Quick mold changing method and system

Abstract

This invention provides a rapid mold change method and system, relating to the field of data processing technology. The method includes acquiring historical mold change data, preprocessing the historical data, establishing a nonlinear correlation between production parameters, mold lifespan, product quality, and mold change time, determining a mold change time prediction model, optimizing hyperparameters through search and verification, and predicting the optimal mold change time. When the difference between the optimal mold change time and the current time is less than a preset trigger threshold, a dynamic adjustment mechanism for production parameters is triggered. A production parameter optimization algorithm is constructed to maximize production efficiency and mold utilization, generating an optimal control strategy and adjusting production parameters. When the optimal mold change time is reached, an intelligent mold switching algorithm identifies the mold to be replaced, obtains the model and status information of the mold, generates an optimal mold selection and replacement scheme based on the model and status information, and performs the mold change operation according to the optimal mold selection and replacement scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a rapid mold changing method and system. Background Art

[0002] In modern manufacturing, the demand for product diversification and personalization is growing. In order to adapt to market changes and meet customer needs, manufacturers need to frequently switch products and production. In the fields of injection molding, die casting, stamping, etc., mold change is a key link in production change, which directly affects production efficiency and product quality. The traditional mold change method mainly relies on manual experience and manual operation, and there are problems such as long mold change time, unstable mold change quality, uncertain production plan, and high labor cost. The rapid mold change method proposed by the industry has shortened the mold change time and improved the mold change efficiency to a certain extent, but it still has shortcomings such as low intelligence level, weak versatility and adaptability, and low system integration. Summary of the invention

[0003] The embodiments of the present invention provide a rapid mold changing method and system, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] A quick mold change method is provided, comprising:

[0006] Acquire historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, optimize the hyperparameters of the mold change time prediction model through search and verification, and predict the optimal mold change time;

[0007] When the difference between the optimal mold change time and the current time is less than a preset trigger threshold, the dynamic adjustment mechanism of production parameters is triggered. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization;

[0008] When the optimal mold change time is reached, the mold to be replaced is identified through the intelligent mold switching algorithm, the model and status information of the mold to be replaced are obtained, and based on the model and the status information, the optimal mold selection and replacement plan is generated, and the mold change operation is performed according to the optimal mold selection and replacement plan.

[0009] In an optional embodiment,

[0010] Obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, optimize the hyperparameters of the mold change time prediction model through search verification, and predict the optimal mold change time including:

[0011] Performing feature engineering on the historical mold change data to obtain a feature list;

[0012] Analyze each feature in the feature list to determine mutually exclusive features, combine the mutually exclusive features into a feature bundle, and combine the independent features corresponding to the non-mutually exclusive features into a feature point set;

[0013] Extracting continuous feature points with continuous values ​​from the feature point set, calculating the value range of each of the continuous feature points, dividing the value range into a plurality of segmented intervals, and combining the discrete intervals corresponding to the discontinuous feature points with discontinuous values ​​in the feature point set into a feature interval;

[0014] Based on the historical mold change data, determine the training samples of the mold change time prediction model, initialize the parameters of the mold change time prediction model, and determine the initial prediction result;

[0015] Based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached to determine the final mold change time prediction model;

[0016] Through grid search and cross validation, the hyperparameters of the final mold change time prediction model are tuned to obtain the optimal mold change time prediction model, and the optimal mold change time is predicted.

[0017] In an optional embodiment,

[0018] Based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached. Determining the final mold change time prediction model includes:

[0019] Based on the initial prediction result, the gradient and the second-order partial derivative matrix of each training sample are calculated, and for each feature interval of each feature point, the total number of training samples in each feature interval, the sum of the gradients corresponding to the total number of training samples, and the sum of the second-order partial derivative matrices corresponding to the total number of training samples are counted to construct a feature point histogram;

[0020] Initialize the root node of the decision tree and put all the training samples into the root node;

[0021] For each non-leaf node, traverse all splitting points of all feature points, calculate the splitting gain, select the feature point and splitting point corresponding to the maximum value of the splitting gain as the optimal splitting, divide the samples into left and right child nodes according to the optimal splitting, update the feature point histogram of the child node, if the depth corresponding to the child node reaches the preset maximum depth, or the number of training samples in the child node is less than the preset sample threshold, mark the child node as a leaf node, stop node splitting, and calculate the predicted value of the leaf node; until all nodes are marked as leaf nodes, determine the decision tree;

[0022] The decision tree is added to the mold change time prediction model to obtain a new prediction result, and a new decision tree is generated based on the new prediction result until a preset number of iterations is reached.

[0023] In an optional embodiment,

[0024] The dynamic adjustment mechanism of production parameters builds a production parameter optimization algorithm to generate an optimal control strategy based on the current production status with the goal of maximizing production efficiency and mold utilization. The production parameters are adjusted as follows:

[0025] Based on the working process of the production system, the state variables are determined, and the state variables are combined into high-order features through feature combinations. The state variables and the high-order features constitute a state space; based on the control actions of the production system, the adjustment range and adjustment granularity of the production parameters are determined, the adjustment range and the adjustment granularity are discretized, and discrete adjustment values ​​are determined. The operation actions corresponding to the discrete adjustment values ​​constitute an action space;

[0026] The production parameter optimization algorithm is based on a deep neural network, constructing an input layer for receiving the current production state, multiple hidden layers for extracting state features, and an output layer for outputting the quality value corresponding to each manipulation action, generating a decision quality network, and creating a target quality network with the same structure as the decision quality network;

[0027] By minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network, the parameters of the decision quality network are updated; according to a preset replication interval step, the parameters of the decision quality network are copied to the target quality network, and finally an optimal target quality network is generated;

[0028] The optimal target quality network determines the optimal production parameters according to the input status.

[0029] In an optional embodiment,

[0030] Updating the parameters of the decision quality network by minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network includes:

[0031] Randomly initialize the weights of the decision quality network and the target quality network, and set the parameters of the target quality network to be copies of the corresponding parameters of the decision quality network;

[0032] According to the current production state received by the input layer, the decision quality value of each operation action is calculated through forward propagation of the decision quality network, and the maximum decision quality value is extracted as the current decision quality value. The operation action corresponding to the current decision quality value is used as the current decision action corresponding to the current production state, forming a current state action pair;

[0033] According to the current state-action pair, a next state and an immediate reward are obtained, the next state is input into a target quality network, a target quality value of each operation action corresponding to the next state is calculated, a maximum target quality value is extracted, and a current target quality value of the current state-action pair is determined based on the immediate reward and the maximum target quality value;

[0034] Based on the current target quality value and the current decision quality value, the TD error is calculated, and the parameters of the decision quality network are updated through back propagation with the goal of minimizing the mean square error between the decision quality value and the target quality value.

[0035] In an optional embodiment,

[0036] Also includes:

[0037] Selecting an operation action according to the current state and a greedy strategy, wherein the greedy strategy includes: generating a random number between 0 and 1, and if the random number is less than a preset exploration rate, randomly selecting an operation action as a decision action; if the random number is greater than or equal to the exploration rate, selecting an operation action corresponding to the maximum decision quality value of the decision quality network as the decision action; wherein the exploration rate decreases as the number of iterations increases;

[0038] Execute the decision action, transfer the current state to the next state, generate state transfer data, obtain an immediate reward, and store the state transfer data in an experience playback buffer, wherein the state transfer data includes the current state, the decision action, the immediate reward, and the next state;

[0039] A set of state transition data is randomly sampled from the experience replay buffer, and for each of the state transition data, a sampled target quality value is calculated; if the next state is a terminal state, the sampled target quality value is equal to the immediate reward; otherwise, the sampled target quality value is equal to the weighted sum of the immediate reward and the maximum target quality value output by the target quality network of the next state.

[0040] In an optional embodiment,

[0041] The mold to be replaced is identified by an intelligent mold switching algorithm, the model and status information of the mold to be replaced are obtained, and based on the model and status information, an optimal mold selection and replacement plan is generated, including:

[0042] According to the model of the mold to be replaced, matching replacement molds are searched in the mold database, the replacement molds are filtered according to the mold status score, the remaining service life and the compatibility index, and a candidate mold set is determined according to a preset filtering quantity;

[0043] According to the preset mold selection benchmark factors, multiple decision attributes are determined, and the decision attributes are compared in pairs by constructing a hierarchical model to determine a judgment matrix, calculate the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, and determine the attribute weights of the decision attributes;

[0044] Normalizing the decision attributes, constructing a decision matrix, adding weights to the decision matrix based on the attribute weights to form a weighted normalized decision matrix, determining a positive ideal solution and a negative ideal solution, respectively calculating the Euclidean distances between the candidate mold and the positive ideal solution and the negative ideal solution, determining the relative closeness of the candidate mold based on the Euclidean distances, and generating a ranking of the candidate molds;

[0045] With maximizing production efficiency as the first goal and maximizing mold utilization as the second goal, a multi-objective optimization algorithm is constructed to generate the optimal mold replacement solution set;

[0046] Based on the ranking of the candidate molds, the candidate mold with the highest ranking is selected from the set of optimal mold replacement solutions to obtain the optimal mold. The optimal mold selection and replacement plan is generated in combination with the mold information in the mold database.

[0047] According to a second aspect of the embodiments of the present invention,

[0048] A quick mold change system is provided, comprising:

[0049] The first unit is used to obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, and optimize the hyperparameters of the mold change time prediction model through search and verification to predict the optimal mold change time;

[0050] The second unit is used to trigger a dynamic adjustment mechanism of production parameters when the difference between the optimal mold change time and the current time is less than a preset trigger threshold. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization;

[0051] The third unit is used to identify the mold to be replaced by an intelligent mold switching algorithm when the optimal mold change time is reached, obtain the model and status information of the mold to be replaced, and generate an optimal mold selection and replacement plan based on the model and the status information, and perform mold change operations according to the optimal mold selection and replacement plan.

[0052] According to a third aspect of the embodiments of the present invention,

[0053] An electronic device is provided, comprising:

[0054] processor;

[0055] a memory for storing processor-executable instructions;

[0056] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the embodiments of the present invention is:

[0058] 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.

[0059] In an embodiment of the present invention, each feature in the feature list is parsed, mutually exclusive features are identified, and they are combined into feature bundles. At the same time, features that are not mutually exclusive are defined as independent features, which can better explore the correlation and influencing factors between features and provide richer feature information for the model; continuous feature points with continuous values ​​are extracted from the feature point set, and the value range of each continuous feature point is calculated, which is divided into multiple segmented intervals. At the same time, for non-continuous feature points with discontinuous values, their possible values ​​are directly used as discrete intervals, which can better handle different types of features and provide more feature combinations for the model; the decision quality network can calculate the decision quality value corresponding to each operation action through the forward propagation process, reflecting the pros and cons of executing each operation action under the current production state; by extracting the maximum decision quality value in the output layer, Determine the optimal operation action in the current state, thereby realizing an intelligent decision-making process; the target quality network can calculate the target quality value corresponding to each operation action through the forward propagation process, and these values ​​reflect the expected quality performance of each operation action in the next state; through the greedy strategy, according to the comparison between the random number and the preset exploration rate, flexibly choose between exploration and utilization, so as to balance the proportion of exploration and utilization; the establishment and management of the experience replay buffer enables the algorithm to effectively utilize previous experience, improving the efficiency and stability of training; by constructing a hierarchical model and pairwise comparison, the weights of multiple decision attributes can be objectively determined, and the importance of each attribute can be fully considered; by calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix, the relative proportion of the attribute weights is determined, and the accuracy and credibility of the decision attribute weights are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the process of the rapid mold changing method according to an embodiment of the present invention;

[0061] Figure 2 It is a schematic structural diagram of a rapid mold change system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] 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.

[0063] 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.

[0064] Figure 1 FIG. 1 is a flow chart of a rapid mold changing method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] S101. Obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, optimize the hyperparameters of the mold change time prediction model through search and verification, and predict the optimal mold change time;

[0066] The historical mold change data includes production parameters, mold service life, product quality, mold change time and other mold change related data;

[0067] Collect historical mold change data, and perform data cleaning and preprocessing, including removing outliers and processing missing values; analyze historical mold change data, build a mold change time prediction model based on machine learning models or time series analysis, and determine the relationship between production parameters, mold service life, product quality and other characteristics and mold change time; optimize the hyperparameters of the established mold change time prediction model through search verification, and use cross-validation technology to determine the best model parameters; evaluate the optimized model to ensure accuracy and generalization ability, and use the trained mold change time prediction model to predict future mold change time and obtain the best mold change time;

[0068] In this embodiment, the quality and availability of the data are improved by preprocessing and feature engineering the historical mold change data, cleaning the data, processing missing values, performing feature selection and conversion, etc.; a modeling method suitable for the problem is adopted to capture the complex relationship between production parameters, mold service life, product quality and mold change time; the established mold change time prediction model is used to dynamically predict the optimal time point for the next mold change, and the production parameters are dynamically adjusted according to the prediction results to optimize production efficiency and product quality and improve model utilization.

[0069] In an optional embodiment, historical mold change data is obtained, the historical mold change data is preprocessed, a nonlinear correlation relationship between production parameters, mold service life, product quality and mold change time is established based on the historical mold change data, a mold change time prediction model is determined, and hyperparameters of the mold change time prediction model are optimized through search verification. Predicting the optimal mold change time includes:

[0070] Performing feature engineering on the historical mold change data to obtain a feature list;

[0071] Analyze each feature in the feature list to determine mutually exclusive features, combine the mutually exclusive features into a feature bundle, and combine the independent features corresponding to the non-mutually exclusive features into a feature point set;

[0072] Extracting continuous feature points with continuous values ​​from the feature point set, calculating the value range of each of the continuous feature points, dividing the value range into a plurality of segmented intervals, and combining the discrete intervals corresponding to the discontinuous feature points with discontinuous values ​​in the feature point set into a feature interval;

[0073] Based on the historical mold change data, determine the training samples of the mold change time prediction model, initialize the parameters of the mold change time prediction model, and determine the initial prediction result;

[0074] Based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached to determine the final mold change time prediction model;

[0075] Through grid search and cross validation, the hyperparameters of the final mold change time prediction model are tuned to obtain the optimal mold change time prediction model, and the optimal mold change time is predicted.

[0076] Clean the historical mold change data, remove missing values, outliers and duplicate data, ensure the integrity and consistency of the data, extract features from the cleaned data, and select features that have a greater impact on the mold change time based on production experience and mold field knowledge, such as mold structure complexity, mold size, processing technology, mold change operation difficulty, etc., to form a feature list; analyze the relationship between each feature in the feature list, and identify mutually exclusive features, such as mold type and product material, mold change operation difficulty and operator proficiency, etc.; combine the identified mutually exclusive features into feature bundles, each feature bundle represents a set of feature values ​​that cannot appear at the same time; define features that do not have a mutually exclusive relationship in the feature list as independent features; combine feature bundles and independent features into a feature point set, each feature point represents a possible feature combination;

[0077] Extract continuous feature points with continuous values ​​from the feature point set, such as the number of times the mold is used, the ambient temperature, etc.; for each continuous feature point, calculate the minimum and maximum values ​​in the historical mold change data to obtain the value range; divide the value range into multiple continuous segmented intervals based on the distribution of continuous feature points and production experience;

[0078] For example, the number of times a mold is used is divided into [0,1000), [1000,2000), [2000,+∞), etc.

[0079] For discontinuous feature points with discontinuous values ​​in the feature point set, such as mold type, material properties, etc., directly use their possible values ​​as discrete intervals; combine the segmented intervals of continuous feature points and the discrete intervals of discontinuous feature points into feature intervals, and each feature interval represents a possible feature value range;

[0080] Based on historical mold change data, the characteristic values ​​of each mold change operation and the corresponding actual mold change time are extracted to form a training sample set. The decision tree is selected as the basic algorithm of the mold change time prediction model, and the model parameters are initialized, including the maximum depth of the tree, the minimum number of samples of leaf nodes, and the node splitting standard. The training samples are fitted with the initialized model to obtain the initial prediction results. Based on the initial prediction results, the samples in each feature interval are counted to construct a feature point histogram. According to the feature point histogram, the optimal splitting feature and splitting threshold are selected by maximizing the node splitting gain, and a decision tree is generated. The generated decision tree is added to the mold change time prediction model to form a decision tree ensemble.

[0081] Use the updated model to re-predict the training samples and obtain new prediction results; repeat the decision tree construction process, iteratively generate multiple decision trees, and continuously improve the prediction accuracy of the model; when the number of iterations reaches the preset threshold or the model performance converges, stop the iteration and obtain the final mold change time prediction model;

[0082] Use the grid search method to optimize the model's hyperparameters, such as the maximum depth of the decision tree, the minimum number of samples of leaf nodes, and the learning rate. Use cross-validation to divide the training samples into multiple subsets, which are used as training sets and validation sets, to evaluate the performance of the model under different hyperparameter combinations. Select the hyperparameter combination with the best average performance in cross-validation as the optimal hyperparameter of the model. Use the optimal hyperparameters to retrain the model and obtain the optimal mold change time prediction model.

[0083] When the production line needs to change molds, the characteristic data of the current production environment is collected, including mold status, product requirements, machine parameters, etc. The collected characteristic data is input into the optimal mold change time prediction model to obtain the predicted mold change time; based on the predicted mold change time, combined with the production plan and actual situation, the optimal mold change time is determined and the mold change operation is arranged.

[0084] In this embodiment, each feature in the feature list is parsed, mutually exclusive features are identified, and they are combined into feature bundles. At the same time, features that are not mutually exclusive are defined as independent features, which can better explore the correlation and influencing factors between features and provide richer feature information for the model; continuous feature points with continuous values ​​are extracted from the feature point set, and the value range of each continuous feature point is calculated, which is divided into multiple segmented intervals. At the same time, for non-continuous feature points with discontinuous values, their possible values ​​are directly used as discrete intervals, which can better handle different types of features and provide more feature combinations for the model; based on historical model change The data is used to initialize the parameters of the mold change time prediction model, and the decision tree is used as the basic algorithm to build the model. By maximizing the node splitting gain principle, a decision tree is generated and added to the mold change time prediction model to form a decision tree integration model, which can integrate the prediction results of multiple decision trees and improve the prediction accuracy and stability of the model. The hyperparameters of the model are optimized through grid search and cross-validation methods, including the maximum depth of the decision tree, the minimum number of samples of leaf nodes, the learning rate, etc. The hyperparameter combination with the best average performance in cross-validation is selected as the optimal hyperparameter of the model, which can improve the generalization ability and prediction performance of the model.

[0085] In an optional embodiment, based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached. Determining the final mold change time prediction model includes:

[0086] Based on the initial prediction result, the gradient and the second-order partial derivative matrix of each training sample are calculated, and for each feature interval of each feature point, the total number of training samples in each feature interval, the sum of the gradients corresponding to the total number of training samples, and the sum of the second-order partial derivative matrices corresponding to the total number of training samples are counted to construct a feature point histogram;

[0087] Initialize the root node of the decision tree and put all the training samples into the root node;

[0088] For each non-leaf node, traverse all splitting points of all feature points, calculate the splitting gain, select the feature point and splitting point corresponding to the maximum value of the splitting gain as the optimal splitting, divide the samples into left and right child nodes according to the optimal splitting, update the feature point histogram of the child node, if the depth corresponding to the child node reaches the preset maximum depth, or the number of training samples in the child node is less than the preset sample threshold, mark the child node as a leaf node, stop node splitting, and calculate the predicted value of the leaf node; until all nodes are marked as leaf nodes, determine the decision tree;

[0089] The decision tree is added to the mold change time prediction model to obtain a new prediction result, and a new decision tree is generated based on the new prediction result until a preset number of iterations is reached.

[0090] Based on the initial prediction results, the gradient and second-order partial derivative matrix of the loss function are calculated for each training sample; the gradient represents the error between the predicted value and the true value, and the second-order partial derivative matrix represents the rate of change of the error, which reflects the fitting of the model at the sample;

[0091] For each feature interval of each feature point, count the total number of training samples falling into the interval; for each feature interval, accumulate the gradients of the training samples in it to obtain the sum of the gradients; for each feature interval, accumulate the second-order partial derivative matrices of the training samples in it to obtain the sum of the second-order partial derivative matrices; combine the above statistical information into a feature point histogram, each feature interval corresponds to a bucket of the histogram, and record the cumulative values ​​of the number of samples, gradients, and second-order partial derivative matrices;

[0092] Create the root node of the decision tree and put all training samples into the root node. The root node represents the starting point of the decision tree and contains the sample set to be split.

[0093] For each non-leaf node, perform the build operation:

[0094] Traverse all feature points of the node and all possible splitting points of each feature point; for each splitting point, calculate the gradient and second-order partial derivative matrix of the left and right child nodes after splitting according to the feature point histogram; use the gradient and second-order partial derivative matrix to calculate the splitting gain of the splitting point, which indicates the degree of error reduction after splitting; select the feature point and splitting point with the largest splitting gain as the optimal splitting, and divide the samples into left and right child nodes; update the feature point histogram of the left and right child nodes, and re-count the number of samples, the cumulative value of the gradient and the second-order partial derivative matrix; determine whether the child node meets the preset stop condition, such as reaching the maximum depth or the number of samples is less than the threshold; if the stop condition is met, mark the child node as a leaf node, calculate the predicted value of the node, and stop splitting; if the stop condition is not met, add the child node to the node queue to be split, and continue recursive splitting;

[0095] Repeat the construction until all nodes are marked as leaf nodes and the decision tree is completed;

[0096] The newly generated decision tree is added to the mold change time prediction model to form an integrated model with the previous tree. The updated model is used to predict the training samples to obtain new prediction results. Based on the new prediction results, the next decision tree is regenerated.

[0097] Each iteration will get an updated prediction result, and build the next tree based on the result until the preset number of iterations is reached or the model performance converges;

[0098] When the iterative optimization is completed, an integrated model consisting of multiple decision trees is obtained; the prediction results of different decision trees are combined to give the final prediction value through weighted average or voting;

[0099] In this embodiment, by calculating the gradient and second-order partial derivative matrix of the loss function, the fitting of the model at each sample can be evaluated more accurately, thereby improving the stability and generalization ability of the model; constructing a feature point histogram can effectively characterize the distribution between different feature points, help the model better understand the characteristics of the data, and make effective splitting decisions during the decision tree construction process; a decision tree is constructed in a recursive manner, the splitting gain is calculated based on the feature point histogram, and the optimal splitting point is selected, thereby effectively improving the construction efficiency and accuracy of the decision tree; through iterative optimization, the model is continuously updated, new decision trees are generated and added to the integrated model, so that the model can continuously learn and improve and adapt to changes in data and changes in the production environment; the final mold change time prediction model is an integrated model composed of multiple decision trees, which combines the prediction results of different decision trees, has stronger generalization ability and robustness, and can more accurately predict the mold change time of new samples.

[0100] S102. When the difference between the optimal mold change time and the current time is less than a preset trigger threshold, the dynamic adjustment mechanism of production parameters is triggered. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization;

[0101] When the difference between the optimal mold change time and the current time is less than a preset trigger threshold, triggering the dynamic adjustment mechanism of production parameters means that when the system detects that a mold change is about to occur, it will automatically start a mechanism to dynamically adjust the production parameters to maximize production efficiency and mold utilization; the trigger condition of the mechanism is that the difference between the optimal mold change time and the current time is less than a preset trigger threshold. The trigger threshold can be set according to production demand, equipment status, mold change preparation time and other factors. It is usually a pre-set time interval. When the system detects that the remaining time is insufficient to complete the scheduled production task and make sufficient preparations, it will start the dynamic adjustment mechanism of production parameters.

[0102] In an optional embodiment, the dynamic adjustment mechanism of production parameters generates an optimal control strategy by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization, and adjusting production parameters includes:

[0103] Based on the working process of the production system, the state variables are determined, and the state variables are combined into high-order features through feature combinations. The state variables and the high-order features constitute a state space; based on the control actions of the production system, the adjustment range and adjustment granularity of the production parameters are determined, the adjustment range and the adjustment granularity are discretized, and discrete adjustment values ​​are determined. The operation actions corresponding to the discrete adjustment values ​​constitute an action space;

[0104] The production parameter optimization algorithm is based on a deep neural network, constructing an input layer for receiving the current production state, multiple hidden layers for extracting state features, and an output layer for outputting the quality value corresponding to each manipulation action, generating a decision quality network, and creating a target quality network with the same structure as the decision quality network;

[0105] By minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network, the parameters of the decision quality network are updated; according to a preset replication interval step, the parameters of the decision quality network are copied to the target quality network, and finally an optimal target quality network is generated;

[0106] The optimal target quality network determines the optimal production parameters according to the input status.

[0107] According to the working process of the production system, variables that can reflect the state of the system, such as temperature, pressure, speed, etc., are determined as state variables; the state variables are combined to form high-order features, such as the mean value, variance, and rate of change of the variables; the state variables and high-order features are combined to form a state space to describe the state of the production system;

[0108] According to the control actions of the production system, determine the adjustable production parameters, such as temperature setting value, pressure threshold, etc.; for each production parameter, determine its adjustable range and adjustment granularity, for example, the temperature can be adjusted between 200-300℃, and the adjustment granularity is 1℃; discretize the adjustment range and adjustment granularity to obtain a set of discrete adjustment values, for example, the temperature can be adjusted to 200℃, 201℃, 202℃, etc.; the operation actions corresponding to the discrete adjustment values ​​constitute an action space, which is used to describe the adjustment method of the production parameters;

[0109] A decision quality network based on a deep neural network is constructed to evaluate the quality value of each operation action. The input layer of the decision quality network receives the current production state, that is, the state variables and high-order features in the state space. The decision quality network also includes multiple hidden layers for extracting state features and learning the relationship between state and action. The output layer of the decision quality network outputs the quality value corresponding to each manipulation action, indicating the degree of performance of the action under the current state.

[0110] Create a target quality network with the same structure as the decision quality network to generate the target quality value, and initialize the parameters of the target quality network to the parameter copy of the decision quality network;

[0111] By minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network, the parameters of the decision quality network are updated; using optimization algorithms such as gradient descent, the weights and biases of the decision quality network are adjusted to make the quality value output by the decision quality network as close to the target quality value as possible;

[0112] According to the preset replication interval step size, such as every 10 steps, the parameters of the decision quality network are copied to the target quality network to keep it relatively stable; until the preset number of iterations is reached, or the decision quality network converges ahead of time, the best target quality network is obtained.

[0113] Use the trained optimal target quality network to predict the quality value of each manipulation action based on the input current production status; select the manipulation action with the highest quality value and use its corresponding discrete adjustment value as the optimal production parameter; apply the optimal production parameter to the production system to guide the subsequent mold change process;

[0114] In this embodiment, by constructing the state space and action space, the state and adjustable parameters of the production system can be fully described, which is helpful to understand the working mechanism and control method of the production system; the decision quality network and target quality network constructed using the deep neural network can effectively learn the complex relationship between the state and the action, and provide a highly flexible and adaptive decision-making mechanism; the existence of the target quality network makes the training process more stable, and by taking the target quality value as the training target, the learning of the decision quality network can be effectively guided, and the convergence speed and performance of the model are improved; through network training, the parameters of the decision quality network can be continuously optimized, so that it can accurately evaluate the quality value of each operation action, thereby providing a reliable basis for the selection of production parameters; the application of optimization algorithms such as gradient descent can effectively reduce the error of the network and improve the prediction accuracy and stability of the model; based on the best target quality network obtained by training, the quality value of each control action can be quickly and accurately predicted, so as to select the optimal production parameters; through the optimization of production parameters, the production efficiency and mold utilization rate can be maximized, resource waste and production costs can be reduced, and the intelligent and automated production process can be realized.

[0115] In an optional embodiment, updating the parameters of the decision quality network by minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network includes:

[0116] Randomly initialize the weights of the decision quality network and the target quality network, and set the parameters of the target quality network to be copies of the corresponding parameters of the decision quality network;

[0117] According to the current production state received by the input layer, the decision quality value of each operation action is calculated through forward propagation of the decision quality network, and the maximum decision quality value is extracted as the current decision quality value. The operation action corresponding to the current decision quality value is used as the current decision action corresponding to the current production state, forming a current state action pair;

[0118] According to the current state-action pair, a next state and an immediate reward are obtained, the next state is input into a target quality network, a target quality value of each operation action corresponding to the next state is calculated, a maximum target quality value is extracted, and a current target quality value of the current state-action pair is determined based on the immediate reward and the maximum target quality value;

[0119] Based on the current target quality value and the current decision quality value, the TD error is calculated, and the parameters of the decision quality network are updated through back propagation with the goal of minimizing the mean square error between the decision quality value and the target quality value.

[0120] The instant reward specifically refers to the reward or feedback obtained immediately after performing an action. In the production system, the instant reward can be calculated based on the production index and used to evaluate the quality of the current action. The instant reward can be a positive value, indicating that a positive effect has been produced after the action has been performed; it can also be a negative value, indicating that a negative effect has been produced after the action has been performed;

[0121] The target quality value specifically refers to the quality level expected to be achieved after performing a certain action in the current state. In the production system, the target quality value is usually estimated based on the expected production indicators, such as maximizing production efficiency and maximizing mold utilization. The target quality value can be used as a decision-making goal to help the system judge the pros and cons of the current decision.

[0122] The decision quality value specifically refers to the quality value of each operation action calculated by the decision quality network according to the current state, which is used to guide the selection of the current decision; the decision quality value can be understood as the potential value of each action in the current state, and the higher the value, the more conducive it is to achieving the goal of the system; the action with the highest decision quality value is selected as the optimal decision action in the current state, so that the system develops in a better direction;

[0123] The current production state is taken as input and received through the input layer of the decision quality network; the decision quality value corresponding to each operation action is calculated through the forward propagation process of the decision quality network; at the output layer, the maximum decision quality value is extracted as the optimal decision quality value under the current state; the current state and the operation action corresponding to the optimal decision quality value constitute a current state action pair;

[0124] According to the current state-action pair, the corresponding operation action is executed to obtain the next production state and immediate reward; the next production state is used as input, and the target quality value corresponding to each operation action is calculated through the forward propagation process of the target quality network; in the output layer of the target quality network, the maximum target quality value is extracted as the optimal target quality value of the next state; based on the immediate reward and the optimal target quality value of the next state, the current target quality value of the current state-action pair is calculated;

[0125] Calculate the TD error, that is, the difference between the current target quality value and the current decision quality value; update the parameters of the decision quality network through the back propagation algorithm with the optimization goal of minimizing the mean square error of the TD error; use optimization algorithms such as gradient descent to adjust the weights and bias of the decision quality network so that the decision quality value output by the decision quality network is as close as possible to the current target quality value.

[0126] In this embodiment, the decision quality network can calculate the decision quality value corresponding to each operation action through the forward propagation process, reflecting the quality of each operation action under the current production state; by extracting the maximum decision quality value in the output layer, the optimal operation action under the current state is determined, thereby realizing an intelligent decision-making process; the target quality network can calculate the target quality value corresponding to each operation action through the forward propagation process, and these values ​​reflect the expected quality performance of each operation action under the next state; the maximum target quality value in the output layer is extracted to determine the optimal target quality value for the next state, providing effective guidance and reference for decision-making; the difference between the current target quality value and the current decision quality value, i.e., the TD error, is calculated to measure the deviation between the current decision quality and the target quality. The calculation of the TD error provides a key guiding direction for subsequent parameter updates, helping the decision quality network to better approach the target quality value.

[0127] In an optional embodiment, it also includes:

[0128] Selecting an operation action according to the current state and a greedy strategy, wherein the greedy strategy includes: generating a random number between 0 and 1, and if the random number is less than a preset exploration rate, randomly selecting an operation action as a decision action; if the random number is greater than or equal to the exploration rate, selecting an operation action corresponding to the maximum decision quality value of the decision quality network as the decision action; wherein the exploration rate decreases as the number of iterations increases;

[0129] Execute the decision action, transfer the current state to the next state, generate state transfer data, obtain an immediate reward, and store the state transfer data in an experience playback buffer, wherein the state transfer data includes the current state, the decision action, the immediate reward, and the next state;

[0130] A set of state transition data is randomly sampled from the experience replay buffer, and for each of the state transition data, a sampled target quality value is calculated; if the next state is a terminal state, the sampled target quality value is equal to the immediate reward; otherwise, the sampled target quality value is equal to the weighted sum of the immediate reward and the maximum target quality value output by the target quality network of the next state.

[0131] According to the current production status, the operation action is selected through the greedy strategy; a random number between 0 and 1 is generated to decide whether to explore or exploit;

[0132] If the random number is less than the preset exploration rate (such as 0.1), exploration is performed and an operation action is randomly selected from the action space as the decision action. Exploration can discover new strategies and avoid premature convergence to a suboptimal solution.

[0133] If the random number is greater than or equal to the exploration rate, then use it, input the current state into the decision quality network, calculate the decision quality value of each operation action, and select the operation action with the largest decision quality value as the decision action; use can make the best decision based on current knowledge and improve benefits;

[0134] As the number of iterations increases, the exploration rate gradually decreases. In the early stage, the exploration rate is high, which encourages the exploration of new strategies. In the later stage, the exploration rate decreases and more existing knowledge is used.

[0135] Execute the selected decision action to transfer the production system from the current state to the next state; record the state transition data, including the current state, decision action, immediate reward and next state;

[0136] The current state refers to the production state before the decision action is executed; the decision action refers to the selected operation action; the immediate reward refers to the immediate feedback given based on production indicators such as quality and efficiency after the decision action is executed; the next state refers to the new state of the production system after the decision action is executed.

[0137] Store state transition data in the experience replay buffer for subsequent training and updating;

[0138] A set of state transition data is randomly sampled from the experience replay buffer; for each sampled state transition data, the corresponding sampling target quality value is calculated.

[0139] If the next state is a terminal state, the sampled target quality value is equal to the immediate reward; if the next state is not a terminal state, the sampled target quality value is the weighted sum of the immediate reward and the maximum target quality value of the next state.

[0140] The target quality network is used to calculate the maximum target quality value of the next state to improve the stability of the target quality value estimation.

[0141] In this embodiment, a greedy strategy is used to flexibly choose between exploration and utilization based on the comparison between random numbers and a preset exploration rate, thereby balancing the ratio of exploration and utilization; the exploration phase can help the system discover new strategies, avoid falling into local optimal solutions, and improve the exploration ability of the algorithm; the utilization phase makes optimal decisions based on current knowledge, improves benefits, and accelerates the convergence speed of the algorithm; the state transition data including the current state, decision action, immediate reward, and next state are recorded, providing a data basis for subsequent training and updating; the establishment and management of the experience replay buffer enables the algorithm to effectively utilize previous experience, thereby improving the efficiency and stability of training; the sampling target quality value is flexibly calculated based on whether the next state is a terminal state, taking into account the weighted sum of the immediate reward and the maximum target quality value of the next state, thereby improving the estimation accuracy and stability of the target quality value; the target quality network is used to calculate the maximum target quality value of the next state, thereby further improving the stability and accuracy of the target quality value.

[0142] S103. When the optimal mold change time is reached, the mold to be replaced is identified through the intelligent mold switching algorithm, the model and status information of the mold to be replaced are obtained, and based on the model and the status information, an optimal mold selection and replacement plan is generated, and the mold change operation is performed according to the optimal mold selection and replacement plan.

[0143] In an optional embodiment, the mold to be replaced is identified by an intelligent mold switching algorithm, the model and status information of the mold to be replaced are obtained, and based on the model and the status information, an optimal mold selection and replacement plan is generated, including:

[0144] According to the model of the mold to be replaced, matching replacement molds are searched in the mold database, the replacement molds are filtered according to the mold status score, the remaining service life and the compatibility index, and a candidate mold set is determined according to a preset filtering quantity;

[0145] According to the preset mold selection benchmark factors, multiple decision attributes are determined, and the decision attributes are compared in pairs by constructing a hierarchical model to determine a judgment matrix, calculate the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, and determine the attribute weights of the decision attributes;

[0146] Normalizing the decision attributes, constructing a decision matrix, adding weights to the decision matrix based on the attribute weights to form a weighted normalized decision matrix, determining a positive ideal solution and a negative ideal solution, respectively calculating the Euclidean distances between the candidate mold and the positive ideal solution and the negative ideal solution, determining the relative closeness of the candidate mold based on the Euclidean distances, and generating a ranking of the candidate molds;

[0147] With maximizing production efficiency as the first goal and maximizing mold utilization as the second goal, a multi-objective optimization algorithm is constructed to generate the optimal mold replacement solution set;

[0148] Based on the ranking of the candidate molds, the candidate mold with the highest ranking is selected from the set of optimal mold replacement solutions to obtain the optimal mold. The optimal mold selection and replacement plan is generated in combination with the mold information in the mold database.

[0149] The positive ideal solution specifically refers to a hypothetical optimal solution, which achieves the optimal value in all evaluation attributes. For benefit-type attributes, the larger the value, the better, the positive ideal solution takes the maximum value of the attribute; for cost-type attributes, the smaller the value, the better, the positive ideal solution takes the minimum value of the attribute; the positive ideal solution represents the best performance of the candidate solution in all attributes and is the best choice in the ideal. The positive ideal solution represents the best performance under ideal conditions and is the ideal state of the optimization target.

[0150] The negative ideal solution specifically refers to a hypothetical worst solution, which reaches the worst value in all evaluation attributes. For benefit-type attributes, the negative ideal solution takes the minimum value of the attribute; for cost-type attributes, the negative ideal solution takes the maximum value of the attribute. The negative ideal solution represents the worst performance of the candidate solution in all attributes and is an ideal choice that should be avoided as much as possible.

[0151] The relative closeness specifically refers to an indicator that measures the degree of closeness of a candidate solution to the positive ideal solution and the negative ideal solution, and is used to evaluate the pros and cons of the candidate solutions. The larger the relative closeness, the closer the solution is to the positive ideal solution and the farther it is from the negative ideal solution, and the higher the degree of pros and cons. Conversely, the smaller the relative closeness, the more the solution deviates from the positive ideal solution and the closer it is to the negative ideal solution, and the lower the degree of pros and cons.

[0152] Identify the mold that needs to be replaced through the intelligent mold switching algorithm; obtain the model and status information of the mold to be replaced, such as mold number, number of uses, degree of wear, etc.;

[0153] According to the model of the mold to be replaced, the matching replacement mold is searched in the mold database; the retrieved replacement molds are filtered to select candidate molds that meet the requirements; the filtering indicators include: the mold status score is used to evaluate the wear degree and maintenance records of the mold to obtain the mold status score; the remaining service life is used to estimate the remaining number of times or time of use of the mold and select a mold with sufficient service life; the compatibility index is used to consider the compatibility of the mold with the production equipment and process parameters to ensure smooth switching;

[0154] According to the preset filtering quantity, such as the top 10, determine the candidate mold set;

[0155] According to the preset mold selection benchmark factors, multiple decision attributes are determined, such as mold quality, switching efficiency, cost, etc.; a hierarchical model is constructed to stratify the decision attributes according to importance; the decision attributes are compared pairwise, and the judgment matrix is ​​determined based on domain knowledge and historical data; the maximum eigenvalue and corresponding eigenvector of the judgment matrix are calculated, and the eigenvector is normalized to obtain the attribute weight of the decision attribute;

[0156] Normalize the decision attributes, convert the attribute values ​​of different dimensions into dimensionless relative values, construct a decision matrix, each row represents a candidate mold, each column represents a decision attribute, and based on the attribute weights, add weights to the decision matrix to form a weighted normalized decision matrix; determine the positive ideal solution and the negative ideal solution, wherein the positive ideal solution represents the optimal value of each attribute, and the negative ideal solution represents the worst value of each attribute; calculate the Euclidean distance between the candidate mold and the positive ideal solution and the negative ideal solution respectively; calculate the relative closeness of the candidate mold based on the Euclidean distance, that is, the closer the mold is to the positive ideal solution and the farther the mold is from the negative ideal solution, the higher the ranking is, and generate the ranking of the candidate molds according to the relative closeness;

[0157] Taking maximizing production efficiency as the first goal, establishing an objective function, taking into account factors such as mold switching time and processing time; taking maximizing mold utilization as the second goal, establishing an objective function, taking into account factors such as the number of mold uses and maintenance frequency, and constructing a multi-objective optimization model; setting decision variables and constraints, such as mold selection, and constraints such as mold life and compatibility, using a multi-objective optimization algorithm, preferably using a genetic algorithm to solve the model, and generating an optimal mold replacement solution set;

[0158] Based on the ranking of candidate molds, the highest-ranked candidate mold is selected as the optimal mold from the mold optimal replacement solution set. Combined with the mold information in the mold database, the optimal mold selection and replacement plan is generated. The mold information includes mold location, instructions for use, replacement temperature, etc.; guide on-site personnel or robotic arms to perform mold replacement.

[0159] In this embodiment, by searching the mold database, a replacement mold that matches the model of the mold to be replaced can be found quickly and accurately; by constructing a hierarchical model and comparing two by two, the weights of multiple decision attributes can be objectively determined, and the importance of each attribute is fully considered; by calculating the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, the relative proportion of the attribute weights is determined, and the accuracy and credibility of the decision attribute weights are improved; based on the relative closeness calculated by Euclidean distance, the degree of proximity between the candidate mold and the ideal solution can be accurately evaluated, and a ranking of the candidate molds is generated, which can help decision makers select the optimal mold more quickly and improve the efficiency and accuracy of mold replacement decisions.

[0160] Figure 2 FIG. 1 is a schematic diagram of the structure of a rapid mold change system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0161] The first unit is used to obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, and optimize the hyperparameters of the mold change time prediction model through search and verification to predict the optimal mold change time;

[0162] The second unit is used to trigger a dynamic adjustment mechanism of production parameters when the difference between the optimal mold change time and the current time is less than a preset trigger threshold. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization;

[0163] The third unit is used to identify the mold to be replaced by an intelligent mold switching algorithm when the optimal mold change time is reached, obtain the model and status information of the mold to be replaced, and generate an optimal mold selection and replacement plan based on the model and the status information, and perform mold change operations according to the optimal mold selection and replacement plan.

[0164] According to a third aspect of the embodiments of the present invention,

[0165] An electronic device is provided, comprising:

[0166] processor;

[0167] a memory for storing processor-executable instructions;

[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] According to a fourth aspect of the embodiments of the present invention,

[0170] 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.

[0171] 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.

[0172] 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. A rapid mold change method, characterized in that: include: Acquire historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, optimize the hyperparameters of the mold change time prediction model through search and verification, and predict the optimal mold change time; When the difference between the optimal mold change time and the current time is less than a preset trigger threshold, the dynamic adjustment mechanism of production parameters is triggered. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization; When the optimal mold change time is reached, the mold to be changed is identified by an intelligent mold switching algorithm, the model and status information of the mold to be changed are obtained, and based on the model and status information, an optimal mold selection and replacement plan is generated, and a mold change operation is performed according to the optimal mold selection and replacement plan; Obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, optimize the hyperparameters of the mold change time prediction model through search verification, and predict the optimal mold change time including: Performing feature engineering on the historical mold change data to obtain a feature list; Analyze each feature in the feature list to determine mutually exclusive features, combine the mutually exclusive features into a feature bundle, and combine the independent features corresponding to the non-mutually exclusive features into a feature point set; Extracting continuous feature points with continuous values ​​from the feature point set, calculating the value range of each of the continuous feature points, dividing the value range into a plurality of segmented intervals, and combining the discrete intervals corresponding to the discontinuous feature points with discontinuous values ​​in the feature point set into a feature interval; Based on the historical mold change data, determine the training samples of the mold change time prediction model, initialize the parameters of the mold change time prediction model, and determine the initial prediction result; Based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached to determine the final mold change time prediction model; By means of grid search and cross validation, the hyper parameters of the final mold change time prediction model are tuned to obtain the optimal mold change time prediction model, and the optimal mold change time is predicted; Based on the initial prediction result, a feature point histogram is constructed, a decision tree is generated by maximizing the node splitting gain, the decision tree is added to the mold change time prediction model, the prediction result is obtained again, and a new decision tree is iteratively generated until a preset number of iterations is reached. Determining the final mold change time prediction model includes: Based on the initial prediction result, the gradient and the second-order partial derivative matrix of each training sample are calculated, and for each feature interval of each feature point, the total number of training samples in each feature interval, the sum of the gradients corresponding to the total number of training samples, and the sum of the second-order partial derivative matrices corresponding to the total number of training samples are counted to construct a feature point histogram; Initialize the root node of the decision tree and put all the training samples into the root node; For each non-leaf node, traverse all splitting points of all feature points, calculate the splitting gain, select the feature point and splitting point corresponding to the maximum value of the splitting gain as the optimal splitting, divide the samples into left and right child nodes according to the optimal splitting, update the feature point histogram of the child node, if the depth corresponding to the child node reaches the preset maximum depth, or the number of training samples in the child node is less than the preset sample threshold, mark the child node as a leaf node, stop node splitting, and calculate the predicted value of the leaf node; until all nodes are marked as leaf nodes, determine the decision tree; The decision tree is added to the mold change time prediction model to obtain a new prediction result, and a new decision tree is generated based on the new prediction result until a preset number of iterations is reached.

2. The method according to claim 1, characterized in that The dynamic adjustment mechanism of production parameters builds a production parameter optimization algorithm to generate an optimal control strategy based on the current production status with the goal of maximizing production efficiency and mold utilization. The production parameters are adjusted as follows: Based on the working process of the production system, the state variables are determined, and the state variables are combined into high-order features through feature combinations. The state variables and the high-order features constitute a state space; based on the control actions of the production system, the adjustment range and adjustment granularity of the production parameters are determined, the adjustment range and the adjustment granularity are discretized, and discrete adjustment values ​​are determined. The operation actions corresponding to the discrete adjustment values ​​constitute an action space; The production parameter optimization algorithm is based on a deep neural network, constructing an input layer for receiving the current production state, multiple hidden layers for extracting state features, and an output layer for outputting the quality value corresponding to each manipulation action, generating a decision quality network, and creating a target quality network with the same structure as the decision quality network; By minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network, the parameters of the decision quality network are updated; according to a preset replication interval step, the parameters of the decision quality network are copied to the target quality network, and finally an optimal target quality network is generated; The optimal target quality network determines the optimal production parameters according to the input status.

3. The method according to claim 2, characterized in that Updating the parameters of the decision quality network by minimizing the error between the quality value output by the target quality network and the quality value output by the decision quality network includes: Randomly initialize the weights of the decision quality network and the target quality network, and set the parameters of the target quality network to be copies of the corresponding parameters of the decision quality network; According to the current production state received by the input layer, the decision quality value of each operation action is calculated through forward propagation of the decision quality network, and the maximum decision quality value is extracted as the current decision quality value. The operation action corresponding to the current decision quality value is used as the current decision action corresponding to the current production state, forming a current state action pair; According to the current state-action pair, a next state and an immediate reward are obtained, the next state is input into a target quality network, a target quality value of each operation action corresponding to the next state is calculated, a maximum target quality value is extracted, and a current target quality value of the current state-action pair is determined based on the immediate reward and the maximum target quality value; Based on the current target quality value and the current decision quality value, the TD error is calculated, and the parameters of the decision quality network are updated through back propagation with the goal of minimizing the mean square error between the decision quality value and the target quality value.

4. The method according to claim 3, characterized in that Also includes: Selecting an operation action according to the current state and a greedy strategy, wherein the greedy strategy includes: generating a random number between 0 and 1, and if the random number is less than a preset exploration rate, randomly selecting an operation action as a decision action; if the random number is greater than or equal to the exploration rate, selecting an operation action corresponding to the maximum decision quality value of the decision quality network as the decision action; wherein the exploration rate decreases as the number of iterations increases; Execute the decision action, transfer the current state to the next state, generate state transfer data, obtain an immediate reward, and store the state transfer data in an experience playback buffer, wherein the state transfer data includes the current state, the decision action, the immediate reward, and the next state; A set of state transition data is randomly sampled from the experience replay buffer, and for each of the state transition data, a sampled target quality value is calculated; if the next state is a terminal state, the sampled target quality value is equal to the immediate reward; otherwise, the sampled target quality value is equal to the weighted sum of the immediate reward and the maximum target quality value output by the target quality network of the next state.

5. The method according to claim 1, characterized in that: The mold to be replaced is identified by an intelligent mold switching algorithm, the model and status information of the mold to be replaced are obtained, and based on the model and status information, an optimal mold selection and replacement plan is generated, including: According to the model of the mold to be replaced, matching replacement molds are searched in the mold database, the replacement molds are filtered according to the mold status score, the remaining service life and the compatibility index, and a candidate mold set is determined according to a preset filtering quantity; According to the preset mold selection benchmark factors, multiple decision attributes are determined, and the decision attributes are compared in pairs by constructing a hierarchical model to determine a judgment matrix, calculate the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, and determine the attribute weights of the decision attributes; Normalizing the decision attributes, constructing a decision matrix, adding weights to the decision matrix based on the attribute weights to form a weighted normalized decision matrix, determining a positive ideal solution and a negative ideal solution, respectively calculating the Euclidean distances between the candidate mold and the positive ideal solution and the negative ideal solution, determining the relative closeness of the candidate mold based on the Euclidean distances, and generating a ranking of the candidate molds; With maximizing production efficiency as the first goal and maximizing mold utilization as the second goal, a multi-objective optimization algorithm is constructed to generate the optimal mold replacement solution set; Based on the ranking of the candidate molds, the candidate mold with the highest ranking is selected from the set of optimal mold replacement solutions to obtain the optimal mold. The optimal mold selection and replacement plan is generated in combination with the mold information in the mold database.

6. A rapid mold change system, used to implement the method according to any one of claims 1 to 5, characterized in that: include: The first unit is used to obtain historical mold change data, pre-process the historical mold change data, establish a nonlinear correlation between production parameters, mold service life, product quality and mold change time based on the historical mold change data, determine a mold change time prediction model, and optimize the hyperparameters of the mold change time prediction model through search and verification to predict the optimal mold change time; The second unit is used to trigger a dynamic adjustment mechanism of production parameters when the difference between the optimal mold change time and the current time is less than a preset trigger threshold. The dynamic adjustment mechanism of production parameters generates an optimal control strategy and adjusts production parameters by constructing a production parameter optimization algorithm according to the current production status with the goal of maximizing production efficiency and mold utilization; The third unit is used to identify the mold to be replaced by an intelligent mold switching algorithm when the optimal mold change time is reached, obtain the model and status information of the mold to be replaced, and generate an optimal mold selection and replacement plan based on the model and the status information, and perform mold change operations according to the optimal mold selection and replacement plan.

7. 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 described in any one of claims 1 to 5.

8. 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 5 is implemented.

Images