Data-driven product collaborative design management method and system

Abstract

The invention discloses a product collaborative design management method and system based on data driving, and relates to the technical field of product design management, and the method comprises the steps: constructing a double-flow heterogeneous quantitative collection network; establishing an incidence matrix and calculating dynamic coupling strength among the parameters; designing a recursive deep belief network based on the incidence matrix and the dynamic coupling strength, encoding product structure parameters and design process data into a probability graph model, and constructing a design knowledge base; dynamically distributing the design rules in the design knowledge base by adopting a swarm intelligent optimization algorithm, generating a collaborative decision-making unit, and establishing a constraint propagation link; and generating a multi-target collaborative optimization scheme group, and screening an optimal scheme group based on Pareto frontier. According to the invention, omnibearing acquisition of product structure parameters and design flow data is realized through the double-flow heterogeneous quantitative acquisition network, and the data integrity is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of product design management, and in particular to a data-driven product collaborative design management method and system. Background Art

[0002] With the continuous improvement of product design complexity, traditional product design management methods have been difficult to meet the needs of modern manufacturing. The existing product collaborative design management methods mainly have the following problems: First, the design data collection method is single, and it is difficult to comprehensively obtain the correlation between product structure parameters and design process data; second, the form of design knowledge expression is fixed and cannot adapt to the dynamically changing design environment; third, the collaborative decision-making process lacks systematic guidance, and the quality of the optimization scheme is difficult to guarantee.

[0003] Most of the design management systems currently used in the industry adopt static rule libraries and simple parameter optimization methods. This method has problems such as low knowledge acquisition efficiency, poor knowledge reuse rate, and local optimality of the optimization results. Although some researchers have tried to introduce artificial intelligence technology to improve the design management system, due to the lack of in-depth analysis of the dynamic characteristics and coupling relationships in the design process, the practicality and scalability of the system are still limited. Summary of the Invention

[0004] In view of the problems existing in the existing product collaborative design management methods in data collection, knowledge expression, decision optimization, etc., the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to construct a data-driven intelligent product collaborative design management method and system to realize the automatic collection, modeling and optimization of the product design process.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides a data-driven product collaborative design management method, which includes constructing a dual-stream heterogeneous quantization acquisition network to respectively perform parallel acquisition on product structure parameters and design process data, and converting the heterogeneous data into a unified feature space representation; using a dynamic adaptive quantization unit to reduce the dimension of the unified feature space representation, extract key feature combinations, establish an association matrix and calculate the dynamic coupling strength between parameters; based on the association matrix and the dynamic coupling strength, design a recursive deep belief network to encode the product structure parameters and design process data into a probabilistic graphical model, and construct a design knowledge base; adopt a swarm intelligence optimization algorithm to dynamically allocate the design rules in the design knowledge base, generate collaborative decision-making units, and establish a constraint propagation link; perform iterative calculations on the collaborative decision-making units on the constraint propagation link, fuse the marginal probability estimation method, generate a multi-objective collaborative optimization solution group, and screen the optimal solution group based on the Pareto front; verify the convergence of the optimal solution group, update the probabilistic graphical model in the design knowledge base, and at the same time perform online learning and adjustment on the dynamic coupling strength.

[0007] As a preferred solution of the data-driven product collaborative design management method of the present invention, the construction of the dual-stream heterogeneous quantization acquisition network includes the following steps: constructing a dual-channel data acquisition unit, the dual-channel data acquisition unit includes a product structure parameter acquisition channel and a design process data acquisition channel, and configure the number of input nodes for each channel; construct a multi-layer convolutional neural network in the product structure parameter acquisition channel, and obtain a structure feature tensor through feature extraction and spatial transformation; construct a long short-term memory network in the design process data acquisition channel, and combine the multi-head temporal attention mechanism to extract the process feature tensor; perform hierarchical tensor decomposition operations on the structure feature tensor and the process feature tensor to obtain the corresponding core tensor and factor matrix; construct a feature projection function according to the core tensor and the factor matrix, map features of different dimensions to a unified feature space, and obtain a unified feature representation matrix; impose sparse constraints and regularization processing on the unified feature representation matrix, and optimize the unified feature matrix by the alternating direction multiplier method to obtain a unified feature space representation; perform normalization preprocessing on the unified feature space representation to generate a standardized feature representation.

[0008] As a preferred solution of the data-driven product collaborative design management method of the present invention, the steps of establishing an association matrix and calculating the dynamic coupling strength between parameters are as follows: constructing a dynamic adaptive quantization unit, which includes an adaptive threshold layer, a dimension reconstruction layer, and a feature selection layer, and using the standardized feature representation as the input; using the adaptive threshold layer to perform piecewise quantization on the standardized feature representation, calculating the optimal piecewise point based on the information entropy criterion, dynamically adjusting the upper and lower threshold values of the quantization interval, and obtaining a quantized feature sequence; in the dimension reconstruction layer, using a manifold learning method to perform non-linear dimensionality reduction on the quantized feature sequence, determining the dominant feature direction through eigenvalue decomposition, and generating a dimensionality-reduced feature combination; introducing a mutual information criterion to the dimensionality-reduced feature combination through the feature selection layer, screening a key feature subset, optimizing the feature weights using a swarm intelligence algorithm, and recombining the selected key feature subset into a multi-dimensional feature combination sequence; calculating the correlation degree between product structure parameters and design process data based on the multi-dimensional feature combination sequence, constructing a two-dimensional association matrix in combination with expert evaluation weights; dynamically weighting the elements of the association matrix, using a time-series sliding window method to analyze the parameter change trend, and calculating the dynamic coupling strength index between parameters.

[0009] As a preferred solution of the data-driven product collaborative design management method of the present invention, the steps of encoding product structure parameters and design process data into a probabilistic graphical model are as follows: constructing a recursive deep belief network structure, which includes a feature encoding layer, a multi-layer restricted Boltzmann machine unit, and a recursive connection layer; inputting the association matrix into the feature encoding layer, completing matrix data reconstruction through the encoding and decoding operations of an autoencoder, and obtaining an initial feature vector containing rule importance parameters; performing contrastive divergence operations within the multi-layer restricted Boltzmann machine unit, setting the network connection weights according to the dynamic coupling strength value, and generating an inter-layer probability distribution; importing a time-series signal sequence into the recursive connection layer, adjusting the network parameters using backpropagation operations, and generating a probabilistic graphical model of product structure parameters and design process data; performing Bayesian inference operations on the probabilistic graphical model, constructing a design rule chain, and generating a design rule set based on a confidence threshold; classifying the design rule set according to the graph structure similarity, establishing a hierarchical retrieval mechanism based on hierarchical hash indexing, and forming a design knowledge base.

[0010] As a preferred solution of the data-driven product collaborative design management method of the present invention, the establishment of the constraint propagation link includes the following steps: constructing a hybrid swarm intelligence optimization algorithm module integrated with two units, the hybrid swarm intelligence optimization algorithm module including an ant colony optimization unit and a particle swarm optimization unit; extracting a design rule set from the design knowledge base and dynamically allocating the design rule set into multiple rule subsets according to the rule attribute values; constructing a collaborative decision-making unit based on the rule subsets, each collaborative decision-making unit including an associated design rule combination and its constraint conditions, forming an initial set of collaborative decision-making units; the ant colony optimization unit setting the pheromone intensity value based on the rule importance parameter, establishing a multi-layer pheromone distribution matrix, and optimizing the organizational structure of the collaborative decision-making unit set through path iteration operations to obtain a collaborative decision-making unit organizational structure matrix; the particle swarm optimization unit mapping the organizational structure matrix to the decision space, combining the constraint conditions of the collaborative decision-making unit, and calculating the velocity vector and position vector of each collaborative decision-making unit according to the preset number of iterations to obtain a probability distribution matrix; calculating the mutual dependence coefficient of each collaborative decision-making unit according to the probability distribution matrix, arranging the dependence coefficients into a probability transfer matrix, and constructing a constraint network including node weights; connecting the collaborative decision-making unit nodes in the constraint network whose dependence coefficients satisfy the threshold condition to form a directed acyclic constraint graph; sorting the collaborative decision-making unit nodes in the directed acyclic constraint graph according to the in-degree value, marking the node numbers, and generating a constraint propagation link table.

[0011] As a preferred solution of the data-driven product collaborative design management method of the present invention, the generation of a multi-objective collaborative optimization solution group includes the following steps: constructing a decision tree structure based on the node order in the constraint propagation link table and setting a node splitting threshold according to historical data; adopting an adaptive sampling strategy and a dimension reduction mechanism, recording the value range of decision variables, and generating a candidate solution space; applying a marginal probability estimation algorithm to the candidate solution space, adjusting the kernel density estimation bandwidth, calculating the probability density distribution, and determining a high-probability decision region; constructing a multi-objective optimization model in the high-probability decision region, determining the objective weights using the analytic hierarchy process, and setting optimization constraint conditions; combining the use of a crossover operator and a local search operator to generate a Pareto optimal solution set and calculating the objective function values of non-dominated solutions; performing hierarchical sorting on the non-dominated solutions based on the Pareto front, and screening the optimal solution group in combination with sensitivity analysis and calculating a comprehensive score.

[0012] As a preferred solution of the data-driven product collaborative design management method of the present invention, wherein: updating the probabilistic graph model in the design knowledge base includes the following steps: constructing a qubit matrix, encoding the optimal solution set into a quantum state sequence through phase encoding, and setting an adaptive annealing parameter; performing quantum annealing iterative operations and local search optimization, recording the change process of the objective function value, and judging the convergence of the solution set; updating the convergence verification result to the probabilistic graph model, and incrementally updating the node probability distribution and edge probability distribution in the design knowledge base; extracting the design rule feature vector based on information gain, calculating the rule importance, and adjusting the structure of the design knowledge base; calculating the correction amount of the dynamic coupling strength by using a composite statistical method, and performing a constrained gradient update; performing piecewise normalization on the corrected dynamic coupling strength, updating the correlation matrix, and completing the optimization iteration.

[0013] In a second aspect, an embodiment of the present invention provides a data-driven product collaborative design management system, which includes a data acquisition and conversion module for constructing a dual-stream heterogeneous quantization acquisition network to perform parallel acquisition of product structure parameters and design process data respectively, and converting the heterogeneous data into a unified feature space representation; a feature analysis module for using a dynamic adaptive quantization unit to reduce the dimension of the unified feature space representation, extract key feature combinations, establish a correlation matrix, and calculate the dynamic coupling strength between parameters; a knowledge modeling module for designing a recursive deep belief network based on the correlation matrix and the dynamic coupling strength, encoding the product structure parameters and design process data into a probabilistic graph model, and constructing a design knowledge base; a decision generation module for dynamically allocating the design rules in the design knowledge base by using a swarm intelligence optimization algorithm, generating collaborative decision-making units, and establishing a constraint propagation link; a solution optimization module for performing iterative calculations on the collaborative decision-making units on the constraint propagation link, fusing the marginal probability estimation method, generating a multi-objective collaborative optimization solution set, and screening the optimal solution set based on the Pareto front; a verification and adjustment module for performing convergence verification on the optimal solution set, updating the probabilistic graph model in the design knowledge base, and at the same time performing online learning and adjustment on the dynamic coupling strength.

[0014] The beneficial effects of the present invention are as follows: The all-round acquisition of product structure parameters and design process data is realized through the dual-stream heterogeneous quantization acquisition network, improving data integrity; by using a recursive deep belief network and a probabilistic graph model, the system has an adaptive learning ability and can dynamically update design knowledge; combining a swarm intelligence optimization algorithm and a marginal probability estimation method, multi-objective collaborative optimization is realized, significantly improving the quality of the design solution; introducing a quantum annealing algorithm for convergence verification ensures the reliability of the optimization result; the dynamic adjustment of the coupling strength through an online learning mechanism enhances the robustness and adaptability of the system. Description of the Drawings

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0016] Figure 1 It is a framework flowchart of a data-driven product collaborative design management method.

[0017] Figure 2 It is a flowchart for constructing a dual-stream heterogeneous quantization acquisition network of a data-driven product collaborative design management method.

[0018] Figure 3 It is a flowchart for establishing a constraint propagation link of a data-driven product collaborative design management method. Specific Embodiments

[0019] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0020] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from this description. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0021] Secondly, the so-called "one embodiment" or "embodiment" herein refers to specific features, structures, or characteristics that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that mutually excludes other embodiments.

[0022] Embodiment 1, referring to Figures 1 to 3 , is the first embodiment of the present invention. This embodiment provides a data-driven product collaborative design management method, and the framework flowchart is as shown in Figure 1 , including: S1: Construct a dual-stream heterogeneous quantization acquisition network, parallelly collect product structure parameters and design process data respectively, and convert the heterogeneous data into a unified feature space representation.

[0023] Specifically, the flowchart for constructing the dual-stream heterogeneous quantization acquisition network is as shown in Figure 2 , including the following steps: S1.1: Construct a dual-channel data acquisition unit, which includes a product structure parameter acquisition channel and a design process data acquisition channel, and configure the number of input nodes for each channel.

[0024] Among them, the product structure parameters include but are not limited to geometric parameters, physical parameters, functional parameters, assembly parameters, process parameters, and system parameters, and the design process data includes but is not limited to timing data, resource data, quality data, collaboration data, decision-making data, knowledge data, and innovation data.

[0025] S1.2: Construct a multi-layer convolutional neural network in the product structure parameter acquisition channel, and obtain the structure feature tensor through feature extraction and spatial transformation.

[0026] S1.3: Construct a long short-term memory network in the design process data acquisition channel, and extract the process feature tensor by combining the multi-head temporal attention mechanism.

[0027] S1.4: Perform hierarchical tensor decomposition operations on the structure feature tensor and the process feature tensor to obtain the corresponding core tensors and factor matrices.

[0028] In the specific implementation process, first construct a tensor decomposition module, including a feature separation unit and a matrix reconstruction unit; then, in the feature separation unit, use the Tucker decomposition algorithm to decompose the structure feature tensor, calculate the first orthogonal basis matrix, and generate the structure feature core tensor; use the same Tucker decomposition algorithm for the process feature tensor, calculate the second orthogonal basis matrix, and generate the process feature core tensor; finally, in the matrix reconstruction unit, perform tensor product operations on the first orthogonal basis matrix and the second orthogonal basis matrix with the corresponding core tensors respectively to obtain the structure feature factor matrix and the process feature factor matrix; perform column vector normalization on the structure feature factor matrix and the process feature factor matrix to generate a group of standardized factor matrices, and at the same time obtain the corresponding core tensor group.

[0029] Preferably, through the hierarchical tensor decomposition operation, the redundancy of the feature data can be effectively reduced, the most representative core feature combination can be extracted, and at the same time the main information of the original data is maintained, significantly improving the computational efficiency of subsequent processing and the accuracy of feature expression. In addition, the obtained core tensors and factor matrices have good interpretability, which is convenient for engineers to understand and analyze the internal relationships between features.

[0030] S1.5: Construct a feature projection function based on the core tensors and factor matrices, map the features of different dimensions to a unified feature space, and obtain a unified feature representation matrix.

[0031] In the feature mapping stage, first, a feature projection unit is constructed, including a mapping function generator and a spatial transformer. Next, in the mapping function generator, based on the element distribution characteristics of the core tensor group, a non-linear mapping function is constructed using the kernel function method, and a scaling factor is introduced to adjust the mapping intensity. The scaling factor is adaptively adjusted based on the statistical characteristics of the feature distribution. The mapping function is used to transform the standardized factor matrix group, calculate the matrix inner product, and generate a feature similarity matrix. Further, in the spatial transformer, a distance metric matrix is constructed based on the feature similarity matrix. The core tensor group is projected into the target feature space through the mapping function and tensor operations are performed with the distance metric matrix to generate a unified feature representation matrix.

[0032] S1.6: Apply sparse constraints and regularization processing to the unified feature representation matrix, and optimize the unified feature matrix by the alternating direction multiplier method to obtain the unified feature space representation.

[0033] S1.7: Perform normalization preprocessing on the unified feature space representation to generate a standardized feature representation.

[0034] S2: Use a dynamic adaptive quantization unit to reduce the dimension of the unified feature space representation, extract key feature combinations, establish an association matrix, and calculate the dynamic coupling strength between parameters.

[0035] Specifically, it includes the following steps: S2.1: Construct a dynamic adaptive quantization unit. The dynamic adaptive quantization unit includes an adaptive threshold layer, a dimension reconstruction layer, and a feature selection layer, and takes the standardized feature representation as the input.

[0036] S2.2: Use the adaptive threshold layer to perform piecewise quantization on the standardized feature representation, calculate the optimal piecewise points based on the information entropy criterion, dynamically adjust the upper and lower threshold values of the quantization interval, and obtain a quantized feature sequence.

[0037] In the quantization processing stage, first, calculate the probability density function of the standardized feature representation, and generate an initial set of piecewise points based on this function. Secondly, optimize the positions of the piecewise points based on the gradient descent method, with the goal of minimizing the conditional entropy, to generate an optimized set of piecewise points. Then, calculate the mean and standard deviation of each sub-interval according to the optimized set of piecewise points, construct a quantization function. Finally, use the quantization function to transform the feature values, generate an initial quantization sequence and calculate the quantization error. Set an error threshold in this process. When the quantization error is less than the error threshold, stop the correction, and correct the quantization coefficient according to the quantization error to generate the final quantized feature sequence.

[0038] Preferably, an adaptive threshold and information entropy criterion-based piecewise quantization method is adopted to achieve intelligent segmentation of feature data, avoiding the limitations of traditional fixed threshold methods. It can automatically adjust the quantization accuracy according to the data distribution characteristics, ensuring both the expression accuracy of the quantized data and significantly reducing the data storage space, thereby improving the overall efficiency of the system.

[0039] S2.3: In the dimensionality reconstruction layer, a manifold learning method is used to perform non-linear dimensionality reduction on the quantized feature sequence. The dominant feature direction is determined through eigenvalue decomposition to generate a dimensionality-reduced feature combination.

[0040] S2.4: The mutual information criterion is introduced into the dimensionality-reduced feature combination through the feature selection layer to screen the key feature subset. A swarm intelligence algorithm is used to optimize the feature weights, and the selected key feature subset is recombined into a multi-dimensional feature combination sequence.

[0041] S2.5: Based on the multi-dimensional feature combination sequence, the correlation degree between the product structure parameters and the design process data is calculated, and a two-dimensional correlation matrix is constructed by combining the expert evaluation weights.

[0042] S2.6: Dynamically weight the elements of the correlation matrix, and use the time series sliding window method to analyze the parameter change trend and calculate the dynamic coupling strength index between parameters.

[0043] For the analysis of the dynamic coupling relationship of parameters, first determine the time series sliding window length according to the data sampling period and engineering actual requirements, and intercept the parameter data at consecutive time points in the correlation matrix; then, based on the parameter data within the sliding window, calculate the correlation coefficient matrix of each parameter pair; next, use the exponential weighting method to perform time series accumulation on the correlation coefficient matrix, giving higher weights to recent data, where the decay coefficient of the exponential weighting method is adaptively adjusted according to the time series characteristics; again, based on the accumulated weights, calculate the dynamic change rate of the parameter pair and establish a parameter change trend matrix; finally, combine the correlation coefficient matrix and the change trend matrix to calculate the dynamic coupling strength index between parameters.

[0044] S3: Based on the correlation matrix and the dynamic coupling strength, design a recursive deep belief network, encode the product structure parameters and design process data into a probabilistic graphical model, and construct a design knowledge base.

[0045] Specifically, it includes the following steps: S3.1: Construct a recursive deep belief network structure, including a feature encoding layer, multiple restricted Boltzmann machine units, and a recursive connection layer.

[0046] S3.2: Input the correlation matrix into the feature encoding layer, and complete the matrix data reconstruction through the encoding and decoding operations of the autoencoder to obtain an initial feature vector containing the rule importance parameters.

[0047] S3.3: Conduct contrastive divergence operation within the multi-layer restricted Boltzmann machine unit, set the network connection weights according to the dynamic coupling strength value, and generate the inter-layer probability distribution.

[0048] Preferably, through the contrastive divergence operation and the setting of the dynamic coupling strength weights, the network can capture the internal laws of the data more accurately, improving the model's expressive ability and generalization performance. At the same time, the generation of the inter-layer probability distribution provides a reliable probability basis for subsequent knowledge reasoning, effectively enhancing the credibility of the design knowledge.

[0049] S3.4: Import the time series signal sequence into the recurrent connection layer, and use backpropagation operation to adjust the network parameters to generate the probability graph model of the product structure parameters and the design process data.

[0050] Specifically, set a time series processing unit in the recurrent connection layer and import the time series signal sequence containing the product structure parameters and the design process data; combine the output result of the feature encoding layer of the recurrent deep belief network with the time series signal sequence to generate a time series feature matrix; based on the inter-layer probability distribution of the multi-layer restricted Boltzmann machine unit, construct a time series feature extraction network to dynamically encode the time series feature matrix; according to the dynamic coupling strength value, set the connection weights of the time series feature extraction network to generate a time series state vector; use backpropagation operation to iteratively optimize the time series state vector and adjust the network parameters until convergence; combine the optimized time series state vector with the inter-layer probability distribution to construct the initial probability graph of the product structure parameters and the design process data; perform structure optimization on the initial probability graph based on the dynamic coupling strength to generate the final probability graph model of the product structure parameters and the design process data.

[0051] S3.5: Conduct Bayesian inference operation on the probability graph model, construct a design rule chain, and generate a design rule set according to the confidence threshold.

[0052] Specifically, construct a Bayesian network structure according to the probability graph model, map the product structure parameters and the design process data to the conditional probability table; set the prior probability distribution in the Bayesian network based on the dynamic coupling strength to construct the conditional dependence relationship between the parameter nodes; perform variable elimination operation on the Bayesian network to extract the node pairs with strong conditional dependence relationships to form the initial rule chain; calculate the posterior probability of each node pair in the initial rule chain and use it as the rule credibility index; compare the rule credibility index with the preset confidence threshold to screen out the high-credibility rule combinations; conduct causality verification on the high-credibility rule combinations based on the inter-layer probability distribution to determine the derivation order between the rules; connect the verified rules according to the derivation order to construct a design rule chain; conduct association analysis on the design rule chain and merge the rules with similar semantics to generate the final design rule set.

[0053] It should be noted that the confidence threshold is determined by combining cross - validation methods with expert experience.

[0054] S3.6: Classify the design rule set according to the graph - structure similarity, establish a hierarchical retrieval mechanism based on hierarchical hash indexing, and form a design knowledge base.

[0055] S3.7: Set the knowledge - base update period, regularly evaluate the effectiveness of the design rule set, and update the design knowledge base according to the evaluation results.

[0056] S4: Use a swarm - intelligence optimization algorithm to dynamically allocate the design rules in the design knowledge base, generate collaborative decision - making units, and establish a constraint - propagation link.

[0057] Specifically, the flowchart for establishing the constraint - propagation link is as Figure 3 shown, including the following steps: S4.1: Construct a hybrid swarm - intelligence optimization algorithm module integrating two units.

[0058] Among them, the hybrid swarm - intelligence optimization algorithm module includes an ant - colony optimization unit and a particle - swarm optimization unit.

[0059] S4.2: Extract the design rule set from the design knowledge base, and dynamically allocate the design rule set into multiple rule subsets according to the rule - attribute values.

[0060] Specifically, establish a rule - attribute feature matrix. The row vectors of the rule - attribute feature matrix represent the design - rule numbers, and the column vectors represent the rule - attribute values, where the rule - attribute values include the design - object type, the constraint - strength coefficient, and the correlation - degree coefficient; use a multi - layer perceptron to extract features from the rule - attribute feature matrix. The hidden layer of the multi - layer perceptron uses a bidirectional gated recurrent unit, and the output layer uses a softmax activation function to generate rule - feature vectors; construct a rule - clustering model, input the rule - feature vectors into the rule - clustering model, and calculate the local density value and relative - distance value of the rule - feature vectors through the density - peak detection algorithm. Furthermore, determine the clustering center points based on the product of the local density value and the relative - distance value, allocate the rule - feature vectors closest to the corresponding clustering center points to form initial rule clusters; calculate the Euclidean distance between the rule - feature vectors within the initial rule clusters, set a dynamic threshold parameter, and merge the rule - feature vectors with Euclidean distances less than the dynamic threshold parameter into rule subsets; conduct a consistency check on the rule subsets, calculate the average correlation - degree coefficient within the rule subsets, and retain the rule subsets when the average correlation - degree coefficient is greater than the preset threshold, otherwise re - allocate the rule subsets.

[0061] S4.3: Construct collaborative decision - making units based on the rule subsets. Each collaborative decision - making unit contains an associated combination of design rules and their constraint conditions, forming an initial set of collaborative decision - making units.

[0062] S4.4: The ant colony optimization unit sets the pheromone intensity value based on the rule importance parameter, establishes a multi-layer pheromone distribution matrix, optimizes the organizational structure of the collaborative decision-making unit set through path iteration operations, and obtains the collaborative decision-making unit organizational structure matrix.

[0063] In the implementation process, calculate the importance score for each rule in the initial collaborative decision-making unit set, which is obtained by the weighted sum of the usage frequency, correlation degree, and constraint strength of the rule, and the score range is within the interval [0, 1]; map the rule importance score to the initial pheromone concentration, and the mapping function uses the sigmoid function for normalization processing to ensure uniform distribution of the pheromone concentration; construct a multi-layer pheromone distribution matrix, where the number of matrix layers is the same as the number of organizational levels of the collaborative decision-making unit, and each layer of the matrix represents the pheromone distribution between decision-making units at that level, and the matrix element value is the pheromone concentration on the corresponding path.

[0064] Correspondingly, initialize the position distribution of the ant colony. Each ant randomly selects a decision-making unit from the initial layer as the starting point, and sets the population size and iteration termination condition; calculate the state transition probability based on the pheromone concentration and heuristic information of adjacent decision-making units in the current layer, where the heuristic information is jointly determined by the similarity between decision-making units and organizational structure constraints; use the roulette wheel selection method, and the ants select the next decision-making unit according to the state transition probability, record the path information during the movement, and update the local pheromone concentration at the same time.

[0065] Accordingly, calculate the fitness value after all ants complete the path search. The fitness value is jointly composed of the path length and the importance score of the decision-making units passed through; sort the fitness of all paths, select the high-fitness paths for global pheromone update, and the update intensity is proportional to the fitness value of the path; determine the termination condition of the path search by judging the change range of the optimal path in consecutive multiple iterations; construct the collaborative decision-making unit organizational structure matrix based on the optimal path, and the element value in the matrix represents the organizational relationship strength between decision-making units, and this strength value is jointly determined by the final pheromone distribution and path fitness.

[0066] Preferably, by introducing the ant colony optimization algorithm and combining the pheromone allocation mechanism of rule importance, the adaptive organizational optimization of decision-making units is realized, which not only improves the organizational efficiency of decision-making units, but also enhances the adaptability of the system to environmental changes. The establishment of the multi-layer pheromone distribution matrix enables the system to optimize simultaneously at different levels, improving the globality of the optimization effect.

[0067] S4.5: The particle swarm optimization unit maps the organizational structure matrix to the decision space, combines the constraint conditions of the collaborative decision-making units, and calculates the velocity vector and position vector of each collaborative decision-making unit according to the preset number of iterations to obtain the probability distribution matrix.

[0068] Specifically, an n-dimensional decision space is constructed, and each collaborative decision unit in the organizational structure matrix is mapped to a dimensional coordinate in the decision space to form an initial decision space representation; a fitness function is constructed based on the constraint conditions of the collaborative decision unit, and the function includes three evaluation indicators: the association strength between decision units, the constraint satisfaction degree, and the decision consistency; the position vector and velocity vector of the particle swarm are initialized. The position vector represents the coordinate point in the decision space, and the velocity vector represents the movement direction and step size of the particle; the fitness value of each particle is calculated, the individual optimal position and the global optimal position are recorded, and an initial state information table of the particle swarm is established.

[0069] On this basis, an inertia weight decreasing strategy is adopted to update the particle velocity vector, where the velocity update formula comprehensively considers the influence of the individual cognitive term and the social cognitive term; the particle position is adjusted based on the updated velocity vector, and at the same time, a position boundary check is performed to ensure that the particle position satisfies the decision space constraint; the state of the particle swarm is re-evaluated according to the fitness function, and the individual optimal position and the global optimal position information are updated; the position distribution information of the particle swarm is converted into a probability distribution form, a probability distribution matrix is constructed, and the matrix element value represents the probability value of the decision unit in a specific state; the probability distribution matrix is normalized to ensure that the sum of the state probabilities of each decision unit is 1, and the final probability distribution matrix is obtained.

[0070] S4.6: Calculate the mutual dependence coefficients of each collaborative decision unit according to the probability distribution matrix, arrange the dependence coefficients into a probability transfer matrix, and construct a constraint network including node weights.

[0071] S4.7: Connect the collaborative decision unit nodes in the constraint network whose dependence coefficients meet the threshold conditions to form a directed acyclic constraint graph.

[0072] It should be added that the dependence coefficients need to meet the preset threshold conditions to establish the connection relationship between nodes in the constraint network. The dependence coefficients reflect the association degree between collaborative decision units, which are obtained by calculating the mutual information and conditional probability between nodes, and the threshold conditions are dynamically adjusted according to the characteristics of the network structure to balance the connectivity and sparsity of the network.

[0073] S4.8: Sort the collaborative decision unit nodes in the directed acyclic constraint graph according to the in-degree value, mark the node numbers, and generate a constraint propagation link table.

[0074] S5: Iteratively calculate the collaborative decision units on the constraint propagation link, fuse the marginal probability estimation method, generate a multi-objective collaborative optimization solution group, and screen the optimal solution group based on the Pareto front.

[0075] Specifically, a decision tree structure is constructed based on the node order in the constraint propagation link table, and the node splitting threshold is set according to historical data; an adaptive sampling strategy and a dimensionality reduction mechanism are adopted to record the value range of decision variables and generate a candidate solution space; the marginal probability estimation algorithm is applied to the candidate solution space, the kernel density estimation bandwidth is adjusted, the probability density distribution is calculated, and the high-probability decision region is determined; a multi-objective optimization model is constructed within the high-probability decision region, the objective weights are determined using the analytic hierarchy process, and the optimization constraint conditions are set; the crossover operator and the local search operator are used in combination to generate a Pareto optimal solution set, and the objective function values of the non-dominated solutions are calculated; the non-dominated solutions are hierarchically sorted based on the Pareto front, and the optimal solution group is screened in combination with sensitivity analysis and the comprehensive score is calculated.

[0076] Preferably, through the integration of iterative calculation of the constraint propagation link and marginal probability estimation, a complete multi-objective optimization scheme generation mechanism is established, which can fully consider the trade-off relationship between multiple optimization objectives while ensuring the feasibility of the scheme. The screening method based on the Pareto front ensures that the final solution group has good balance and diversity, providing more choice space for decision-makers.

[0077] S6: Conduct convergence verification on the optimal solution group, update the probability graph model in the design knowledge base, and at the same time perform online learning and adjustment of the dynamic coupling strength.

[0078] Specifically, it includes the following steps: S6.1: Construct a qubit matrix, encode the optimal solution group into a quantum state sequence through phase encoding, and set the adaptive annealing parameter.

[0079] S6.2: Perform quantum annealing iterative operations and local search optimization, record the change process of the objective function value, and judge the convergence of the solution group.

[0080] In the implementation process, the initial state of quantum annealing is constructed. The qubit matrix of the optimal solution set is initialized according to the Boltzmann distribution at the initial annealing temperature, and the system energy operator is established. The time-dependent Hamiltonian evolution equation is designed, and the problem Hamiltonian and the driving Hamiltonian are linearly combined according to the annealing time to construct a quantum state evolution model. The time-dependent Schrödinger equation is solved to obtain the evolution trajectory of the quantum state over time, and the objective function value at each time step is calculated. Local search is performed within each annealing cycle: the quantum tunneling effect is used to explore the local optimal solution in the energy space, and the state transition path during the search process is recorded. According to the simulated annealing criterion, it is judged whether to accept the new solution: when the new solution is better than the current solution or meets the probability acceptance condition, the current optimal solution and its objective function value are updated. The change of the objective function value in multiple consecutive annealing cycles is calculated. When the change rate is less than the preset threshold, it is determined that the solution set has converged; otherwise, the system temperature is reduced and the iteration continues. The convergence judgment result is output, including the final convergence flag, the change sequence of the objective function value at the time of convergence, the optimal solution, and its corresponding objective function value.

[0081] Preferably, the quantum annealing algorithm is combined with the local search strategy, which significantly improves the accuracy and efficiency of convergence verification and can effectively avoid falling into local optimal solutions. Through the dynamic monitoring of the change of the objective function value, the system can timely judge the convergence state of the optimization process, ensuring the stability and reliability of the final solution.

[0082] S6.3: Update the convergence verification result to the probability graph model, and incrementally update the node probability distribution and edge probability distribution in the design knowledge base.

[0083] S6.4: Extract the design rule feature vector based on information gain, calculate the rule importance, and adjust the structure of the design knowledge base.

[0084] Among them, the rule importance is calculated through the following process: extract the projection coefficient of the rule in the feature space; calculate the sensitivity matrix of the rule to the design objective; fuse the usage frequency and timeliness weight of the rule; determine the combined weight of the rule importance based on fuzzy hierarchical analysis; and obtain the comprehensive importance index of the rule by weighted summation.

[0085] S6.5: Calculate the dynamic coupling strength correction amount using a composite statistical method and perform a constrained gradient update.

[0086] Specifically, based on the convergence data of the optimal solution set, a probability distribution model of design parameters is constructed using the kernel density estimation method to extract the statistical correlation between parameters. Combining with the adaptive weighted moving average algorithm, time series analysis is performed on historical design iteration data to calculate the dynamic change trend of the parameter coupling relationship. The parameter statistical correlation and the dynamic change trend are decomposed at multiple levels, and the multi-scale representation of coupling features is extracted through wavelet transform. The Bayesian inference framework is used to fuse the multi-scale coupling features, calculate the correction amount of the dynamic coupling strength, and set the upper and lower limit constraints for correction. Based on the confidence interval theory, the gradient update step size is determined to perform a restricted update on the dynamic coupling strength to ensure the stability of the correction process. The cross-validation method is used to evaluate the corrected coupling strength, calculate the prediction error, and adjust the correction strategy according to the error feedback. The corrected dynamic coupling strength and its reliability index are output, providing a basis for subsequent parameter normalization and correlation matrix update.

[0087] S6.6: Perform piecewise normalization on the corrected dynamic coupling strength, update the correlation matrix, and complete the optimization iteration.

[0088] Furthermore, this embodiment also provides a data-driven product collaborative design management system, including a data acquisition and conversion module for constructing a dual-stream heterogeneous quantization acquisition network to parallelly acquire product structure parameters and design process data respectively, and convert the heterogeneous data into a unified feature space representation; a feature analysis module for using a dynamic adaptive quantization unit to reduce the dimension of the unified feature space representation, extract key feature combinations, establish a correlation matrix, and calculate the dynamic coupling strength between parameters; a knowledge modeling module for designing a recursive deep belief network based on the correlation matrix and the dynamic coupling strength, encoding the product structure parameters and design process data into a probabilistic graphical model, and constructing a design knowledge base; a decision generation module for dynamically allocating the design rules in the design knowledge base using a swarm intelligence optimization algorithm to generate collaborative decision-making units and establish a constraint propagation link; a solution optimization module for iteratively calculating the collaborative decision-making units on the constraint propagation link, fusing the marginal probability estimation method, generating a multi-objective collaborative optimization solution set, and screening the optimal solution set based on the Pareto front; a verification and adjustment module for performing convergence verification on the optimal solution set, updating the probabilistic graphical model in the design knowledge base, and simultaneously performing online learning and adjustment on the dynamic coupling strength.

[0089] In summary, the present invention realizes the comprehensive acquisition of product structure parameters and design process data through a dual-stream heterogeneous quantization acquisition network, improving data integrity; adopts a recursive deep belief network and a probabilistic graphical model to endow the system with an adaptive learning ability and dynamically update design knowledge; combines a swarm intelligence optimization algorithm and a marginal probability estimation method to achieve multi-objective collaborative optimization, significantly improving the quality of design solutions; introduces a quantum annealing algorithm for convergence verification to ensure the reliability of the optimization results; and dynamically adjusts the coupling strength through an online learning mechanism to enhance the robustness and adaptability of the system.

[0090] Example 2. Refer to Figures 1 to 3 , which is the second embodiment of the present invention. This embodiment provides a data-driven product collaborative design management method. To verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0091] To verify the effectiveness of the method of the present invention, a certain aeroengine blade design project is selected as the experimental object. This project involves complex structural parameter optimization and design process collaboration. The experimental environment uses a high-performance workstation with an Intel Xeon Gold 6338 processor and 512 GB of memory, and the operating system is Ubuntu 22.04 LTS. The software platform uses MATLABR2023b and Python 3.9 for algorithm implementation, and uses the PyTorch 2.0 deep learning framework to build a neural network model.

[0092] In the data acquisition stage, design data for 6 consecutive months are collected from a certain type of engine blade design project, including 4,526 sets of structural parameter records and 3,862 design process data. The structural parameter data cover 62 feature dimensions such as blade geometric features (leading edge radius, trailing edge thickness, twist angle, etc.), material properties (elastic modulus, Poisson's ratio, density, etc.), and stress distribution; the design process data include 46 process features such as task assignment, scheme review, and iterative optimization. The Savitzky-Golay filter is used to preprocess the original data, and after removing outliers, the z-score normalization method is used to normalize the data.

[0093] In the configuration of the dual-stream heterogeneous quantization acquisition network, the ResNet-50 is used as the basic network structure for the product structure parameter acquisition channel, and the output dimension of the last fully connected layer is adjusted to 256. The bidirectional GRU network is used for the design process data acquisition channel, the hidden layer dimension is 256, the time step is set to 32, and the dropout rate is 0.4. In the tensor decomposition module, the core tensor rank of Tucker decomposition is set to (64, 48, 32), the feature projection function uses a Gaussian kernel function, and the bandwidth parameter is adaptively determined by the Silverman criterion.

[0094] To scientifically evaluate the performance of the method of the present invention, a widely used deep learning method is selected as the control group for comparative experiments, as shown in Table 1.

[0095] Table 1 Performance Comparison Table From the comparison results in Table 1, it can be seen that the method of the present invention is superior to the traditional deep learning method in all key performance indicators. Especially in terms of the convergence time of the design scheme and the number of design iterations, the method of the present invention has achieved significant improvements of 63.6% and 59.0% respectively, which is mainly due to the efficient processing of structural parameters and process data by the dual-stream heterogeneous quantization acquisition network, and the accurate feature extraction of the dynamic adaptive quantization unit.

[0096] In terms of the configuration of the dynamic adaptive quantization unit, the adaptive threshold layer adopts a multi-level threshold segmentation method based on the OTSU algorithm, the initial number of segments is set to 12, and the information entropy threshold is 0.82. The dimension reconstruction layer uses the UMAP algorithm for non-linear dimensionality reduction, the number of neighbors is set to 15, and the minimum distance is 0.1. The mutual information threshold of the feature selection layer is set to 0.65, and the improved bat algorithm is used to optimize the feature weights, the population size is 150, and the maximum number of iterations is 100.

[0097] The recursive deep belief network adopts a 5-layer structure, and the number of hidden layer nodes of each restricted Boltzmann machine is 1024, 512, 256, 128, and 64 respectively. The Adam optimizer is used, and the learning rate is 0.0008. , . In the hybrid swarm intelligence optimization algorithm, the number of ants in the ant colony optimization unit is set to 300, the pheromone evaporation coefficient is 0.15, and the global update intensity is 0.1; the number of particles in the particle swarm optimization unit is 200, and a non-linear decreasing inertia weight strategy is adopted, and both the social cognitive factor and the individual cognitive factor are 2.05.

[0098] The finally generated multi-objective collaborative optimization scheme shows that the blade aerodynamic efficiency has increased by 9.2%, the structural reliability has increased by 7.8%, the vibration characteristics have improved by 5.6%, and the design cycle has been shortened by 52%. Through the quantum annealing algorithm for convergence verification, the initial temperature is set to 150, the termination temperature is 0.001, and the cooling coefficient is 0.96. After 186 iterations, the convergence condition is met. In the online learning process of the dynamic coupling strength, the AdaBelief optimizer is used, the initial learning rate is 0.0003, and the weight decay coefficient is 1e-4. , , the upper limit of parameter correction is 0.25, achieving an accurate characterization and real-time optimization of the coupling relationship. The experimental results fully demonstrate the significant advantages of the method of the present invention in improving design quality, shortening the design cycle, and enhancing system reliability, etc.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A data-driven product collaborative design management method, characterized by: include, Construct a dual-stream heterogeneous quantitative acquisition network to collect product structure parameters and design process data in parallel, and convert heterogeneous data into a unified feature space representation; Using a dynamic adaptive quantization unit to reduce the dimension of the unified feature space representation, extract key feature combinations, establish a correlation matrix and calculate the dynamic coupling strength between parameters; Based on the association matrix and dynamic coupling strength, a recursive deep belief network is designed to encode product structure parameters and design process data into a probabilistic graphical model to construct a design knowledge base; Using a swarm intelligence optimization algorithm to dynamically allocate design rules in the design knowledge base, generate collaborative decision-making units, and establish a constraint propagation link; Iteratively calculate the collaborative decision-making unit on the constraint propagation link, integrate the marginal probability estimation method, generate a multi-objective collaborative optimization solution group, and screen the optimal solution group based on the Pareto front; The convergence of the optimal solution group is verified, the probabilistic graphical model in the design knowledge base is updated, and the dynamic coupling strength is adjusted through online learning.

2. The data-driven product collaborative design management method according to claim 1, characterized in that: The construction of the dual-stream heterogeneous quantization acquisition network includes the following steps: Constructing a dual-channel data acquisition unit, the dual-channel data acquisition unit includes a product structure parameter acquisition channel and a design process data acquisition channel, and configuring the number of input nodes of each channel; Constructing a multi-layer convolutional neural network in the product structure parameter acquisition channel, and obtaining a structural feature tensor through feature extraction and spatial transformation; Constructing a long short-term memory network in the design process data collection channel, and extracting process feature tensors in combination with a multi-head temporal attention mechanism; Performing a hierarchical tensor decomposition operation on the structural feature tensor and the process feature tensor to obtain corresponding core tensors and factor matrices; Constructing a feature projection function according to the core tensor and the factor matrix, mapping features of different dimensions to a unified feature space, and obtaining a unified feature representation matrix; Applying sparse constraints and regularization processing to the unified feature representation matrix, optimizing the unified feature matrix by an alternating direction multiplication method, and obtaining a unified feature space representation; The unified feature space representation is normalized and preprocessed to generate a standardized feature representation.

3. The data-driven product collaborative design management method according to claim 1, characterized in that: The establishment of the correlation matrix and calculation of the dynamic coupling strength between parameters comprises the following steps: Constructing a dynamic adaptive quantization unit, the dynamic adaptive quantization unit comprising an adaptive threshold layer, a dimension reconstruction layer, and a feature selection layer, taking the standardized feature representation as input; The standardized feature representation is segmented and quantized using an adaptive threshold layer, the optimal segmentation point is calculated based on the information entropy criterion, and the upper and lower limit thresholds of the quantization interval are dynamically adjusted to obtain a quantized feature sequence; In the dimension reconstruction layer, a manifold learning method is used to perform nonlinear dimensionality reduction on the quantized feature sequence, and the dominant feature direction is determined by eigenvalue decomposition to generate a dimensionality reduction feature combination; The mutual information criterion is introduced into the dimensionality reduction feature combination through the feature selection layer, the key feature subset is screened, the feature weight is optimized by using the swarm intelligence algorithm, and the screened key feature subset is reorganized into a multi-dimensional feature combination sequence; Calculate the correlation between product structure parameters and design process data based on the multi-dimensional feature combination sequence, and construct a two-dimensional correlation matrix in combination with expert evaluation weights; The elements of the correlation matrix are dynamically weighted, the time series sliding window method is used to analyze the parameter variation trend, and the dynamic coupling strength index between parameters is calculated.

4. The data-driven product collaborative design management method according to claim 1, characterized in that: The encoding of product structure parameters and design process data into a probabilistic graphical model comprises the following steps: Construct a recursive deep belief network structure, including a feature encoding layer, a multi-layer restricted Boltzmann machine unit, and a recursive connection layer; The association matrix is ​​input into the feature coding layer, and the matrix data is reconstructed through the encoding and decoding operation of the autoencoder to obtain the initial feature vector containing the rule importance parameter; Performing contrast divergence calculations in the multi-layer restricted Boltzmann machine unit, setting network connection weights according to the dynamic coupling strength values, and generating inter-layer probability distributions; In the recursive connection layer, a time series signal sequence is introduced, and a back propagation operation is used to adjust the network parameters to generate a probabilistic graph model of product structure parameters and design process data; Performing Bayesian reasoning on the probability graph model, constructing a design rule chain, and generating a design rule set according to a confidence threshold; The design rule set is divided into levels according to the similarity of the graph structure, and a hierarchical retrieval mechanism based on a hierarchical hash index is established to form a design knowledge base.

5. The data-driven product collaborative design management method according to claim 1, characterized in that: The establishment of the constraint propagation link comprises the following steps: Constructing a dual-unit integrated hybrid swarm intelligence optimization algorithm module, wherein the hybrid swarm intelligence optimization algorithm module includes an ant colony optimization unit and a particle swarm optimization unit; Extracting a design rule set from a design knowledge base, and dynamically allocating the design rule set into a plurality of rule subsets according to rule attribute values; Constructing collaborative decision units based on the rule subset, each collaborative decision unit includes an associated design rule combination and its constraint conditions, forming an initial collaborative decision unit set; The ant colony optimization unit sets the pheromone strength value based on the rule importance parameter, establishes a multi-layer pheromone distribution matrix, optimizes the organizational structure of the collaborative decision-making unit set through path iteration operation, and obtains the collaborative decision-making unit organizational structure matrix; The particle swarm optimization unit maps the organizational structure matrix to the decision space, calculates the velocity vector and position vector of each collaborative decision unit according to a preset number of iterations, and obtains a probability distribution matrix in combination with the constraint conditions of the collaborative decision unit; Calculating the interdependence coefficients of the collaborative decision-making units according to the probability distribution matrix, arranging the interdependence coefficients into a probability transfer matrix, and constructing a constraint network including node weights; Connecting the collaborative decision-making unit nodes whose dependency coefficients satisfy the threshold condition in the constraint network to form a directed acyclic constraint graph; The collaborative decision-making unit nodes in the directed acyclic constraint graph are sorted according to the in-degree values, the node numbers are marked, and a constraint propagation link table is generated.

6. The data-driven product collaborative design management method according to claim 1, characterized in that: The generating of the multi-objective collaborative optimization solution group comprises the following steps: Construct a decision tree structure based on the node order in the constraint propagation link table, and set the node splitting threshold according to historical data; Adopting adaptive sampling strategy and dimensionality reduction mechanism, recording the value range of decision variables and generating candidate solution space; Applying a marginal probability estimation algorithm to the candidate solution space, adjusting the kernel density estimation bandwidth, calculating the probability density distribution, and determining a high probability decision region; Constructing a multi-objective optimization model in the high-probability decision-making area, determining the objective weights using the analytic hierarchy process, and setting optimization constraints; Combine the crossover operator and the local search operator to generate the Pareto optimal solution set and calculate the objective function value of the non-dominated solution; The non-dominated solutions are hierarchically sorted based on the Pareto front, and the optimal solution group is screened and the comprehensive score is calculated based on sensitivity analysis.

7. The data-driven product collaborative design management method according to claim 1, characterized in that: The updating of the probability graph model in the design knowledge base comprises the following steps: Construct a qubit matrix, encode the optimal solution group into a quantum state sequence through phase encoding, and set adaptive annealing parameters; Perform quantum annealing iterative operations and local search optimization, record the change process of the objective function value, and judge the convergence of the solution group; Update the convergence verification results to the probabilistic graph model, and incrementally update the node probability distribution and edge probability distribution in the design knowledge base; Extract design rule feature vectors based on information gain, calculate rule importance, and adjust the structure of the design knowledge base; A composite statistical method is used to calculate the dynamic coupling strength correction and perform restricted gradient update; The corrected dynamic coupling strength is piecewise normalized, the correlation matrix is ​​updated, and the optimization iteration is completed.

8. A data-driven product collaborative design management system, based on the data-driven product collaborative design management method according to any one of claims 1 to 7, characterized in that: Also includes, The data acquisition conversion module is used to build a dual-stream heterogeneous quantitative acquisition network to collect product structure parameters and design process data in parallel, and convert heterogeneous data into a unified feature space representation; A feature analysis module, used to reduce the dimension of the unified feature space representation using a dynamic adaptive quantization unit, extract key feature combinations, establish a correlation matrix and calculate the dynamic coupling strength between parameters; A knowledge modeling module is used to design a recursive deep belief network based on the association matrix and the dynamic coupling strength, encode product structure parameters and design process data into a probabilistic graphical model, and construct a design knowledge base; A decision generation module, used to dynamically allocate the design rules in the design knowledge base using a swarm intelligence optimization algorithm, generate collaborative decision units, and establish a constraint propagation link; A solution optimization module, used to iteratively calculate the collaborative decision-making unit on the constraint propagation link, integrate the marginal probability estimation method, generate a multi-objective collaborative optimization solution group, and screen the optimal solution group based on the Pareto front; The verification and adjustment module is used to verify the convergence of the optimal solution group, update the probabilistic graphical model in the design knowledge base, and perform online learning adjustment on the dynamic coupling strength.

Images