer" (first role) and a "designer" (second role). The engineer has hard-locked permissions for functional positions, while the designer has soft-preference weights for styling positions. After independent evaluation by both, a weighted Borda count method is used to fuse preferences (engineer weight 2, designer weight 1) to form a comprehensive score. Substitute the values ​​into the fitness function. The remaining steps are the same as in Example 1, thus achieving multi-role cooperation based on dynamic constraint evolution.

[0063] Example 3

[0064] This embodiment uses a home service robot as the customized object, and three family members (father, mother and child) participate in the evaluation as three users.

[0065] During the fitness evaluation phase, three users independently rate the individuals in the population. The system records each user's historical rating data and calculates the correlation coefficient between each user's rating and the final aggregated rating. If a user's rating consistently shows low consistency with other users (correlation coefficient < 0.3), the system automatically reduces that user's weight coefficient; conversely, it increases it.

[0066] Through the above mechanism, the system can adaptively identify "reliable evaluators" and assign them higher weights, thereby reducing the impact of "noisy evaluators" on the final solution.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for product appearance customization with dynamic constraint evolution, characterized in that, Includes the following steps S1. Demand Acquisition and Semantic Mapping: Based on a preset "text-image-feature encoding" visual semantic mapping library, the system receives the current user's selection of preset intention semantics, generates and displays representative samples corresponding to the intention semantics; after the user confirms that the sample meets expectations, the system decodes the sample to generate a feature demand vector containing the user's preferences. S2. Embodied Constraint Matrix Construction and Solution Space Filtering: Constructing a Parametric Relationship Matrix F based on a Pre-defined Knowledge Graph K The matrix defines the compatibility constraints between various design parameters, including morphological constraints, functional constraints, and environmental constraints; the feature requirement vector generated in step S1 is validated using the relation matrix, and physically infeasible design schemes are automatically intercepted to generate a valid requirement set; S3. Demand-guided population initialization: The effective demand set is used as the preference-guided individuals and mixed with randomly generated individuals to form the initial population of the interactive genetic algorithm (IGA). S4. Interactive Evaluation and Dynamic Constraint Evolution: Receive at least one user's interaction, and based on the comprehensive fitness function: Make a selection, among which Rate users To constrain the degree of violation, Weighted ratings for multiple users This represents the constraint weight; in the early stages of evolution... Higher weights are used to filter feasible solutions in the later stages of evolution. Higher weighting is used to optimize user preferences; At the same time, an adaptive mutation probability function is adopted: in, Represents the probability of individual variation. Represents the basic mutation probability. This represents the decay coefficient, where t is the generation number. During the evolution process, monitor users' manual unlocking behavior for individuals who are unable to enter the next generation due to constraint blocking; Based on the unlocking behavior records, the parameter relationship matrix is ​​dynamically adjusted according to a preset update formula. The penalty weight for the corresponding constraint type is adjusted to reduce the penalty for constraint types that are frequently unlocked by the user in subsequent evolutions. S5. Elite Retention and Termination Judgment: Receives the user's lock command for a satisfactory individual and directly retains that individual to the next generation; when the user confirms the final satisfactory solution or the preset maximum number of evolution generations is reached, the process terminates and the customized solution is output.

2. The customization method according to claim 1, characterized in that, The feature demand vector mentioned in S1 is a set of personalized demands formed by the user through multiple rounds of cross-selection of representative samples corresponding to different semantic orientations. .

3. The customization method according to claim 1, characterized in that, The relation matrix mentioned in step S2 It is an M×N matrix, where M is the number of classification items and N is the number of design objectives; the matrix representation is as follows: Where the row vector is: Matrix elements Defined as: For each column Summation yields If all If the requirement set is valid, then the requirement set is valid; if any Make If the requirement set is invalid, automatic interception will be triggered.

4. The customization method according to claim 1, characterized in that, The functional constraint conflict detection described in S2 is specifically implemented through a spatial occupancy matrix: a binary occupancy matrix is ​​established on the same parameterized discrete grid for different functional components. and ; Calculate the overlap matrix ,in For the Hadamard product; calculate the sum of the elements of the overlapping matrix: in This refers to the grid size; if If a spatial conflict exists, the design is deemed invalid.

5. The customization method according to claim 1, characterized in that, The weight allocation strategy for the comprehensive fitness function mentioned in step S4 is as follows: in, The term is used to increase the fitness penalty for infeasible solutions; in the early stages of evolution, the focus is on feasibility, and in the later stages of evolution, the focus is on user preference.

6. The customization method according to claim 1, characterized in that, The preset update formula for dynamically adjusting the penalty weight in S4 is: For constraint types Its penalty weight The update rules are as follows: in, For the first Substitute weight, For the updated weights, The learning rate has a range of values. , This represents the cumulative number of times constraint type K has been manually unlocked by the first role in the current session. The total number of times automatic interception is triggered for constraint type K. To prevent extremely small positive numbers from being divided by zero; the updated Directly substituted into the constraint violation degree It participates in fitness calculation. According to the customization method of claim 1, in step S4, when there are multiple users participating in the interaction, the multiple users are pre-assigned as at least one first role and at least one second role, wherein the first role has hard-locking authority over functional constraint gene loci, and the second role has soft preference weight adjustment authority over styling gene loci; in each generation of evolution, the preferences of multiple users are weighted and fused through a preset social selection function to generate a unified fitness evaluation, which participates in the calculation of the comprehensive fitness function.

7. The customization method according to claim 1, characterized in that, The visual semantic mapping, validity verification, and population evolution process are all completed through simple interactive operations such as browsing and clicking by the user, without requiring the user to perform numerical scoring or parameter configuration.

8. The customization method according to claim 1, characterized in that, The product is a unibody robot or mining equipment. The morphological constraints include geometric size constraints and material property constraints. The functional constraints include kinematic feasibility constraints. The environmental constraints include task scenario adaptability constraints.

9. A product appearance customization system with dynamic constraint evolution, characterized in that, include: The requirement acquisition and semantic mapping module is used to receive the current user's selection of the preset intention semantics based on the preset "text-image-feature encoding" visual semantic mapping library, generate and display representative samples corresponding to the intention semantics, and decode the samples to generate feature requirement vectors containing user preferences after the user confirms that the samples meet expectations. The embodied constraint matrix construction and solution space filtering module is used to construct parameter relationship matrices based on a preset knowledge graph. The relationship matrix is ​​used to verify the validity of the feature requirement vector, automatically intercepting physically infeasible design schemes and generating a set of valid requirements. The population initialization module is used to mix the effective demand set as preference-guided individuals with randomly generated individuals to form the initial population of the interactive genetic algorithm (IGA). The interactive evaluation and dynamic constraint evolution module receives interactive operations from at least one user, selects individuals based on a comprehensive fitness function, performs mutation operations using an adaptive mutation probability function, and dynamically adjusts the parameter relationship matrix F based on the user's manual unlocking behavior when an individual is blocked from entering the next generation due to constraint. K The penalty weight corresponding to the constraint type; The Elite Retention and Termination Judgment Module is used to receive the user's lock command for a satisfactory individual, directly retain the individual to the next generation, and terminate the process when the user confirms the final satisfactory solution or the preset maximum number of evolution generations is reached, and output the customized solution.

Images