Model training method, flow chart-based parameter prediction method and device

Abstract

The invention relates to a model training method and a parameter prediction method and device based on a flow chart. The method comprises the steps of obtaining key configuration data of a current process in a process flow chart; a flow chart data set is generated according to the key configuration data, the flow chart data set serves as a training sample, and the flow chart data set comprises flow convergence state information; and constructing a network model based on a variational auto-encoder of any condition, training the network model by using the training sample, and constructing a flow chart network model adaptive to the process flow chart. Performing missing value prediction on the key configuration data by using the flow chart network model, and performing parameter adjustment on the key configuration data according to the predicted missing value; and inputting the adjusted configuration data into the process flow diagram to carry out convergence prediction. Therefore, the problem of non-convergence of the optimized chemical process flow diagram is solved, the convergence probability of the flow diagram is improved, and meanwhile, the simulation stability and efficiency are improved.

Description

Technical Field

[0001] This application mainly relates to the field of process flow simulation calculation, and particularly relates to a method for model training, a method and device for parameter prediction based on a flow chart. Background Art

[0002] In the production processes of enterprises in fields such as oil refining, chemical industry, safety, and energy, the real-time operation of process flow charts requires accuracy, stability, and speed. In the actual process of chemical process simulation calculation, the convergence problem often involves iterative unit operations, cyclic processes, design specifications, and optimization steps. Configuring appropriate iterative algorithm parameters and accurate initial variable values is the key. If these parameters are set inappropriately, it may lead to the inability of the simulation process to converge, reduce the construction efficiency of the process model, and affect the stability of the flow chart operation. The existing method is to set parameters manually. However, due to the complexity of the chemical process, configuring the flow chart and managing convergence usually requires professional knowledge and a large amount of time, relying on the experience of operators, and the efficiency is low. Especially when dealing with complex processes with multi-layer convergence calculations, the convergence behavior is often opaque, making it more difficult and complex to reach the convergence state. Summary of the Invention

[0003] An object of this application is to provide a method for model training, a method and device for parameter prediction based on a flow chart, so as to solve the problems of non-convergence of the process flow chart, low efficiency, and complex operation caused by manual configuration of flow chart parameters in the prior art.

[0004] According to one aspect of this application, a method for training a flow chart network model is provided, which is applied to a process flow. The method includes:

[0005] Obtain the key configuration data of the current process in the process flow chart;

[0006] Generate a flow chart data set according to the key configuration data, and use the flow chart data set as a training sample, where the flow chart data set includes process convergence state information;

[0007] Construct a network model based on a variational autoencoder under arbitrary conditions, and use the training sample to train the network model to construct a flow chart network model adapted to the flow chart for parameter prediction of the process flow chart.

[0008] Optionally, obtaining the key configuration data of the current process in the process flow chart includes:

[0009] Judge whether the convergence state of the current process in the process flow chart is a failure. If so, determine the key parameters that cause the convergence failure, use the key parameters as the key configuration data of the current process, and set corresponding parameter step sizes and value ranges for the key configuration data.

[0010] Optionally, generate a flowchart data set according to the key configuration data, including:

[0011] Sample the key configuration data according to the parameter step size and value range to generate an input data combination;

[0012] Input the input data combination into the process flowchart for calculation to obtain a flowchart data set.

[0013] Optionally, input the input data combination into the process diagram for calculation to obtain a flowchart data set, including:

[0014] Initialization data of the current process obtained from the database;

[0015] Input the input data combination into the process flowchart to replace the parameters at the corresponding positions, and calculate the convergence state information of the input data combination in combination with the initialization data;

[0016] Create a flowchart data set based on the input data combination and the corresponding convergence state information.

[0017] Optionally, construct a network model based on a variational autoencoder with arbitrary conditions, including:

[0018] Build a prior network in the network model based on a variational autoencoder with arbitrary conditions, where the prior network uses a residual network as a hidden layer, and the prior network is used to receive observed features and missing value masks to learn to generate latent variables;

[0019] Build a proposal network in the network model, where the proposal network uses a residual network as a hidden layer, and the proposal network is used to receive all features and output proposals for latent variables;

[0020] Build a generation network in the network model to decode missing values from latent variables and multi-scale features through long skip connections;

[0021] Set a loss function, introduce the relative entropy divergence between the latent variable distributions between the prior network and the proposal network, and construct an optimizer;

[0022] Add a dropout layer in each residual network block of the generation network to prevent overfitting.

[0023] Optionally, the probability of dropout is dynamically allocated in each residual network block and satisfies the following formula:

[0024]

[0025] Wherein, i is the index of the residual network block, P1 represents the dropout probability of the residual network block that only receives one long skip connection, Depth is the number of hidden layers, and N isc,i is the number of long skip connections received by the residual network block.

[0026] Optionally, training the network model using the training samples includes:

[0027] Dividing the training samples into a training set and a validation set, and inputting the training set into the network model;

[0028] Evaluating the metrics of the validation set, and selecting the optimal network model as the flowchart network model.

[0029] Optionally, training the network model using the training samples includes:

[0030] Creating masked data for missing values of the features in the training samples based on the probability generation mechanism of the component Bernoulli distribution;

[0031] Inputting the training samples and the masked data into the network model to train the network model.

[0032] Optionally, after training includes:

[0033] Evaluating the trained flowchart network model to determine the proportion of successfully converged cases in the originally non-converging cases.

[0034] According to another aspect of the present application, a method for parameter prediction based on a flowchart is provided, and the method includes:

[0035] Obtaining the key configuration data of the current process in the process flowchart;

[0036] Using the flowchart network model to predict missing values of the key configuration data, and adjusting the parameters of the key configuration data according to the predicted missing values, wherein the network model is constructed based on a variational autoencoder under arbitrary conditions and is obtained by training the network model using training samples, and the training samples are determined from a flowchart dataset generated according to the key configuration data;

[0037] Inputting the adjusted configuration data into the process flowchart for convergence prediction.

[0038] Optionally, using the flowchart network model to predict missing values of the key configuration data includes:

[0039] Obtaining the target parameters in the process flowchart selected by the user, and adding a mask to the target parameters;

[0040] Replace the target parameter with a missing value placeholder, where the missing value placeholder is used to simulate a missing state;

[0041] Use the flowchart network model to predict the missing value for the missing value placeholder under the mask.

[0042] According to another aspect of the present application, there is also provided a flowchart network model training device, which includes:

[0043] One or more processors; and

[0044] A memory storing computer-readable instructions, and when the computer-readable instructions are executed, the processor executes the operations of the aforementioned method.

[0045] According to yet another aspect of the present application, there is also provided a device for parameter prediction based on a flowchart, which includes:

[0046] One or more processors; and

[0047] A memory storing computer-readable instructions, and when the computer-readable instructions are executed, the processor executes the operations of the aforementioned method.

[0048] According to still another aspect of the present application, there is also provided a computer-readable medium, on which computer instructions are stored, and the computer-readable instructions can be executed by a processor to implement the aforementioned method.

[0049] Compared with the prior art, the present application obtains the key configuration data of the current process in the process flowchart; generates a flowchart data set according to the key configuration data, and uses the flowchart data set as a training sample, where the flowchart data set includes process convergence state information; constructs a network model based on a variational autoencoder under any conditions, uses the training sample to train the network model, and constructs a flowchart network model adapted to the process flowchart for parameter prediction of the process flowchart. Use the flowchart network model to predict the missing value of the key configuration data, and adjust the parameters of the key configuration data according to the predicted missing value; input the adjusted configuration data into the process flowchart for convergence prediction, thereby solving the non-convergence problem of optimizing the chemical process flowchart, improving the convergence probability of the process flowchart, and at the same time improving the stability and efficiency of the simulation. Description of the Drawings

[0050] To make the above objects, features, and advantages of the present application more obvious and understandable, the following detailed description of the specific embodiments of the present application is provided in conjunction with the accompanying drawings, where:

[0051] Figure 1Shows a schematic flow chart of a method for training a flowchart network model provided according to an aspect of the present application;

[0052] Figure 2 Shows a method flow chart of parameter prediction based on a flowchart provided according to an aspect of the present application;

[0053] Figure 3 Shows a performance graph of a decreasing arbitrary conditional variational autoencoder model used in an embodiment of the present application;

[0054] Figure 4 Shows a schematic flow chart of a method for excluding and predicting incorrect configurations of a complex chemical process flowchart based on an arbitrary conditional variational autoencoder in an embodiment of the present application;

[0055] Figure 5 Shows a schematic framework diagram of a device provided according to another aspect of the present application.

[0056] Like reference numerals in the drawings represent like or similar components. Detailed Embodiments

[0057] To make the above objects, features, and advantages of the present application more obvious and understandable, the following provides a detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings.

[0058] Many specific details are set forth in the following description to facilitate a thorough understanding of the present application, but the present application may also be implemented in other ways different from those described herein. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0059] As shown in the present application and the claims, unless the context clearly indicates an exception, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "including" and "comprising" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0060] The present application provides a method for training a flowchart network model adapted to a flowchart, and a method for parameter prediction based on a flowchart, specifically targeting the convergence problems caused by improper settings of iterative algorithm parameters and initial values of iterative variables, which can greatly reduce the time and effort for manual inspection and exclusion of parameters, provide diverse predicted values for the simulation results of chemical process flowcharts, improve the construction efficiency of online process models, and ensure the stability of online operation flowcharts. The specific solutions are as follows:

[0061] Figure 1A schematic flowchart showing a method for training a flowchart network model provided according to an aspect of the present application, applied to a process flow. The method includes steps S11 to S13.

[0062] Step S11: Obtain the key configuration data of the current process in the process flowchart.

[0063] Determine the process flowchart for which convergence prediction is required. According to the execution order of the processes in the process flowchart, obtain the configuration data of the processes that already exist in the flowchart. Select some parameters that affect the convergence of the flowchart from the obtained configuration data as the key configuration data. The key configuration data is the configuration data that affects the convergence of the process flowchart. Subsequently, use the selected key configuration data to generate a relevant data set.

[0064] Step S12: Generate a flowchart data set based on the key configuration data, and use the flowchart data set as a training sample. Among them, the flowchart data set includes process convergence status information.

[0065] Sample the key configuration data to generate different groups of data combinations. Each group of data combinations includes parameters of different dimensions. Use the generated groups of data combinations to determine the flowchart data set. Among them, the flowchart data set is a data set composed of input data combinations and data containing process convergence status labels formed through process calculations. This data set is associated with process convergence. Use the data set associated with convergence as a training sample for subsequent model training, and then perform convergence prediction.

[0066] Step S13: Construct a network model based on a variational autoencoder under arbitrary conditions. Use the training sample to train the network model to construct a flowchart network model adapted to the process flowchart for parameter prediction of the process flowchart.

[0067] The variational autoencoder under arbitrary conditions is an improvement of the variational autoencoder, including a prior network, a proposal network, and a generation network. Based on the variational autoencoder under arbitrary conditions, further optimize and improve to obtain a network model. Thus, use the obtained flowchart data set as a training sample to train the network model, and then train a flowchart network model adapted to the process flowchart for parameter prediction and convergence evaluation of the process flowchart.

[0068] In an embodiment of the present application, in step S11, determine whether the convergence status of the current process in the process flowchart is a failure. If so, determine the key parameters that cause the convergence failure, use the key parameters as the key configuration data of the current process, and set corresponding parameter step sizes and value ranges for the key configuration data.

[0069] The configuration data can be diagnostically screened to identify the key parameters affecting convergence as the required key configuration data. The model training end directly uses the diagnostically processed key configuration data and then processes the key configuration data using the parameter step size and value range.

[0070] Specifically, after the process flow diagram is built and the calculation is completed, the user can diagnose the convergence of the process. For example, after the process calculation is completed, if the user finds that the process cannot converge, they can choose to trigger the convergence diagnosis, analyze the parameters that may cause the process not to converge, and select these key parameters. Then, the corresponding step size and value range are configured for these key parameters, so that a better sampling data set can be obtained by setting the step size and value range later. The model training end can also receive a request from the user end. This request is set to carry the unique identification information (ID) of the process, and the selected parameter content and parameters are passed through the request. The data structure for passing is as follows:

[0072] {path: "Apath", value: [Alowerbound, Aupperbound, Astep]},

[0073] {path: "Bpath", value: [Blowerbound, Bupperbound, Bstep]},

[0075] [[ID=~16]]Among them, Apath is the path in the variable manager corresponding to the selected parameter, Alowerbound is the lower bound of the parameter set by the user, Aupperbound1 is the upper bound of the parameter set by the user, Astep is the step size of the parameter set by the user. The variable manager is a storage system that contains all process flows and their parameters. In this variable manager, all parameters have a unique path for storing and accessing these parameters.

[0076] In some embodiments of the present application, in step S12, the key configuration data is sampled according to the parameter step size and value range to generate an input data combination; the input data combination is input into the process flow diagram for calculation to obtain a flowchart data set.

[0077] The Cartesian product method can be used to sample the key configuration data of the current process to generate a diagnostic parameter sampling list and obtain different multiple groups of input data combinations. The obtained key configuration data includes configuration data that affects the calculation convergence, configurations that directly affect the process calculation results, and the convergence states of the entire process, each unit operation, and the loop circuit. The diagnostic parameter sampling list generated by the Cartesian product method is used as the input data, that is, an input data combination is generated, satisfying the following formula:​​

[0078] X = A1 × A2 × … × A n ;

[0079] where A1, A2, …, A n are the respective dimensions of the input data set, X represents all possible parameter combinations, and × represents the Cartesian product operation.

[0080] Multiple processes can be used to group and process multiple sets of input data combinations. Each process fills a set of input data combinations into the process flow chart, generates multiple process tasks, and calls the initial process flow for calculation, thereby obtaining the corresponding flow chart data set.

[0081] Continuing from the above embodiment, the initialization data of the current process obtained from the database; the input data combination is input into the process flow chart to replace the parameters at the corresponding positions, and the convergence state information of the input data combination is calculated in combination with the initialization data; a flow chart data set is created based on the input data combination and the corresponding convergence state information.

[0082] Use the python library and methods to obtain the initialization data of the current process in the database. This initialization data includes chemical components, petroleum analysis data selection, physical property method selection, and flash algorithm selection; then, implement multi-process parallel computing, and use the multi-process parallel computing method to group and process the input data combinations. Each process fills a set of input data combinations into the flow chart built by the user, calls the initial industrial process, replaces the sampled input data combinations into the corresponding parameter positions in the process respectively, generates multiple process tasks for parallel computing, records the calculation results and convergence status of each set of parameters during the calculation process, obtains the convergence status of the entire process, each unit operation, and the loop circuit, and creates a comprehensive data set containing various convergence success and failure data. Specifically, in the sampling stage, set the sampling range for the selected parameters. For example, set the reflux ratio sampling range of the distillation column to 1 - 2, and sample at a step size of 0.2. Another example is that the total sampling range at the top of the distillation column is set to 100 - 200, and the step size is 20; replace these sampling values into the corresponding parameter positions in the current process configuration in sequence. For example, replace 1, 1.2, 1.4, 1.6, 1.8, 2 into the current process configuration data respectively, and allow process simulation calculation.

[0083] In some embodiments of the present application, in step S13, a prior network in the network model of the variational autoencoder based on arbitrary conditions is constructed. Among them, the prior network uses a residual network as a hidden layer. The prior network is used to receive the observed features and the missing value mask to learn to generate latent variables; a proposal network in the network model is constructed. Among them, the proposal network uses a residual network as a hidden layer. The proposal network is used to receive all features and output a proposal of the latent variable; a generation network in the network model is constructed to decode the missing value from the latent variable and multi-scale features through long skip connections; a loss function is set, and the relative entropy divergence between the latent variable distributions is introduced between the prior network and the proposal network, and an optimizer is constructed; a dropout layer is added to each residual network block in the generation network to prevent overfitting. Among them, the probability of the dropout is dynamically allocated in each residual network block and satisfies the following formula:

[0084]

[0085] where i is the index of the residual network block, P1 represents the probability of dropout of the residual network block that only receives one long skip connection, Depth is the number of hidden layers, and N isc,i is the number of long skip connections received by the residual network block.

[0086] A network model is constructed based on the variational autoencoder under arbitrary conditions. The constructed network model includes a prior network, a proposal network, a generation network, and an optimizer. Among them, the prior network uses a residual network as a hidden layer. The prior network is used to receive the observed features and the missing value mask, and replace the missing features and abnormal features with NaN to learn to generate latent variables. Among them, NaN serves as a missing value placeholder to simulate the missing state and is used to mark the data that needs to be imputed; a proposal network in the network model is constructed. Among them, the proposal network uses a residual network as a hidden layer. The proposal network is used to receive all features and output a proposal of the latent variable; a generation network in the network model is constructed to decode the missing value from the latent variable and multi-scale features through long skip connections; a loss function is set, and the relative entropy divergence between the latent variable distributions is introduced between the prior network and the proposal network, and an optimizer is constructed; a dropout layer is added to each residual network block in the generation network to prevent overfitting.

[0087] Only the observed features and masks are input into the prior network to learn the ability to generate reasonable latent variables; the generative network decodes the missing values from the latent variables and multi-scale features through long skip connections, and among the three structures of "dense", "decreasing" and "increasing" of the long skip connections, the decreasing long skip connection is experimentally determined to be the best model for the flowchart convergence dataset, and the accuracy of feature imputation is improved by integrating multi-scale information; among them, the multi-scale features include different parameter information of the input data dimension, high-level feature representations of the latent variables, and local and global features extracted from different levels of the neural network, etc. By introducing the relative entropy (KL) divergence between the latent variable distributions of the proposal network and the prior network into the loss function, the latent vectors of the prior network gradually contain the information of the missing features. The training of the model aims to maximize the variational lower bound (VLB), and combines the reconstruction error, negative KL divergence and regularization term for optimization.

[0088] Specifically, the prior network: is used to learn the distribution of latent variables, only receives the observed features and masks, and replaces the missing features and abnormal features with NaN; through the prior network, the model can learn how to generate reasonable latent variables. The proposal network: is used to generate proposals for latent variables, receives all features and does not discard any masked features; the network objective is to assist the generative network in inferring missing data by outputting proposals for latent variables. The generative network: adopts a decreasing long skip connection structure to decode the missing values from the latent variables and multi-scale features; through the decreasing long skip connection structure, the generative network can better recover the missing feature information. An optimizer is constructed between the proposal network and the prior network, and through the optimization objective design, the KL divergence between the latent variable distributions is introduced to achieve the reasonable generation and learning of latent variables. The loss function is set, and the training objective of the model is to maximize the variational lower bound. For this purpose, the loss function includes the following three parts: the reconstruction error, which measures the difference between the data generated by the model and the original data; the negative KL divergence, which ensures that the latent variable distributions of the prior network and the proposal network are close; the regularization term, which reduces the overfitting of the model on the training data and improves its generalization ability. To improve the robustness of the model, a dropout layer is added to each residual network block (ResNet) of the generative network. In the variational autoencoder framework under any conditions, all hidden layers adopt residual networks with a depth of 10 and a width of 256, and the latent space of each network is represented by a 64-dimensional latent vector. The dataset is divided into a training set and a test set, and their size ratio is 3:1.

[0089] Continuing from the above embodiments, by setting the dropout probability to decrease with the network depth, the parameter sensitivity can be effectively reduced in the deep structure, thereby alleviating the overfitting problem. The probability of dropout is dynamically allocated in each residual network block to ensure that the network effectively reduces the dependence on a single neuron in the deep structure.

[0090] In some embodiments of the present application, in step S13, the training samples are divided into a training set and a validation set, and the training set is input into the network model; the metrics of the validation set are evaluated, and the optimal network model is selected as the flowchart network model.

[0091] When generating training samples, first preprocess the flowchart data set, process the flowchart data set, handle continuous and discrete variables, and complete the normalization operation to generate training samples. Divide the training samples into a training set and a validation set, iteratively update the parameters of the network model, and record the variational lower bound, reconstruction loss, and validation importance weighted autoencoder; use the validation importance weighted autoencoder to evaluate the metrics of the validation set, and select the optimal network model as the flowchart network model.

[0092] Specifically, when preprocessing the flowchart data set, distinguish the continuous variables and discrete variables in the data, perform normalization on the continuous variables; perform one-hot encoding on the discrete variables. The normalization formula is as follows:

[0093]

[0094] where X′ is the value after normalization and x is the original value.

[0095] Divide the training data into a training set and a validation set, iteratively update the model parameters, and record the variational lower bound, reconstruction loss, and validation importance weighted autoencoder (IWAE) at the same time; periodically estimate the IWAE metrics of the validation set, select the optimal model, and save its parameters and training status for subsequent use.

[0096] In some embodiments of the present application, in step S13, a probability generation mechanism based on component Bernoulli distribution creates mask data of missing values for features in the training samples; the training samples and mask data are input into the network model to train the network model. Here, after the flowchart data set is preprocessed as described above, a missing value mask is dynamically generated to simulate data missing scenarios to ensure the generalization ability of the model, thereby generating training samples and corresponding mask data. A probability generation mechanism based on component Bernoulli distribution creates a missing value mask for the features of the preprocessed data, wherein the missingness of the preprocessed data is independent of other features in the data, and the missing probability of each data point is the same; the missing value mask and the preprocessed data are input into the network model to determine the missing feature data and predict the missing value.

[0097] Specifically, the following processes are performed: 1) Generation of missing value masks: Dynamically create missing value masks through a probability generation mechanism based on component Bernoulli distribution, where the probability of each feature being missing in the data is controlled by the parameter p of the Bernoulli distribution; 2) Simulate the "completely random missing" pattern: The generated mask follows the completely random missing pattern, that is, the missing data is independent of other features or variables, and the missing probability of each data point is the same. This method simulates the situation in actual applications where the missing data is independent of other features; 3) Combination of mask and data: During the training process, the missing value mask is input into the network model together with the normalized real data. The network uses the mask to determine which data is missing and infers the missing values based on other features and training samples. In the embodiment of the present application, the use of masks not only helps the model learn how to deal with missing data, but also enhances the robustness and accuracy of the model in dealing with missing data in real-world scenarios.

[0098] The mask distribution is generated from a component-wise Bernoulli distribution, and the missing features introduced by the mask follow a "completely missing at random" pattern. The training samples and masked data are input into the network model for training. The ResNet blocks of the generative network integrate multi-scale information and learn the global patterns of latent variables to ensure accurate feature interpolation and robust data generation. This global pattern learning involves the generative network learning the distributional characteristics of the entire data through a prior network and a proposal network, ensuring that the interpolation results conform to the overall pattern of the data.

[0099] In some embodiments of the present application, after the flowchart network model is trained, the trained flowchart network model is evaluated to determine the proportion of successful convergence in the cases that originally failed to converge.

[0100] After training is completed, the imputation ability of the model is evaluated, and the proportion of cases that successfully converge among the original cases that could not converge after imputation is evaluated; the convergence probability is calculated to evaluate the effectiveness of the model as the percentage of the number of cases that successfully converge after optimization to the total number of cases. For example, there are 100 cases in the original dataset that could not converge. After optimization, 85 cases successfully converge, so the convergence rate is 85%; if the convergence rate after optimization is greater than 70%, it indicates that the overall performance of the model is good; if the convergence rate after optimization is still less than 50%, it means that the model fails to effectively improve the convergence of the process flow and needs to be retrained.

[0101] Figure 2 FIG. shows a schematic flow chart of a method for parameter prediction based on a flow chart according to another aspect of the present application. The method includes: steps S21 to S23.

[0102] Step S21, obtaining key configuration data of the current process in the process flow chart; Step S22, using the flow chart network model to predict missing values of the key configuration data, and adjusting parameters of the key configuration data according to the predicted missing values. Among them, the network model is constructed based on a variational autoencoder under arbitrary conditions and is trained using training samples. The training samples are determined from a flow chart dataset generated according to the key configuration data.

[0103] Process the key configuration data of the current process. The key configuration data can be configuration data that affects the convergence of the process flow chart, generate a flow chart dataset used as training samples, train a network model constructed based on a variational autoencoder under arbitrary conditions, and then train a flow chart network model adapted to the process flow chart for parameter prediction and convergence evaluation of the process flow chart.

[0104] Use the generated flow chart network model to predict the key configuration data of the original process flow chart. The original process flow chart is the initially obtained required current flow chart, predict the missing values of the parameters, and then exclude abnormal parameters, and then correct the wrong configurations that cause the flow chart to not converge. It should be noted that when the flow chart network model performs parameter prediction, it will predict the missing values of the missing parameters and treat abnormal parameters as missing parameters, and then predict the corresponding missing values.

[0105] Step S23, input the adjusted configuration data into the process flow chart for convergence prediction. Here, use the adjusted configuration data to input into the corresponding process flow chart for convergence prediction. The prediction is imputed data, and then generate a set of prediction data to obtain new configuration data, and the prediction data can be evaluated.

[0106] In some embodiments of the present application, in step S22, the target parameters in the process flow chart selected by the user are obtained, and a mask is added to the target parameters; the target parameters are replaced with missing value placeholders, where the missing value placeholders are used to simulate the missing state; the flowchart network model is used to predict the missing values of the missing value placeholders under the mask. Here, by using the generated flowchart network model combined with the application of a specific mask, the parameter exclusion and prediction of the configuration data of the original process flow chart are carried out to generate a set of optimized process flow configuration data, and the optimized process flow configuration data is evaluated to increase the convergence probability of the process flow. Thus, by using the mask, the key parameters are fixed, and only the key parameters under the mask are predicted, and the parameters without the mask are not predicted, making the prediction scheme more in line with the actual situation and supporting global and partial predictions.

[0107] Specifically: a specific parameter mask is set, and NaN is used to replace it to simulate the missing state and added to the data set, where NaN serves as the missing value placeholder and is used to mark the data that needs to be imputed; the trained flowchart network model is imported and applied to missing value filling and feature imputation; the generated filled values are combined with the non-missing features in the original data set to form complete optimized process flow configuration data, which is input into the process flow chart for convergence testing. Evaluate the proportion of successfully converged cases in the original non-converging cases after modifying the mask configuration.

[0108] In the above embodiments, the user can add a mask and set the parameter values for the configuration data selected in step S21. After setting, the best model trained in step S22 and adapted to the flowchart can be used for data generation. Among them, the data structure of the prediction list input into the model is expressed as:

[0109] Apath:{"status":"1","value":"60"}

[0110] Among them, Apath is the path in the variable manager corresponding to the selected parameter; status is whether to select to add a mask. If the value is 1, a mask is added; if the value is 0, there is no mask; value is the parameter value.

[0111] Combining the generated filled values with the non-missing features in the original data set to form complete optimized process flow configuration data and inputting it into the process flow chart for convergence testing can evaluate the proportion of successfully converged cases in the original non-converging cases after modifying the mask configuration.

[0112] Figure 3A performance graph of the decreasing arbitrary conditional variational autoencoder model used in an embodiment of the present application is shown, presenting the training and validation loss curves of the decreasing arbitrary conditional variational autoencoder model (VAEAC). The model performs well on both the training set and the validation set, and the performance on the training set and the validation set tends to be consistent.

[0113] Figure 4 A schematic diagram of the method flow for excluding and predicting incorrect configurations of a complex chemical process flow diagram based on an arbitrary conditional variational autoencoder in an embodiment of the present application is shown, which is used to optimize the non-convergence problem of the chemical process flow diagram, improve the convergence probability of the flow diagram, and simultaneously improve the stability and efficiency of the simulation. The method specifically includes the following steps: S1. Obtain the process flow configuration data, and generate different input data combinations through sampling; S2. Use multi-process computing to obtain the flow diagram data set, which is associated with the flow convergence; S3. Use the variational autoencoder to train the created data set to obtain a network model adapted to the process flow diagram; S4. Use the generated network model and combine the application of a specific mask to exclude and predict the parameters of the configuration data of the original process flow diagram, generate a set of prediction data, and evaluate the prediction data.

[0114] Figure 5 A schematic framework diagram of a device provided according to another aspect of the present application is shown. The device can be a device for training a flow diagram network model, and the device at least includes a processor 501 and a memory 502.

[0115] The processor 501 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 501 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor 501 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as the CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 501 may be integrated with a GPU (Graphics Processing Unit), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 501 may further include an AI (Artificial Intelligence) processor, and the AI processor is used to process computational operations related to machine learning.

[0116] The memory 502 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 502 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 502 is used to store at least one instruction, and the at least one instruction is used to be executed by the processor 501 to implement a method for training a flowchart network model provided in the method embodiments of the present application.

[0117] In some embodiments, the device may optionally further include: a peripheral device interface and at least one peripheral device. The processor 501, the memory 502, and the peripheral device interface may be connected through a bus or signal lines. Each peripheral device may be connected to the peripheral device interface through a bus, signal lines, or a circuit board. Schematically, the peripheral devices include but are not limited to: a radio frequency circuit, a touch display screen, an audio circuit, and a power supply, etc.

[0118] Of course, the device may also include fewer or more components, and this embodiment does not limit this.

[0119] Continue to refer to Figure 5, the device can also be a device for parameter prediction based on a flowchart; in some embodiments, the non-transitory computer-readable storage medium in the memory 502 is used to store at least one instruction, and the at least one instruction is used to be executed by the processor 501 to implement a method for parameter prediction based on a flowchart provided in the method embodiments of the present application.

[0120] The present application also provides a computer-readable medium, on which computer instructions are stored, and the computer-readable instructions can be executed by a processor to implement a method for training a flowchart network model as described above.

[0121] When the method for training a flowchart network model is implemented as a computer program, it can also be stored in a computer-readable storage medium as an article of manufacture. For example, the computer-readable storage medium can include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic strips), optical disks (such as compact discs (CDs), digital versatile discs (DVDs)), smart cards, and flash memory devices (such as electrically erasable programmable read-only memories (EPROMs), cards, sticks, key drives). In addition, the various storage media described herein can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" can include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data.

[0122] The present application also provides a computer-readable medium, on which computer instructions are stored, and the computer-readable instructions can be executed by a processor to implement a method for parameter prediction based on a flowchart as described above.

[0123] When the method for parameter prediction based on a flowchart is implemented as a computer program, it can also be stored in a computer-readable storage medium as an article of manufacture. For example, the computer-readable storage medium can include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic strips), optical disks (such as compact discs (CDs), digital versatile discs (DVDs)), smart cards, and flash memory devices (such as electrically erasable programmable read-only memories (EPROMs), cards, sticks, key drives). In addition, the various storage media described herein can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" can include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data.

[0124] It should be understood that the embodiments described above are merely illustrative. The embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or any combination thereof. For a hardware implementation, the processor can be implemented in one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, and / or other electronic units designed to perform the functions described herein, or a combination thereof.

[0125] Some aspects of the present application can be executed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software can all be referred to as "data blocks", "modules", "engines", "units", "components", or "systems". The processor can be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DAPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, or a combination thereof. In addition, aspects of the present application may be embodied as a computer product located in one or more computer readable media, which includes computer readable program code. For example, the computer readable media may include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic tapes...), optical discs (such as compact discs CD, digital versatile discs DVD...), smart cards, and flash memory devices (such as cards, sticks, key drives...).

[0126] The computer readable media may contain a propagated data signal having computer program code therein, for example, on a baseband or as part of a carrier wave. The propagated signal may have various forms of manifestation, including electromagnetic form, optical form, etc., or a suitable combination thereof. The computer readable media can be any computer readable media other than a computer readable storage media, which can be connected to an instruction execution system, apparatus, or device to implement communication, propagation, or transmission for use of the program. The program code located on the computer readable media can be propagated through any suitable media, including radio, cable, fiber optic cable, radio frequency signal, or similar media, or any combination of the above media.

[0127] The basic concepts have been described above. Obviously, for those skilled in the art, the above application disclosure is merely an example and does not constitute a limitation to the present application. Although not explicitly stated here, those skilled in the art may make various modifications, improvements, and corrections to the present application. Such modifications, improvements, and corrections are proposed in the present application, so such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of the present application.

[0128] Meanwhile, the present application uses specific terms to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" mean a certain feature, structure, or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that the "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more at different positions in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of the present application can be appropriately combined.

[0129] In some embodiments, numbers are used to describe components and the quantity of attributes. It should be understood that such numbers used to describe embodiments are, in some examples, modified by the modifiers "about", "approximately", or "substantially". Unless otherwise stated, "about", "approximately", or "substantially" indicate that the said numbers allow a variation of ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, and these approximate values can change according to the characteristics required by individual embodiments. In some embodiments, the numerical parameters should consider the specified significant digits and adopt the method of retaining the general number of digits. Although the numerical ranges and parameters used in some embodiments of the present application to confirm the breadth of their scope are approximate values, in specific embodiments, such numerical settings are made as precise as possible within the feasible range.

Claims

1. A method for training a flowchart network model, applied to a process flow, characterized in that, The method includes: Obtaining the key configuration data of the current process in the process flow chart; Generating a flow chart data set according to the key configuration data, and using the flow chart data set as a training sample, wherein the flow chart data set includes process convergence state information; Constructing a network model based on a variational autoencoder under arbitrary conditions, training the network model using the training sample, and constructing a flow chart network model adapted to the process flow chart for parameter prediction of the process flow chart.

2. The method according to claim 1, characterized in that Obtaining the key configuration data of the current process in the process flow chart includes: Judging whether the convergence state of the current process in the process flow chart is a failure. If so, determining the key parameters that cause the convergence failure, using the key parameters as the key configuration data of the current process, and setting corresponding parameter step sizes and value ranges for the key configuration data.

3. The method according to claim 2, wherein Generating a flow chart data set according to the key configuration data includes: Sampling the key configuration data according to the parameter step size and value range to generate an input data combination; Inputting the input data combination into the process flow chart for calculation to obtain a flow chart data set.

4. The method according to claim 3, characterized in that Inputting the input data combination into the process chart for calculation to obtain a flow chart data set includes: Initialization data of the current process obtained from the database; Inputting the input data combination into the process flow chart to replace the parameters at the corresponding positions, and calculating the convergence state information of the input data combination in combination with the initialization data; Creating a flow chart data set based on the input data combination and the corresponding convergence state information.

5. The method according to claim 1, characterized in that, Constructing a network model based on a variational autoencoder under arbitrary conditions includes: Building a prior network in the network model based on a variational autoencoder under arbitrary conditions, wherein the prior network uses a residual network as a hidden layer, and the prior network is used to receive observed features and missing value masks to learn to generate latent variables; Building a proposal network in the network model, wherein the proposal network uses a residual network as a hidden layer, and the proposal network is used to receive all features and output a proposal of latent variables; Building a generation network in the network model for decoding missing values from latent variables and multi-scale features through long skip connections; Setting a loss function, introducing the relative entropy divergence between the latent variable distributions between the prior network and the proposal network, and constructing an optimizer; Adding a dropout layer in each residual network block of the generation network to prevent overfitting.

6. The method according to claim 5, wherein The probability of the dropout is dynamically allocated in each residual network block and satisfies the following formula: where i is the index of the residual network block, P1 represents the probability of dropout of the residual network block that only receives one long skip connection, Depth is the number of hidden layers, and Nisc,i is the number of long skip connections received by the residual network block.

7. The method according to claim 1, wherein Training the network model using the training sample includes: Dividing the training sample into a training set and a validation set, and inputting the training set into the network model; Evaluating the metrics of the validation set, and selecting the optimal network model as the flow chart network model.

8. The method according to claim 1, characterized in that, Training the network model using the training samples includes: Creating masked data of missing values for the features in the training samples based on the probability generation mechanism of the component Bernoulli distribution; Inputting the training samples and the masked data into the network model to train the network model.

9. The method according to claim 1, characterized in that After training includes: Evaluating the trained flowchart network model to determine the proportion of successfully converged cases in the originally non-converging cases.

10. A method for parameter prediction based on a flowchart, characterized in that, The method includes: Obtaining the key configuration data of the current process in the process flowchart; Using the flowchart network model to predict the missing values of the key configuration data, and adjusting the parameters of the key configuration data according to the predicted missing values, where the flowchart network model is a network model constructed based on the variational autoencoder under arbitrary conditions and is obtained by training the network model using training samples, and the training samples are determined from the flowchart dataset generated according to the key configuration data; Inputting the adjusted configuration data into the process flowchart for convergence prediction.

11. The method according to claim 10, wherein Using the flowchart network model to predict the missing values of the key configuration data includes: Obtaining the target parameters in the process flowchart selected by the user and adding a mask to the target parameters; Replacing the target parameters with missing value placeholders, where the missing value placeholders are used to simulate the missing state; Using the flowchart network model to predict the missing values of the missing value placeholders under the mask.

12. A flowchart network model training device, characterized in that, The training device includes: One or more processors; and A memory storing computer-readable instructions that, when executed, cause the processor to perform the operations of the method according to any one of claims 1 to 9.

13. An apparatus for parameter prediction based on a flowchart, characterized in that, The device includes: One or more processors; and A memory storing computer-readable instructions that, when executed, cause the processor to perform the operations of the method according to claim 10 or 11.

14. A computer-readable medium having computer instructions stored thereon, the computer-readable instructions being executable by a processor to implement the method according to any one of claims 1 to 11.

Images