Three-dimensional industrial design material AI intelligent recommendation management system

Abstract

The invention relates to the technical field of management systems, and particularly discloses a three-dimensional industrial design material AI intelligent recommendation management system. Comprising a dynamic knowledge graph module, a multi-source perception input module, a neural symbol fusion engine, a situation self-adaptive output module, a self-evolution optimization module, a recommendation rule updating module, a cold start processing module and a multi-target decision module. The neural symbol fusion engine realizes symbol logical reasoning through Transform, the cold start processing module generates a user portrait through gradient embedding and generates virtual user features through cross-map feature propagation, the situation adaptive output module dynamically switches rendering modes according to a design stage, and the self-evolution optimization module performs meta-learning and concept drift detection; through multi-technology fusion and module cooperation, intelligence, precision and high efficiency of three-dimensional industrial design material recommendation are realized, the design efficiency is remarkably improved, and the project landing risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of management system technology, specifically to a 3D industrial design material AI intelligent recommendation management system. Background Technology

[0002] In the field of 3D industrial design, the intelligence and efficiency of material recommendation systems are crucial to design efficiency and quality. Existing technologies mainly suffer from the following shortcomings:

[0003] Insufficient material association modeling capability: Traditional recommendation systems are mostly based on static databases or simple semantic associations, which makes it difficult to capture the geometric features, functional semantics and dynamic associations of 3D materials in real time, resulting in a low degree of matching between recommendation results and actual design needs;

[0004] Limited ability to fuse and reason about multi-source data: The industrial design process involves multi-source information such as user behavior data, design environment parameters, and physical simulation feedback. However, existing systems lack effective neural symbol fusion mechanisms and cannot combine the feature extraction capabilities of neural networks with the reasoning capabilities of symbolic logic, making it difficult to generate recommendation decisions that balance accuracy and interpretability.

[0005] The cold start problem is prominent: For new users or new projects, due to the lack of historical interaction data, traditional recommendation systems have difficulty building user profiles quickly, resulting in poor initial recommendation performance, which requires a large amount of data accumulation to optimize.

[0006] Insufficient context adaptability: The design process is usually divided into stages such as conceptual design, detailed design, and simulation verification. Existing systems cannot dynamically adjust the recommended strategies and output accuracy according to the design stage, and have weak performance adaptability to different devices, which affects the user experience.

[0007] Lack of model self-optimization and risk assessment: Industrial design needs evolve dynamically with technological development and changing scenarios, but existing systems lack closed-loop optimization mechanisms, making it difficult to cope with concept drift; at the same time, the feasibility of materials, manufacturing compatibility and performance risks are not fully considered in the recommendation process, resulting in obstacles to the implementation of recommendation results in practical applications. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a 3D industrial design material AI intelligent recommendation and management system, which solves the problems mentioned in the background technology.

[0009] To achieve the above objectives, this invention provides the following technical solution: a 3D industrial design material AI intelligent recommendation management system, comprising: a dynamic knowledge graph module, a multi-source sensing input module, a neural symbol fusion engine, a context-adaptive output module, a self-evolving optimization module, and a recommendation rule update module; the dynamic knowledge graph module is used to construct and update a weighted material association graph in real time. , where nodes Represents design materials, edges Indicates the semantic relationship between materials, weight The system represents the strength of associations; the multi-source sensing input module collects user interaction data, design environment parameters, and physical simulation feedback data; the neural symbolic fusion engine generates recommendation decisions through hybrid neural networks and symbolic reasoning; the context-adaptive output module dynamically switches rendering modes according to the design stage and adaptively adjusts output accuracy based on device performance; the self-evolutionary optimization module achieves closed-loop iterative optimization of the recommendation model through meta-learning and concept drift detection; and the recommendation rule update module confirms the updated recommendation rules for the target user when the recommendation rule fit of the target user is less than the set fit.

[0010] Preferably, the construction of the material association map includes the following steps:

[0011] Node feature generation: via formula Integrating text metadata semantic features and three-dimensional geometric features Topology information, specifically including:

[0012] Extract word vectors from text metadata using the BERT model; The network extracts local geometric features from the 3D mesh; the two types of features are then concatenated and input. Generate node embedding vectors ;

[0013] Dynamic edge weight update: via formula Iteratively update edge weights, where Quantitative design functional gradient similarity specifically includes: calculating historical co-occurrence frequencies. Statistical characteristics Solving for adjacent material design functions The gradient cosine of the angle between the gradients is used; historical weights, co-occurrence features, and gradient similarity are concatenated to form a feature vector; and then the feature vector is generated through a fully connected layer. and Sigmoid activation function Output new weights .

[0014] Preferably, the neural symbol fusion engine performs the following steps:

[0015] Symbolic logical reasoning: through formulas Implement tensor matching between user queries and historical behavior rules, where Tensor contraction operations specifically include:

[0016] Parsing user queries This is a predicate logic expression; the predicate logic expression is then compared with the design rules in the knowledge base. Perform tensor matching; generate symbolic inference results through the Transformer encoding rule inference chain. ;

[0017] Neural network feature extraction: through formulas Based on the map Graph attention network extracts material features Specifically, this includes: running a multi-head attention mechanism on the material association graph; and aggregating the feature information of neighboring nodes to generate neural network features. ;

[0018] Feature fusion: via formula This involves concatenating the channels of symbolic reasoning results with neural network features, where... This indicates that the feature concatenation operation is used in the subsequent comprehensive calculation of the recommendation score.

[0019] Preferably, generating a recommendation score includes the following steps:

[0020] Calculate the degree of conformity with design intent: using the formula In parameter space Inner integration, minimizing the user-constrained gradient Gradient of material properties The differences specifically include: discretizing the design parameter space into a set of grid points. ; Calculate the gradient difference norm at each grid point The differences are mapped using a Gaussian kernel function and then summed and normalized.

[0021] Calculate cross-project novelty: using the formula use Divergence measures target domain With source domain The differences in feature distribution specifically include: aligning the feature spaces of the source and target domains using the optimal transmission algorithm; and calculating the conditional distribution of material features in the target domain. ; Calculate the empirical distribution mean of the source domain features;

[0022] Solve Divergence value; generating fusion score, using a formula Balancing precision, innovation, and manufacturability, among which Express material With manufacturing process Compatibility rating.

[0023] Preferably, the self-evolutionary optimization module performs the following steps:

[0024] Meta-learning training: through formulas Optimize model parameters, where For task-specific parameters, Controlling model complexity specifically includes: sampling multiple design tasks. Calculate the parameters after gradient update for each task. Minimize the weighted sum of task loss and parameter offset;

[0025] Concept drift detection: via formula Discriminant feature distribution changes include: training the discriminator. Distinguish current features With historical characteristics When fighting losses The model is re-reconstructed on demand; the recommendation model is retrained using the latest data.

[0026] Preferably, the physical simulation verification interface performs the following steps:

[0027] Finite element analysis preprocessing: recommending materials Import simulation environment; set boundary conditions and physical parameters;

[0028] Feasibility quantification: through formula Feasibility of calculation simulation, among which For simulation response, As per specifications, this specifically includes: performing finite element analysis to obtain stress and strain distribution. ; Calculate the normalized difference norm with respect to specifications; through function Mapped to feasibility score;

[0029] Dynamic filtering: when Remove the material from the recommended list.

[0030] Preferably, the 3D industrial design material AI intelligent recommendation management system also includes a cold start processing module, which is used to generate user profiles through gradient embedding and generate virtual user features through cross-graph feature propagation in the case of new users or insufficient data, so as to provide an initial basis for subsequent recommendations. The user features generated are the basis for multi-objective decision-making to optimize recommendations in the cold start stage.

[0031] The startup processing module executes the following steps:

[0032] Gradient embedding generation: via formula The process of constructing user profiles by aggregating the gradient directions of historical design decisions includes: extracting basic attribute features of new users; and retrieving historical design sequences of similar users. ; Calculate the mean gradient of each design function;

[0033] Cross-spectral feature propagation: via formula Generating virtual user characteristics specifically includes: locating neighbor nodes on a cross-project graph. Aggregate neighbor features using a cross-domain graph neural network; generate virtual user features using the READOUT function. .

[0034] Preferably, the context-adaptive output module performs the following steps:

[0035] Design phase awareness, through piecewise functions:

[0036] ;

[0037] Dynamically adjust rendering precision, specifically including: monitoring design time. and milestone events; initial output simplified model Intermediate output parameterized model Later output of physical accuracy model ;

[0038] Device adaptive compression: via formula Dynamically adjusting the model compression ratio specifically includes: real-time acquisition of available compression ratios on the user end. Memory Based on network latency Dynamically adjust the compression ratio; start chunked transmission when the compression ratio is <0.8.

[0039] Preferably, the 3D industrial design material AI intelligent recommendation management system also includes a multi-objective decision-making module, which is used to comprehensively consider multiple objectives such as risk and similarity during the recommendation process, and optimize the recommendation results through risk value calculation and dynamic weight adjustment, so that the recommendation is more in line with actual needs;

[0040] The multi-objective decision-making module performs the following steps:

[0041] Value at Risk (VaR) calculation: using the formula Quantifying recommendation risks, among which The historical performance variation at risk value specifically includes: statistical data on historical performance fluctuations. ; Calculate confidence level Value at risk; through The function maps to a risk score;

[0042] Dynamic weight adjustment: via formula Adaptive adjustment of similarity weights, specifically including: real-time scoring of the cumulative risk value of used materials; and adjustment based on a decay coefficient. Reduce similarity weight; constrain weight range .

[0043] Preferably, the 3D industrial design material AI intelligent recommendation management system also includes a cross-project knowledge distillation module, which is used to compress the capabilities of expert knowledge and complex models into a lightweight model through teacher model training and student model distillation, and deploy it to edge devices to improve the system's operating efficiency and response speed, while ensuring the accuracy of recommendations;

[0044] The cross-project knowledge distillation module performs the following steps:

[0045] Teacher model training: through formulas Optimizing the teacher model specifically includes: training the teacher model using expert-annotated data; and minimizing the recommendation score. With expert rating The mean square error;

[0046] Student model distillation: via formula The compressed model specifically includes: extracting the probability distribution of the teacher model output. Minimize and student model distribution of Divergence; Add sparsity regularization term Compress model parameters; deploy lightweight student models to edge devices.

[0047] This invention provides an AI-powered intelligent recommendation and management system for 3D industrial design materials, which has the following beneficial effects:

[0048] 1. The dynamic knowledge graph module constructs a weighted material association graph in real time, integrating textual semantic features with three-dimensional geometric topology information (such as extracting features through BERT and PointNet and generating node embeddings through GraphSAGE), and dynamically updates edge weights based on historical co-occurrence frequency and design function gradient similarity. This can capture deep semantic relationships and functional dependencies between materials, improving the accuracy and relevance of recommendations. For example, in automotive parts design, it can accurately recommend connectors or sensor models that match the current structure.

[0049] 2. The neural symbolic fusion engine uses Transformer to achieve symbolic logic reasoning and combines graph features extracted by Graph Attention Network (GAT) to achieve dual recommendation decisions based on "semantic understanding and data-driven" approaches. For example, when a user queries "lightweight high-strength aluminum alloy bracket", the system can simultaneously match design rules (such as strength threshold and weight limit) through symbolic reasoning and analyze the geometric features of similar materials through neural networks to generate a recommendation list that combines rule constraints and data support, thereby improving the interpretability and reliability of the decision.

[0050] 3. The cold start processing module generates user profiles (aggregating design decision gradients of similar users) through gradient embedding and generates virtual user features through cross-graph feature propagation, which can provide initial recommendations for new users without a large amount of historical data. For example, when a new designer uses the system for the first time, initial recommendations can be generated based on the historical design sequences of senior designers in the same field, solving the problem of recommendation gaps in the "zero data" scenario.

[0051] 4. The context-adaptive output module dynamically switches rendering modes according to the design stage (e.g., outputting low-poly models in the concept stage and physically accurate models in the simulation stage), and adaptively adjusts the model compression rate based on the device's GPU memory and network latency. This not only optimizes the work efficiency at different design stages (e.g., rapid preview in the early stage and accurate verification in the later stage), but also ensures smooth display of complex models on low-end devices, improving cross-terminal compatibility.

[0052] 5. The self-evolutionary optimization module monitors changes in design requirements and reconstructs the model in real time through meta-learning and concept drift detection. For example, when a new material standard is released, the system can automatically update the recommended weights of relevant materials to maintain the timeliness of the model. The multi-objective decision-making module introduces risk value calculation and dynamic weight adjustment to comprehensively evaluate the design intent compliance, cross-project novelty, and manufacturing compatibility of materials (such as verifying feasibility through finite element analysis), avoiding the recommendation of high-risk or unmanufacturable materials. For example, in aerospace design, component models that have been verified by simulation and have small performance fluctuations are given priority.

[0053] 6. The cross-project knowledge distillation module, through a teacher-student model architecture, compresses complex expert models into lightweight models for deployment on edge devices. This reduces cloud computing pressure and improves response speed while maintaining recommendation accuracy. For example, field engineers can quickly obtain recommendations via mobile devices without relying on high-bandwidth networks to transmit complex model data.

[0054] In summary, this invention achieves intelligent, accurate, and efficient 3D industrial design material recommendation through the integration of multiple technologies and module collaboration, significantly improving design efficiency and reducing engineering implementation risks. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of an AI-powered intelligent recommendation and management system for 3D industrial design materials as described in this invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1 As shown, the present invention provides a technical solution: a three-dimensional industrial design material AI intelligent recommendation management system, comprising: an auxiliary support layer, a neural symbol fusion engine, a core processing layer and a data input layer;

[0058] The auxiliary support layer includes: a recommendation rule update module and a cold start processing module;

[0059] The core processing layer includes: a multi-objective decision-making module and a self-evolutionary optimization module;

[0060] The data input layer includes: a multi-source perception input module, a dynamic knowledge graph module, and a context-adaptive output module;

[0061] The dynamic knowledge graph module is used to build and update weighted material association graphs in real time. , where nodes Represents design materials, edges Indicates the semantic relationship between materials, weight The system is structured as follows: Characterizing association strength; a multi-source sensing input module collects user interaction data, design environment parameters, and physical simulation feedback data; a neural symbolic fusion engine generates recommendation decisions through hybrid neural networks and symbolic reasoning; a context-adaptive output module dynamically switches rendering modes based on the design stage and adaptively adjusts output accuracy based on device performance; a self-evolutionary optimization module achieves closed-loop iterative optimization of the recommendation model through meta-learning and concept drift detection; a recommendation rule update module confirms updated recommendation rules for the target user when the recommendation rule fit is less than a set fit; a cold start processing module generates user profiles through gradient embedding and virtual user features through cross-graph feature propagation in scenarios with new users or insufficient data, providing an initial basis for subsequent recommendations. The generated user features form the foundation for multi-objective decision-making in the cold start phase of recommendation optimization; and a multi-objective decision-making module comprehensively considers multiple objectives such as risk and similarity during the recommendation process, optimizing the recommendation results through risk value calculation and dynamic weight adjustment to make the recommendations more in line with actual needs.

[0062] More specifically, the construction of the material association map includes the following steps:

[0063] Node feature generation: via formula Integrating text metadata semantic features and three-dimensional geometric features Topology information, specifically including:

[0064] Extract word vectors from text metadata using the BERT model; The network extracts local geometric features from the 3D mesh; the two types of features are then concatenated and input. Generate node embedding vectors ;

[0065] Dynamic edge weight update: via formula Iteratively update edge weights, where Quantitative design functional gradient similarity specifically includes: calculating historical co-occurrence frequencies. Statistical characteristics Solving for adjacent material design functions The gradient cosine of the angle between the gradients is used; historical weights, co-occurrence features, and gradient similarity are concatenated to form a feature vector; and then the feature vector is generated through a fully connected layer. and Sigmoid activation function Output new weights .

[0066] Historical co-occurrence frequency statistical characteristic calculation: Calculate the frequency of co-occurrence of adjacent materials in historical use. The frequency data was then statistically analyzed to generate corresponding statistical features. These statistical characteristics can reflect the degree of correlation between adjacent materials in past design practices. For example, the mean and variance of the frequency of occurrence can provide important information about the historical collaborative use of materials for updating edge weights.

[0067] Solving for the cosine value of the gradient angle of the design function: Design function for adjacent materials and Solve for their gradients and Then, the cosine of the angle between their gradients is calculated; this cosine value It can measure the similarity of adjacent materials in the trend of design function change. The closer the cosine value is to 1, the more similar the gradient directions of the design functions of the two materials are, and their trends of design function change are relatively similar; conversely, the closer the cosine value is to -1, the greater the difference in the trends of design function change between the two. This similarity is quantified into a specific numerical value, which is used for edge weight update calculation.

[0068] Feature vector concatenation: combining historical edge weights Historical co-occurrence frequency statistical characteristics and design functional gradient similarity The features are concatenated to form a feature vector containing multiple aspects of information. This feature vector comprehensively considers the historical weight information of the edges, the historical co-occurrence relationship of the materials, and the similarity of the current design functions, providing a rich data foundation for a comprehensive assessment of the importance of the edges.

[0069] New weight output: The feature vector obtained by the above concatenation is input into the fully connected layer W and the sigmoid activation function. In the network structure, the fully connected layer W performs a linear transformation on the input feature vector, adjusting the weights of each feature dimension to highlight features that significantly influence edge weight updates; the sigmoid activation function... The output of the fully connected layer is then mapped to the (0,1) interval to obtain new edge weights. In this way, the edge weights are dynamically updated, enabling them to adapt to new material relationships as the design data changes and the design process progresses. This more accurately reflects the real-time correlation strength between materials, providing more precise graph structure support for material recommendation.

[0070] More specifically, the neural symbol fusion engine performs the following steps:

[0071] Symbolic logical reasoning: through formulas Implement tensor matching between user queries and historical behavior rules, where Tensor contraction operations specifically include:

[0072] Parsing user queries This is a predicate logic expression; the predicate logic expression is then compared with the design rules in the knowledge base. Perform tensor matching; generate symbolic inference results through the Transformer encoding rule inference chain. ;

[0073] Neural network feature extraction: through formulas Based on the map Graph attention network extracts material features Specifically, this includes: running a multi-head attention mechanism on the material association graph; and aggregating the feature information of neighboring nodes to generate neural network features. ;

[0074] Feature fusion: via formula This involves concatenating the channels of symbolic reasoning results with neural network features, where... This indicates that the feature concatenation operation is used in the subsequent comprehensive calculation of the recommendation score.

[0075] Symbolic logic reasoning through formulas Implement tensor matching between user queries and historical behavior rules, where The tensor shrinkage operation is represented, and the specific execution flow is as follows:

[0076] User query parsing: The user-input query statement Q is parsed into a predicate logic expression; for example, the user query "lightweight aluminum alloy bracket" will be broken down into logical predicates such as "material = aluminum alloy ∧ property = lightweight", forming a structured logical expression that is easy to match with the rules in the knowledge base.

[0077] Tensor matching operation: Matches the parsed predicate logic expression with the design rules in the knowledge base. Perform tensor shrinkage operations; tensor shrinkage operations can capture the semantic relationship between query logic and historical rules in high-dimensional space, achieve efficient matching between logical expressions and rule base through matrix operations, and filter out a subset of design rules related to the query.

[0078] Transformer Encoding Inference: The Transformer model is used to encode the matched rule inference chain. Transformer captures long-distance dependencies between rules through a self-attention mechanism, encoding the rule sequence into symbolic inference results. For example, for the rule that "aluminum alloy materials must meet the requirement of tensile strength ≥200MPa", Transformer will encode it into a symbol vector containing material property constraints, providing a logical constraint basis for subsequent recommendations.

[0079] Through formula Graph attention networks based on graph G extract material features M, specifically including:

[0080] Multi-head attention mechanism operation: On the material association graph G, a multi-head attention mechanism is run for each node (material); the multi-head attention mechanism captures the association weights between the node and its neighboring nodes from different dimensions through multiple parallel attention heads; for example, each attention head will focus on different feature dimensions such as the geometric features and functional labels of the material, and calculate the attention score between the current node and its neighboring nodes.

[0081] Neighbor Feature Aggregation: The feature information of neighboring nodes is weighted and aggregated based on the attention score. For each node, the features of its neighboring nodes in the graph (such as 3D geometric features, text semantic features, etc.) are aggregated to generate a neural network feature N containing local graph structure information; this feature not only includes the node's own features M, but also integrates its contextual association information in the graph, which can capture implicit associations between materials.

[0082] Through formula This involves concatenating the channels of symbolic reasoning results with neural network features, where... The feature concatenation operation has the following specific functions:

[0083] Feature dimension concatenation: combining symbolic reasoning results (e.g., logical constraint vectors) and neural network features N (e.g., spectral semantic vectors) are concatenated along the feature dimension to form fused features. For example, the symbolic constraint vector of "tensile strength ≥ 200MPa" is concatenated with the geometric feature vector of the aluminum alloy bracket to form a comprehensive vector that simultaneously contains logical rules and data features.

[0084] Recommended score calculation: fused features It participates in the calculation of subsequent recommendation scores; by combining the logical constraints of symbolic reasoning (such as design rule compliance) and the data analysis of neural networks (such as material similarity), the recommendation decision satisfies both the interpretability of design rules and the accuracy of data-driven decision-making. For example, when recommending aluminum alloy brackets, the fusion features will simultaneously consider rule constraints (material strength) and graph association (historical usage of similar structures) to generate a comprehensive recommendation score.

[0085] More specifically, generating a recommendation score involves the following steps:

[0086] Calculate the degree of conformity with design intent: using the formula In parameter space Inner integration, minimizing the user-constrained gradient Gradient of material properties The differences specifically include: discretizing the design parameter space into a set of grid points. ; Calculate the gradient difference norm at each grid point The differences are mapped using a Gaussian kernel function and then summed and normalized.

[0087] Calculate cross-project novelty: using the formula use Divergence measures target domain With source domain The differences in feature distribution specifically include: aligning the feature spaces of the source and target domains using the optimal transmission algorithm; and calculating the conditional distribution of material features in the target domain. ; Calculate the empirical distribution mean of the source domain features;

[0088] Solve Divergence value; generating fusion score, using a formula Balancing precision, innovation, and manufacturability, among which Express material With manufacturing process Compatibility rating.

[0089] Through formula In parameter space Inner integration, minimizing the user-constrained gradient Gradient of material properties The differences are explained in the following steps:

[0090] Discretization of design parameter space: Discretizing the continuous design parameter space Discretize into a finite set of grid points For example, in the design of mechanical parts, the parameter space can include dimensions such as size and material properties. After discretization, it forms regular grid points, which facilitates subsequent numerical calculations.

[0091] Gradient difference norm calculation: for each grid point Calculate the user-constrained gradient Gradient of material properties difference norm Taking lightweight automotive design as an example, This represents the user's constraint gradient on the part weight. This indicates the strength characteristic gradient of a certain aluminum alloy material; the difference between the two reflects the degree of matching between the material and the design requirements.

[0092] Gaussian kernel mapping and normalization: using Gaussian kernel functions The gradient difference values ​​of each grid point are mapped to similarity scores, where The weighting parameter is used. The scores of all grid points are summed and then normalized (divided by the normalization constant Z) to obtain the design intent compliance rate. The range of values ​​is A higher value indicates that the material better matches the user's design intent.

[0093] Through formula Using KL divergence to measure the target domain With source domain The characteristic distribution differences include:

[0094] Feature space alignment: Align the source domain using the optimal transfer algorithm (such as the Sinkhorn algorithm). (Set of historical project features) and target domain The feature space of the current project feature set ensures that features from different projects are comparable in the same semantic space. For example, new material features in the aerospace field and historical material features in the automotive field can be mapped to a common functional semantic space through optimal transfer.

[0095] Conditional distribution calculation: Calculating material characteristics In the target domain Conditional distribution This reflects the probability of material features appearing in the current project context. For example, in the design of battery brackets for new energy vehicles, the target domain... It may contain a combination of features such as "lightweight + corrosion resistance", and the conditional distribution measures the probability of the material being compatible with this combination.

[0096] Calculation of the empirical distribution mean of the source domain: Calculation of the source domain Features of all materials The mean of the empirical distribution is used as the baseline distribution for historical project characteristics.

[0097] KL divergence calculation: The KL divergence formula measures the difference between the mean of the conditional distribution in the target domain and the mean of the empirical distribution in the source domain. The larger the KL divergence value, the higher the novelty of the material in the current project. For example, if a certain composite material feature is rarely seen in historical automotive projects (low mean of the source domain distribution), but has high demand in the current new energy project (high conditional distribution in the target domain), then a large KL divergence value indicates high novelty.

[0098] Through formula Balancing precision, innovation, and manufacturability, among which:

[0099] Integrating Design Intent with Novelty: Ensuring the Conformity of Design Intent With novelty Through the index term Integration To adjust the parameters. When When the index is high, the novelty factor amplifies the impact of novelty on the recommendation score, encouraging the recommendation of innovative materials; conversely, it places more emphasis on matching design intent.

[0100] Creating compatibility filters: via the ReLU function Filter out materials that are incompatible with manufacturing process P. For example, if the structural complexity of a certain 3D printing material exceeds the current factory equipment capabilities, If the ReLU value is 0, the material will be directly excluded from the recommendation list; only when... At that time, the compatibility score is retained for product calculation.

[0101] Comprehensive scoring mechanism: final score Simultaneously consider design intent matching (accuracy), cross-project innovation (novelty), and production feasibility (manufacturability). For example, in the recommendation of aero-engine parts, if a certain titanium alloy material simultaneously meets the design strength requirements... Adopting new coating technology And it meets the requirements of the forging process. Then its fusion score It will be significantly higher than other materials and will be given priority in recommendations.

[0102] More specifically, the self-evolutionary optimization module performs the following steps:

[0103] Meta-learning training: through formulas Optimize model parameters, where For task-specific parameters, Controlling model complexity specifically includes: sampling multiple design tasks. Calculate the parameters after gradient update for each task. Minimize the weighted sum of task loss and parameter offset;

[0104] Concept drift detection: via formula Discriminant feature distribution changes include: training the discriminator. Distinguish current features With historical characteristics When fighting losses The model is re-reconstructed on demand; the recommendation model is retrained using the latest data.

[0105] Through formula Optimize model parameters, where For task-specific parameters, To control model complexity, the specific execution steps are as follows:

[0106] Multi-task sampling: Sampling multiple different design tasks from historical design data. For example, tasks in different scenarios such as automotive parts design and aerospace structural design ensure coverage of diverse design needs.

[0107] Task-specific parameter calculation: For each sampling task Based on the current model parameters Calculate task-specific parameters using gradient descent. Specifically ,in For learning rate, For the task The gradient of the loss function is used to adaptively adjust the model parameters for the specific task.

[0108] Weighted sum minimization: Constructing a meta-learning loss function It consists of two parts: mission loss and parameter offset terms By optimizing this loss function, the optimal shared parameters can be found across different tasks. ,in This is used to balance task loss and parameter stability, avoid overfitting the model on specific tasks, and ensure the model's generalization ability.

[0109] Through formula The discriminant feature distribution changes specifically include:

[0110] Discriminator training: Training a discriminator Its goal is to distinguish the current features. (e.g., the material characteristics of the current design task) and historical characteristics (As in the design features of the previous stage). The discriminator D typically uses a binary classification neural network, taking the feature vector as input and outputting the probability of belonging to the current feature.

[0111] Countermeasure Loss Calculation: Countermeasure Loss Used to measure the difference in distribution between current features and historical features. When ( When the threshold is set to a preset value, it indicates a significant change in design requirements or data distribution, i.e., "concept drift" has occurred. For example, when new material standards are released or design processes are updated, the characteristic distribution of related materials will change, leading to increased resistance losses.

[0112] Model Restructuring and Retraining: When concept drift is detected, a model restructuring mechanism is triggered. The recommendation model is retrained using the latest design data, and the model parameters are updated so that the model can adapt to the new design feature distribution, maintaining the accuracy and timeliness of the recommendations. For example, in the design of new energy vehicles, if a new battery material emerges, the system will detect a change in feature distribution and retrain the model using the new material data to ensure that the recommendation results conform to the latest design trends.

[0113] More specifically, the physical simulation verification interface performs the following steps:

[0114] Finite element analysis preprocessing: recommending materials Import simulation environment; set boundary conditions and physical parameters;

[0115] Feasibility quantification: through formula Feasibility of calculation simulation, among which For simulation response, As per specifications, this specifically includes: performing finite element analysis to obtain stress and strain distribution. ; Calculate the normalized difference norm with respect to specifications; through function Mapped to feasibility score;

[0116] Dynamic filtering: when Remove the material from the recommended list.

[0117] Import recommended materials: Import materials recommended by the system. Import (e.g., 3D model files) into the finite element analysis simulation environment. For example, in the automotive component design scenario, the recommended aluminum alloy bracket model can be imported into simulation software such as ANSYS or ABAQUS in STL or STEP format.

[0118] Boundary condition settings: Define the boundary conditions of the simulation model. This includes geometric constraints (such as fixed supports and hinge points) and load conditions (such as pressure, torque, and vibration loads). For example, a bottom fixed constraint is set for the support model, and a dynamic load equivalent to the actual working condition is applied.

[0119] Physical parameter configuration: Set material physical parameters (such as elastic modulus, Poisson's ratio, density), environmental parameters (such as temperature, humidity), etc. For example, set the elastic modulus to 70 GPa, the Poisson's ratio to 0.3, and the operating temperature range to -20℃ to 80℃ for aluminum alloy materials.

[0120] Through formula The feasibility of the simulation is calculated, including:

[0121] Finite Element Analysis Execution: Run the Finite Element Analysis (FEA) algorithm on the material. Under the set boundary conditions The simulation was performed to obtain the distribution results of physical quantities such as stress, strain, and displacement. For example, calculate the maximum stress distribution cloud map of the support under load.

[0122] Normalized difference calculation: Calculating the simulation response In accordance with specifications The difference norm, and by dividing by the scale parameter Normalization is performed. For example, if the specification requires the maximum stress to not exceed 200 MPa, and the simulation yields a maximum stress of 250 MPa, then the difference norm is... After normalization, it becomes Typically, the specification tolerance range is taken.

[0123] Sigmoid function mapping: via the Sigmoid function Map the normalized difference to The feasibility score for the interval. When When it approaches 0, The score approaches 1 (high feasibility); when the difference is significant, the score approaches 0 (low feasibility).

[0124] Feasibility threshold assessment: Setting a feasibility threshold (e.g., 0.7), when the feasibility score of the material is... If so, the material is deemed not to meet the design requirements.

[0125] Recommendation list update: Automatically remove materials with insufficient feasibility from the recommendation list to avoid recommending design solutions that cannot pass physical verification. For example, if the feasibility score of a support model is 0.65 < 0.7, it will be removed from the recommendation results, and the user will be notified that the material may have insufficient strength under the current working conditions.

[0126] More specifically, the processing module will execute the following steps:

[0127] Gradient embedding generation: via formula The process of constructing user profiles by aggregating the gradient directions of historical design decisions includes: extracting basic attribute features of new users; and retrieving historical design sequences of similar users. ; Calculate the mean gradient of each design function;

[0128] Cross-spectral feature propagation: via formula Generating virtual user characteristics specifically includes: locating neighbor nodes on a cross-project graph. Aggregate neighbor features using a cross-domain graph neural network; generate virtual user features using the READOUT function. .

[0129] Through formula The user profile is constructed by aggregating the gradient directions of historical design decisions. The specific steps are as follows:

[0130] New User Basic Attribute Extraction: Collect basic attribute characteristics of new users, including but not limited to their industry (such as automotive manufacturing, aerospace), design field (such as structural design, component design), commonly used material types, and target product specifications. For example, if a new user is a "new energy vehicle battery bracket designer," then attributes such as "automotive industry" and "battery bracket design" will be extracted.

[0131] Similar User Historical Sequence Retrieval: Based on the basic attributes of new users, retrieve the historical design sequences of similar users from the historical user database. Similarity matching rules can include industry tag matching, design task type matching, and material overlap. For example, retrieving the historical design function sequences of 100 senior designers specializing in "car battery bracket design".

[0132] Calculation of the gradient mean of the design function: For each historical design function Calculate its value in parameter space gradient at This characterizes the direction and intensity of changes in design decisions. The gradient embedding vector is obtained by arithmetically averaging the T gradients using a formula. This vector reflects the design preference trends of similar user groups. For example, in battery bracket design, the gradient mean may reflect a design tendency of "lightweight over strength".

[0133] Through formula Generating virtual user characteristics, specifically including:

[0134] Cross-project knowledge graph neighbor location: In a cross-project knowledge graph, locate the neighbor nodes of a new user by using their basic attributes as query conditions. Neighbor nodes can include resource nodes for similar design tasks, design case nodes from experts in the same field, etc. For example, a new user's "battery bracket design" attribute will be associated with all car battery bracket-related resource nodes and historical design case nodes across the project graph.

[0135] Cross-domain graph neural network feature aggregation: utilizing cross-domain graph neural networks For neighboring nodes eigenvectors Weighted aggregation is performed. This network can learn feature mapping relationships between different project domains, such as establishing association weights between lightweight design features in the aerospace field and manufacturing process features in the automotive field. During the aggregation process, the semantic association strength between nodes (such as functional similarity and geometric compatibility) is considered as attention weights.

[0136] Feature generation using the READOUT function: The READOUT function performs global pooling on the aggregated neighbor features to generate virtual user features. The READOUT function typically uses summation, averaging, or max pooling operations to compress graph structure information into a fixed-dimensional vector. For example, it aggregates geometric, material, and process features related to battery brackets from across project graphs into a 128-dimensional virtual user feature vector, which serves as the basis for recommendations during the cold start phase.

[0137] More specifically, the context-adaptive output module performs the following steps:

[0138] Design phase awareness, through piecewise functions:

[0139] ;

[0140] Dynamically adjust rendering precision, specifically including: monitoring design time. and milestone events; initial output simplified model Intermediate output parameterized model Later output of physical accuracy model ; For the preset stage time threshold (e.g.) Deadline for the conceptual design phase (Detailed design phase deadline).

[0141] Real-time monitoring of design duration and milestone events (such as scheme review nodes, simulation start nodes);

[0142] when When the design is in its early stages, a low-poly simplified model is output to improve the efficiency of interactive preview.

[0143] when When the design is considered to be in the mid-stage, a parametric model is output, supporting dimension-driven structural modifications.

[0144] when When the design is in its later stages, a physical accuracy model (Physics-based) is output, which includes the parameters required for simulation, such as material properties and mechanical characteristics.

[0145] The model compression ratio is adaptively adjusted using the following formula:

[0146] ;

[0147] Real-time acquisition of available GPU memory on the user side (unit: MB); The original model file size (in MB); This is the delay impact coefficient (an empirical value, usually taken as 0.1-0.5). Network latency (unit: milliseconds); specifically includes real-time acquisition of client availability. Memory Based on network latency Dynamically adjust the compression ratio; initiate chunked transmission when the compression ratio is <0.8.

[0148] The model is decomposed into several sub-blocks according to its geometric structure (such as splitting it into assembly components); a streaming protocol is used to prioritize the transmission of visible sub-blocks, while the remaining parts are loaded asynchronously in the background to ensure smooth rendering.

[0149] In the context of automotive body design: the concept design stage ( As in the previous 3 days), the system outputs a Low-poly model (face count ≤ 10k), supporting rapid modeling iteration; detailed design stage ( (e.g., days 4-15), switch to the Parametric model, including parametric sections of the vehicle body frame (e.g., beam cross-section dimensions can be adjusted in real time); simulation verification stage ( (e.g., after day 16), a Physics-based model is generated, integrating material mechanical properties (e.g., yield strength of steel plates) and connection constraints (e.g., weld stiffness). If the user accesses the system via a mobile device (GPU memory ≤ 256MB, network latency ≥ 150ms), the system calculates the compression ratio and initiates chunked transmission to ensure that the model loading time is ≤ 3 seconds.

[0150] More specifically, the multi-objective decision-making module performs the following steps:

[0151] The risk of recommendations is quantified using the Logistic function, as shown in the following formula:

[0152] ;

[0153] This is the i-th recommended material; For material Historical performance fluctuation data (such as peak stress deviation, dimensional tolerance, etc.); For confidence level The value at risk indicates that... The maximum possible performance loss of the material within the probability range; This is the risk sensitivity coefficient (an empirical value, usually taken as 2-5).

[0154] The implementation steps are as follows:

[0155] Collect historical performance index fluctuation data of the materials, such as temperature tolerance fluctuations of aero-engine blades and fatigue strength deviations of automotive chassis parts; calculate confidence levels based on historical data. The value at risk is, for example, if the maximum tensile strength loss of an aluminum alloy part is 15 MPa at a 95% confidence level, then... The VaR value is mapped to a risk score in the [0,1] interval using the Logistic function. As VaR increases... A value closer to 1 indicates a higher risk.

[0156] The similarity weight is dynamically adjusted by accumulating risk through points, as shown in the following formula:

[0157] ;

[0158] for Time-time similarity weights; Initial similarity weight (preset value, such as 0.7); This is the attenuation coefficient (an empirical value, usually taken as 0.01-0.05). This is the set of recommended materials that have already been used; constraints: ,in The minimum weight is 0.3.

[0159] The implementation steps are as follows: Calculate the cumulative risk value of used materials in real time. For example, if the risk scores of the first three used materials are 0.2, 0.3, and 0.1 respectively, then... According to the attenuation coefficient Reduce similarity weight, such as If the calculated Then the forced assignment is To avoid excessive weight decay.

[0160] Application example: In the scenario of recommending aerospace parts:

[0161] Value at Risk (VaR) calculation: Historical performance data for a certain titanium alloy connector shows that, at a 99% confidence level, the maximum fatigue strength loss is 20 MPa, i.e. Substitute into the formula:

[0162] This indicates that the material carries extremely high risk and should be recommended with caution.

[0163] Dynamic weight adjustment: initial similarity weight attenuation coefficient , If the cumulative risk score of the first 5 used materials is 1.5, then: The weight remains within the range of [0.2, 0.6], indicating an effective reduction in weight. If the accumulated risk score reaches 20, then... Once the lower limit is reached, it will no longer decay.

[0164] When the system detects that high-risk materials are used frequently, the similarity weight is automatically reduced, and low-risk but functionally similar alternative materials are recommended instead. For example, high-risk titanium alloy parts can be replaced with ceramic matrix composite materials to reduce the risk of project implementation.

[0165] More specifically, the 3D industrial design material AI intelligent recommendation management system also includes a cross-project knowledge distillation module, which is used to compress the capabilities of expert knowledge and complex models into a lightweight model through teacher model training and student model distillation, and deploy it to edge devices to improve the system's operating efficiency and response speed, while ensuring the accuracy of recommendations;

[0166] The cross-project knowledge distillation module performs the following steps:

[0167] Teacher model training: through formulas Optimizing the teacher model specifically includes: training the teacher model using expert-annotated data; and minimizing the recommendation score. With expert rating The mean square error;

[0168] Student model distillation: via formula The compressed model specifically includes: extracting the probability distribution of the teacher model output. Minimize and student model distribution of Divergence; Add sparsity regularization term Compress model parameters; deploy lightweight student models to edge devices.

[0169] Teacher model training mechanism:

[0170] The objective definition is optimized, and the teacher model is optimized using the mean squared error (MSE) loss function, as shown in the following formula:

[0171] ;

[0172] Let be the teacher model's recommendation score for the i-th material; Reference scores (such as design compliance and engineering feasibility scores) are provided by domain experts.

[0173] Implementation steps: Collect scores from domain experts on typical design scenarios, such as material compatibility scores for engine parts by aerospace experts and fatigue strength scores for chassis components by automotive engineers. Train a teacher model (e.g., a Transformer-based graph neural network model) using expert-annotated data, and minimize the mean squared error between the recommended score and the expert score through backpropagation to ensure that the teacher model has expert-level reasoning capabilities.

[0174] Student model distillation mechanism:

[0175] The compression optimization formula, combining KL divergence and the L1 regularized compression model, is as follows:

[0176] ;

[0177] The output of the teacher model is the probability distribution of recommended materials (such as the probability of each material being recommended). The probability distribution output by the student model; For student model parameters, This is an L1 regularization term used for parameter sparsity. This is the regularization coefficient (an empirical value, usually taken as 0.001-0.01).

[0178] Implementation steps: Extract the output probability distribution of the teacher model on the test set. As the target signal for knowledge distillation; training student models (such as lightweight GNN models) to make their output distribution With teacher distribution The KL divergence is minimized to preserve the inference logic of the teacher model; L1 regularization is used to constrain the parameters of the student model, reducing redundant connections and compressing the model size to less than 1 / 10 of the original teacher model (e.g., from 100MB to 10MB); the lightweight student model is deployed to edge devices (e.g., the field engineer's tablet, industrial tablet) to support real-time recommendations in offline or low-bandwidth environments.

[0179] Application example: In the design of battery brackets for new energy vehicles:

[0180] Teacher Model Training: Using 100,000 sets of data (including scores for material strength, heat dissipation performance, etc.) annotated by battery experts, a Transformer-based teacher model was trained to make its recommendation scores match the expert scores. .

[0181] Student model distillation: The student model adopts a 3-layer graph convolutional network with a parameter scale of 15% of that of the teacher model; through KL divergence optimization, the KL divergence between the student model's recommendation probability distribution and the teacher model is ≤0.1, preserving the core recommendation logic; after L1 regularization, the model file size is compressed to 8MB, which can be deployed to the production line inspection tablet (memory ≤2GB); when field engineers call the recommendation, the model response time is shortened from 3 seconds in the cloud to 500ms on the edge device.

[0182] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A 3D industrial design material AI intelligent recommendation management system, characterized in that, include: The dynamic knowledge graph module is used to build and update weighted material association graphs in real time. , where nodes Represents design materials, edges Indicates the semantic relationship between materials, weight Characterizes the strength of the association; The multi-source sensing input module is used to collect user interaction behavior data, design environment parameters, and physical simulation feedback data; The neural symbolic fusion engine generates recommendation decisions through hybrid neural networks and symbolic reasoning. The context-adaptive output module dynamically switches rendering modes according to the design phase and adaptively adjusts output precision based on device performance. The self-evolutionary optimization module achieves closed-loop iterative optimization of the recommendation model through meta-learning and concept drift detection; The recommendation rule update module is used to confirm the update of recommendation rules for the target user when the recommendation rule fit of the target user is less than the set fit.

2. The AI-powered intelligent recommendation and management system for 3D industrial design materials according to claim 1, characterized in that, Construction of the material association map Includes the following steps: Node feature generation: via formula Integrating text metadata semantic features and three-dimensional geometric features Topology information, specifically including: Extract word vectors from text metadata using the BERT model; pass The network extracts local geometric features from the 3D mesh; The two types of features are concatenated and then input. Generate node embedding vectors ; Dynamic edge weight update: via formula Iteratively update edge weights, where Quantitative design functional gradient similarity, specifically including: Calculate historical co-occurrence frequency Statistical characteristics ; Solving for adjacent material design functions The cosine value of the gradient angle; The feature vector is formed by concatenating historical weights, co-occurrence features, and gradient similarity. Through the fully connected layer and Sigmoid activation function Output new weights .

3. The AI-powered intelligent recommendation and management system for 3D industrial design materials according to claim 2, characterized in that, The neural symbol fusion engine performs the following steps: Symbolic logical reasoning: through formulas Implement tensor matching between user queries and historical behavior rules, where Tensor contraction operations specifically include: Parsing user queries For predicate logic expressions; Integrate predicate logic expressions with design rules from the knowledge base Perform tensor matching; Symbolic inference results are generated using the Transformer encoding rule inference chain. ; Neural network feature extraction: through formulas Based on the map Graph attention network extracts material features Specifically, it includes: Run a multi-head attention mechanism on the material association graph; Generate neural network features by aggregating the feature information of neighboring nodes ; Feature fusion: via formula This involves concatenating the channels of symbolic reasoning results with neural network features, where... This indicates that the feature concatenation operation is used in the subsequent comprehensive calculation of the recommendation score.

4. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 3, characterized in that, Generating a recommendation score involves the following steps: Calculate the degree of conformity with design intent: using the formula In parameter space Inner integration, minimizing the user-constrained gradient Gradient of material properties The differences specifically include: Discretize the design parameter space into a set of grid points. ; Calculate the gradient difference norm at each grid point. ; The differences are mapped using a Gaussian kernel function and then summed and normalized. Calculate cross-project novelty: using the formula use Divergence measures target domain With source domain The characteristic distribution differences include: Align the feature spaces of the source and target domains using the optimal transmission algorithm; Calculate the conditional distribution of material features in the target domain. ; Calculate the mean of the empirical distribution of source domain features; Solve Divergence value; Generate a fusion score using a formula. Balancing precision, innovation, and manufacturability, among which Express material With manufacturing process Compatibility rating.

5. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 4, characterized in that, The self-evolutionary optimization module performs the following steps: Meta-learning training: through formulas Optimize model parameters, where For task-specific parameters, Controlling model complexity specifically includes: Sampling multiple design tasks ; Calculate the parameters after gradient update for each task. ; Minimize the weighted sum of task loss and parameter offset; Concept drift detection: via formula The discriminant feature distribution changes specifically include: Training the discriminator Distinguish current features With historical characteristics ; When the fight against loss Time-triggered model reconstruction; The recommendation model was retrained using the latest data.

6. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 5, characterized in that, The physical simulation verification interface performs the following steps: Finite element analysis preprocessing: Recommended materials Import simulation environment; Set boundary conditions and physical parameters; Feasibility quantification: through formulas Feasibility of calculation simulation, among which For simulation response, The specifications include: Run finite element analysis to obtain stress and strain distribution ; Calculate the normalized difference norm between the specification and the actual value; pass function Mapped to feasibility score; Dynamic filtering: when Remove the material from the recommended list.

7. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 6, characterized in that, The AI-powered intelligent recommendation management system for 3D industrial design materials also includes a cold start processing module, which is used to generate user profiles and virtual user features through gradient embedding and cross-graph feature propagation in scenarios with new users or insufficient data, providing an initial basis for subsequent recommendations. The user features generated are the basis for multi-objective decision-making to optimize recommendations in the cold start stage. The startup processing module executes the following steps: Gradient embedding generation: via formula Constructing user profiles by aggregating the gradient directions of historical design decisions, specifically including: Extract the basic attribute characteristics of new users; Retrieving historical design sequences of similar users ; Calculate the mean gradient of each design function; Cross-spectral feature propagation: via formula Generating virtual user characteristics, specifically including: Locating neighbor nodes on a cross-project graph ; Aggregate neighbor features using a cross-domain graph neural network; Use the READOUT function to generate virtual user characteristics. .

8. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 7, characterized in that, The context-adaptive output module performs the following steps: Design phase awareness, through piecewise functions: ； Dynamically adjusting rendering precision includes: Monitoring design duration and milestone events; Initial output simplified model Intermediate output parameterized model Later output of physical accuracy model ; Device adaptive compression: via formula Dynamically adjusting the model compression ratio specifically includes: Real-time acquisition of user-end availability Memory ; Based on network latency Dynamically adjust compression ratio; Chunked transmission is initiated when the compression ratio is less than 0.

8.

9. The AI-powered intelligent recommendation and management system for three-dimensional industrial design materials according to claim 8, characterized in that, The AI-powered intelligent recommendation management system for 3D industrial design materials also includes a multi-objective decision-making module, which comprehensively considers multiple objectives such as risk and similarity during the recommendation process. It optimizes the recommendation results through risk value calculation and dynamic weight adjustment to make the recommendations more in line with actual needs. The multi-objective decision-making module performs the following steps: Value at Risk (VaR) calculation: using the formula Quantifying recommendation risks, among which The historical performance variation at risk value specifically includes: Statistical data on historical performance fluctuations of materials ; Calculate confidence level The value under risk; pass The function maps to a risk score; Dynamic weight adjustment: via formula Adaptive adjustment of similarity weights, specifically including: Real-time points are the accumulated risk value of materials used for material processing; According to the attenuation coefficient Reduce similarity weight; Constraint weight range .

10. The AI-powered intelligent recommendation management system for three-dimensional industrial design materials according to claim 9, characterized in that, The 3D industrial design material AI intelligent recommendation management system also includes a cross-project knowledge distillation module, which is used to compress the capabilities of expert knowledge and complex models into a lightweight model through teacher model training and student model distillation, and deploy it to edge devices to improve the system's operating efficiency and response speed, while ensuring the accuracy of recommendations. The cross-project knowledge distillation module performs the following steps: Teacher model training: through formulas Optimizing the teacher model specifically includes: Train teacher models using expert-annotated data; Minimize recommendation score With expert rating The mean square error; Student model distillation: via formula Compression models, specifically including: Extract the probability distribution of the teacher model output ; Minimize and student model distribution of Divergence; Add sparse regularization terms Compress model parameters; Deploy lightweight student models to edge devices.

Images