Die casting parameter optimization and processing method and system based on orthogonal experiment

Abstract

The present invention provides a die-casting parameter optimization and processing method and system based on orthogonal experiment, which relates to the field of die-casting optimization technology, including calculating the importance of die-casting parameters through a clustering algorithm based on a die-casting process knowledge base and die-casting production history data, obtaining a die-casting parameter set, constructing a nonlinear mapping model between die-casting parameters and corresponding value ranges through a gradient boosting decision tree algorithm, and generating a die-casting parameter optimization level set in combination with the importance of die-casting parameters through a die-casting parameter optimization algorithm; generating an intelligent orthogonal experiment plan through a preset orthogonal experiment design method, controlling die-casting equipment to perform multiple batches of experiments, obtaining experimental data, and optimizing the die-casting parameter configuration strategy through a deep reinforcement learning model to obtain the optimal die-casting parameter setting, and producing the optimal die-casting workpiece through the die-casting equipment.

Description

Technical Field

[0001] The invention relates to the technical field of die-casting optimization, and in particular to a die-casting parameter optimization and processing method and system based on orthogonal experiment. Background Art

[0002] Die casting is a common metal parts forming process, widely used in the automotive, aviation, electronics and other industries. The selection and optimization of die casting process parameters have an important impact on the quality and production efficiency of die castings. At present, the optimization of die casting parameters mainly depends on the experience of process personnel and repeated tests, which is very blind and uncertain. It is difficult to obtain the optimal parameter combination, resulting in unstable quality of die castings and high production costs.

[0003] In the prior art, attempts have been made to optimize die casting parameters using orthogonal experimental design methods. Orthogonal experimental design can obtain more comprehensive test data with fewer tests by reasonably arranging the number of tests and parameter level combinations, and determine the optimal value of each parameter by analyzing the test results. In addition, most of the existing die casting parameter optimization methods optimize a single quality indicator.

[0004] In summary, the orthogonal test design of the prior art mainly relies on manual experience to select test factors and levels, and the test results are analyzed by simple intuitive analysis methods, which makes it difficult to accurately obtain the interaction and nonlinear effects between parameters, and the optimization effect is limited. At the same time, there are many uncertain factors and interferences in the die-casting production process, such as raw material quality fluctuations, equipment status changes, etc., which make it difficult for the optimized parameter configuration to achieve the expected effect in actual production. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a die-casting parameter optimization and processing method and system based on orthogonal experiment, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a die casting parameter optimization and processing method based on orthogonal experiment, including:

[0008] Based on the die-casting process knowledge base and die-casting production history data, the importance of die-casting parameters is calculated through clustering algorithm to obtain the die-casting parameter set. The nonlinear mapping model between die-casting parameters and corresponding value ranges is constructed through gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through die-casting parameter optimization algorithm to generate the die-casting parameter optimization level set.

[0009] Based on the die-casting parameter optimization level set, an intelligent orthogonal test plan is generated through a preset orthogonal test design method, the die-casting equipment is controlled to conduct multiple batch tests, test data is obtained, and the die-casting parameter configuration strategy is optimized through a deep reinforcement learning model to obtain the optimal die-casting parameter settings, and the optimal die-casting workpiece is produced through the die-casting equipment;

[0010] During the production process of die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through an incremental learning algorithm to generate updated experience knowledge, which is accumulated in the die-casting process knowledge graph to form a die-casting knowledge update mechanism.

[0011] In an optional embodiment,

[0012] Based on the die-casting process knowledge base and die-casting production history data, the core die-casting parameters are extracted through clustering algorithm, and the die-casting parameter set is obtained, including:

[0013] Obtain die-casting production history data, clean and standardize it, and obtain die-casting parameter numerical data;

[0014] Based on the die-casting parameter numerical data, by introducing a multi-perspective similarity measurement mechanism, the similarity between samples in the die-casting parameter numerical data is calculated from three perspectives: numerical similarity, trend similarity and semantic similarity, and a multi-perspective die-casting parameter similarity matrix is ​​constructed;

[0015] Based on the quality indicators in the die-casting production history data, an adaptive fusion strategy is determined, and the weights of the multi-view die-casting parameter similarity matrix are dynamically adjusted to obtain a dynamic die-casting parameter similarity matrix through fusion;

[0016] Performing spectral decomposition on the dynamic die-casting parameter similarity matrix to obtain a low-dimensional embedding representation of the die-casting parameters, and mapping the numerical data of the die-casting parameters to a corresponding low-dimensional embedding space of the die-casting parameters;

[0017] In the die-casting parameter low-dimensional embedding space, based on a preset initial cluster number range, the silhouette coefficient corresponding to the data sample is calculated, and the cluster number range of the clustering is determined according to the change of the silhouette coefficient;

[0018] Traversing each potential cluster number within the range of the cluster number;

[0019] Based on the local edge-preserving sampling strategy, data samples in the local neighborhood are selected in the low-dimensional embedding space of the die-casting parameters to construct a local similarity matrix, and the local similarity matrix is ​​reconstructed in combination with the random walk mechanism to construct a compressed Laplace matrix. Based on the compressed Laplace matrix, the spectral embedding representation of the die-casting parameters is obtained through spectral decomposition, and K-means clustering is performed. By maximizing the silhouette coefficient under the corresponding number of clusters, the local optimal clustering result under the current number of clusters is obtained;

[0020] Compare the local optimal clustering results under different cluster numbers, select the local optimal clustering result with the largest silhouette coefficient, and obtain the global optimal clustering result;

[0021] Based on the global optimal clustering results, the preset Shapley value method is used to calculate the contribution of the die-casting parameters in each cluster to the clustering results, and the importance ranking results of the die-casting parameters in the cluster are generated;

[0022] Based on the ranking results of the importance of die-casting parameters within a cluster, an inter-cluster similarity matrix is ​​constructed. Based on the inter-cluster similarity matrix, a cluster importance propagation network is constructed. By adjusting the importance propagation intensity, the propagation fusion of the importance of die-casting parameters between clusters is achieved, and the global importance of die-casting parameters is obtained.

[0023] According to the global importance of die-casting parameters, the die-casting parameters in the global optimal clustering results are sorted in descending order, duplicate parameters are removed, and the die-casting parameter set is output.

[0024] In an optional embodiment,

[0025] Through the gradient boosting decision tree algorithm, a nonlinear mapping model between die casting parameters and corresponding value ranges is constructed, including:

[0026] Based on the die-casting parameter set and in combination with the die-casting production history data, a die-casting parameter sample set is determined, wherein the die-casting parameter sample set includes a die-casting parameter value vector of a plurality of die-casting parameter corresponding value combinations and a corresponding die-casting workpiece quality;

[0027] Initializing a strong learner, and determining a predicted value of the strong learner based on a mass mean of the die-cast workpiece;

[0028] For each sample in the die-casting parameter sample set, the negative gradient under the current strong learner is calculated to determine the residual estimate, and based on the die-casting parameter value vector and the corresponding residual estimate, a CART regression tree is generated by least squares fitting to obtain the leaf node area of ​​the CART regression tree, and for the leaf node area, based on a preset loss function, the minimum value of the loss function is determined, and the best fitting value is calculated, and based on the best fitting value, the strong learner is updated to re-determine the new prediction value;

[0029] Repeat the strong learner update until the preset number of iterations is reached, obtain the die-casting parameter value range, and build a nonlinear mapping model of die-casting parameters and values.

[0030] In an optional embodiment,

[0031] The die casting parameter optimization algorithm includes:

[0032] The die-casting parameter set is divided into a plurality of sub-populations, each sub-population corresponds to a type of die-casting parameter, each sub-population is initialized, the die-casting parameter value is randomly determined in combination with the die-casting parameter value range, the initial sub-population is determined, and the representative individual is randomly determined;

[0033] Process all initial subpopulations to determine the current subpopulation, combine the individuals in the current subpopulation with the representative individuals of the remaining subpopulations, substitute them into the nonlinear mapping model of die-casting parameters and values, calculate the fitness values ​​corresponding to the individuals in the current subpopulation, calculate the novelty values ​​corresponding to each individual in the current subpopulation based on neighborhood comparison, determine the individual evaluations in the current subpopulation based on the fitness values ​​and novelty values, and determine the new representative individuals after weighted calculation according to the preset weights;

[0034] Iterate the individuals of each initial subpopulation according to the preset crossover probability and mutation probability; when the individuals in an initial subpopulation do not meet the preset diversity index, reinitialize the corresponding number of individuals according to the preset number of new individuals and add them to the corresponding initial subpopulation;

[0035] During the subpopulation iteration process, two initial subpopulations are randomly selected, and the representative individual of one initial subpopulation is migrated to the other initial subpopulation to replace the individual with the lowest individual evaluation in the other initial subpopulation;

[0036] Until the preset number of iterations is reached, all final sub-populations are determined, representative individuals in each final sub-population are selected, and the optimal level combination of each die-casting parameter is determined.

[0037] In an optional embodiment,

[0038] Based on the die casting parameter optimization level set, the intelligent orthogonal test scheme is generated through the preset orthogonal test design method, including:

[0039] Based on the die-casting parameter optimization level set, the factors and levels of the orthogonal test are determined, where the factors are various die-casting parameters and the levels are the value ranges of the die-casting parameters. Combined with the preset die-casting process limitation knowledge, a matching orthogonal table is selected;

[0040] Based on the association rule mining algorithm, the association strength between the die-casting parameters is determined by calculating the support, confidence and lift of the die-casting parameter combinations, and the optimization difficulty corresponding to each die-casting parameter is determined in combination with the preset process constraints. According to the association strength and the optimization difficulty, in combination with the preset association strength threshold and the preset optimization difficulty threshold, the column distribution of each factor in the orthogonal table is adjusted to obtain an adjusted orthogonal table, and the parameter value combination in the die-casting parameter optimization level concentration is filled into the adjusted orthogonal table to generate an initial orthogonal test plan;

[0041] The initial orthogonal test scheme is encoded as the initial population of the neighborhood search algorithm. According to the objective function value of the individuals in the initial population, the corresponding quality evaluation index is calculated, and the quality evaluation index optimization weight vector is adaptively generated. For the individuals in the initial population, hyperspherical neighborhoods of different radii are generated according to the values ​​of the die-casting parameters, and new solutions are generated by sampling. The weight vector is optimized in combination with the quality evaluation index, and the weighted quality evaluation index is calculated. The new solution and the original solution are compared by non-dominated sorting, and the maximum value between the two is retained.

[0042] The initial population is updated repeatedly until the preset number of iterations is reached, and an intelligent orthogonal test plan is output.

[0043] In an optional embodiment,

[0044] Control the die-casting equipment to conduct multiple batch tests, obtain test data, and optimize the die-casting parameter configuration strategy through the deep reinforcement learning model to obtain the optimal die-casting parameter settings. The production of the optimal die-casting workpiece through the die-casting equipment includes:

[0045] Based on the intelligent orthogonal test scheme, the initial parameter configuration of the die-casting equipment is set, the die-casting equipment is controlled to conduct multiple batches of test production, and the quality evaluation data of the die-casting workpieces under different die-casting parameter configurations is obtained;

[0046] The die-casting parameter configuration is mapped to the state space of the deep reinforcement learning model, the adjustment actions corresponding to the die-casting parameters are mapped to the action space, and the die-casting workpiece quality evaluation data is mapped to the reward function to construct a Markov decision process for die-casting parameter optimization.

[0047] Fit the state-action value function through a deep neural network, use the Q-learning algorithm to train the deep reinforcement learning model, and learn the die-casting parameter configuration strategy;

[0048] The ε-greedy strategy is used to balance exploration and exploitation, and the experience replay mechanism is combined for model training to obtain the optimal deep reinforcement learning model;

[0049] Based on the optimal deep reinforcement learning model, the current die-casting process status is input to obtain the optimal die-casting parameter configuration strategy;

[0050] The optimal die-casting parameter configuration strategy is decoded into parameter setting instructions executable by the die-casting equipment, and the die-casting equipment is controlled to produce the optimal die-casting workpiece.

[0051] In an optional embodiment,

[0052] In the production process of die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through the incremental learning algorithm to generate updated experience knowledge, which is accumulated into the die-casting process knowledge graph to form a die-casting knowledge update mechanism including:

[0053] In the production process of die-casting equipment, die-casting parameter data and die-casting quality data are collected in real time and organized into sample data of state-action-reward triples in time series;

[0054] The newly collected sample data is added to the original training data set to form an incremental data set. The deep reinforcement learning model is adjusted based on the incremental data set using an incremental learning algorithm to obtain an updated deep reinforcement learning model.

[0055] The incremental learning algorithm determines the experience replay priority according to the time axis position and error size of the sample data, and sets a forgetting factor to eliminate expired samples;

[0056] Extract relational implicit knowledge from the updated deep reinforcement learning model to generate die-casting updated experience knowledge, formalize the updated experience knowledge into ontology, rules and cases, integrate it into the original die-casting process knowledge graph, and determine the dynamic evolution of the die-casting process knowledge graph, including ontology fusion, rule reasoning and case update;

[0057] Based on natural language processing, the die-casting process knowledge graph is structured.

[0058] According to a second aspect of the embodiments of the present invention,

[0059] Provided is a die casting parameter optimization and processing system based on orthogonal experiment, including:

[0060] The first unit is used to calculate the importance of die-casting parameters based on the die-casting process knowledge base and die-casting production history data through a clustering algorithm to obtain a die-casting parameter set, and to construct a nonlinear mapping model between die-casting parameters and corresponding value ranges through a gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through a die-casting parameter optimization algorithm to generate an optimized level set of die-casting parameters;

[0061] The second unit is used to generate intelligent orthogonal test schemes based on the die-casting parameter optimization level set through a preset orthogonal test design method, control the die-casting equipment to conduct multiple batch tests, obtain test data, and optimize the die-casting parameter configuration strategy through a deep reinforcement learning model to obtain the optimal die-casting parameter settings, and produce the optimal die-casting workpiece through the die-casting equipment;

[0062] The third unit is used to collect data in real time during the production process of the die-casting equipment and dynamically update the deep reinforcement learning model through an incremental learning algorithm to generate updated experience knowledge, accumulate the updated experience knowledge into the die-casting process knowledge graph, and form a die-casting knowledge update mechanism.

[0063] According to a third aspect of the embodiments of the present invention,

[0064] An electronic device is provided, comprising:

[0065] processor;

[0066] a memory for storing processor-executable instructions;

[0067] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] According to a fourth aspect of the embodiments of the present invention,

[0069] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0070] In the embodiment of the present invention, by calculating and analyzing the importance of die-casting parameters, the parameters that have the greatest impact on the quality of the workpiece can be effectively identified and optimized preferentially, avoiding the disadvantages of blindly adjusting all parameters; through the combination of optimized parameter levels, the quality of the workpiece can be significantly improved without increasing the production cost, while reducing the scrap rate and process instability; using a deep reinforcement learning model, through the analysis and learning of multiple batches of test data, the die-casting parameter configuration strategy is gradually optimized to ensure that the parameter setting can be continuously improved; combining the test data and the reinforcement learning model, the optimal die-casting parameter setting is finally obtained, which improves the accuracy of production and the stability of the process; by With the application of intelligent parameter optimization and optimal parameter setting, die-casting equipment can produce higher quality workpieces, reduce defects, and improve product consistency; through real-time data collection and incremental learning, the deep reinforcement learning model can dynamically adapt to changes in the production process, adjust die-casting parameters in time, and improve the flexibility and response speed of the production process; the constantly updated empirical knowledge is accumulated in the die-casting process knowledge graph, so that the die-casting process can be gradually optimized in actual production, improving workpiece quality and production efficiency; through the knowledge update mechanism, the die-casting process knowledge graph is gradually enriched, providing more accurate parameter settings and decision support for subsequent production, reducing the need for experimental adjustments. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic flow chart of a die-casting parameter optimization and processing method based on orthogonal experiment according to an embodiment of the present invention;

[0072] Figure 2 It is a schematic diagram of the structure of the die-casting parameter optimization and processing system based on orthogonal experiment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of a die casting parameter optimization and processing method based on orthogonal experiment according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S101. Based on the die-casting process knowledge base and die-casting production history data, the importance of die-casting parameters is calculated through a clustering algorithm to obtain a die-casting parameter set. A nonlinear mapping model between die-casting parameters and corresponding value ranges is constructed through a gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through a die-casting parameter optimization algorithm to generate an optimized level set of die-casting parameters;

[0077] The die-casting parameters specifically refer to a series of technical parameters that need to be set and controlled during the die-casting process. These parameters directly affect the quality, production efficiency and cost of the final workpiece. Common die-casting parameters include pouring temperature, injection speed, holding time, holding pressure, mold temperature, etc. The setting of die-casting parameters needs to be optimized according to the material properties, geometric structure and process requirements of the workpiece to ensure the production of products that meet quality requirements.

[0078] The die casting parameter importance specifically refers to the degree of influence of each die casting parameter on the final workpiece quality and the production process. Through clustering algorithms and production history data, the relative importance of different parameters in the entire production process can be quantified. By calculating the parameter importance, it is possible to identify which parameters are critical to the process results, so that these high-importance parameters can be given priority in the parameter optimization process.

[0079] The optimal level combination specifically refers to the best setting values ​​of each die-casting parameter determined during the die-casting process after optimization algorithm and parameter importance analysis. The setting values ​​are combined together to produce the best workpiece quality and production efficiency under specific die-casting conditions. The optimal level combination represents the die-casting parameter settings obtained through comprehensive analysis and optimization under the current production environment and conditions, which can maximize product quality, reduce scrap rate, and optimize production efficiency.

[0080] In this embodiment, by calculating and analyzing the importance of die-casting parameters, the parameters that have the greatest impact on the quality of the workpiece can be effectively identified and optimized preferentially, avoiding the disadvantages of blindly adjusting all parameters; through the combination of optimized parameter levels, the quality of the workpiece can be significantly improved without increasing production costs, while reducing scrap rate and process instability; it integrates data-driven and knowledge-driven optimization methods, embodies the idea of ​​intelligent manufacturing, and helps to realize the automation and intelligent regulation of the die-casting process.

[0081] In an optional embodiment, based on the die-casting process knowledge base and the die-casting production history data, the die-casting core parameters are extracted by a clustering algorithm to obtain a die-casting parameter set including:

[0082] Obtain die-casting production history data, clean and standardize it, and obtain die-casting parameter numerical data;

[0083] Based on the die-casting parameter numerical data, by introducing a multi-perspective similarity measurement mechanism, the similarity between samples in the die-casting parameter numerical data is calculated from three perspectives: numerical similarity, trend similarity and semantic similarity, and a multi-perspective die-casting parameter similarity matrix is ​​constructed;

[0084] Based on the quality indicators in the die-casting production history data, an adaptive fusion strategy is determined, and the weights of the multi-view die-casting parameter similarity matrix are dynamically adjusted to obtain a dynamic die-casting parameter similarity matrix through fusion;

[0085] Performing spectral decomposition on the dynamic die-casting parameter similarity matrix to obtain a low-dimensional embedding representation of the die-casting parameters, and mapping the numerical data of the die-casting parameters to a corresponding low-dimensional embedding space of the die-casting parameters;

[0086] In the die-casting parameter low-dimensional embedding space, based on a preset initial cluster number range, the silhouette coefficient corresponding to the data sample is calculated, and the cluster number range of the clustering is determined according to the change of the silhouette coefficient;

[0087] Traversing each potential cluster number within the range of the cluster number;

[0088] Based on the local edge-preserving sampling strategy, data samples in the local neighborhood are selected in the low-dimensional embedding space of the die-casting parameters to construct a local similarity matrix, and the local similarity matrix is ​​reconstructed in combination with the random walk mechanism to construct a compressed Laplace matrix. Based on the compressed Laplace matrix, the spectral embedding representation of the die-casting parameters is obtained through spectral decomposition, and K-means clustering is performed. By maximizing the silhouette coefficient under the corresponding number of clusters, the local optimal clustering result under the current number of clusters is obtained;

[0089] Compare the local optimal clustering results under different cluster numbers, select the local optimal clustering result with the largest silhouette coefficient, and obtain the global optimal clustering result;

[0090] Based on the global optimal clustering results, the preset Shapley value method is used to calculate the contribution of the die-casting parameters in each cluster to the clustering results, and the importance ranking results of the die-casting parameters in the cluster are generated;

[0091] Based on the ranking results of the importance of die-casting parameters within a cluster, an inter-cluster similarity matrix is ​​constructed. Based on the inter-cluster similarity matrix, a cluster importance propagation network is constructed. By adjusting the importance propagation intensity, the propagation fusion of the importance of die-casting parameters between clusters is achieved, and the global importance of die-casting parameters is obtained.

[0092] According to the global importance of die-casting parameters, the die-casting parameters in the global optimal clustering results are sorted in descending order, duplicate parameters are removed, and the die-casting parameter set is output.

[0093] Obtain die-casting production history data, including die-casting process parameter setting values ​​and corresponding die-casting quality inspection results. Clean the acquired data, remove missing values, outliers and other noise data, and standardize the data to unify die-casting parameters of different dimensions to the same numerical scale to obtain numerical data of die-casting parameters.

[0094] Based on the numerical data of die-casting parameters, the numerical similarity, trend similarity and semantic similarity between different die-casting parameters are calculated. The numerical similarity calculates the closeness of the die-casting parameter values ​​through metrics such as Euclidean distance; the trend similarity calculates the consistency of the die-casting parameter change trend through metrics such as Pearson correlation coefficient; the semantic similarity calculates the semantic relevance of the die-casting parameter names through metrics such as cosine similarity. The three similarity calculation results are combined to form a multi-perspective die-casting parameter similarity matrix.

[0095] According to the die-casting quality inspection results in the die-casting production history data, such as surface defects, internal defects, dimensional accuracy, etc., the weight adjustment strategy of each perspective similarity matrix is ​​determined. When the die-casting quality performance is good, the weight of the trend similarity matrix is ​​increased; when the die-casting has specific defects, the weight of the corresponding perspective similarity matrix is ​​increased. By weighted fusion of multi-perspective die-casting parameter similarity matrices, a dynamic die-casting parameter similarity matrix is ​​obtained.

[0096] The dynamic die-casting parameter similarity matrix is ​​eigendecomposed, and the low-dimensional embedding representation of the die-casting parameters is obtained by calculating the eigenvalues ​​and eigenvectors. The original high-dimensional die-casting parameter numerical data is mapped to the corresponding low-dimensional embedding space.

[0097] In the low-dimensional embedding space of die casting parameters, set the initial cluster number range, such as 2 to 20. For each potential cluster number within the range, calculate the silhouette coefficient of the data sample. The silhouette coefficient measures the similarity between the sample and the cluster to which it belongs and the difference between it and other clusters. The larger the value, the better the clustering effect. According to the change of the silhouette coefficient with the number of clusters, such as the occurrence of inflection points or stable intervals, determine the final evaluation range of the number of clusters.

[0098] Each potential cluster number within the cluster number evaluation range is traversed to perform local edge-preserving spectral clustering. Specifically, in the low-dimensional embedding space of die-casting parameters, for each data sample, data samples in its local neighborhood are selected to construct a local similarity matrix. The local similarity matrix is ​​reconstructed through a random walk mechanism to obtain a compressed Laplace matrix. The compressed Laplace matrix can enhance the preservation of local structure and improve the robustness of clustering.

[0099] The compressed Laplace matrix is ​​spectrally decomposed to obtain the spectral embedding representation of the die-casting parameters. In the spectral embedding space, the K-means clustering algorithm is applied to divide the die-casting parameter data samples into different clusters under the current number of clusters. By maximizing the silhouette coefficient corresponding to the clustering result, the local optimal clustering division under the current number of clusters is obtained.

[0100] The local optimal clustering results under different numbers of clusters are compared, and the clustering result with the largest silhouette coefficient is selected as the global optimal clustering division of die-casting parameters.

[0101] Based on the global optimal clustering division, the contribution of die-casting parameters within each cluster to the clustering results is calculated. The Shapley value method is used to quantify the importance of die-casting parameters to the clustering results by considering their marginal contribution and combined contribution. The die-casting parameters within each cluster are sorted from high to low according to the Shapley value to obtain the ranking result of the importance of die-casting parameters within the cluster.

[0102] The similarity between different clusters is calculated by using the ranking results of the importance of die-casting parameters within the cluster, and the inter-cluster similarity matrix is ​​constructed. Based on the inter-cluster similarity matrix, a cluster importance propagation network is constructed, and the propagation and fusion process of importance between clusters is controlled by setting the propagation intensity parameter. After multiple rounds of iterative propagation, the global importance evaluation results of die-casting parameters are obtained.

[0103] According to the global importance of die-casting parameters, the die-casting parameters in the global optimal clustering results are arranged in descending order to obtain the die-casting parameter sorting results from high to low importance. The duplicate parameters in the sorting list are removed, and finally the optimal die-casting parameter set is output.

[0104] In this embodiment, by comprehensively considering numerical similarity, trend similarity and semantic similarity, a more comprehensive parameter association analysis method is provided, which can more accurately reflect the complex relationship between different die-casting parameters; through local edge-preserving spectral clustering and compressed Laplace matrix, the preservation of local data structure is enhanced, the robustness and stability of clustering are improved, and the classification of die-casting parameters is more accurate; the silhouette coefficient is maximized to ensure the global optimality of the clustering results, effectively improve the quality of clustering, and ensure that the final die-casting parameter importance assessment is more accurate; through the Shapley value method, the contribution of die-casting parameters to the clustering results is quantified, and the importance of each parameter can be scientifically and reasonably evaluated; by using the inter-cluster similarity matrix and the importance propagation network, the cross-cluster parameter importance fusion is realized, the integrity and accuracy of the importance assessment are improved, and the final output die-casting parameter set is ensured to have the highest priority in the overall optimization.

[0105] In an optional embodiment, constructing a nonlinear mapping model between die-casting parameters and corresponding value ranges by using a gradient boosting decision tree algorithm includes:

[0106] Based on the die-casting parameter set and in combination with the die-casting production history data, a die-casting parameter sample set is determined, wherein the die-casting parameter sample set includes a die-casting parameter value vector of a plurality of die-casting parameter corresponding value combinations and a corresponding die-casting workpiece quality;

[0107] Initializing a strong learner, and determining a predicted value of the strong learner based on a mass mean of the die-cast workpiece;

[0108] For each sample in the die-casting parameter sample set, the negative gradient under the current strong learner is calculated to determine the residual estimate, and based on the die-casting parameter value vector and the corresponding residual estimate, a CART regression tree is generated by least squares fitting to obtain the leaf node area of ​​the CART regression tree, and for the leaf node area, based on a preset loss function, the minimum value of the loss function is determined, and the best fitting value is calculated, and based on the best fitting value, the strong learner is updated to re-determine the new prediction value;

[0109] Repeat the strong learner update until the preset number of iterations is reached, obtain the die-casting parameter value range, and build a nonlinear mapping model of die-casting parameters and values.

[0110] Based on the die-casting parameter set selected previously and combined with the die-casting production history data, a die-casting parameter sample set for modeling is determined. Specifically, the parameter values ​​corresponding to the selected die-casting parameters and the corresponding die-casting workpiece quality data are extracted from the die-casting production history data. Each sample consists of a die-casting parameter value vector and a corresponding die-casting workpiece quality value, forming a complete die-casting parameter sample set.

[0111] Initialize the GBDT strong learner model. Set the initial prediction value of the strong learner according to the mean value of the die-casting workpiece quality value in the die-casting parameter sample set. The initial prediction value serves as the starting point of the strong learner in the iterative optimization process.

[0112] For each sample in the die casting parameter sample set, the negative gradient value under the current strong learner prediction value is calculated. The negative gradient represents the residual between the true mass value of the sample and the current prediction value, which is used to measure the degree of fit of the current model. Based on the calculated negative gradient value, the residual estimate of each sample is determined.

[0113] Based on the die casting parameter value vector and the corresponding residual estimation value in the die casting parameter sample set, a CART regression tree is generated by least squares fitting. The CART regression tree divides the sample space into multiple leaf node regions by recursively selecting the optimal segmentation features and segmentation points. Each leaf node region contains a part of the samples and corresponds to a constant prediction value.

[0114] For each leaf node region generated by the CART regression tree, the loss function value of the samples in the region is calculated based on a preset loss function, such as the mean square error loss function. By minimizing the loss function, the best fitting value of the leaf node region is determined, that is, the constant value that can minimize the sum of squared residuals of the samples in the region.

[0115] Based on the best fit value of each leaf node area of ​​the CART regression tree, the prediction value of the strong learner is updated. The best fit value of the leaf node area to which each sample belongs is added to the current prediction value of the strong learner to obtain the updated prediction value of the strong learner. By combining the prediction results of the weak learner (CART regression tree), the prediction performance of the strong learner is gradually optimized.

[0116] Perform multiple rounds of iterations to update the strong learner. Each round of iteration is based on the prediction results of the previous round of the strong learner, calculates the residual estimate, generates a new CART regression tree, and updates the prediction value of the strong learner. The iterative process continues until the preset number of iterations is reached or the predetermined stopping condition is met.

[0117] After completing all iterations, the final strong learner model is obtained. The final strong learner model is actually a combination of multiple CART regression trees, each tree corresponding to a weak learner. The die-casting parameter value vector in the die-casting parameter sample set is input into the final strong learner model to obtain the corresponding die-casting workpiece quality prediction value.

[0118] Based on the final strong learner model, the nonlinear relationship between die-casting parameter values ​​and die-casting workpiece quality is analyzed. A large number of die-casting parameter value combinations are generated by grid search or random sampling of die-casting parameter values, and the strong learner model is used to predict the corresponding die-casting workpiece quality values. According to the prediction results, the value range of each die-casting parameter is determined, and a nonlinear mapping model between die-casting parameters and values ​​is constructed.

[0119] The constructed nonlinear mapping model is used to predict the quality of new die-casting parameter value combinations. The new die-casting parameter values ​​are input into the nonlinear mapping model to obtain the corresponding die-casting workpiece quality prediction value. By analyzing the prediction results, the impact of different die-casting parameter value combinations on the quality of die-casting workpieces can be evaluated, providing decision support for the optimization of die-casting process parameters.

[0120] In this embodiment, the GBDT model can accurately capture the nonlinear relationship between die-casting parameters and workpiece quality; in each round of iteration, the model gradually optimizes the prediction results by calculating the negative gradient (residual) and generating a CART regression tree, so that the fitting ability of the model is continuously enhanced and the prediction accuracy is improved; the strong learner is updated through multiple rounds of iterations, and the relationship model between die-casting parameters and workpiece quality is gradually optimized, thereby improving the stability and reliability of the model; the nonlinear mapping model finally generated can predict the quality of different die-casting parameter combinations, help find the optimal parameter value range, and improve the optimization efficiency of the die-casting process; through this nonlinear mapping model, the impact of new die-casting parameter combinations on workpiece quality can be effectively predicted, which helps to quickly screen out the optimal parameter combination and reduce experimental costs.

[0121] In an optional embodiment, the die casting parameter optimization algorithm includes:

[0122] The die-casting parameter set is divided into a plurality of sub-populations, each sub-population corresponds to a type of die-casting parameter, each sub-population is initialized, the die-casting parameter value is randomly determined in combination with the die-casting parameter value range, the initial sub-population is determined, and the representative individual is randomly determined;

[0123] Process all initial subpopulations to determine the current subpopulation, combine the individuals in the current subpopulation with the representative individuals of the remaining subpopulations, substitute them into the nonlinear mapping model of die-casting parameters and values, calculate the fitness values ​​corresponding to the individuals in the current subpopulation, calculate the novelty values ​​corresponding to each individual in the current subpopulation based on neighborhood comparison, determine the individual evaluations in the current subpopulation based on the fitness values ​​and novelty values, and determine the new representative individuals after weighted calculation according to the preset weights;

[0124] Iterate the individuals of each initial subpopulation according to the preset crossover probability and mutation probability; when the individuals in an initial subpopulation do not meet the preset diversity index, reinitialize the corresponding number of individuals according to the preset number of new individuals and add them to the corresponding initial subpopulation;

[0125] During the subpopulation iteration process, two initial subpopulations are randomly selected, and the representative individual of one initial subpopulation is migrated to the other initial subpopulation to replace the individual with the lowest individual evaluation in the other initial subpopulation;

[0126] Until the preset number of iterations is reached, all final sub-populations are determined, representative individuals in each final sub-population are selected, and the optimal level combination of each die-casting parameter is determined.

[0127] The die-casting parameter set obtained by optimization is divided into multiple sub-populations. Each sub-population corresponds to a type of die-casting parameters, such as temperature parameter sub-population, pressure parameter sub-population, time parameter sub-population, etc. The division of sub-populations is usually carried out according to the physical properties or process characteristics of the die-casting parameters.

[0128] Initialize each subpopulation. According to the value range of each type of die-casting parameter, randomly generate a certain number of parameter value combinations as the initial individuals of the subpopulation. An individual represents a combination of a set of die-casting parameter values. At the same time, randomly select an individual from each subpopulation as the representative individual of the subpopulation.

[0129] All initialized subpopulations are evaluated and updated. For each subpopulation, the individuals in it are combined with representative individuals of other subpopulations to form a complete die-casting parameter value combination. The combined die-casting parameter values ​​are input into the previously constructed die-casting parameter-value nonlinear mapping model, and the die-casting workpiece quality prediction value corresponding to each individual is calculated as the individual fitness value.

[0130] Within each subpopulation, the novelty value of each individual is calculated based on the neighborhood comparison strategy. Novelty reflects the uniqueness and diversity of an individual in a subpopulation. Specifically, for each individual, it is compared with other individuals in its neighborhood, and the number of individuals with similarities below a preset threshold is counted as the novelty value of the individual. The higher the novelty value, the more unique the individual is in the subpopulation.

[0131] Taking into account the fitness value and novelty value of each individual, the comprehensive evaluation index of the individual is calculated according to the preset weighting coefficient. The weighting coefficient reflects the relative importance of fitness and novelty in the evaluation. The higher the comprehensive evaluation index, the better the overall quality of the individual in the subpopulation.

[0132] In each subpopulation, the individual with the highest comprehensive evaluation index is selected as the new representative individual. The representative individual is used to combine and interact with individuals in other subpopulations.

[0133] Genetic operations are performed on individuals in each subpopulation. According to the preset crossover probability, two individuals are randomly selected for crossover operations to generate new individuals. The crossover operation generates individuals with new characteristics by exchanging some gene fragments of two individuals. According to the preset mutation probability, mutation operations are performed on each individual to randomly change some gene values ​​of the individual and introduce new diversity.

[0134] After the genetic manipulation, the diversity level of each subpopulation is evaluated. The average similarity or genetic diversity index between individuals in the subpopulation is calculated. If the diversity level is lower than the preset threshold, it indicates that the subpopulation is becoming homogeneous and new individuals need to be introduced. According to the preset number of new individuals, new individuals are randomly generated and added to the corresponding subpopulation to maintain the diversity of the subpopulation.

[0135] During the subpopulation iteration process, individual migration between subpopulations is performed regularly. Two subpopulations are randomly selected, and the representative individuals of one subpopulation are migrated to the other subpopulation, replacing the individuals with the lowest comprehensive evaluation index in the other subpopulation. Individual migration helps to spread excellent characteristics between subpopulations and promote information exchange between subpopulations.

[0136] Multiple rounds of iterative optimization are performed on all subpopulations. The iterative process continues until the preset number of iterations is reached or the predetermined convergence conditions are met. During the iterative process, individuals are continuously updated and optimized within the subpopulation through fitness evaluation, novelty evaluation, and genetic operations, and information exchange and collaborative optimization are achieved between subpopulations through individual migration.

[0137] After reaching the iteration termination condition, the individual with the highest comprehensive evaluation index is selected from each subpopulation as the optimal solution of the subpopulation. The optimal solutions of all subpopulations are combined to form the optimal level combination of die-casting parameters. The optimal level combination contains the optimal value of each type of die-casting parameter, which represents the parameter setting scheme that can obtain the best die-casting workpiece quality under given constraints.

[0138] Apply the optimal level combination of die-casting parameters to actual die-casting production. Set the various parameters of the die-casting machine according to the optimal combination, and conduct trial production and quality inspection. Through feedback from production practice, the setting of die-casting parameters can be further verified and optimized, and the stability of the die-casting process and product quality can be continuously improved.

[0139] In this embodiment, the die-casting parameter set is divided into multiple sub-populations, and different types of parameters are optimized independently to ensure that each type of parameter can receive sufficient attention and optimization; the diversity of the sub-populations is maintained by calculating the novelty value and introducing new individuals, and the algorithm is prevented from falling into the local optimal solution, thereby improving the effect of the global search. In each round of iteration, the sub-population individuals are continuously optimized through fitness, novelty evaluation and genetic operations, and information exchange and collaborative optimization are promoted through individual migration between sub-populations to achieve progressive optimization of the die-casting parameters; new features and randomness are introduced through crossover and mutation operations in the genetic algorithm, so that the optimization process has strong adaptability and can handle complex nonlinear die-casting parameter relationships; individual migration between sub-populations promotes the propagation of excellent features between different populations, enhances the overall optimization effect, and improves the efficiency of die-casting parameter optimization.

[0140] S102. Based on the die-casting parameter optimization level set, an intelligent orthogonal test plan is generated through a preset orthogonal test design method, the die-casting equipment is controlled to perform multiple batch tests, test data is obtained, and the die-casting parameter configuration strategy is optimized through a deep reinforcement learning model to obtain the optimal die-casting parameter setting, and the optimal die-casting workpiece is produced through the die-casting equipment;

[0141] In this embodiment, the orthogonal experimental design method is used to systematically arrange experiments, reduce the number of parameter combinations that need to be tested, improve experimental efficiency, and ensure the representativeness of each parameter combination; the deep reinforcement learning model is used to gradually optimize the die-casting parameter configuration strategy through analysis and learning of multiple batches of test data to ensure that the parameter settings can be continuously improved; the test data and the reinforcement learning model are combined to ultimately obtain the optimal die-casting parameter settings, thereby improving production accuracy and process stability; through intelligent parameter optimization and the application of optimal parameter settings, the die-casting equipment can produce higher quality workpieces, reduce defects, and improve product consistency.

[0142] In an optional embodiment, based on the die casting parameter optimization level set, generating an intelligent orthogonal test scheme through a preset orthogonal test design method includes:

[0143] Based on the die-casting parameter optimization level set, the factors and levels of the orthogonal test are determined, where the factors are various die-casting parameters and the levels are the value ranges of the die-casting parameters. Combined with the preset die-casting process limitation knowledge, a matching orthogonal table is selected;

[0144] Based on the association rule mining algorithm, the association strength between the die-casting parameters is determined by calculating the support, confidence and lift of the die-casting parameter combinations, and the optimization difficulty corresponding to each die-casting parameter is determined in combination with the preset process constraints. According to the association strength and the optimization difficulty, in combination with the preset association strength threshold and the preset optimization difficulty threshold, the column distribution of each factor in the orthogonal table is adjusted to obtain an adjusted orthogonal table, and the parameter value combination in the die-casting parameter optimization level concentration is filled into the adjusted orthogonal table to generate an initial orthogonal test plan;

[0145] The initial orthogonal test scheme is encoded as the initial population of the neighborhood search algorithm. According to the objective function value of the individuals in the initial population, the corresponding quality evaluation index is calculated, and the quality evaluation index optimization weight vector is adaptively generated. For the individuals in the initial population, hyperspherical neighborhoods of different radii are generated according to the values ​​of the die-casting parameters, and new solutions are generated by sampling. The weight vector is optimized in combination with the quality evaluation index, and the weighted quality evaluation index is calculated. The new solution and the original solution are compared by non-dominated sorting, and the maximum value between the two is retained.

[0146] The initial population is updated repeatedly until the preset number of iterations is reached, and an intelligent orthogonal test plan is output.

[0147] According to the die-casting parameter optimization level set obtained previously, the factors and levels used to construct the orthogonal test scheme are determined. Each factor corresponds to a type of die-casting parameter, such as temperature parameter, pressure parameter, time parameter, etc. The level of each factor corresponds to the value range of the die-casting parameter, that is, the selectable value of each type of parameter in the optimization level set.

[0148] Combined with the preset die-casting process limitation knowledge, select the orthogonal table that matches the number of factors and levels. The die-casting process limitation knowledge includes prior information such as the value constraints of each parameter and the mutual influence between parameters. According to the number of factors and levels, select the orthogonal table that meets the experimental requirements from the standard orthogonal table library.

[0149] The association rule mining algorithm is applied to analyze the association relationship between die-casting parameters. For each pair of die-casting parameter combinations, the support of co-occurrence, the confidence that one parameter also appears when the other parameter appears, and the probability lift of one parameter appearing when the other parameter appears are calculated. The calculation of support, confidence, and lift is based on the historical data of die-casting production.

[0150] Combined with the preset process constraints, the optimization difficulty of each die casting parameter is evaluated. The optimization difficulty takes into account factors such as parameter value range, parameter sensitivity, and the degree of coupling between the parameter and other parameters.

[0151] Considering the correlation strength and optimization difficulty between die casting parameters, the column distribution of factors in the orthogonal table is adjusted. According to the preset correlation strength threshold and optimization difficulty threshold, the parameters with high correlation strength and high optimization difficulty are arranged in the front column of the orthogonal table, and the parameters with low correlation strength and low optimization difficulty are arranged in the back column. By adjusting the column distribution, the experimental plan can explore the influence of key parameters more specifically.

[0152] Fill the parameter value combinations in the die casting parameter optimization level set into the adjusted orthogonal table to generate the initial orthogonal test plan. Each row represents an experimental condition, that is, a combination of die casting parameter values.

[0153] The generated initial orthogonal test scheme is encoded as the initial population of the neighborhood search algorithm. Each test condition corresponds to an individual in the population. The encoding method of the individual preferably uses discrete encoding to represent the value combination of the die casting parameters.

[0154] For each individual in the initial population, the corresponding die-casting workpiece quality prediction value is calculated according to the die-casting parameter-quality nonlinear mapping model. The quality prediction value is used as the objective function value of the individual to evaluate the quality of the test conditions.

[0155] Based on the objective function value of the initial population, the optimization weight vector of the quality evaluation index is adaptively generated. The quality evaluation index includes multiple aspects such as the dimensional accuracy, surface quality, and internal defects of the die-casting workpiece. The optimization weight vector reflects the relative importance of different quality indicators in the optimization target and can be adaptively adjusted according to production needs and historical data.

[0156] For each individual in the initial population, hyperspherical neighborhoods of different sizes are generated according to the die-casting parameter values. The neighborhood size is controlled by the search radius, which reflects the scope of the local search. Within the neighborhood of each individual, new experimental conditions are randomly sampled and generated as new solutions.

[0157] Evaluate the quality performance of each new solution. Input the die casting parameter values ​​corresponding to the new solution into the die casting parameter-quality nonlinear mapping model to calculate its quality prediction value. Optimize the weight vector in combination with the quality evaluation index and calculate the weighted quality evaluation index of the new solution. The weighted quality evaluation index comprehensively considers various quality indicators and provides a comprehensive measure of the quality of the test conditions.

[0158] The new solution is compared with the original solution, and the non-dominated sorting strategy is applied to select the best solution. Non-dominated sorting divides the solution into different non-dominated levels according to the objective function value and constraint violation degree of the solution. For the new solution and the original solution, the non-dominated solution between them is retained as the next generation of individuals.

[0159] The initial population is updated through multiple rounds of iterations. In each round of iteration, a new solution is generated through neighborhood search, and an excellent solution is selected through non-dominated sorting. The iterative process continues until the preset number of iterations is reached or the predetermined convergence condition is met.

[0160] After the iterative optimization is completed, a set of non-dominated solutions is selected from the final population as the output of the intelligent orthogonal test scheme. Each non-dominated solution corresponds to an optimized test condition, that is, a combination of die-casting parameter values. These test conditions can better balance the quality of die-casting workpieces and production efficiency based on various quality indicators and constraints.

[0161] Apply the intelligent orthogonal test scheme to the actual die-casting production test. Set the various parameters of the die-casting machine according to the test scheme, and conduct trial production and quality inspection. By analyzing the test results, the effectiveness of the intelligent orthogonal test scheme can be verified, and the die-casting parameter settings can be further optimized.

[0162] According to the test results and production feedback, the intelligent orthogonal test plan is adjusted and improved as necessary. The test factors and level settings can be adjusted to regenerate the orthogonal test plan for specific quality problems or production bottlenecks. Through iterative optimization and production verification, the die-casting parameter optimization plan is continuously improved to improve the stability of the die-casting process and product quality.

[0163] In this embodiment, an adaptive weight and non-dominated sorting strategy is adopted to comprehensively consider multiple quality indicators, balance various quality requirements of die-casting workpieces, and ensure that the optimization scheme can improve product quality in multiple aspects; through neighborhood search and iterative update, the value combination of die-casting parameters is flexibly adjusted, the local optimal solution is fully explored, and the global optimal solution is continuously approached, thereby improving the effect of parameter optimization; by applying the intelligent orthogonal test scheme to actual production, the scheme is iteratively optimized in combination with production feedback, thereby effectively improving the stability of the die-casting process and product quality, thereby ensuring the feasibility and effectiveness of the optimization scheme in actual production.

[0164] In an optional embodiment, controlling the die-casting equipment to conduct multiple batch tests, obtaining test data, and optimizing the die-casting parameter configuration strategy through a deep reinforcement learning model to obtain the optimal die-casting parameter setting, and producing the optimal die-casting workpiece through the die-casting equipment includes:

[0165] Based on the intelligent orthogonal test scheme, the initial parameter configuration of the die-casting equipment is set, the die-casting equipment is controlled to conduct multiple batches of test production, and the quality evaluation data of the die-casting workpieces under different die-casting parameter configurations is obtained;

[0166] The die-casting parameter configuration is mapped to the state space of the deep reinforcement learning model, the adjustment actions corresponding to the die-casting parameters are mapped to the action space, and the die-casting workpiece quality evaluation data is mapped to the reward function to construct a Markov decision process for die-casting parameter optimization.

[0167] Fit the state-action value function through a deep neural network, use the Q-learning algorithm to train the deep reinforcement learning model, and learn the die-casting parameter configuration strategy;

[0168] The ε-greedy strategy is used to balance exploration and exploitation, and the experience replay mechanism is combined for model training to obtain the optimal deep reinforcement learning model;

[0169] Based on the optimal deep reinforcement learning model, the current die-casting process status is input to obtain the optimal die-casting parameter configuration strategy;

[0170] The optimal die-casting parameter configuration strategy is decoded into parameter setting instructions executable by the die-casting equipment, and the die-casting equipment is controlled to produce the optimal die-casting workpiece.

[0171] According to the intelligent orthogonal test scheme, the initial parameter configuration of the die-casting equipment is set, and the die-casting equipment is controlled to conduct multiple batches of test production. Each batch corresponds to a test condition, that is, a set of die-casting parameter configurations. In the production process of each batch, the parameter settings are kept unchanged, and a certain number of die-casting workpiece samples are produced; the quality of the die-casting workpieces produced in each batch is evaluated, and the die-casting parameter configurations and the corresponding die-casting workpiece quality evaluation data are collected and sorted to form a data set for deep reinforcement learning; the die-casting parameter configurations are mapped to the state space of the deep reinforcement learning model, and each state corresponds to a set of die-casting parameter value combinations. The adjustment actions of the die-casting parameters are mapped to the action space, and each action corresponds to an increase or decrease operation on a certain parameter. The die-casting workpiece quality evaluation data is mapped to a reward function, and the reward value reflects the quality improvement or decrease obtained after taking a specific action in a given state.

[0172] Design a deep neural network structure to fit the state-action value function. The deep neural network takes the die-cast parameter configuration as input and outputs the expected long-term reward of taking each action in that state.

[0173] The deep reinforcement learning model is trained using the Q-learning algorithm. Q-learning is a reinforcement learning algorithm with value function approximation that learns the optimal strategy by iteratively updating the state-action value function. In each training step, an action is selected based on the current state, and after the action is executed, the new state and reward are observed, and the parameters of the deep neural network are updated according to the Q-learning update formula.

[0174] During the training process, the ε-greedy strategy is used to balance exploration and exploitation. Actions are randomly selected with probability ε for exploration, and the optimal action estimated by the current value function is selected with probability 1-ε for exploitation. As training progresses, the value of ε is gradually reduced to encourage the model to make more use of the learned optimal strategy in the later stage.

[0175] The state, action, reward, and next state in each training step are stored in the experience replay buffer. When updating the model parameters, a batch of experience data is randomly sampled from the buffer and used to perform gradient descent optimization on the model. The experience replay mechanism can break the correlation between data and stabilize the training process.

[0176] The deep reinforcement learning model is trained in multiple rounds of iterations. Each iteration includes steps such as interacting with the environment, storing experience data, sampling experience from the buffer, and updating model parameters. The training process continues until the model converges or reaches the preset performance indicators.

[0177] After the training is completed, the optimal deep reinforcement learning model is obtained; the optimal deep reinforcement learning model obtained through training is deployed in the die-casting production system. When the die-casting equipment is in a certain process state, the state is input into the model, and the model outputs the optimal parameter configuration strategy under this state.

[0178] Decode the optimal parameter configuration strategy into parameter setting instructions executable by the die-casting equipment. Generate corresponding equipment control instructions based on the parameter adjustment actions in the strategy, and send the instructions to the control system of the die-casting equipment through the equipment control interface.

[0179] In this embodiment, through the deep reinforcement learning model, the die-casting parameter configuration strategy can be automatically learned and optimized, making parameter adjustment more intelligent and reducing the need for human intervention; the model can perform online learning and optimization based on real-time quality feedback data, adapt to changes in process conditions and external interference in the production process, and maintain the optimal state of the production process; the optimized parameter configuration strategy is directly converted into control instructions executable by the equipment, realizing the automated adjustment of the die-casting equipment, simplifying the production process, and reducing errors caused by manual operation; through the closed-loop feedback mechanism of the production system and the deep reinforcement learning model, continuous optimization of the die-casting parameters is achieved, ensuring quality improvement and continuous improvement of production efficiency in the production process.

[0180] S103. During the production process of the die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through an incremental learning algorithm to generate updated experience knowledge, which is accumulated into the die-casting process knowledge graph to form a die-casting knowledge update mechanism.

[0181] In this embodiment, through real-time data collection and incremental learning, the deep reinforcement learning model can dynamically adapt to changes in the production process, adjust the die-casting parameters in time, and improve the flexibility and response speed of the production process; the continuously updated empirical knowledge is accumulated in the die-casting process knowledge graph, so that the die-casting process can be gradually optimized in actual production, improving workpiece quality and production efficiency; through the knowledge update mechanism, the die-casting process knowledge graph is gradually enriched, providing more accurate parameter settings and decision support for subsequent production, reducing the need for experimental adjustments.

[0182] In an optional embodiment, during the production process of the die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through an incremental learning algorithm to generate updated experience knowledge, and the updated experience knowledge is accumulated into the die-casting process knowledge graph to form a die-casting knowledge update mechanism including:

[0183] In the production process of die-casting equipment, die-casting parameter data and die-casting quality data are collected in real time and organized into sample data of state-action-reward triples in time series;

[0184] The newly collected sample data is added to the original training data set to form an incremental data set. The deep reinforcement learning model is adjusted based on the incremental data set using an incremental learning algorithm to obtain an updated deep reinforcement learning model.

[0185] The incremental learning algorithm determines the experience replay priority according to the time axis position and error size of the sample data, and sets a forgetting factor to eliminate expired samples;

[0186] Extract relational implicit knowledge from the updated deep reinforcement learning model to generate die-casting updated experience knowledge, formalize the updated experience knowledge into ontology, rules and cases, integrate it into the original die-casting process knowledge graph, and determine the dynamic evolution of the die-casting process knowledge graph, including ontology fusion, rule reasoning and case update;

[0187] Based on natural language processing, the die-casting process knowledge graph is structured.

[0188] In the production process of die-casting equipment, die-casting parameter data and die-casting quality data are collected in real time. Die-casting parameter data include injection speed, injection pressure, pouring temperature, mold temperature, etc.; die-casting quality data include dimensional accuracy, surface quality, internal defects, etc. The collected data is organized into sample data of state-action-reward triples according to time series. The state represents the current die-casting process parameter combination, the action represents the adjustment measures taken in this state, and the reward represents the degree of improvement of the die-casting quality after adjustment.

[0189] The newly collected sample data is added to the original training data set to form an incremental data set. The incremental data set contains the original data and the newly added data. The incremental learning algorithm is used to adjust the parameters of the deep reinforcement learning model based on the incremental data set. During the incremental learning process, the experience replay priority is determined according to the time axis position and error size of the sample data. Higher priority is given to recently collected samples that have a greater impact on the model. A forgetting factor is set to eliminate samples that are too old and have little guiding significance for the current process status, so as to improve the model update efficiency and reduce the risk of overfitting. The updated deep reinforcement learning model is obtained through incremental learning, which can better adapt to the current die-casting process status and make more accurate predictions on the process parameter adjustment strategy.

[0190] Extract implicit die-casting process knowledge from the updated deep reinforcement learning model to form die-casting updated experience knowledge. This knowledge includes the relationship between specific process parameter combinations and die-casting quality, the causes of different defects and countermeasures, etc. Formalize the updated experience knowledge into ontology, rules and cases. Ontology describes the key concepts in the die-casting process and the semantic relationship between concepts; rules characterize the logical correspondence between process parameters and between process parameters and defects; cases record typical defect instances and treatment processes. The formalized updated knowledge is integrated into the original die-casting process knowledge graph. Ontology fusion should deal with the semantic mapping relationship between the newly added concepts and the original concepts; rule reasoning should determine whether there is a conflict between the new rules and the original rules and coordinate them; case update should classify the new cases into the corresponding defect categories. Through knowledge fusion, the dynamic evolution of the die-casting process knowledge graph is realized, and the knowledge graph can continuously absorb new knowledge generated in process practice and continuously expand and improve.

[0191] Using natural language processing technology, the fused and updated die-casting process knowledge graph is structured to make it easier for computers to understand and process. The structured representation of the knowledge graph usually takes the form of RDF triples, namely (entity, relationship, entity) or (entity, attribute, attribute value). The concepts and their relationships in the ontology are described in triples, such as (injection speed, influence, filling quality), (injection pressure, acting on, molten metal), etc. The rule knowledge is represented by predicate logic, such as IF (pouring temperature is too high, AND, injection speed is too fast) THEN produces (cold shut defect). The case knowledge is represented by a framework structure, and the framework slots such as defect type, defect description, cause, and solution are designed to organize the case content. The structured knowledge graph can be used to support intelligent applications such as semantic query and knowledge reasoning, and provide decision support for die-casting process optimization.

[0192] In this embodiment, through real-time data collection and incremental learning, the deep reinforcement learning model can continuously update and adapt to the current die-casting process status, adjust parameters in time, and improve production efficiency and product quality; extract new process knowledge and integrate it into the die-casting process knowledge graph, so that the graph continues to evolve, gradually improve and enrich it, and provide more accurate parameter settings and optimization strategies for subsequent production; by introducing the forgetting factor, old data that is no longer instructive is eliminated, reducing the risk of model overfitting, and ensuring that the model always meets the current process requirements.

[0193] Figure 2 FIG. 1 is a schematic diagram of the structure of a die casting parameter optimization and processing system based on orthogonal experiment according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0194] The first unit is used to calculate the importance of die-casting parameters based on the die-casting process knowledge base and die-casting production history data through a clustering algorithm to obtain a die-casting parameter set, and to construct a nonlinear mapping model between die-casting parameters and corresponding value ranges through a gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through a die-casting parameter optimization algorithm to generate an optimized level set of die-casting parameters;

[0195] The second unit is used to generate intelligent orthogonal test schemes based on the die-casting parameter optimization level set through a preset orthogonal test design method, control the die-casting equipment to conduct multiple batch tests, obtain test data, and optimize the die-casting parameter configuration strategy through a deep reinforcement learning model to obtain the optimal die-casting parameter settings, and produce the optimal die-casting workpiece through the die-casting equipment;

[0196] The third unit is used to collect data in real time during the production process of the die-casting equipment and dynamically update the deep reinforcement learning model through an incremental learning algorithm to generate updated experience knowledge, accumulate the updated experience knowledge into the die-casting process knowledge graph, and form a die-casting knowledge update mechanism.

[0197] According to a third aspect of the embodiments of the present invention,

[0198] An electronic device is provided, comprising:

[0199] processor;

[0200] a memory for storing processor-executable instructions;

[0201] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0202] According to a fourth aspect of the embodiments of the present invention,

[0203] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0204] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Die casting parameter optimization and processing method based on orthogonal experiment, characterized in that: include: Based on the die-casting process knowledge base and die-casting production history data, the importance of die-casting parameters is calculated through clustering algorithm to obtain the die-casting parameter set. The nonlinear mapping model between die-casting parameters and corresponding value ranges is constructed through gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through die-casting parameter optimization algorithm to generate the die-casting parameter optimization level set. Based on the die-casting parameter optimization level set, an intelligent orthogonal test plan is generated through a preset orthogonal test design method, the die-casting equipment is controlled to conduct multiple batch tests, test data is obtained, and the die-casting parameter configuration strategy is optimized through a deep reinforcement learning model to obtain the optimal die-casting parameter settings, and the optimal die-casting workpiece is produced through the die-casting equipment; In the production process of die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through the incremental learning algorithm to generate updated experience knowledge, which is accumulated into the die-casting process knowledge graph to form a die-casting knowledge update mechanism; Based on the die casting parameter optimization level set, the intelligent orthogonal test scheme is generated through the preset orthogonal test design method, including: Based on the die-casting parameter optimization level set, the factors and levels of the orthogonal test are determined, where the factors are various die-casting parameters and the levels are the value ranges of the die-casting parameters. Combined with the preset die-casting process limitation knowledge, a matching orthogonal table is selected; Based on the association rule mining algorithm, the association strength between the die-casting parameters is determined by calculating the support, confidence and lift of the die-casting parameter combinations, and the optimization difficulty corresponding to each die-casting parameter is determined in combination with the preset process constraints. According to the association strength and the optimization difficulty, in combination with the preset association strength threshold and the preset optimization difficulty threshold, the column distribution of each factor in the orthogonal table is adjusted to obtain an adjusted orthogonal table, and the parameter value combination in the die-casting parameter optimization level concentration is filled into the adjusted orthogonal table to generate an initial orthogonal test plan; The initial orthogonal test scheme is encoded as the initial population of the neighborhood search algorithm. According to the objective function value of the individuals in the initial population, the corresponding quality evaluation index is calculated, and the quality evaluation index optimization weight vector is adaptively generated. For the individuals in the initial population, hyperspherical neighborhoods of different radii are generated according to the values ​​of the die-casting parameters, and new solutions are generated by sampling. The weight vector is optimized in combination with the quality evaluation index, and the weighted quality evaluation index is calculated. The new solution and the original solution are compared by non-dominated sorting, and the maximum value between the two is retained. The initial population is updated repeatedly until the preset number of iterations is reached, and an intelligent orthogonal test plan is output.

2. The method according to claim 1, characterized in that Based on the die-casting process knowledge base and die-casting production history data, the core die-casting parameters are extracted through clustering algorithm, and the die-casting parameter set is obtained, including: Obtain die-casting production history data, clean and standardize it, and obtain die-casting parameter numerical data; Based on the die-casting parameter numerical data, by introducing a multi-perspective similarity measurement mechanism, the similarity between samples in the die-casting parameter numerical data is calculated from three perspectives: numerical similarity, trend similarity and semantic similarity, and a multi-perspective die-casting parameter similarity matrix is ​​constructed; Based on the quality indicators in the die-casting production history data, an adaptive fusion strategy is determined, and the weights of the multi-view die-casting parameter similarity matrix are dynamically adjusted to obtain a dynamic die-casting parameter similarity matrix through fusion; Performing spectral decomposition on the dynamic die-casting parameter similarity matrix to obtain a low-dimensional embedding representation of the die-casting parameters, and mapping the numerical data of the die-casting parameters to a corresponding low-dimensional embedding space of the die-casting parameters; In the die-casting parameter low-dimensional embedding space, based on a preset initial cluster number range, a silhouette coefficient corresponding to the data sample is calculated, and according to a change in the silhouette coefficient, a cluster number range is determined; Traversing each potential cluster number within the range of the cluster number; Based on the local edge-preserving sampling strategy, data samples in the local neighborhood are selected in the low-dimensional embedding space of the die-casting parameters to construct a local similarity matrix, and the local similarity matrix is ​​reconstructed in combination with the random walk mechanism to construct a compressed Laplace matrix. Based on the compressed Laplace matrix, the spectral embedding representation of the die-casting parameters is obtained through spectral decomposition, and K-means clustering is performed. By maximizing the silhouette coefficient under the corresponding number of clusters, the local optimal clustering result under the current number of clusters is obtained; Compare the local optimal clustering results under different cluster numbers, select the local optimal clustering result with the largest silhouette coefficient, and obtain the global optimal clustering result; Based on the global optimal clustering results, the preset Shapley value method is used to calculate the contribution of the die-casting parameters in each cluster to the clustering results, and the importance ranking results of the die-casting parameters in the cluster are generated; Based on the ranking results of the importance of die-casting parameters within a cluster, an inter-cluster similarity matrix is ​​constructed. Based on the inter-cluster similarity matrix, a cluster importance propagation network is constructed. By adjusting the importance propagation intensity, the propagation fusion of the importance of die-casting parameters between clusters is achieved, and the global importance of die-casting parameters is obtained. According to the global importance of die-casting parameters, the die-casting parameters in the global optimal clustering results are sorted in descending order, duplicate parameters are removed, and the die-casting parameter set is output.

3. The method according to claim 2, characterized in that Through the gradient boosting decision tree algorithm, a nonlinear mapping model between die casting parameters and corresponding value ranges is constructed, including: Based on the die-casting parameter set and in combination with the die-casting production history data, a die-casting parameter sample set is determined, wherein the die-casting parameter sample set includes a die-casting parameter value vector of a plurality of die-casting parameter corresponding value combinations and a corresponding die-casting workpiece quality; Initializing a strong learner, and determining a predicted value of the strong learner based on a mass mean of the die-cast workpiece; For each sample in the die-casting parameter sample set, the negative gradient under the current strong learner is calculated to determine the residual estimate, and based on the die-casting parameter value vector and the corresponding residual estimate, a CART regression tree is generated by least squares fitting to obtain the leaf node area of ​​the CART regression tree, and for the leaf node area, based on a preset loss function, the minimum value of the loss function is determined, and the best fitting value is calculated, and based on the best fitting value, the strong learner is updated to re-determine the new prediction value; Repeat the strong learner update until the preset number of iterations is reached, obtain the die-casting parameter value range, and build a nonlinear mapping model of die-casting parameters and values.

4. The method according to claim 3, characterized in that The die casting parameter optimization algorithm includes: The die-casting parameter set is divided into a plurality of sub-populations, each sub-population corresponds to a type of die-casting parameter, each sub-population is initialized, the die-casting parameter value is randomly determined in combination with the die-casting parameter value range, the initial sub-population is determined, and the representative individual is randomly determined; Process all initial subpopulations to determine the current subpopulation, combine the individuals in the current subpopulation with the representative individuals of the remaining subpopulations, substitute them into the nonlinear mapping model of die-casting parameters and values, calculate the fitness values ​​corresponding to the individuals in the current subpopulation, calculate the novelty values ​​corresponding to each individual in the current subpopulation based on neighborhood comparison, determine the individual evaluations in the current subpopulation based on the fitness values ​​and novelty values, and determine the new representative individuals after weighted calculation according to the preset weights; Iterate the individuals of each initial subpopulation according to the preset crossover probability and mutation probability; when the individuals in an initial subpopulation do not meet the preset diversity index, reinitialize the corresponding number of individuals according to the preset number of new individuals and add them to the corresponding initial subpopulation; During the subpopulation iteration process, two initial subpopulations are randomly selected, and the representative individual of one initial subpopulation is migrated to the other initial subpopulation to replace the individual with the lowest individual evaluation in the other initial subpopulation; Until the preset number of iterations is reached, all final sub-populations are determined, representative individuals in each final sub-population are selected, and the optimal level combination of each die-casting parameter is determined.

5. The method according to claim 1, characterized in that Control the die-casting equipment to conduct multiple batch tests, obtain test data, and optimize the die-casting parameter configuration strategy through the deep reinforcement learning model to obtain the optimal die-casting parameter settings. The production of the optimal die-casting workpiece through the die-casting equipment includes: Based on the intelligent orthogonal test scheme, the initial parameter configuration of the die-casting equipment is set, the die-casting equipment is controlled to conduct multiple batches of test production, and the quality evaluation data of the die-casting workpieces under different die-casting parameter configurations is obtained; The die-casting parameter configuration is mapped to the state space of the deep reinforcement learning model, the adjustment actions corresponding to the die-casting parameters are mapped to the action space, and the die-casting workpiece quality evaluation data is mapped to the reward function to construct a Markov decision process for die-casting parameter optimization. Fit the state-action value function through a deep neural network, use the Q-learning algorithm to train the deep reinforcement learning model, and learn the die-casting parameter configuration strategy; The ε-greedy strategy is used to balance exploration and exploitation, and the experience replay mechanism is combined for model training to obtain the optimal deep reinforcement learning model; Based on the optimal deep reinforcement learning model, the current die-casting process status is input to obtain the optimal die-casting parameter configuration strategy; The optimal die-casting parameter configuration strategy is decoded into parameter setting instructions executable by the die-casting equipment, and the die-casting equipment is controlled to produce the optimal die-casting workpiece.

6. The method according to claim 1, characterized in that In the production process of die-casting equipment, data is collected in real time and the deep reinforcement learning model is dynamically updated through the incremental learning algorithm to generate updated experience knowledge, which is accumulated into the die-casting process knowledge graph to form a die-casting knowledge update mechanism including: In the production process of die-casting equipment, die-casting parameter data and die-casting quality data are collected in real time and organized into sample data of state-action-reward triples in time series; The newly collected sample data is added to the original training data set to form an incremental data set. The deep reinforcement learning model is adjusted based on the incremental data set using an incremental learning algorithm to obtain an updated deep reinforcement learning model. The incremental learning algorithm determines the experience replay priority according to the time axis position and error size of the sample data, and sets a forgetting factor to eliminate expired samples; Extract relational implicit knowledge from the updated deep reinforcement learning model to generate die-casting updated experience knowledge, formalize the updated experience knowledge into ontology, rules and cases, integrate it into the original die-casting process knowledge graph, and determine the dynamic evolution of the die-casting process knowledge graph, including ontology fusion, rule reasoning and case update; Based on natural language processing, the die-casting process knowledge graph is structured.

7. A die casting parameter optimization and processing system based on orthogonal experiment, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to calculate the importance of die-casting parameters based on the die-casting process knowledge base and die-casting production history data through a clustering algorithm to obtain a die-casting parameter set, and to construct a nonlinear mapping model between die-casting parameters and corresponding value ranges through a gradient boosting decision tree algorithm. Combined with the importance of die-casting parameters, the optimal level combination of each die-casting parameter is determined through a die-casting parameter optimization algorithm to generate an optimized level set of die-casting parameters; The second unit is used to generate intelligent orthogonal test schemes based on the die-casting parameter optimization level set through a preset orthogonal test design method, control the die-casting equipment to conduct multiple batch tests, obtain test data, and optimize the die-casting parameter configuration strategy through a deep reinforcement learning model to obtain the optimal die-casting parameter settings, and produce the optimal die-casting workpiece through the die-casting equipment; The third unit is used to collect data in real time during the production process of the die-casting equipment and dynamically update the deep reinforcement learning model through an incremental learning algorithm to generate updated experience knowledge, accumulate the updated experience knowledge into the die-casting process knowledge graph, and form a die-casting knowledge update mechanism.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images