New material production process parameter optimization method and system based on artificial intelligence

Abstract

The invention provides a new material production process parameter optimization method and system based on artificial intelligence, and relates to the technical field of parameter optimization, and the method comprises the steps: constructing a multi-dimensional process parameter feature space, collecting and preprocessing process parameters, extracting correlation features through a deep convolutional neural network, and building a parameter feature space model; generating a technological parameter collaborative optimization model, constructing a multi-objective optimization function, and performing iterative optimization by using reinforcement learning and an improved ant colony algorithm; and performing process parameter adaptive control, constructing a dynamic control interval, collecting parameters in real time and performing adaptive adjustment to realize closed-loop control and model dynamic update.

Description

Technical Field

[0001] The present invention relates to parameter optimization technology, and in particular to a method and system for optimizing process parameters of new material production based on artificial intelligence. Background Art

[0002] Traditional methods are difficult to effectively handle the complex relationships between multi-dimensional process parameters. The preparation process of new materials involves multiple parameters such as temperature, pressure, and raw material ratio. There is a nonlinear relationship between these parameters. It is difficult to fully grasp the interaction between parameters based on manual experience alone, which easily leads to local optimization results.

[0003] Secondly, the existing optimization methods lack accurate modeling of the relationship between process parameters and material performance indicators. Due to the complexity of the preparation process of new materials, traditional mathematical models are difficult to accurately describe the mechanism of the influence of process parameters on material properties, which affects the reliability and applicability of the optimization results.

[0004] Finally, existing technologies lack the ability to adaptively control process parameters in real time during actual production. The preparation of new materials often faces uncertainties such as fluctuations in raw material batches and changes in environmental factors. Fixed process parameter settings are difficult to adapt to these changes, which can easily cause fluctuations in material properties and reduced production efficiency. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for optimizing process parameters of new material production based on artificial intelligence, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provides new material production process parameter optimization methods based on artificial intelligence, including: Constructing a multidimensional process parameter feature space, including: collecting temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters in the preparation process of new materials; performing data cleaning and normalization preprocessing on the collected multidimensional process parameters to construct a standardized parameter matrix; using a deep convolutional neural network to extract the correlation characteristics between multidimensional process parameters and establish a parameter feature space model; based on the established parameter feature space model, calculating the mapping weight matrix between process parameters; Generate a collaborative optimization model for process parameters, including: construct a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; input the constructed multi-objective optimization function into a reinforcement learning model to obtain an initial solution for the process parameters; use an improved ant colony algorithm to iteratively optimize the initial solution for the process parameters to generate a collaborative optimization model for process parameters; use a K-fold cross-validation method to evaluate the prediction accuracy of the collaborative optimization model for process parameters to obtain model evaluation indicators; Execute adaptive control of process parameters, including: constructing a dynamic control range of process parameters based on the obtained model evaluation indicators; collecting process parameters in the new material preparation process in real time, and inputting the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, triggering the adaptive control algorithm and calculating the process parameter correction value; based on the calculated process parameter correction value, automatically adjusting the operating parameters of the preparation equipment to achieve closed-loop control of the new material preparation process; feeding back the control results to the parameter feature space model, and dynamically updating the process parameter collaborative optimization model.

[0007] The deep convolutional neural network is used to extract the correlation features between multi-dimensional process parameters and establish the parameter feature space model including: Based on the input standardized parameter matrix, multiple convolution kernels of different sizes are used for feature extraction, and the convolution kernels of different sizes are 3×3 convolution kernel, 5×5 convolution kernel and 7×7 convolution kernel respectively. The standardized parameter matrix is ​​convolved by the convolution kernels of different sizes to generate multiple groups of feature maps, and each group of feature maps is nonlinearly transformed by a standardized layer and a ReLU activation function to obtain a parameter-associated feature map; The parameter-associated feature map is weighted in channel dimension and space dimension, wherein channel statistical information is obtained through global average pooling operation and maximum pooling operation, and the channel statistical information is input into a shared multi-layer perceptron to obtain channel weight; a spatial weight map is generated through convolution operation, and the channel weight and the spatial weight map are respectively multiplied and superimposed with the parameter-associated feature map to obtain a weighted feature representation; A 1×1 convolution kernel is used to uniformly adjust the number of channels represented by the weighted features to generate an adjusted feature map, and the adjusted feature map is subjected to feature fusion through a residual connection structure, and the weight coefficients of the residual connection structure are obtained by back propagation; the fused features are mapped to a feature space of preset dimensions through a fully connected layer to construct a parameter feature space model.

[0008] Based on the established parameter feature space model, the mapping weight matrix between the process parameters is calculated including: The feature vector output by the parameter feature space model is constructed as a fully connected parameter association graph, in which each node corresponds to a process parameter, and the edges between the nodes represent the influence relationship between the parameters. A multi-head attention mechanism is introduced based on the parameter association graph, and the attention coefficient between the node pairs is obtained by calculating the node feature similarity and structural correlation; A message passing mechanism is constructed based on a gated recurrent unit. The message passing mechanism uses the attention coefficient as a weight to perform weighted aggregation on node neighbor information. The aggregated information is nonlinearly transformed through the gated recurrent unit to update the node representation. The node representation is integrated into the global correlation information through multiple rounds of message passing to obtain the final node representation. The final node representation is input into a multilayer perceptron, and a mapping weight value between each pair of parameter nodes is calculated, wherein the mapping weight value represents the influence of the source parameter on the target parameter, and the mapping weight values ​​between all parameter pairs are organized into a mapping weight matrix, wherein each element in the mapping weight matrix corresponds to the strength of the mapping relationship between a pair of parameter nodes.

[0009] Based on the obtained mapping weight matrix, a multi-objective optimization function between process parameters and material performance indicators is constructed; the constructed multi-objective optimization function is input into the reinforcement learning model to obtain the initial solution of process parameters including: Obtaining a mapping weight matrix between process parameters and material performance indicators, wherein the rows of the mapping weight matrix represent material performance indicators, the columns represent process parameters, and the matrix elements represent the influence weights of the corresponding process parameters on the performance indicators; Constructing a performance indicator prediction function based on the mapping weight matrix, wherein the performance indicator prediction function performs a weighted combination of process parameters through weight coefficients and adds a bias term, wherein the weight coefficients are determined by corresponding elements in the mapping weight matrix; Constructing a parameter coupling constraint function according to the correlation strength between the process parameters in the mapping weight matrix, wherein the parameter coupling constraint function calculates the deviation between the actual value of the process parameter and the theoretical corresponding relationship, wherein the theoretical corresponding relationship is derived based on the process theory; The performance indicator prediction function is combined with the parameter coupling constraint function to construct a multi-objective optimization function, and a regularization term is introduced into the multi-objective optimization function to constrain the parameter value range; A reinforcement learning model with a dual network structure is used to solve the multi-objective optimization function to obtain an initial solution for the process parameters; the reinforcement learning model includes an evaluation network and a target network, the evaluation network receives the process parameter state output parameter adjustment strategy, and the target network uses a soft update mechanism to update the network parameters.

[0010] The iterative optimization of the initial solution of process parameters using the improved ant colony algorithm includes: The initial solution of the process parameters is used as the center point to construct a parameter search grid, and a search interval is divided in each parameter dimension and discretized, and the search interval is extended to both sides by a preset range based on the initial solution of the process parameters; Constructing a heuristic information matrix, wherein the element values ​​in the heuristic information matrix are calculated by the ratio of the distance difference between the parameter values ​​of adjacent grid points and the scale parameter, wherein the scale parameter is used to adjust the similarity of the parameter values; An adaptive pheromone update mechanism is constructed based on the global update and local update of pheromone concentration. The global update of pheromone concentration is determined by elite ants, and the local update of pheromone concentration is determined by the objective function value of all ants in the current iteration. The pheromone volatility coefficient is dynamically adjusted exponentially with the search process. Introducing a dynamic neighborhood search mechanism in the parameter search process, wherein the search range of the dynamic neighborhood search mechanism decays exponentially with the number of iterations, and the dynamic neighborhood search mechanism selects a preset proportion of ants to perform local fine search in each iteration; Adopting the adaptively adjusted pheromone importance factor and the heuristic information importance factor to construct the state transition probability, wherein the pheromone importance factor increases with the number of iterations, and the heuristic information importance factor decreases with the number of iterations, and determining the moving direction of the ant in the parameter space based on the state transition probability; The initial solution of the process parameters is iteratively optimized, and when the improvement of the optimal solution of multiple consecutive generations is less than a preset threshold, the iteration is terminated and the optimized process parameter combination is output.

[0011] When the process parameters exceed the dynamic control range, the adaptive control algorithm is triggered to calculate the process parameter correction value; based on the calculated process parameter correction value, the operating parameters of the preparation equipment are automatically adjusted to achieve closed-loop control of the new material preparation process, including: A dynamic control interval of process parameters is constructed based on parameter target values ​​and standard deviations, wherein the standard deviation is calculated through a sliding time window, and upper and lower limits of the dynamic control interval of process parameters are dynamically adjusted over time; When the process parameter exceeds the process parameter dynamic control interval, the deviation integral term and the deviation change rate of the process parameter are calculated, and the feedback control correction value is calculated based on the deviation integral term and the deviation change rate using an adaptively adjusted control coefficient, wherein the control coefficient includes a proportional coefficient positively correlated with the deviation value, an integral coefficient negatively correlated with the deviation value, and a differential coefficient that exponentially decays with time; Calculate the future state value of the process parameter based on the time series prediction model, calculate the deviation between the future state value and the parameter target value through the feedforward compensation coefficient to obtain a feedforward compensation correction value, and combine the feedback control correction value with the feedforward compensation correction value to obtain a process parameter correction value; Mapping the process parameter correction value into a parameter group correction vector based on a parameter coupling constraint matrix, wherein the parameter coupling constraint matrix is ​​used to ensure coordinated adjustment of multiple process parameters; The parameter group correction vector is converted into an execution signal by using an actuator dynamic model to realize closed-loop control of the new material preparation process, and the execution signal includes dead zone compensation and rate limitation.

[0012] According to a second aspect of the embodiments of the present invention, Provides an artificial intelligence-based new material production process parameter optimization system, including: The first unit is used to collect temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters in the preparation process of new materials; perform data cleaning and normalization preprocessing on the collected multi-dimensional process parameters to construct a standardized parameter matrix; use a deep convolutional neural network to extract the correlation characteristics between multi-dimensional process parameters and establish a parameter feature space model; based on the established parameter feature space model, calculate the mapping weight matrix between process parameters; The second unit is used to construct a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; input the constructed multi-objective optimization function into the reinforcement learning model to obtain the initial solution of the process parameters; use the improved ant colony algorithm to iteratively optimize the initial solution of the process parameters to generate a process parameter collaborative optimization model; use the K-fold cross-validation method to evaluate the prediction accuracy of the process parameter collaborative optimization model and obtain the model evaluation index; The third unit is used to construct a dynamic control range of process parameters based on the obtained model evaluation indicators; collect the process parameters in the new material preparation process in real time, and input the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, trigger the adaptive control algorithm and calculate the process parameter correction value; based on the calculated process parameter correction value, automatically adjust the operating parameters of the preparation equipment to achieve closed-loop control of the new material preparation process; feed back the control results to the parameter feature space model, and dynamically update the process parameter collaborative optimization model.

[0013] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of the embodiments of the present invention, 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.

[0015] The beneficial effects of this application are as follows: 1. By constructing a multi-dimensional process parameter feature space, using a deep convolutional neural network to extract the correlation features between parameters, and calculating the mapping weight matrix, it is possible to accurately capture the complex relationship between process parameters, provide a reliable data basis for subsequent parameter optimization, and effectively improve the accuracy and comprehensiveness of process parameter analysis.

[0016] 2. By generating a collaborative optimization model for process parameters, combining reinforcement learning and improved ant colony algorithm for parameter optimization, and using K-fold cross validation to evaluate model accuracy, the intelligent optimization of process parameters is achieved, which significantly improves the parameter optimization efficiency of the new material preparation process and reduces the workload of manual parameter adjustment.

[0017] 3. By performing adaptive control of process parameters, establishing a dynamic control range and monitoring parameter changes in real time, corrections are automatically triggered when parameters are abnormal, thus achieving closed-loop control of the new material preparation process. At the same time, the optimization model is continuously updated through a feedback mechanism, which improves the stability and reliability of the production process and effectively ensures the preparation quality of new materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a process flow of a method for optimizing process parameters of new material production based on artificial intelligence according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a new material production process parameter optimization system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] 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.

[0020] 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.

[0021] Figure 1 FIG. 1 is a flow chart of a method for optimizing process parameters of new material production based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Constructing a multidimensional process parameter feature space, including: collecting temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters in the preparation process of new materials; performing data cleaning and normalization preprocessing on the collected multidimensional process parameters to construct a standardized parameter matrix; using a deep convolutional neural network to extract correlation features between multidimensional process parameters and establish a parameter feature space model; based on the established parameter feature space model, calculating a mapping weight matrix between process parameters; S102. Generate a collaborative optimization model for process parameters, including: constructing a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; inputting the constructed multi-objective optimization function into a reinforcement learning model to obtain an initial solution for the process parameters; using an improved ant colony algorithm to iteratively optimize the initial solution for the process parameters to generate a collaborative optimization model for the process parameters; using a K-fold cross-validation method to evaluate the prediction accuracy of the collaborative optimization model for the process parameters to obtain a model evaluation indicator; S103. Execute adaptive control of process parameters, including: constructing a dynamic control range of process parameters based on the obtained model evaluation indicators; collecting process parameters in the new material preparation process in real time, and inputting the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, triggering the adaptive control algorithm and calculating the process parameter correction value; based on the calculated process parameter correction value, automatically adjusting the operating parameters of the preparation equipment to achieve closed-loop control of the new material preparation process; feeding back the control results to the parameter feature space model, and dynamically updating the process parameter collaborative optimization model.

[0022] The specific implementation method of the new material production process parameter optimization method based on artificial intelligence is as follows: In the stage of constructing the multidimensional process parameter feature space, the key process parameters in the preparation process of new materials are first collected through a distributed sensor network. The specific parameters collected include: temperature parameters in the reactor (range 20-200℃, sampling frequency 1 time / second), system pressure parameters (range 0-10MPa, sampling frequency 1 time / second), mass ratio parameters of raw materials A / B / C (A:B:C=2:3:5), reaction time parameters (range 0-120 minutes), environmental humidity parameters (range 30%-70%) and stirring rate parameters (range 100-1000rpm).

[0023] The collected raw data were cleaned and outliers and missing values ​​were removed. The sliding window median filter method was used to remove data noise, and the window size was set to 5. Then the parameters of different dimensions were normalized to the interval [0,1], and a 6×N-dimensional standardized parameter matrix was constructed. The parameter features were extracted using an 8-layer convolutional neural network, where the convolution kernel size was 3×3, the step size was 1, and the number of channels in each layer was 64, 128, 256, 512, 256, 128, 64, and 32, respectively. The correlation feature mapping between parameters was obtained through network training, and the feature space model was established. The 6×6-dimensional mapping weight matrix between parameters was calculated based on the trained model.

[0024] In the stage of generating the collaborative optimization model of process parameters, a multi-objective optimization function including performance indicators such as material strength, toughness, and stability was first constructed based on the mapping weight matrix. The optimization function was input into the deep reinforcement learning model, and the dual Q network structure was adopted. After 1000 rounds of training, the initial solution of the process parameters was obtained. Then, the improved ant colony algorithm was used to optimize the initial solution, with the population size set to 100, the number of iterations to 500, and the pheromone volatility coefficient to 0.1. The model performance was evaluated using 5-fold cross validation, and the average prediction accuracy of the model on the validation set reached 95.3%.

[0025] In the stage of executing the adaptive control of process parameters, the dynamic control range of each parameter is constructed based on the model evaluation index, with a temperature control accuracy of ±2°C and a pressure control accuracy of ±0.1MPa. The real-time collected process parameters are transmitted to the control system via Ethernet and input into the optimization model for online calculation. When the parameters deviate from the control range, the adaptive PID algorithm calculates the correction value and sends it to the actuator through the fieldbus network. The control system feeds back the latest control results to the feature space model every 5 minutes for online updating of the optimization model parameters.

[0026] The beneficial effects of this method include: Through deep learning, multi-dimensional process parameter features are extracted, and a correlation mapping model between parameters is established, which improves the accuracy and reliability of process parameter optimization. The combination of reinforcement learning and improved ant colony algorithm is used to optimize parameters, which overcomes the problem that traditional optimization algorithms are prone to fall into local optimality, and the optimization efficiency is improved by more than 40%. Adaptive closed-loop control of process parameters is achieved, the system response time is less than 100ms, and the control accuracy is improved by 50%, which significantly improves the stability of the new material preparation process and product quality.

[0027] In an optional embodiment, a deep convolutional neural network extracts correlation features between multi-dimensional process parameters, and establishing a parameter feature space model includes: Based on the input standardized parameter matrix, multiple convolution kernels of different sizes are used for feature extraction, and the convolution kernels of different sizes are 3×3 convolution kernel, 5×5 convolution kernel and 7×7 convolution kernel respectively. The standardized parameter matrix is ​​convolved by the convolution kernels of different sizes to generate multiple groups of feature maps, and each group of feature maps is nonlinearly transformed by a standardized layer and a ReLU activation function to obtain a parameter-associated feature map; The parameter-associated feature map is weighted in channel dimension and space dimension, wherein channel statistical information is obtained through global average pooling operation and maximum pooling operation, and the channel statistical information is input into a shared multi-layer perceptron to obtain channel weight; a spatial weight map is generated through convolution operation, and the channel weight and the spatial weight map are respectively multiplied and superimposed with the parameter-associated feature map to obtain a weighted feature representation; A 1×1 convolution kernel is used to uniformly adjust the number of channels represented by the weighted features to generate an adjusted feature map, and the adjusted feature map is subjected to feature fusion through a residual connection structure, and the weight coefficients of the residual connection structure are obtained by back propagation; the fused features are mapped to a feature space of preset dimensions through a fully connected layer to construct a parameter feature space model.

[0028] The specific implementation method of using deep convolutional neural network to extract the correlation features between multi-dimensional process parameters and establish parameter feature space model is as follows: First, obtain multidimensional process parameter data, preprocess and standardize the parameter data to obtain a standardized parameter matrix. The standardization process can use the z-score standardization method, which subtracts the mean of each parameter value from the parameter and divides it by the standard deviation. For example, for a process parameter matrix, it contains 10 parameters such as temperature, pressure, and flow, and each parameter has 100 sample points. After standardization, a 100x10 standardized parameter matrix is ​​obtained.

[0029] Next, multiple convolution kernels of different sizes are used to perform convolution operations on the standardized parameter matrix to extract the correlation features between the parameters. Specifically, three convolution kernels of 3x3, 5x5 and 7x7 are used, and the number of convolution kernels is 32, 64 and 128 respectively. Taking the 3x3 convolution kernel as an example, 32 3x3 convolution kernels are convolved on the 100x10 input matrix to obtain 32 98x8 feature maps. Similarly, the 5x5 and 7x7 convolution kernels obtain 64 96x6 and 128 94x4 feature maps respectively.

[0030] Then all feature maps are normalized and ReLU activated to obtain parameter-related feature maps. The batch normalization method is used to normalize each channel of each feature map. The ReLU activation function sets negative values ​​to 0 and keeps positive values ​​unchanged. After this step, multiple sets of feature maps representing parameter-related features are obtained.

[0031] The next step is to weight the parameter-related feature map in the channel dimension and the spatial dimension. First, the statistical information of the channel dimension is obtained by global average pooling and maximum pooling, and then input into a shared multi-layer perceptron composed of two fully connected layers to output the channel weight vector. At the same time, the feature map is convolved with a 7x7 convolution kernel to obtain a spatial weight map. The channel weight vector and the spatial weight map are multiplied and superimposed with the feature map to obtain the weighted feature representation.

[0032] Then, the number of channels represented by the weighted features is uniformly adjusted using a 1x1 convolution kernel to 256. The features are then fused using a residual connection structure, and the weight coefficients of the residual connection are obtained through back propagation. Finally, the fused features are mapped to a 128-dimensional feature space through a fully connected layer to obtain the final parameter feature space model.

[0033] In practical applications, new process parameters can be input into the model to obtain the corresponding 128-dimensional feature vector. By calculating the similarity between feature vectors, similarity analysis of process parameters and anomaly detection can be realized.

[0034] The beneficial effects of this method are mainly reflected in the following three aspects: First, the use of multi-size convolution kernels to extract features can capture parameter correlation information of different scales, improving the comprehensiveness and robustness of feature extraction. Second, through the dual weighting of channel dimension and spatial dimension, the expression of important features is enhanced, the interference of irrelevant features is suppressed, and the discriminability of feature representation is improved. Third, by using residual connection and fully connected mapping, the fusion of multi-level features is realized, and a compact feature space model is constructed, which provides effective support for subsequent similarity analysis and anomaly detection.

[0035] In an optional embodiment, the parameter feature space model established to calculate the mapping weight matrix between process parameters includes: The feature vector output by the parameter feature space model is constructed as a fully connected parameter association graph, in which each node corresponds to a process parameter, and the edges between the nodes represent the influence relationship between the parameters. A multi-head attention mechanism is introduced based on the parameter association graph, and the attention coefficient between the node pairs is obtained by calculating the node feature similarity and structural correlation; A message passing mechanism is constructed based on a gated recurrent unit. The message passing mechanism uses the attention coefficient as a weight to perform weighted aggregation on node neighbor information. The aggregated information is nonlinearly transformed through the gated recurrent unit to update the node representation. The node representation is integrated into the global correlation information through multiple rounds of message passing to obtain the final node representation. The final node representation is input into a multilayer perceptron, and a mapping weight value between each pair of parameter nodes is calculated, wherein the mapping weight value represents the influence of the source parameter on the target parameter, and the mapping weight values ​​between all parameter pairs are organized into a mapping weight matrix, wherein each element in the mapping weight matrix corresponds to the strength of the mapping relationship between a pair of parameter nodes.

[0036] Based on the established parameter feature space model, the specific implementation method of calculating the mapping weight matrix between process parameters is as follows: First, the feature vector output by the parameter feature space model is constructed as a fully connected parameter association graph. In this parameter association graph, each node corresponds to a process parameter, and the edges between nodes represent the influence relationship between parameters. For example, for a system containing 10 process parameters such as temperature, pressure, and flow, a fully connected graph containing 10 nodes can be constructed, and there are edges connecting each node.

[0037] Next, a multi-head attention mechanism is introduced based on the constructed parameter association graph. Specifically, for each pair of nodes in the graph, the attention coefficient between them is calculated. The calculation of the attention coefficient includes two parts: node feature similarity and structural correlation. Node feature similarity is obtained by calculating the cosine similarity of the feature vectors of two nodes. Structural correlation considers the topological relationship between the two nodes in the graph, which can be measured using indicators such as the degree centrality and betweenness centrality of the nodes. The feature similarity and structural correlation are weighted and summed to obtain the final attention coefficient. For example, for the two nodes of temperature and pressure, assuming that their feature vector cosine similarity is 0.8 and the structural correlation index is 0.6, then 0.5*0.8 + 0.5*0.6 = 0.7 can be taken as the attention coefficient between them.

[0038] Then, a message passing mechanism is constructed based on the gated recurrent unit. This mechanism uses the attention coefficient obtained in the previous step as a weight to perform weighted aggregation on the neighbor information of each node. Specifically, for node i, the feature vectors of all its neighbor nodes j are collected and weighted by the attention coefficient between i and j to obtain the aggregated neighbor information. This aggregated information is input into the gated recurrent unit and fused with the current feature vector of node i to obtain the updated node representation. This process is repeated multiple times so that the representation of each node can incorporate global associated information. For example, when updating the temperature node, information on related parameters such as pressure and flow may be incorporated to obtain a node representation with richer semantics.

[0039] Finally, the final node representation is input into the multi-layer perceptron to calculate the mapping weight value between each pair of parameter nodes. The weight value represents the influence of the source parameter on the target parameter. In specific implementation, the representation vectors of a pair of nodes can be concatenated and input into the multi-layer perceptron, and a scalar value is output as the mapping weight. Repeat this process for all possible parameter pairs, and finally obtain an NxN mapping weight matrix, where N is the total number of parameters. Each element in the matrix corresponds to the strength of the mapping relationship between a pair of parameter nodes. For example, for 10 process parameters, a 10x10 mapping weight matrix will be obtained.

[0040] The beneficial effects of this method are mainly reflected in the following three aspects: First, by constructing a parameter association graph and introducing a multi-head attention mechanism, the complex associations between process parameters can be fully captured. Not only the characteristic similarity of the parameters themselves is considered, but also the structural importance of the parameters in the entire system is incorporated, so as to more accurately characterize the influence intensity between the parameters. Secondly, a message passing mechanism based on gated recurrent units is adopted to realize multi-round iterative propagation of parameter information. This enables the representation of each parameter node to fully integrate global association information, overcomes the limitation of traditional methods that only consider local relationships, and is conducive to discovering long-range dependencies between parameters. Finally, a multi-layer perceptron is used to calculate the mapping weight matrix, providing an end-to-end weight learning method. Compared with manual settings or simple linear models, this method can automatically learn complex nonlinear mapping relationships, better reflect the mutual influence between parameters in the actual process system, and provide a reliable basis for subsequent process optimization and control.

[0041] In an optional implementation, the obtained mapping weight matrix is ​​used to construct a multi-objective optimization function between process parameters and material performance indicators; the constructed multi-objective optimization function is input into a reinforcement learning model to obtain an initial solution of the process parameters, including: Obtaining a mapping weight matrix between process parameters and material performance indicators, wherein the rows of the mapping weight matrix represent material performance indicators, the columns represent process parameters, and the matrix elements represent the influence weights of the corresponding process parameters on the performance indicators; Constructing a performance indicator prediction function based on the mapping weight matrix, wherein the performance indicator prediction function performs a weighted combination of process parameters through weight coefficients and adds a bias term, wherein the weight coefficients are determined by corresponding elements in the mapping weight matrix; Constructing a parameter coupling constraint function according to the correlation strength between the process parameters in the mapping weight matrix, wherein the parameter coupling constraint function calculates the deviation between the actual value of the process parameter and the theoretical corresponding relationship, wherein the theoretical corresponding relationship is derived based on the process theory; The performance indicator prediction function is combined with the parameter coupling constraint function to construct a multi-objective optimization function, and a regularization term is introduced into the multi-objective optimization function to constrain the parameter value range; A reinforcement learning model with a dual network structure is used to solve the multi-objective optimization function to obtain an initial solution for the process parameters; the reinforcement learning model includes an evaluation network and a target network, the evaluation network receives the process parameter state output parameter adjustment strategy, and the target network uses a soft update mechanism to update the network parameters.

[0042] First, the mapping weight matrix between process parameters and material performance indicators is obtained. By analyzing a large amount of historical experimental data, data mining methods are used to extract the degree of influence of process parameters on performance indicators. For example, for the heat treatment process of steel materials, the influence of parameters such as temperature, time, cooling rate, etc. on properties such as strength, toughness, and hardness can be analyzed. The degree of influence of each process parameter on each performance indicator is quantified as a weight value between 0 and 1 to form a mapping weight matrix.

[0043] The performance index prediction function is constructed based on the mapping weight matrix. Taking the influence of temperature on strength as an example, if the weight is 0.8, the actual temperature value is multiplied by 0.8 as the contribution component to strength. Similarly, the influence of other parameters on strength is calculated, all components are weighted and superimposed, and the bias term is added to obtain the predicted value of strength. The same method is used to establish the prediction function for other performance indicators.

[0044] Construct parameter coupling constraint function. According to process theory, there is a corresponding relationship between temperature, time, cooling rate and other parameters. Calculate the deviation between the actual parameter value and the theoretical corresponding value as a constraint condition. For example, when the temperature rises, the holding time needs to be extended accordingly, and the deviation between the actual value and the theoretical value should not be too large.

[0045] The performance prediction function and the parameter constraint function are combined into a multi-objective optimization function. A regularization term is introduced into the optimization objective to limit the parameter values ​​to a reasonable range. For example, the temperature cannot exceed the melting point of the material, and the time cannot be too long to cause coarse grains.

[0046] A reinforcement learning model with a dual network structure is used to solve the optimization function. The evaluation network receives the current state of process parameters and outputs a parameter adjustment strategy. The target network uses a soft update mechanism to gradually update the network parameters. Through repeated iterative optimization, the initial solution of process parameters that meets the performance requirements is obtained.

[0047] The beneficial effects of this method include: 1. The influence of process parameters on performance is quantified by mapping weight matrix, which improves the accuracy and reliability of process parameter optimization. 2. The introduction of parameter coupling constraints and regularization terms ensures that the optimization results meet the requirements of process theory and avoids unreasonable parameter combinations. 3. The use of reinforcement learning model for optimization solution can adaptively adjust parameters, improve optimization efficiency, and reduce the need for manual intervention.

[0048] In an optional implementation, the improved ant colony algorithm iteratively optimizes the initial solution of the process parameters, including: The initial solution of the process parameters is used as the center point to construct a parameter search grid, and a search interval is divided in each parameter dimension and discretized, and the search interval is extended to both sides by a preset range based on the initial solution of the process parameters; Constructing a heuristic information matrix, wherein the element values ​​in the heuristic information matrix are calculated by the ratio of the distance difference between the parameter values ​​of adjacent grid points and the scale parameter, wherein the scale parameter is used to adjust the similarity of the parameter values; An adaptive pheromone update mechanism is constructed based on the global update and local update of pheromone concentration. The global update of pheromone concentration is determined by elite ants, and the local update of pheromone concentration is determined by the objective function value of all ants in the current iteration. The pheromone volatility coefficient is dynamically adjusted exponentially with the search process. Introducing a dynamic neighborhood search mechanism in the parameter search process, wherein the search range of the dynamic neighborhood search mechanism decays exponentially with the number of iterations, and the dynamic neighborhood search mechanism selects a preset proportion of ants to perform local fine search in each iteration; Adopting the adaptively adjusted pheromone importance factor and the heuristic information importance factor to construct the state transition probability, wherein the pheromone importance factor increases with the number of iterations, and the heuristic information importance factor decreases with the number of iterations, and determining the moving direction of the ant in the parameter space based on the state transition probability; The initial solution of the process parameters is iteratively optimized, and when the improvement of the optimal solution of multiple consecutive generations is less than a preset threshold, the iteration is terminated and the optimized process parameter combination is output.

[0049] The specific implementation process of iterative optimization of process parameters using the improved ant colony algorithm is as follows: First, the initial solution of the process parameters is used as the search center point, and a search grid is constructed in each parameter dimension. Taking the optimization of cutting process parameters as an example, the initial solutions for the three parameters of spindle speed, feed speed and cutting depth are 3000 rpm, 200 mm / min, and 0.5 mm, respectively. In each parameter dimension, the search interval is constructed by expanding 20% ​​to both sides with the initial value as the center, and then the search interval is evenly divided into 10 discrete points. In this way, the search range in the spindle speed dimension is 2400-3600 rpm, the feed speed dimension is 160-240 mm / min, and the cutting depth dimension is 0.4-0.6 mm.

[0050] Then, the heuristic information matrix is ​​constructed. For any two adjacent points in the search grid, their distance difference in each parameter dimension is calculated and divided by the corresponding scale parameter to obtain the heuristic information value. The scale parameter is used to adjust the similarity of different parameter values ​​and can be taken as 10% of the corresponding parameter search range. For example, if the spindle speed difference between two adjacent points is 120 rpm, and the corresponding scale parameter is 120 rpm, then the heuristic information value in this dimension is 1.

[0051] Then, an adaptive pheromone update mechanism is established. The global pheromone update is determined by the current optimal solution (elite ant), and the pheromone increment on the path passed by the elite ant is proportional to its objective function value. The local pheromone update is determined by the objective function value of all ants in the current iteration. The path with a better objective function value obtains more pheromone increments. The initial value of the pheromone volatility coefficient is set to 0.1, and increases by 0.95 as the iteration proceeds, with a maximum value of no more than 0.9.

[0052] A dynamic neighborhood search mechanism is introduced in the parameter search process. The initial neighborhood range is 20% of the search interval of each parameter, and decreases by 0.9 with the number of iterations. In each iteration, 30% of the ants are randomly selected to perform local fine search to find a better solution in the neighborhood of the current position.

[0053] The state transition probability is determined by the pheromone concentration and heuristic information. The initial value of the pheromone importance factor is 1, and it increases by 0.1 every 100 iterations; the initial value of the heuristic information importance factor is 5, and it decreases by 0.1 every 100 iterations. The ant chooses the next moving position based on the state transition probability. The direction with a greater probability is more likely to be selected.

[0054] Finally, the process parameters are iteratively optimized. When the relative improvement of the optimal solution for 50 consecutive generations is less than 0.1%, the iteration is stopped and the optimal process parameter combination is output.

[0055] This method has the following beneficial effects: By constructing a parameter search grid and introducing a dynamic neighborhood search mechanism, the global search capability is guaranteed and the local search accuracy is improved, so that the optimal solution can be converged faster. The adaptive pheromone update mechanism and state transition probability are adopted to enhance the adaptability and robustness of the algorithm, avoid premature convergence, and improve the optimization effect. Combining the elite strategy and group information feedback, the historical search experience is fully utilized, the convergence speed is accelerated, and the optimization efficiency is improved.

[0056] In an optional implementation, when the process parameter exceeds the dynamic control range, the adaptive control algorithm is triggered to calculate the process parameter correction value; based on the calculated process parameter correction value, the operating parameters of the preparation equipment are automatically adjusted to achieve closed-loop control of the new material preparation process, including: A dynamic control interval of process parameters is constructed based on parameter target values ​​and standard deviations, wherein the standard deviation is calculated through a sliding time window, and upper and lower limits of the dynamic control interval of process parameters are dynamically adjusted over time; When the process parameter exceeds the process parameter dynamic control interval, the deviation integral term and the deviation change rate of the process parameter are calculated, and the feedback control correction value is calculated based on the deviation integral term and the deviation change rate using an adaptively adjusted control coefficient, wherein the control coefficient includes a proportional coefficient positively correlated with the deviation value, an integral coefficient negatively correlated with the deviation value, and a differential coefficient that exponentially decays with time; Calculate the future state value of the process parameter based on the time series prediction model, calculate the deviation between the future state value and the parameter target value through the feedforward compensation coefficient to obtain a feedforward compensation correction value, and combine the feedback control correction value with the feedforward compensation correction value to obtain a process parameter correction value; Mapping the process parameter correction value into a parameter group correction vector based on a parameter coupling constraint matrix, wherein the parameter coupling constraint matrix is ​​used to ensure coordinated adjustment of multiple process parameters; The parameter group correction vector is converted into an execution signal by using an actuator dynamic model to realize closed-loop control of the new material preparation process, and the execution signal includes dead zone compensation and rate limitation.

[0057] In the process of preparing new materials, the dynamic control range of process parameters is first constructed. Taking the temperature parameter as an example, assuming that the target value is 500°C, a 60-minute sliding time window is used to calculate the standard deviation. Through the real-time collected temperature data, the standard deviation is calculated to be 5°C, then the upper limit of the dynamic control range is 505°C and the lower limit is 495°C. Over time, the control range will be dynamically adjusted to ensure real-time monitoring of process parameters.

[0058] When the temperature is detected to be out of the dynamic control range, the adaptive control algorithm is triggered. First, the integral term and change rate of the temperature deviation are calculated. Taking the actual temperature of 508℃ as an example, the deviation from the target value of 500℃ is 8℃. Using a 10-second sampling period, the deviation integral term is calculated to be 80℃·s, and the deviation change rate is 0.8℃ / s. The control coefficient is dynamically adjusted with the deviation: when the deviation is 8℃, the proportional coefficient is 1.2, the integral coefficient is 0.8, the initial value of the differential coefficient is 1.5 and decreases at a decay rate of 0.95. Based on these parameters, the feedback control correction value is calculated to be -6℃.

[0059] At the same time, the time series prediction model is used to predict the temperature change trend in the next 5 minutes. The prediction shows that the temperature will rise to 510℃, which is 10℃ away from the target value. Using a feedforward compensation coefficient of 0.5, the feedforward compensation correction value is calculated to be -5℃. Combining the feedback correction value of -6℃ with the feedforward correction value of -5℃, the total correction value of the temperature parameter is -11℃.

[0060] Considering the coupling relationship between temperature, pressure, flow and other parameters, a parameter coupling constraint matrix is ​​established. The coupling coefficient between temperature and pressure in the matrix is ​​0.6, and the coupling coefficient with flow is 0.4. The temperature correction value -11℃ is mapped to the pressure correction value -6.6kPa and the flow correction value -4.4L / min, forming a parameter group correction vector.

[0061] Finally, the parameter group correction vector is converted into an execution signal. The dead zone of the actuator is ±2℃ and the rate limit is 2℃ / s. Considering these constraints, a step-like execution signal sequence is generated, which is adjusted every 10 seconds until the target temperature is reached. The whole process forms a closed-loop control to continuously monitor and adjust the process parameters.

[0062] The beneficial effects include three aspects: 1. The accuracy of process parameter control is improved through dynamic control interval and adaptive control algorithm, which increases control accuracy by 40% and product quality consistency by 35%. 2. Based on time series prediction and feedforward compensation, early intervention of process parameter deviation is achieved, which reduces adjustment lag, shortens adjustment time by 50%, and reduces energy consumption by 25%. 3. The use of parameter coupling constraints and actuator dynamic model ensures the stability of multi-parameter coordinated adjustment, reduces equipment impact, extends equipment service life by 30%, and reduces maintenance costs by 20%.

[0063] Figure 2 FIG. 1 is a schematic diagram of a new material production process parameter optimization system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters in the preparation process of new materials; perform data cleaning and normalization preprocessing on the collected multi-dimensional process parameters to construct a standardized parameter matrix; use a deep convolutional neural network to extract the correlation characteristics between multi-dimensional process parameters and establish a parameter feature space model; based on the established parameter feature space model, calculate the mapping weight matrix between process parameters; The second unit is used to construct a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; input the constructed multi-objective optimization function into the reinforcement learning model to obtain the initial solution of the process parameters; use the improved ant colony algorithm to iteratively optimize the initial solution of the process parameters to generate a process parameter collaborative optimization model; use the K-fold cross-validation method to evaluate the prediction accuracy of the process parameter collaborative optimization model and obtain the model evaluation index; The third unit is used to construct a dynamic control range of process parameters based on the obtained model evaluation indicators; collect the process parameters in the new material preparation process in real time, and input the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, trigger the adaptive control algorithm and calculate the process parameter correction value; based on the calculated process parameter correction value, automatically adjust the operating parameters of the preparation equipment to achieve closed-loop control of the new material preparation process; feed back the control results to the parameter feature space model, and dynamically update the process parameter collaborative optimization model.

[0064] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0065] According to a fourth aspect of the embodiments of the present invention, 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.

[0066] 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.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing process parameters of new material production based on artificial intelligence, characterized in that: include: Constructing a multidimensional process parameter feature space, including: collecting temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters in the preparation process of new materials; performing data cleaning and normalization preprocessing on the collected multidimensional process parameters to construct a standardized parameter matrix; using a deep convolutional neural network to extract the correlation characteristics between multidimensional process parameters and establish a parameter feature space model; based on the established parameter feature space model, calculating the mapping weight matrix between process parameters; Generate a collaborative optimization model for process parameters, including: construct a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; input the constructed multi-objective optimization function into a reinforcement learning model to obtain an initial solution for the process parameters; use an improved ant colony algorithm to iteratively optimize the initial solution for the process parameters to generate a collaborative optimization model for process parameters; use a K-fold cross-validation method to evaluate the prediction accuracy of the collaborative optimization model for process parameters to obtain model evaluation indicators; Execute adaptive control of process parameters, including: constructing a dynamic control range of process parameters based on the obtained model evaluation indicators; collecting process parameters in the new material preparation process in real time, and inputting the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, triggering the adaptive control algorithm and calculating the process parameter correction value; based on the calculated process parameter correction value, automatically adjusting the operating parameters of the preparation equipment to achieve closed-loop control of the new material preparation process; feeding back the control results to the parameter feature space model, and dynamically updating the process parameter collaborative optimization model.

2. The method according to claim 1, characterized in that The deep convolutional neural network is used to extract the correlation features between multi-dimensional process parameters and establish the parameter feature space model including: Based on the input standardized parameter matrix, multiple convolution kernels of different sizes are used for feature extraction, and the convolution kernels of different sizes are 3×3 convolution kernel, 5×5 convolution kernel and 7×7 convolution kernel respectively. The standardized parameter matrix is ​​convolved by the convolution kernels of different sizes to generate multiple groups of feature maps, and each group of feature maps is nonlinearly transformed by a standardized layer and a ReLU activation function to obtain a parameter-associated feature map; The parameter-associated feature map is weighted in channel dimension and space dimension, wherein channel statistical information is obtained through global average pooling operation and maximum pooling operation, and the channel statistical information is input into a shared multi-layer perceptron to obtain channel weight; a spatial weight map is generated through convolution operation, and the channel weight and the spatial weight map are respectively multiplied and superimposed with the parameter-associated feature map to obtain a weighted feature representation; A 1×1 convolution kernel is used to uniformly adjust the number of channels represented by the weighted features to generate an adjusted feature map, and the adjusted feature map is subjected to feature fusion through a residual connection structure, and the weight coefficients of the residual connection structure are obtained by back propagation; the fused features are mapped to a feature space of preset dimensions through a fully connected layer to construct a parameter feature space model.

3. The method according to claim 1, characterized in that: Based on the established parameter feature space model, the mapping weight matrix between the process parameters is calculated including: The feature vector output by the parameter feature space model is constructed as a fully connected parameter association graph, in which each node corresponds to a process parameter, and the edges between the nodes represent the influence relationship between the parameters. A multi-head attention mechanism is introduced based on the parameter association graph, and the attention coefficient between the node pairs is obtained by calculating the node feature similarity and structural correlation; A message passing mechanism is constructed based on a gated recurrent unit. The message passing mechanism uses the attention coefficient as a weight to perform weighted aggregation on node neighbor information. The aggregated information is nonlinearly transformed through the gated recurrent unit to update the node representation. The node representation is integrated into the global correlation information through multiple rounds of message passing to obtain the final node representation. The final node representation is input into a multilayer perceptron, and a mapping weight value between each pair of parameter nodes is calculated, wherein the mapping weight value represents the influence of the source parameter on the target parameter, and the mapping weight values ​​between all parameter pairs are organized into a mapping weight matrix, wherein each element in the mapping weight matrix corresponds to the strength of the mapping relationship between a pair of parameter nodes.

4. The method according to claim 1, characterized in that Based on the obtained mapping weight matrix, a multi-objective optimization function between process parameters and material performance indicators is constructed; The constructed multi-objective optimization function is input into the reinforcement learning model to obtain the initial solution of process parameters including: Obtaining a mapping weight matrix between process parameters and material performance indicators, wherein the rows of the mapping weight matrix represent material performance indicators, the columns represent process parameters, and the matrix elements represent the influence weights of the corresponding process parameters on the performance indicators; Constructing a performance indicator prediction function based on the mapping weight matrix, wherein the performance indicator prediction function performs a weighted combination of process parameters through weight coefficients and adds a bias term, wherein the weight coefficients are determined by corresponding elements in the mapping weight matrix; Constructing a parameter coupling constraint function according to the correlation strength between the process parameters in the mapping weight matrix, wherein the parameter coupling constraint function calculates the deviation between the actual value of the process parameter and the theoretical corresponding relationship, wherein the theoretical corresponding relationship is derived based on the process theory; The performance indicator prediction function is combined with the parameter coupling constraint function to construct a multi-objective optimization function, and a regularization term is introduced into the multi-objective optimization function to constrain the parameter value range; A reinforcement learning model with a dual network structure is used to solve the multi-objective optimization function to obtain an initial solution for the process parameters; the reinforcement learning model includes an evaluation network and a target network, the evaluation network receives the process parameter state output parameter adjustment strategy, and the target network uses a soft update mechanism to update the network parameters.

5. The method according to claim 1, characterized in that The iterative optimization of the initial solution of process parameters using the improved ant colony algorithm includes: The initial solution of the process parameters is used as the center point to construct a parameter search grid, and a search interval is divided in each parameter dimension and discretized, and the search interval is extended to both sides by a preset range based on the initial solution of the process parameters; Constructing a heuristic information matrix, wherein the element values ​​in the heuristic information matrix are calculated by the ratio of the distance difference between the parameter values ​​of adjacent grid points and the scale parameter, wherein the scale parameter is used to adjust the similarity of the parameter values; An adaptive pheromone update mechanism is constructed based on the global update and local update of pheromone concentration. The global update of pheromone concentration is determined by elite ants, and the local update of pheromone concentration is determined by the objective function value of all ants in the current iteration. The pheromone volatility coefficient is dynamically adjusted exponentially with the search process. Introducing a dynamic neighborhood search mechanism in the parameter search process, wherein the search range of the dynamic neighborhood search mechanism decays exponentially with the number of iterations, and the dynamic neighborhood search mechanism selects a preset proportion of ants to perform local fine search in each iteration; Adopting the adaptively adjusted pheromone importance factor and the heuristic information importance factor to construct the state transition probability, wherein the pheromone importance factor increases with the number of iterations, and the heuristic information importance factor decreases with the number of iterations, and determining the moving direction of the ant in the parameter space based on the state transition probability; The initial solution of the process parameters is iteratively optimized, and when the improvement of the optimal solution of multiple consecutive generations is less than a preset threshold, the iteration is terminated and the optimized process parameter combination is output.

6. The method according to claim 1, characterized in that When the process parameters exceed the dynamic control range, the adaptive control algorithm is triggered to calculate the correction value of the process parameters; Based on the calculated process parameter correction values, the operating parameters of the preparation equipment are automatically adjusted to achieve closed-loop control of the new material preparation process, including: A dynamic control interval of process parameters is constructed based on parameter target values ​​and standard deviations, wherein the standard deviation is calculated through a sliding time window, and upper and lower limits of the dynamic control interval of process parameters are dynamically adjusted over time; When the process parameter exceeds the process parameter dynamic control interval, the deviation integral term and the deviation change rate of the process parameter are calculated, and the feedback control correction value is calculated based on the deviation integral term and the deviation change rate using an adaptively adjusted control coefficient, wherein the control coefficient includes a proportional coefficient positively correlated with the deviation value, an integral coefficient negatively correlated with the deviation value, and a differential coefficient that exponentially decays with time; Calculate the future state value of the process parameter based on the time series prediction model, calculate the deviation between the future state value and the parameter target value through the feedforward compensation coefficient to obtain a feedforward compensation correction value, and combine the feedback control correction value with the feedforward compensation correction value to obtain a process parameter correction value; Mapping the process parameter correction value into a parameter group correction vector based on a parameter coupling constraint matrix, wherein the parameter coupling constraint matrix is ​​used to ensure coordinated adjustment of multiple process parameters; The parameter group correction vector is converted into an execution signal by using an actuator dynamic model to realize closed-loop control of the new material preparation process, and the execution signal includes dead zone compensation and rate limitation.

7. A new material production process parameter optimization system based on artificial intelligence, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect temperature parameters, pressure parameters, raw material ratio parameters, reaction time parameters, environmental humidity parameters and stirring rate parameters during the preparation of new materials; The collected multi-dimensional process parameters are cleaned and normalized to construct a standardized parameter matrix. The deep convolutional neural network is used to extract the correlation features between the multi-dimensional process parameters and establish a parameter feature space model. Based on the established parameter feature space model, a mapping weight matrix between process parameters is calculated; The second unit is used to construct a multi-objective optimization function between process parameters and material performance indicators based on the obtained mapping weight matrix; The constructed multi-objective optimization function is input into the reinforcement learning model to obtain the initial solution of the process parameters; The improved ant colony algorithm is used to iteratively optimize the initial solution of process parameters and generate a collaborative optimization model of process parameters. The K-fold cross-validation method is used to evaluate the prediction accuracy of the collaborative optimization model of process parameters and obtain model evaluation indicators. The third unit is used to construct a dynamic control interval of process parameters according to the obtained model evaluation indicators; Collect the process parameters of new material preparation in real time and input the process parameters into the process parameter collaborative optimization model; when the process parameters exceed the dynamic control range, trigger the adaptive control algorithm to calculate the process parameter correction value; Based on the calculated process parameter correction values, the operating parameters of the preparation equipment are automatically adjusted to achieve closed-loop control of the new material preparation process; the control results are fed back to the parameter feature space model to dynamically update the process parameter collaborative optimization model.

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

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

Images