Abnormity detection method based on hierarchical topological graph isomorphic contrast learning

Abstract

The invention discloses an anomaly detection method based on hierarchical topological graph isomorphic comparative learning. The anomaly detection method comprises the following steps: 1) acquiring and integrating data; 2) constructing a topological graph and a positive and negative sample pair: constructing a variable topological graph according to industrial process knowledge, performing graph enhancement on the constructed topological graph, and simultaneously creating a plurality of samples, each sample comprising two enhanced topological graphs, two mappings of the same sample being regarded as a positive pair, and projections between different samples being regarded as a negative pair; 3) modeling training and model testing: constructing a hierarchical topological graph isomorphic comparative learning model to train the training data set, and injecting topological professional knowledge into graph comparative learning; and 4) data reconstruction: reconstructing the high-dimensional embedding of the output, and locating the anomaly by using the reconstruction loss. According to the method, the positive and negative sample pairs are constructed by using graph enhancement on graph data samples, topological professional knowledge is injected into graph contrast learning to help the model to mine interaction information of process variables, and the unsupervised anomaly detection capability of the model is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of process industry anomaly detection modeling, and in particular to an anomaly detection method based on hierarchical topology graph isomorphism contrast learning. Background Art

[0002] In industrial applications, anomaly detection is crucial to ensure process safety. The main methods include first-principles-based and data-driven models. As industrial processes become more complex, data-driven models based on deep learning (DL) have gradually become mainstream, especially unsupervised learning, which shows great potential in the absence of labeled data. Traditional unsupervised methods, such as prediction-based and reconstruction-based methods, have difficulty extracting effective features when faced with high-dimensional, multivariate time series data, thus affecting detection accuracy.

[0003] As an unsupervised learning method, contrastive learning (CL) can effectively extract discriminative features in high-dimensional data for unlabeled anomaly detection by enhancing the ability to distinguish positive and negative sample pairs. However, the CL model lacks interpretability, especially in complex industrial processes, and cannot fully utilize the connections between variables and the process knowledge of anomaly propagation.

[0004] Graph Neural Networks (GNN) combined with CL produces a graph contrastive learning model (GCL), which can effectively solve the problem of insufficient interpretability of contrastive learning. Prior knowledge is used to construct a topological graph, which transmits variable information, models anomaly propagation, and identifies the root cause of anomalies to improve the transparency of the model. Although GCL has made progress in modeling topological structures, its ability to distinguish complex topological structures is still limited. Experiments show that the introduction of graph topology expertise, especially through the one-dimensional Weisfeiler-Lehman (WL) test, can significantly improve the performance of the GCL model.

[0005] Based on the above problems, the present invention designs a hierarchical topological isomorphic expertise embedded graph contrastive learning (HTGICL) model. Starting from process knowledge, the model constructs a topological graph, generates positive and negative sample pairs through graph enhancement, and introduces topological isomorphic expertise using the WL test to enhance the model's ability to discern graph structures. By reconstructing the embedding and calculating the reconstruction error through a multilayer perceptron (MLP), HTGICL performs well in unsupervised anomaly detection. Summary of the invention

[0006] In view of the problems of limited labeled samples and complex unit interactions in the existing process industry anomaly detection, the purpose of the present invention is to provide a process anomaly detection modeling method based on HTGICL, construct a variable topology graph based on prior knowledge, and then use graph enhancement technology to help the model learn a more generalized graph representation. Subsequently, topological isomorphism expertise is injected into the graph comparison task to guide the encoder to mine distinctive anomaly information from process data and graph representation. Taking the sugar production process of Lublin Sugar Factory as an example, it is compared with the traditional method to demonstrate the effectiveness of the proposed method.

[0007] The technical solution of the present invention is as follows:

[0008] An anomaly detection method based on hierarchical topological graph isomorphic contrast learning includes the following steps:

[0009] 1) Data acquisition and integration

[0010] First, data on various variables during normal operation and abnormal conditions are collected based on the chemical production process. Normal operation usually refers to the state of the system when it is operating stably under predetermined operating conditions, when various variables (such as temperature, pressure, flow, etc.) are within the set safety range. Abnormal operation refers to the state of the system under abnormal conditions, which may include equipment failure, operating errors, etc., causing the variables to fluctuate beyond the safety range. The data of normal operation is used as the training data set, and the data of abnormal operation is used as the validation set and test set and normalized;

[0011] Step 1.1: The collected normal industrial process data set is used as the training data set, and the abnormal industrial process data is used as the test data set. The ratio of the training data set, the verification data set, and the test data set is 2:1:1;

[0012] Step 1.2: Normalize the training data set, validation data set, and test data set. The data normalization formula is:

[0013]

[0014] Among them, X' is the standardized data, X is the original data without standardization, μ is the mean of X, and σ is the variance of X.

[0015] 2) Construct topological graph and positive and negative sample pairs

[0016] The variable topology map is constructed based on the industrial process knowledge, and the constructed topology map is enhanced. The process of topology map construction and map enhancement is as follows: Figure 1As shown, the process data is divided into several samples, each sample corresponds to two enhanced topological maps, the sample and the enhanced topological map obtain two graph representations through the encoder, and the first two graph representations are mapped into a new graph representation through projection. The two mappings of the same sample are regarded as positive pairs, and the projections between different samples are regarded as negative pairs.

[0017] Step 2.1: First, find out the relationship between variables based on the prior knowledge of the chemical process, and construct a variable topology map based on the relationship between variables;

[0018] Step 2.2: Use the edge perturbation graph enhancement method to disrupt the connectivity in the original topology graph by randomly adding or deleting a certain proportion of edges, thereby generating two enhanced topology graphs.

[0019] Step 2.3: Process data is defined as Where K is the number of samples, V is the number of variables, and T is the number of time steps for each sample. The enhanced graph G is obtained by performing edge perturbations on the topological graph of each sample. k,i and G k,j , where k represents the kth sample. This helps the model build positive and negative sample pairs.

[0020] 3) Modeling training and model testing

[0021] Construct a hierarchical topology graph isomorphism contrast learning model to train the training data set. The model structure of HTGICL is as follows: Figure 2 As shown, after training, the test data set is predicted;

[0022] Step 3.1: The graph encoder F[·] combines the process data X and its enhanced graph G k,i and G k,j Mapping to Embedding Representation and And the variables follow the topological relationship in the topological graph. The process is expressed as follows:

[0023] F[X|G k,i ,G k,j ]→Z k,i ,Z k,j (2)

[0024] Step 3.2: prioj(·) is called the projection head, which uses a two-layer MLP to map the enhanced graph representation to another latent space. Using prioj(·), Z k,i and Z k,j is mapped to Z' k,i and Z' k,j , calculate the contrast loss L c , the calculation process is as follows:

[0025]

[0026] where sim(·) represents the cosine similarity, which can be calculated as sim(Z' k,i ,Z' k,j )=(Z' k,i ·(Z' k,j ) T ) / (∥Z' k,i ∥·∥(Z' k,j ) T ∥), ∥·∥ means taking the second norm, k means the kth sample, and τ means the temperature coefficient.

[0027] Step 3.3: In order to enable the graph encoder F[·] to learn the layer topology isomorphism expertise, a layer topology isomorphism system S is established. iso To measure the isomorphic similarity between two graphs, the system uses a one-dimensional WL test to measure the isomorphic similarity between two graphs. The enhanced graph G of G is sampled from k samples k,i and G k,j , through the one-dimensional WL test, the layer topology isomorphic system S iso Get node label set A i and A j The number of iterations of the one-dimensional WL test is consistent with the number of layers of the encoder F[·]. The Jaccard coefficient is used to calculate the layer isomorphism similarity between graphs:

[0028]

[0029] where y iso Represents G k,i and G k,j The topological isomorphism similarity between them.

[0030] An unsupervised distillation task is proposed to introduce layer topology isomorphism expertise into the GCL method. Specifically, we train the graph representation obtained based on F[·] to predict G k,i and G k,j For prediction, G obtained by F[·] k,i and G k,j The representations are concatenated and fed into the MLP to obtain the predicted isomorphic similarity y' iso . This process can be described as:

[0031] y i ' so =MLP([Z k,i ; Z k,j ]) (5)

[0032] The mean square error is used to measure the ability of F[·] to learn the expertise of layer topology isomorphism and improve the ability of F[·] to capture topological discriminative information:

[0033]

[0034] Among them, L iso represents the layer homogeneity similarity loss, y k,iso Indicates the kth sample G k,i and G k,j The isomorphic similarity of y' k,iso Represents the isomorphic similarity of the kth sample obtained through prediction.

[0035] Step 3.4: Learn the expertise of sub-layer topology isomorphism;

[0036] Sublayer topology isomorphic system S subiso is introduced to guide F[·] to explore the topological isomorphism expertise of the sub-layers. The structural coefficient is introduced to capture the fine-grained topological information in the graph. Given a node v in a graph G, which is independently sampled from a dataset with m graphs, the neighborhood subgraph of node v (denoted as O v ) consists of itself, the nodes directly connected to v, and the corresponding edges. For two adjacent nodes v and u, the overlapping subgraph between them is defined as the intersection of their neighborhood subgraphs, i.e., O vu =O v ∩O u Formally, the structural coefficients w of v and u are vu This can be achieved by:

[0037]

[0038] Where m Vvu and m Evu Respectively represent O vu The number of nodes and edges in the cluster, and λ>0 is a hyperparameter. Then, w vu pass Normalize, where N(v) is the set of neighbors of v. In order to guide the encoder F[·] to learn fine-grained topological expertise, the structural coefficient matrix of G is derived to form the sub-layer topological isomorphic expertise y subiso We further use the nodes obtained from the first layer of F[·] to represent H G To fit the structural coefficient matrix:

[0039] y′ subiso =MLP(AUTOCOR(MLP(H G ))) (8)

[0040] where y' subisorepresents the predicted sub-layer topological isomorphism, MLP can be considered as the projection head, and AUTOCOR(·) represents the autocorrelation matrix calculation function. The mean square error loss is used to align the predictions of the sub-layer topological isomorphism system and F[·], thus deriving

[0041]

[0042] Among them, L subiso represents the sub-layer homogeneity similarity loss, y k,subiso Indicates the kth sample G k,i and G k,j The isomorphic similarity of y' k,subiso Represents the isomorphic similarity of the kth sample obtained through prediction.

[0043] To jointly learn layer and sub-layer topological isomorphic expertise, the loss of HTGICL is set to L c ,L iso and L subiso The weighted sum of is as follows:

[0044] L=L c +αL iso +βL subiso (10)

[0045] where α and β are the balance L during training iso and L subiso The coefficient of influence. Generally, α=10, β=1000.

[0046] Step 3.5: Input the test data set into the trained HTGICL model for testing, and use the trained comparison model to test the test data set;

[0047] 4) Data reconstruction and anomaly location

[0048] Reconstruct the high-dimensional embedding of the output and use the reconstruction loss to locate anomalies. Anomaly detection frameworks such as Figure 3 As shown;

[0049] Step 4.1: Use MLP to reconstruct the output embedding. The reconstructed data is consistent with the input data dimension. The reconstruction is as follows:

[0050] X1=MLP(Z) (11)

[0051] Among them, X1 represents the reconstructed output, and Z is the embedding of the input data X after training with the HTGICL model.

[0052] Using reconstruction loss to locate anomalies, the process is as follows:

[0053] Lrec =X-X1 (12)

[0054] Where L rec Represents the reconstruction loss matrix, which is used to locate the loss.

[0055] Step 4.2: Compare the test results of the HTGICL model with those of other comparison models.

[0056] The beneficial effects of the present invention are mainly manifested in: the present invention proposes an anomaly detection method based on hierarchical topological graph isomorphism contrast learning, constructs a variable topological graph based on prior knowledge, and then uses graph enhancement technology to help the model learn a more generalized graph representation. Subsequently, topological isomorphism expertise is injected into the graph comparison task to guide the encoder to mine distinguishable anomaly information from process data and graph representation. Compared with traditional anomaly detection models, the proposed model has better anomaly detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the process of topological graph construction and graph enhancement of the present invention;

[0058] Figure 2 It is a framework diagram of the present invention based on hierarchical topological graph isomorphism contrast learning;

[0059] Figure 3 is the anomaly detection framework of the present invention;

[0060] Figure 4 This is a schematic diagram of the sugar production process of the Lublin Sugar Factory of the present invention;

[0061] Figure 5 The present invention is based on the variable topology diagram of the sugar production process of Lublin Sugar Factory;

[0062] Figure 6 is the anomaly detection effect of the model proposed in this invention and the baseline model at 35 seconds of failure, Figure 6 Among them (a) VAE, (b) LSTMAE, (c) Transformer, (d) GCL, (e) HTGICL. DETAILED DESCRIPTION

[0063] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but the protection scope of the present invention is not limited thereto.

[0064] Example

[0065] (1) Divide the sugar production process data set of Lublin Sugar Factory as follows:

[0066] Step 1.1: Get the normal and abnormal data of the Lublin Sugar Factory dataset (Damadics benchmark data set, DMDS). A total of 86,400 normal data points are used for training, and the same number of abnormal data points are used for validation and testing, with a ratio of 1:1. The process data is divided into sliding windows with a length of 100 and a step size of 10. The training set, validation set, and test set of DMDS are 8631×100×32, 4315×100×32, and 4316×100×32 (samples×time steps×variables), respectively;

[0067] Step 1.2: Add noise (50% Gaussian noise) to the DMDS abnormal data and normal data.

[0068] Step 1.3: Normalize the data after adding noise.

[0069] (2) Constructing topological graphs and positive and negative sample pairs

[0070] Step 2.1: The DMDS was collected from an evaporation station of a sugar factory in Poland, which consists of 7 evaporators, 5 of which are natural circulation and 2 are pump circulation. It records the process data of the evaporation station in November 2001, including 32 process variables sampled per second, such as pressure, temperature, liquid density and flow rate. The recorded measured variables are shown in Table 1. When constructing the topology map, no time information is used. Figure 4 The schematic diagram of the DMDS process is shown, based on which the variable topology diagram is constructed, such as Figure 5 shown.

[0071] Step 2.2: Use edge perturbation to enhance the topological map corresponding to each sample, and obtain two enhanced topological maps for each sample.

[0072] Table 1 Sampling feature description table

[0073]

[0074]

[0075] Step 2.3: Process data is defined as Where K is the number of samples, V is the number of variables, and T is the number of time steps for each sample. The enhanced graph G is obtained by performing edge perturbations on the topological graph of each sample. k,i and G k,j , where k represents the kth sample. This helps the model build positive and negative sample pairs.

[0076] (3) Model training and model testing

[0077] Step 3.1: Establish the HTGICL model and input the constructed positive and negative sample pairs into HTGICL. HTGICL uses hierarchical topological isomorphism expertise to learn the representation between data.

[0078] Step 3.2: Train the model by optimizing the loss function composed of the loss obtained from the contrastive learning task and the loss obtained from the hierarchical topological isomorphism expertise.

[0079] Step 3.3: Input the test data set into the trained HTGICL model for testing, and use the trained comparison model to test the test data set;

[0080] (4) Data reconstruction and anomaly location

[0081] Step 4.1: Reconstruct the high-dimensional embedding of the output, and use the difference between the input data and the reconstructed data to calculate the reconstruction loss, calculate the mean and standard deviation of the reconstructed data, and identify data points that exceed the mean ± 3 standard deviations as abnormal;

[0082] Step 4.2: Compare the test results of the HTGICL model with those of other comparison models.

[0083] This model uses precision, recall, and F1 score to evaluate model performance:

[0084] Precision can be expressed as:

[0085]

[0086] Recall can be expressed as:

[0087]

[0088] The F1 score can be expressed as:

[0089]

[0090] Where TP represents the number of true positive instances (instances that actually occur and are correctly detected as anomalies), FP represents the number of false positive instances (instances that do not occur but are incorrectly labeled as anomalies), and FN represents the number of false negative instances (instances that actually occur but are incorrectly identified as normal). Precision refers to the proportion of truly abnormal instances among instances detected as anomalies. Recall refers to the proportion of instances detected as anomalies among instances where anomalies occur. The F1 score represents the harmonic mean of Precision and Recall and is usually used to evaluate the performance of anomaly detection models. F1∈[0,1], values ​​close to 1 indicate good performance.

[0091] Figure 6The anomaly detection effects of the proposed HTGICL model and the baseline comparison model under the 35-second pressure drop anomaly are shown, where the rectangular box represents the time and variable when the anomaly occurs, and the gray area represents the anomaly detected by the model. It can be seen that the anomaly detection effect of HTGICL is significant. According to the Precision, Recall and F1 scores in Table 2, it can be seen that the proposed HTGICL model has the best anomaly detection effect. Variational autoencoder (VAE), long short-term memory autoencoder (LSTMAE), Transformer, and GCL are selected as baseline models. The F1 score plays a vital role in evaluating the overall performance of the model. In practice, it is usually more valued than precision and recall. For example, compared with the baseline model, the F1 score of HTGICL is improved by 5.9-9.2%. This shows that the topological relationship between variables and the isomorphic expertise of the graph can effectively capture the abnormal information in industrial process data. Compared with HTGICL, GCL has significantly more misjudgments, which shows that the addition of topological isomorphic prior knowledge can enhance the model's ability to capture the propagation of variable anomaly information, thereby reducing misjudgments. At the same time, the model performance of HTGICL and GCL is better than LSTMAE, which is due to the lack of LSTMAE to capture the dependency of variables, which is prone to potential misjudgment. Although the baseline models have certain capabilities in detecting and locating anomalies, they often produce a large number of false positives. Therefore, they usually show the characteristics of high precision and low recall, which is unacceptable in practical applications. In contrast, HTGICL shows high precision and recall performance, which greatly facilitates the decision-making process of engineers. HTGICL shows good performance in low-noise scenarios, proving the rationality of its design.

[0092] Table 2 Performance comparison of five anomaly detection models for DMDS anomaly detection

[0093]

[0094] The anomaly detection method based on hierarchical topological graph isomorphic contrast learning proposed in the present invention has been fully verified on the Lublin sugar production process dataset, which improves the accuracy of anomaly detection and has universality and versatility.

[0095] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. An anomaly detection method based on hierarchical topological graph isomorphic contrast learning, characterized in that: The steps include: Step 1) Data acquisition and integration: According to the data of various variables collected during normal operation and abnormal operation in the chemical production process, the data of normal operation is used as the training data set, and the data of abnormal operation is used as the verification set and test set and normalized; Normal operation means that all variables are within the set range when the system is running; abnormal operation means that variables exceed the set range when the system is running; Step 2) Construct topological graph and positive and negative sample pairs: A variable topology map is constructed according to the industrial process knowledge, and a graph enhancement is constructed according to the constructed topology map. The process data is divided into several samples, each sample corresponds to two enhanced topology maps, and the sample and the enhanced topology map obtain two graph representations through an encoder. The first two graph representations are mapped into a new graph representation through projection. The two mappings of the same sample are regarded as positive pairs, and the projections between different samples are regarded as negative pairs. Step 3) Model training and model testing: Construct a hierarchical topology graph isomorphism comparison learning model to train the training data set, and at the same time establish a comparison model for model training. After the training is completed, predict the test data set; Step 4) Data reconstruction and anomaly location: Reconstruct the high-dimensional embedding of the output and use the reconstruction loss to locate anomalies.

2. The anomaly detection modeling method based on hierarchical topological graph isomorphism contrast learning as claimed in claim 1, characterized in that: The specific process of step 1) includes the following steps: Step 1.1: Use the collected normal industrial process data set as the training data set, and the abnormal industrial process data as the test data set. The ratio of the training data set, the verification data set, and the test data set is 2:1:1; Step 1.2: Normalize the training data set, validation data set, and test data set. The data normalization formula is: Among them, X' is the standardized data, X is the original data without standardization, μ is the mean of X, and σ is the variance of X.

3. The anomaly detection modeling method based on hierarchical topology isomorphism contrast learning as claimed in claim 2 is characterized in that: The specific process of step 2) includes the following steps: Step 2.1: First, find out the relationship between variables based on the prior knowledge of the chemical process, and construct a variable topology map based on the relationship between variables; Step 2.2: Use edge perturbation to enhance the graph connectivity by randomly adding or deleting edges to generate two enhanced topological graphs. Step 2.3: Process data is defined as Where K is the number of samples, V is the number of variables, and T is the number of time steps for each sample; the enhanced graph G is obtained by performing edge perturbations on the topological graph of each sample k,i and G k,j , where k represents the kth sample, thus helping the model to construct positive and negative sample pairs.

4. The anomaly detection modeling method based on hierarchical topology isomorphism contrast learning as claimed in claim 3 is characterized in that: The specific operation process of step 3) includes the following steps: Step 3.1: The graph encoder F[·] combines the process data X and its enhanced graph G k,i and G k,j Mapping to Embedding Representation and And the variables follow the topological relationship in the topological graph; the above process is expressed as follows: F[X|G k,i ,G k,j ]→Z k,i ,Z k,j (2) Step 3.2: prioj(·) is called the projection head, which uses a two-layer multi-layer perceptron MLP to map the enhanced graph representation to another latent space; using prioj(·), Z k,i and Z k,j is mapped to Z' k,i and Z' k,j , to calculate the contrast loss L c , the process is as follows: where sim(·,·) represents the cosine similarity, which can be calculated as sim(Z' k,i ,Z' k,j )=(Z' k,i ·(Z' k,j ) T ) / (∥Z' k,i ∥·∥(Z' k,j ) T ∥), ∥·∥ means taking the second norm, k means the kth sample, and τ means the temperature coefficient; Step 3.3: Use the one-dimensional Weissfeller-Lehman WL test to measure the isomorphism similarity between two graphs; sample the augmented graph G of G from k samples k,i and G k,j , through the one-dimensional WL test, the layer topology isomorphic system S iso Get node label set A i and A j ; The number of iterations of the one-dimensional WL test is consistent with the number of layers of the encoder F[·], and the Jaccard coefficient is used to calculate the layer isomorphism similarity between graphs: where y iso Represents G k,i and G k,j The topological isomorphism similarity between them; Train the graph encoder F[·] to predict G k,i and G k,j The isomorphic similarity of G obtained by the graph encoder F[·] k,i and G k,j The representations of are concatenated and fed into the MLP, which is regarded as a prediction head, and then the predicted isomorphic similarity y' is obtained iso The process is described as and i ′ so =MLP([Z k,i ;WITH k,j ]) (5) The mean square error is used to measure the ability of the graph encoder F[·] to learn the expertise of layer topology isomorphism and improve the ability of the graph encoder F[·] to capture topological discriminative information: Among them, L iso represents the layer homogeneity similarity loss, y k,iso Indicates the kth sample G k,i and G k,j The isomorphic similarity of y' k,iso Represents the isomorphic similarity of the kth sample obtained by prediction; Step 3.4: Learn the expertise of sub-layer topology isomorphism; The structural coefficient is introduced to capture the fine-grained topological information in the graph; given a node v in the graph G, node v is independently sampled from a dataset with m graphs, the neighborhood subgraph O of node v is v It consists of itself, the nodes directly connected to v, and the corresponding edges; for two adjacent nodes v and u, the overlapping subgraph between them is defined as the intersection of their neighborhood subgraphs, that is, O vu =O v ∩O u ; Formally, the structural coefficient w of v and u vu This is accomplished by: Where m Vvu and m Evu Respectively represent O vu The number of nodes and edges that are aggregated, and λ>0 is a hyperparameter; then, w vu pass Normalize, where N(v) is the set of neighbors of v; In order to guide the encoder F[·] to learn fine-grained topological expertise, the structural coefficient matrix of G is derived to form the sub-layer topological isomorphic expertise y subiso ; Further use the nodes obtained from the first layer of F[·] to represent H G To fit the structural coefficient matrix: y′ subiso =MLP(AUTOCOR(MLP(H G ))) (8) where y' subiso represents the predicted sub-layer topological isomorphism, MLP is considered as the projection head, AUTOCOR(·) represents the autocorrelation matrix calculation function; the mean square error loss is used to align the predictions of the sub-layer topological isomorphism system and F[·], thus deriving Among them, L subiso represents the sub-layer homogeneity similarity loss, y k,subiso Indicates the kth sample G k,i and G k,j The isomorphic similarity of y' k,subiso Represents the isomorphic similarity of the kth sample obtained by prediction; In order to jointly learn layer and sub-layer topological isomorphism expertise, the loss of the hierarchical topological graph isomorphism contrastive learning model HTGICL is set to L c ,L iso and L subiso The weighted sum of is as follows: L=L c +αL iso +βL subiso (10) where α and β are the balance L during training iso and L subiso The coefficient of influence; Step 3.5: Input the test data set into the trained HTGICL model for testing, and use the trained comparison model to test the test data set.

5. The method for anomaly detection based on hierarchical topology isomorphism contrast learning as claimed in claim 4, characterized in that: The specific operation process of step 4) includes the following steps: Step 4.1: Use MLP to reconstruct the output embedding. The reconstructed data is consistent with the input data dimension. The reconstruction is as follows: X1=MLP(Z) (11) Among them, X1 represents the reconstructed output, and Z is the embedding obtained after the input data X is trained by the HTGICL model; Using reconstruction loss to locate anomalies, the process is as follows: L rec =X-X1 (12) Where L rec Represents the reconstruction loss matrix, which is used to locate the loss; Step 4.2: Compare the test results of the HTGICL model and the comparison model.

Images