Abnormal variable identification method and device based on causal-embedded reconstruction network

Abstract

The invention discloses an abnormal variable identification method and device based on a causal-embedded reconstruction network. The method comprises the steps of constructing the causal-embedded reconstruction network, setting network structure parameters, randomly initializing connection parameters of the network structure parameters and defining a loss function; inputting the preprocessed historical data under the normal working condition and a causal relationship network constructed based on process knowledge into a causal-embedding reconstruction network, and converting multivariable time sequence data of the chemical process into graph structure data with a causal relationship; calculating an abnormal score of each variable of the verification set monitoring data, and calculating an abnormal score threshold value of each variable under a given confidence degree; and identifying an abnormal variable according to the relationship between the abnormal score of the variable and a threshold value during online monitoring. According to the abnormal variable identification method and device based on the causal-embedded reconstruction network, a data attribute and causal relationship dual decoding mechanism is adopted, the pollution effect can be overcome, high-accuracy identification of fault variables under various types of abnormal working conditions is achieved, and the low false alarm rate is ensured.

Description

Technical Field

[0001] The present invention relates to the field of process control in chemical processes, and particularly to the field of fault diagnosis in process control. By using the disclosed method and device for identifying abnormal variables based on a causal-embedded reconstruction network, abnormal variable identification with high accuracy and low false alarm rate can be achieved for various different types of abnormal operating conditions in chemical processes. Background Art

[0002] With the booming development of artificial intelligence and industrial big data technologies, many data-driven process control technologies have been successively proposed and applied to the field of fault diagnosis in chemical processes.

[0003] Fault condition diagnosis requires timely and accurately identifying the variables that are abnormal in process monitoring (i.e., abnormal variable identification); the existing methods for abnormal variable identification mainly include techniques based on multivariate statistical analysis and techniques based on deep learning. The inventor has found through research that the commonly used methods in the existing abnormal variable identification techniques based on multivariate statistical analysis are mainly the method based on contribution degree and the contribution method based on reconstruction, but they have the disadvantage of contamination effect, resulting in inaccurate or incomplete identification of abnormal variables and prone to misjudging normal variables as abnormal variables; while the existing abnormal variable identification techniques based on deep learning also need to use the contribution degree or reconstruction method to identify abnormal variables after fault detection, so they also have the disadvantage of low identification accuracy due to the contamination effect, and the interpretability of the model is poor.

[0004] The information disclosed in this background art section is only intended to increase the understanding of the overall background of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art already known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to improve the accuracy of abnormal variable identification and reduce the false alarm rate.

[0006] The present invention provides a method for identifying abnormal variables based on a causal-embedded reconstruction network, and its steps for offline modeling include:

[0007] S11. Preprocess the original historical data under normal operating conditions, including: data standardization, dividing the training set and the validation set, and using the sliding window technique to create a time series window to construct the first time-delay matrix of the monitoring data in the training set at time k;

[0008] S12. Define the causal relationship between variables based on process knowledge, and use a directed graph to construct the corresponding causal relationship network;

[0009] S13. Set the structural parameters of the causal-embedded reconstruction network, randomly initialize its connection parameters, and define the loss function of the causal-embedded reconstruction network;

[0010] S14. Use the training set and the causal relationship network among variables to jointly train the causal-embedding reconstruction network, and obtain the optimal connection parameters of the causal-embedding reconstruction network by learning the data features under normal working conditions;

[0011] S15. Input the monitoring data of the validation set into the trained causal-embedding reconstruction network, and calculate the anomaly score values of each variable;

[0012] S16. Calculate the anomaly score threshold s of each variable according to the kernel density estimation method under a given confidence level η cl

[0013] Preferably, the online monitoring step in the present invention includes:

[0014] S21. Obtain the real-time operation data in the chemical process and perform data standardization to obtain the input set; according to the sliding window technique, construct the second time-delay matrix of the monitoring data at time k as the online input set of the causal-embedding reconstruction network;

[0015] S22. For each input sample in the online input set, perform feature representation based on the causal-embedding reconstruction network and calculate the anomaly score values of each variable;

[0016] S23. If the anomaly score value of a certain variable exceeds the corresponding anomaly score threshold, then determine the variable as an abnormal variable; otherwise, continue data collection and process monitoring.

[0017] On the other hand of the embodiment of the present invention, there is also provided an abnormal variable identification device based on a causal-embedding reconstruction network. The abnormal variable identification device based on a causal-embedding reconstruction network includes a computer program stored on a medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the methods described in the above aspects and achieves the same technical effects.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] ​Through research, the inventor found that the main reasons for the false alarms and missed detections in the existing abnormal variable recognition technologies include: In the existing abnormal variable recognition technologies based on multivariate statistical analysis, the commonly used methods are mainly the contribution-based method and the reconstruction-based contribution method. These two types of methods are mostly based on latent space projection technology. In the process of dimensionality reduction and reconstruction, the model adopts complex non-linear transformations, that is, the features after projection integrate the information of all input variables. Then, an abnormal change in one input variable may affect all other variables after projection, resulting in the risk that normal variables are misjudged as abnormal variables due to the increase in contribution degree under the influence of abnormal variables when abnormal working conditions occur, and there is a contamination effect. Further, due to the application of non-linear transformations in the latent space projection process, it is difficult to intuitively establish the mapping relationship between the low-dimensional features of the projection space and the original high-dimensional input data, and there is a disadvantage that the contamination effect is difficult to eliminate. The existing abnormal variable recognition technologies based on deep learning use the contribution degree or reconstruction method to identify abnormal variables after fault detection, so there is also the disadvantage of low identification accuracy due to the contamination effect. Further, due to the non-linear feature extraction of the deep network, the interpretability of the model is poor.

[0020] When determining abnormal variables, the existing abnormal variable recognition technologies will select several variables with the largest contribution degree as abnormal variables, which is somewhat subjective and cannot quantitatively determine how much the contribution of a variable should be judged as abnormal. More importantly, when some variables in the process have small abnormal changes or slow drifts, their contribution degrees may be relatively small, but they are also abnormal variables, which may lead to missed detections of abnormal variables. In short, the abnormal variable recognition methods in the existing technologies are prone to false alarms or missed detections, and the recognition results are not comprehensive and accurate enough.

[0021] Based on the above understanding, the present invention proposes a new abnormal variable recognition method: an abnormal variable recognition method based on a causal-embedded reconstruction network. Using a graph deep network to transform the multivariate variable time series data of a chemical process into graph-structured data, and extract the dynamic and spatial relationships between variables; due to the embedding of the causal relationship network, the features in the hidden space are restricted to be generated only based on the variables that have a causal relationship with it; by simultaneously reconstructing the variable time series data and the causal relationship network between variables, the contamination effect can be overcome, and high-accuracy and low-false-alarm-rate abnormal variable recognition can be achieved for various different types of abnormal working conditions, especially for scenarios such as small-amplitude changes and slow drifts.

[0022] Through the present invention, when a fault occurs, the abnormal variables and their abnormal degrees can be accurately presented to the operator, which has important practical significance for the transparent process diagnosis of the actual chemical process.

[0023] The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly and be implemented according to the content of the specification, and in order to make the above and other objects, technical features and advantages of the present invention more understandable, one or more preferred embodiments are listed below and described in detail with reference to the accompanying drawings as follows. Description of the Drawings

[0024] Figure 1 It is a schematic structural diagram of a causal-embedded reconstruction network;

[0025] Figure 2 It is a flow chart for identifying abnormal variables based on the causal-embedded reconstruction network method;

[0026] Figure 3 It is a step diagram of the method for identifying abnormal variables based on the causal-embedded reconstruction network;

[0027] Figure 4 It is a schematic diagram of a continuous stirred tank reactor model with a control loop;

[0028] Figure 5 It is a causal relationship network of the continuous stirred tank reactor model;

[0029] Figure 6 It is the data of the continuous stirred tank reactor model under normal operating conditions;

[0030] Figure 7 It is the statistical distribution of the abnormal indexes of each variable under normal operating conditions by three methods;

[0031] Figure 8 It is the data of the continuous stirred tank reactor under abnormal condition 1;

[0032] Figure 9 It is a heat map of the abnormal scores of each variable in the continuous stirred tank reactor failure 1 by three methods;

[0033] Figure 10 It is the proportion of the abnormal scores of each variable 100 seconds after the occurrence of failure 1 by three methods;

[0034] Figure 11 It is the abnormal score curve of each variable in abnormal condition 1 by three methods;

[0035] Figure 12 It is the abnormal score curve of each variable in abnormal condition 1 by the fully connected causal-embedded reconstruction network;

[0036] Figure 13 It is the data of each variable in the continuous stirred tank reactor under abnormal condition 2;

[0037] Figure 14They are the heat maps of the anomaly scores of each variable in the continuous stirred tank reactor fault 2 by three methods;

[0038] Figure 15 They are the anomaly score ratios of each variable when the abnormal condition 2 occurs for 100 seconds by three methods;

[0039] Figure 16 They are the anomaly score curves of each variable in the fault 2 by three methods;

[0040] Figure 17 They are the anomaly score curves of each variable in the abnormal condition 2 by the FC-Causal-Embedding Reconstruction Network;

[0041] Figure 18 They are the data of each variable in the continuous stirred tank reactor fault 3;

[0042] Figure 19 They are the anomaly score ratios of each variable after the abnormal condition 3 occurs for 100 seconds by three methods;

[0043] Figure 20 They are the anomaly score curves of each variable in the abnormal condition 3 by three methods;

[0044] Figure 21 It is the structural schematic diagram of the abnormal variable identification device based on the Causal-Embedding Reconstruction Network. Detailed implementation manners

[0045] Next, in combination with the attached drawings, the detailed implementation manners of the present invention will be described in detail, but it should be understood that the protection scope of the present invention is not limited by the detailed implementation manners.

[0046] Unless otherwise clearly stated, in the whole specification and claims, the term "comprise" or its variations such as "comprises" or "including" etc. will be understood to include the stated elements or components, without excluding other elements or other components.

[0047] In this article, for the convenience of description, spatial relative terms such as "below", "beneath", "under", "above", "over", "on" etc. can be used to describe the relationship between one element or feature and another element or feature in the attached drawings. It should be understood that the spatial relative terms are intended to include different directions of the object in use or operation in addition to the directions depicted in the figures. For example, if the object in the figure is flipped, the element described as "below" or "under" other elements or features will be oriented "above" the element or feature. Therefore, the exemplary term "below" can include both the below and above directions. The object can also have other orientations (rotated 90 degrees or other orientations) and the corresponding explanations should be made for the spatial relative terms used herein.

[0048] In this text, terms such as "first", "second", etc. are used to distinguish two different components or parts, rather than to limit a specific position or relative relationship. In other words, in some embodiments, terms such as "first", "second", etc. can also be interchanged with each other.

[0049] Embodiment 1

[0050] In order to improve the accuracy of abnormal variable recognition and reduce the false alarm rate, as Figure 1 shown, in an embodiment of the present invention, an abnormal variable recognition method based on a causal-embedded reconstruction network is provided. Specifically:

[0051] The embodiment of the present invention is for abnormal variable recognition in a chemical process, and the main problems to be solved are the accuracy and efficiency of abnormal variable recognition in the multi-variable process monitoring of process chemical industry, especially the recognition in scenarios such as small amplitude changes and slow drifts is very good. When an abnormal working condition occurs, the embodiment of the present invention can accurately present the abnormal variable and its abnormal degree to the operator, realizing transparent process diagnosis.

[0052] The chemical process data often shows complex characteristics such as high dimensionality, non-linearity, and non-steadiness. Therefore, the abnormal variable recognition method based on the causal-embedded reconstruction network proposed in the embodiment of the present invention uses a graph neural network for feature representation learning. The network structure of the causal-embedded reconstruction network in the embodiment of the present invention can refer to Figure 1 (Schematic diagram of the structure of the causal-embedded reconstruction network); in the application scenario of the chemical process, the flow chart of abnormal variable recognition based on the causal-embedded reconstruction network method can refer to Figure 2 (Flow chart of abnormal variable recognition based on the causal-embedded reconstruction network method).

[0053] The causal-embedded reconstruction network in the embodiment of the present invention consists of 3 main modules: a graph encoder, a causal decoder, and an attribute decoder.

[0054] (1) Graph encoder:

[0055] The graph encoder is composed of a graph recurrent convolutional network and a graph sampling aggregation network connected in sequence to capture the connection relationship between adjacent nodes and model the temporal data characteristics in the chemical process. Given the process data set of the chemical process First, the causal relationship between variables is constructed according to process knowledge. A is an n×n square matrix, and its element A ij = 1 represents that there is an edge pointing from the i-th variable to the j-th variable; the observed sequence X and the causal relationship A are input into the graph convolutional recurrent neural network to obtain the encoded feature H, where H (0) = GCRN(X,A).

[0056] The graph convolutional recurrent network can capture the connection characteristics between nodes and their neighbors on the attribute network and the dynamic characteristics of the attribute data, and then learn the features of the data. The specific calculation process is as follows:

[0057]

[0058]

[0059]

[0060]

[0061] h t = otamh ⊙ (c t )

[0062] Wherein, is the input time step, is the cell state vector, is the output vector; ⊙ represents the Hadamard product operation. σ(·) represents the sigmoid function; i, f, o are the input gate, forget gate, and output gate respectively; the weights W and biases b are trainable network parameters.

[0063] Subsequently, through the graph sampling aggregation network, the feature representations of the input data at different scales are encoded and generated, that is, the neighbor data of the nodes are sampled, and their representations at different scales are learned through node aggregation, and then the embedding of the target node is learned:

[0064] Z = H (1) = ReLU[GraphSAGE(H (0) , A),

[0065] At this time is the latent space feature extracted by the graph encoder. The present invention selects the mean aggregation method, that is, the current feature of the node itself and the features of the neighbor nodes are linearly transformed as follows:

[0066]

[0067] (2) Causal decoding:

[0068] The task of the causal decoder is to reconstruct the causal relationship structure of the input data through the features in the latent space. Z in the latent space represents the features of each node. The features of two nodes with a causal relationship should show a higher correlation. Therefore, the product of the features of nodes with a causal relationship should be larger than the product of the features of nodes without a causal relationship. Based on this idea, the present invention adopts matrix multiplication operation to restore the adjacency matrix of the input data. The calculation method for reconstructing the adjacency matrix can be as follows:

[0069]

[0070] where the sigmoid function is defined as sigmoid(x) = 1 / (1 + e -x ), and the diagonal function represents extracting the diagonal element matrix of a matrix. Setting it as a symmetric matrix is to try to restore the causal connection relationship between different variables through the latent space feature Z; at the same time, since the diagonal elements of sigmoid(ZZ T ) represent the dot product of each component of Z with itself, this kind of self-correlation is not the most concerned issue of the present invention, so it is subtracted during the calculation. Similarly, the corresponding processing is also performed on the original adjacency matrix:

[0071] A g = A - diagonal(A).

[0072] In the above formula, the subscript g of A g represents the ground truth. At this time, the error of causal reconstruction is defined as,

[0073]

[0074] where the operator ||·|| F represents the Frobenius norm of the matrix, that is, the square root of the sum of the squares of each element of the matrix, that is

[0075] R c The larger it is, the greater the abnormality of the causal relationship decoded by the causal decoder. At this time, the causal relationship of the data is difficult to be reconstructed, so there may be abnormal working conditions during the process.

[0076] (3) Attribute decoding:

[0077] After receiving the latent space feature Z, the attribute decoder inputs it into graph sampling aggregation to reconstruct the attributes of the input data.

[0078]

[0079] At this time, the reconstruction error of the input data is defined as:

[0080]

[0081] The larger the reconstruction error, the more difficult it is to reconstruct the data attributes of the node. At this time, the system may be operating under abnormal working conditions.

[0082] (4) Loss function:

[0083] The loss function of the causality-embedded reconstruction network is composed of two terms: the reconstruction error of causality and the reconstruction error of attributes, to comprehensively optimize the ability of the encoder and decoder to reconstruct the causality structure and attribute data.

[0084] The reconstruction error of the attribute decoder mainly focuses on whether the data of certain variables are abnormal, and the reconstruction error of the causality decoder pays more attention to whether the causality structure between certain variables is abnormal. To achieve a comprehensive judgment of abnormal working conditions, the loss function needs to take into account both reconstruction errors, which is defined as

[0085]

[0086] where α is used to control the attribute error of the input data and the causality error The relative importance of, and α is an adjustable hyperparameter. Since this loss function is essentially the sum of the number of anomalies of each node, the anomaly score value of each node can be correspondingly defined as:

[0087]

[0088] where, is the reconstruction of the attributes of each node by the attribute decoder, is the vector obtained by summing the rows of the reconstruction matrix of the causality decoder. If the anomaly score of a certain variable is very high, it means that its time series data value and its causal connection relationship with other variables have changed significantly compared with the normal working condition, so it is more likely to be an abnormal variable. By calculating its threshold value under normal working conditions, variables exceeding the threshold can be determined as abnormal variables during abnormal working conditions.

[0089] During the online monitoring stage, real-time online monitoring data is input into the causality-embedded reconstruction network model to calculate the anomaly score of each variable. If the anomaly score of a certain variable exceeds the threshold, then this variable is determined as an abnormal variable. Otherwise, continue data acquisition and process monitoring.

[0090] The method for identifying abnormal variables based on the causality-embedded reconstruction network described in the embodiments of the present invention, as Figure 3 shown, the offline modeling process it includes specifically may include the following steps:

[0091] S11. Preprocess the original historical data under normal working conditions, including: data standardization, dividing the training set and the validation set, and using the sliding window technique to create a time series window to construct the first time-delay matrix of the monitoring data in the training set at time k;

[0092] In this step, preprocessing the original historical data under normal working conditions may specifically include:

[0093] Divide the original historical data into a training set and a validation set Next, obtain the training set and the validation set through data standardization. The formula for data standardization can include:

[0094] Perform pre - processing on the training set to obtain

[0095]

[0096] where is the mean of the training set, is the standard deviation of the training set.

[0097] Perform pre - processing on the validation set to obtain

[0098]

[0099] Use the sliding window technique to create a time - series window and construct the time - lag matrix of the monitoring data at time k in the training set where k ∈ [w, n ; which is composed of train and its previous w - 1 samples {x n and k}, and the width of the sliding window is w; train,k-w+1 ,…,x train,k-2 ,x train,k-1}

[0100]

[0101] S12. Define the causal relationship between variables based on process knowledge and construct the corresponding causal relationship network based on a directed graph;

[0102] This step can specifically include:

[0103] Define the causal relationship between variables based on process knowledge and construct an adjacency matrix A ∈ R m×m , where A is an n×n square matrix, and its element A ij = 1 represents that there is an edge pointing from the i - th variable to the j - th variable;

[0104] The causal relationship network is constructed based on a directed graph. The nodes on the directed graph correspond to the monitoring variables in the process; the edges on the directed graph correspond to the connection relationship between variables, defined as where is a set of n nodes in the graph, and ε is the set of edges in the graph.

[0105] S13. Set the structural parameters of the causal-embedding reconstruction network and randomly initialize its connection parameters; define the loss function of the causal-embedding reconstruction network. Specifically:

[0106] The structural parameters of the causal-embedding reconstruction network include: the number of input layer units, the number of hidden layer units, the size of each batch of data BATCH_SIZE, the size of EPOCH for training all samples in the training set once, and the dropout value to avoid overfitting; randomly initialize its connection parameter weights W and biases b.

[0107] The loss function of the causal-embedding reconstruction network is:

[0108]

[0109] In the formula, α is used to control the relative importance of the attribute error of the input data and the causal relationship error , and α is an adjustable hyperparameter.

[0110] S14. Use the training set and the causal relationship network between variables to jointly train the causal-embedding reconstruction network, and obtain the optimal connection parameters of the causal-embedding reconstruction network by learning the data features under normal working conditions.

[0111] The unsupervised training of the causal-embedding reconstruction network is to jointly train with the data X of the training set train,k and the causal relationship network A between variables. The model of the causal-embedding reconstruction network is:

[0112] H (0) = GCRN(X, A).

[0113] Z = H (1) = ReLU[GraphSAGE(H (0) , A)],

[0114]

[0115]

[0116] where is the latent space feature extracted by the graph encoder, is the reconstructed adjacency matrix, is the reconstructed input data, and the optimal network connection parameter weights W and biases b are obtained.

[0117] S15. Input the monitoring data of the validation set into the trained causal-embedding reconstruction network, and calculate the anomaly score values of each variable. Specifically:

[0118] The anomaly score value of each variable The calculation process includes:

[0119]

[0120] Among them, is the reconstruction of each node attribute by the attribute decoder, is the vector obtained by summing the rows of the matrix reconstructed by the causal decoder.

[0121] S16. At a given confidence level η, calculate the anomaly threshold s of each variable according to the kernel density estimation method cl .

[0122] Furthermore, in the embodiment of the present invention, the steps of online monitoring specifically include:

[0123] S21. Obtain the real-time operation data in the chemical process and perform data standardization to obtain the input set; according to the sliding window technology, construct the second time-delay matrix of the monitoring data at time k as the online input set of the causal-embedding reconstruction network;

[0124] Construct a data set for the online operation data in the chemical process Use the mean and variance of the training set data to perform data standardization processing to obtain the input set of online data Use the sliding window technology to create a time series window and construct the time-delay matrix X of the monitoring data at time k test,k ∈R w×m , k ∈ [w, n test ;

[0125] The formula for data standardization processing includes:

[0126]

[0127] Among them, is the mean of the training set, is the standard deviation of the training set;

[0128] Based on the sliding window, construct the second time-delay matrix X of the monitoring data in the input set at time k test,k , which is composed of x testk , and its previous w - 1 samples {x test,k-w+1 , …, x test,k-2 , x test,k-1}:

[0129]

[0130] S22. For each input sample in the online input set, perform feature representation based on the causal-embedding reconstruction network and calculate the anomaly score values of each variable;

[0131] For each input sample X test,k , perform feature representation based on the trained causal-embedding reconstruction network and calculate the anomaly score s i of each variable;

[0132] Input sample X test,k The calculation process of the anomaly score s test,i of each variable includes:

[0133]

[0134] where is the reconstruction of each node attribute by the attribute decoder, is the vector obtained by summing the rows of the reconstruction matrix of the causal decoder.

[0135] S23. If the anomaly score value s test,i of a certain variable exceeds the corresponding anomaly score threshold s cl , then determine the variable as an abnormal variable; otherwise, continue data collection and process monitoring.

[0136] In summary, the present invention proposes a new method for identifying abnormal variables: a method for identifying abnormal variables based on a causal-embedding reconstruction network. The method uses a graph deep network to transform the multivariate variable time-series data of a chemical process into graph-structured data, and extracts the dynamic and spatial relationships between variables; due to the embedding of the causal relationship network, the features in the latent space are restricted to be generated only based on the variables that have a causal relationship with it; by simultaneously reconstructing the variable time-series data and the causal relationship network between variables, the pollution effect can be overcome, and high-accuracy and low false-alarm-rate identification of abnormal variables for various different types of abnormal working conditions can be achieved, especially for scenarios such as small-amplitude changes and slow drifts, the identification is very good.

[0137] Through the present invention, when a fault occurs, the abnormal variables and their degrees of abnormality can be accurately presented to the operator. Therefore, high-accuracy and low false-alarm-rate identification of abnormal variables for various different types of abnormal working conditions can be achieved, especially for scenarios such as small-amplitude changes and slow drifts, the identification is very good, which has important practical significance for the transparent process diagnosis of actual chemical processes.

[0138] To verify the feasibility and effectiveness of the method proposed in the embodiments of the present invention, a specific example of the present invention is disclosed, and the effectiveness of the present invention is verified through experiments in a continuous stirred tank reactor process, and the performance is compared with classical methods for identifying abnormal variables: the squared prediction error contribution method and the integrated gradient method.

[0139] For the sake of unifying concepts and facilitating narration, in the experiment, the anomaly variable identification method based on the causality-embedded reconstruction network is denoted as CERN (Causality-embedded Reconstruction Network), the anomaly variable identification method based on the integrated gradient method is denoted as IG (Integrated Gradient), and the anomaly variable identification method based on the squared prediction error contribution is denoted as SPE contribution (Squared Prediction Error); the variable contribution given by the squared prediction error method, the variable integrated gradient given by the integrated gradient method, and the anomaly variable score based on the causality-embedded reconstruction network are collectively referred to as "anomaly indicators".

[0140] In this example, the continuous stirred tank reactor is a closed-loop reactor widely used in polymerization chemical reactions. The material undergoes a first-order exothermic reaction inside it and is discharged, belonging to a typical and highly nonlinear chemical reaction system in the process industry. In the experiment of this example, the Simulink model of the continuous stirred tank reactor can be downloaded at https: / / ww2.mathworks.cn / matlabcentral / fileexchange / 65091-cascade-controlled- continuous stirred tank reactor-for-fault-simulation?s_tid=FX_rc1_behav. This reactor has 2 control loops. The temperature inside the reactor is controlled by cooling water, and the liquid level of the reactor is controlled by the flow rate at the reactor outlet, as shown in Figure 4 (Schematic diagram of the continuous stirred tank reactor model with control loops). The 10 variables involved in the Simulink model are shown in Table 1, including 4 input variables and 6 output variables.

[0141] Table 1: Variable list of the continuous stirred tank reactor with control loops

[0142]

[0143] The relationships between the above variables are represented by the following mass and energy balance equations:

[0144]

[0145]

[0146]

[0147]

[0148]

[0149] The causal relationship network of the continuous stirred tank reactor model can be obtained from the above equations as shown in Figure 5 (Causal relationship network of the continuous stirred tank reactor model); the steady-state operating point of the reactor is determined by the parameters in Table 2.

[0150] Table 2: Parameters and steady-state values of the continuous stirred tank reactor

[0151]

[0152] By adding random perturbations to the 4 input variables, the data around the operating point can be obtained. Under normal operating conditions, the data of 10 monitoring variables are shown in Figure 6 (Data under normal operating conditions of the continuous stirred tank reactor model). In this invention, 4500 normal operating condition data are selected to construct the training set, and 1000 normal operating condition data are used to construct the validation set. The sampling time is 1 per second, and the time window length is set to 30. The value of α is 0.6.

[0153] By adding a perturbation module to Simulink, different abnormal operating condition scenarios can be simulated. This case study generates 6 abnormal operating condition scenarios as the research objects, as shown in Table 3. Each abnormal operating condition includes 1000 data points, and the fault is introduced at the 200th data point.

[0154] Table 3: Abnormal operating conditions and their introductions in the continuous stirred tank reactor model

[0155]

[0156] First, calculate the anomaly indices of each variable for the three methods under normal operating conditions, as shown in Figure 7 (Statistical distribution of the anomaly indices of each variable for the three methods under normal operating conditions). At this time, the anomaly indices of each variable are generally stable at relatively small values. Next, specifically examine the anomaly variable identification effects of the three methods under three scenarios: abnormal operating condition 1, abnormal operating condition 2, and abnormal operating condition 3.

[0157] (1) Abnormal operating condition 1

[0158] Abnormal operating condition 1 is a simple step fault of the liquid level sensor, and the value of h shows a step increase. Since this is only a reading fault of the sensor and the process itself is not affected abnormally, the other variables are all normal, as shown in Figure 8 (Data of abnormal operating condition 1 of the continuous stirred tank reactor). Before and after the fault occurs, the heat maps of the anomaly indices of each variable calculated by the causal-embedded reconstruction network, the squared prediction error method, and the integrated gradient method are shown in Figure 9(Heat maps of the anomaly scores of each variable in the continuous stirred tank reactor failure 1 by three methods). All three methods successfully identified h as the variable with the highest degree of anomaly, proving the effectiveness of the three methods as a whole.

[0159] Under abnormal condition 1, the proportions of the anomaly indexes of all variables calculated by the three methods are shown in Figure 10 (Proportions of the anomaly scores of each variable 100 seconds after the occurrence of failure 1 by three methods). At this time, only the proportion of the anomaly index of h exceeded 30% in each method, and the contribution degrees of the remaining variables were relatively low, indicating that the anomaly degree of h was far beyond that of other variables at this time. Overall, the integrated gradient method is more sensitive to variables with a large degree of failure, indicating that the integrated gradient method may be more effective in identifying the main abnormal variables.

[0160] The changes in the index values of each variable by the three methods and their relative relationships with the thresholds are shown in Figure 11 (Anomaly score curves of each variable in abnormal condition 1 by three methods). The horizontal line represents the anomaly index threshold of each variable. The part below the horizontal line represents normal, and the part above the horizontal line represents abnormal. In the anomaly score curve given by the causal-embedded reconstruction network, only the anomaly score of h significantly exceeded the threshold, while the anomaly scores of all other variables were stable and all below the threshold. This is completely consistent with the facts.

[0161] In the contribution degree method of the squared prediction error method, although h is the variable with the largest increase in contribution degree, the contribution degrees of the squared prediction error method of the three variables F, Ci, and Fc also increased significantly and exceeded the threshold, and the anomaly indexes of C, T, and Tc also showed abnormal fluctuations. In the integrated gradient method, in addition to h, the integrated gradient values of F, T, Tc, and Fc also exceeded the threshold. Therefore, if only the main abnormal variables are to be qualitatively given, all three methods can do it. However, if the threshold is to be used to specifically determine the abnormal variables, only the causal-embedded reconstruction network can give accurate results, and false alarms will occur in the squared prediction error method and the integrated gradient method.

[0162] The core reason for the above phenomenon is that the conventional autoencoder structure uses the data of all variables indiscriminately to generate reconstructed data. The reconstruction of the data of other normal variables may use the data of h, so the large anomaly of h drives the reconstruction of other variables to be abnormal.

[0163] To prove this, a control experiment was conducted on the modification of the causal-embedded reconstruction network structure: (1) Instead of setting the causal relationship network between variables, a connection relationship was set between every two variables, that is, the causal connection was changed to an ordinary full connection; (2) Set α = 1, that is, like an ordinary autoencoder, only consider the error of the reconstructed data as the anomaly score; (3) Set the graph convolutional recurrent network to aggregate information from multi-order neighbors during sampling, rather than only aggregating information from 1st-order neighbors. This structure is called the FC-causal-embedded reconstruction network, that is, the fully connected causal-embedded reconstruction network. The fully connected causal-embedded reconstruction network loses the information of the graph structure and is functionally similar to an ordinary autoencoder. The anomaly scores of each variable in the FC-causal-embedded reconstruction network under abnormal condition 1 are shown in Figure 12 (Anomaly score curves of each variable in the fully connected causal-embedded reconstruction network under abnormal condition 1). At this time, except for h, the anomaly scores of variables such as T, Tc, and Fc also exceed the threshold, and the pollution effect in the ordinary autoencoder begins to appear. The fully connected causal-embedded reconstruction network can no longer accurately identify abnormal variables. This control experiment strongly proves that the key reason why the causal-embedded reconstruction network is more accurate than the traditional autoencoder in identifying abnormal variables lies in its effective use of the causal connection relationship between variables.

[0164] Abnormal condition 1 verifies that in the actual chemical process application, if the squared prediction error method or the integrated gradient method is used to construct an abnormal variable identification model, it is possible that the misreport of other variables may be caused by the abnormality of one sensor. However, in the identification result given by the causal-embedded reconstruction network, only h is abnormal and no other changes are caused, which makes it easy for operators to judge that it is a sensor abnormality and thus take corresponding solutions.

[0165] (2) Abnormal condition 2

[0166] Abnormal condition 2 represents an increase in the feed flow rate Fi, which is a very complex situation. The data of each variable at this time are shown in Figure 13 (Data of each variable in the continuous stirred tank reactor under abnormal condition 2). The liquid level control loop causes the outlet flow rate F to increase synchronously to control h to remain unchanged. Since the overall flow rate of the continuous stirred tank reactor becomes larger, although more reactions occur in general, the degree of reaction may decrease. On the one hand, this leads to an increase in the outlet concentration C, and on the other hand, it leads to an increase in the reaction heat generation. In order to maintain the stability of T, the cooling water flow rate Fc will increase accordingly. The increase in heat generation and the increase in cooling water flow rate make it more difficult to control the stability of the temperature T and the cooling water outlet temperature Tc compared to the normal condition. Therefore, although the mean values of T and Tc can be maintained stable at this time, their later fluctuation degrees will be more intense, showing a different pattern from the normal condition.

[0167] In this abnormal operating condition, the changes in Fi and F are significant, but the changes in C and Fc are only a small and slow increase on the basis of the normal values, while the fluctuations in T and Tc are only slightly increased on the basis of the normal values. In actual chemical processes, this phenomenon of slow drift and slightly increased fluctuation is very difficult to detect for DCS alarms because the change in the variable value is very slow and the difference from the normal operating condition is not obvious.

[0168] Before and after the occurrence of abnormal operating condition 2, the heat maps of the abnormal indexes of each variable calculated by the causal-embedded reconstruction network, the square prediction error method, and the integrated gradient method are shown in Figure 14 (Heat maps of the abnormal scores of each variable in the continuous stirred tank reactor fault 2 by the three methods). At this time, all three methods successfully identified that the variables with the largest abnormal degree are Fi and F, indicating that the results given by the three methods all have good overall reference value in the identification of abnormal variables.

[0169] At the 100th second when this abnormal operating condition occurred, the percentage of abnormal scores given by the three methods is shown in Figure 15 (Proportion of abnormal scores of each variable when the three methods occur at the 100th second of abnormal operating condition 3). The law in the data is similar to that in the above example. The proportions given by the causal-embedded reconstruction network and the square prediction error method are relatively uniform, while the integrated gradient method is still more sensitive to variables with a large abnormal degree.

[0170] The changes in the index values of each variable by the three methods are shown in Figure 16 (Abnormal score curves of each variable in fault 2 by the three methods). In the abnormal score curve of the causal-embedded reconstruction network, in addition to the significantly increased Fi and F, the abnormal scores of C and Fc also exceed the threshold at about 400 seconds. After the abnormal fluctuations of T and Tc occur at about 600 seconds, their abnormal indexes also exceed the threshold. This shows that the causal-embedded reconstruction network can not only identify variables with large abnormal changes, but is also very sensitive to difficult-to-identify scenarios such as slow drift and increased fluctuation degree, and can accurately identify various different types of abnormal variables. In addition to the above variables, the abnormal score curves of the unchanged variables h, Ci, Ti, and Tci are always below the threshold and no false alarms will occur.

[0171] In the contribution method of the squared prediction error method, except for Ti, the contributions of almost all variables exceed the threshold under normal operating conditions. For h, Ci, and Tci, although their contributions by the squared prediction error method are very small, they still increase significantly due to the influence of other abnormal variables. The performance of the integrated gradient method is similar to that of the squared prediction error method, and variables except for Ci and Ti almost all exceed the threshold. Thus, it can be seen that only the large abnormal changes in these two important variables, F and Fi, can lead to an increase in the reconstruction errors of almost all other variables. This contamination effect makes it impossible for the squared prediction error method and the integrated gradient method to accurately identify abnormal variables, and they can only be used to macroscopically examine variables with significantly larger abnormal variables.

[0172] From the case study of abnormal operating condition 2, it can be seen that the causal-embedded reconstruction network method does not misreport normal variables as abnormal variables. Similarly, a control experiment was conducted on the modification of the causal-embedded reconstruction network structure. The causal connection was changed to a fully connected connection, and the identification results of the fully connected causal-embedded reconstruction network were examined. The fully connected causal-embedded reconstruction network showed similar phenomena to ordinary autoencoders, with false alarms occurring for a large number of variables, as shown in Figure 17 (Abnormal score curves of each variable in abnormal operating condition 2 for the fully connected causal-embedded reconstruction network). The result comparison between the causal-embedded reconstruction network and the fully connected causal-embedded reconstruction network further proves the effectiveness of the causal connection design in the causal-embedded reconstruction network. Further, the alarm times of the causal-embedded reconstruction network for each alarm variable in abnormal operating condition 3 were examined, as shown in Table 4.

[0173] Table 4 Alarm times of the causal-embedded reconstruction network for each variable in abnormal operating condition 2

[0174]

[0175]

[0176] Since the causal-embedded reconstruction network accurately identifies abnormal variables, based on this alarm sequence, we can diagnose that the root cause variable of abnormal operating condition 2 is the variable Fi that alarms first. The increase in Fi directly causes the increase in F, which subsequently leads to complex changes in other variables in the process.

[0177] In actual chemical processes, the DCS may be slow in detecting slow-drift variables, and it may not issue any alarms for fluctuations near the steady state, resulting in missed alarms for abnormal variables. If the squared prediction error method or the integrated gradient method is used, false alarms may occur, misidentifying normal variables as abnormal variables. These phenomena limit the application effects of the above methods in practice.

[0178] (3) Abnormal operating condition 3

[0179] Abnormal condition 3 means that scaling gradually occurs at the heat transfer position of cooling water. The values ​​of the variables at this time are shown in Figure 18 (Data of various variables in continuous stirred tank reactor fault 3). Scaling will cause the cooling water heat exchange capacity to gradually decrease, and thus the cooling water outlet temperature Tc will decrease. In order to control the reactor temperature to be stable, the cooling water flow rate Fc must be increased. At this time, the temperature in the reactor becomes difficult to control, and the temperature T fluctuates significantly after about 800 seconds. The change in temperature affects the kinetic constants of the reaction, causing the reactant concentration C to fluctuate accordingly. Similarly, the abnormal score ratio after 100 seconds is shown in Figure 19 (The abnormal score ratio of each variable of the three methods 100 seconds after abnormal condition 3 occurs) At this time, Tc is obviously the most critical abnormal variable.

[0180] The variable anomaly indicators given by the three methods are shown in Figure 20 (Abnormal score curves of the three methods for each variable in abnormal condition 3). In the abnormal score curve of the causal-embedded reconstruction network, Tc and Fc were successfully identified as obviously abnormal variables; several variables such as h, Fi, F, Ci, Ti, Tci were not affected and were all within the normal range; the abnormal fluctuations of C and T in the abnormal condition 3 in the later period were also correctly reflected in the curve. Therefore, the causal-embedded reconstruction network also gave accurate identification results for the abnormal variables in this case. The performance of the squared prediction error method and the integral gradient method was not satisfactory, and almost all variables showed abnormal change trends. If a large number of variables have false alarms in the actual process, this is obviously not conducive to the elimination of abnormal conditions.

[0181] The alarm time of each variable in abnormal condition 3 by the causal-embedded reconstruction network is shown in Table 5. At this time, the causal-embedded reconstruction network diagnosed that the root cause variable of abnormal condition 3 was Tc, and then gradually transmitted to the three variables of Fc, T, and C.

[0182] Table 5: Alarm time of each variable in abnormal condition 3 by causal-embedded reconstruction network

[0183]

[0184] For other abnormal working conditions, we will not analyze them one by one. Table 6 summarizes the abnormal variables identified by the causal-embedded reconstruction network method under various abnormal working conditions, and calculates the average index of the variables as their abnormal degree. Comparing the recognition results in Table 6 with the actual situation, it is found that the abnormal variable recognition results given by the causal-embedded reconstruction network are almost completely consistent with the actual situation.

[0185] Table 6: Results of abnormal variable identification for various abnormal conditions of continuous stirred tank reactor using the causal-embedded reconstruction network method

[0186]

[0187] In summary, the causality-embedded reconstruction network can comprehensively and accurately identify abnormal variables for all abnormal working conditions, and there are almost no false alarms or missed alarms. This good performance has broad application prospects in practice.

[0188] Embodiment 2

[0189] Corresponding to the method embodiment, the embodiment of the present invention further provides an abnormal variable recognition device based on the causality-embedded reconstruction network, such as a terminal, a server, etc. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc., but is not limited thereto.

[0190] An example diagram of the hardware structure block diagram of the abnormal variable recognition device based on the causality-embedded reconstruction network provided by the embodiment of the present invention is as Figure 21 shown, and may include:

[0191] Processor 1, communication interface 2, memory 3, and communication bus 4;

[0192] Among them, the processor 1, the communication interface 2, and the memory 3 complete mutual communication through the communication bus 4;

[0193] Optionally, the communication interface 2 can be an interface of a communication module, such as an interface of a GSM module;

[0194] The processor 1 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0195] The memory 3 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk memory.

[0196] Among them, the processor 1 is specifically configured to execute the computer program stored in the memory 3 to perform the following steps:

[0197] Its steps of offline modeling include:

[0198] S11. Preprocess the original historical data under normal working conditions, including: data standardization, dividing the training set and the validation set, and using the sliding window technique to create a time series window to construct the first time delay matrix of the monitoring data at time k in the training set;

[0199] S12. Define the causal relationship between variables based on process knowledge, and construct the corresponding causal relationship network using a directed graph;

[0200] S13. Set the structural parameters of the causal-embedding reconstruction network, randomly initialize its connection parameters, and define the loss function of the causal-embedding reconstruction network;

[0201] S14. Use the training set and the causal relationship network between variables to jointly train the causal-embedding reconstruction network, and obtain the optimal connection parameters of the causal-embedding reconstruction network by learning the data characteristics under normal working conditions;

[0202] S15. Input the monitoring data of the validation set into the trained causal-embedding reconstruction network, and calculate the anomaly score values of each variable;

[0203] S16. Calculate the anomaly score threshold s of each variable according to the kernel density estimation method under the given confidence level η cl 。

[0204] Preferably, the online monitoring steps include:

[0205] S21. Obtain the real-time operation data in the chemical process and perform data standardization to obtain the input set; according to the sliding window technique, construct the second time delay matrix of the monitoring data at time k as the online input set of the causal-embedding reconstruction network;

[0206] S22. For each input sample in the online input set, perform feature representation based on the causal-embedding reconstruction network and calculate the anomaly score values of each variable;

[0207] S23. If the anomaly score value of a certain variable exceeds the corresponding anomaly score threshold, then determine the variable as an abnormal variable; otherwise, continue data collection and process monitoring.

[0208] The above product can execute the method provided by the embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. For the technical details not described in detail in this embodiment, reference can be made to the abnormal variable identification method based on the causal-embedding reconstruction network provided by the embodiment of the present invention.

[0209] Those of ordinary skill in the art will appreciate that the various example units and algorithm steps described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled professionals may use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0210] In several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. Additionally, the couplings, direct couplings, or communication connections shown or discussed among each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0211] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0212] In addition, in each embodiment of this application, the various functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0213] It should be understood that in the embodiments of this application, the dependent claims, various embodiments, and features can be combined with each other to achieve the solution of the foregoing technical problems.

[0214] If the described functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of this application. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs, etc., which can store program codes.

[0215] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An abnormal variable recognition method based on a causal-embedding reconstruction network, characterized in that The steps of offline modeling include: S11. Preprocess the original historical data under normal working conditions, including: data standardization, dividing the training set and the validation set, and using the sliding window technique to create a time series window to construct the first time-delay matrix of the monitoring data at time k in the training set; S12. Define the causal relationship between variables based on process knowledge and construct a corresponding causal relationship network using a directed graph; S13. Set the structure parameters of the causal-embedding reconstruction network, randomly initialize its connection parameters, and define the loss function of the causal-embedding reconstruction network; S14. Use the training set and the causal relationship network between variables to jointly train the causal-embedding reconstruction network, and obtain the optimal connection parameters of the causal-embedding reconstruction network by learning the data characteristics under normal working conditions; S15. Input the monitoring data of the validation set into the trained causal-embedding reconstruction network and calculate the anomaly score values of each variable; S16. Calculate the anomaly score threshold s of each variable according to the kernel density estimation method at a given confidence level η cl .

2. The abnormal variable recognition method based on a causal-embedding reconstruction network according to claim 1, wherein The steps of online monitoring include: S21. Obtain the real-time operation data in the chemical process and perform data standardization to obtain the input set; according to the sliding window technique, construct the second time-delay matrix of the monitoring data at time k as the online input set of the causal-embedding reconstruction network; S22. For each input sample in the online input set, perform feature representation based on the causal-embedding reconstruction network and calculate the anomaly score values of each variable; S23. If the anomaly score value of a certain variable exceeds the corresponding anomaly score threshold, then determine the variable as an abnormal variable; otherwise, continue with data collection and process monitoring.

3. The abnormal variable recognition method based on the causal-embedded reconstruction network according to claim 1, characterized in that The preprocessing of the original historical data under normal working conditions includes: Divide the original historical data into a training set and a validation set Data standardization processing of the original historical data training set: Among them, is the mean of the training set, is the standard deviation of the training set; Construct a training set The time-delay matrix X of the monitoring data at time k train,k ∈R w×m , k ∈ [w, n train ; composed of x train,k and its previous w - 1 samples {x train,k-w+1 , …, x train,k-2 , x train,k-1}; the w is the width of the sliding window; 4. The abnormal variable recognition method based on the causal-embedding reconstruction network according to claim 3, characterized in that, The defining of the causal relationship between variables based on process knowledge and constructing a corresponding causal relationship network using a directed graph includes: Define the adjacency matrix \(A\in\mathbb{R}\) based on process knowledge m×m , where \(A\) is an \(n\times n\) square matrix, and its element \(A\) ij \(= 1\) indicates that there is an edge pointing from the \(i\)-th variable to the \(j\)-th variable; Construct a causal relationship network based on a directed graph, where the nodes on the directed graph correspond to the monitoring variables in the process; the edges on the directed graph correspond to the connection relationships between variables; defined as Among them, is the set of n nodes in the graph, and ε is the set of edges in the graph.

5. The abnormal variable recognition method based on the causal-embedded reconstruction network according to claim 4, characterized in that The setting of the structure parameters of the causal-embedding reconstruction network, randomly initializing its connection parameters, and defining the loss function of the causal-embedding reconstruction network includes: The structure parameters of the causal-embedding reconstruction network include: the number of input layer units, the number of hidden layer units, the size of each batch of data Batch_size, the size of Epoch for training all samples in the training set once, the Dropout value to avoid overfitting; randomly initialize the weights W and biases b of its connection parameters; Loss function of the causality-embedded reconstruction network is as follows: Where α is used to control the attribute error of the input data and the causal relationship error of the relative importance, and α is an adjustable hyperparameter.

6. The method for identifying abnormal variables based on the causal-embedded reconstruction network according to claim 5, wherein The using of the training set and the causal relationship network between variables to jointly train the causal-embedding reconstruction network and obtaining the optimal connection parameters of the causal-embedding reconstruction network by learning the data characteristics under normal working conditions includes: Unsupervised training of the causal-embedded reconstruction network is to jointly train using the training set data X train,k and the causal relationship network A between variables. The model of the causal-embedded reconstruction network is as follows: H (0) = GCRN(X,A). Z = H (1) = ReLU[GraphSAGE(H (0) , A)], Among them, is the latent space feature extracted by the graph encoder, is the reconstructed adjacency matrix, is the reconstructed input data attribute. By training the causal-embedding reconstruction network, the optimal network connection parameters, i.e., the weights W and the bias b, are obtained.

7. The abnormal variable recognition method based on the causal-embedded reconstruction network according to claim 6, characterized in that Inputting the monitoring data of the validation set into the trained causal-embedding reconstruction network and calculating the anomaly score values of each variable includes: Data standardization processing of the original historical data validation set: Among them, is the mean of the training set, is the standard deviation of the training set; Construct the validation set The time-delay matrix X of the monitoring data at time k val,k ∈R w×m , where k ∈ [w, n val ; composed of x valk , and its previous w - 1 samples {x val,k-w+1 , …, x val,k-2 , x val,k-1}; where w is the width of the sliding window Abnormal score values of each variable The calculation process is as follows: Among them, is the reconstruction of each node attribute by the attribute decoder, is the vector obtained by summing the rows of the reconstruction matrix of the causal decoder.

8. The abnormal variable recognition method based on the causal-embedded reconstruction network according to claim 7, characterized in that The obtaining of the real-time operation data in the chemical process and performing data standardization to obtain the input set; According to the sliding window technique, constructing the second time-delay matrix of the monitoring data at time k as the online input of the causal-embedding reconstruction network includes: Construct a data set from the on-line operation data in the chemical process Perform data standardization processing using the mean and variance of the training set data to obtain the input set Use the sliding window technique to create a time series window and construct the time-delay matrix X of the monitoring data at time k test,k ∈R w×m , k ∈ [w, n test ; Data standardization processing of the online operation data includes: Among them, is the mean of the training set, is the standard deviation of the training set; Construct the second time-delay matrix \(X\) of the monitoring data in the input set at time \(k\) based on a sliding window test,k , which consists of \(x\) testk , and its previous \(w - 1\) samples as follows:

9. The abnormal variable recognition method based on a causal-embedded reconstruction network according to claim 8, characterized in that, For each input sample in the input set, performing feature representation based on the causal-embedding reconstruction network and calculating the anomaly scores of each variable, including: For each input sample X test,k , feature representation and the anomaly score s of each variable are calculated based on the trained causal-embedding reconstruction network i ; Input sample X test,k The anomaly score s of each variable test,i The calculation process includes: Among them, it is the reconstruction of each node attribute by the attribute decoder, and it is the vector obtained by summing the rows of the matrix reconstructed by the causal decoder.

10. An abnormal variable recognition device based on a causal-embedded reconstruction network, characterized in that, Including: a memory for storing a computer program; a processor for calling and executing the computer program to implement the steps of the anomaly variable recognition method based on the causal-embedding reconstruction network according to any one of claims 1-9.

Images