A method and system for automatic optimization of durene parameters based on deep learning

Abstract

The present invention relates to the technical field of machine learning, and specifically to a method and system for automatic optimization of durenyl parameters based on deep learning. The first historical parameter data is obtained by collecting historical production data of durenyl gas-phase oxidation to prepare pyromellitic anhydride, and the first pre-processed historical parameter data is obtained by pre-processing. It is divided into first positive example data and first negative example data according to by-product data. The first positive example data is analyzed to calculate the first parameter range of each parameter. The first real-time parameter data is obtained to determine whether it is within the first parameter range. If not, an early warning is given; if so, the first real-time parameter data is input into a first pyromellitic anhydride regression model to obtain first predicted output data, and the data is input into a first pyromellitic anhydride classification model to obtain a first predicted category. The reaction parameters are automatically adjusted according to the first predicted category to achieve intelligent optimization and effective control of preparing pyromellitic anhydride from durenyl.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method and system for automatically optimizing durenyl parameters based on deep learning. Background Art

[0002] Durene is a chemical substance also known as 1,2,4,5-tetramethylbenzene. Durene is a derivative of benzene with four methyl groups attached to different positions of the benzene ring. Its chemical formula is C 10 H 14 . Durene is soluble in ethanol, ether and benzene, but insoluble in water. It is widely used in many fields such as medicine, pesticides, dyes and fragrances. Durene is the raw material for preparing pyromellitic anhydride. The gas phase oxidation method of durene has the characteristics of simple process, which can omit the dehydration process to anhydride. It does not require other oxidants (such as nitric acid, permanganate and chromic acid, etc.) except air, and does not require the catalyst separation process required in the liquid phase method. It can achieve continuous production and is easy to realize automated operation. This method is widely used at home and abroad. Pyromellitic anhydride, also known as 1,2,4,5-benzoic acid anhydride, is mainly used as a raw material for polyimide, used to manufacture curing matting agents for epoxy resins and cross-linking agents for polyester resins. It is a widely used chemical product.

[0003] In the existing related research on the preparation of pyromellitic anhydride by gas phase oxidation of durenyl as raw material, scholar Zhang Wenjie studied the process of synthesizing pyromellitic anhydride by gas phase oxidation of durenyl in "Preparation of pyromellitic anhydride by air oxidation of durenyl", focusing on the selection of oxide catalysts. The catalyst carrier, active components and reaction conditions were investigated, and a catalyst with high catalytic activity was found through screening, and the optimal reaction process adjustment parameters were determined. Although this study has achieved certain results, it lacks in-depth analysis of catalyst performance and reaction mechanism, and requires more detailed research; in addition, the optimization space of process conditions has not been fully explored, and the yield and purity can still be further improved; the analysis of by-products and catalyst stability is insufficient, and there is a lack of comprehensive understanding; finally, the reliability and scientificity of the data need to be further confirmed. Scholar Li Qi used durenyl as raw material in "Preparation of pyromellitic anhydride by air gas phase oxidation of durenyl". However, this method has limitations, focusing only on the catalyst itself, without comprehensive consideration of other factors. Scholar Cao Maohong proposed a theoretical and experimental research method for the reaction system of gas-phase oxidation of durenyl to prepare pyromellitic anhydride in the "Kinetics of Selective Oxidation of Durenyl to Prepare Pyromellitic Anhydride and Simulation of Fixed-Bed Reactor". Through thermodynamic analysis, kinetic modeling and reactor simulation, the mechanism and condition optimization of the gas-phase oxidation reaction of durenyl were deeply explored, providing important theoretical support and technical guidance for the efficient preparation of pyromellitic anhydride. However, there are still some shortcomings that affect the depth and comprehensiveness of the research. In the study, only the main by-product carbon dioxide in the reaction system of gas-phase oxidation of durenyl to prepare dianhydride was analyzed, while other organic by-products were treated as a whole, resulting in insufficient comprehensive understanding of the reaction system; the research on the long-term stability and activity decay of the catalyst was not in-depth enough, and the specific mechanism and influencing factors of catalyst deactivation were not revealed; in the simulation of the fixed-bed reactor, the calculation of heat and mass transfer parameters was based on empirical formulas, lacking experimental data support, which may affect the accuracy and reliability of the simulation results; at the same time, the treatment of the reactor model was relatively simplified, and the heterogeneous properties of the particle phase and the gas phase were not considered, which may affect the authenticity and practicality of the simulation results.

[0004] Therefore, in summary of the above, the gas phase oxidation of durenyl to prepare pyromellitic anhydride is a complex multiphase catalytic process, and its product is a mixture of a wide range of oxygen-containing compounds. The composition of the reaction product mainly depends on the extreme temperature of the catalyst. In order to improve the yield and selectivity of the main product and reduce the generation of impurities in the reaction mixture, it is necessary to optimize the initial concentration of durenyl, the reaction temperature, the molten salt temperature and the contact time of the reaction system. However, current research is mostly focused on the regulation of a single parameter (such as the type of catalyst), which has certain limitations. At the same time, the existing parameter adjustment methods are often simple and rough, lack intelligence and accuracy, and are difficult to efficiently measure the complex interactions between the parameters. Therefore, the efficiency and economy of the chemical reaction process cannot be guaranteed, thereby restricting its further development in practical applications.

[0005] To this end, a method and system for automatic optimization of durenyl parameters based on deep learning are proposed. Summary of the invention

[0006] The object of the present invention is to provide a method and system for automatically optimizing durenyl parameters based on deep learning, and to realize parameter optimization and control of the process of preparing pyromellitic anhydride by gas phase oxidation of durenyl in a data-driven manner. First, historical production data is collected to obtain the first historical parameter data. The first pre-processed historical parameter data of normalization is obtained through pre-processing. The first pre-processed historical parameter data is divided into positive examples (meeting quality standards) and negative examples (not meeting quality standards) according to the by-product data, and a quality standard label is established. The first positive example data is analyzed to define the reasonable value range of the parameter. Then, in the production process, the reaction conditions and product performance are monitored in real time, the first real-time parameter data is obtained, and it is determined whether it is within the preset reasonable parameter range. If it is not within the range, an early warning is issued. If it is within the range, the first real-time parameter data is input into the first pyromellitic anhydride regression model to obtain the first predicted output data, and the first predicted output data is input into the first pyromellitic anhydride classification model to obtain the first predicted category. If the first prediction category is up to standard, the parameters remain unchanged; if it is not up to standard, the reaction is divided into stages according to the reaction time, product type and content in the historical data, and an independent deep learning model is trained for each stage, and the reaction parameters are adjusted through the expert experience database. This method uses a deep learning model to process historical data and real-time data, and combines it with an expert experience database to achieve intelligent optimization and automated control of the process of gas-phase oxidation of durene to produce pyromellitic anhydride, improving production efficiency and the stability and consistency of product quality.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for automatic optimization of durene parameters based on deep learning, comprising the following steps:

[0009] Collect historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene, obtain chemical reaction parameters under various reaction conditions, and obtain first historical parameter data, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content;

[0010] Preprocessing the first historical parameter data, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data;

[0011] Dividing the first preprocessing historical parameter data into first positive data and first negative data with quality standard labels according to the byproduct data;

[0012] Analyze the first positive example data to determine the mean and standard deviation of each parameter; calculate the first parameter range according to the mean and the standard deviation;

[0013] In the production process of preparing pyromellitic anhydride by gas phase oxidation of durene, real-time monitoring of reaction conditions and product performance is performed to obtain first real-time parameter data;

[0014] Determining whether the first real-time parameter data is within the first parameter range;

[0015] If the judgment is no, a warning is issued;

[0016] If the judgment is yes, the first real-time parameter data is input into a first pyromellitic anhydride regression model to predict the mass of the pyromellitic anhydride product at the end of the reaction to obtain first predicted output data;

[0017] Inputting the first predicted production data into a first pyromellitic anhydride classification model to obtain a first predicted category;

[0018] A reaction parameter is automatically adjusted based on the first predicted category.

[0019] Furthermore, the first historical parameter data also includes: reaction temperature, reaction pressure, oxygen flow rate, catalyst type, catalyst dosage, reaction time and reaction material ratio.

[0020] Furthermore, the mass data of pyromellitic dianhydride specifically refers to the mass of the generated pyromellitic dianhydride product.

[0021] Furthermore, the first preprocessing historical parameter data is divided according to the by-product data into:

[0022] The quality standard label is a binary classification label, including positive examples and negative examples;

[0023] If the by-product content in the first historical parameter data is greater than a preset threshold and / or the number of by-product types is greater than a preset threshold and / or the by-product types do not meet the requirements, the quality standard label with the category of negative example is added to the first preprocessed historical parameter data to obtain the first negative example data; the quality standard label with the category of positive example is added to the remaining first historical parameter data to obtain the first positive example data.

[0024] Furthermore, by analyzing the first positive example data to obtain the mean and the standard deviation of each parameter, the first parameter range is calculated as follows:

[0025] For each parameter, calculate its mean value and standard deviation in the first positive example data;

[0026] The first parameter range of the parameter is defined as [μ-3σ, μ+3σ];

[0027] Wherein, μ represents the mean value, and σ represents the standard deviation.

[0028] Furthermore, the first pyromellitic anhydride regression model is specifically:

[0029] The first isophthalic anhydride regression model is based on a fusion framework of a deep neural network and a graph neural network. The deep neural network processes continuous variables and adopts a multi-layer fully connected network structure. The graph neural network focuses on processing discrete variables and captures complex interactions between substances through a graph convolutional network. The first isophthalic anhydride regression model is trained by combining transfer learning and self-supervised learning, using a pre-trained model to accelerate convergence and generating pseudo labels through unlabeled data. The deep neural network adopts the Adam optimizer and the mean square error loss function, and the graph neural network uses a graph attention network to automatically learn the weights between nodes. The dependent variable of the first isophthalic anhydride regression model is the mass of isophthalic anhydride produced after the reaction is completed, and the independent variable is the first preprocessing historical parameter data except the dependent variable.

[0030] Furthermore, the first pyromellitic anhydride classification model is specifically:

[0031] Using the first positive example data and the first negative example data to train a classification model based on an RNN model; the RNN model includes an input layer, an embedding layer, a recurrent neural network layer, a fully connected layer, and an output layer, the output layer includes 2 neurons, and uses a Softmax activation function to output the probabilities of two classifications of product quality meeting the standard and not meeting the standard;

[0032] Integrate the first positive example data and the first negative example data, and divide them into a first training set and a first test set;

[0033] The model is trained using the first training set, the training label is the quality standard label, and the cross entropy loss function is used to optimize the model parameters. After the training is completed, the performance of the model is evaluated through the first test set to obtain the optimal first pyromellitic anhydride classification model; the first pyromellitic anhydride classification model classifies the first preprocessed historical parameter data into a qualified category and a non-qualified category.

[0034] Further, automatically adjusting the reaction conditions according to the first prediction category includes:

[0035] If the first prediction category is the target-reaching category, the reaction parameters are maintained without adjustment;

[0036] If the first prediction category is a non-compliant category, the chemical reaction is divided into multiple stages according to the reaction time, product type and product content in the first historical parameter data, and the chemical reaction parameters are used as different input features. An independent deep learning model is trained for each stage to obtain the stage chemical reaction parameter range; the reaction parameters are adjusted according to the stage chemical reaction parameter range and the expert experience database.

[0037] Further, adjusting the reaction parameters according to the stage chemical reaction parameter range and the expert experience database includes:

[0038] Collect and organize expert solutions and expert experience data for the chemical reaction of preparing isophthalic acid dianhydride by gas-phase oxidation of various types of tetramethylbenzene; store the expert solutions and the expert experience data in a database in a structured form to obtain the expert experience database; retrieve multiple expert candidate solutions that match the current reaction conditions and the first real-time parameter data from the expert experience database based on the first real-time parameter data; comprehensively consider the chemical reaction parameter range of the stage, and screen the first parameter adjustment solution from the multiple expert candidate solutions; and automatically adjust the reaction parameters based on the first parameter adjustment solution.

[0039] A deep learning-based automatic optimization system for durene parameters, comprising: a historical data acquisition module, a data preprocessing module, a positive and negative example classification module, a parameter range calculation module, a real-time data acquisition module, a condition judgment module, an early warning module, a regression model prediction module, a classification model prediction module and a parameter adjustment module;

[0040] The historical data acquisition module is used to collect historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene, obtain chemical reaction parameters under various reaction conditions, and obtain first historical parameter data, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content;

[0041] The data preprocessing module is used to preprocess the first historical parameter data, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data;

[0042] The positive and negative example division module is used to divide the first preprocessing historical parameter data into first positive example data and first negative example data with quality standard labels according to the byproduct data;

[0043] The parameter range calculation module is used to analyze the first positive example data to determine the mean and standard deviation of each parameter; and calculate the first parameter range according to the mean and the standard deviation;

[0044] The real-time data acquisition module is used to monitor the reaction conditions and product performance in real time and obtain the first real-time parameter data during the production process of preparing pyromellitic anhydride by gas-phase oxidation of durene;

[0045] The condition judgment module is used to judge whether the first real-time parameter data is within the first parameter range;

[0046] The early warning module is used to issue an early warning when the first real-time parameter data is not within the first parameter range;

[0047] The regression model prediction module is used for inputting the first real-time parameter data into a first pyromellitic anhydride regression model when the first real-time parameter data is within the first parameter range, predicting the mass of the pyromellitic anhydride product at the end of the reaction, and obtaining first predicted output data;

[0048] The classification model prediction module is used to input the first predicted output data into a first pyromellitic anhydride classification model to obtain a first predicted category;

[0049] The parameter adjustment module is used to automatically adjust the reaction parameters according to the first prediction category.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention divides the historical parameter data according to the by-product data and marks the quality standard label, which provides a good data basis for the subsequent steps, making the training process of the subsequent classification model more effective and accurate. By dividing the historical parameter data into positive examples and negative examples, the first pyromellitic anhydride classification model can better learn and understand which parameter combinations will lead to ideal product quality. Such label division helps the model to accurately capture the key features that affect product quality. In addition, this method improves the generalization ability and prediction accuracy of the classification model. By considering the impact of by-product content and type on quality labels, the model can learn more abundant and comprehensive information during the training process, thereby more accurately predicting the product quality under unknown parameter combinations. The application effect and stability of the model in actual production can be effectively improved, thereby fully ensuring the accuracy of parameter adjustment.

[0052] 2. The first pyromellitic anhydride regression model of the present invention adopts a fusion framework of deep neural network and graph neural network, which can effectively process continuous variables and discrete variables at the same time. The deep neural network is responsible for processing continuous parameters and learning data features through a multi-layer fully connected network structure; while the graph neural network focuses on processing discrete parameters, using graph convolutional networks to capture the complex interaction relationship between substances, significantly improving the model's ability to extract data features. In addition, the model is trained by combining transfer learning and self-supervised learning, using pre-trained models to accelerate the model convergence speed, and generating pseudo labels through unlabeled data, thereby improving the generalization ability and training effect of the model. The deep neural network uses the Adam optimizer and the mean square error loss function to effectively improve the convergence speed and training effect of the model; at the same time, the graph neural network uses the graph attention network to automatically learn the weights between nodes, enhancing the model's ability to process graph data. In summary, the first pyromellitic anhydride regression model of the present invention has efficient, accurate and reliable prediction capabilities in the optimization of chemical reaction parameters, and provides important support and guarantee for the automatic optimization of parameters in chemical production processes.

[0053] 3. The parameter adjustment method in the present invention is discussed by different situations to make the whole system more intelligent and flexible. According to different prediction categories, corresponding strategies are adopted to achieve more accurate, efficient and reliable chemical reaction parameter optimization and product quality control. When the first prediction category is up to standard, the system maintains the current reaction parameters, indicating that the current conditions have reached the expected product quality requirements. This strategy avoids unnecessary adjustments and saves resources and time. When the first prediction category is not up to standard, the system decomposes the chemical reaction into multiple stages and trains an independent deep learning model for each stage to determine the stage chemical reaction parameter range. Finally, the system uses the stage chemical reaction parameter range and the expert experience database to adjust the reaction parameters. This stage-based adjustment strategy effectively improves the accuracy and controllability of the reaction parameters, can better cope with the difference between theoretical parameters and actual conditions, thereby improving the accuracy and operability of parameter adjustment, and ensuring the stability and optimization performance of product quality. In summary, this automatic adjustment strategy not only improves the efficiency and economy of the chemical reaction, but also enhances the system's intelligent processing ability for complex reaction processes, making the parameter automatic optimization method for preparing pyromellitic anhydride from durenyl more intelligent and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic diagram of a pyromellitic dianhydride sample provided in an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of a durene sample provided in an embodiment of the present invention;

[0056] Figure 3 A flow chart of a method for automatic optimization of durene parameters based on deep learning provided in an embodiment of the present invention;

[0057] Figure 4 A process flow chart of preparing pyromellitic dianhydride by gas phase oxidation of durene provided in an embodiment of the present invention;

[0058] Figure 5 A structural diagram of a system for automatic optimization of durenyl parameters based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0060] Embodiment 1:

[0061] Pyromellitic dianhydride (such as Figure 1 As shown in Figure 2, PMDA is abbreviated as PMDA. With the rapid development of the petrochemical industry, the output of C9 fraction has increased dramatically, bringing high-quality and high-quality durenyl (such as Figure 2 As shown). Since the gas phase oxidation method of tetramethylbenzene has the characteristics of high atomic utilization and low environmental pollution, the air gas phase oxidation method is currently commonly used. However, due to the characteristics of large heat release, easy over-oxidation and fast reaction rate, it is necessary to select a heat-resistant catalyst with good thermal conductivity. In addition, the product needs to be desorbed from the catalyst surface in time after the reaction. Therefore, it is necessary to consider both the activity of the catalyst and the depth of the oxidation reaction. The oxidation process is usually a complex reaction network formed by series, parallel or a combination of the two. Different catalysts and reaction conditions may cause the reaction to pass through different paths and produce different reaction products. Therefore, determining the appropriate process parameters to improve the yield of PMDA is a challenging problem.

[0062] In this embodiment, a factory A in a certain city adopts a method for automatic optimization of durene parameters based on deep learning, including the following steps:

[0063] See also Figure 3 S10 in which historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene is collected, chemical reaction parameters under various reaction conditions are obtained, and first historical parameter data are obtained, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content; such as Figure 4 The process flow of preparing pyromellitic dianhydride by gas phase oxidation of durene is shown in Table 1, which includes four major steps: oxidation, hydrolysis, dehydration and drying. In the oxidation stage, a catalyst is added in addition to durene.

[0064] Table 1. Four major steps of preparing pyromellitic anhydride by gas phase oxidation of durene

[0065]

[0066] Furthermore, the first historical parameter data also includes: reaction temperature, reaction pressure, oxygen flow rate, catalyst type, catalyst dosage, reaction time and reaction material ratio.

[0067] The first historical parameter data fully describes the various reaction conditions and operating parameters involved in the historical production data. These parameters cover the key operating conditions and influencing factors in the reaction process, and can fully analyze and understand the physical and chemical environment of the reaction. These parameters have an important impact on the production process of preparing pyromellitic anhydride by gas phase oxidation of durene, and directly affect the quality and output of the product. It helps to fully analyze and understand the key factors in the reaction process, thereby realizing intelligent optimization and effective control of the production process, and improving production efficiency and product quality.

[0068] Furthermore, the mass data of pyromellitic dianhydride specifically refers to the mass of the generated pyromellitic dianhydride product.

[0069] The quality data of the generated pyromellitic dianhydride product can be used for quality control and evaluation to help determine whether the product quality in the production process meets the standards and expected requirements.

[0070] See also Figure 3 In S20, the first historical parameter data is preprocessed, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data; it is necessary to remove abnormal data points whose units or dimensions are obviously inconsistent with other data due to data entry errors or equipment failures.

[0071] In the chemical reaction in which durene participates, the main reaction is durene (C 10 H 14 ) reacts with oxygen to form pyromellitic anhydride and water. There are many parallel reactions simultaneously with the main reaction to produce by-products such as pyromellitic acid, trimellitic anhydride, xylene terephthalic acid, dimethylphthalic anhydride, 5-methyltrimellitic anhydride, terephthalic acid and maleic acid, as well as carbon monoxide and carbon dioxide generated by the complete and incomplete combustion of tetramethylene.

[0072] See also Figure 3 In S30, the first preprocessing historical parameter data is divided according to the byproduct data into first positive data and first negative data with quality standard labels;

[0073] Furthermore, the first preprocessing historical parameter data is divided according to the by-product data into:

[0074] The quality standard label is a binary classification label, including positive examples and negative examples;

[0075] If the by-product content in the first historical parameter data is greater than a preset threshold and / or the number of by-product types is greater than a preset threshold and / or the by-product types do not meet the requirements, the quality standard label with the category of negative example is added to the first preprocessed historical parameter data to obtain the first negative example data; the quality standard label with the category of positive example is added to the remaining first historical parameter data to obtain the first positive example data.

[0076] First, by setting preset thresholds and combining the restrictions on by-product content and types, it is possible to more accurately screen out abnormal or unqualified situations in the production process and mark these situations as negative data. Doing so helps to detect potential problems or abnormalities in the production process early and improve the stability and reliability of the production process. Secondly, clearly distinguishing positive and negative data and processing them in a targeted manner can provide a clearer and more accurate data basis for subsequent data analysis and model training. This data division method can help the system better understand and learn the differences between normal production conditions and abnormal situations, thereby improving the accuracy and predictive ability of the model. In addition, marking abnormal situations as negative examples and dividing negative data can help to provide timely early warning and processing, reduce the risk of quality problems in the production process, and thus improve the quality of anhydride products and production efficiency.

[0077] See also Figure 3 In S40, analyzing the first positive example data to determine the mean and standard deviation of each parameter; calculating the first parameter range according to the mean and the standard deviation;

[0078] Furthermore, by analyzing the first positive example data to obtain the mean and the standard deviation of each parameter, the first parameter range is calculated as follows:

[0079] For each parameter, calculate its mean value and standard deviation in the first positive example data;

[0080] The first parameter range of the parameter is defined as [μ-3σ, μ+3σ];

[0081] Wherein, μ represents the mean value, and σ represents the standard deviation.

[0082] Table 2. Parameter examples of the first positive example data

[0083] Serial number stage Reaction temperature (℃) ... <![CDATA[Air volume (m 3 / h)]]> 001 Oxidation 435 ... 2100 002 Oxidation 440 ... 2200 003 Oxidation 440 ... 2200 004 Oxidation 445 ... 2280 005 Oxidation 440 ... 2300

[0084] This embodiment takes five pieces of data as an example to demonstrate the specific process of calculating the first parameter range of the two parameters of reaction temperature and component (the calculation result of the first parameter range retains integer digits). The specific data is shown in Table 2, and the intermediate calculation steps are as follows:

[0085]

[0086] Among them, μ1 represents the mean of reaction temperature, μ2 represents the mean of air volume, σ1 represents the standard deviation of reaction temperature, and σ2 represents the standard deviation of air volume.

[0087] The first parameter range of the reaction temperature is [440-3×3.54, 440+3×3.54], that is, [429, 451].

[0088] The first parameter range of the air volume is [2216-3×79.25,2216+3×79.25], that is, [1978,2454].

[0089] In the oxidation stage, the mean of the reaction temperature is 440, the standard deviation is 3.54, and the first parameter range is [429, 451]. The mean of the air volume is 2216, the standard deviation is 79.25, and the first parameter range is [1978, 2454].

[0090] Calculating the mean and standard deviation of parameters under normal production conditions helps to understand the typical performance and variation of parameters in the production process of preparing homoanhydride from durene. The mean value provides the central position of the parameter under normal conditions, while the standard deviation reflects the degree of variation of the parameter value around the mean value. Such analysis can help determine a reasonable parameter range to ensure that the production process is carried out under stable conditions. Secondly, defining a reasonable parameter value range can provide an effective reference for parameter setting in the production process. Through the parameter range calculated based on the positive example data, the upper and lower limits of the parameters in the production process can be set to avoid excessive or too small parameter values, thereby ensuring the stability and reliability of production. In addition, the parameter range setting based on the three sigma principle helps to establish an effective production control strategy. In the case of normal distribution (or approximate normal distribution), about 99.7% of the data points fall within the range of the mean plus or minus three times the standard deviation. Therefore, limiting the parameter value range to three times the standard deviation can effectively cover the vast majority of normal situations and exclude abnormal situations, which is conducive to ensuring the stability of production and quality controllability.

[0091] See also Figure 3 S50, in the production process of preparing pyromellitic anhydride by gas phase oxidation of durene, real-time monitoring of reaction conditions and product properties to obtain first real-time parameter data;

[0092] See also Figure 3 S60, determining whether the first real-time parameter data is within the first parameter range;

[0093] Table 3. Example of first real-time parameter data judgment

[0094]

[0095] As shown in Table 3, the vaporizer temperature and the molten salt temperature in the first real-time parameter data are both outside the specified first parameter range.

[0096] See also Figure 3 In S70, if the judgment is no, a warning is issued; some of the first real-time parameter data provided in the embodiment are shown in Table 3, and a warning is directly issued in this case.

[0097] See also Figure 3 In S80, if the judgment is yes, inputting the first real-time parameter data into a first pyromellitic anhydride regression model to predict the mass of the pyromellitic anhydride product at the end of the reaction to obtain first predicted yield data;

[0098] Furthermore, the first pyromellitic anhydride regression model is specifically:

[0099] The first isophthalic anhydride regression model is based on a fusion framework of a deep neural network and a graph neural network. The deep neural network processes continuous variables and adopts a multi-layer fully connected network structure. The graph neural network focuses on processing discrete variables and captures complex interactions between substances through a graph convolutional network. The first isophthalic anhydride regression model is trained by combining transfer learning and self-supervised learning, using a pre-trained model to accelerate convergence and generating pseudo labels through unlabeled data. The deep neural network adopts the Adam optimizer and the mean square error loss function, and the graph neural network uses a graph attention network to automatically learn the weights between nodes. The dependent variable of the first isophthalic anhydride regression model is the mass of isophthalic anhydride produced after the reaction is completed, and the independent variable is the first preprocessing historical parameter data except the dependent variable.

[0100] First, the regression model of isophthalic anhydride adopts a fusion framework of deep neural network and graph neural network, which can make full use of the advantages of these two networks. Deep neural network is suitable for processing continuous variables, while graph neural network focuses on processing discrete variables, such as interactions between substances. By fusing them together, it is possible to more comprehensively capture and understand the complex characteristics and relationships in chemical reactions, and improve the accuracy and generalization ability of the model. Secondly, the model is trained by combining transfer learning and self-supervised learning. The pseudo labels are generated by using pre-trained models and unlabeled data, which helps to accelerate the convergence of the model and improve the generalization ability of the model. Transfer learning can apply learned knowledge to new tasks, while self-supervised learning can effectively use the data's own information for training, reduce dependence on labeled data, and improve the model's adaptability. In addition, the Adam optimizer and mean square error loss function, as well as the graph attention network, are used in the model to automatically learn the weights between nodes. These optimizations and technical means help to improve the efficiency and accuracy of model training, thereby better predicting the product quality of isophthalic anhydride, and provide important support and guarantee for parameter optimization and quality control in the production process of isophthalic anhydride.

[0101] See also Figure 3 In S90, the first predicted output data is input into a first pyromellitic anhydride classification model to obtain a first predicted category;

[0102] Furthermore, the first pyromellitic anhydride classification model is specifically:

[0103] Using the first positive example data and the first negative example data to train a classification model based on an RNN model; the RNN model includes an input layer, an embedding layer, a recurrent neural network layer, a fully connected layer, and an output layer, the output layer includes 2 neurons, and uses a Softmax activation function to output the probabilities of two classifications of product quality meeting the standard and not meeting the standard;

[0104] Integrate the first positive example data and the first negative example data, and divide them into a first training set and a first test set;

[0105] The model is trained using the first training set, the training label is the quality standard label, and the cross entropy loss function is used to optimize the model parameters. After the training is completed, the performance of the model is evaluated through the first test set to obtain the optimal first pyromellitic anhydride classification model; the first pyromellitic anhydride classification model classifies the first preprocessed historical parameter data into a qualified category and a non-qualified category.

[0106] The first classification model of isophthalic anhydride utilizes the structure of RNN, which can effectively process data with sequential properties. In the process of preparing isophthalic anhydride by gas phase oxidation of isophthalic anhydride, the judgment of product quality is often affected by the previous reaction conditions and parameters. Therefore, the use of RNN can better capture this temporal correlation and improve the model's understanding and prediction ability of the relationship between parameters and product quality. Secondly, the design of the model structure can effectively extract features and perform classification and discrimination, thereby achieving accurate prediction and classification of product quality. In addition, the model uses a cross entropy loss function to optimize model parameters during training, trains the model through the first training set, and evaluates the performance of the model through the first test set, thereby obtaining the optimal classification model. This training and evaluation method helps to improve the generalization ability and accuracy of the model, so that it can be effectively applied and guide parameter adjustment in actual production processes.

[0107] See also Figure 3 In S100, a reaction parameter is automatically adjusted according to the first prediction category.

[0108] Further, automatically adjusting the reaction conditions according to the first prediction category includes:

[0109] If the first prediction category is the target-reaching category, the reaction parameters are maintained without adjustment;

[0110] If the first prediction category is a non-compliant category, the chemical reaction is divided into multiple stages according to the reaction time, product type and product content in the first historical parameter data, and the chemical reaction parameters are used as different input features. An independent deep learning model is trained for each stage to obtain the stage chemical reaction parameter range; the reaction parameters are adjusted according to the stage chemical reaction parameter range and the expert experience database.

[0111] For the prediction results of the qualified category, the applied automatic optimization method can automatically identify that the current reaction conditions have reached the expected quality standards, so no adjustment is required. This automated judgment can save human resources, improve production efficiency, and ensure a stable production state when the quality is up to standard. For the prediction results of the unqualified category, the method adopts a more refined adjustment strategy. By analyzing the information such as reaction time, product type and product content in the first historical parameter data, the chemical reaction is decomposed into multiple stages, and the chemical reaction parameters are used as different input features to train independent deep learning models for each stage. This staged adjustment strategy can more accurately identify the problem and adjust the reaction conditions in a targeted manner, thereby improving the control ability and response speed of the production process. Finally, the reaction parameters are adjusted according to the stage chemical reaction parameter range and the expert experience database. The system comprehensively considers the prediction ability of the deep learning model and the practical guidance of expert experience, making the adjustment process more scientific and accurate, and can be effectively applied in actual production. This comprehensive adjustment strategy can improve the stability of the production process and the consistency of product quality, providing beneficial support and guarantee for production efficiency and quality.

[0112] Further, adjusting the reaction parameters according to the stage chemical reaction parameter range and the expert experience database includes:

[0113] Collect and organize expert solutions and expert experience data for the chemical reaction of preparing isophthalic acid dianhydride by gas-phase oxidation of various types of tetramethylbenzene; store the expert solutions and the expert experience data in a database in a structured form to obtain the expert experience database; retrieve multiple expert candidate solutions that match the current reaction conditions and the first real-time parameter data from the expert experience database based on the first real-time parameter data; comprehensively consider the chemical reaction parameter range of the stage, and screen the first parameter adjustment solution from the multiple expert candidate solutions; and automatically adjust the reaction parameters based on the first parameter adjustment solution.

[0114] By collecting and collating expert solutions and experience data, this method can adjust reaction parameters with the help of rich professional knowledge and experience. Such a method can not only ensure the scientificity and reliability of parameter adjustment, but also inherit and apply the valuable experience of experts to improve the accuracy and efficiency of parameter adjustment. According to real-time parameter data, the matching expert candidate solutions can be quickly retrieved from the expert experience database, which can effectively save the time and energy of parameter adjustment. The system can quickly obtain multiple candidate solutions that match the current reaction conditions, providing reference and selection for subsequent parameter adjustment. At the same time, considering the range of stage chemical reaction parameters and multiple candidate solutions in the expert experience database, the system can comprehensively consider multiple factors, including historical data, real-time parameters and expert experience data, so as to formulate a more comprehensive and effective parameter adjustment plan. By automatically adjusting the reaction parameters, the system can optimize the production process, improve production efficiency and product quality. Parameter adjustment based on expert experience and real-time data helps to reduce production costs, reduce scrap rates, and ensure that the production process is stable and controllable.

[0115] In general, the deep learning-based automatic optimization method for tetramethylbenzene parameters used in Factory A can first accurately determine the effective range of chemical reaction parameters by analyzing historical production data and by-product information, and after data preprocessing and positive and negative example division, thereby optimizing the production process. Secondly, through real-time monitoring and prediction models, timely monitoring and early warning of reaction conditions can be achieved, and adverse reaction trends can be quickly discovered to ensure that the production process is stable and the product quality meets the standards. Most importantly, this method can automatically adjust the reaction parameters according to the predicted categories, realize automatic optimization of parameters, and improve production efficiency and product quality. This intelligent and efficient production method will significantly improve product quality and output, bringing significant benefits to factory operations.

[0116] Embodiment 2:

[0117] In this embodiment, a factory B in a certain city applied a deep learning-based automatic optimization system for durene parameters, such as Figure 5 As shown in Table 4, the system includes: historical data acquisition module, data preprocessing module, positive and negative example classification module, parameter range calculation module, real-time data acquisition module, condition judgment module, early warning module, regression model prediction module, classification model prediction module and parameter adjustment module; the main equipment included in the system is shown in Table 4.

[0118] Table 4. List of main equipment of the system

[0119] Chemical trough valve Delivery pump Measuring tank Filters Vapor mixer Roots blower Air buffer tank Catcher Air preheater Molten salt cooler Oxidation Reactor Molten salt tank Water washing tower Heat Exchanger Crystallization kettle Concentration kettle Sublimation kettle Vacuum Pump Air compressor

[0120] The historical data acquisition module is used to collect historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene, obtain chemical reaction parameters under various reaction conditions, and obtain first historical parameter data, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content;

[0121] The data preprocessing module is used to preprocess the first historical parameter data, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data;

[0122] The positive and negative example division module is used to divide the first preprocessing historical parameter data into first positive example data and first negative example data with quality standard labels according to the byproduct data;

[0123] The parameter range calculation module is used to analyze the first positive example data to determine the mean and standard deviation of each parameter; and calculate the first parameter range according to the mean and the standard deviation;

[0124] The real-time data acquisition module is used to monitor the reaction conditions and product performance in real time and obtain the first real-time parameter data during the production process of preparing pyromellitic anhydride by gas-phase oxidation of durene;

[0125] The condition judgment module is used to judge whether the first real-time parameter data is within the first parameter range;

[0126] The early warning module is used to issue an early warning when the first real-time parameter data is not within the first parameter range;

[0127] The regression model prediction module is used for inputting the first real-time parameter data into a first pyromellitic anhydride regression model when the first real-time parameter data is within the first parameter range, predicting the mass of the pyromellitic anhydride product at the end of the reaction, and obtaining first predicted output data;

[0128] The classification model prediction module is used to input the first predicted output data into a first pyromellitic anhydride classification model to obtain a first predicted category;

[0129] The parameter adjustment module is used to automatically adjust the reaction parameters according to the first prediction category.

[0130] Table 5. Comparison data before and after the application of the system in Factory B

[0131]

[0132] As shown in Table 5, the application of the deep learning-based automatic optimization system for durenyl parameters in Factory B in a certain city has significantly improved production efficiency, reduced energy consumption and by-product generation, increased the output of homoanhydride products, and reduced abnormalities and failures in production by optimizing production process parameters and conditions, thus bringing significant economic benefits. The application of the system makes the production process more intelligent, stable and sustainable, and improves the competitiveness and sustainable development capabilities of the factory.

[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic optimization of durene parameters based on deep learning, characterized in that: The following steps are involved: Collect historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene, obtain chemical reaction parameters under various reaction conditions, and obtain first historical parameter data, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content; Preprocessing the first historical parameter data, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data; Dividing the first preprocessing historical parameter data into first positive data and first negative data with quality standard labels according to the byproduct data; Analyze the first positive example data to determine the mean and standard deviation of each parameter; calculate the first parameter range according to the mean and the standard deviation; In the production process of preparing pyromellitic anhydride by gas phase oxidation of durene, real-time monitoring of reaction conditions and product performance is performed to obtain first real-time parameter data; Determining whether the first real-time parameter data is within the first parameter range; If the judgment is no, a warning is issued; If the judgment is yes, the first real-time parameter data is input into a first pyromellitic anhydride regression model to predict the mass of the pyromellitic anhydride product at the end of the reaction to obtain first predicted output data; Inputting the first predicted production data into a first pyromellitic anhydride classification model to obtain a first predicted category; automatically adjusting a reaction parameter based on the first predicted category; The first preprocessing historical parameter data is divided according to the by-product data as follows: The quality standard label is a binary classification label, including positive examples and negative examples; If the by-product content in the first historical parameter data is greater than a preset threshold value and / or the number of the by-product types is greater than a preset threshold value and / or the by-product types do not meet the requirements, the quality standard label of the negative example category is added to the first pre-processed historical parameter data to obtain the first negative example data; the quality standard label of the positive example category is added to the rest of the first historical parameter data to obtain the first positive example data; Wherein, the first pyromellitic anhydride regression model is specifically: The first isophthalic anhydride regression model is based on a fusion framework of a deep neural network and a graph neural network. The deep neural network processes continuous variables and adopts a multi-layer fully connected network structure. The graph neural network focuses on processing discrete variables and captures complex interactions between substances through a graph convolutional network. The first isophthalic anhydride regression model is trained by combining transfer learning and self-supervised learning, using a pre-trained model to accelerate convergence, and generating pseudo labels through unlabeled data. The deep neural network uses an Adam optimizer and a mean square error loss function, and the graph neural network uses a graph attention network to automatically learn the weights between nodes. The dependent variable of the first isophthalic anhydride regression model is the mass of isophthalic anhydride produced after the reaction is completed, and the independent variable is the first pre-processed historical parameter data except the dependent variable. Wherein, automatically adjusting the reaction conditions according to the first prediction category includes: If the first prediction category is the target-reaching category, the reaction parameters are maintained without adjustment; If the first prediction category is a non-compliant category, the chemical reaction is divided into multiple stages according to the reaction time, product type and product content in the first historical parameter data, and the chemical reaction parameters are used as different input features. An independent deep learning model is trained for each stage to obtain the stage chemical reaction parameter range; the reaction parameters are adjusted according to the stage chemical reaction parameter range and the expert experience database.

2. The method for automatic optimization of durene parameters based on deep learning according to claim 1, characterized in that: The first historical parameter data also includes: reaction temperature, reaction pressure, oxygen flow rate, catalyst type, catalyst dosage, reaction time and reaction material ratio.

3. The method for automatic optimization of durene parameters based on deep learning according to claim 1, characterized in that: The mass data of pyromellitic dianhydride specifically refers to the mass of the generated pyromellitic dianhydride product.

4. The method for automatic optimization of durene parameters based on deep learning according to claim 1, characterized in that: The mean and the standard deviation of each parameter are obtained by analyzing the first positive example data, and the first parameter range is calculated as follows: For each parameter, calculate its mean value and standard deviation in the first positive example data; The first parameter range of the parameter is defined as [μ-3σ, μ+3σ]; Wherein, μ represents the mean value, and σ represents the standard deviation.

5. The method for automatic optimization of durene parameters based on deep learning according to claim 1, characterized in that: The first pyromellitic anhydride classification model is specifically: Using the first positive example data and the first negative example data to train a classification model based on an RNN model; the RNN model includes an input layer, an embedding layer, a recurrent neural network layer, a fully connected layer, and an output layer, the output layer includes 2 neurons, and uses a Softmax activation function to output the probabilities of two classifications of product quality meeting the standard and not meeting the standard; Integrate the first positive example data and the first negative example data, and divide them into a first training set and a first test set; The model is trained using the first training set, the training label is the quality standard label, and the cross entropy loss function is used to optimize the model parameters. After the training is completed, the performance of the model is evaluated through the first test set to obtain the optimal first pyromellitic anhydride classification model; the first pyromellitic anhydride classification model classifies the first preprocessed historical parameter data into a qualified category and a non-qualified category.

6. The method for automatic optimization of durene parameters based on deep learning according to claim 1, characterized in that: Adjusting the reaction parameters according to the chemical reaction parameter range of the stage and the expert experience database includes: Collect and organize expert solutions and expert experience data for the chemical reaction of preparing isophthalic acid dianhydride by gas-phase oxidation of various types of tetramethylbenzene; store the expert solutions and the expert experience data in a database in a structured form to obtain the expert experience database; retrieve multiple expert candidate solutions that match the current reaction conditions and the first real-time parameter data from the expert experience database based on the first real-time parameter data; comprehensively consider the chemical reaction parameter range of the stage, and screen the first parameter adjustment solution from the multiple expert candidate solutions; and automatically adjust the reaction parameters based on the first parameter adjustment solution.

7. A deep learning-based automatic optimization system for durene parameters, characterized in that: include: Historical data collection module, data preprocessing module, positive and negative example classification module, parameter range calculation module, real-time data collection module, condition judgment module, early warning module, regression model prediction module, classification model prediction module and parameter adjustment module; The historical data acquisition module is used to collect historical production data of preparing pyromellitic anhydride by gas phase oxidation of durene, obtain chemical reaction parameters under various reaction conditions, and obtain first historical parameter data, including pyromellitic anhydride mass data and by-product data; the by-product data includes by-product type and by-product content; The data preprocessing module is used to preprocess the first historical parameter data, including outlier removal and missing value processing, to obtain first preprocessed historical parameter data; The positive and negative example division module is used to divide the first preprocessing historical parameter data into first positive example data and first negative example data with quality standard labels according to the byproduct data; The parameter range calculation module is used to analyze the first positive example data to determine the mean and standard deviation of each parameter; and calculate the first parameter range according to the mean and the standard deviation; The real-time data acquisition module is used to monitor the reaction conditions and product performance in real time and obtain the first real-time parameter data during the production process of preparing pyromellitic anhydride by gas-phase oxidation of durene; The condition judgment module is used to judge whether the first real-time parameter data is within the first parameter range; The early warning module is used to issue an early warning when the first real-time parameter data is not within the first parameter range; The regression model prediction module is used for inputting the first real-time parameter data into a first pyromellitic anhydride regression model when the first real-time parameter data is within the first parameter range, predicting the mass of the pyromellitic anhydride product at the end of the reaction, and obtaining first predicted output data; The classification model prediction module is used to input the first predicted output data into a first pyromellitic anhydride classification model to obtain a first predicted category; The parameter adjustment module is used to automatically adjust the reaction parameters according to the first prediction category; wherein the first preprocessing historical parameter data is divided according to the by-product data as follows: The quality standard label is a binary classification label, including positive examples and negative examples; If the by-product content in the first historical parameter data is greater than a preset threshold value and / or the number of the by-product types is greater than a preset threshold value and / or the by-product types do not meet the requirements, the quality standard label of the negative example category is added to the first pre-processed historical parameter data to obtain the first negative example data; the quality standard label of the positive example category is added to the rest of the first historical parameter data to obtain the first positive example data; Wherein, the first pyromellitic anhydride regression model is specifically: The first isophthalic anhydride regression model is based on a fusion framework of a deep neural network and a graph neural network. The deep neural network processes continuous variables and adopts a multi-layer fully connected network structure. The graph neural network focuses on processing discrete variables and captures complex interactions between substances through a graph convolutional network. The first isophthalic anhydride regression model is trained by combining transfer learning and self-supervised learning, using a pre-trained model to accelerate convergence, and generating pseudo labels through unlabeled data. The deep neural network uses an Adam optimizer and a mean square error loss function, and the graph neural network uses a graph attention network to automatically learn the weights between nodes. The dependent variable of the first isophthalic anhydride regression model is the mass of isophthalic anhydride produced after the reaction is completed, and the independent variable is the first pre-processed historical parameter data except the dependent variable. Wherein, automatically adjusting the reaction conditions according to the first prediction category includes: If the first prediction category is the target-reaching category, the reaction parameters are maintained without adjustment; If the first prediction category is a non-compliant category, the chemical reaction is divided into multiple stages according to the reaction time, product type and product content in the first historical parameter data, and the chemical reaction parameters are used as different input features. An independent deep learning model is trained for each stage to obtain the stage chemical reaction parameter range; the reaction parameters are adjusted according to the stage chemical reaction parameter range and the expert experience database.

Images