Mold temperature adaptive adjustment method and system based on machine learning

Abstract

The present invention provides a method and system for adaptively adjusting mold temperature based on machine learning, which relates to the field of temperature adjustment technology, including collecting real-time temperature data and temperature-related process parameters in a mold production process, performing feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and constructing a multi-objective optimization model of mold temperature uniformity and product quality according to the obtained target temperature adjustment key feature set to obtain the optimal solution of the multi-objective optimization model; constructing a temperature adjustment model, and obtaining an initial temperature adjustment strategy by solving the temperature adjustment model, and sending it to an edge computing node at a workshop site at the same time, and when an abnormality is detected, triggering a behavior cloning-based online strategy update mechanism, using the newly collected temperature data at the time of the abnormality to adaptively optimize the initial temperature adjustment strategy, and generating a target temperature adjustment strategy for guiding the real-time control of the mold temperature.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature regulation, and in particular to a method and system for adaptively regulating mold temperature based on machine learning. Background Art

[0002] In the die-casting and other mold processing processes, mold temperature is one of the key factors affecting product quality. Fluctuations in mold temperature can lead to problems such as product dimensional deviation and reduced surface quality, so mold temperature needs to be precisely controlled. Traditional mold temperature regulation mainly relies on manual experience, which has shortcomings such as low efficiency and difficulty in ensuring control accuracy. In order to improve the automation level of mold temperature control, it is urgent to develop intelligent mold temperature adaptive regulation technology.

[0003] At present, the automatic adjustment technology of mold temperature mainly includes two categories: PID control and intelligent optimization algorithm. The method based on PID control realizes closed-loop control through temperature feedback, but its adaptability is poor and it is difficult to cope with the dynamic changes of production conditions. The method based on intelligent optimization algorithm, such as mold flow analysis optimization and intelligent algorithm optimization, can establish a mathematical model of mold temperature and process parameters to realize intelligent control of mold temperature, but it is limited by the accuracy and scope of application of the model in practical applications.

[0004] Therefore, the purpose of the present invention is to provide a mold temperature adaptive adjustment method and system based on machine learning. By introducing a machine learning algorithm, the intelligence and adaptability of the mold temperature control process can be realized, the accuracy and efficiency of mold temperature control can be further improved, and product quality can be stabilized. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for adaptively adjusting mold temperature based on machine learning, which can solve the problems in the prior art.

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

[0007] A method for adaptively adjusting mold temperature based on machine learning is provided, comprising:

[0008] Collecting real-time temperature data and temperature-related process parameters during the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, performing feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and using a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features;

[0009] According to the obtained target temperature adjustment key feature set, a multi-objective optimization model of mold temperature uniformity and product quality is constructed, and the multi-objective optimization model is divided into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it, and based on the solutions to each sub-problem, the optimal solution of the multi-objective optimization model is obtained;

[0010] The optimal solution of the multi-objective optimization model is taken as the optimization target, and the temperature regulation key feature set is taken as the state space. A reward function for multi-objective optimization is designed, and a temperature regulation model is constructed. By solving the temperature regulation model, an initial temperature regulation strategy is obtained and simultaneously sent to the edge computing nodes at the workshop site. The edge computing nodes collect real-time temperature data during the production process, and detect temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, a strategy online update mechanism based on behavior cloning is triggered, and the initial temperature regulation strategy is adaptively optimized using the newly collected temperature data during the anomaly, so as to generate a target temperature regulation strategy for guiding the real-time control of the mold temperature.

[0011] In an optional embodiment,

[0012] Feature extraction is performed on real-time temperature data and process parameters related to temperature to obtain multiple candidate features. The key feature selection algorithm is used to select the target temperature regulation key feature set from multiple candidate features, including:

[0013] The real-time temperature data collected by each temperature sensor is organized into a two-dimensional time series matrix, where each row of the matrix represents the temperature time series of a sensor; an adjacency matrix is ​​constructed based on the geometric structure of the mold and the sensor layout, where each element of the adjacency matrix represents the spatial connection relationship between two sensors; the process parameters are converted into an attribute matrix, where each row of the attribute matrix represents the local process parameters corresponding to a sensor position; a unified attribute graph is constructed based on the two-dimensional time series matrix, the adjacency matrix and the attribute matrix, where the node attributes of the attribute graph are the temperature time series and the local process parameters, and the edge attributes are the spatial connection relationship between sensors;

[0014] The attribute graph is input into the graph attention network. The attention weight between each node and its neighbor nodes is calculated through the attention mechanism. The features of the neighbor nodes are aggregated using the attention weight to obtain the output features of the node. The receptive field of the node is gradually expanded by stacking multiple layers of graph attention networks to aggregate high-order neighborhood information. The output features of the last layer of graph attention network are globally averaged and pooled to obtain graph-level embedded features that integrate the dynamic association patterns between temperature sensors.

[0015] The obtained graph-level embedded features are input into the pre-trained graph convolutional network, and local smoothing and transformation are performed through multi-layer graph convolution operations to extract multi-scale high-order features. The output features of the last layer of graph convolutional network are subjected to global maximum pooling operations to obtain multiple candidate features reflecting different high-order distribution patterns of the temperature field.

[0016] The obtained multiple candidate features are input into the gated self-attention network, and the attention score matrix is ​​obtained by calculating the attention scores between the candidate features. The forget gate and the update gate are introduced to generate the gating signal matrix. The attention score matrix is ​​weighted by the gating signal matrix to obtain the feature importance weight matrix. Based on the feature importance weight matrix, multiple candidate features are weighted to obtain weighted feature representations. The weighted feature representations are average pooled to obtain the final target temperature regulation key feature set.

[0017] In an optional embodiment,

[0018] The steps before inputting the obtained graph-level embedded features into the pre-trained graph convolutional network also include:

[0019] Discretize the heat conduction equation into the graph domain. The heat conduction equation describes the physical relationship between material density, specific heat capacity, thermal conductivity, temperature field and internal heat source, and obtain its finite difference form. Based on the finite difference form, construct a physical regularization term. Combine the physical regularization term with the task loss function of the graph convolutional network to obtain a loss function that integrates physical constraints.

[0020] Training the graph convolutional network by minimizing the joint loss function of the fused physical constraints;

[0021] Among them, the calculation formula of the joint loss function is as follows:

[0022] ;

[0023] Among them, L represents the joint loss function, N represents the number of nodes, and T represents the time step. represents the temperature of node m at time step t in the real temperature field, represents the temperature of node m at time step t in the predicted temperature field, λ represents the equilibrium coefficient, η represents the material density, c p represents the specific heat capacity, represents the temperature of node m at time step t+1, △t represents the time interval between two consecutive time steps, μ represents thermal conductivity, N m represents the set of neighboring nodes of node m, represents the temperature of node n at time step t, d mn represents the distance between nodes m and n, represents the internal heat source of node m at time step t.

[0024] In an optional embodiment,

[0025] The multi-objective optimization model is divided into a plurality of loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it. Based on the solutions of each sub-problem, the optimal solution of the multi-objective optimization model is obtained, including:

[0026] According to the temperature control area of ​​the mold, the multi-objective optimization model of mold temperature uniformity and product quality is divided into multiple loosely coupled sub-problems. Each sub-problem corresponds to a temperature control area, and coordination variables are introduced to realize information interaction between sub-problems.

[0027] For each of the subproblems, a proximal term is added to its objective function to construct a proximal subgradient subproblem, wherein the proximal term includes the coordination variable of the previous iteration of the current subproblem, the deviation between the optimization variable of the subproblem and the coordination variable, and the subgradient direction of the objective function of the subproblem;

[0028] Use the proximal subgradient method to solve each subproblem. By calculating the subgradient of the subproblem objective function at the current solution, use the subgradient descent method to update the optimization variable of the subproblem, and repeat the iteration until the preset subproblem convergence condition is met to obtain the approximate optimal solution of each subproblem.

[0029] Using the obtained approximate optimal solution of each subproblem, the coordination variables are updated through the linear combination of the optimization variables of the subproblem and the subgradient direction, and the Lagrange multipliers are updated based on the optimization variables of the subproblem and the updated coordination variables;

[0030] Repeat the steps of proximal subgradient subproblem construction, subproblem solution and coordination variable update until the difference between the optimization variable and the coordination variable of the subproblem between two consecutive iterations is less than the preset convergence threshold, and output the local optimal solution of each subproblem;

[0031] The local optimal solutions of each of the sub-problems are integrated to obtain the optimal solution of the multi-objective optimization model.

[0032] In an optional embodiment,

[0033] The calculation formula for updating the optimization variables of the subproblem using the subgradient descent method is as follows:

[0034] ;

[0035] in, represents the optimization variable of subproblem i at the k+1th iteration, Indicates that in Pi The induced norm downward feasible domain X i The projection operation, P i represents the projection matrix of subproblem i, X i represents the feasible domain of subproblem i, represents the optimization variable of subproblem i at the kth iteration, α k represents the step size at the kth iteration, Denotes the objective function f of subproblem i i The subgradient at the current optimization variable, f i (·) represents the objective function, ρ represents the penalty term weight, represents the coordination variable of subproblem i, represents the Lagrange multiplier in subproblem i, ▽ represents the gradient symbol, h i Represents the constraint function of subproblem i.

[0036] In an optional embodiment,

[0037] The edge computing node collects real-time temperature data during the production process, detects temperature anomalies through a pre-built temperature anomaly detection model, and triggers an online strategy update mechanism based on behavior cloning when an anomaly is detected. The initial temperature adjustment strategy is adaptively optimized using the newly collected temperature data during the anomaly, and a target temperature adjustment strategy is generated to guide the real-time control of the mold temperature, including:

[0038] Build a temperature anomaly detection model based on LSTM to predict the temperature trend in the future. The deviation between the actual temperature value and the predicted temperature value is used as the anomaly score. When the anomaly score exceeds the set threshold, it is determined as a temperature anomaly, triggering an early warning and strategy update.

[0039] Collect historical temperature data under normal working conditions, use a sliding time window to intercept temperature time series fragments to construct a training set, and use the training set to pre-train the temperature anomaly detection model by minimizing the mean square error between the predicted temperature series and the actual temperature series;

[0040] Based on the pre-trained temperature anomaly detection model, the temperature anomaly detection model is fine-tuned using the pre-labeled temperature anomaly data to obtain the final trained temperature anomaly detection model;

[0041] Deploy the trained temperature anomaly detection model to the edge computing node, perform anomaly detection on the sliding time window of the real-time collected temperature data, identify temperature anomalies in real time through prediction errors, and trigger an early warning.

[0042] A feedforward neural network architecture is used to build a behavioral cloning model of the temperature control strategy. The input of the behavioral cloning model is the operating condition characteristics when the temperature is abnormal, and the output is the control strategy parameters. It is used to learn the expert's strategy decision mapping relationship by imitating the expert's control log in different abnormal scenarios.

[0043] Collect the action data of experts in different temperature abnormality scenarios, build a training set for the behavior cloning model, and minimize the deviation between the output of the behavior cloning model and the expert action data to enable the model to learn the expert's control strategy under abnormal conditions;

[0044] After the temperature anomaly detection model finds the temperature anomaly, the abnormal working condition characteristics are input into the behavior cloning model to generate the corresponding control strategy parameters, the initial temperature adjustment strategy is adaptively optimized, and the target temperature adjustment strategy is generated to guide the real-time control of the mold temperature.

[0045] In an optional embodiment,

[0046] The method further comprises:

[0047] The temperature data of different workshops, production lines and equipment are collected on multiple edge nodes respectively. Based on the collected temperature data, a local temperature adjustment strategy model is constructed on each edge node using a feedforward neural network structure, and local model parameters are obtained by minimizing the strategy loss function and training the local temperature adjustment strategy model.

[0048] The local model parameters obtained by training each edge node are uploaded to the central server, and the central server determines the weight of the corresponding local model parameters according to the data volume of each edge node, and generates the global temperature adjustment strategy model parameters by aggregating the local model parameters;

[0049] The central server sends the generated global temperature regulation strategy model parameters to each edge node. Each edge node initializes the local temperature regulation strategy model using the received global temperature regulation strategy model parameters, and retrains the local temperature regulation strategy model based on the temperature data to update the local model parameters.

[0050] Repeat the iteration until the parameters of the global temperature control strategy model converge, and finally obtain the target temperature control strategy.

[0051] According to a second aspect of the embodiments of the present invention,

[0052] A mold temperature adaptive adjustment system based on machine learning is provided, comprising:

[0053] The first unit is used to collect real-time temperature data and temperature-related process parameters in the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, perform feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and use a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features;

[0054] The second unit is used to adjust the key feature set of the target temperature, build a multi-objective optimization model of mold temperature uniformity and product quality according to the obtained target temperature, divide the multi-objective optimization model into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables, each sub-problem corresponds to a temperature control area of ​​the mold, and solve each sub-problem by using a distributed optimization algorithm based on the alternating direction multiplier method, and obtain the optimal solution of the multi-objective optimization model based on the solutions of each sub-problem;

[0055] The third unit is used to take the optimal solution of the multi-objective optimization model as the optimization target, the temperature regulation key feature set as the state space, design a reward function for multi-objective optimization, and construct a temperature regulation model. By solving the temperature regulation model, an initial temperature regulation strategy is obtained, and the initial temperature regulation strategy is sent to the edge computing node at the workshop site. The edge computing node collects real-time temperature data in the production process and detects temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, the behavior cloning-based strategy online update mechanism is triggered, and the initial temperature regulation strategy is adaptively optimized using the newly collected temperature data at the time of the anomaly, and a target temperature regulation strategy is generated to guide the real-time control of the mold temperature.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] A fourth aspect of the embodiments of the present invention is:

[0062] 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.

[0063] In this embodiment, the key factors affecting the temperature distribution are captured by feature extraction and key feature selection of real-time temperature data and process parameters. By using a multi-objective optimization model to simultaneously consider temperature uniformity and product quality, the mold temperature is precisely controlled, the temperature distribution uniformity is improved, and the temperature fluctuation is reduced, thereby improving the consistency of product quality. By using a temperature regulation model to generate an initial regulation strategy, and by using a behavior cloning mechanism to implement online updates of the strategy, it can automatically adapt to changes and abnormal conditions in the production process, reduce manual intervention, and improve production efficiency and stability. When an abnormality is detected, the strategy online update mechanism is triggered to quickly identify and respond to temperature anomalies, reduce the generation of defective products, and improve the reliability of the production process. Through domain decomposition and distributed optimization algorithms, complex problems are decomposed into sub-problems that can be solved in parallel, which improves the efficiency of solving large-scale mold temperature control problems, is applicable to complex mold systems, and shortens the optimization time. It can not only optimize the production process, but also reduce energy consumption and production costs, improve product quality, and enhance the competitiveness of enterprises. At the same time, this intelligent and adaptive method also helps manufacturing companies achieve digital transformation and sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic flow chart of a method for adaptively adjusting mold temperature based on machine learning according to an embodiment of the present invention;

[0065] Figure 2 It is a structural schematic diagram of a mold temperature adaptive adjustment system based on machine learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] 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.

[0067] 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.

[0068] Figure 1 FIG. 1 is a flow chart of a method for adaptively adjusting mold temperature based on machine learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] S101. Collect real-time temperature data and temperature-related process parameters during the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, perform feature extraction on the real-time temperature data and temperature-related process parameters to obtain multiple candidate features, and use a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features.

[0070] Among them, high-precision temperature sensors are used, and measuring points are arranged in different areas of the mold to collect temperature data in real time during the mold production process. At the same time, through the communication interface with the die-casting machine control system, process parameters closely related to temperature are collected, including die-casting pressure, die-casting speed, material temperature and ambient temperature. Among them, die-casting pressure reflects the pressure change of the melt during the die-casting process and has a direct relationship with the temperature distribution inside the mold.

[0071] Feature engineering methods, such as statistical features, frequency domain features, and time domain features, are used to extract features from the preprocessed process parameter data to obtain multiple candidate features. Key feature selection algorithms, such as Relief and mRMR, are used to screen out a preliminary feature subset from the candidate features. The Sequential Forward Selection (SFS) algorithm is further used to optimize the preliminary feature subset. By iteratively adding features forward, the performance of the new feature subset is evaluated, and finally an optimal set of key features for temperature regulation is obtained.

[0072] In an optional embodiment,

[0073] Feature extraction is performed on real-time temperature data and process parameters related to temperature to obtain multiple candidate features. The key feature selection algorithm is used to select the target temperature regulation key feature set from multiple candidate features, including:

[0074] The real-time temperature data collected by each temperature sensor is organized into a two-dimensional time series matrix, where each row of the matrix represents the temperature time series of a sensor; an adjacency matrix is ​​constructed based on the geometric structure of the mold and the sensor layout, where each element of the adjacency matrix represents the spatial connection relationship between two sensors; the process parameters are converted into an attribute matrix, where each row of the attribute matrix represents the local process parameters corresponding to a sensor position; a unified attribute graph is constructed based on the two-dimensional time series matrix, the adjacency matrix and the attribute matrix, where the node attributes of the attribute graph are the temperature time series and the local process parameters, and the edge attributes are the spatial connection relationship between sensors;

[0075] The attribute graph is input into the graph attention network. The attention weight between each node and its neighbor nodes is calculated through the attention mechanism. The features of the neighbor nodes are aggregated using the attention weight to obtain the output features of the node. The receptive field of the node is gradually expanded by stacking multiple layers of graph attention networks to aggregate high-order neighborhood information. The output features of the last layer of graph attention network are globally averaged and pooled to obtain graph-level embedded features that integrate the dynamic association patterns between temperature sensors.

[0076] The obtained graph-level embedded features are input into the pre-trained graph convolutional network, and local smoothing and transformation are performed through multi-layer graph convolution operations to extract multi-scale high-order features. The output features of the last layer of graph convolutional network are subjected to global maximum pooling operations to obtain multiple candidate features reflecting different high-order distribution patterns of the temperature field.

[0077] The obtained multiple candidate features are input into the gated self-attention network, and the attention score matrix is ​​obtained by calculating the attention scores between the candidate features. The forget gate and the update gate are introduced to generate the gating signal matrix. The attention score matrix is ​​weighted by the gating signal matrix to obtain the feature importance weight matrix. Based on the feature importance weight matrix, multiple candidate features are weighted to obtain weighted feature representations. The weighted feature representations are average pooled to obtain the final target temperature regulation key feature set.

[0078] For example, the collected raw temperature data first needs to be organized and represented. Specifically, the real-time temperature data collected by each temperature sensor is arranged in chronological order to form a temperature time series. Then, the temperature time series of all sensors are combined into a two-dimensional matrix, where each row of the matrix corresponds to the temperature time series of a sensor, and different rows represent data from different sensors. In this way, a two-dimensional time series matrix representing the dynamic changes of the entire mold temperature field is obtained.

[0079] Next, we need to consider the spatial correlation between temperature sensors. Based on the geometry of the mold and the layout information of the sensors, an adjacency matrix is ​​constructed to represent the connection relationship between the sensors. The adjacency matrix is ​​a square matrix, and each element of the matrix indicates whether there is a connection between two sensors. If there is a direct physical connection between the two sensors or they are close in space, the corresponding adjacency matrix element is set to 1, otherwise it is set to 0. The adjacency matrix describes the topological structure of the temperature sensors in space and reflects the correlation of temperature changes at different locations.

[0080] In addition to temperature data and spatial connection relationships, process parameters related to the temperature field also need to be included in the feature representation. Various process parameters in the die-casting process, such as injection temperature, injection pressure, injection speed, holding time, etc., are divided according to the position of the sensor to obtain a set of local process parameters corresponding to each sensor position. Then, these local process parameters are organized into an attribute matrix, where each row of the matrix corresponds to a sensor position and different columns represent different types of process parameters. The attribute matrix reflects the process status at different spatial positions and plays an important role in explaining local temperature changes.

[0081] After obtaining the temperature time series matrix, adjacency matrix and attribute matrix, a unified attribute graph model can be constructed. The attribute graph is a graph structure that integrates node attributes and edge attributes, which can simultaneously represent the intrinsic characteristics and topological relationships of the data. Each sensor position is regarded as a node in the graph, and the attributes of the node are the temperature time series and local process parameters corresponding to the position. The edge connection relationship between nodes is determined by the adjacency matrix, which represents the spatial correlation between sensors. In this way, an attribute graph representation containing rich temperature field information is obtained.

[0082] In order to extract key features from the attribute graph, the graph attention network is used to process the graph structure data. The graph attention network is a graph neural network model based on the attention mechanism, which aggregates neighborhood information by learning the attention weights between nodes. Specifically, for each node in the attribute graph, the attention mechanism is used to calculate the attention weights between the node and all its neighboring nodes. The attention weights measure the importance of different neighboring nodes to the current node. The larger the weight, the greater the influence of the neighboring nodes. Then, these attention weights are used to weighted sum the features of the neighboring nodes to obtain the output features of the current node. This process can be regarded as a process in which a node updates its own features by selectively aggregating neighbor information.

[0083] An important feature of the graph attention network is that multiple layers of the network can be stacked to gradually expand the receptive field of the node. Each layer of the graph attention network aggregates the first-order neighborhood information of the node. By stacking multiple layers, higher-order neighborhood information can be gradually integrated into the node features. In this way, the representation of the node can contain contextual information in a wider range and capture the global distribution pattern of the temperature field. In the last layer of the graph attention network, the output features of all nodes are globally averaged and pooled to aggregate them into a graph-level embedded feature vector. This graph-level feature integrates the temperature information of all sensor locations and the dynamic association patterns between them, representing a high-level semantic representation of the entire temperature field.

[0084] After obtaining the graph-level features output by the graph attention network, the graph convolution network is further used to extract multi-scale high-order features. The graph convolution network is a graph neural network model based on convolution operations. It realizes local feature extraction and fusion by defining convolution kernels on the graph structure. The graph-level features obtained by the graph attention network are input into the pre-trained graph convolution network. After multiple layers of graph convolution operations, the high-order distribution patterns of the temperature field can be extracted at different scales. Each layer of graph convolution aggregates the features of the current node with the features of its neighboring nodes. By propagating layer by layer, local features reflecting the temperature distribution in different regions can be obtained. In the last layer of the graph convolution network, the output features are subjected to global maximum pooling operations to extract the most significant local features as candidate features. These candidate features represent the key distribution patterns of the temperature field at different scales.

[0085] The last step is to use the gated self-attention network to fuse the candidate features and generate the final set of key features for temperature regulation. The gated self-attention network is a neural network model that combines the attention mechanism and the gating unit, which can adaptively adjust the importance weights between features. First, multiple candidate features obtained by the graph convolutional network are input into the gated self-attention network, and an attention score matrix is ​​obtained by calculating the attention scores between the candidate features. The attention score measures the correlation and importance between different candidate features. Then, two gating units, the forget gate and the update gate, are introduced to generate a gated signal matrix. The forget gate controls the degree to which the feature information of the previous moment is retained, and the update gate controls the degree to which the feature information of the current moment is added. By element-by-element multiplication of the gated signal matrix and the attention score matrix, a feature importance weight matrix is ​​obtained, which represents the importance of different candidate features.

[0086] Based on the feature importance weight matrix, the candidate features are weighted and summed to obtain a weighted feature representation. This weighted feature representation integrates the temperature distribution information at different scales and locations, and optimizes the combination according to their importance. Finally, the weighted feature representation is averaged and pooled to obtain a compact feature vector as the final key feature set for temperature regulation. This feature set extracts the most essential and critical patterns and laws in the temperature field, which can be used as the basis for optimizing the temperature regulation strategy.

[0087] In an optional embodiment,

[0088] The steps before inputting the obtained graph-level embedded features into the pre-trained graph convolutional network also include:

[0089] Discretize the heat conduction equation into the graph domain. The heat conduction equation describes the physical relationship between material density, specific heat capacity, thermal conductivity, temperature field and internal heat source, and obtain its finite difference form. Based on the finite difference form, construct a physical regularization term. Combine the physical regularization term with the task loss function of the graph convolutional network to obtain a loss function that integrates physical constraints.

[0090] Training the graph convolutional network by minimizing the joint loss function of the fused physical constraints;

[0091] Among them, the calculation formula of the joint loss function is as follows:

[0092] ;

[0093] Among them, L represents the joint loss function, N represents the number of nodes, and T represents the time step. represents the temperature of node m at time step t in the real temperature field, represents the temperature of node m at time step t in the predicted temperature field, λ represents the equilibrium coefficient, η represents the material density, c p represents the specific heat capacity, represents the temperature of node m at time step t+1, △t represents the time interval between two consecutive time steps, μ represents thermal conductivity, N m represents the set of neighboring nodes of node m, represents the temperature of node n at time step t, d mn represents the distance between nodes m and n, represents the internal heat source of node m at time step t.

[0094] In this embodiment, through advanced deep learning models such as graph attention network, graph convolution network, gated self-attention network, etc., efficient representation and feature extraction of die-casting mold temperature field data are achieved. Graph attention network can capture the dynamic correlation pattern between sensors, graph convolution network can extract multi-scale high-order features, and gated self-attention network can adaptively adjust the importance of features. The organic combination of these three models makes full use of the spatiotemporal characteristics and process parameter information of temperature data, extracts the most critical and valuable feature representations from the complex temperature field, and lays a solid data foundation for the subsequent optimization of temperature regulation strategies. By integrating the knowledge of the physical mechanism of heat conduction and introducing regularization constraints based on partial differential equations, the physical interpretability of graph neural network temperature field modeling is significantly improved, making the model prediction results more in line with the physical laws of heat conduction. In engineering applications, temperature field prediction models with high physical consistency can provide more reliable decision support for the optimization of die-casting molding processes, reduce trial and error costs, and improve production efficiency and product quality.

[0095] S102. Based on the obtained target temperature adjustment key feature set, a multi-objective optimization model of mold temperature uniformity and product quality is constructed, and the multi-objective optimization model is divided into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it, and based on the solutions to the sub-problems obtained, the optimal solution of the multi-objective optimization model is obtained.

[0096] For each sub-problem, a corresponding intelligent agent (such as a reinforcement learning agent) is designed and solved using a distributed optimization algorithm based on the alternating direction multiplier method (ADMM). By iteratively updating the optimization variables and multiplier variables of each intelligent agent, the objective function value of each sub-problem is continuously converged, and the optimal solution of each sub-problem is finally obtained. The optimized solution of each sub-problem is used as the final solution of the multi-objective optimization model, that is, the optimal setting value of mold temperature adjustment.

[0097] In an optional embodiment,

[0098] The multi-objective optimization model is divided into a plurality of loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it. Based on the solutions of each sub-problem, the optimal solution of the multi-objective optimization model is obtained, including:

[0099] According to the temperature control area of ​​the mold, the multi-objective optimization model of mold temperature uniformity and product quality is divided into multiple loosely coupled sub-problems. Each sub-problem corresponds to a temperature control area, and coordination variables are introduced to realize information interaction between sub-problems.

[0100] For each of the subproblems, a proximal term is added to its objective function to construct a proximal subgradient subproblem, wherein the proximal term includes the coordination variable of the previous iteration of the current subproblem, the deviation between the optimization variable of the subproblem and the coordination variable, and the subgradient direction of the objective function of the subproblem;

[0101] Use the proximal subgradient method to solve each subproblem. By calculating the subgradient of the subproblem objective function at the current solution, use the subgradient descent method to update the optimization variable of the subproblem, and repeat the iteration until the preset subproblem convergence condition is met to obtain the approximate optimal solution of each subproblem.

[0102] Using the obtained approximate optimal solution of each subproblem, the coordination variables are updated through the linear combination of the optimization variables of the subproblem and the subgradient direction, and the Lagrange multipliers are updated based on the optimization variables of the subproblem and the updated coordination variables;

[0103] Repeat the steps of proximal subgradient subproblem construction, subproblem solution and coordination variable update until the difference between the optimization variable and the coordination variable of the subproblem between two consecutive iterations is less than the preset convergence threshold, and output the local optimal solution of each subproblem;

[0104] The local optimal solutions of each of the sub-problems are integrated to obtain the optimal solution of the multi-objective optimization model.

[0105] For example, the optimization model first needs to be decomposed according to the temperature control area of ​​the mold. The entire mold is divided into multiple relatively independent temperature control areas, each area corresponds to an optimization sub-problem. The purpose of this is to transform the original large-scale optimization problem into multiple smaller and easier-to-solve sub-problems, reducing the complexity of the optimization. At the same time, in order to ensure the consistency and coordination between different sub-problems, coordination variables are introduced as a bridge for information exchange between sub-problems. Coordination variables reflect the coupling relationship between different temperature control areas, and the coordinated optimization of sub-problems is achieved by optimizing coordination variables.

[0106] After the subproblems are divided, for each subproblem, the proximal term is introduced into its objective function to construct the proximal subgradient subproblem. The proximal term is a regularization term used to constrain the consistency between the current subproblem and other subproblems. The proximal term consists of three parts: the coordination variable obtained in the previous iteration of the current subproblem, the deviation between the optimization variable of the subproblem and the coordination variable, and the subgradient direction of the subproblem objective function. Among them, the coordination variable represents the coupling relationship between the current subproblem and other subproblems, the deviation between the optimization variable and the coordination variable reflects the degree of difference between the current subproblem and the overall optimization goal, and the subgradient direction indicates the optimization direction of the current subproblem.

[0107] After constructing the proximal subgradient subproblem, the proximal subgradient method is used to solve each subproblem. The proximal subgradient method is an optimization algorithm based on subgradient descent, which is suitable for solving non-smooth and non-convex optimization problems. For each subproblem, the subgradient of its objective function at the current solution is first calculated. The subgradient is a generalized gradient of the objective function at a non-smooth point, indicating the direction in which the objective function descends. Then, the subgradient descent method is used to update the optimization variables of the subproblem, that is, the optimization variables are updated in the opposite direction of the subgradient with a certain step size. Through multiple iterative optimizations, the optimization variables are continuously updated until the preset subproblem convergence conditions are met, such as the change in the optimization variables is less than a certain threshold or the number of iterations reaches the upper limit. In this way, the approximate optimal solution of each subproblem is obtained, that is, the local optimal point of the subproblem objective function under the constraints of the current coordination variables and the proximal terms.

[0108] After obtaining the approximate optimal solution for each subproblem, the coordination variables need to be updated to achieve information interaction and coordination between different subproblems. The update of the coordination variables is achieved through a linear combination of the optimization variables of the subproblems and the subgradient direction. Specifically, the optimization variables of each subproblem are weighted summed with the subgradient direction of its objective function at the optimal solution to obtain the update direction of the coordination variables. Then, along the update direction, the coordination variables are updated with a certain step size. This process can be regarded as fusing the local optimization information of each subproblem into the coordination variables, so that the coordination variables can reflect the global information of the overall optimization problem. While updating the coordination variables, it is also necessary to update the Lagrange multipliers, which are auxiliary variables used to process isoform constraints in optimization problems. The value of the Lagrange multipliers is calculated by the optimization variables of the subproblems and the updated coordination variables.

[0109] After completing a round of subproblem solving and coordination variable updating, the above steps need to be repeated until the global convergence condition is reached. In each round of iteration, the proximal subgradient subproblem is reconstructed, and each subproblem is solved using the updated coordination variables and Lagrange multipliers to obtain a new local optimal solution. Then, the coordination variables and Lagrange multipliers are updated again to transfer the optimization information between different subproblems. Repeat this iterative process until the difference between the optimization variables and coordination variables of all subproblems between two adjacent iterations is less than the preset convergence threshold, indicating that the optimization process has reached a stable state. At this time, the local optimal solution of each subproblem is output as the approximate global optimal solution of the multi-objective optimization model.

[0110] The last step is to integrate the local optimal solutions of each sub-problem to obtain the optimal solution of the entire multi-objective optimization model. The optimization variables of each sub-problem are combined according to their corresponding temperature control areas to form a complete optimization solution for mold temperature uniformity and product quality.

[0111] In an optional embodiment,

[0112] The calculation formula for updating the optimization variables of the subproblem using the subgradient descent method is as follows:

[0113] ;

[0114] in, represents the optimization variable of subproblem i at the k+1th iteration, Indicates that in P i The induced norm downward feasible domain X i The projection operation, P i represents the projection matrix of subproblem i, X i represents the feasible domain of subproblem i, represents the optimization variable of subproblem i at the kth iteration, αk represents the step size at the kth iteration, Denotes the objective function f of subproblem i i The subgradient at the current optimization variable, f i (·) represents the objective function, ρ represents the penalty term weight, represents the coordination variable of subproblem i, represents the Lagrange multiplier in subproblem i, ▽ represents the gradient symbol, h i Represents the constraint function of subproblem i.

[0115] In this embodiment, the complexity of the optimization problem is effectively reduced by dividing the multi-objective optimization model into multiple loosely coupled sub-problems and introducing coordination variables to realize information interaction between sub-problems. In each sub-problem, by constructing a proximal sub-gradient sub-problem and solving it using the proximal sub-gradient method, the approximate optimal solution of the sub-problem is obtained. Then, by updating the coordination variables and Lagrange multipliers, the coordinated optimization between different sub-problems is realized. Through multiple rounds of iterations, the solutions and coordination variables of the sub-problems are continuously updated, and finally the global optimal solution of the multi-objective optimization model is obtained. This scheme makes full use of the characteristics of the problem, and through the idea of ​​divide and conquer, the large-scale optimization problem is converted into multiple small-scale sub-problems, and the coupling between the sub-problems is realized through coordination variables, which not only improves the optimization efficiency, but also ensures the quality of the optimization results. At the same time, the introduction of the proximal sub-gradient method and Lagrange multipliers enables the scheme to effectively handle non-smooth and non-convex optimization problems, and has strong robustness and adaptability.

[0116] S103. The optimal solution of the multi-objective optimization model is used as the optimization target, and the temperature control key feature set is used as the state space. A reward function for multi-objective optimization is designed to construct a temperature control model. By solving the temperature control model, an initial temperature control strategy is obtained and sent to the edge computing node at the workshop site. The edge computing node collects real-time temperature data during the production process and detects temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, an online strategy update mechanism based on behavior cloning is triggered, and the initial temperature control strategy is adaptively optimized using the newly collected temperature data at the time of the anomaly, so as to generate a target temperature control strategy for guiding the real-time control of the mold temperature.

[0117] In order to achieve multi-objective optimization of mold temperature uniformity and product quality, it is necessary to build an optimization model containing multiple objective functions. The first objective function is used to measure the uniformity of mold temperature, which can be in the form of the weighted sum of squares of the difference between the actual temperature and the target temperature of each temperature control area. The second objective function is used to measure product quality, which can be in the form of the weighted sum of squares of the difference between the key quality indicators of the product and their target values.

[0118] In order to ensure the feasibility and rationality of the optimization results, it is necessary to set appropriate constraints on the optimization variables. Common constraints include the value range constraints of the temperature adjustment strategy, the thermal balance constraints between the temperature control areas, and the constraints of the product quality indicators. By combining the objective function and the constraints, a complete multi-objective optimization model can be obtained.

[0119] In order to achieve adaptive optimization of the temperature control strategy, a temperature control model needs to be constructed. First, the optimal solution of the multi-objective optimization model is used as the optimization target, and the key feature set of temperature control is used as the state space. The state space can include the following features: the actual temperature of each temperature control area, the target temperature of each temperature control area, the heating power of each temperature control area, the real-time value of the key quality indicators of the product, the ambient temperature and other external factors.

[0120] In order to guide the optimization direction of the temperature control model, a reasonable reward function needs to be designed. The reward function should comprehensively consider the two optimization goals of temperature uniformity and product quality. The weighted combination of the two objective functions can be used. Using the state space and reward function, a temperature control model can be constructed. The model can be solved based on the reinforcement learning algorithm to obtain the initial temperature control strategy. Common reinforcement learning algorithms include Q-learning, SARSA, PolicyGradient, etc.

[0121] In an optional embodiment,

[0122] The edge computing node collects real-time temperature data during the production process, detects temperature anomalies through a pre-built temperature anomaly detection model, and triggers an online strategy update mechanism based on behavior cloning when an anomaly is detected. The initial temperature adjustment strategy is adaptively optimized using the newly collected temperature data during the anomaly, and a target temperature adjustment strategy is generated to guide the real-time control of the mold temperature, including:

[0123] Build a temperature anomaly detection model based on LSTM to predict the temperature trend in the future. The deviation between the actual temperature value and the predicted temperature value is used as the anomaly score. When the anomaly score exceeds the set threshold, it is determined as a temperature anomaly, triggering an early warning and strategy update.

[0124] Collect historical temperature data under normal working conditions, use a sliding time window to intercept temperature time series fragments to construct a training set, and use the training set to pre-train the temperature anomaly detection model by minimizing the mean square error between the predicted temperature series and the actual temperature series;

[0125] Based on the pre-trained temperature anomaly detection model, the temperature anomaly detection model is fine-tuned using the pre-labeled temperature anomaly data to obtain the final trained temperature anomaly detection model;

[0126] Deploy the trained temperature anomaly detection model to the edge computing node, perform anomaly detection on the sliding time window of the real-time collected temperature data, identify temperature anomalies in real time through prediction errors, and trigger an early warning.

[0127] A feedforward neural network architecture is used to build a behavioral cloning model of the temperature control strategy. The input of the behavioral cloning model is the operating condition characteristics when the temperature is abnormal, and the output is the control strategy parameters. It is used to learn the expert's strategy decision mapping relationship by imitating the expert's control log in different abnormal scenarios.

[0128] Collect the action data of experts in different temperature abnormality scenarios, build a training set for the behavior cloning model, and minimize the deviation between the output of the behavior cloning model and the expert action data to enable the model to learn the expert's control strategy under abnormal conditions;

[0129] After the temperature anomaly detection model finds the temperature anomaly, the abnormal working condition characteristics are input into the behavior cloning model to generate the corresponding control strategy parameters, the initial temperature adjustment strategy is adaptively optimized, and the target temperature adjustment strategy is generated to guide the real-time control of the mold temperature.

[0130] For example, we first need to build a temperature anomaly detection model based on a long short-term memory network (LSTM). LSTM is a special type of recursive neural network that is good at processing time series data and can effectively learn and capture long-term dependencies in data. By training the LSTM model to learn the dynamic change law of the temperature time series, a mathematical model of temperature change is established to predict the temperature trend in the future. The predicted temperature value of the LSTM model is compared with the actual collected temperature value, and the deviation between them is calculated. The deviation is used as a measure of temperature anomaly, which is called the anomaly score. When the anomaly score exceeds the preset threshold, it is determined that the current temperature is abnormal, triggering the early warning mechanism, and starting the update of the temperature control strategy.

[0131] In order to train the LSTM temperature anomaly detection model, it is necessary to first collect historical temperature data under normal working conditions. Using the sliding time window method, the continuous temperature time series is divided into segments of fixed length, each of which contains temperature sampling points within a certain time range. These temperature segments are used as training samples to construct the training set of the anomaly detection model. The input of the training set is the temperature segment, and the output is the temperature value of the next moment of the corresponding segment. By minimizing the mean square error between the temperature sequence predicted by the LSTM model and the actual temperature sequence, the model parameters are optimized so that it can learn the dynamic characteristics of temperature changes. This process is called pre-training of the model, which aims to enable the model to grasp the general change law of the temperature time series.

[0132] After completing the pre-training, in order to further improve the performance of the anomaly detection model, it is necessary to fine-tune it based on the pre-training. Collect some pre-labeled temperature anomaly data, which contains various typical anomalies, such as temperature mutations and temperature fluctuations. Add these anomaly data to the training set and retrain the LSTM model so that it can better identify and distinguish anomalies. After the training is completed, the optimized LSTM temperature anomaly detection model is deployed on the edge computing node to perform anomaly detection on the real-time collected temperature data. Edge computing nodes refer to computing devices close to the data source, such as embedded systems, smart gateways, etc., which can process and analyze data locally in real time. By running the anomaly detection model on the edge computing node, the delay of data transmission can be greatly reduced, and real-time monitoring and early warning of temperature anomalies can be achieved. Specifically, the sliding time window method is used to continuously input the latest collected temperature data into the anomaly detection model, and the anomaly score is calculated by comparing the difference between the model's predicted temperature and the actual temperature. When the anomaly score exceeds the set threshold, an early warning is triggered immediately to notify relevant personnel to take countermeasures.

[0133] After detecting a temperature anomaly, it is not enough to simply issue an early warning. The temperature control strategy also needs to be adaptively adjusted to restore the temperature to the normal range as soon as possible. To this end, a behavioral cloning model based on a feedforward neural network is introduced to learn the control strategy of experts under temperature anomalies. Behavioral cloning is a method of imitation learning. By learning the expert's behavioral data, a mapping relationship from state to action is established, thereby cloning the expert's decision-making ability. Specifically, the operating characteristics of the temperature anomaly are used as the input of the behavioral cloning model, such as the temperature value, temperature change rate, current control parameters, etc. when the anomaly occurs, and the control strategy adopted by the expert is used as the output of the model, such as adjusting the pressure, speed, and temperature setting value of the die-casting machine. By training the behavioral cloning model, it can learn the strategic decision-making rules of experts in different abnormal scenarios.

[0134] In order to train the behavior cloning model, it is necessary to collect the decision data of experts when dealing with various temperature anomalies. This data includes the working condition characteristics when the anomaly occurs and the corresponding control actions taken by the experts. This data is organized into state-action pairs to construct the training set of the behavior cloning model. Then, the supervised learning method is used to optimize the parameters of the model by minimizing the difference between the output of the behavior cloning model and the expert's actions. This process can be regarded as the model imitating the decision-making behavior of the expert and learning the expert's control experience and strategy under abnormal conditions.

[0135] When the anomaly detection model finds a temperature anomaly, it inputs the current abnormal operating condition characteristics into the trained behavior cloning model, which generates the corresponding control strategy parameters. These control strategy parameters are dynamically adjusted according to the current abnormal situation and the experience of experts, and can adaptively optimize the initial temperature control strategy. The generated control strategy parameters are passed to the control system of the die-casting machine as the target temperature adjustment strategy to guide the real-time adjustment of the temperature. In this way, rapid response and strategy optimization in the case of temperature anomalies are achieved, and the temperature can be effectively maintained within a reasonable range, ensuring the stability of the die-casting process and product quality.

[0136] In this embodiment, by constructing a temperature anomaly detection model based on LSTM and a behavioral cloning model based on a feedforward neural network, real-time anomaly detection and adaptive control of the die-casting mold temperature are achieved. The anomaly detection model distinguishes and warns temperature anomalies in real time by learning the dynamic change law of the temperature time series. The behavioral cloning model learns the control strategy of experts, generates optimized control strategy parameters when anomalies occur, and adaptively adjusts the temperature control strategy. The collaborative work of these two models constitutes a closed-loop temperature anomaly detection and control system, which can detect temperature anomalies in time, and dynamically adjust the control strategy according to the abnormal situation to ensure the stability of the die-casting process and the consistency of product quality. This solution makes full use of the advantages of machine learning and expert experience, and provides an intelligent and adaptive temperature control method with strong practicality and robustness.

[0137] In an optional embodiment,

[0138] The method further comprises:

[0139] The temperature data of different workshops, production lines and equipment are collected on multiple edge nodes respectively. Based on the collected temperature data, a local temperature adjustment strategy model is constructed on each edge node using a feedforward neural network structure, and local model parameters are obtained by minimizing the strategy loss function and training the local temperature adjustment strategy model.

[0140] The local model parameters obtained by training each edge node are uploaded to the central server, and the central server determines the weight of the corresponding local model parameters according to the data volume of each edge node, and generates the global temperature adjustment strategy model parameters by aggregating the local model parameters;

[0141] The central server sends the generated global temperature regulation strategy model parameters to each edge node. Each edge node initializes the local temperature regulation strategy model using the received global temperature regulation strategy model parameters, and retrains the local temperature regulation strategy model based on the temperature data to update the local model parameters.

[0142] Repeat the iteration until the parameters of the global temperature control strategy model converge, and finally obtain the target temperature control strategy.

[0143] For example, it is necessary to first collect temperature data from different workshops, production lines, and equipment on multiple edge nodes. Edge nodes refer to computing devices close to data sources, such as industrial gateways and smart sensors, which can process and analyze data locally in real time. By deploying a temperature acquisition module on each edge node, temperature data from various areas and equipment can be obtained in real time, providing a data basis for subsequent temperature strategy learning.

[0144] After collecting temperature data, each edge node uses local data to build a local temperature regulation strategy model. The model uses a feedforward neural network structure, with the current temperature state and operating conditions as input and the temperature regulation control strategy as output. The local temperature regulation strategy model is trained by defining a strategy loss function, measuring the difference between the control strategy output by the model and the optimal strategy, and minimizing the loss function. During the training process, the parameters of the model are continuously updated using local temperature data so that it can learn the optimal control strategy for the local environment. Each edge node independently completes the training of the local model and obtains a set of local model parameters.

[0145] When all edge nodes complete local model training, they upload the trained local model parameters to the central server. The central server is a powerful cloud server that coordinates and manages the learning process of all edge nodes. In order to balance the contributions of different edge nodes, the central server calculates the weights of the corresponding local model parameters based on the amount of data uploaded by each node. Nodes with larger data volumes are assigned higher weights to their local model parameters, while nodes with smaller data volumes are assigned relatively lower parameter weights. Then, the central server aggregates the local model parameters of all nodes by weighted averaging to generate the parameters of the global temperature control strategy model.

[0146] After generating the global model parameters, the central server sends them to all edge nodes. After receiving the global parameters, each edge node uses them to initialize its own local temperature adjustment strategy model and replace the original local model parameters. In this way, the local models of all nodes inherit the knowledge of the global model and have preliminary temperature adjustment capabilities. However, since the global model is trained based on the data of all nodes, it may not be able to fully adapt to the local characteristics of each node. Therefore, after initializing the local model, each node also needs to retrain the local model based on its own temperature data to further optimize the model parameters so that it can better adapt to the local environment.

[0147] After completing a round of global parameter distribution and local model retraining, each edge node re-uploads the updated local model parameters to the central server to start a new round of global aggregation and update. The central server recalculates the weights of the parameters of each node, generates new global model parameters, and distributes them to all nodes. After receiving the new global parameters, the node updates the local model again and retrains. This cycle is repeated, and the global model and local model are continuously updated until the global model parameters converge, that is, the difference between the global parameters obtained in the previous and subsequent iterations is very small, or the preset number of iterations is reached. At this point, the global temperature adjustment strategy model obtained is the final target model, which integrates the data and knowledge of all nodes and can effectively guide the temperature adjustment of each area and device.

[0148] In practical applications, the target temperature adjustment strategy model can be deployed on each edge node to perform real-time reasoning on local temperature data and generate corresponding control instructions. At the same time, the central server can also use the target model to monitor and evaluate the global temperature adjustment effect, and update and optimize the model as needed. Through federated learning, the effective use of distributed temperature data is achieved, which not only protects data privacy but also realizes knowledge sharing and integration. Through collaborative learning, each node obtains a globally optimized temperature adjustment strategy that can adapt to different environments and working conditions and improve the temperature control level of the entire system.

[0149] In this embodiment, a distributed temperature control strategy optimization system is constructed by adopting the framework of federated learning. By collecting temperature data on multiple edge nodes and training the local temperature control strategy model on each node, and then performing global aggregation and updating through the central server, a comprehensive optimized global temperature control strategy is finally obtained. This distributed learning method makes full use of the data and computing resources scattered in each node, and achieves the dual goals of data privacy protection and knowledge sharing. At the same time, through the iterative optimization process, the accuracy and robustness of the temperature control strategy are continuously improved, so that it can adapt to complex and changeable industrial environments. This solution provides an intelligent and adaptive solution for industrial temperature control, which has broad application prospects.

[0150] Figure 2 FIG. 1 is a schematic diagram of the structure of a mold temperature adaptive adjustment system based on machine learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0151] The first unit is used to collect real-time temperature data and temperature-related process parameters in the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, perform feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and use a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features;

[0152] The second unit is used to adjust the key feature set of the target temperature, build a multi-objective optimization model of mold temperature uniformity and product quality according to the obtained target temperature, divide the multi-objective optimization model into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables, each sub-problem corresponds to a temperature control area of ​​the mold, and solve each sub-problem by using a distributed optimization algorithm based on the alternating direction multiplier method, and obtain the optimal solution of the multi-objective optimization model based on the solutions of each sub-problem;

[0153] The third unit is used to take the optimal solution of the multi-objective optimization model as the optimization target, the temperature regulation key feature set as the state space, design a reward function for multi-objective optimization, and construct a temperature regulation model. By solving the temperature regulation model, an initial temperature regulation strategy is obtained, and the initial temperature regulation strategy is sent to the edge computing node at the workshop site. The edge computing node collects real-time temperature data in the production process and detects temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, the behavior cloning-based strategy online update mechanism is triggered, and the initial temperature regulation strategy is adaptively optimized using the newly collected temperature data at the time of the anomaly, and a target temperature regulation strategy is generated to guide the real-time control of the mold temperature.

[0154] According to a third aspect of the embodiments of the present invention,

[0155] An electronic device is provided, comprising:

[0156] processor;

[0157] a memory for storing processor-executable instructions;

[0158] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0159] A fourth aspect of the embodiments of the present invention is:

[0160] 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.

[0161] 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.

[0162] 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 adaptively adjusting mold temperature based on machine learning, characterized in that: include: Collecting real-time temperature data and temperature-related process parameters during the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, performing feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and using a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features; According to the obtained target temperature adjustment key feature set, a multi-objective optimization model of mold temperature uniformity and product quality is constructed, and the multi-objective optimization model is divided into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it, and based on the solutions to each sub-problem, the optimal solution of the multi-objective optimization model is obtained; The optimal solution of the multi-objective optimization model is taken as the optimization target, and the target temperature regulation key feature set is taken as the state space. A reward function for multi-objective optimization is designed, and a temperature regulation model is constructed. By solving the temperature regulation model, an initial temperature regulation strategy is obtained and simultaneously sent to the edge computing nodes at the workshop site. The edge computing nodes collect real-time temperature data during the production process and detect temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, a strategy online update mechanism based on behavior cloning is triggered, and the initial temperature regulation strategy is adaptively optimized using the newly collected temperature data during the anomaly, so as to generate a target temperature regulation strategy for guiding the real-time control of the mold temperature.

2. The method according to claim 1, characterized in that: Feature extraction is performed on real-time temperature data and process parameters related to temperature to obtain multiple candidate features. The key feature selection algorithm is used to select the target temperature regulation key feature set from multiple candidate features, including: The real-time temperature data collected by each temperature sensor is organized into a two-dimensional time series matrix, where each row of the matrix represents the temperature time series of a sensor; an adjacency matrix is ​​constructed based on the geometric structure of the mold and the sensor layout, where each element of the adjacency matrix represents the spatial connection relationship between two sensors; the process parameters are converted into an attribute matrix, where each row of the attribute matrix represents the local process parameters corresponding to a sensor position; a unified attribute graph is constructed based on the two-dimensional time series matrix, the adjacency matrix and the attribute matrix, where the node attributes of the attribute graph are the temperature time series and the local process parameters, and the edge attributes are the spatial connection relationship between sensors; The attribute graph is input into the graph attention network. The attention weight between each node and its neighbor nodes is calculated through the attention mechanism. The features of the neighbor nodes are aggregated using the attention weight to obtain the output features of the node. The receptive field of the node is gradually expanded by stacking multiple layers of graph attention networks to aggregate high-order neighborhood information. The output features of the last layer of graph attention network are globally averaged and pooled to obtain graph-level embedded features that integrate the dynamic association patterns between temperature sensors. The obtained graph-level embedded features are input into the pre-trained graph convolutional network, and local smoothing and transformation are performed through multi-layer graph convolution operations to extract multi-scale high-order features. The output features of the last layer of graph convolutional network are subjected to global maximum pooling operations to obtain multiple candidate features reflecting different high-order distribution patterns of the temperature field. The obtained multiple candidate features are input into the gated self-attention network, and the attention score matrix is ​​obtained by calculating the attention scores between the candidate features. The forget gate and the update gate are introduced to generate the gating signal matrix. The attention score matrix is ​​weighted by the gating signal matrix to obtain the feature importance weight matrix. Based on the feature importance weight matrix, multiple candidate features are weighted to obtain weighted feature representations. The weighted feature representations are average pooled to obtain the final target temperature regulation key feature set.

3. The method according to claim 2, characterized in that The steps before inputting the obtained graph-level embedded features into the pre-trained graph convolutional network also include: Discretize the heat conduction equation into the graph domain. The heat conduction equation describes the physical relationship between material density, specific heat capacity, thermal conductivity, temperature field and internal heat source, and obtain its finite difference form. Based on the finite difference form, construct a physical regularization term. Combine the physical regularization term with the task loss function of the graph convolutional network to obtain a loss function that integrates physical constraints. Training the graph convolutional network by minimizing the joint loss function of the fused physical constraints; Among them, the calculation formula of the joint loss function is as follows: ； Among them, L represents the joint loss function, N represents the number of nodes, and T represents the time step. represents the temperature of node m at time step t in the real temperature field, represents the temperature of node m at time step t in the predicted temperature field, λ represents the equilibrium coefficient, η represents the material density, c p represents the specific heat capacity, represents the temperature of node m at time step t+1, △t represents the time interval between two consecutive time steps, μ represents thermal conductivity, N m represents the set of neighboring nodes of node m, represents the temperature of node n at time step t, d mn represents the distance between nodes m and n, represents the internal heat source of node m at time step t.

4. The method according to claim 1, characterized in that The multi-objective optimization model is divided into a plurality of loosely coupled sub-problems by domain decomposition and introduction of coordination variables. Each sub-problem corresponds to a temperature control area of ​​the mold. For each sub-problem, a distributed optimization algorithm based on the alternating direction multiplier method is used to solve it. Based on the solutions of each sub-problem, the optimal solution of the multi-objective optimization model is obtained, including: According to the temperature control area of ​​the mold, the multi-objective optimization model of mold temperature uniformity and product quality is divided into multiple loosely coupled sub-problems. Each sub-problem corresponds to a temperature control area, and coordination variables are introduced to realize information interaction between sub-problems. For each of the sub-problems, a proximal term is added to the objective function thereof, wherein the proximal term includes the coordination variable of the last iteration of the current sub-problem, the deviation between the optimization variable of the sub-problem and the coordination variable, and the sub-gradient direction of the objective function of the sub-problem; Use the proximal subgradient method to solve each subproblem. By calculating the subgradient of the subproblem objective function at the current solution, use the subgradient descent method to update the optimization variable of the subproblem, and repeat the iteration until the preset subproblem convergence condition is met to obtain the approximate optimal solution of each subproblem. Using the obtained approximate optimal solution of each subproblem, the coordination variables are updated through the linear combination of the optimization variables of the subproblem and the subgradient direction, and the Lagrange multipliers are updated based on the optimization variables of the subproblem and the updated coordination variables; Repeat the steps of proximal subgradient subproblem construction, subproblem solution and coordination variable update until the difference between the optimization variable and the coordination variable of the subproblem between two consecutive iterations is less than the preset convergence threshold, and output the local optimal solution of each subproblem; The local optimal solutions of each of the sub-problems are integrated to obtain the optimal solution of the multi-objective optimization model.

5. The method according to claim 4, characterized in that The calculation formula for updating the optimization variables of the subproblem using the subgradient descent method is as follows: ； in, represents the optimization variable of subproblem i at the k+1th iteration, Indicates that in P i The induced norm downward feasible domain X i The projection operation, P i represents the projection matrix of subproblem i, X i represents the feasible domain of subproblem i, represents the optimization variable of subproblem i at the kth iteration, α k represents the step size at the kth iteration, represents the objective function f of subproblem i i The subgradient at the current optimization variable, f i (·) represents the objective function, ρ represents the penalty term weight, represents the coordination variable of subproblem i, represents the Lagrange multiplier in subproblem i, ▽ represents the gradient symbol, h i Represents the constraint function of subproblem i.

6. The method according to claim 1, characterized in that The edge computing node collects real-time temperature data during the production process, detects temperature anomalies through a pre-built temperature anomaly detection model, and triggers an online strategy update mechanism based on behavior cloning when an anomaly is detected. The initial temperature adjustment strategy is adaptively optimized using the newly collected temperature data during the anomaly, and a target temperature adjustment strategy is generated to guide the real-time control of the mold temperature, including: Build a temperature anomaly detection model based on LSTM to predict the temperature trend in the future. The deviation between the actual temperature value and the predicted temperature value is used as the anomaly score. When the anomaly score exceeds the set threshold, it is determined as a temperature anomaly, triggering an early warning and strategy update. Collect historical temperature data under normal working conditions, use a sliding time window to intercept temperature time series fragments to construct a training set, and use the training set to pre-train the temperature anomaly detection model by minimizing the mean square error between the predicted temperature series and the actual temperature series; Based on the pre-trained temperature anomaly detection model, the temperature anomaly detection model is fine-tuned using the pre-labeled temperature anomaly data to obtain the final trained temperature anomaly detection model; Deploy the trained temperature anomaly detection model to the edge computing node, perform anomaly detection on the sliding time window of the real-time collected temperature data, identify temperature anomalies in real time through prediction errors, and trigger an early warning. A feedforward neural network architecture is used to build a behavioral cloning model of the temperature control strategy. The input of the behavioral cloning model is the operating condition characteristics when the temperature is abnormal, and the output is the control strategy parameters. It is used to learn the expert's strategy decision mapping relationship by imitating the expert's control log in different abnormal scenarios. Collect the action data of experts in different temperature abnormality scenarios, build a training set for the behavior cloning model, and minimize the deviation between the output of the behavior cloning model and the expert action data to enable the model to learn the expert's control strategy under abnormal conditions; After the temperature anomaly detection model finds the temperature anomaly, the abnormal working condition characteristics are input into the behavior cloning model to generate the corresponding control strategy parameters, the initial temperature adjustment strategy is adaptively optimized, and the target temperature adjustment strategy is generated to guide the real-time control of the mold temperature.

7. The method according to claim 1, characterized in that The method further comprises: The temperature data of different workshops, production lines and equipment are collected on multiple edge nodes respectively. Based on the collected temperature data, a local temperature adjustment strategy model is constructed on each edge node using a feedforward neural network structure, and local model parameters are obtained by minimizing the strategy loss function and training the local temperature adjustment strategy model. The local model parameters obtained by training each edge node are uploaded to the central server, and the central server determines the weight of the corresponding local model parameters according to the data volume of each edge node, and generates the global temperature adjustment strategy model parameters by aggregating the local model parameters; The central server sends the generated global temperature regulation strategy model parameters to each edge node. Each edge node initializes the local temperature regulation strategy model using the received global temperature regulation strategy model parameters, and retrains the local temperature regulation strategy model based on the temperature data to update the local model parameters. Repeat the iteration until the parameters of the global temperature control strategy model converge, and finally obtain the target temperature control strategy.

8. A mold temperature adaptive adjustment system based on machine learning, 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 real-time temperature data and temperature-related process parameters in the mold production process, wherein the process parameters include at least one of die-casting pressure, die-casting speed, material temperature and ambient temperature, perform feature extraction on the real-time temperature data and the temperature-related process parameters to obtain multiple candidate features, and use a key feature selection algorithm to screen out a target temperature adjustment key feature set from the multiple candidate features; The second unit is used to adjust the key feature set of the target temperature, build a multi-objective optimization model of mold temperature uniformity and product quality according to the obtained target temperature, divide the multi-objective optimization model into multiple loosely coupled sub-problems by domain decomposition and introduction of coordination variables, each sub-problem corresponds to a temperature control area of ​​the mold, and solve each sub-problem by using a distributed optimization algorithm based on the alternating direction multiplier method, and obtain the optimal solution of the multi-objective optimization model based on the solutions of each sub-problem; The third unit is used to take the optimal solution of the multi-objective optimization model as the optimization target, the target temperature regulation key feature set as the state space, design a reward function for multi-objective optimization, build a temperature regulation model, and obtain the initial temperature regulation strategy by solving the temperature regulation model. The initial temperature regulation strategy is sent to the edge computing node at the workshop site. The edge computing node collects real-time temperature data during the production process and detects temperature anomalies through a pre-built temperature anomaly detection model. When an anomaly is detected, the behavior cloning-based strategy online update mechanism is triggered, and the initial temperature regulation strategy is adaptively optimized using the newly collected temperature data at the time of the anomaly, to generate a target temperature regulation strategy for guiding the real-time control of the mold temperature.

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