Characteristic variable causal relationship discovery system for high-purity carbonate process fault diagnosis process

Abstract

The invention provides a feature variable causal relationship discovery system for a high-purity carbonate process fault diagnosis process, which comprises the following steps of: according to time sequence data of feature variables in the high-purity carbonate process, performing preprocessing noise reduction on the time sequence data through ensemble empirical mode decomposition, training an inference model for inferring a reason variable associated with each feature variable, and obtaining a feature variable causal relationship of the high-purity carbonate process fault diagnosis process; and determining a reason variable of each feature variable by using a parameter value obtained by the reasoning model, eliminating a pseudo causal relationship of the reason variable associated with each feature variable by using an arrangement feature importance strategy, and further realizing the causal relationship of the feature variables in the fault diagnosis process of the high-purity carbonate process through the association between each feature variable and the reason variable.

Description

Technical Field

[0001] This invention relates to a characteristic variable causal relationship discovery system for fault diagnosis in high-purity carbonate processes, belonging to the fields of intelligent manufacturing of chemical processes, monitoring of chemical process operation modes, and process systems engineering. Background Technology

[0002] The efficient and safe operation of chemical processes is crucial for ensuring product quality, reducing energy and material consumption, and minimizing safety accidents. While the development of process control systems and safety instrumented systems has significantly improved the automation level and production efficiency of the chemical industry, complex chemical processes, such as the production of high-purity carbonates, still lack the ability to identify and self-heal under fault conditions. The construction of fault diagnosis systems can promptly detect faults and determine their propagation paths within the chemical process, thereby quickly pinpointing the causes and providing scientific decision-making to eliminate the faults and prevent accidents.

[0003] The purpose of causal discovery is to uncover the causal relationships between variables in chemical processes, thereby constructing a causal relationship network to help operators and decision-makers efficiently and quickly identify and resolve faults. Causal relationship methods can be divided into two categories: methods based on prior knowledge and data-driven methods. Methods based on prior knowledge rely on prior knowledge of the process to construct causal relationships between characteristic variables, mainly including symbolic directed graphs and expert systems. However, these methods, largely based on process knowledge and expert experience, suffer from drawbacks such as strong subjectivity of expert knowledge and the need for regular maintenance and updates to expert systems and knowledge bases. Data-driven methods determine causal relationships based on time-series data of characteristic variables, primarily using Granger causality analysis, transfer entropy, Bayesian networks, and convergent cross-mapping. However, these methods are ineffective for inferring causal relationships between process characteristic variables in chemical processes with strong nonlinearity, strong coupling of characteristic variables, and complex causal driving factors, such as the high-purity carbonate process system for lithium battery electrolyte solvents. Deep learning models, capable of fully mining the potential relationships between variables in high-dimensional time-series data and revealing the causal relationships between process characteristic variables at the time of fault formation, have become an effective approach for constructing causal relationship inference models for characteristic variables. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing methods for inferring causal relationships of characteristic variables in the fault diagnosis of high-purity carbonate processes, and proposes a system for discovering causal relationships of characteristic variables in the fault diagnosis process of high-purity carbonate processes. This invention uses time-series data of process characteristic variables as samples, trains a causal-gated time-series Transformer model to identify causal variables associated with the characteristic variables, and establishes the relationships between all characteristic variables from the perspective of the process system by training trend prediction models for each characteristic variable.

[0005] A system for discovering causal relationships of characteristic variables in the fault diagnosis process of high-purity carbonate production includes the following steps:

[0006] (1) Obtain the time series data of process characteristic variables of high-purity carbonate process, and use ensemble empirical mode decomposition to denoise the time series data;

[0007] (2) Study the time series X of the characteristic variable of the j-th process. j When considering the causal variables, construct a causal gated time series Transformer model G. j Using n variables as input predictor variables X j X j The corresponding causal gate vector uses a one-dimensional vector c. j =[c 1,j ,c 2,j ,...,c n,j Each causal gate value corresponds to several potential causal variables;

[0008] (3) The n causal gate values ​​are converted to values ​​between 0 and 1 using the Softmax function, c′ j =softmax([c 1,j ,c 2,j ,...,c n,j ]);

[0009] (4) c j The elements in X' are multiplied one by one with each of the n variables, and the output is fed into a time series Transformer network to predict the target variable X'. j , X′ j =TST([c′ 1,j ⊙X1,c′ 2,j ⊙X2,...,c′ n,j ⊙X n ])+error;

[0010] (5) For each of the n process feature variables, a causal-gated time series Transformer model needs to be trained independently to discover the causal variable. After the n trained causal-gated time series Transformer models are completed, the causal gate vectors of all target variables are combined to form a causal gate matrix C. If the elements c in the causal gate matrix C are... i,j >1, then X i Determined as X j Potential causal variables;

[0011] (6) Use the method of arranging feature importance to eliminate spurious causal relationships between feature variables.

[0012] This invention provides a system for discovering the causal relationship of characteristic variables in the fault diagnosis process of high-purity carbonate production, the features and beneficial effects of which are as follows:

[0013] This invention addresses a series of important characteristic variables characterizing the process operation status during fault diagnosis in high-purity carbonate processes. Based on time-series data denoising using ensemble empirical mode decomposition, a causal-gated time-series Transformer model is used to train the prediction model for each characteristic variable. Causal threshold vectors are used to identify potential causal variables associated with important characteristic variables, and a method of arranging feature importance is employed to eliminate spurious causal relationships between process characteristic variables. Finally, a causal relationship network between various characteristic variables is constructed from a global perspective of the high-purity carbonate process system. This enables fault tracing and fault propagation path determination during carbonate process fault diagnosis. This invention can solve the problem of discovering causal relationships among characteristic variables in the fault diagnosis process of high-purity carbonate processes characterized by strong nonlinearity, strong coupling of characteristic variables, and complex causal driving factors. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the characteristic variable causal relationship discovery system in the high-purity carbonate process fault diagnosis process of the present invention;

[0015] Figure 2 This is a high-purity carbonate production process with a PI control loop;

[0016] Figure 3 A causal-gated time series Transformer model for inferring the correlation between characteristic variables and causal variables in the fault diagnosis process of high-purity carbonate production;

[0017] Figure 4 This is a threshold iteration diagram of the causal relationship of a reactive distillation column unit during the fault diagnosis process of high-purity carbonate production.

[0018] Figure 5 A method for eliminating spurious causal relationships by arranging the importance of process features in the fault diagnosis of high-purity carbonate processes;

[0019] Figure 6 The threshold adjacency matrix of reactive distillation column units is used to eliminate falsified causal relationships during the fault diagnosis process of high-purity carbonate processes using the eigenvalue importance method.

[0020] Figure 7 This is a causal relationship diagram obtained from the characteristic variables of the reactive distillation column in the fault diagnosis process of high-purity carbonate process according to the present invention;

[0021] Figure 8 This is a causal relationship diagram obtained from the characteristic variables of the dimethyl carbonate and methanol azeotropic separation and extraction distillation column in the fault diagnosis process of high-purity carbonate process according to the present invention.

[0022] Figure 9 This is a causal relationship diagram obtained from the characteristic variables of the dimethyl carbonate and methanol azeotrope separation extractant recovery tower in the fault diagnosis process of high-purity carbonate process according to the present invention.

[0023] Figure 10 This is a causal relationship diagram obtained from the characteristic variables of the dimethyl carbonate recovery tower in the fault diagnosis process of high-purity carbonate process according to the present invention;

[0024] Figure 11 This is a causal relationship diagram obtained by the present invention for the characteristic variables of the methyl ethyl carbonate and diethyl carbonate separation tower in the fault diagnosis process of high-purity carbonate process. Detailed Implementation

[0025] This invention proposes a system for discovering causal relationships of characteristic variables in the fault diagnosis process of high-purity carbonate production, specifically including the following steps:

[0026] (1) Obtain the time series data of process characteristic variables of high-purity carbonate process, and use ensemble empirical mode decomposition to denoise the time series data;

[0027] (2) Study the time series X of the characteristic variable of the j-th process. j When considering the causal variables, construct a causal gated time series Transformer model G. j Using n variables as input predictor variables X j X j The corresponding causal gate vector uses a one-dimensional vector c. j =[c 1,j ,c 2,j ,...,c n,j Each causal gate value corresponds to several potential causal variables;

[0028] (3) The n causal gate values ​​are converted to values ​​between 0 and 1 using the Softmax function, c′ j =softmax([c 1,j ,c 2,j ,...,c n,j ]);

[0029] (4) c j The elements in X' are multiplied one by one with each of the n variables, and the output is fed into a time series Transformer network to predict the target variable X'. j , X′ j =TST([c′ 1,j ⊙X1,c′ 2,j ⊙X2,...,c′ n,j ⊙X n])+error;

[0030] (5) For each of the n process feature variables, a causal-gated time series Transformer model needs to be trained independently to discover the causal variable. After the n trained causal-gated time series Transformer models are completed, the causal gate vectors of all target variables are combined to form a causal gate matrix C. If the elements c in the causal gate matrix C are... i,j >1, then X i Determined as X j Potential causal variables;

[0031] (6) The method of arranging feature importance is used to eliminate the spurious causal relationship between feature variables, and a causal relationship network between each feature variable is constructed from each feature variable.

[0032] Referring to the accompanying drawings, a specific embodiment of the present invention is as follows:

[0033] (1) The process of producing high-purity carbonate, such as Figure 2 As shown, the system includes a reactive distillation column C1, an extractive distillation column C2, an extractant recovery column C3, a dimethyl carbonate recovery column C4, and a methyl ethyl carbonate and diethyl carbonate separation column C5. Dimethyl carbonate and ethanol, as feedstocks, enter the reactive distillation column C1 at a molar ratio of 1:1.2. A continuous transesterification reaction occurs in the reactive distillation column to produce methyl ethyl carbonate, diethyl carbonate, and methanol as a byproduct. Dimethyl carbonate and methanol form an azeotrope, which is collected from the top of the reactive distillation column C1. The azeotrope of dimethyl carbonate and methanol, along with the extractant aniline, enters the extractive distillation column C2 for extractive distillation separation. The mixture of dimethyl carbonate and aniline collected from the bottom of the extractive distillation column C2 enters the extractant recovery column C3. Dimethyl carbonate is recovered from the top of the extractant recovery column C3 and recycled back to the reactive distillation column C1. Aniline collected from the bottom of recovery tower C3 is recycled to extractive distillation tower C2; a mixture of dimethyl carbonate, methyl ethyl carbonate, and diethyl carbonate collected from the bottom of reactive distillation tower C1 enters dimethyl carbonate recovery tower C4. Dimethyl carbonate is collected from the top of dimethyl carbonate recovery tower C4 and recycled to reactive distillation tower C1. A mixture of methyl ethyl carbonate and diethyl carbonate collected from the bottom of dimethyl carbonate recovery tower C4 enters methyl ethyl carbonate and diethyl carbonate separation tower C5. Methyl ethyl carbonate is collected from the top of methyl ethyl carbonate and diethyl carbonate separation tower C5, and diethyl carbonate is collected from the bottom of methyl ethyl carbonate and diethyl carbonate separation tower C5. The PI control loop of the high-purity carbonate process includes 6 temperature controllers, 10 level controllers, 5 pressure controllers, and 2 flow controllers. The high-purity carbonate process includes 55 process characteristic variables, the specific meanings of which are shown in Figure 1.

[0034] Table 1 Characteristic variables of high-purity carbonate process

[0035] The feature variable dataset contains 3000 sampling time steps, each with a sampling time step of 0.01 h. The first 200 time steps are stationary data, while the last 2800 time steps are perturbation data under the change of dimethyl carbonate feed rate. The time series data of the feature variables are subjected to ensemble empirical mode decomposition for data denoising and normalization. The processed data are then processed by sliding window with a sliding window length of 10, and the training set and validation set of the feature variable prediction model are constructed in an 8:2 ratio.

[0036] For the process characteristic variables of the reactive distillation column in the high-purity carbonate process, the following methods are used: Figure 3 The causal-gated time series Transformer model shown is trained, with one model trained for each feature variable. Historical window data from all feature variables is used to predict future window data for that variable. The causal-gated time series Transformer model has a dimension of 76, a self-attention mechanism head of 4, an encoder layer of 3, and 100 training epochs. After training, the magnitudes of the elements in the causal gate vector are used to determine the potential causal variables of the feature variables. Figure 4 This is a threshold iteration diagram of the causal relationship of the reactive distillation column unit during the fault diagnosis process of high-purity carbonate.

[0037] Using the importance of arrangement features, such as Figure 5 To eliminate spurious causal relationships between feature variables, the potential causal variables are randomly rearranged along the time dimension and then used as a new validation set. This new set is then input into the trained causal-gated time series Transformer model. By comparing the loss function value with that of the un-randomized potential causal variables, if the loss function value of the un-randomized potential causal variables and the loss function value obtained by randomly rearranging the potential causal variables along the time dimension are both less than the threshold of 0.4, then the potential variables are considered true causal variables. Figure 6 This is a gate adjacency matrix for reactive distillation column units, which eliminates falsified causal relationships during fault diagnosis of high-purity carbonate processes using a feature importance method.

[0038] By performing empirical mode decomposition of the dataset on 55 process characteristic variables of each unit module in the high-purity carbonate process using the above steps, training the feature variable causal gated time series Transformer model to determine latent variables, and using the ranking feature importance method to identify real variables, the causal relationship among the 55 process characteristic variables in the high-purity carbonate process can be determined.

Claims

1. A system for discovering causal relationships of characteristic variables in the fault diagnosis process of high-purity carbonate production, characterized in that... Causal-gated time series Transformer models were used to discover causal relationships among process characteristic variables in fault diagnosis of high-purity carbonate production processes. (1-1) Obtain the time series data of process characteristic variables of high-purity carbonate process, and use ensemble empirical mode decomposition to denoise the time series data; (1-2) Study the time series X of the characteristic variable of the j-th process. j When considering the causal variables, construct a causal gated time series Transformer model G. j Using n variables as input predictor variables X j X j The corresponding causal gate vector uses a one-dimensional vector c. j =[c 1,j ,c 2,j ,...,c n,j Each causal gate value corresponds to several potential causal variables; (1-3) The n causal gate values ​​are converted to values ​​between 0 and 1 using the Softmax function, c′ j =softmax([c 1,j ,c 2,j ,...,c n,j ]); (1-4) c j The elements in the matrix are multiplied one by one with each of the n variables, and the output is fed into a causal gated time series Transformer network to predict the target variable X. j ′, X j =TST([c′) 1,j ⊙X1,c′ 2,j ⊙X2,...,c′ n,j ⊙X n ])+error; (1-5) For each of the n process feature variables, a causal-gated time series Transformer model needs to be trained independently to discover the causal variable. After the n causal-gated time series Transformer models are trained, the causal gate vectors of all target variables are combined to form a causal gate matrix C. If the elements c in the causal gate matrix C are... i,j >1, then X i Determined as X j Potential causal variables; (1-6) The method of arranging feature importance is used to eliminate spurious causal relationships between feature variables.

2. A system for discovering causal relationships of characteristic variables in the fault diagnosis process of high-purity carbonate production, characterized in that... The model described in claim 1 can discover the causal relationship between characteristic variables of the high-purity carbonate process, thereby enabling the construction of a relationship network between various process characteristic variables from a global perspective of the high-purity carbonate process system, and thus tracing the source of fault diagnosis in the high-purity carbonate process.

Images