Personalized design intelligent interaction method based on topological optimization

Abstract

The invention discloses a personalized design intelligent interaction method based on topological optimization, and the method comprises the steps: S1, constructing a user behavior monitoring system based on multi-modal data collection, and generating a high-dimensional data set containing time sequence features; s2, based on an adaptive deep neural network model, performing dynamic prediction and hierarchical modeling on the personalized demand of the user; s3, generating an optimal solution of the design parameters in real time through a dynamic topology reconstruction algorithm; s4, iteratively updating a design scheme generation rule through a reinforcement learning mechanism; s5, constructing a multi-objective design space exploration and optimization framework supported by the neural network, and rapidly screening and optimizing design parameters; s6, dynamic updating and local optimization of the design scheme are achieved through multi-dimensional data mapping and parameterization regulation and control means; and S7, establishing a data-driven continuous learning and evolution mechanism, continuously optimizing the user demand prediction model and designing an optimization algorithm. The method has the advantages of being high in dynamic adaptability, high in intelligent level and high in personalized meeting precision.

Description

Technical Field

[0001] The present invention relates to the field of intelligent interaction technologies, and in particular, to an intelligent interaction method for personalized design based on topology optimization. Background Art

[0002] With the rapid development of intelligent technologies, personalized design has become an important requirement in many fields, especially in user experience, product customization, and intelligent interaction systems. However, existing personalized design methods often have difficulty effectively dealing with complex and changing user needs and dynamic design optimization scenarios, resulting in low efficiency in the design process and an inability to accurately meet the personalized needs of users.

[0003] In the prior art, traditional personalized design methods mainly rely on static rules or predefined templates and optimize them by combining limited user behavior data. These methods have significant deficiencies in terms of dynamics and adaptability, which are specifically manifested in the following problems:

[0004] 1. Lack of dynamic adaptation ability: Traditional personalized design methods usually adopt predefined design templates or fixed optimization rules and cannot be dynamically adjusted according to the real-time changes in user needs, making it difficult to meet complex and diverse personalized needs.

[0005] 2. Lack of intelligent interaction mechanism: Most existing design systems rely on manual user input or a single data source and lack an intelligent feedback mechanism based on multi-modal interaction data, making it difficult to accurately capture the implicit needs and behavior characteristics of users, resulting in inaccurate design adjustments.

[0006] 3. Low design optimization efficiency: In the design optimization process, existing methods often rely on single-objective or static optimization algorithms and cannot comprehensively explore and quickly optimize high-dimensional complex design spaces, resulting in low design generation efficiency and insufficient accuracy.

[0007] 4. Difficulty in achieving continuous learning and optimization: Traditional methods usually do not make full use of user historical interaction data and design optimization records and lack a continuous learning mechanism, making it impossible to dynamically optimize the user demand prediction model and design optimization algorithm according to new user behaviors and historical data, restricting the adaptive ability and evolution ability of the system.

[0008] Therefore, how to provide an intelligent interaction method for personalized design based on topology optimization is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose an intelligent interaction method for personalized design based on topology optimization. The present invention fully combines topology optimization, deep learning, multimodal interaction, and continuous learning technologies, and details the implementation methods of dynamic personalized design adjustment, user demand prediction, multi-objective optimization, and intelligent feedback mechanism, with the advantages of strong dynamic adaptability, high optimization efficiency, high intelligence level, and high accuracy in meeting personalization.

[0010] An intelligent interaction method for personalized design based on topology optimization according to an embodiment of the present invention includes the following steps:

[0011] S1. Construct a user behavior monitoring system based on multimodal data collection, and obtain the operation trajectory, preference characteristics, and multi-source feedback information of the user during the interaction in real time, and generate a high-dimensional data set containing temporal characteristics;

[0012] S2. Based on an adaptive deep neural network model, dynamically predict and hierarchically model the user's personalized needs by integrating user behavior characteristics and context semantic information;

[0013] S3. Adopt a personalized design adjustment module driven by topology optimization, transform the user demand prediction result into a multi-objective optimization problem, and generate the optimal solution of the design parameters in real time through a dynamic topology reconstruction algorithm;

[0014] S4. Through a reinforcement learning mechanism, design an intelligent feedback algorithm based on the user's real-time operations and interaction behaviors, and iteratively update the design scheme generation rules;

[0015] S5. Construct a multi-objective design space exploration and optimization framework supported by a neural network, and quickly screen and optimize the design parameters through a high-dimensional non-linear solution and parallel computing strategy;

[0016] S6. Adopt a technology based on dynamic visualization and multimodal interaction to construct a real-time design scheme presentation and interaction interface, and realize the dynamic update and local optimization of the design scheme through multi-dimensional data mapping and parameterized regulation means;

[0017] S7. Establish a data-driven continuous learning and evolution mechanism, and continuously optimize the user demand prediction model and design optimization algorithm by integrating new user interaction data and historical design optimization information.

[0018] Optionally, the S1 specifically includes:

[0019] S11. Based on multimodal perception devices, collect user operation behavior data, and generate a dynamic data set in combination with interaction context information, including the operation trajectory T op ={t i |=1,2,…,n}, biometric signals B sig ={b j| = 1, 2, …, m} and environmental parameter E ctx = {e k | = 1, 2, …, l};

[0020] S12. Construct a modal fusion algorithm to normalize and feature-align multi-modal data with a dynamic weight allocation mechanism to form a fused feature matrix C fusion = [c ijk , where c ijk represents the normalized value of the i-th trajectory dimension and the j-th biological signal at the k-th time step;

[0021] S13. Extract the dynamic feature vector F of user behavior based on an adaptive time series model time = [f m1 , f m2 , …, f mp , where f mp represents the user behavior pattern weight under the time dimension p;

[0022] S14. Construct a user preference feature clustering model and extract user preference feature P through recursive hierarchical clustering analysis user = [p a1 , p a2 , …, p aq , where p aq represents the user preference feature vector under the category q;

[0023] S15. Design a multi-source feedback perception module to collect the user's interaction feedback on the current design scheme in real time and generate a feedback matrix R feedback = [r st , where r st represents the weight value of the feedback category s at the time step t;

[0024] S16. Fuse the behavior feature vector F time , the preference feature P user and the feedback matrix R feedback to generate a dynamic feature set D through a high-dimensional feature interaction model combined = [d uvw , where d uvw represents the comprehensive weight value of the behavior dimension u, the preference dimension v, and the feedback category w;

[0025] S17. Perform dimensionality reduction on D combined based on a sparse optimization and deep feature extraction algorithm to generate an optimized user behavior feature vector V optimized = [v z1 , v z2 , …, v zn , where v zn represents the n-th key feature.

[0026] Optionally, S2 specifically includes:

[0027] S21. Construct a deep neural network model M based on adaptive parameter regulation. The model structure includes multiple layers of non-linear activation units, a recursive enhancement module, and a context feature weighting unit, and the model weights W adaptive are updated in real time through a dynamic optimization strategy m and bias B m ;

[0028] S22. Input the user behavior optimization feature vector Q behavior = [q1, q2,..., q u and the context semantic feature matrix U context = [u ij into M adaptive , where q u represents the weight of the behavior feature dimension u, and u ij is the weighted value of the context category i in the semantic dimension j;

[0029] S23. Through the feature cross-fusion module, jointly embed Q behavior and U context to generate a high-dimensional joint feature representation matrix:

[0030] Z joint = [z kl ;

[0031] z kl = g(q k , u il );

[0032] where g is a feature fusion function, representing the interaction relationship between behavior and semantic features;

[0033] S24. According to the demand hierarchical prediction unit, generate a demand vector based on Z joint through multi-channel feature extraction:

[0034] D predict = [d1, d2,..., d p ;

[0035] where d p = h(W d ·z kl + B d ), W d and B d respectively represent the prediction layer weights and biases, and h is a non-linear activation function;

[0036] S25. In the time-step recursive analysis module, based on the user historical behavior sequence matrix Hhistory = [h mn and the current context feature matrix U context , construct a time-step prediction model through a recurrent neural network to generate the dynamic demand distribution T at time step t dynamic = [t1, t2, …, t v , where t v = k(h mn , u ij , W t ), and k is a recursive relationship mapping function;

[0037] S26. Construct an optimization objective function:

[0038]

[0039] Among them, Φ represents the set of model parameters, λ i and μ i are the demand weight and the time correlation coefficient respectively, and are solved through an adaptive gradient optimization algorithm;

[0040] S27. Generate the demand prediction result matrix P output = [p xy , where p xy represents the predicted value of demand category x at time step y, and the matrix is optimized and iterated in combination with a dynamic update mechanism, and output to the topology optimization module as the basis for dynamic adjustment of design parameters.

[0041] Optionally, the S3 specifically includes:

[0042] S31. Based on the user demand prediction result matrix P output = [p ab , construct a multi-objective optimization model, define the objective function set F opt = {f1(Y), f2(Y), …, f u (Y)}, where f i (Y) represents the optimization performance of the design parameter set Y = {y1, y2, …, y v} in the target dimension i, and at the same time set the constraint condition set G constraints = {g1(Y), g2(Y), …, g w (Y)} to limit the search range of the design space;

[0043] S32. Use the non-uniform topology grid generation method to discretize the design parameter set Y into a multi-scale grid structure M topo = [m ij , where m ij represents the state variable of the i, j grid cell, and the dynamic topology variable φ ij is used to represent the grid cell mij Based on the weight distribution, construct a grid optimization function:

[0044]

[0045] where β k is the priority weight of the objective function f k , δC(φ ij ) is the grid complexity regularization term, and use a non-linear constraint function to dynamically optimize φ ij ;

[0046] S33. A design sensitivity analysis module, which calculates the sensitivity of the topological state to the optimization objective based on high-order partial derivatives, and defines a sensitivity matrix S topo = [s pq , where Optimize the sensitivity matrix S through a dynamic adjustment strategy topo and enhance the global search ability of the optimization model;

[0047] S34. Use a distributed computing framework combined with a dynamic topology reconstruction algorithm to perform iterative optimization on M topo to generate an optimization result matrix R opt = [r ij , where r ij represents the final state of the grid cell after multi-objective optimization;

[0048] S35. Through the inverse mapping relationship from the grid to the design parameters, map the topological optimization result matrix R opt to the design parameter set to generate the final optimized design parameter set as the output of the personalized design scheme.

[0049] Optionally, the S4 specifically includes:

[0050] S41. Construct a deep reinforcement learning model F RL based on multi-source interaction data, which includes a state space Z state = {z1, z2,..., z m}, an action space K action = {k1, k2,..., k n}, and a reward function V reward (z, k), where the state z i represents the user's real-time interaction characteristics, and the action k j represents the optimization strategy adjusted by the system;

[0051] S42. Construct a state transition probability model based on the user's interaction behavior, and the state transition relationship is P change (z t+1 |, k t ), and through the policy network σω (k|z) Output of the optimal action in state z t to dynamically adjust the design parameters;

[0052] S43. Design a multi-layer composite reward function:

[0053] V reward (z t ,k t ) = κ1·L efficiency +κ2·L precision -κ3·R overhead -κ4·D latency ;

[0054] where L efficiency represents the optimization efficiency of the interaction operation, L precision represents the improvement in the accuracy of user feedback, R overhead represents the resource consumption cost, D latency represents the response delay, and κ1, κ2, κ3, κ4 are reward weight coefficients;

[0055] S44. Through a dual-policy network structure based on deep reinforcement learning, define the evaluation network Q eval (z,k) and the target network Q target (z,k), and adopt a recursive temporal difference to update the policy parameter ω:

[0056]

[0057] where ν is the learning rate, ρ is the discount factor, and k ′ is the optimal action in the next state;

[0058] S45. Apply the optimized policy network to the dynamic generation rule of the design scheme, and combine the state set Z state updated in real time and the action set K action to adjust the design parameter space and output a personalized design scheme that meets the real-time needs of users.

[0059] Optionally, the S5 specifically includes:

[0060] S51. Construct a design space exploration model T based on multi-objective non-linear expression explore , define the design parameter set Ω params = {ω1,…,ω m}、the objective function set Λ objectives = {λ1(Ω),…,λ p (Ω)} and the constraint set Γ constraints = {γ1(Ω),…,γ q ​(Ω)}, where λ i (Ω) represents the optimization objective of the parameter set in dimension i;

[0061] S52. Construct a high-dimensional non-linear target prediction model N based on a deep neural network optimize , with the input being the design parameter set Ω params , and the output being the set of predicted target values By optimizing the network weight set Θ optimize Minimize the prediction loss function:

[0062]

[0063] where λ i (Ω) represents the true target value;

[0064] S53. Adopt a multi-objective optimization algorithm based on the dynamic Pareto front, and define the Pareto boundary solution set Δ Pareto ={δ1, δ2, …, δ r}, where δ k represents the non-dominated solution:

[0065]

[0066] where O multi_goal is the multi-objective optimization function, ζ i represents the objective weight, and optimize the solution set according to the dynamic update strategy;

[0067] S54. Adopt a distributed parallel optimization strategy to partition the design parameter set Ω params into computational subsets P k ={ω k1 , ω k2 , …, ω kn}, and calculate the objective function values and Pareto boundary conditions in parallel in a multi-threaded environment, and aggregate the results to generate an optimized solution set;

[0068] S55. Based on the global update of the Pareto solution set Δ Pareto , generate an optimized design parameter set and use it as the output of the personalized design scheme.

[0069] Optionally, the S6 specifically includes:

[0070] S61. Construct a multi-modal interaction data acquisition and fusion module Q interaction , and collect the set of user operation behaviors U action ={u1, u2, …, u s}, the real-time feedback set V feedback ={v1, v2, …, vt} and the set of environmental parameters W context = {w1, w2, …, w r}, and fuse them to generate the dynamic data stream Q dynamic = {U action , V feedback , W context};

[0071] S62. Design a real-time visualization engine R based on dynamic multi-dimensional data mapping visualize , and define the set of mapping functions Ψ map = {ψ1(Q), ψ2(Q), …, ψ h (Q)}, where ψ k (Q) represents the dynamic mapping relationship from the interactive data Q dynamic to the visualization dimension R view , and generate the set of multi-dimensional visualization schemes R design = {ρ1, ρ2, …, ρ h};

[0072] S63. Through the parameterized regulation module P adaptive , dynamically adjust the set of design parameters Θ parameters = {θ1, θ2, …, θ z}, and construct the local optimization objective function:

[0073]

[0074] where τ j is the optimization weight coefficient, and H(θ j , Q dynamic ) represents the non-linear relationship between the parameter θ j and the dynamic data Q dynamic ;

[0075] S64. Utilize the set of local optimization requirements K of the user local = {k1, k2, …, k l}, establish the mapping matrix Φ adjus t = {φ ij | = G(k i , θ j )}, where υ ij represents the influence value of the requirement k i on the parameter θ j , G is the dynamic mapping function, and locally update the design parameters through the matrix Φ adjust to generate the optimized set of parameters

[0076] S65. Through the real-time visualization engine R visualize display the optimized design scheme Θoptimal Presented to the user interface I interactive , provide a dynamic adjustment interface to support users to perform secondary personalized optimization based on real-time input of local parameters.

[0077] Optionally, the S7 specifically includes:

[0078] S71. Construct a continuous learning module L based on data fusion fusion , define the user real-time interaction data set Ξ realtime = {ξ1, ξ2, …, ξ a} and the historical optimization record set Θ history = {θ1, θ2, …, θ b}, fuse and generate the learning data set Π training = {Ξ realtime , Θ history} for dynamic model update;

[0079] S72. Design a user demand prediction model U based on a deep recurrent network predict , with the input being the learning data set Π training , and the output being the user demand prediction result set Λ output = {λ1, λ2, …, λ c}, and minimize the prediction error function by optimizing the model parameter set Γ weights :

[0080]

[0081] where λ i is the actual demand value, is the model prediction value;

[0082] S73. Adopt a dynamic incremental learning mechanism to continuously expand the data set Π training , and adjust the demand prediction model parameter set Γ weights based on the multi-step gradient descent method:

[0083]

[0084] where η is the learning rate and t is the update step;

[0085] S74. Construct a design optimization evolution module G based on the genetic algorithm optimize , define the optimization parameter set Δ parameters = {δ1, δ2, …, δ d} and the objective function set Φ targets = {υ1, υ2, …, υ e}, and the optimization objective function is:

[0086]

[0087] Among them, ζ j is the target weight, and T(δ j , Π training ) represents the fitness of the optimization parameters for the comprehensive dataset;

[0088] S75. Based on the optimized user demand prediction model and the set of design optimization parameters dynamically adjust the personalized design generation rules, and apply the optimization results to the design scheme output module in real time to achieve adaptive evolution and continuous optimization.

[0089] The beneficial effects of the present invention are as follows:

[0090] (1) By combining topology optimization, deep learning, adaptive neural networks, and multi-modal interaction technologies, the present invention can capture user behavior and demand changes in real time, dynamically transform the demands into multi-objective optimization problems, and generate optimal design parameters through the topology reconstruction algorithm, solving the deficiencies of traditional methods in terms of dynamic adaptability and personalized satisfaction, and effectively improving the accuracy and efficiency of design optimization.

[0091] (2) Through the intelligent feedback mechanism and reinforcement learning algorithm, the present invention dynamically adjusts the design scheme generation rules based on real-time user interaction data, making the design process more intelligent and user-friendly. At the same time, it supports multi-modal interaction and dynamic local optimization, overcoming the problem of the lack of an intelligent interaction mechanism in traditional systems, and significantly improving the satisfaction and participation of the user experience.

[0092] (3) Using the design evolution module based on the genetic algorithm and the continuous learning mechanism, the present invention integrates new user data and historical design optimization records, continuously optimizes the user demand prediction model and the design optimization algorithm, has high self-adaptability and evolution ability, not only reduces the dependence on manual intervention, but also enables the system to continuously learn and adapt to complex and changing personalized demands, realizing the long-term optimization and intelligent upgrade of the design system. Description of the Drawings

[0093] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0094] Figure 1 is the overall framework diagram of a personalized design intelligent interaction method based on topology optimization proposed by the present invention. Detailed Embodiments

[0095] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only schematically showing the basic structure of the present invention, and therefore only showing the components related to the present invention.

[0096] Reference Figure 1 , a personalized design intelligent interaction method based on topological optimization, characterized by comprising the following steps:

[0097] S1. Construct a user behavior monitoring system based on multi-modal data acquisition, and in real time obtain the operation trajectory, preference characteristics, and multi-source feedback information of the user during the interaction process, and generate a high-dimensional data set containing time series characteristics;

[0098] In this embodiment, S1 specifically includes:

[0099] S11. Based on multi-modal perception devices, collect user operation behavior data, and combine the interaction context information to generate a dynamic data set, including the operation trajectory T op ={t i |=1, 2, …, n}, the biometric signal B sig ={b j |=1, 2, …, m} and the environmental parameter E ctx ={e k |=1, 2, …, l};

[0100] S12. Construct a modal fusion algorithm to normalize and feature-align multi-modal data with a dynamic weight allocation mechanism, and form a fusion feature matrix C fusion =[c ijk , where c ijk represents the normalized value of the i-th trajectory dimension and the j-th biological signal at the k-th time step;

[0101] S13. Based on an adaptive time series model, extract the dynamic feature vector F time =[f m1 , f m2 , …, f mp , where f mp represents the user behavior pattern weight in the time dimension p;

[0102] S14. Construct a user preference feature clustering model, and extract the user preference feature P user =[p a1 , p a2 , …, p aq through recursive hierarchical clustering analysis, where p aq represents the user preference feature vector under the category q;

[0103] S15. Design a multi-source feedback perception module to collect the interaction feedback of users on the current design scheme in real time and generate a feedback matrix R feedback = [r st , where r st represents the weight value of feedback category s at time step t;

[0104] S16. Integrate the behavior feature vector F time , preference feature P user and feedback matrix R feedback to generate a dynamic feature set D combined = [d uvw , where d uvw represents the comprehensive weight value of behavior dimension u, preference dimension v and feedback category w;

[0105] S17. Perform dimensionality reduction on D combined based on the sparse optimization and deep feature extraction algorithm to generate an optimized user behavior feature vector V optimized = [v z1 , v z2 , …, v zn , where v zn represents the nth key feature.

[0106] S2. Based on the adaptive deep neural network model, dynamically predict and hierarchically model the personalized needs of users by integrating user behavior features and context semantic information;

[0107] In this embodiment, S2 specifically includes:

[0108] S21. Construct a deep neural network model M adaptive with adaptive parameter regulation. The model structure includes multiple layers of non-linear activation units, a recursive enhancement module and a context feature weighting unit, and dynamically updates the model weights W m and biases B m in real time through a dynamic optimization strategy;

[0109] S22. Input the optimized user behavior feature vector Q behavior = [q1, q2, …, q u and the context semantic feature matrix U context = [u ij into M adaptive , where q u represents the weight of behavior feature dimension u, and u ij is the weighted value of context category i on semantic dimension j;

[0110] S23. Through the feature cross-fusion module, fuse Q behavior and U contextPerform joint embedding to generate a high-dimensional joint feature representation matrix:

[0111] Z joint =[z kl ;

[0112] z kl =g(q k ,u il );

[0113] Among them, g is a feature fusion function, representing the interaction relationship between behavior and semantic features;

[0114] S24. According to the demand hierarchical prediction unit, based on Z joint Generate a demand vector through multi-channel feature extraction:

[0115] D predict =[d1,d2,…,d p ;

[0116] Among them, d p =h(W d ·z kl +B d ), W d and B d represent the prediction layer weights and biases respectively, and h is a non-linear activation function;

[0117] S25. In the time step recursive analysis module, based on the user historical behavior sequence matrix H history =[h mn and the current context feature matrix U context , construct a time step prediction model through a recursive neural network to generate the dynamic demand distribution T dynamic =[t1,t2,…,t v , where t v =k(h mn ,u ij ,W t ), and k is a recursive relationship mapping function;

[0118] S26. Construct an optimization objective function:

[0119]

[0120] Among them, Φ represents the set of model parameters, λ i and μ i are the demand weight and the time correlation coefficient respectively, and are solved through an adaptive gradient optimization algorithm;

[0121] S27. Generate a demand prediction result matrix P output =[p xy , where pxy Represents the predicted value of demand category x at time step y, optimizes and iterates the matrix in combination with the dynamic update mechanism, and outputs it to the topology optimization module as the basis for dynamically adjusting the design parameters.

[0122] S3. Adopt a personalized design adjustment module driven by topology optimization, transform the user demand prediction result into a multi-objective optimization problem, and generate the optimal solution of the design parameters in real time through the dynamic topology reconstruction algorithm;

[0123] In this embodiment, S3 specifically includes:

[0124] S31. Based on the user demand prediction result matrix P output =[p ab , construct a multi-objective optimization model, define the objective function set F opt ={f1(Y), f2(Y), …, f u (Y)}, where f i (Y) represents the optimization performance of the design parameter set Y = {y1, y2, …, y v} in the target dimension i, and at the same time set the constraint condition set G constraints ={g1(Y), g2(Y), …, g w (Y)}, restricting the search range of the design space;

[0125] S32. Use the non-uniform topology grid generation method to discretize the design parameter set Y into a multi-scale grid structure M topo =[m ij , where m ij represents the state variable of the i,j grid cell, and use the dynamic topology variable φ ij to represent the weight distribution of the grid cell m ij , and construct a grid optimization function:

[0126]

[0127] Among them, β k is the priority weight of the objective function f k , δC(υ ij ) is the grid complexity regularization term, and use the non-linear constraint function to dynamically optimize υ ij ;

[0128] S33. Design sensitivity analysis module, based on the high-order partial derivative, calculate the sensitivity of the topology state to the optimization objective, and define the sensitivity matrix S topo =[s pq , where Optimize the sensitivity matrix S through the dynamic adjustment strategy topo and enhance the global search ability of the optimization model;

[0129] S34. Use a distributed computing framework combined with a dynamic topology reconstruction algorithm to perform iterative optimization on M topo to generate an optimized result matrix R opt =[r ij , where r ij represents the final state of the grid cell after multi-objective optimization;

[0130] S35. Through the inverse mapping relationship from the grid to the design parameters, map the topology optimization result matrix R opt to the set of design parameters to generate the final optimized design parameter set as the output of the personalized design scheme.

[0131] S4. Through a reinforcement learning mechanism, design an intelligent feedback algorithm based on the user's real-time operations and interaction behaviors, and iteratively update the design scheme generation rules;

[0132] In this embodiment, S4 specifically includes:

[0133] S41. Construct a deep reinforcement learning model F RL based on multi-source interaction data, including a state space Z state ={z1,z2,…,z m}, an action space K action ={k1,k2,…,k n}, and a reward function V reward (z,k), where the state z i represents the user's real-time interaction characteristics, and the action k j represents the optimization strategy adjusted by the system;

[0134] S42. Construct a state transition probability model based on the user's interaction behavior, and the state transition relationship is P change (z t+1 |,k t ), and output the optimal action ω (k|z) under the state z t to dynamically adjust the design parameters;

[0135] S43. Design a multi-layer composite reward function:

[0136] V reward (z t ,k t )=κ1·L efficiency +κ2·L precision -κ3·R overhead -κ4·D latency ;

[0137] Among them, L efficiency ​Indicates the optimized efficiency of the interactive operation, L precision Indicates the improved accuracy of user feedback, R overhead Indicates the resource consumption cost, D latency Indicates the response delay, κ1, κ2, κ3, κ4 are reward weight coefficients;

[0138] S44. Through the dual-policy network structure based on deep reinforcement learning, respectively define the evaluation network Q eval (z,k) and the target network Q target (z,k), and adopt the recursive temporal difference to update the policy parameter ω:

[0139]

[0140] where ν is the learning rate, ρ is the discount factor, and k ′ is the optimal action in the next state;

[0141] S45. Apply the optimized policy network to the dynamic generation rules of the design scheme, and combine the state set Z state updated in real time and the action set K action to adjust the design parameter space and output a personalized design scheme that meets the real-time needs of users.

[0142] S5. Construct a multi-objective design space exploration and optimization framework supported by a neural network, and quickly screen and optimize design parameters through high-dimensional nonlinear solution and parallel computing strategies;

[0143] In this embodiment, S5 specifically includes:

[0144] S51. Construct a design space exploration model T explore based on multi-objective nonlinear expressions, and define the design parameter set Ω params ={ω1,…,ω m}, the objective function set Λ objectives ={λ1(Ω),…,λ p (Ω)}, and the constraint set Γ constraints ={γ1(Ω),…,γ q (Ω)}, where λ i (Ω) represents the optimization objective of the parameter set in dimension i;

[0145] S52. Construct a high-dimensional nonlinear target prediction model N optimize based on a deep neural network. The input is the design parameter set Ω params , and the output is the predicted target value set . Minimize the prediction loss function by optimizing the network weight set Θ optimize :

[0146]

[0147] Among them, λ i (Ω) represents the true target value;

[0148] S53. Adopt a multi-objective optimization algorithm based on the dynamic Pareto front, and define the Pareto boundary solution set Δ Pareto ={δ1, δ2, …, δ r}), where δ k represents the non-dominated solution:

[0149]

[0150] Among them, O multi_goal is the multi-objective optimization function, ζ i represents the objective weight, and the solution set is optimized according to the dynamic update strategy;

[0151] S54. Adopt a distributed parallel optimization strategy to partition the design parameter set Ω params into computational subsets P k ={ω k1 , ω k2 , …, ω kn}), and calculate the objective function values and Pareto boundary conditions in parallel in a multi-threaded environment, and aggregate the results to generate an optimized solution set;

[0152] S55. Based on the global update of the Pareto solution set Δ Pareto , generate an optimized design parameter set and use it as the output of the personalized design scheme.

[0153] S6. Adopt a technology based on dynamic visualization and multi-modal interaction to construct a real-time design scheme presentation and interaction interface, and realize the dynamic update and local optimization of the design scheme through multi-dimensional data mapping and parameterized regulation means;

[0154] In this embodiment, S6 specifically includes:

[0155] S61. Construct a multi-modal interaction data acquisition and fusion module Q interaction , and collect the user operation behavior set U action ={u1, u2, …, u s}), the real-time feedback set V feedback ={v1, v2, …, v t}), and the environmental parameter set W context ={w1, w2, …, w r}), and fuse them to generate a dynamic data stream Q dynamic ={U action , Vfeedback ,W context};

[0156] S62. Design a real-time visualization engine R based on dynamic multi-dimensional data mapping visualize , define a set of mapping functions Ψ map = {ψ1(Q), ψ2(Q), …, ψ h (Q)}, where ψ k (Q) represents the dynamic mapping relationship from the interaction data Q dynamic to the visualization dimension R view , and generate a set of multi-dimensional visualization schemes R design = {ρ1, ρ2, …, ρ h};

[0157] S63. Through the parameterized regulation module P adaptive , dynamically adjust the set of design parameters Θ parameters = {θ1, θ2, …, θ z}, and construct a local optimization objective function:

[0158]

[0159] where τ j is the optimization weight coefficient, and H(θ j , Q dynamic ) represents the non-linear relationship between the parameter θ j and the dynamic data Q dynamic ;

[0160] S64. Utilize the set of local optimization requirements K local = {k1, k2, …, k l} of the user to establish a mapping matrix Φ adjust = {υ ij | = G(k i , θ j )}, where φ ij represents the influence value of the requirement k i on the parameter θ j , G is a dynamic mapping function, and the design parameters are locally updated through the matrix Φ adjust to generate an optimized set of parameters

[0161] S65. Through the real-time visualization engine R visualize present the optimized design scheme Θ optimal to the user interface I interactive , provide a dynamic adjustment interface, and support the user to perform secondary personalized optimization based on the real-time input of local parameters.

[0162] S7. Establish a data-driven continuous learning and evolution mechanism. By integrating new user interaction data and historical design optimization information, continuously optimize the user demand prediction model and design optimization algorithm.

[0163] In this embodiment, S7 specifically includes:

[0164] S71. Construct a continuous learning module L based on data fusion fusion , define the set Ξ of real-time user interaction data realtime = {ξ1, ξ2, …, ξ a} and the set Θ of historical optimization records history = {θ1, θ2, …, θ b}, and fuse them to generate the learning data set Π training = {Ξ realtime , Θ history} for dynamic model update;

[0165] S72. Design a user demand prediction model U based on a deep recurrent network predict , with the input being the learning data set Π training , and the output being the set Λ of user demand prediction results output = {λ1, λ2, …, λ c}. By optimizing the set Γ of model parameters weights , minimize the prediction error function:

[0166]

[0167] where λ i is the actual demand value, is the model prediction value;

[0168] S73. Adopt a dynamic incremental learning mechanism to continuously expand the data set Π training , and adjust the set Γ of demand prediction model parameters based on the multi-step gradient descent method weights :

[0169]

[0170] where η is the learning rate and t is the number of update steps;

[0171] S74. Construct a design optimization evolution module G based on the genetic algorithm optimize , define the set Δ of optimization parameters parameters = {δ1, δ2, …, δ d} and the set Φ of objective functions targets = {υ1, υ2, …, υ e}, and the optimization objective function is:

[0172]

[0173] Among them, ζ j is the target weight, and T(δ j , Π training ) represents the fitness of the optimization parameters for the comprehensive data set;

[0174] S75. Based on the optimized user demand prediction model and the set of design optimization parameters dynamically adjust the personalized design generation rules, and apply the optimization results to the design scheme output module in real time to achieve adaptive evolution and continuous optimization.

[0175] Example 1:

[0176] To verify the feasibility of the present invention, it is applied to a well-known domestic smart home customization company, which is mainly engaged in the design and production of personalized customized products. Due to the complexity and diversity of customer needs, traditional design methods have great limitations in terms of dynamic response ability, design efficiency, and user satisfaction, and it is difficult to meet market demands. Therefore, the company introduces the intelligent interaction method for personalized design based on topology optimization proposed by the present invention to optimize its design process, improve customer satisfaction, and reduce design costs.

[0177] In an actual application scenario, the company tested the customized design task of a smart wardrobe. The customer hopes that the wardrobe has highly personalized functions, including dynamically adjusting the size according to the room layout, providing modular storage design, and integrating an intelligent lighting system. In the traditional method, designers need to communicate with customers multiple times and complete the task by manually drawing sketches and adjusting the design model. This process usually takes more than two weeks, and the accuracy of the design scheme depends on the designer's subjective understanding of customer needs and is prone to errors.

[0178] Through the implementation of the present invention, the company established a user interaction system based on multi-modal data collection, which can capture the operation behaviors of customers in the interface in real time, including click actions, voice inputs, and text descriptions, etc. These data are integrated into the deep neural network model for demand modeling to generate hierarchical personalized demand prediction results. Subsequently, based on the topology optimization algorithm, the system transforms the demands into multi-objective optimization problems and generates design parameters that meet the demands through topology reconstruction. At the same time, the intelligent feedback mechanism can dynamically capture the adjustment behaviors of users, such as adjusting the module size or adding new functions, and update the design rules in real time through the reinforcement learning model.

[0179] In actual operation, the customer first selects the initial style and color of the wardrobe through the system interface and inputs the dimensions and layout plan of their room. The system generates a preliminary design plan within 10 seconds, which includes a recommended modular storage layout and intelligent lighting configuration. The customer then adjusts the internal structure of the wardrobe through the interface, such as increasing the number of drawers and changing the opening and closing method of the doors. The system captures these adjustments in real time and updates the design plan through a topology optimization algorithm. Finally, after two rounds of adjustments, a final plan that meets the customer's needs is generated. The following are the data results collected and optimized by the system in this case, as shown in Table 1.

[0180] Table 1 Comparison Table of Design Efficiency and Effectiveness of Smart Home Customization Company

[0181]

[0182] As can be seen from Table 1, the traditional design method takes 15 days to complete a design plan, while the present invention shortens the design cycle to 3 days. At the same time, the user satisfaction score has increased from 6.8 points in the traditional method to 9.4 points; the number of design adjustments has decreased from an average of 7 times to 2 times, and the accuracy rate of the design plan has increased from 70% to 95%. In addition, the response time of the system has been shortened from the traditional 2 minutes to 8 seconds, significantly improving the interaction experience.

[0183] In the 3-month application test, the company completed more than 200 personalized customization design tasks, covering multiple scenarios such as smart wardrobes, bookcases, and kitchen storage systems. Through user survey feedback, more than 93% of customers expressed high satisfaction with the design plans generated by this system. At the same time, the company saved nearly 30% of the design cost due to the shortened design cycle and improved plan accuracy rate.

[0184] The present invention provides an intelligent, efficient, and dynamically adaptable solution for the field of personalized design through the combination of topology optimization, multi-modal interaction, deep learning, and continuous learning technologies, significantly improving design efficiency and user satisfaction, and creating significant economic benefits and market competitive advantages for enterprises.

[0185] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A personalized design intelligent interaction method based on topology optimization, characterized in that: The steps include: S1. Build a user behavior monitoring system based on multimodal data collection to obtain the user's operation trajectory, preference characteristics and multi-source feedback information in real time during the interaction process, and generate a high-dimensional data set containing time series characteristics; S2, based on the adaptive deep neural network model, dynamically predicts and hierarchically models user personalized needs by integrating user behavior characteristics and contextual semantic information; S3, using the personalized design adjustment module driven by topology optimization, converting the user demand prediction results into multi-objective optimization problems, and generating the optimal solution of design parameters in real time through the dynamic topology reconstruction algorithm; S4. Through the reinforcement learning mechanism, an intelligent feedback algorithm is designed based on the user's real-time operation and interactive behavior, and the design solution generation rules are iteratively updated; S5. Build a multi-objective design space exploration and optimization framework supported by neural networks, and quickly screen and optimize design parameters through high-dimensional nonlinear solutions and parallel computing strategies; S6. Adopt dynamic visualization and multimodal interaction technology to build a real-time design scheme presentation and interactive interface, and realize dynamic update and local optimization of the design scheme through multi-dimensional data mapping and parametric control methods; S7. Establish a data-driven continuous learning and evolution mechanism to continuously optimize the user demand prediction model and design optimization algorithm by integrating new user interaction data and historical design optimization information.

2. According to claim 1, a personalized design intelligent interaction method based on topology optimization is characterized in that: The S1 specifically includes: S11. Based on multimodal sensing devices, collect user operation behavior data and generate dynamic data sets based on interactive context information, including operation trajectory T op ={t i |=1,2,…,n}, biometric signal B sig = {b j |=1,2,…,m} and environmental parameters E ctx ={e k |=1,2,…,l}; S12. Construct a modal fusion algorithm to normalize and align multimodal data using a dynamic weight allocation mechanism to form a fusion feature matrix C fusion =[c ijk ], where c ijk represents the normalized value of the i-th trajectory dimension and the j-th biological signal at the k-th time step; S13. Extract dynamic feature vectors of user behavior based on adaptive time series model in Represents the weight of user behavior pattern under time dimension p; S14, construct a user preference feature clustering model, and extract user preference features P through recursive hierarchical clustering analysis user =[p a1 ,p a2 ,…,p aq ], where p aq Represents the user's preference feature vector under category q; S15. Design a multi-source feedback perception module to collect user interactive feedback on the current design solution in real time and generate a feedback matrix R feedback =[r st ], where r st Represents the weight value of feedback category s at time step t; S16, the behavior feature vector F time , preference feature P user and the feedback matrix R feedback Fusion, generating dynamic feature set D through high-dimensional feature interaction model combined =[d uvw ], where d uvw Represents the comprehensive weight value of behavior dimension u, preference dimension v and feedback category w; S17, based on sparse optimization and deep feature extraction algorithm combined Perform dimensionality reduction processing to generate user behavior optimization feature vector V optimized =[v z1 ,v z2 ,…,v zn ], where v zn Represents the nth key feature.

3. According to claim 1, a personalized design intelligent interaction method based on topology optimization is characterized in that: The S2 specifically includes: S21. Constructing a deep neural network model based on adaptive parameter control M adaptive The model structure includes multi-layer nonlinear activation units, recursive enhancement modules and context feature weighting units, and the model weights W are updated in real time through dynamic optimization strategies. m and bias B m ; S22. Optimize the user behavior feature vector Q behavior =[q1,q2,…,q u ] and contextual semantic feature matrix U context =[u ij ] Input to M adaptive , where q u represents the weight of the behavioral feature dimension u, u ij is the weighted value of context category i on semantic dimension j; S23, through the feature cross fusion module, Q behavior and U context Perform joint embedding to generate a high-dimensional joint feature representation matrix: WITH joint =[of kl ]; z kl =g(q k ,u il ); Among them, g is the feature fusion function, which represents the interactive relationship between behavior and semantic features; S24, according to the demand layer forecast unit, based on Z joint Generate demand vector through multi-channel feature extraction: D predict =[d1,d2,…,d p ]; Among them, d p =h(W d ·z kl +B d ), W d and B d They represent the prediction layer weight and bias respectively, and h is the nonlinear activation function; S25, in the time step recursive analysis module, based on the user historical behavior sequence matrix H history =[h mn ] and the current context feature matrix U context , a time step prediction model is constructed through a recurrent neural network to generate the dynamic demand distribution T at time step t dynamic =[t1,t2,…,t v ], where t v = k(h mn ,u ij ,W t ), k is a recursive relation mapping function; S26. Construct optimization objective function: Among them, Φ represents the model parameter set, λ i and μ i are the demand weight and time correlation coefficient, respectively, which are solved by adaptive gradient optimization algorithm; S27, generate demand forecast result matrix P output =[p xy ], where p xy It represents the predicted value of demand category x at time step y. The matrix is ​​optimized and iterated in combination with the dynamic update mechanism, and is output to the topology optimization module as the basis for dynamic adjustment of design parameters.

4. According to claim 1, a personalized design intelligent interaction method based on topology optimization is characterized in that: The S3 specifically includes: S31, based on the user demand prediction result matrix P output =[p ab ], build a multi-objective optimization model, define the objective function set F opt ={f1(Y),f2(Y),…,f u (Y)}, where f i (Y) represents the design parameter set Y = {y1, y2, ..., y v }Optimize performance under target dimension i and set constraint set G constraints ={g1(Y),g2(Y),…,g w (Y)}, limiting the search scope of the design space; S32, using the non-uniform topology grid generation method, discretize the design parameter set Y into a multi-scale grid structure M topo =[m ij ], where m ij Represents the state variable of the i,j grid unit, using the dynamic topological variable φ ij Represents the grid unit m ij The weight distribution of , construct the grid optimization function: Among them, β k is the objective function f k The priority weight, δC(φ ij ) is the grid complexity regularization term, and a nonlinear constraint function is used to dynamically optimize φ ij ; S33. Design a sensitivity analysis module to calculate the sensitivity of the topological state to the optimization target based on high-order partial derivatives and define the sensitivity matrix S topo =[s pq ],in Optimizing the sensitivity matrix S by dynamically adjusting the strategy topo And enhance the global search capability of the optimization model; S34, using a distributed computing framework combined with a dynamic topology reconstruction algorithm to topo Perform iterative optimization to generate the optimization result matrix R opt =[r ij ], where r ij Represents the final state of the grid unit after multi-objective optimization; S35, through the inverse mapping relationship from mesh to design parameters, the topology optimization result matrix R opt Map to the design parameter set to generate the final optimized design parameter set As the output of personalized design solutions.

5. The personalized design intelligent interaction method based on topology optimization according to claim 1, characterized in that: The S4 specifically includes: S41. Constructing a deep reinforcement learning model based on multi-source interaction data F RL , containing the state space Z state ={z1,z2,…,z m }, action space K action ={k1,k2,…,k n } and reward function V reward (z,k), where state z i represents the user's real-time interaction features, action k j Indicates the optimization strategy for system adjustment; S42. Construct a state transition probability model based on user interaction behavior. The state transition relationship is P change (z t+1 |,k t ), and through the policy network σ ω (k|z) output in state z t The best action under To dynamically adjust design parameters; S43. Design a multi-layer composite reward function: V reward (z t ,k t )=κ1·L efficiency +κ2·L precision -κ3·R overhead -k4·D latency ; Among them, L efficiency represents the optimization efficiency of the interactive operation, L precision Indicates the improvement in the accuracy of user feedback, R overhead represents the resource consumption cost, D latency represents the response delay, κ1, κ2, κ3, κ4 are the reward weight coefficients; S44, through the dual-strategy network structure based on deep reinforcement learning, define the evaluation network Q eval (z,k) and the target network Q target (z,k), and use recursive time difference to update the strategy parameter ω: Among them, ν is the learning rate, ρ is the discount factor, and k ′ is the optimal action in the next state; S45. Optimize the strategy network Dynamic generation rules applied to design solutions, combined with real-time updated state set Z state and the action set K action Adjust the design parameter space and output a personalized design solution that meets the user's real-time needs.

6. The personalized design intelligent interaction method based on topology optimization according to claim 1 is characterized in that: The S5 specifically includes: S51. Constructing a design space exploration model based on multi-objective nonlinear expression T explore , define the design parameter set Ω params ={ω1,…,ω m }、Objective function set Λ objectives ={λ1(Ω),…,λ p (Ω)} and the constraint set Γ constraints ={γ1(Ω),…,γ q (Ω)}, where λ i (Ω) represents the optimization target of the parameter set in dimension i; S52. Constructing a high-dimensional nonlinear target prediction model based on deep neural network N optimize , the input is the design parameter set Ω params , the output is the predicted target value set By optimizing the network weight set Θ optimize Minimize the prediction loss function: Among them, λ i (Ω) represents the true target value; S53, using a multi-objective optimization algorithm based on dynamic Pareto frontier, define the Pareto frontier solution set Δ Pareto ={δ1,δ2,…,δ r }, where δ k Represents a non-inferior solution: Among them, O multi_goal is a multi-objective optimization function, ζ i represents the target weight, and optimizes the solution set according to the dynamic update strategy; S54, using distributed parallel optimization strategy to optimize the design parameter set Ω params Partition into computing subsets P k ={ω k1 ,ω k2 ,…,ω kn }, parallelly calculate the objective function value and Pareto boundary conditions in a multi-threaded environment, and aggregate the results to generate an optimized solution set; S55, based on Pareto solution set Δ Pareto Global update of the optimized design parameter set Take it as the output of personalized design plan.

7. The personalized design intelligent interaction method based on topology optimization according to claim 1 is characterized in that: The S6 specifically includes: S61. Construct multimodal interaction data collection and fusion module Q interaction , collect user operation behavior set U action ={u1,u2,…,u s }、Real-time feedback set V feedback ={v1,v2,…,v t } and the environmental parameter set W context ={w1,w2,…,w r }, fusion generates dynamic data stream Q dynamic = {U action ,V feedback ,W context }; S62. Design a real-time visualization engine based on dynamic multidimensional data mapping visualize , define the mapping function set Ψ map ={ψ1(Q),ψ2(Q),…,ψ h (Q)}, where ψ k (Q) represents the interaction data Q dynamic To the visualization dimension R view Dynamic mapping relationship, real-time generation of multi-dimensional visualization solution set R design ={ρ1,ρ2,…,ρ h }; S63, through parameterized control module P adaptive , dynamically adjust the design parameter set Θ parameters ={θ1,θ2,…,θ z }, construct the local optimization objective function: Among them, τ j To optimize the weight coefficient, H(θ j ,Q dynamic ) represents the parameter θ j With dynamic dataQ dynamic The nonlinear relationship between S64. Using the user's local optimization demand set K local ={k1,k2,…,k l }, establish the mapping matrix Φ between requirements and parameters adjust ={φ ij |=G(k i ,θ j )}, where φ ij Indicates the demand k i For parameter θ j The influence value of G is the dynamic mapping function, through the matrix Φ adjust Locally update the design parameters to generate an optimized parameter set S65, through the real-time visualization engine R visualize The optimized design solution Θ optimal Presented to the user interface I interactive , provides a dynamic adjustment interface, and supports users to perform secondary personalized optimization based on real-time input of local parameters.

8. The personalized design intelligent interaction method based on topology optimization according to claim 1 is characterized in that: The S7 specifically includes: S71. Constructing a continuous learning module based on data fusion fusion , define the user real-time interaction data set Ξ realtime ={ξ1,ξ2,…,ξ a } and historical optimization record set Θ history ={θ1,θ2,…,θ b }, fusion generates learning dataset Π training ={Ξ realtime ,Θ history }, used for dynamic model update; S72. Design a user demand prediction model based on deep recurrent network predict , the input is the learning data set Π training , the output is the user demand prediction result set Λ output ={λ1,λ2,…,λ c }, by optimizing the model parameter set Γ weights Minimize the prediction error function: Among them, λ i is the actual demand value, is the model prediction value; S73, using dynamic incremental learning mechanism to learn data set Π training Continue to expand and adjust the demand forecasting model parameter set Γ based on the multi-step gradient descent method weights : Where η is the learning rate and t is the number of update steps; S74. Constructing a design optimization evolution module based on genetic algorithm G optimize , define the optimization parameter set Δ parameters ={δ1,δ2,…,δ d } and the objective function set Φ targets ={υ1,υ2,…,υ e }, the optimization objective function is: Among them, j is the target weight, T(δ j ,Π training ) represents the fitness of the optimized parameters to the comprehensive data set; S75, based on the optimized user demand prediction model and design optimization parameter set Dynamically adjust the personalized design generation rules and apply the optimization results to the design solution output module in real time to achieve adaptive evolution and continuous optimization.

Images