A causal modeling optimization method for the chloride process titanium dioxide production of titanium dioxide based on a neural network surrogate model, along with electronic devices and storage media.

Abstract

This invention discloses a causal modeling optimization method, electronic device, and storage medium for the chloride process titanium dioxide production process based on a neural network surrogate model. The method includes acquiring and normalizing preprocessing data of the chloride process titanium dioxide production process; using recursive feature elimination based on cross-validation to obtain confounding factor features strongly correlated with the target output; training a surrogate model using the obtained feature data to predict the potential effects of each manipulated variable on the target output; and calculating the causal benefits through the potential results output by the surrogate model. This invention reduces the computational cost of causal modeling for complex chloride process titanium dioxide production processes, and can efficiently evaluate the causal benefits of manipulated variables when the system model is unclear or system parameters are lacking, providing a scientific basis for auxiliary decision-making and fault diagnosis in the chloride process titanium dioxide production process.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a causal modeling optimization method for the chloride process titanium dioxide production based on a neural network surrogate model, as well as electronic equipment and storage media. Background Technology

[0002] In the chloride process for titanium dioxide production, the production process typically involves complex reactions, heat transfer, mass transfer, and fluid flow, exhibiting high nonlinearity and strong coupling. Traditional chemical process modeling methods rely on physical models and empirical formulas; however, in practical applications, due to system complexity, data noise, and experimental cost limitations, relying solely on physical models often fails to fully understand and optimize the process. In recent years, causal inference has been increasingly applied to chemical process modeling. By identifying causal relationships rather than simple correlations, it is possible to more accurately predict the system's response under different interventions and evaluate the effects of different operating conditions. This approach is particularly suitable for solving problems related to key variable selection, parameter tuning, and uncertainty analysis.

[0003] Traditional methods for causal inference mainly include randomized controlled trials (RCTs), matching methods, regression analysis, instrumental variable methods, differencing methods, and propensity score matching. Each method has its advantages; for example, RCTs have high intrinsic validity, instrumental variable methods can address endogeneity issues, and differencing methods are suitable for observational data. However, they also have many limitations, such as high cost, ethical concerns, overly stringent model assumptions, insufficient control over confounding factors, and poor performance on high-dimensional data. Furthermore, these methods typically rely on strong assumptions (such as linear relationships and independence), making it difficult to handle the complexities of the real world. They are also sensitive to unobserved confounding factors or external shocks, potentially leading to biased causal inference results or insufficient generalization. Summary of the Invention

[0004] Based on the above background, this invention proposes a method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model. Based on causal inference theory, and combining a potential outcome framework and a surrogate model, this method constructs an efficient and accurate causal modeling method for chemical processes. While improving model accuracy, this method also considers computational efficiency and operability in the application of chloride process titanium dioxide production, and has high practical value.

[0005] To achieve the above effects, the technical solution adopted by the present invention is as follows:

[0006] The first aspect of this invention provides a causal modeling optimization method for the chloride process titanium dioxide production of titanium dioxide based on a neural network surrogate model, comprising the following sub-steps:

[0007] S1: Obtain the production process data of titanium dioxide produced by the chloride process, select the product output variable in the production process data of titanium dioxide produced by the chloride process as the target variable, and use the recursive feature elimination method based on cross-validation to eliminate features of the production process data of titanium dioxide produced by the chloride process to obtain the confounding factor features related to the target variable.

[0008] S2: After obtaining the characteristics of confounding factors, causal modeling is performed on the target variable and the characteristics of confounding factors to estimate the causal benefits of the target variable. At the same time, a rule-constrained four-layer multilayer perceptron (MLP) neural network surrogate model is established to replace the causal model building process.

[0009] S3: Use the confounding features and operational variables obtained in S1 as input to the surrogate model, the target variable as output to the surrogate model, and train the surrogate model using a loss function with added rule constraints until the loss function converges and the number of iterations is reached, thus obtaining the trained surrogate model.

[0010] S4: Copy the two surrogate models trained in S3, and connect the outputs of the two surrogate models with a fully connected layer. The two surrogate models are given the same confounding variables and different operational variables. The fully connected layer can output the causal effect on the target variable caused by the change of the operational variables.

[0011] A further preferred embodiment is a method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model, the method specifically including the following sub-steps:

[0012] S1: Obtain process data for the chloride process titanium dioxide production flow, including 19 process measurement variables and 14 process operation variables. Select the product output variable in the process flow as the target output result of the study, and use recursive feature elimination based on cross-validation to obtain confounding factor features that are strongly correlated with the target output result.

[0013] S2: Causal modeling is performed using a potential outcome architecture, and a surrogate model is introduced for optimization; the surrogate model directly obtains the potential outcomes of each operational variable on the target output to evaluate causal benefits.

[0014] S3: The surrogate model is modeled using a four-layer multilayer perceptron (MLP) neural network with rule constraints. It is trained by mapping the input confounding features to the output target result. During model construction, rule constraints and physical information are introduced to enhance the model's reliability and generalization ability. During training, the input is the confounding features obtained in step S1, and the output is the corresponding target output result. Predictive regression is used to obtain the target output result.

[0015] S4: An MLP proxy model with different operational variables was designed. By integrating rule constraints and combining the outputs of the two models with a fully connected layer, the target causal benefits were calculated, thereby achieving accurate modeling and evaluation of the causal relationship in the chloride process of titanium dioxide production.

[0016] In step S1, real data on the production process of titanium dioxide using the chloride process is obtained and categorized by operational variable: the normal label is 0, and different operational variables are categorized by labels 1, 2, 3, 4, and 5. The total number of samples is 10,000.

[0017] TiO2 particle size distribution was selected as the target variable for evaluating causal benefits. The genetic algorithm and greedy algorithm in the recursive feature elimination algorithm were used to obtain the importance of each feature to the target variable. Then, features with low importance to the target variable were eliminated. Finally, variables such as raw material flow rate, titanium ore purity, coke particle size distribution, chlorine flow rate, reactor temperature, titanium tetrachloride (TiCl4) formation rate, TiCl4 purity, hydrochloric acid (HCl) recovery rate, feeder speed, raw material preheating temperature, chlorine feed temperature, coke addition amount, and reactor stirring speed were selected as confounding factors.

[0018] A rule-constrained four-layer multilayer perceptron (MLP) neural network surrogate model was trained. During model training, the different operational variables in step 2 and the confounding factor features obtained in step 3 were used as inputs to the surrogate model, and the target variable selected in step 3 was used as the output. The root mean square error combined with the prediction confidence and the physical information penalty term were used as the loss function, and R0 was selected. 2 The coefficient of determination is used as a performance metric for surrogate models.

[0019] The prediction confidence level is estimated using a Bayesian neural network to assess the uncertainty of the model's predictions. Samples with high uncertainty have lower confidence levels, while samples with low uncertainty have higher confidence levels. The confidence level is represented by α. The physical information penalty term is the mathematical expression of the chloride process titanium dioxide production model, which incorporates the actual physical laws governing the chloride process. The gradient descent method is used to minimize the total loss function, which includes the root mean square error combining the prediction confidence level and the physical information penalty term. Represented as:

[0020]

[0021] Where y represents the actual TiO2 particle size distribution output data; This represents the output data of the TiO2 particle size distribution predicted by the model; α is the confidence level; β is the weighting coefficient of the physical loss. This is the root mean square error term; This is a penalty item for physical information rules.

[0022] The trained surrogate model is used to estimate the causal benefit. The method involves designing two surrogate models with different operands: one for when the operand is normal, and the other for when the operand is faulty. A fully connected layer is trained, whose input is the output of the two surrogate models, and whose output is the difference between the two surrogate model outputs. The difference between the two surrogate model outputs represents the causal benefit to the target variable when a fault occurs, expressed as:

[0023] CE=Y intervention -Y no_intervention

[0024] Where CE represents causal benefit; Y intervention For output during a fault; Y no_intervention This is the output under normal conditions.

[0025] A second aspect of the present invention provides an electronic device including a memory and a processor, the memory being coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model.

[0026] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model.

[0027] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in:

[0028] By combining causal inference with surrogate models, this study achieves accurate modeling of complex causal relationships in the chloride process of titanium dioxide production. This overcomes the shortcomings of traditional causal inference methods in process design, enabling rapid causal benefit analysis of variables under different operational variables. Through optimizing the potential outcome framework, the surrogate model can generate corresponding causal benefit estimates for different fault types, providing important references for process optimization, fault detection, and fault diagnosis, and possessing broad practical application value. Attached Figure Description

[0029] Figure 1 This is a schematic diagram showing the distribution of normal and faulty data.

[0030] Figure 2 This is a schematic diagram of the feature elimination process of the method of the present invention;

[0031] Figure 3This is a schematic diagram of a proxy model used to replace part of the process in the production of titanium dioxide via the chloride process.

[0032] Figure 4 The performance of the method is optimized by modeling the potential outcome architecture based on the neural network agent model;

[0033] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, design concepts, and technical solutions of the embodiments of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings.

[0035] Reference Figures 1-4 This invention proposes a method for constructing a causal model of the chloride process for titanium dioxide production based on a neural network surrogate model. The method includes the following steps:

[0036] Step S1: Obtain process data for the production of titanium dioxide using the chloride process, such as... Figure 1 As shown, it includes normal sample data and fault 1 to 5 sample data. Normal data is labeled with the normal type, and fault data is labeled with the fault type. The operation variable is different fault types.

[0037] The dataset originates from the production process of titanium dioxide using the chloride process, which mainly comprises five core units: a chlorination reactor, a condenser, a gas-liquid separator, a compressor, and an oxidation reactor. It also includes a tail gas absorption tower and a supporting circulation system. The entire process uses titanium ore (such as rutile or high-titanium slag) and chlorine as the main raw materials, supplemented by coke as a reducing agent. Several intermediate materials and byproducts are also involved in the process.

[0038] The main reactants include solid titanium ore and gaseous chlorine, which react to produce titanium tetrachloride (TiCl4) and byproducts ferric chloride (FeCl3) and silicon chloride (SiCl4). Titanium tetrachloride is converted into titanium dioxide (TiO2) in the downstream oxidation stage, while hydrogen chloride gas is produced as a byproduct; after absorption, it forms hydrochloric acid solution. The feed also contains small amounts of inert substances, such as nitrogen and unreacted carbon dioxide.

[0039] The core reaction equation for the entire process is as follows:

[0040] TiO2(s)+2C(s)+2Cl2(g)→TiCl4(g)+2CO(g) (1)

[0041] TiCl4(g)+O2(g)→TiO2(s)+2Cl2(g) (2)

[0042] The main process of titanium dioxide production via the chloride process includes the following steps: First, in the chlorination reaction stage, titanium ore, coke, and chlorine react at high temperature in a chlorination reactor to produce TiCl4 and byproducts. Subsequently, the mixed gas generated after the reaction is cooled by a condenser, and liquid TiCl4 is separated in a gas-liquid separator. The remaining gas is either recycled via a compressor or partially discharged as byproducts. In the TiCl4 refining and oxidation stage, liquid TiCl4 is distilled to remove impurities and then fed into an oxidation reactor to react with oxygen to produce TiO2 powder. Simultaneously, the byproduct chlorine is recycled back to the chlorination process. Finally, hydrogen chloride in the tail gas reacts with water in an absorption tower to produce hydrochloric acid, which is collected, stored, and then sold. In this process dataset, the sample data includes measured variables and operational variables. Measured variables include flow rate, temperature, pressure, and component concentration, such as chlorine flow rate, reactor temperature, and titanium tetrachloride level. Operational variables include raw material feed rate and condenser cooling water flow rate. In addition, the process exhibits several typical failure types, such as excessively high chlorination reactor temperature, which may lead to increased byproducts or reduced main product efficiency; blockage of the gas-liquid separator, causing system pressure fluctuations; and decreased condenser cooling efficiency, resulting in incomplete condensation of high-temperature gases. Each failure type and its occurrence time are recorded in detail in the sample dataset for analysis and optimization of the chlorination process for titanium dioxide production.

[0043] The dataset was obtained from actual chemical processes and consists of 33 dimensions, including five types of fault anomalies.

[0044] 1. For each of the five fault scenarios, select 1000 normal and fault data points.

[0045] 2. Save all data and categorize it by label: 0 for normal data, 1 for fault 1, and so on. The total number of samples is 10,000.

[0046] 3. Preprocess the data by performing normalization. The operation is as follows:

[0047]

[0048] Where, x new x represents the normalized data; x represents the data before normalization; x max x represents the maximum value in the data. min It is the minimum value in the data.

[0049] 4. The TiO2 particle size distribution is selected as the target output. Recursive feature elimination based on cross-validation is used to obtain confounding features strongly correlated with the target output. Redundant and low-correlation features increase the complexity of causal inference and reduce its accuracy. Therefore, features with a significant impact on the target output should be selected. While ensuring the accuracy of causal inference, feature selection can reduce the dimensionality of the dataset and alleviate complexity. Recursive feature elimination based on cross-validation (RFECV) is a greedy algorithm designed to find the optimal subset of features. The process is as follows: Figure 2 As shown. First, all features in the sample set are retrieved. A classifier is built and trained, and the importance of each feature is calculated. Features with lower importance are then removed. The remaining features are iteratively retrieved until all features have been traversed. The desired number of features is retained to obtain a feature subset. Two models, genetic algorithm and greedy algorithm, are selected for feature selection. The judgment of feature importance is related to the dataset. When some features are eliminated, the model will recalculate the importance of each feature based on the new dataset. Therefore, the RFECV process is more accurate than training a single model and selecting based on the feature importance of the results. This accuracy is reflected in the fact that each time a feature is eliminated, the remaining features are compared in the new environment. Using this recursive feature elimination algorithm, the following contamination factors are selected: raw material flow rate, titanium ore purity, coke particle size distribution, chlorine flow rate, reactor temperature, titanium tetrachloride formation rate, TiCl4 purity, HCl recovery rate, feeder speed, raw material preheating temperature, chlorine feed temperature, coke addition amount, and reactor stirring speed.

[0050] Step S2: Perform causal modeling using a latent outcome architecture, a causal inference method used to estimate the causal benefits of an intervention or operation on a system. In the chloride process for titanium dioxide production, the failure type is set as the operational variable, and the latent outcomes are defined for both intervention and non-intervention scenarios.

[0051] The modeling method is as follows:

[0052] S201. Determine the target output: Select the variable that has the most significant impact on system performance as the target output (such as titanium ore purity, chlorine flow rate, etc.).

[0053] S202. Define intervention strategies: Simulate system behavior under intervention and non-intervention by setting different types of control faults.

[0054] S203. Data collection: Based on simulation models or actual operating data, record the potential output of the system under intervention and non-intervention states.

[0055] S204. The final causal benefit is obtained by calculating the difference between the potential outputs under intervention and non-intervention states, using the following formula:

[0056] CE=Y intervention -Y no_intervention (4)

[0057] Where CE represents causal benefit; Y intervention The state after intervention; Y no_intervention This represents a state of no intervention.

[0058] Due to the high complexity of actual processes, directly modeling potential outcomes may suffer from insufficient data or high computational costs. Therefore, surrogate models are introduced as approximation tools to quickly predict potential outcomes. For example... Figure 3 As shown, this process is a flowchart for producing titanium dioxide from hydrochloric acid, a byproduct of the chloride process. When a fault is triggered, we can use a proxy model to replace part of the process and directly obtain the potential results of the titanium dioxide product after the fault occurs.

[0059] Using MLP neural networks as surrogate models is suitable for modeling complex chemical processes due to their powerful capabilities in nonlinear mapping.

[0060] Step 3: The surrogate model employs a four-layer MLP neural network: the input layer receives the relevant feature variables obtained from S1. The hidden layers (2 layers) each contain an appropriate number of neurons (128 and 64), using the ReLU activation function. The output layer predicts the TiO2 particle size distribution, also using the ReLU activation function. The data is divided into training and validation sets, and the model is trained using K-fold cross-validation.

[0061] K-fold cross-validation is a commonly used method for evaluating model performance. First, it divides the data into K mutually exclusive subsets (folds), ensuring that each subset serves as the validation set in one iteration, while the other K subsets serve as the training set. This fully utilizes the data and improves the stability and reliability of the evaluation results. Second, K-fold cross-validation effectively reduces evaluation bias caused by uneven data partitioning, providing a more accurate estimate of model performance.

[0062] A custom loss function is defined to comprehensively consider data error, physical consistency constraints, and confidence assessment in a weighted manner. The coefficient of determination (R-squared) is used as the model performance metric, and a composite loss function of the following form is constructed:

[0063]

[0064]

[0065] Where y represents the actual TiO2 particle size distribution output data; This represents the output data of the TiO2 particle size distribution predicted by the model; n represents the number of output data; y i, Let represent the output data of the i-th actual TiO2 particle size distribution and the output data of the TiO2 particle size distribution predicted by the model, respectively; α is the confidence level; β is the weighting coefficient of the physical loss. This is the root mean square error term; This is a penalty item for physical information rules.

[0066] Introducing physical information and confidence scores into the loss function can significantly improve the predictive performance and practical application value of neural network models. First, physical information constraints ensure that the model's predictions conform to scientific laws (such as conservation laws and dynamic equations), effectively avoiding results that violate physical principles and improving the model's reliability and accuracy. Second, guided by physical information, the model can efficiently learn key relationships even with scarce or incomplete data, thus reducing its dependence on large-scale, high-quality data. Simultaneously, the confidence score evaluation introduces uncertainty weights into the prediction results, enabling the model to impose stricter constraints on low-confidence predictions, thereby reducing the impact of uncertainty on the overall results. The confidence score is estimated using a Bayesian neural network to assess the uncertainty of the model's predictions; samples with high uncertainty have lower confidence scores, and samples with low uncertainty have higher confidence scores. Furthermore, the combination of physical information and confidence scores significantly enhances the model's generalization ability, enabling it to provide stable predictions even under unknown conditions or new scenarios. This design also optimizes the training process, improving the model's convergence speed by balancing data error, physical constraints, and confidence scores. The closer the loss function is to 0, the closer the model's predictions are to the true values. In this invention, R is selected 2 Both the coefficient of determination and the absolute error (MAE) characterize the model accuracy, and the formulas are as follows:

[0067]

[0068] Where y represents the actual TiO2 particle size distribution output data; This represents the output data of the TiO2 particle size distribution predicted by the model; n represents the number of actual output data; y i , These represent the output data of the actual TiO2 particle size distribution and the output data of the TiO2 particle size distribution predicted by the model, respectively. This represents the mean of the actual output data.

[0069] Regression squared coefficient of determination (R²) 2Based on the definition of the residual sum of squares, the R-squared can quantify the model's explanatory power for the target variable. Its value is typically between 0 and 1, making it easy to understand and compare. It weights the error using a squared form, making it more sensitive to large errors and thus helping the model focus on extreme prediction problems. Furthermore, the coefficient of determination is directly compatible with the least squares method and can reflect the proportion of change in the target output explained by the model, possessing good interpretability. Meanwhile, as a standardization metric, R-squared... 2 It is insensitive to the units and scale of the data, making it suitable for datasets of different sizes. Its formula can also decompose the error into explained and unexplained components, facilitating the diagnosis of model problems.

[0070]

[0071] Where MAE represents mean absolute error; n represents the number of true output data; y i , These represent the output data of the actual TiO2 particle size distribution and the output data of the TiO2 particle size distribution predicted by the model, respectively.

[0072] MAE has the following advantages as an evaluation metric for predictive models: its result unit is consistent with the target variable, making it easy to interpret; it is insensitive to outliers and will not amplify the impact of large errors; it is applicable to any data distribution and is more robust to skewed data; it can directly measure the average error of the model's predictions, fairly reflecting the overall predictive performance.

[0073] Step 4: Designed MLP proxy models with different output operation variables, as follows:

[0074] S401, Normal Output Proxy Model: The input is the confounding factor features under normal conditions, and the output is the target output result under normal conditions.

[0075] S402: Fault Output Proxy Model: The input is the characteristics of confounding factors under fault conditions, and the output is the target output result under fault conditions.

[0076] The outputs of the trained normal output proxy model and the fault output proxy model are integrated through a fully connected layer, and the target causal benefit is calculated based on the outputs of the normal and fault models.

[0077] This invention uses a causal modeling example in the chloride process for titanium dioxide production to illustrate the effectiveness of a causal benefit assessment method based on a potential outcome architecture and a proxy model. Figure 4 The performance of the causal benefit evaluation method based on the potential outcome architecture and the proxy model is demonstrated. Figure 4(a) The iteration rounds of the validation loss for each fold show the decreasing trend of the loss function of various methods during the training of the surrogate model. The straight line is the training loss of fold 1, the long dashed line is the training loss of fold 2, the dashed line with dots is the training loss of fold 3, and the short dashed line is the training loss of fold 4. Figure 4 (b) The curves showing the predicted and actual values ​​demonstrate the prediction results output by the surrogate model proposed in this invention. The results show that the predicted values ​​can roughly fit the actual values.

[0078] In modeling the complex chlorination process of titanium dioxide production, directly using traditional causal inference methods (such as latent outcome architectures) for causal benefit assessment results in high computational complexity and significant dependence on system parameters. The method proposed in this invention effectively reduces computational complexity by introducing a surrogate model, making causal inference more efficient. Furthermore, the method exhibits strong performance under various surrogate models (such as multilayer perceptrons and long short-term memory networks). Experimental results show that the causal benefit assessment method based on latent outcome architectures and surrogate models proposed in this invention significantly improves the accuracy of causal benefit prediction and reduces the computational costs of model training and inference, demonstrating significant advantages in training time and result quality, especially considering the complex causal relationships in the chlorination process of titanium dioxide production.

[0079] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the above-described method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model. Figure 4 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in an embodiment of the present invention to construct a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0080] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the aforementioned method for constructing a causal model of the chloride process titanium dioxide production process based on a neural network surrogate model. The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0081] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0082] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for process causal modeling optimization of chlorination process of titanium dioxide based on a neural network proxy model, characterized in that, The method comprises the following sub-steps: S1: Obtain chlorination titanium dioxide production process data, select a fault type in the chlorination titanium dioxide production process data as an operation variable, select a product output variable in the chlorination titanium dioxide production process data as a target variable, perform feature elimination on the chlorination titanium dioxide production process data based on a recursive feature elimination method based on cross-validation, and obtain a mixed factor feature; The product output variable is a TiO2 particle size distribution; S2: Constrain a four-layer multilayer perception neural network through confidence and physical information rules to establish a proxy model; S3: Take the mixed factor feature and the operation variable obtained in S1 as proxy model inputs, take the target variable as a proxy model output, and perform proxy model training by using a loss function constrained by confidence and physical information rules until the loss function converges and an iteration number is reached, to obtain a trained proxy model, which is copied into two, as a first proxy model and a second proxy model; The mixed factor variable input into the first proxy model and the second proxy model is the same, the operation variable input into the first proxy model and the second proxy model is different, and the output of the first proxy model and the output of the second proxy model are connected by a full connection layer and output, and the full connection layer outputs a causal effect of the target variable caused by a change in the operation variable; The operation variable input into the first proxy model and the second proxy model is different, and specifically includes: The input of the first proxy model is a normal operation variable, and the input of the second proxy model is a fault operation variable; The causal effect of the target variable caused by a change in the operation variable output by the full connection layer specifically includes: The difference between the outputs of the two proxy models is the causal effect of the target variable caused by a fault, and is expressed as: ； wherein is the causal effect; is the output at failure; is the output at normal.

2. The method of claim 1, wherein the method is a method of neural network proxy model-based chlorination process of titanium dioxide production process causal modeling optimization. In step S1, the recursive feature elimination method based on cross-validation is used to eliminate features from the chlorination titanium dioxide production process data to obtain a mixed factor feature, which specifically includes: The genetic algorithm and the greedy algorithm in the recursive feature elimination algorithm are used to obtain the importance of each feature in the chlorination titanium dioxide production process data to the target variable, and irrelevant features to the target variable are removed, and finally, the raw material conveying flow, the titanium ore purity, the coke particle size distribution, the chlorine flow, the reactor temperature, the titanium tetrachloride generation rate, the TiCl4 purity, the hydrochloric acid recovery rate, the feeder speed, the raw material preheating temperature, the chlorine feeding temperature, the coke addition amount, and the reactor stirring speed variable are obtained as the mixed factor feature. 3.The method of claim 1, wherein the method is characterized by, In step S2, the four-layer multilayer perception neural network is constrained by confidence and physical information rules to establish a proxy model, which specifically includes: The confidence is obtained using a Bayesian neural network, and the confidence size is represented by The confidence is obtained using a Bayesian neural network, and the confidence size is represented by The confidence is obtained using a Bayesian neural network, and the confidence size is represented by 4. The method according to claim 3, wherein the method is characterized by, In step S3, the mixed factor feature and the operation variable obtained in S1 are taken as proxy model inputs, the target variable is taken as a proxy model output, and proxy model training is performed by using a loss function constrained by confidence and physical information rules until the loss function converges and an iteration number is reached, to obtain a trained proxy model, which specifically includes: The gradient descent method is used to minimize the total loss function, the total loss function is represented as: ； wherein, represents the true TiO2 particle size distribution output data; represents the model predicted TiO2 particle size distribution output data; is the confidence; is the weight coefficient of physical loss; is the root mean square error term, is the physical information rule penalty term; Until the loss function converges and an iteration number is reached, a trained proxy model is obtained.

5. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled with the processor; wherein the memory is configured to store program data, and the processor is configured to execute the program data to implement the method for neural network agent model based causal modeling optimization of chlorination process of titanium dioxide production process according to any one of claims 1-4.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method for neural network agent model based causal modeling optimization of chlorination process of titanium dioxide production process according to any one of claims 1-4.

Images