Neighborhood anti-signal-to-noise unwrapping anomaly detection method, computer and storage medium

Abstract

The invention discloses a neighborhood anti-signal-to-noise unwrapping anomaly detection method. The anomaly detection method aims at overcoming the limitation of a traditional deep learning method and providing higher-level automatic and intelligent support for equipment and system operation monitoring. According to the method, how to automatically screen out a standard set of normal data in an unsupervised anomaly detection scene is emphasized, a normal data mode is learned by mining essential characteristics of the standard set through neighborhood confrontation and signal-to-noise unwrapping, and anomaly detection is realized by using the difference between abnormal data and normal data mode reconstruction. By introducing neighborhood confrontation and signal-to-noise unwrapping into a deep learning model, data polluted by noise and abnormal data in complex equipment and systems are modeled, an active anomaly detection method is provided, the anomaly detection capability in a complex environment is improved, and the reliability, efficiency and performance of the equipment and the systems are improved.

Description

Technical Field

[0001] The present invention relates to the field of anomaly detection technology, and in particular to a neighborhood-resistant signal-to-noise untangling anomaly detection method, a computer device capable of implementing the method, and a computer-readable storage medium capable of implementing the method. Background Art

[0002] Intelligent operation and maintenance is an operation and maintenance management method that comprehensively utilizes cutting-edge technologies such as big data and artificial intelligence. It learns and summarizes the rules of massive operation and maintenance data to make intelligent decisions, thereby improving the reliability, efficiency and performance of equipment and systems, while reducing maintenance costs. Anomaly detection is an important direction in intelligent operation and maintenance, which can help enterprises detect and identify abnormal conditions in equipment or systems in a timely manner, so that they can take necessary maintenance or optimization measures to ensure the normal production and operation.

[0003] Traditional anomaly detection methods have some problems in the application stage, such as poor adaptability to complex systems, difficulty in processing high-dimensional data and nonlinear relationships, etc. In order to overcome these problems, deep learning technology has become a research hotspot in the field of intelligent operation and maintenance anomaly detection. However, the application of deep learning in the field of intelligent operation and maintenance anomaly detection still faces the problem of "data pollution" in actual scene data, that is, a small amount of abnormal data is mixed in the normal data. Since the data cannot be labeled in advance, abnormal data may interfere with the model building process. Summary of the invention

[0004] In order to improve the above technical problems, the technical solution of the first aspect of the present invention provides a neighborhood-resistant signal-noise disentanglement anomaly detection method for achieving more accurate, efficient and practical anomaly detection.

[0005] The neighborhood anti-noise signal-to-noise unwrapping anomaly detection method provided by the present invention, referred to as the anomaly detection method, comprises the following steps:

[0006] Step S1, neighborhood construction: obtain operation and maintenance data, extract data features from the operation and maintenance data, remove abnormal data in the operation and maintenance data according to the data features, obtain a standard set of normal data, and filter out a set of samples with similar features in the standard set according to a preset threshold r as a sample neighborhood;

[0007] Step S2, neighborhood confrontation and signal-noise disentanglement: Neighborhood confrontation and signal-noise disentanglement are introduced into the deep learning model to model the data polluted by noise and abnormal data in the operation and maintenance equipment and system, and the impact of environmental disturbance on the stability of the model is reduced by training the model;

[0008] Step S3, anomaly detection: Apply the trained model to the actual intelligent operation and maintenance scenario, and determine whether the input operation and maintenance data is mixed with abnormal data based on the output results.

[0009] Wherein, step S2 includes the following sub-steps:

[0010] Step S21, constructing a model including a primary feature extractor netE1, a data reconstructor netG, a second-order feature extractor netE2, and a data discriminator netD;

[0011] Step S22, initializing the parameters of the primary feature extractor netE1, the data reconstructor netG, the second-order feature extractor netE2, and the data discriminator netD in the model;

[0012] Step S23, randomly select m samples from the standard set as true samples x, and randomly select m samples from the sample neighborhoods corresponding to the m true samples x as anchor samples x′;

[0013] Step S24, generating a virtual sample according to the anchor sample x′ and the preset noise;

[0014] Step S25, fixing the parameters of the primary feature extractor netE1, the data reconstructor netG, and the second-order feature extractor netE2, and training the data discriminator netD, to improve the ability of the data discriminator netD to distinguish between real samples and virtual samples;

[0015] Step S26, fixing the parameters of the data discriminator netD, training the primary feature extractor netE1, the data reconstructor netG, and the second-order feature extractor netE2, so that the generated virtual samples are close to the real samples, and thus the features extracted by the second-order feature extractor netE2 are the same as the output after passing through the data discriminator netD;

[0016] Step S27, obtaining the maximum value of the mutual information between the features extracted by the primary feature extractor netE1 and the features extracted by the secondary feature extractor netE2;

[0017] Step S28, repeating steps S23 to S27 until the maximum value of the mutual information is less than the preset entropy value; or the number of virtual samples is close to the number of real samples in the judgment result of the data discriminator netD on the virtual samples for a preset number of consecutive times; or the number of iterations is greater than the preset number.

[0018] The anomaly detection method provided by the technical solution learns normal data patterns by mining the essential features of the standard set through neighborhood confrontation and signal-noise disentanglement, and realizes anomaly detection by using the difference between abnormal data and normal data pattern reconstruction. Among them, the negative impact of abnormal data is filtered out by using adversarial learning of similar samples in the standard set, and the essential features of similar samples are mined to realize neighborhood confrontation, and the essential features of the data are disentangled from the noise by maximizing the mutual information between the features extracted by the two feature extractors.

[0019] In the above technical solution, step S1 includes the following sub-steps:

[0020] Record the operation and maintenance data generated by intelligent operation and maintenance equipment or intelligent operation and maintenance systems;

[0021] Build a data set based on operation and maintenance data;

[0022] Extract data features from the dataset;

[0023] Divide the data set into coarse-grained form according to data features to obtain several category clusters;

[0024] According to the category cluster, the category cluster with less sample number than the preset threshold is regarded as abnormal data, and the abnormal data in the data set is removed to obtain the standard set, so that the operation and maintenance data in the standard set are all normal data;

[0025] In the standard set, according to the preset threshold r, a set of samples with similar features in the normal data is screened out to construct the sample neighborhood of the normal data in the standard set.

[0026] Furthermore, the following sub-steps are included between "constructing a data set according to operation and maintenance data" and "extracting data features from the data set" in step S1: preprocessing the data set; wherein the preprocessing method includes data cleaning and normalization processing.

[0027] In the above technical solution, in step S21, the primary feature extractor netE1 is used to extract feature v from the anchor sample x′; the data reconstructor netG is used to reconstruct the data of the temporary feature z generated by the feature v and the noise noise to obtain a virtual sample; the second-order feature extractor netE2 is used to extract feature v′1 from the real sample, and to extract feature v′2 from the virtual sample; the data discriminator netD is used to discriminate whether the sample is a real sample or a virtual sample according to the features extracted by the second-order feature extractor.

[0028] In the above technical solution, step S24 includes:

[0029] Pass the anchor sample x′ through the primary feature extractor netE1 to extract the feature v;

[0030] Generate noise according to normal distribution, and use feature v and noise to generate temporary feature z;

[0031] The temporary feature z is passed through the data reconstructor netG to generate a virtual sample.

[0032] Furthermore, the real sample x follows a normal distribution, the temporary feature z follows a normal distribution, and the noise noise follows a normal distribution.

[0033] In the above technical solution, in step S27, the maximum value of the mutual information is expressed as:

[0034] maxI(v;v′)=max(H(v)+H(v′)-H(v,v′))

[0035] Among them, I(v; v′) represents the mutual information between features v and v′, H(v) and H(v′) represent the information entropy of features v and v′ respectively, and H(v, v′) represents the joint entropy. In addition, it can be understood that v′ in this formula is the feature v′2 extracted from the virtual sample by the second-order feature extractor netE2.

[0036] In the above technical solution, in step S3, steps S1 and S2 are repeatedly executed to continuously update the parameters of the model.

[0037] In summary, the neighborhood-resistant signal-noise disentanglement anomaly detection method provided by the present invention can at least have the following beneficial effects:

[0038] 1. The generative adversarial network architecture is used as the basic model, which can better model nonlinear relationships and has strong adaptability to complex equipment and systems.

[0039] 2. Through neighborhood confrontation, this method learns the common characteristics of the neighborhood while avoiding abnormal data pollution, reduces the impact of environmental disturbances on model stability, and improves the robustness and generalization ability of the model.

[0040] 3. This anomaly detection method shows strong resistance to noise interference, and further improves the generalization ability of the model through signal-noise disentanglement, thereby improving the stability and reliability of anomaly detection in complex environments.

[0041] 4. The present invention provides an accurate, efficient and practical anomaly detection method, which can help operation and maintenance personnel to promptly discover potential problems and take preventive maintenance measures. The beneficial effects cover equipment reliability, cost, data security, maintenance efficiency and other aspects.

[0042] 5. This anomaly detection method has a wide range of applications, including industrial equipment, power systems, network operations, and manufacturing, and can provide a general anomaly detection solution for intelligent operation and maintenance systems in different industries.

[0043] The technical solution of the second aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the neighborhood anti-noise detangling anomaly detection method in any one of the technical solutions of the first aspect of the present invention can be implemented. Therefore, the computer device provided by the technical solution of the second aspect of the present invention has all the beneficial effects of any of the technical solutions of the first aspect of the present invention, which will not be repeated here.

[0044] The technical solution of the third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed, the steps of the neighborhood anti-noise detangling anomaly detection method in any one of the technical solutions of the first aspect of the present invention can be implemented. Therefore, the computer-readable storage medium provided by the technical solution of the third aspect of the present invention has all the beneficial effects of any of the technical solutions of the first aspect of the present invention, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The flowchart is a flowchart of a neighborhood-based signal-to-noise disentanglement anomaly detection method in one embodiment of the present invention.

[0046] Figure 2 The figure is a schematic diagram of the process of neighborhood construction in the anomaly detection method in one embodiment of the present invention.

[0047] Figure 3 The figure is a flow chart of the model architecture and model training in the anomaly detection method in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0049] An embodiment of the present invention provides a neighborhood-resistant signal-to-noise disentanglement anomaly detection method.

[0050] Example 1

[0051] like Figure 1 As shown, the neighborhood anti-signal-noise disentanglement anomaly detection method in this embodiment includes the following steps:

[0052] Step S10, recording sensor data, log records, operation records and other data generated by the device or system, constructing an unlabeled data set D(x), and performing preprocessing operations such as data cleaning and normalization on the data set;

[0053] Step S20, constructing a neighborhood for the preprocessed data set;

[0054] Step S30, performing neighborhood confrontation combined with signal-noise disentanglement on the sample domain set;

[0055] Step S40, applying the trained model to the actual intelligent operation and maintenance scenario. During the operation and maintenance process, new data is fed back to the model to continuously update the model parameters.

[0056] Example 2

[0057] On the basis of Example 1, Figure 2 As shown, in step S20, firstly, statistical analysis, data encoding and decoding, principal component analysis and other methods are used to extract data features. Secondly, a coarse-grained data set is divided using a clustering method based on distance and density to obtain several category clusters. Abnormal data is removed by rule-based or manual methods to obtain a standard set of normal data. Then, the process establishes a sample neighborhood according to the set threshold r through fine-grained clustering.

[0058] This step aims to filter out a set of samples with similar characteristics in the normal data and construct a neighborhood of normal samples in the standard set. The neighborhood involved in the present invention is defined as Indicates that through data enhancement method Generate a set of samples whose distance from sample x is r. r is an artificially set threshold used for further fine-grained clustering based on the coarse-grained data set division to establish the sample neighborhood.

[0059] The samples in the data set include both normal data and abnormal data. It can be considered that the category clusters with a sample number less than a preset threshold include abnormal data, thereby achieving rapid removal of abnormal data.

[0060] Example 3

[0061] On the basis of Example 1 or 2, Figure 3 As shown, the model architecture in step S30 includes a primary feature extractor netE1, a data reconstructor netG, a second-order feature extractor netE2, and a data discriminator netD.

[0062] like Figure 1 , Figure 3 As shown, step S30 includes the following process:

[0063] Step P10, initializing the parameters of netE1, netG, netE2, and netD in the model.

[0064] Step P20, randomly select m samples x from the standard set as real samples, randomly select m samples x' from the neighborhood corresponding to the m samples as anchor samples. Generate virtual samples through netE1 and netG using the anchor samples x' and noise noise.

[0065] Step P30, fixing netE1, netG, and netE2, and training netD to enhance the ability of netD to distinguish between real samples and generated virtual samples;

[0066] Step P40, fix the parameters of netD, train netE1, netG, and netE2 so that the model minimizes the gap between virtual samples and real samples. At the same time, by using signal-noise disentanglement, the data essential features and noise are disentangled by maximizing the mutual information between vv′ extracted by the netE1 and netE2 feature extractors to ensure that the disentangled essential features contain sufficient information related to the reconstructed data, discard redundant information related to specific fields, and improve the generalization ability of the features;

[0067] Step P50, back propagation and parameter update, repeat steps P20 to P40 until the fake samples generated by the generator are realistic enough and the discriminator cannot easily distinguish between real and fake samples, or the predetermined number of training iterations is reached.

[0068] This step includes a neighborhood feature learning method based on the generative adversarial network architecture, which forces the model to learn the common features of neighborhood samples, reduce the impact of environmental disturbances on model stability, and improve the robustness and generalization ability of the model. Figure 3 As shown, it includes the primary feature extractor netE1, the data reconstructor netG, the second-order feature extractor netE2, and the data discriminator netD. Neighborhood adversarial learning is performed through the maximum and minimum values ​​of the objective function between the data reconstructor netG and the data discriminator netD.

[0069] In addition, this step includes a technical means that revolves around the data not obeying the independent and identically distributed assumption. By maximizing the mutual information between v and v′ extracted by the netE1 and netE2 feature extractors, the essential features of the data and the noise are disentangled to ensure that the disentangled essential features contain sufficient information related to the reconstructed data, discard unnecessary information related to specific fields, and improve the generalization ability of the features.

[0070] It should be noted that the anchor samples are also samples in the standard set. They can be considered as real samples, but they are not selected in the random selection of P20.

[0071] In some specific embodiments, the sample may be document data, video data, image data, or audio data. For document data, the features extracted by the feature extractor may be specific paragraphs, keywords, etc. For video data, the features extracted by the feature extractor may be specific frames, images with key elements, etc. For image data, the features extracted by the feature extractor may be partial pixel maps, specific color areas, etc. For audio data, the features extracted by the feature extractor may be specific frames, audio within a specific frequency range, etc. In different scenarios (such as industrial equipment, power systems, network operations, and manufacturing), the data type of the sample and the features extracted by the feature extractor may be set accordingly.

[0072] Example 4

[0073] On the basis of Example 3, the "using anchor sample x′ and noise noise to generate virtual samples through netE1 and netG" in step P20 specifically includes: first, the anchor sample x′ is passed through netE1 to extract feature v, then noise noise is generated according to the normal distribution, and v and noise are used to generate a temporary feature z, z=netE1(x′)+noise, and finally the temporary feature z is passed through netG to generate a virtual sample.

[0074] Example 5

[0075] Based on Example 3, the objective function of the data discriminator netD in step P30 is:

[0076]

[0077] In the formula, V(netD, netG) represents the objective function, sample x follows the normal distribution μ, temporary feature z follows the normal distribution γ, and other represents the model part except netD.

[0078] Example 6

[0079] Based on Example 3, the objective function of the mutual information in step P40 is:

[0080] maxI(v;v′)=max(H(v)+H(v′)-H(v,v′))

[0081] Among them, I(v; v′) represents the mutual information between features v and v′, H(v) and H(v′) represent the information entropy of features v and v′ respectively, and H(v, v′) represents the joint entropy. In this way, the correlation or information sharing degree between v and v′ is quantified by mutual information, and signal-noise disentanglement removes the influence of noise on the reconstructed data by improving the correlation between v and v′.

[0082] Example 7

[0083] The difference from Example 6 is that the objective function of the mutual information in step P40 can also be expressed as:

[0084] max(netD(netE2(netG(z)))+I(v;v′)).

[0085] Example 8

[0086] Based on Example 3, the step P50 of “until the fake samples generated by the generator are realistic enough and the discriminator cannot easily distinguish between real and fake samples” can be specifically expressed as:

[0087] The maximum value of the mutual information is less than the preset entropy value.

[0088] That is, when the objective function of the mutual information between features v, v′ is less than the preset entropy value, it is considered that the features of the virtual sample after being extracted twice by the primary feature extractor netE1 and the second-order feature extractor netE2 are similar to the features of the real sample after being extracted by the second-order feature extractor netE2, which is sufficient to be confused.

[0089] Example 9

[0090] The difference from Example 8 is that in step P50, "until the fake samples generated by the generator are realistic enough and the discriminator cannot easily distinguish between true and false samples" can also be specifically expressed as:

[0091] In the result of the data discriminator netD judging the virtual samples for a preset number of consecutive times, the number of virtual samples is close to the number of real samples.

[0092] That is, after meeting the preset number of discriminations, the discrimination results of the data discriminator netD for virtual samples are close to 50% for both true and false. This indicates that the data discriminator netD cannot make discriminations based on the actual situation of the virtual samples, which also means that the features extracted from the virtual samples are similar to those of the real samples, which is enough to confuse the data discriminator netD.

[0093] In some specific embodiments, the final output result of the data discriminator netD is 0 or 1, 0 represents the virtual sample (false), and 1 represents the real sample (true). For example, after testing 100 or more samples, the number of 0s and 1s output by the data discriminator netD are close (the difference between the number of 0s and the number of 1s is no more than 5). For another example, after testing 300 or 500 samples, the number of 0s and 1s output by the data discriminator netD are basically the same (the difference between the number of 0s and the number of 1s is no more than 2). That is, it can be foreseen that the output of the data discriminator netD at this time is not based on the actual situation of the input sample, so that the output result is closer to a random number of 1 or 0.

[0094] In addition, another embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can implement the steps of the neighborhood-based signal-to-noise disentanglement anomaly detection method of any one of the above-mentioned embodiments.

[0095] In addition, another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, can implement the steps of the neighborhood-resistant signal-to-noise disentanglement anomaly detection method of any one of the above-mentioned embodiments.

[0096] It should be noted that those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, a magnetic tape storage, or any other medium of a computer scale that can be used to carry or store data.

[0097] In summary, the neighborhood adversarial signal-noise disentanglement anomaly detection method provided by the present invention aims to overcome the limitations of traditional deep learning methods and provide a higher level of automation and intelligent support for equipment and system operation monitoring. This method focuses on how to automatically screen out a standard set of normal data in an unsupervised anomaly detection scenario, learns normal data patterns by mining the essential characteristics of the standard set through neighborhood adversarial and signal-noise disentanglement, and realizes anomaly detection by utilizing the differences in reconstruction of abnormal data and normal data patterns. By introducing neighborhood adversarial and signal-noise disentanglement into the deep learning model, the invention realizes modeling of data contaminated by noise and abnormal data in complex equipment and systems, provides an active anomaly detection method, improves anomaly detection capabilities in complex environments, and achieves improvements in equipment and system reliability, efficiency, and performance.

[0098] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "specific embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0099] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A neighborhood-based anomaly detection method for signal-to-noise disentanglement, characterized in that: The steps include: Step S1, neighborhood construction: obtain operation and maintenance data, extract data features from the operation and maintenance data, remove abnormal data in the operation and maintenance data according to the data features, obtain a standard set of normal data, and filter out a set of samples with similar features in the standard set according to a preset threshold r as a sample neighborhood; Step S2, neighborhood confrontation and signal-to-noise detangling, includes the following sub-steps: Step S21, constructing a model including a primary feature extractor netE1, a data reconstructor netG, a second-order feature extractor netE2, and a data discriminator netD; Step S22, initializing the parameters of the primary feature extractor netE1, the data reconstructor netG, the second-order feature extractor netE2, and the data discriminator netD in the model; Step S23, randomly select m samples from the standard set as true samples x, and randomly select m samples from the sample neighborhoods corresponding to the m true samples x as anchor samples x′; Step S24, generating a virtual sample according to the anchor sample x′ and the preset noise; Step S25, fixing the parameters of the primary feature extractor netE1, the data reconstructor netG, and the second-order feature extractor netE2, and training the data discriminator netD, to improve the ability of the data discriminator netD to distinguish between real samples and virtual samples; Step S26, fixing the parameters of the data discriminator netD, training the primary feature extractor netE1, the data reconstructor netG, and the second-order feature extractor netE2, so that the generated virtual samples are close to the real samples; Step S27, obtaining the maximum value of the mutual information between the features extracted by the primary feature extractor netE1 and the features extracted by the secondary feature extractor netE2; Step S28, repeating steps S23 to S27 until the maximum value of the mutual information is less than the preset entropy value; or the number of virtual samples is close to the number of real samples in the result of the data discriminator netD judging the virtual samples for a preset number of consecutive times; or the number of iterations is greater than the preset number; Step S3, anomaly detection: Apply the trained model to the actual intelligent operation and maintenance scenario, and determine whether the input operation and maintenance data is mixed with abnormal data based on the output results.

2. The method for detecting anomalies by neighborhood-based signal-to-noise disentanglement according to claim 1, characterized in that: Step S1 includes the following sub-steps: Record the operation and maintenance data generated by intelligent operation and maintenance equipment or intelligent operation and maintenance systems; Build a data set based on operation and maintenance data; Extract data features from the dataset; Divide the data set into coarse-grained form according to data features to obtain several category clusters; According to the category cluster, the category cluster with less sample number than the preset threshold is regarded as abnormal data, and the abnormal data in the data set is removed to obtain the standard set, so that the operation and maintenance data in the standard set are all normal data; In the standard set, according to the preset threshold r, a set of samples with similar features in the normal data is screened out to construct the sample neighborhood of the normal data in the standard set.

3. The method for detecting anomalies in a neighborhood against signal-to-noise disentanglement according to claim 2, characterized in that: The following sub-steps are included between "building a data set based on operation and maintenance data" and "extracting data features from the data set" in step S1: Preprocess the data set; the preprocessing methods include data cleaning and normalization.

4. The method for detecting anomalies by neighborhood-based signal-to-noise disentanglement according to claim 1, characterized in that: In step S21, The primary feature extractor netE1 is used to extract feature v from the anchor sample x′; The data reconstructor netG is used to reconstruct the data of the temporary feature z generated by the feature v and the noise noise to obtain a virtual sample; The second-order feature extractor netE2 is used to extract feature v′1 from real samples and to extract feature v′2 from virtual samples; The data discriminator netD is used to discriminate whether the sample is a real sample or a virtual sample based on the features extracted by the second-order feature extractor.

5. The method for detecting anomalies by neighborhood-based signal-to-noise disentanglement according to claim 1, characterized in that: Step S24 includes: Pass the anchor sample x′ through the primary feature extractor netE1 to extract the feature v; Generate noise according to normal distribution, and use feature v and noise to generate temporary feature z; The temporary feature z is passed through the data reconstructor netG to generate a virtual sample.

6. The method for detecting anomalies in a neighborhood against signal-to-noise disentanglement according to claim 5, characterized in that: The real sample x follows a normal distribution, the temporary feature z follows a normal distribution, and the noise noise follows a normal distribution.

7. The neighborhood-resistant signal-noise disentanglement anomaly detection method according to claim 1, characterized in that: In step S27, the maximum value of the mutual information is expressed as: maxI(v;v′)=max(H(v)+H(v′)-H(v,v′)) Among them, I(v; v′) represents the mutual information between features v and v′, H(v) and H(v′) represent the information entropy of features v and v′ respectively, and H(v, v′) represents the joint entropy.

8. The method for detecting anomalies by using neighborhood-based signal-to-noise disentanglement according to any one of claims 1 to 7, characterized in that: In step S3, steps S1 and S2 are repeatedly executed to continuously update the parameters of the model.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the neighborhood anti-noise signal-to-noise disentanglement anomaly detection method as described in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the steps of the neighborhood anti-noise signal-to-noise disentanglement anomaly detection method as described in any one of claims 1 to 8 can be implemented.

Images