Energy model-based die-casting die heat balance distribution optimization method and system

Abstract

The invention provides a die-casting die heat balance distribution optimization method and system based on an energy model, and relates to the technical field of die-casting dies, and the method comprises the steps: collecting heat distribution data, and building a heat energy transfer mapping relation; extracting space-time correlation features by using a graph attention network; constructing a heat distribution dynamic evolution model; constructing a temperature field predictor based on a deep belief network; training a predictor by using a multi-agent coevolution algorithm; constructing a temperature field optimization strategy network; generating an optimal temperature field control sequence; establishing a regional adaptive control strategy; constructing a temperature field controller based on deep reinforcement learning; and a dynamic fuzzy immune optimization algorithm is adopted to adjust cooling control parameters, and intelligent optimization control over heat distribution of the die-casting die is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of die-casting dies, and in particular to a method and system for optimizing heat balance distribution of a die-casting die based on an energy model. Background Art

[0002] Die casting is an efficient metal parts manufacturing process. The uniformity and stability of the mold temperature field directly affect the quality of die castings. The heat transfer process of the die casting mold involves complex thermal-mechanical coupling effects, including heat conduction, convection heat transfer, radiation heat transfer and other heat transfer methods between different areas of the mold. As the requirements of the die casting process continue to increase, higher requirements are placed on the control of the mold temperature field. The traditional empirical temperature control method can no longer meet the needs of high-quality die casting production.

[0003] In the existing technology, the temperature field control technology of die-casting molds has the following main deficiencies: the traditional thermodynamic model is too simplified, resulting in insufficient accuracy in heat transfer modeling; there is a lack of effective extraction and characterization methods for the spatiotemporal evolution characteristics of the temperature field, making it difficult to accurately describe the dynamic interaction relationship between temperature changes in various regions of the mold; it is difficult to handle the coupling relationship between multiple regions, and it is impossible to achieve balanced control of the global temperature field; fixed control rules are used, lack of adaptive adjustment capabilities, and it is difficult to cope with process fluctuations and uncertainty interference in the production process.

[0004] To sum up, in view of the problems existing in the prior art, there is an urgent need for a method for optimizing the thermal balance distribution of die-casting molds based on an energy model, which can realize accurate modeling of the heat energy transfer process of the die-casting mold, provide an effective method for extracting the spatiotemporal characteristics of the temperature field, realize collaborative optimization control of multi-region temperature fields, and establish a real-time control strategy with adaptive capabilities. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for optimizing the heat balance distribution of a die-casting mold based on an energy model, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, A method for optimizing heat balance distribution of a die casting mold based on an energy model is provided, comprising: Collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation characteristics of the die-casting mold; establish a nonlinear state equation of the heat distribution of the die-casting mold based on the spatiotemporal correlation characteristics, use a hybrid particle swarm quantum genetic algorithm to solve the nonlinear state equation, and construct a dynamic evolution model of the heat distribution of the die-casting mold; The heat distribution dynamic evolution model is input into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; a multi-agent collaborative evolution algorithm is used to train the temperature field predictor to determine a hierarchical optimization function of the die-casting mold temperature field; a temperature field optimization strategy network for the die-casting mold is constructed based on the hierarchical optimization function, and an adaptive ant colony optimization algorithm based on taboo search is used to generate an optimal temperature field control sequence for the die-casting mold; a sub-region adaptive control strategy for the die-casting mold is established according to the optimal temperature field control sequence; The sub-region adaptive control strategy is input into the dynamic evolution model of heat distribution to obtain the temperature field response characteristics of the die-casting mold; according to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are obtained in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the dynamic fuzzy immune optimization algorithm is used to adjust the cooling control parameters to realize the intelligent optimization control of the heat distribution of the die-casting mold.

[0007] In an optional embodiment, Collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into the graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation features of the die-casting mold, including: Construct a temperature sensor grid array on the surface of the die-casting mold, wherein the temperature sensor grid array includes a preset number of temperature acquisition points, records temperature values, acquisition timestamps, and spatial coordinate information; continuously acquires temperature data at a preset sampling frequency based on the temperature sensor grid array, and organizes the temperature data into a time series data set; calculates the temperature gradient vector between adjacent temperature acquisition points in the time series data set, calculates the heat flux density, and establishes a heat energy transfer mapping relationship based on the temperature gradient vector and the heat flux density; The time series data set is subjected to Fourier transformation to obtain the frequency domain characteristics of the temperature field, and the frequency domain characteristics of the temperature field are combined with the temperature gradient vector to construct a first mapping feature vector; the thermal conductivity coefficient distribution is calculated based on the thermal energy transfer mapping relationship, and the thermal conductivity coefficient distribution is combined with the first mapping feature vector to construct a second mapping feature vector; the second mapping feature vectors are arranged according to the spatial distribution of the temperature collection points to construct a feature vector matrix; A graph structure is constructed based on the spatial distribution of the temperature sensor grid array, the temperature collection points are used as graph structure nodes, and the thermal conductivity coefficient matrix is ​​used as graph structure edge weights; an unsupervised adversarial learning network is constructed, wherein the generator of the unsupervised adversarial learning network generates simulated heat transfer data based on the feature vector matrix, and the corresponding discriminator compares the simulated heat transfer data with the time series data set for training; Based on the graph structure, the attention weight coefficients between nodes are calculated, and the attention weight coefficients are weightedly fused with the thermal energy transfer mapping relationship to obtain the heat conduction weights; the feature vector matrix is ​​reconstructed according to the heat conduction weights to obtain the spatiotemporal correlation features.

[0008] In an optional embodiment, Based on the time-space correlation characteristics, a nonlinear state equation of the heat distribution of the die-casting mold is established, and a hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation. The construction of a dynamic evolution model of the heat distribution of the die-casting mold includes: Based on the time-space correlation characteristics, the heat spatial diffusion rate, the heat time cumulative change rate, and the mold material heat transfer coefficient are extracted and calculated, and the state variable vector is constructed in combination; based on the heating power parameters and cooling rate parameters of the die-casting mold, a control variable vector is generated; and the state variable vector and the control variable vector are used to establish a nonlinear state equation that describes the heat transfer law inside the die-casting mold; Converting the solution parameters of the nonlinear state equation into quantum state codes, constructing an initial particle swarm including quantum position genes and quantum speed genes, substituting the quantum position genes and quantum speed genes of the initial particle swarm into the nonlinear state equation, calculating the fitness value, and determining the individual optimal solution and the global optimal solution of the initial particle swarm based on the fitness value; Based on the individual optimal solution and the global optimal solution, a quantum rotation angle is calculated, based on the quantum rotation angle, a quantum rotation gate operation is performed on the quantum position gene to obtain a crossover quantum position gene, and a quantum NOT gate operation is performed on the quantum speed gene to obtain a mutated quantum speed gene; the crossover quantum position gene and the mutated quantum speed gene are combined to construct an optimized particle swarm; Substituting the quantum position gene and the quantum speed gene of the optimized particle swarm into the nonlinear state equation to calculate a new fitness value, updating the individual optimal solution and the global optimal solution according to the new fitness value, and repeating the iterative update until the change rate of the new fitness value is less than a preset convergence threshold, thereby obtaining the optimal solution; According to the optimal solution, the temperature state transfer matrix of the die-casting mold in a discrete time series is constructed, and the heat transfer response function between the various regions of the die-casting mold is established based on the temperature state transfer matrix. The temperature state transfer matrix is ​​combined with the heat transfer response function to construct a dynamic evolution model of the thermal distribution of the die-casting mold.

[0009] In an optional embodiment, The heat distribution dynamic evolution model is input into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; a multi-agent collaborative evolution algorithm is used to train the temperature field predictor to determine the hierarchical optimization function of the die-casting mold temperature field, including: Converting the temperature state transfer matrix in the thermal distribution dynamic evolution model into a characteristic graph, converting the heat transfer response function in the thermal distribution dynamic evolution model into a time series vector, normalizing the characteristic graph and the time series vector, and obtaining input data of a die-casting mold temperature field predictor; Construct a knowledge distillation framework consisting of a teacher network and a student network, wherein the teacher network adopts a deep residual network structure, and the student network adopts a three-layer deep belief network structure. The input data is input into the teacher network to obtain a soft label of the temperature field distribution. The soft label of the temperature field distribution is used as a training target of the student network. Each layer of the student network is pre-trained, and the restricted Boltzmann machine of each layer is trained using a contrastive divergence algorithm. The trained student network is used as the basic network structure of the die-casting mold temperature field predictor to obtain the die-casting mold temperature field predictor. The die-casting mold temperature field predictor is divided into multiple sub-predictors, each of which is responsible for the temperature prediction of its own corresponding area, and an information transmission channel is established between the sub-predictors through an attention mechanism; a multi-agent collaborative evolution algorithm is used to train the sub-predictors, and a fitness function including local prediction error and prediction consistency is constructed. The prediction performance of the sub-predictors is evaluated based on the fitness function, and the prediction performance is used as an evolutionary indicator. Based on the evolutionary indicator, the network parameters of each sub-predictor are optimized to obtain the optimized temperature field prediction performance indicator; a hierarchical optimization function of the die-casting mold temperature field is constructed according to the temperature field prediction performance indicator.

[0010] In an optional embodiment, Based on the hierarchical optimization function, a temperature field optimization strategy network for the die-casting mold is constructed, and an adaptive ant colony optimization algorithm based on taboo search is used to generate an optimal temperature field control sequence for the die-casting mold, including: Based on the hierarchical optimization function, a temperature field optimization strategy network for the casting mold is constructed, the temperature field optimization strategy network is divided into a temperature field uniformity layer, a cooling efficiency layer and an energy consumption control layer, and an evaluation index function is determined; Initialize the position of the ant colony, encode the temperature field control parameters of the die-casting mold into ant position information, determine the temperature field control sequence, input the temperature field control sequence corresponding to each ant position into the temperature field optimization strategy network, calculate the path evaluation value through the evaluation index function, and update the pheromone concentration on the path according to the path evaluation value, wherein the pheromone increment is positively correlated with the path evaluation value; Constructing a dynamic taboo table, recording the searched temperature field control sequences, calculating the similarity between the newly generated temperature field control sequence and the temperature field control sequence in the taboo table, and adding the temperature field control sequence with a similarity higher than a preset threshold to the dynamic taboo table; Based on the pheromone concentration on the path and the dynamic taboo table, the state transition probability is calculated to guide the search direction of the ant colony; A variable neighborhood search structure is constructed near the current position of each ant, wherein the variable neighborhood search structure includes a single parameter adjustment space and a multi-parameter linkage adjustment space, and the parameters in the temperature field control sequence are adjusted independently and in combination respectively; the candidate solutions corresponding to the temperature field control sequence generated in the variable neighborhood search structure are input into the temperature field optimization strategy network, the path evaluation value of each candidate solution is calculated, the temperature field control sequence with the highest path evaluation value is selected, and the ant position is updated; Repeat the iteration until the preset maximum number of iterations is reached to obtain the optimal temperature field control sequence.

[0011] In an optional embodiment, The sub-region adaptive control strategy is input into the dynamic evolution model of heat distribution to obtain the temperature field response characteristics of the die-casting mold; according to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are obtained in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the cooling control parameters are adjusted by a dynamic fuzzy immune optimization algorithm to realize the intelligent optimization control of the heat distribution of the die-casting mold, including: The regional adaptive control strategy is input into the dynamic evolution model of heat distribution, the temperature field state equation of the die-casting mold is established based on the law of conservation of energy, the Kalman filter algorithm is used to perform state estimation on the temperature field state equation, and the temperature field response characteristics of the die-casting mold are obtained by iteratively updating the state estimation error covariance matrix and the Kalman gain matrix; A temperature field controller based on deep reinforcement learning is constructed according to the temperature field response characteristics. The temperature field controller adopts a dual network architecture, including a policy network for generating control actions and a value network for evaluating state values. The state space of the policy network and the value network includes the current temperature distribution, temperature change rate and control input history, and the action space includes the control parameter adjustment amount of each area. The policy network and the value network are trained using the dominant actor-critic algorithm, and an immediate reward function is constructed based on the temperature uniformity value and the control stability value. The temperature field change parameters of the die-casting mold are obtained by optimizing the policy loss function and the value loss function. The temperature field change parameters are input into the temperature field predictor for online correction, a dynamic fuzzy rule base is constructed according to the corrected prediction results, and fuzzy control rules are designed based on temperature deviation and temperature change rate; a dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, the adjustment strategy set is constructed into an antibody group, and the antibody affinity is calculated based on the temperature control error. Through cloning mutation operations, the antibody group is optimized, and finally the optimal cooling control parameters are obtained, and the thermal distribution of the die-casting mold is intelligently adjusted.

[0012] In an optional embodiment, A dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, which is constructed as an antibody group. The antibody affinity is calculated based on the temperature control error. The antibody group is optimized through cloning and mutation operations, and the optimal cooling control parameters are finally obtained. Intelligent adjustment of the heat distribution of the die-casting mold includes: Acquiring cooling control parameters of the die-casting mold, encoding the cooling control parameters to generate an initial antibody group, each antibody individual comprising a parameter vector corresponding to the cooling control parameters; Collecting real-time temperature data of each measuring point of the die-casting mold, calculating the temperature deviation and the temperature deviation change rate according to the real-time temperature data, taking the temperature deviation and the temperature deviation change rate as input variables, establishing a dynamic fuzzy rule mapping relationship, and constructing a fuzzy rule library for temperature field control; Based on the square sum of the deviations between the real-time temperature data and the target temperature, a temperature control error index is constructed, the temperature control error index is converted into an antibody affinity value through an exponential function, the exponential function includes an adjustment coefficient for controlling the selection pressure, and the antibody individuals in the antibody group are sorted from high to low according to the antibody affinity value; according to the preset rule of decreasing clone number, the clone number is sequentially assigned to each antibody individual to generate a clone antibody group, and a Gaussian mutation operation is performed on the clone antibody group, the variation step length of the Gaussian mutation is inversely proportional to the antibody affinity value, and a variant antibody group is obtained; Calculating the affinity value of each antibody individual in the variant antibody group, determining candidate variant antibody individuals from the variant antibody group according to a preset affinity selection threshold, and determining candidate antibody individuals from the initial antibody group, merging the candidate variant antibody individuals and the candidate candidate antibody individuals to generate a new generation antibody group, and at the same time, introducing randomly generated new antibody individuals into the new generation antibody group according to a preset number; A new generation of antibody groups is generated by repeated iterations until a preset number of iterations is reached, the antibody individual corresponding to the maximum affinity value is selected, the optimal antibody individual is determined, the optimal antibody individual is decoded, the optimal cooling control parameters are obtained, and the heat distribution of the die-casting mold is adjusted.

[0013] According to a second aspect of the embodiments of the present invention, A heat balance distribution optimization system for die casting molds based on an energy model is provided, comprising: The first unit is used to collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation characteristics of the die-casting mold; establish a nonlinear state equation of the heat distribution of the die-casting mold based on the spatiotemporal correlation characteristics, use a hybrid particle swarm quantum genetic algorithm to solve the nonlinear state equation, and construct a dynamic evolution model of the heat distribution of the die-casting mold; The second unit is used to input the dynamic evolution model of heat distribution into a deep belief network based on knowledge distillation to construct a temperature field predictor for a die-casting mold; adopt a multi-agent collaborative evolution algorithm to train the temperature field predictor to determine a hierarchical optimization function of the temperature field of the die-casting mold; construct a temperature field optimization strategy network for the die-casting mold based on the hierarchical optimization function, and adopt an adaptive ant colony optimization algorithm based on taboo search to generate an optimal temperature field control sequence for the die-casting mold; and establish a sub-region adaptive control strategy for the die-casting mold according to the optimal temperature field control sequence; The third unit is used to input the sub-region adaptive control strategy into the dynamic evolution model of heat distribution to obtain the temperature field response characteristics of the die-casting mold; according to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are obtained in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the cooling control parameters are adjusted by a dynamic fuzzy immune optimization algorithm to realize intelligent optimization control of the heat distribution of the die-casting mold.

[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the embodiments of the present invention is: 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.

[0016] In the embodiments of the present invention, the heat distribution can be predicted more accurately through the extraction of spatiotemporal correlation features and nonlinear modeling; the dynamic evolution model describes the spatiotemporal changes of the heat distribution, making the system more adaptable when adjusting the control strategy; the model supports precise control of the temperature field, reduces energy consumption, extends the life of the mold, and improves production efficiency; the temperature field predictor is constructed by a deep belief network based on knowledge distillation, which can more accurately predict the dynamic changes of the mold temperature field; the temperature field optimization strategy network is trained through a multi-agent collaborative evolution algorithm and a hierarchical optimization function to improve the accuracy and adaptability of temperature control; an adaptive ant colony taboo search algorithm is used to generate the optimal temperature control sequence, and formulate regional adaptive control strategies based on this, to achieve precise thermal management of the mold and further improve the quality of die-casting; real-time monitoring and feedback adjustment significantly improve the control accuracy of the mold temperature field, ensure the uniform distribution of temperature in each area, and effectively avoid mold loss and product defects caused by thermal unevenness; the correction function of the temperature field response characteristics and the real-time data input predictor enables the control system to adapt to temperature changes in the die-casting process, and improves the system's response speed and robustness to fluctuations; the dynamic fuzzy immune optimization algorithm automatically adjusts the cooling control parameters according to real-time data, realizes the intelligence and efficiency of the cooling process, and reduces cooling time and energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of a method for optimizing heat balance distribution of a die casting mold based on an energy model according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of a die-casting mold heat balance distribution optimization system based on an energy model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] 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.

[0019] 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.

[0020] Figure 1 FIG. 1 is a flow chart of a method for optimizing heat balance distribution of a die casting mold based on an energy model according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation characteristics of the die-casting mold; establish a nonlinear state equation of the heat distribution of the die-casting mold based on the spatiotemporal correlation characteristics, use a hybrid particle swarm quantum genetic algorithm to solve the nonlinear state equation, and construct a dynamic evolution model of the heat distribution of the die-casting mold; Collect heat distribution data on the surface and inside of the die-casting mold, capture the temperature change information of the mold under different working conditions, and ensure the integrity and accuracy of the data; establish a mapping relationship of heat energy transfer inside and outside the mold based on the collected heat distribution data, and describe the transfer rules of heat flow and temperature in different areas; mathematically model the heat energy transfer mapping relationship and convert it into a feature vector matrix so that it can be effectively input into the subsequent feature extraction network.

[0021] The feature vector matrix is ​​input into the graph attention network based on unsupervised adversarial learning to extract spatiotemporal correlation features and capture the dynamic changes of the mold's thermal distribution and the correlation between regions. Based on the extracted spatiotemporal correlation features, a nonlinear state equation is established to describe the thermal distribution state of the mold at different times and regions for subsequent solution and prediction. A hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation in order to construct a dynamic evolution model to accurately describe the changes in the mold's thermal distribution over time. A dynamic evolution model of thermal distribution is constructed: The solution results are used to construct a dynamic evolution model of thermal distribution to comprehensively describe the spatiotemporal changes in the thermal distribution inside the mold, providing a basis for subsequent temperature control strategies.

[0022] In this embodiment, through the extraction of spatiotemporal correlation features and nonlinear modeling, the heat distribution can be predicted more accurately; the dynamic evolution model describes the spatiotemporal changes of the heat distribution, making the system more adaptable when adjusting the control strategy; the model supports precise control of the temperature field, reduces energy consumption, extends mold life, and improves production efficiency.

[0023] In an optional embodiment, heat distribution data of the die-casting mold is collected, a heat energy transfer mapping relationship of the die-casting mold is established, the heat energy transfer mapping relationship is converted into a feature vector matrix, and the feature vector matrix is ​​input into a graph attention network based on unsupervised adversarial learning for feature extraction, and the spatiotemporal correlation features of the die-casting mold are obtained, including: Construct a temperature sensor grid array on the surface of the die-casting mold, wherein the temperature sensor grid array includes a preset number of temperature acquisition points, records temperature values, acquisition timestamps, and spatial coordinate information; continuously acquires temperature data at a preset sampling frequency based on the temperature sensor grid array, and organizes the temperature data into a time series data set; calculates the temperature gradient vector between adjacent temperature acquisition points in the time series data set, calculates the heat flux density, and establishes a heat energy transfer mapping relationship based on the temperature gradient vector and the heat flux density; The time series data set is subjected to Fourier transformation to obtain the frequency domain characteristics of the temperature field, and the frequency domain characteristics of the temperature field are combined with the temperature gradient vector to construct a first mapping feature vector; the thermal conductivity coefficient distribution is calculated based on the thermal energy transfer mapping relationship, and the thermal conductivity coefficient distribution is combined with the first mapping feature vector to construct a second mapping feature vector; the second mapping feature vectors are arranged according to the spatial distribution of the temperature collection points to construct a feature vector matrix; A graph structure is constructed based on the spatial distribution of the temperature sensor grid array, the temperature collection points are used as graph structure nodes, and the thermal conductivity coefficient matrix is ​​used as graph structure edge weights; an unsupervised adversarial learning network is constructed, wherein the generator of the unsupervised adversarial learning network generates simulated heat transfer data based on the feature vector matrix, and the corresponding discriminator compares the simulated heat transfer data with the time series data set for training; Based on the graph structure, the attention weight coefficients between nodes are calculated, and the attention weight coefficients are weightedly fused with the thermal energy transfer mapping relationship to obtain the heat conduction weights; the feature vector matrix is ​​reconstructed according to the heat conduction weights to obtain the spatiotemporal correlation features.

[0024] Specifically, a temperature sensor grid array is constructed on the surface of the die-casting mold. The temperature sensor grid array adopts a 12x12 matrix layout, with a spacing of 50 mm between adjacent sensors. Each sensor node includes a temperature acquisition unit, a time synchronization unit, and a three-dimensional coordinate positioning unit. The measurement range of the temperature acquisition unit is 0-800 degrees Celsius, and the measurement accuracy is ±0.1 degrees Celsius; the time synchronization unit records millisecond-level timestamps; and the three-dimensional coordinate positioning unit records the spatial position coordinates of the sensor on the mold surface, with a coordinate accuracy of 0.1 mm.

[0025] Temperature data is collected continuously at a sampling frequency of 100 Hz. Each sensor node collects 1,000 temperature data points in one die-casting cycle to form a multi-dimensional time series data set containing timestamps, spatial coordinates, and temperature values. The temperature difference between adjacent temperature collection points is calculated to obtain the temperature gradient components in the x and y directions, and the temperature gradient vector is obtained in combination with the sensor spacing. The heat flux density is calculated based on the temperature gradient vector, and the heat energy transfer mapping function is established through the ratio relationship between the temperature gradient vector and the heat flux density.

[0026] Fast Fourier transform is performed on the time series data set to extract the frequency domain features of the temperature field, including the main frequency component, harmonic component, and phase information. The frequency domain features are mapped one by one to the temperature gradient vector according to the sensor node position to construct a 144-dimensional first mapping feature vector. Based on the heat energy transfer mapping function, a 144x144 thermal conductivity coefficient distribution matrix is ​​calculated, and the thermal conductivity coefficient distribution matrix is ​​combined with the first mapping feature vector to construct a second mapping feature vector containing thermal conductivity characteristics. The second mapping feature vector is rearranged according to the 12x12 spatial distribution of the temperature sensor to form a feature vector matrix.

[0027] A graph structure is constructed based on the temperature sensor grid array, with 144 sensor nodes as the vertices of the graph. Connection edges are established between adjacent nodes, and the weight of the edge is set to the corresponding thermal conductivity coefficient. An adversarial learning network architecture consisting of a generator and a discriminator is constructed. The generator inputs a feature vector matrix and generates a simulated heat transfer data sequence through a multi-layer convolutional neural network. The discriminator compares the generated simulated data with the actual collected time series data set, calculates the discrimination loss, and back-propagates to optimize the network parameters.

[0028] In the graph structure, the attention score between each pair of connected nodes is calculated to obtain a 144x144 attention weight coefficient matrix. The attention weight coefficient is weighted and combined with the heat transfer mapping function, with weight coefficients of 0.6 and 0.4, to obtain the comprehensive heat conduction weight. The heat conduction weight is used to weight each element in the feature vector matrix to generate a spatiotemporal feature representation that integrates the time evolution feature and the spatial correlation feature.

[0029] In practical applications, for a typical die-casting mold, a single die-casting cycle is 30 seconds, and a 144x1000 temperature time series data matrix is ​​collected. After Fourier transform, the main frequency component is 0.033 Hz, containing 3-5 significant harmonic components. The amplitude range of the temperature gradient vector is 0.5-2.0 degrees Celsius / mm, and the numerical range of the heat flux density is 50-200 kilowatts / square meter. The final reconstructed spatiotemporal characteristics can accurately characterize the dynamic evolution law of the temperature field in the die-casting process, providing a basis for the precise control of the mold temperature field.

[0030] In this embodiment, the sensor grid array and high-precision data acquisition ensure the detailed monitoring of the temperature field and capture the subtle changes in the surface temperature of the die-casting mold; through the calculation of temperature gradient and heat flux density, the heat energy transfer mapping is accurately established, and the accuracy of the temperature field model is improved; frequency domain analysis and adversarial learning network are combined with graph attention mechanism to generate spatiotemporal feature representation, which helps to have a deeper understanding of the dynamic changes of the temperature field in time and space; the generated high-dimensional feature vector matrix provides comprehensive spatiotemporal features, which provides an important reference basis for the precise control and adjustment of the temperature field, improves the mold quality and prolongs the mold life; through fast Fourier transform and feature vector rearrangement, the system can efficiently process a large amount of temperature data, extract the main frequency component and harmonic component, and effectively analyze the dynamic changes of the temperature field.

[0031] In an optional embodiment, a nonlinear state equation of the heat distribution of the die-casting mold is established based on the spatiotemporal correlation characteristics, and a hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation. Constructing a dynamic evolution model of the heat distribution of the die-casting mold includes: Based on the time-space correlation characteristics, the heat spatial diffusion rate, the heat time cumulative change rate, and the mold material heat transfer coefficient are extracted and calculated, and the state variable vector is constructed in combination; based on the heating power parameters and cooling rate parameters of the die-casting mold, a control variable vector is generated; and the state variable vector and the control variable vector are used to establish a nonlinear state equation that describes the heat transfer law inside the die-casting mold; Converting the solution parameters of the nonlinear state equation into quantum state codes, constructing an initial particle swarm including quantum position genes and quantum speed genes, substituting the quantum position genes and quantum speed genes of the initial particle swarm into the nonlinear state equation, calculating the fitness value, and determining the individual optimal solution and the global optimal solution of the initial particle swarm based on the fitness value; Based on the individual optimal solution and the global optimal solution, a quantum rotation angle is calculated, based on the quantum rotation angle, a quantum rotation gate operation is performed on the quantum position gene to obtain a crossover quantum position gene, and a quantum NOT gate operation is performed on the quantum speed gene to obtain a mutated quantum speed gene; the crossover quantum position gene and the mutated quantum speed gene are combined to construct an optimized particle swarm; Substituting the quantum position gene and the quantum speed gene of the optimized particle swarm into the nonlinear state equation to calculate a new fitness value, updating the individual optimal solution and the global optimal solution according to the new fitness value, and repeating the iterative update until the change rate of the new fitness value is less than a preset convergence threshold, thereby obtaining the optimal solution; According to the optimal solution, the temperature state transfer matrix of the die-casting mold in a discrete time series is constructed, and the heat transfer response function between the various regions of the die-casting mold is established based on the temperature state transfer matrix. The temperature state transfer matrix is ​​combined with the heat transfer response function to construct a dynamic evolution model of the thermal distribution of the die-casting mold.

[0032] Heat transfer characteristic parameters are extracted from the time-space correlation characteristics, including the heat spatial diffusion rate, the heat time cumulative change rate and the mold material heat transfer coefficient. The heat spatial diffusion rate reflects the speed of heat propagation between different areas of the mold, and the value range is 0.5-2.0 m / s; the heat time cumulative change rate represents the degree of heat accumulation per unit time, and the value range is 20-100 kJ / s; the mold material heat transfer coefficient describes the thermal conductivity of the material, and the typical value is 30-50 W / m·K. These three parameters are combined to form a nine-dimensional state variable vector, which is used to characterize the thermodynamic state of the mold.

[0033] Obtain the process control parameters of the die-casting mold, including heating power parameters and cooling rate parameters. The heating power parameters include heater power, heating time, and heating area distribution. The power range is 10-50 kilowatts and the heating time is 5-15 seconds. The cooling rate parameters include cooling medium flow, cooling channel size, and cooling time. The flow range is 10-30 liters / minute and the cooling time is 10-30 seconds. Integrate these parameters into a six-dimensional control variable vector. Based on the state variable vector and the control variable vector, establish a nonlinear state equation that describes the heat transfer law inside the die-casting mold.

[0034] The solution parameters of the nonlinear state equation are represented by quantum coding, and each parameter is encoded with 8-bit quantum bits. An initial particle swarm containing 100 particles is constructed, and each particle contains a quantum position gene and a quantum speed gene. The quantum position gene represents the parameter value, and the quantum speed gene represents the parameter change direction and step size. The gene information of the initial particle swarm is substituted into the nonlinear state equation to calculate the fitness value of each particle. The fitness value uses the inverse of the temperature field prediction error. Based on the fitness value, the individual historical optimal position of each particle and the global optimal position of the entire group are determined.

[0035] According to the difference between the individual optimal position and the global optimal position, the quantum rotation angle is calculated, and the angle range is 0.01π to 0.05π. The quantum rotation gate operation is performed on the quantum position gene to achieve gene crossover by adjusting the phase of the quantum bit; the quantum NOT gate operation is performed on the quantum speed gene to achieve gene mutation by flipping the quantum bit. The quantum position gene after crossover and the quantum speed gene after mutation are recombined to construct an optimized particle group.

[0036] Substitute the genetic information of the optimized particle swarm into the nonlinear state equation to calculate the new fitness value, and update the individual optimal position and the global optimal position. When the relative change rate of the fitness value in 50 consecutive iterations is less than 0.1%, the optimization process is considered to have converged and the optimal solution is obtained. A typical optimization process requires 200-300 iterations to converge.

[0037] Based on the optimal solution, a 100x100 temperature state transfer matrix is ​​constructed. The matrix elements represent the temperature transfer relationship between the various regions of the mold at adjacent moments. The eigenvalue distribution range of the temperature state transfer matrix is ​​0.85-1.15, reflecting the stability of the system. The temperature state transfer matrix is ​​used to establish the heat transfer response function between the various regions. The rise time of the response function is 1-3 seconds and the adjustment time is 5-8 seconds. The temperature state transfer matrix is ​​combined with the heat transfer response function to obtain a mathematical model that can accurately predict the dynamic evolution of the temperature field of the die-casting mold.

[0038] In actual application verification, the prediction error of the model for the mold temperature field is less than 3%, which can effectively capture the dynamic characteristics of heat transfer and provide a reliable theoretical basis for achieving accurate temperature field control. The temperature field distribution predicted by the model is in good agreement with the actual measurement results, which verifies the effectiveness and practicality of the modeling method.

[0039] In this embodiment, through the application of quantum coding optimization and temperature state transfer matrix, the temperature field prediction error of the model is less than 3%, which ensures the accuracy of the prediction results; by extracting key parameters such as heat spatial diffusion rate, time cumulative change rate and material heat transfer coefficient, the model can effectively characterize the spatiotemporal transfer law of heat in the mold; quantum rotating gate and quantum NOT gate operations are used to optimize the particle swarm, improve the solution efficiency, and enable the optimization to converge within 200-300 generations; the temperature state transfer matrix and heat transfer response function reflect the stability of the system, provide a smooth response for heat transfer between regions, and help maintain uniform temperature distribution during the die-casting process; the model can predict the dynamic evolution process of the temperature field, provide a reliable basis for the precise control of the temperature field of the die-casting mold, and effectively improve the processing quality and life of the mold.

[0040] S102. Input the dynamic evolution model of heat distribution into the deep belief network based on knowledge distillation to construct a temperature field predictor for the die-casting mold; use a multi-agent collaborative evolution algorithm to train the temperature field predictor to determine the hierarchical optimization function of the temperature field of the die-casting mold; construct a temperature field optimization strategy network for the die-casting mold based on the hierarchical optimization function, and use an adaptive ant colony optimization algorithm based on taboo search to generate an optimal temperature field control sequence for the die-casting mold; establish a sub-region adaptive control strategy for the die-casting mold according to the optimal temperature field control sequence; In this embodiment, a temperature field predictor is constructed by a deep belief network based on knowledge distillation, which can more accurately predict the dynamic changes of the mold temperature field; the temperature field optimization strategy network is trained through a multi-agent collaborative evolution algorithm and a hierarchical optimization function to improve the accuracy and adaptability of temperature control; an adaptive ant colony taboo search algorithm is used to generate the optimal temperature control sequence, and a regional adaptive control strategy is formulated based on this to achieve precise thermal management of the mold and further improve the die-casting quality; regional control reduces the temperature adjustment response time, helps to quickly adapt to temperature changes during the die-casting process, and improves production efficiency and mold life.

[0041] In an optional embodiment, the heat distribution dynamic evolution model is input into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; a multi-agent collaborative evolution algorithm is used to train the temperature field predictor, and determining a hierarchical optimization function of the die-casting mold temperature field includes: Converting the temperature state transfer matrix in the thermal distribution dynamic evolution model into a characteristic graph, converting the heat transfer response function in the thermal distribution dynamic evolution model into a time series vector, normalizing the characteristic graph and the time series vector, and obtaining input data of a die-casting mold temperature field predictor; Construct a knowledge distillation framework consisting of a teacher network and a student network, wherein the teacher network adopts a deep residual network structure, and the student network adopts a three-layer deep belief network structure. The input data is input into the teacher network to obtain a soft label of the temperature field distribution. The soft label of the temperature field distribution is used as a training target of the student network. Each layer of the student network is pre-trained, and the restricted Boltzmann machine of each layer is trained using a contrastive divergence algorithm. The trained student network is used as the basic network structure of the die-casting mold temperature field predictor to obtain the die-casting mold temperature field predictor. The die-casting mold temperature field predictor is divided into multiple sub-predictors, each of which is responsible for the temperature prediction of its own corresponding area, and an information transmission channel is established between the sub-predictors through an attention mechanism; a multi-agent collaborative evolution algorithm is used to train the sub-predictors, and a fitness function including local prediction error and prediction consistency is constructed. The prediction performance of the sub-predictors is evaluated based on the fitness function, and the prediction performance is used as an evolutionary indicator. Based on the evolutionary indicator, the network parameters of each sub-predictor are optimized to obtain the optimized temperature field prediction performance indicator; a hierarchical optimization function of the die-casting mold temperature field is constructed according to the temperature field prediction performance indicator.

[0042] Specifically, the 100x100 temperature state transfer matrix in the thermal distribution dynamic evolution model is processed and converted into a 64-channel feature map through a convolution operation, with each channel size of 32x32. The heat transfer response function is sampled into 1000 points in the time dimension to form a time series vector. The feature map and time series vector are normalized to the maximum and minimum values, and the numerical range is uniformly scaled to between 0 and 1 as the standardized input data of the temperature field predictor.

[0043] A temperature field prediction network based on the knowledge distillation framework is constructed, in which the teacher network adopts a 50-layer deep residual network structure, including 5 residual blocks, each residual block contains 10 convolutional layers, the convolution kernel size is 3x3, and the number of channels increases from 64 to 512. The student network adopts a three-layer deep belief network structure, the first layer contains 1024 neurons, the second layer contains 512 neurons, and the third layer contains 256 neurons. The normalized input data is sent to the teacher network to obtain the soft label prediction results of the temperature field distribution.

[0044] A layer-by-layer pre-training strategy is adopted for the student network, and a restricted Boltzmann machine structure is constructed for each layer. The number of visible layer units of the first layer restricted Boltzmann machine is the same as the input data dimension, and the number of hidden layer units is 1024; the number of visible layer units of the second layer is 1024, and the number of hidden layer units is 512; the number of visible layer units of the third layer is 512, and the number of hidden layer units is 256. The contrastive divergence algorithm is used to train each layer of the restricted Boltzmann machine, with the learning rate set to 0.01, the momentum factor to 0.9, and the number of training rounds to 100. The three-layer network that has been trained is superimposed and combined to form the basic network structure of the die-casting mold temperature field predictor.

[0045] The die-casting mold surface is divided into 16 prediction regions, each of which corresponds to a sub-predictor. The sub-predictor inherits the basic network structure and takes the local features of the corresponding region as input. An information transmission channel of a multi-head attention mechanism is constructed between adjacent sub-predictors, with 8 attention heads and a dimension of 32 for each attention head to capture the correlation of temperature fields between regions.

[0046] The multi-agent co-evolution algorithm was used to train 16 sub-predictors, with the population size set to 50 and the number of evolution generations set to 200. A fitness function was constructed to evaluate the performance of the sub-predictors. The fitness function consists of two components: the local prediction error accounts for 60% and the prediction consistency accounts for 40%. The local prediction error is calculated using the mean absolute error, and the prediction consistency is calculated using the correlation coefficient of the prediction values ​​of adjacent regions.

[0047] During the evolution process, the network parameters of the sub-predictor are crossover and mutation operations are performed once every 10 generations. The crossover operation uses a uniform crossover method with a crossover probability of 0.8; the mutation operation uses a Gaussian mutation method with a mutation probability of 0.1 and a mutation intensity of 0.05. The prediction performance after crossover and mutation is evaluated based on the fitness function, and individuals with excellent performance are retained to enter the next generation of evolution.

[0048] After 200 generations of evolutionary optimization, the average prediction error of the sub-predictor was reduced to less than 1.5%, and the correlation coefficient of the prediction values ​​of adjacent regions was increased to above 0.95. The optimized prediction performance indicators were integrated into a hierarchical optimization function to guide the precise control of the temperature field of the die-casting mold. The hierarchical optimization function reflects the overall temperature field distribution characteristics at the macro level and reflects the local temperature gradient change law at the micro level.

[0049] In this embodiment, through the combination of the knowledge distillation framework, the regional sub-predictor and the multi-agent collaborative evolution algorithm, the prediction error is reduced to below 1.5%, which significantly improves the accuracy of temperature field prediction; the multi-head attention mechanism is used to capture the temperature correlation of adjacent prediction areas, so that the prediction consistency correlation coefficient of adjacent areas reaches above 0.95, which helps to achieve coordinated control of the overall temperature field; evolutionary algebraic optimization and crossover mutation operations make the training process more efficient, ensure rapid convergence to the optimal solution, and improve the adaptability and stability of the sub-predictor in a dynamic environment; the hierarchical optimization function integrates the macroscopic temperature field distribution and the local temperature gradient details, provides multi-level guidance for the precise temperature field control of the die-casting mold, and improves the quality and stability of the die-casting process.

[0050] In an optional embodiment, constructing a temperature field optimization strategy network for a die-casting mold based on the hierarchical optimization function, and using an adaptive ant colony optimization algorithm based on taboo search to generate an optimal temperature field control sequence for the die-casting mold includes: Based on the hierarchical optimization function, a temperature field optimization strategy network for the casting mold is constructed, the temperature field optimization strategy network is divided into a temperature field uniformity layer, a cooling efficiency layer and an energy consumption control layer, and an evaluation index function is determined; Initialize the position of the ant colony, encode the temperature field control parameters of the die-casting mold into ant position information, determine the temperature field control sequence, input the temperature field control sequence corresponding to each ant position into the temperature field optimization strategy network, calculate the path evaluation value through the evaluation index function, and update the pheromone concentration on the path according to the path evaluation value, wherein the pheromone increment is positively correlated with the path evaluation value; Constructing a dynamic taboo table, recording the searched temperature field control sequences, calculating the similarity between the newly generated temperature field control sequence and the temperature field control sequence in the taboo table, and adding the temperature field control sequence with a similarity higher than a preset threshold to the dynamic taboo table; Based on the pheromone concentration on the path and the dynamic taboo table, the state transition probability is calculated to guide the search direction of the ant colony; A variable neighborhood search structure is constructed near the current position of each ant, wherein the variable neighborhood search structure includes a single parameter adjustment space and a multi-parameter linkage adjustment space, and the parameters in the temperature field control sequence are adjusted independently and in combination respectively; the candidate solutions corresponding to the temperature field control sequence generated in the variable neighborhood search structure are input into the temperature field optimization strategy network, the path evaluation value of each candidate solution is calculated, the temperature field control sequence with the highest path evaluation value is selected, and the ant position is updated; Repeat the iteration until the preset maximum number of iterations is reached to obtain the optimal temperature field control sequence.

[0051] Specifically, a three-layer temperature field optimization strategy network is constructed. The temperature field uniformity layer focuses on the consistency of the mold surface temperature distribution. The temperature standard deviation of 16 temperature measurement points is used as the evaluation index, and the standard deviation threshold is set to 3 degrees Celsius. The cooling efficiency layer focuses on the cooling rate and cooling time, requiring the cooling rate not to exceed 50 degrees Celsius / second and the cooling time to be controlled within the range of 15-30 seconds. The energy consumption control layer focuses on the heating power and cooling water flow, with the heating power limited to 40 kilowatts and the cooling water flow limited to 25 liters / minute. The three layers of evaluation indicators are combined in a weight ratio of 4:3:3 to construct a comprehensive evaluation index function.

[0052] Create an ant colony containing 100 ants. The position information of each ant contains 10 temperature field control parameters: heater power, heating time, heating area distribution coefficient, cooling water flow, cooling channel opening, cooling time, temperature setting value, control cycle, proportional coefficient, and integration time. Encode these parameters into a 200-bit binary string to form a temperature field control sequence. During initialization, the ant position is randomly generated within the parameter feasible domain, and the feasible domain boundary is determined according to the process requirements.

[0053] The temperature field control sequence corresponding to each ant is input into the optimization strategy network to calculate the path evaluation value. The higher the path evaluation value, the better the control effect. The pheromone concentration on the path is updated according to the path evaluation value. The pheromone volatility coefficient is set to 0.1, the pheromone increment is proportional to the path evaluation value, and the proportional coefficient is 0.2. The update cycle of the pheromone concentration is 10 iterations.

[0054] The dynamic taboo table adopts a circular queue structure with a maximum capacity of 1000 control sequences. For the newly generated control sequence, the Hamming distance with the existing sequence in the taboo table is calculated. If the minimum Hamming distance is less than 5% of the length of the control sequence, the sequence is added to the taboo table. When the taboo table reaches the maximum capacity, the earliest added sequence is deleted. The taboo table is updated every 50 iterations.

[0055] The state transition probability is calculated based on the path pheromone concentration and taboo table information. The influence weight of pheromone concentration on transition probability is 0.7, and the influence weight of taboo table information is 0.3. The direction with a larger transition probability is more likely to be selected by ants for search. In each iteration, 20% of the ants perform random search to increase the exploration ability of the algorithm.

[0056] A variable neighborhood search structure is constructed around the current position of the ant. The single parameter adjustment space adjusts each control parameter up and down with a step size of 5% near the current value; the multi-parameter linkage adjustment space selects a combination of parameters with strong correlation, such as heating power and heating time, cooling water flow and cooling time, and adjusts these parameters at the same time. 50 candidate solutions are randomly generated from the variable neighborhood structure, their path evaluation values ​​are calculated, and the candidate solution with the highest evaluation value is selected to update the ant position.

[0057] The algorithm terminates after 300 iterations, and the optimal temperature field control sequence is obtained.

[0058] In this embodiment, by setting the temperature standard deviation threshold and temperature control parameters, the consistency of the temperature distribution on the mold surface is guaranteed and the quality problems caused by temperature difference are reduced; the rate and time limit of the cooling efficiency layer ensure that the cooling process is smooth and fast, meet the process requirements and avoid material stress problems caused by too fast cooling; the energy consumption control layer effectively limits the heating and cooling power to achieve energy-saving operation, while ensuring the temperature control accuracy and optimizing resource utilization; the ant colony algorithm combines taboo search and dynamic pheromone concentration update to prevent local optimality, ensure that the temperature control parameters gradually approach the optimal solution after multiple iterations, and improve the stability of the optimization results; through random search and variable neighborhood search strategies, the global exploration ability of the algorithm is increased, and the local adjustment is refined, which helps to obtain a more comprehensive and optimized temperature field control sequence.

[0059] S103. Input the sub-region adaptive control strategy into the dynamic evolution model of heat distribution to obtain the temperature field response characteristics of the die-casting mold; construct a temperature field controller based on deep reinforcement learning according to the temperature field response characteristics, and obtain the temperature field change parameters of the die-casting mold in real time; input the temperature field change parameters into the temperature field predictor for online correction, and use the dynamic fuzzy immune optimization algorithm to adjust the cooling control parameters to achieve intelligent optimization control of the heat distribution of the die-casting mold.

[0060] In this embodiment, real-time monitoring and feedback adjustment significantly improve the control accuracy of the mold temperature field, ensure the uniform distribution of temperature in each area, and effectively avoid mold loss and product defects caused by thermal unevenness; the correction function of the temperature field response characteristics and the real-time data input predictor enables the control system to adapt to temperature changes in the die-casting process, and improves the system's response speed and robustness to fluctuations; the dynamic fuzzy immune optimization algorithm automatically adjusts the cooling control parameters according to real-time data, realizes the intelligence and efficiency of the cooling process, and reduces cooling time and energy consumption; the optimized temperature field distribution control reduces the thermal fatigue of the mold, reduces thermal stress, and improves the mold service life and production stability.

[0061] In an optional embodiment, the sub-region adaptive control strategy is input into the dynamic evolution model of the heat distribution to obtain the temperature field response characteristics of the die-casting mold; according to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are obtained in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the cooling control parameters are adjusted by a dynamic fuzzy immune optimization algorithm to realize the intelligent optimization control of the heat distribution of the die-casting mold, which includes: The regional adaptive control strategy is input into the dynamic evolution model of heat distribution, the temperature field state equation of the die-casting mold is established based on the law of conservation of energy, the Kalman filter algorithm is used to perform state estimation on the temperature field state equation, and the temperature field response characteristics of the die-casting mold are obtained by iteratively updating the state estimation error covariance matrix and the Kalman gain matrix; A temperature field controller based on deep reinforcement learning is constructed according to the temperature field response characteristics. The temperature field controller adopts a dual network architecture, including a policy network for generating control actions and a value network for evaluating state values. The state space of the policy network and the value network includes the current temperature distribution, temperature change rate and control input history, and the action space includes the control parameter adjustment amount of each area. The policy network and the value network are trained using the dominant actor-critic algorithm, and an immediate reward function is constructed based on the temperature uniformity value and the control stability value. The temperature field change parameters of the die-casting mold are obtained by optimizing the policy loss function and the value loss function. The temperature field change parameters are input into the temperature field predictor for online correction, a dynamic fuzzy rule base is constructed according to the corrected prediction results, and fuzzy control rules are designed based on temperature deviation and temperature change rate; a dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, the adjustment strategy set is constructed into an antibody group, and the antibody affinity is calculated based on the temperature control error. Through cloning mutation operations, the antibody group is optimized, and finally the optimal cooling control parameters are obtained, and the thermal distribution of the die-casting mold is intelligently adjusted.

[0062] The die-casting mold surface is divided into 16 control areas, each of which is equipped with independent heating and cooling channels to achieve regional temperature control. The temperature field state equation is established based on the law of conservation of energy. The state variables include the temperature value of each area, heat flux density and boundary heat exchange coefficient. The Kalman filter algorithm is used for state estimation. The observation data comes from the real-time acquisition value of the temperature sensor buried on the mold surface, and the sampling period is 100 milliseconds.

[0063] The temperature field response characteristics are obtained by iteratively calculating the state estimation error covariance matrix and the Kalman gain matrix. The initial value of the error covariance matrix is ​​set to the unit matrix, and the system noise covariance matrix and the measurement noise covariance matrix are set to 0.01 and 0.05 times the unit matrix, respectively. After 1000 iterations, the root mean square error of the state estimation is reduced to less than 2 degrees Celsius.

[0064] A deep reinforcement learning controller with a dual network architecture is constructed, including a policy network and a value network. The policy network adopts a four-layer fully connected neural network structure: the input layer has 1024 neurons, the two hidden layers have 512 and 256 neurons respectively, and the output layer has 48 neurons corresponding to the heating power, cooling water flow and channel opening adjustment of 16 regions. The value network adopts a three-layer structure: the input layer has 1024 neurons, the hidden layer has 512 neurons, and the output layer has 1 neuron for state value evaluation.

[0065] The state space dimension is 1024, including the current temperature value of 16 regions, the temperature change rate of the previous three moments, and the history of the last 10 control inputs. The action space dimension is 48, corresponding to the adjustment of the three control parameters of each region, the heating power adjustment range is plus or minus 5 kilowatts, the cooling water flow adjustment range is plus or minus 2 liters / minute, and the channel opening adjustment range is plus or minus 10%.

[0066] The dual network is trained using the dominant actor-critic algorithm, and the immediate reward function is constructed based on temperature uniformity and control stability. Temperature uniformity is represented by the standard deviation of the temperature values ​​of the 16 regions, and the smaller the standard deviation, the greater the reward; control stability is represented by the sum of squares of the control parameter adjustments, and the smaller the adjustment, the greater the reward. The reward weight ratio is 7:3.

[0067] The dominant actor-critic algorithm contains two core components during the training process: the policy network and the value network. The policy network is responsible for generating control actions, and the value network is responsible for evaluating state values. The two networks interact and cooperate to complete the training process.

[0068] The construction of the immediate reward function considers two key indicators: temperature uniformity and control stability. The temperature uniformity is characterized by calculating the temperature deviation of the 16 control areas. The specific method is to sum the absolute value of the difference between the actual temperature and the target temperature of each area. The smaller the value, the more uniform the temperature field. The control stability is characterized by calculating the change amplitude of three consecutive control actions. The specific method is to sum the absolute value of the difference between the two control parameter adjustments. The smaller the value, the more stable the control.

[0069] In actual applications, the target threshold of the temperature average value is set to 5 degrees Celsius, and the target threshold of the control stability value is set to 10% of the adjustment amount. When the temperature average value is less than 5 degrees Celsius and the control stability value is less than 10%, a positive reward value of 1.0 is given; when one of them exceeds the threshold, a negative reward value of -0.5 is given; when both exceed the threshold, a negative reward value of -1.0 is given.

[0070] The policy loss function is used to optimize the policy network. Its design idea is to make the control actions generated by the network maximize the future cumulative rewards. In the specific implementation, the advantage value obtained by executing an action in the current state is first calculated. The advantage value is equal to the actual cumulative reward minus the state value predicted by the value network. If the advantage value is positive, it means that the action is better than the average level, and the output probability of the action is increased during training; if the advantage value is negative, the output probability of the action is reduced.

[0071] The value loss function is used to optimize the value network. Its design idea is to make the network accurately predict the value of each state. In specific implementation, the deviation between the predicted value output by the value network and the actual cumulative reward is calculated, and the network parameters are optimized by reducing this deviation.

[0072] During the training process, an alternating optimization method is used: first, the value network is fixed and the policy network is optimized for several steps; then the policy network is fixed and the value network is optimized for several steps. The optimization step size is dynamically adjusted according to the training progress. A larger step size such as 0.001 is used in the early stage of training, and gradually reduced to 0.0001 in the later stage of training to ensure convergence stability.

[0073] In specific implementation, 1,000 state-action-reward samples are collected in each round of training, and the samples are randomly divided into 10 batches. For each batch, the state value is first calculated using the value network, then the advantage value is calculated based on the actual reward, and finally the parameters of the policy network and the value network are updated respectively. This process is repeated until convergence, which generally requires 5,000-10,000 rounds of training.

[0074] To avoid local optimality, an exploration mechanism is introduced during the training process: the control actions generated by the policy network will be superimposed with a random disturbance, and the disturbance intensity will gradually decrease as the training progresses, from the initial 20% to the final 5%. At the same time, the experience replay mechanism is used to store historical training samples, and historical samples are randomly selected and mixed with new samples for training each time to improve sample utilization efficiency and training stability.

[0075] After training and optimization, the policy network can output appropriate control actions according to the current temperature field state to achieve precise control of the temperature field. Under typical working conditions, the accuracy of the control action can reach more than 85%, and the temperature field adjustment time is shortened to less than 15 seconds, meeting actual production needs.

[0076] The training adopts the experience replay mechanism, the experience pool capacity is 10,000, and the sampling batch size is 128. The learning rates of the policy network and the value network are set to 0.0001 and 0.0002 respectively, and the discount factor is 0.99. The training lasts for 10,000 rounds and stops when the average reward value changes by less than 1% for 100 consecutive rounds.

[0077] The temperature field change parameters obtained through training are input into the temperature field predictor for online correction. A dynamic rule base containing 49 fuzzy rules is constructed, and control rules are designed based on temperature deviation and temperature change rate. The temperature deviation is divided into seven levels, the temperature change rate is divided into seven levels, and the control output is divided into seven levels.

[0078] The dynamic fuzzy immune algorithm is used to optimize the cooling control parameters. 100 antibodies are initialized, each containing 48 gene positions, corresponding to the cooling control parameters of 16 regions. The antibody affinity is calculated by the inverse of the temperature control error. The smaller the error, the higher the affinity. The 20 antibodies with the highest affinity are selected for cloning and amplification. The number of clones is proportional to the affinity.

[0079] The antibodies produced by the clones were subjected to Gaussian mutation operation, with a mutation probability of 0.1, and the mutation strength was inversely proportional to the affinity. The mutated antibody group was subjected to suppression calculation, and antibodies with high affinity and low similarity were retained. The cloning, mutation and suppression operations were repeated for 50 generations to obtain the optimal cooling control parameter combination.

[0080] In this embodiment, based on deep reinforcement learning and regional control, the temperature of each area can be automatically adjusted under a complex temperature field, and ultimately the temperature uniformity target is achieved, with the temperature difference controlled within 5 degrees Celsius; the Kalman filter estimates the temperature state in real time, improves the response speed of the temperature field control, and shortens the adjustment time to less than 15 seconds to meet the production cycle requirements; the dynamic fuzzy immune algorithm flexibly adjusts the control parameters in the cooling parameter optimization, ensures the stability of the cooling process, and reduces energy consumption, cooling time and energy consumption; the control strategy can adapt to temperature field fluctuations, and perform predictor corrections in an online state, maintaining high control accuracy and strong robustness under complex working conditions.

[0081] In an optional embodiment, a dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, the adjustment strategy set is constructed as an antibody group, and the antibody affinity is calculated based on the temperature control error. The antibody group is optimized through cloning mutation operations, and the optimal cooling control parameters are finally obtained. Intelligently adjusting the heat distribution of the die-casting mold includes: Acquiring cooling control parameters of the die-casting mold, encoding the cooling control parameters to generate an initial antibody group, each antibody individual comprising a parameter vector corresponding to the cooling control parameters; Collecting real-time temperature data of each measuring point of the die-casting mold, calculating the temperature deviation and the temperature deviation change rate according to the real-time temperature data, taking the temperature deviation and the temperature deviation change rate as input variables, establishing a dynamic fuzzy rule mapping relationship, and constructing a fuzzy rule library for temperature field control; Based on the square sum of the deviations between the real-time temperature data and the target temperature, a temperature control error index is constructed, the temperature control error index is converted into an antibody affinity value through an exponential function, the exponential function includes an adjustment coefficient for controlling the selection pressure, and the antibody individuals in the antibody group are sorted from high to low according to the antibody affinity value; according to the preset rule of decreasing clone number, the clone number is sequentially assigned to each antibody individual to generate a clone antibody group, and a Gaussian mutation operation is performed on the clone antibody group, the variation step length of the Gaussian mutation is inversely proportional to the antibody affinity value, and a variant antibody group is obtained; Calculating the affinity value of each antibody individual in the variant antibody group, determining candidate variant antibody individuals from the variant antibody group according to a preset affinity selection threshold, and determining candidate antibody individuals from the initial antibody group, merging the candidate variant antibody individuals and the candidate candidate antibody individuals to generate a new generation antibody group, and at the same time, introducing randomly generated new antibody individuals into the new generation antibody group according to a preset number; A new generation of antibody groups is generated by repeated iterations until a preset number of iterations is reached, the antibody individual corresponding to the maximum affinity value is selected, the optimal antibody individual is determined, the optimal antibody individual is decoded, the optimal cooling control parameters are obtained, and the heat distribution of the die-casting mold is adjusted.

[0082] First, the initial cooling control parameters of the die-casting mold are obtained, including three types of parameters: cooling water flow, cooling channel opening, and cooling time in 16 control areas. These parameters are encoded into a 48-dimensional vector, and each parameter is encoded in 8-bit binary. 100 antibody individuals are initialized to form an antibody group, and each antibody corresponds to a complete set of cooling control parameter configuration schemes.

[0083] 32 temperature sensors are pre-buried on the mold surface with a sampling frequency of 10Hz. The real-time collected temperature data is used to calculate the temperature deviation and the temperature deviation change rate. The temperature deviation is defined as the difference between the measured temperature and the target temperature, and the temperature deviation change rate is the ratio of the temperature deviation difference between two adjacent samples to the sampling time interval.

[0084] According to actual production experience, the temperature deviation is divided into seven fuzzy subsets: negative large, negative medium, negative small, zero, positive small, positive medium, positive large, and the corresponding specific intervals are: less than -10 degrees Celsius, -10 to -5 degrees Celsius, -5 to -2 degrees Celsius, -2 to 2 degrees Celsius, 2 to 5 degrees Celsius, 5 to 10 degrees Celsius, and greater than 10 degrees Celsius. The temperature deviation change rate is also divided into seven fuzzy subsets, with specific intervals: less than -2 degrees Celsius / second, -2 to -1 degrees Celsius / second, -1 to -0.3 degrees Celsius / second, -0.3 to 0.3 degrees Celsius / second, 0.3 to 1 degrees Celsius / second, 1 to 2 degrees Celsius / second, and greater than 2 degrees Celsius / second.

[0085] Based on the above fuzzy subsets, a rule base containing 49 fuzzy rules is constructed. For example, when the temperature deviation is positive and the temperature deviation change rate is positive, the cooling intensity is adjusted to the maximum enhancement; when the temperature deviation is negative and the temperature deviation change rate is negative, the cooling intensity is adjusted to the maximum reduction. In this way, dynamic fuzzy control of the temperature field is achieved.

[0086] An exponential function was used to construct the calculation criteria for antibody affinity. The smaller the temperature control error, the larger the corresponding affinity value. The adjustment coefficient for controlling the selection pressure was set to 0.5 to adjust the uniformity of the fitness distribution. The affinity values ​​of 100 antibody individuals were calculated and sorted. The antibody with the highest affinity obtained the most clones, specifically in a linear decreasing manner: 20 clones for the first place, 19 clones for the second place, and so on.

[0087] Gaussian mutation operation is performed on the antibody group generated by cloning, and the variable step length is inversely proportional to the antibody affinity value. For antibodies with affinity values ​​above 0.9, the variable step length is set to 0.01; for antibodies with affinity values ​​between 0.7-0.9, the variable step length is set to 0.05; for antibodies with affinity values ​​between 0.5-0.7, the variable step length is set to 0.1; for antibodies with affinity values ​​below 0.5, the variable step length is set to 0.2. This adaptive mutation mechanism can increase the diversity of the search space while maintaining high-quality solutions.

[0088] After mutation, the affinity values ​​of each antibody are recalculated, and the affinity selection threshold is set to 0.6. Antibodies with an affinity greater than 0.6 are selected from the mutant antibody group as candidate mutant antibodies, and antibodies with an affinity greater than 0.8 are selected from the initial antibody group as candidate antibodies. These two parts of antibodies are combined to form a new generation population, and 10 new antibodies are randomly generated and injected into the population, keeping the population size at 100.

[0089] The above evolution process is repeated, and the number of iterations is set to 50 generations. In practical applications, when the change in the optimal affinity value for 10 consecutive generations is less than 0.1%, the iteration can be terminated in advance. Finally, the antibody individual with the highest affinity value is selected and decoded to obtain the optimal cooling control parameters.

[0090] Typical working condition verification shows that the optimal control parameters obtained by this method are: cooling water flow 18-25 liters / minute, cooling channel opening 60%-85%, cooling time 8-12 seconds. The precise control of the mold temperature field is achieved, the temperature control accuracy is better than plus or minus 1.5 degrees Celsius, and the temperature field uniformity is controlled within 3 degrees Celsius, which significantly improves the quality stability of die castings.

[0091] In this embodiment, the optimization and adjustment of the cooling control parameters enables the control accuracy of the mold temperature field to reach ±1.5 degrees Celsius, meeting the needs of high-precision temperature control; through the dynamic adjustment of multi-zone cooling water flow, channel opening and cooling time, the temperature field uniformity is controlled within 3 degrees Celsius, which helps to improve the stability of the quality of die-castings; based on the fuzzy rule base, an intelligent response to real-time temperature deviations and change rates is achieved, and the cooling intensity can be adaptively adjusted to ensure stable control effects under different production conditions; the antibody affinity adaptive variation mechanism increases the diversity of the search space while maintaining a high affinity solution, accelerates the convergence of the optimal control parameters, and significantly improves the parameter search efficiency; through the continuous optimization and evolution of the cooling control parameters, the method has good adaptability and is suitable for a variety of die-casting conditions, ultimately improving the quality consistency and production efficiency of die-castings.

[0092] Figure 2 FIG. 1 is a schematic diagram of a heat balance distribution optimization system for a die casting mold based on an energy model according to an embodiment of the present invention. Figure 2As shown, the system comprises: The first unit is used to collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation characteristics of the die-casting mold; establish a nonlinear state equation of the heat distribution of the die-casting mold based on the spatiotemporal correlation characteristics, use a hybrid particle swarm quantum genetic algorithm to solve the nonlinear state equation, and construct a dynamic evolution model of the heat distribution of the die-casting mold; The second unit is used to input the dynamic evolution model of heat distribution into a deep belief network based on knowledge distillation to construct a temperature field predictor for a die-casting mold; adopt a multi-agent collaborative evolution algorithm to train the temperature field predictor to determine a hierarchical optimization function of the temperature field of the die-casting mold; construct a temperature field optimization strategy network for the die-casting mold based on the hierarchical optimization function, and adopt an adaptive ant colony optimization algorithm based on taboo search to generate an optimal temperature field control sequence for the die-casting mold; and establish a sub-region adaptive control strategy for the die-casting mold according to the optimal temperature field control sequence; The third unit is used to input the sub-region adaptive control strategy into the dynamic evolution model of the heat distribution to obtain the temperature field response characteristics of the die-casting mold; according to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed to obtain the temperature field change parameters of the die-casting mold in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the dynamic fuzzy immune optimization algorithm is used to adjust the cooling control parameters to achieve intelligent optimization control of the heat distribution of the die-casting mold. The third aspect of the embodiment of the present invention is An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0093] A fourth aspect of the embodiments of the present invention is: 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.

[0094] 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.

[0095] 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 method for optimizing heat balance distribution of die casting molds based on energy model, characterized in that: include: Collecting heat distribution data of the die-casting mold, establishing a heat energy transfer mapping relationship of the die-casting mold, converting the heat energy transfer mapping relationship into a feature vector matrix, inputting the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtaining spatiotemporal correlation features of the die-casting mold; A nonlinear state equation of the heat distribution of the die-casting mold is established based on the time-space correlation characteristics, and a hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation to construct a dynamic evolution model of the heat distribution of the die-casting mold; Inputting the thermal distribution dynamic evolution model into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; The temperature field predictor is trained by using a multi-agent collaborative evolution algorithm to determine a hierarchical optimization function of the temperature field of the die-casting mold; Based on the hierarchical optimization function, a temperature field optimization strategy network of the die-casting mold is constructed, and an optimal temperature field control sequence of the die-casting mold is generated by an adaptive ant colony optimization algorithm based on taboo search; and a sub-region adaptive control strategy of the die-casting mold is established according to the optimal temperature field control sequence; Inputting the sub-region adaptive control strategy into the thermal distribution dynamic evolution model to obtain the temperature field response characteristics of the die-casting mold; According to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are acquired in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the cooling control parameters are adjusted using a dynamic fuzzy immune optimization algorithm to achieve intelligent optimization control of the thermal distribution of the die-casting mold.

2. The method according to claim 1, characterized in that Collect the heat distribution data of the die-casting mold, establish the heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into the graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation features of the die-casting mold, including: Construct a temperature sensor grid array on the surface of the die-casting mold, wherein the temperature sensor grid array includes a preset number of temperature acquisition points, records temperature values, acquisition timestamps, and spatial coordinate information; continuously acquires temperature data at a preset sampling frequency based on the temperature sensor grid array, and organizes the temperature data into a time series data set; calculates the temperature gradient vector between adjacent temperature acquisition points in the time series data set, calculates the heat flux density, and establishes a heat energy transfer mapping relationship based on the temperature gradient vector and the heat flux density; The time series data set is subjected to Fourier transformation to obtain the frequency domain characteristics of the temperature field, and the frequency domain characteristics of the temperature field are combined with the temperature gradient vector to construct a first mapping feature vector; the thermal conductivity coefficient distribution is calculated based on the thermal energy transfer mapping relationship, and the thermal conductivity coefficient distribution is combined with the first mapping feature vector to construct a second mapping feature vector; the second mapping feature vectors are arranged according to the spatial distribution of the temperature collection points to construct a feature vector matrix; A graph structure is constructed based on the spatial distribution of the temperature sensor grid array, the temperature collection points are used as graph structure nodes, and the thermal conductivity coefficient matrix is ​​used as graph structure edge weights; an unsupervised adversarial learning network is constructed, wherein the generator of the unsupervised adversarial learning network generates simulated heat transfer data based on the feature vector matrix, and the corresponding discriminator compares the simulated heat transfer data with the time series data set for training; Based on the graph structure, the attention weight coefficients between nodes are calculated, and the attention weight coefficients are weightedly fused with the thermal energy transfer mapping relationship to obtain the heat conduction weights; the feature vector matrix is ​​reconstructed according to the heat conduction weights to obtain the spatiotemporal correlation features.

3. The method according to claim 1, characterized in that Based on the time-space correlation characteristics, a nonlinear state equation of the heat distribution of the die-casting mold is established, and a hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation. The construction of a dynamic evolution model of the heat distribution of the die-casting mold includes: Based on the time-space correlation characteristics, the heat spatial diffusion rate, the heat time cumulative change rate, and the mold material heat transfer coefficient are extracted and calculated, and the state variable vector is constructed in combination; based on the heating power parameters and cooling rate parameters of the die-casting mold, a control variable vector is generated; and the state variable vector and the control variable vector are used to establish a nonlinear state equation that describes the heat transfer law inside the die-casting mold; Converting the solution parameters of the nonlinear state equation into quantum state codes, constructing an initial particle swarm including quantum position genes and quantum speed genes, substituting the quantum position genes and quantum speed genes of the initial particle swarm into the nonlinear state equation, calculating the fitness value, and determining the individual optimal solution and the global optimal solution of the initial particle swarm based on the fitness value; Based on the individual optimal solution and the global optimal solution, a quantum rotation angle is calculated, based on the quantum rotation angle, a quantum rotation gate operation is performed on the quantum position gene to obtain a crossover quantum position gene, and a quantum NOT gate operation is performed on the quantum speed gene to obtain a mutated quantum speed gene; the crossover quantum position gene and the mutated quantum speed gene are combined to construct an optimized particle swarm; Substituting the quantum position gene and the quantum speed gene of the optimized particle swarm into the nonlinear state equation to calculate a new fitness value, updating the individual optimal solution and the global optimal solution according to the new fitness value, and repeating the iterative update until the change rate of the new fitness value is less than a preset convergence threshold, thereby obtaining the optimal solution; According to the optimal solution, the temperature state transfer matrix of the die-casting mold in a discrete time series is constructed, and the heat transfer response function between the various regions of the die-casting mold is established based on the temperature state transfer matrix. The temperature state transfer matrix is ​​combined with the heat transfer response function to construct a dynamic evolution model of the thermal distribution of the die-casting mold.

4. The method according to claim 1, characterized in that: Inputting the thermal distribution dynamic evolution model into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; The temperature field predictor is trained by using a multi-agent collaborative evolution algorithm to determine the hierarchical optimization function of the temperature field of the die casting mold, including: Converting the temperature state transfer matrix in the thermal distribution dynamic evolution model into a characteristic graph, converting the heat transfer response function in the thermal distribution dynamic evolution model into a time series vector, normalizing the characteristic graph and the time series vector, and obtaining input data of a die-casting mold temperature field predictor; Construct a knowledge distillation framework consisting of a teacher network and a student network, wherein the teacher network adopts a deep residual network structure, and the student network adopts a three-layer deep belief network structure. The input data is input into the teacher network to obtain a soft label of the temperature field distribution. The soft label of the temperature field distribution is used as a training target of the student network. Each layer of the student network is pre-trained, and the restricted Boltzmann machine of each layer is trained using a contrastive divergence algorithm. The trained student network is used as the basic network structure of the die-casting mold temperature field predictor to obtain the die-casting mold temperature field predictor. The die-casting mold temperature field predictor is divided into multiple sub-predictors, each of which is responsible for the temperature prediction of its own corresponding area, and an information transmission channel is established between the sub-predictors through an attention mechanism; a multi-agent collaborative evolution algorithm is used to train the sub-predictors, and a fitness function including local prediction error and prediction consistency is constructed. The prediction performance of the sub-predictors is evaluated based on the fitness function, and the prediction performance is used as an evolutionary indicator. Based on the evolutionary indicator, the network parameters of each sub-predictor are optimized to obtain the optimized temperature field prediction performance indicator; a hierarchical optimization function of the die-casting mold temperature field is constructed according to the temperature field prediction performance indicator.

5. The method according to claim 1, characterized in that Based on the hierarchical optimization function, a temperature field optimization strategy network for the die-casting mold is constructed, and an adaptive ant colony optimization algorithm based on taboo search is used to generate an optimal temperature field control sequence for the die-casting mold, including: Based on the hierarchical optimization function, a temperature field optimization strategy network for the casting mold is constructed, the temperature field optimization strategy network is divided into a temperature field uniformity layer, a cooling efficiency layer and an energy consumption control layer, and an evaluation index function is determined; Initialize the position of the ant colony, encode the temperature field control parameters of the die-casting mold into ant position information, determine the temperature field control sequence, input the temperature field control sequence corresponding to each ant position into the temperature field optimization strategy network, calculate the path evaluation value through the evaluation index function, and update the pheromone concentration on the path according to the path evaluation value, wherein the pheromone increment is positively correlated with the path evaluation value; Constructing a dynamic taboo table, recording the searched temperature field control sequences, calculating the similarity between the newly generated temperature field control sequence and the temperature field control sequence in the taboo table, and adding the temperature field control sequence with a similarity higher than a preset threshold to the dynamic taboo table; Based on the pheromone concentration on the path and the dynamic taboo table, the state transition probability is calculated to guide the search direction of the ant colony; A variable neighborhood search structure is constructed near the current position of each ant, wherein the variable neighborhood search structure includes a single parameter adjustment space and a multi-parameter linkage adjustment space, and the parameters in the temperature field control sequence are adjusted independently and in combination respectively; the candidate solutions corresponding to the temperature field control sequence generated in the variable neighborhood search structure are input into the temperature field optimization strategy network, the path evaluation value of each candidate solution is calculated, the temperature field control sequence with the highest path evaluation value is selected, and the ant position is updated; Repeat the iteration until the preset maximum number of iterations is reached to obtain the optimal temperature field control sequence.

6. The method according to claim 1, characterized in that Input the sub-region adaptive control strategy into the dynamic evolution model of heat distribution to obtain the temperature field response characteristics of the die-casting mold; construct a temperature field controller based on deep reinforcement learning according to the temperature field response characteristics, and obtain the temperature field change parameters of the die-casting mold in real time; Inputting the temperature field change parameters into the temperature field predictor for online correction, and using the dynamic fuzzy immune optimization algorithm to adjust the cooling control parameters to achieve intelligent optimization control of the heat distribution of the die casting mold includes: The regional adaptive control strategy is input into the dynamic evolution model of heat distribution, the temperature field state equation of the die-casting mold is established based on the law of conservation of energy, the Kalman filter algorithm is used to perform state estimation on the temperature field state equation, and the temperature field response characteristics of the die-casting mold are obtained by iteratively updating the state estimation error covariance matrix and the Kalman gain matrix; A temperature field controller based on deep reinforcement learning is constructed according to the temperature field response characteristics. The temperature field controller adopts a dual network architecture, including a policy network for generating control actions and a value network for evaluating state values. The state space of the policy network and the value network includes the current temperature distribution, temperature change rate and control input history, and the action space includes the control parameter adjustment amount of each area. The policy network and the value network are trained using the dominant actor-critic algorithm, and an immediate reward function is constructed based on the temperature uniformity value and the control stability value. The temperature field change parameters of the die-casting mold are obtained by optimizing the policy loss function and the value loss function. The temperature field change parameters are input into the temperature field predictor for online correction, a dynamic fuzzy rule base is constructed according to the corrected prediction results, and fuzzy control rules are designed based on temperature deviation and temperature change rate; a dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, the adjustment strategy set is constructed into an antibody group, and the antibody affinity is calculated based on the temperature control error. Through cloning mutation operations, the antibody group is optimized, and finally the optimal cooling control parameters are obtained, and the thermal distribution of the die-casting mold is intelligently adjusted.

7. The method according to claim 6, characterized in that A dynamic fuzzy immune optimization algorithm is used to generate a cooling parameter adjustment strategy set, which is constructed as an antibody group. The antibody affinity is calculated based on the temperature control error. The antibody group is optimized through cloning and mutation operations, and the optimal cooling control parameters are finally obtained. Intelligent adjustment of the heat distribution of the die-casting mold includes: Acquiring cooling control parameters of the die-casting mold, encoding the cooling control parameters to generate an initial antibody group, each antibody individual comprising a parameter vector corresponding to the cooling control parameters; Collecting real-time temperature data of each measuring point of the die-casting mold, calculating the temperature deviation and the temperature deviation change rate according to the real-time temperature data, taking the temperature deviation and the temperature deviation change rate as input variables, establishing a dynamic fuzzy rule mapping relationship, and constructing a fuzzy rule library for temperature field control; Based on the square sum of the deviations between the real-time temperature data and the target temperature, a temperature control error index is constructed, the temperature control error index is converted into an antibody affinity value through an exponential function, the exponential function includes an adjustment coefficient for controlling the selection pressure, and the antibody individuals in the antibody group are sorted from high to low according to the antibody affinity value; according to the preset rule of decreasing clone number, the clone number is sequentially assigned to each antibody individual to generate a clone antibody group, and a Gaussian mutation operation is performed on the clone antibody group, the variation step length of the Gaussian mutation is inversely proportional to the antibody affinity value, and a variant antibody group is obtained; Calculating the affinity value of each antibody individual in the variant antibody group, determining candidate variant antibody individuals from the variant antibody group according to a preset affinity selection threshold, and determining candidate antibody individuals from the initial antibody group, merging the candidate variant antibody individuals and the candidate candidate antibody individuals to generate a new generation antibody group, and at the same time, introducing randomly generated new antibody individuals into the new generation antibody group according to a preset number; A new generation of antibody groups is generated by repeated iterations until a preset number of iterations is reached, the antibody individual corresponding to the maximum affinity value is selected, the optimal antibody individual is determined, the optimal antibody individual is decoded, the optimal cooling control parameters are obtained, and the heat distribution of the die-casting mold is adjusted.

8. A die casting mold heat balance distribution optimization system based on an energy model, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect heat distribution data of the die-casting mold, establish a heat energy transfer mapping relationship of the die-casting mold, convert the heat energy transfer mapping relationship into a feature vector matrix, input the feature vector matrix into a graph attention network based on unsupervised adversarial learning for feature extraction, and obtain the spatiotemporal correlation features of the die-casting mold; A nonlinear state equation of the heat distribution of the die-casting mold is established based on the time-space correlation characteristics, and a hybrid particle swarm quantum genetic algorithm is used to solve the nonlinear state equation to construct a dynamic evolution model of the heat distribution of the die-casting mold; The second unit is used to input the thermal distribution dynamic evolution model into a deep belief network based on knowledge distillation to construct a die-casting mold temperature field predictor; The temperature field predictor is trained by using a multi-agent collaborative evolution algorithm to determine a hierarchical optimization function of the temperature field of the die-casting mold; Based on the hierarchical optimization function, a temperature field optimization strategy network of the die-casting mold is constructed, and an optimal temperature field control sequence of the die-casting mold is generated by an adaptive ant colony optimization algorithm based on taboo search; and a sub-region adaptive control strategy of the die-casting mold is established according to the optimal temperature field control sequence; The third unit is used to input the sub-region adaptive control strategy into the thermal distribution dynamic evolution model to obtain the temperature field response characteristics of the die-casting mold; According to the temperature field response characteristics, a temperature field controller based on deep reinforcement learning is constructed, and the temperature field change parameters of the die-casting mold are acquired in real time; the temperature field change parameters are input into the temperature field predictor for online correction, and the cooling control parameters are adjusted using a dynamic fuzzy immune optimization algorithm to achieve intelligent optimization control of the thermal distribution of the die-casting mold.

9. 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 7.

10. 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 7 is implemented.

Images