Abnormality detection intelligent recommendation method supporting online rapid evaluation

Through an intelligent recommendation method based on meta-learning and collaborative filtering, the algorithm performance evaluation problem in spacecraft anomaly detection is solved, and online rapid evaluation and recommendation is realized under the conditions of limited computing resources. It is suitable for spacecraft autonomous health management and intelligent operation and maintenance Scene.

CN119046828APending Publication Date: 2024-11-29BEIJING INST OF CONTROL ENG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411032569.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The prior art lacks intelligent recommendation methods for all components and systems in spacecraft anomaly detection and fault diagnosis, and cannot effectively evaluate algorithm performance online quickly, especially when computing resources are limited and data marking is difficult to obtain.

Method used

Using an intelligent recommendation algorithm based on meta-learning, combined with collaborative filtering method, a sublinear calculation method for data timing characteristics is designed to support the rapid online evaluation of the performance of anomaly detection algorithm, and the data similarity and The inherent relationship of algorithm matching.

Benefits of technology

It realizes intelligent recommendation of abnormal detection algorithms under the conditions of limited computing resources, supports spacecraft's independent health management and intelligent and precise operation and maintenance, and provides unified evaluation indicators and efficient intelligent recommendation solutions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046828A_ABST
    Figure CN119046828A_ABST
Patent Text Reader

Abstract

The invention relates to an anomaly detection intelligent recommendation method supporting online rapid evaluation, and belongs to the field of spacecraft anomaly detection and fault diagnosis. Comprising the following steps: step 1, determining a required representation symbol; step 2, taking # imgabs0 # and Ddec as input, substituting the # imgabs0 # and Ddec into a detection data preprocessing model # imgabs1 # to output preprocessed data # imgabs2 # and # imgabs3 #; step 3, taking data # imgabs4 # as input, substituting the data # imgabs4 # into a calculation distribution parameter model # imgabs5 #, and outputting p; 4, taking the preprocessed data # imgabs6 # and # imgabs7 # as input, and substituting the preprocessed data # imgabs6 # and # imgabs7 # into a filtering model # imgabs8 # to judge whether the preprocessed data # imgabs9 # is filtered or not; and step 5, establishing an anomaly detection recommendation model # imgabs10 # to obtain a final recommendation method, realizing an intelligent recommendation algorithm based on meta learning, fusing data features of a collaborative filtering method, and supporting rapid online evaluation of the performance of an anomaly detection algorithm based on time series data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of spacecraft anomaly detection and fault diagnosis, and relates to an anomaly detection intelligent recommendation method supporting online rapid evaluation. Background Art

[0002] In the field of spacecraft anomaly detection and fault diagnosis, a single algorithm is difficult to apply to all components, systems and working conditions. It is necessary to combine the characteristics of components, system characteristics and actual working conditions to design an intelligent recommendation method for anomaly detection and fault diagnosis that responds. At the same time, the working environment of a spacecraft is usually open, and it is difficult to form a complete algorithm to handle all conditions in advance. The performance of anomaly detection and fault diagnosis algorithms for different working environments and mission characteristics varies. When dealing with environments and states that the system has never encountered, it is necessary to combine the characteristics of various algorithms and data characteristics to design an intelligent recommendation method for anomaly detection for corresponding scenarios. Considering the extremely high cost of obtaining operation and maintenance data and the extremely limited computing resources of spacecraft, it is necessary to design an intelligent recommendation method for anomaly detection that supports online rapid evaluation.

[0003] At present, the intelligent recommendation technology for anomaly detection algorithms is in its infancy. The corresponding theoretical methods lack completeness, no practical solutions have been formed, and the needs for rapid quantitative evaluation of online performance have not been fully considered. There are the following deficiencies:

[0004] (1) Intelligent recommendation of algorithms is a difficult problem in the field of artificial intelligence. It is necessary to fully study the impact of data characteristics and algorithm characteristics on generalization ability and design effective generalization performance evaluation indicators. Limited by the current development status of artificial intelligence research, there is no effective intelligent recommendation method to meet the urgent needs of autonomous health management and intelligent and precise operation and maintenance of spacecraft. Recommendation algorithms designed for other general scenarios do not take into account the special needs of spacecraft operation and maintenance and are difficult to apply directly.

[0005] (2) Due to the technical difficulties of spacecraft information exchange and high cost, it is very difficult for anomaly detection algorithms to obtain data labels. Many existing intelligent algorithms are unsupervised learning algorithms. The evaluation standards of these algorithms are not unified, making it difficult to directly compare performance. In addition, the classic model evaluation method relies on the real labeling of the data, and it is also difficult to obtain data labels during the online operation of the operation and maintenance algorithm. It is difficult to use the classic model evaluation indicators to select the operation and maintenance algorithm.

[0006] (3) Since the computing resources of spacecraft are extremely limited, and the long-distance and long-delay operating characteristics make online quantitative evaluation based on algorithm performance the only option, algorithm intelligent recommendation technology needs to operate under the condition of limited computing resources, and it is necessary to design quantitative indicators and recommendation algorithms that support online rapid evaluation. Summary of the invention

[0007] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, propose an intelligent recommendation method for anomaly detection that supports online rapid evaluation, and can realize intelligent recommendation of algorithms under the condition of limited computing resources. According to the detection capability (detectability) of the anomaly detection algorithm, an intelligent recommendation algorithm based on meta-learning is realized, the data features of the collaborative filtering method are integrated, the algorithm performance is quantitatively ranked, and a sublinear calculation method considering the time series characteristics of the data is designed to support the rapid online evaluation of the anomaly detection algorithm based on time series data.

[0008] The solution of the present invention is:

[0009] An intelligent recommendation method for anomaly detection supporting online rapid evaluation, comprising:

[0010] Step 1: Determine the required representation symbols, including the data set used to build the anomaly detection intelligent model Represents a collection of anomaly detection models Represents the data set D to be tested in the online detection process dec , represents the anomaly detection model and data matching information set

[0011] Step 2: The data set used to build the anomaly detection intelligent model in step 1 And the dataset D to be tested dec Preprocess the detection data and output the preprocessed data and

[0012] Step 3: Calculate the preprocessed data in step 2 Similarity score; where a is the preset value; the output is p;

[0013] Step 4: Preprocess the data in step 2 and Perform filtering and judgment processing; wherein, It is the data set to be tested that has been preprocessed in the online detection process; the preprocessed data is judged Whether it is filtered;

[0014] Step 5: According to the output p in step 3 and the preprocessed data in step 4 Whether it is filtered or not, establish an anomaly detection recommendation model Method to get final recommendation.

[0015] In the above-mentioned anomaly detection intelligent recommendation method supporting online rapid evaluation, in step 1, the method for determining the representation symbol is:

[0016] make represents the data set used to build an anomaly detection intelligent model; where h represents the number of data; D i is the i-th data;

[0017] make represents the set of anomaly detection models; k represents the number of anomaly detection models; A j is the jth anomaly detection model;

[0018] Through offline training and testing, each anomaly detection model A is obtained j exist Performance on all data sets in ;

[0019] Let the anomaly detection model and data matching information set be gather Each of the j ,D i >The matching pairs of each item are the anomaly detection model A j In the dataset D i The performance is better than other models.

[0020] Let parameter S represent the discretization parameter required for the preprocessing process; S is a positive integer; set the accuracy requirement of the model filtering algorithm to ε; ε∈(0,1); set the filtering threshold of discrete distribution data items to γ, γ∈(0,1); set the parameter for controlling the granularity of discrete distribution description to α, α>0; set the parameter used to control the sampling probability of the training data set to β, β∈[0,1]; set the parameter used to control the size of the sampling block of the data to be detected to η, The parameter used to control the upper limit of uneven bucketing is set to κ, κ>0.

[0021] In the above-mentioned intelligent recommendation method for anomaly detection supporting online rapid evaluation, the D i Including training and testing data.

[0022] In the above-mentioned anomaly detection intelligent recommendation method supporting online rapid evaluation, each anomaly detection model A j They are all built using machine learning algorithms on different data.

[0023] In the above-mentioned method for intelligent recommendation of anomaly detection algorithms supporting online rapid evaluation, in step 2, the data set used for constructing anomaly detection intelligent model in step 1 is And the dataset D to be tested dec The method for preprocessing the detection data is:

[0024] Establishing a detection data preprocessing model ​

[0025] S11. Settings The input of the model is a data set And the dataset D to be tested dec , the output is the preprocessed data and

[0026] S12. For each data Do the following processing: Data D i The time series data obtained by the anomaly detection process is represented by {x1,x2,...,x n}; where n represents the detection time length; the detection data corresponding to time t∈[1,n] is an m-dimensional real vector; let x t,q Represents x t The q∈[1,m]th dimension data; each dimension data is a real number between 0 and 1;

[0027] S13. Data D i Discretize, calculate and output data The data corresponding to each moment t and dimension q in is

[0028] S14. Data Perform windowed transformation representation, that is, combine the sliding window data of length l; for time t∈[1,n], let the corresponding data for The superscript T indicates the transpose operation; is a vector of length; when t<l, the insufficient data is filled with a constant 0;

[0029] S15. Obtaining preprocessed data

[0030] S16. Data to be detected D dec Do the same processing as steps S12 to S15 to obtain the preprocessed data

[0031] S17. Output preprocessed data and

[0032] In the above-mentioned method for intelligent recommendation of anomaly detection algorithms supporting online rapid evaluation, in step 3, the data preprocessed in step 2 is calculated. The method for the similarity score is:

[0033] Establishing a model for calculating distributed parameters

[0034] Let n D For the dataset Scale;

[0035] a is the preset value; traverse Calculate the number of data that is exactly the same as a, denoted as c a Output

[0036] In the above-mentioned method for intelligent recommendation of anomaly detection algorithms supporting online rapid evaluation, in step 4, the data preprocessed in step 2 is and The method for filtering and judging is:

[0037] Building a Filter Model

[0038] S41. Order

[0039] S42, from Delete the data that falls in the set Z to construct the data set D1, and let Z′ be the set of the remaining data values;

[0040] S43, Order Where n = |Z′| is the size of the effective value range, and α is a parameter used to adjust the granularity of data distribution description;

[0041] S44, from Evenly select samples, and remove the data that falls in set Z to obtain data D2;

[0042] S45, completing the bucket construction of the data value range;

[0043] S46, Order

[0044] S47, obtain n data samples from D1, denoted as D S ;

[0045] S48, to D S Each data sample d in is processed as follows:

[0046] S481, from Find the bucket B containing d in i , add d to the set S i ;

[0047] S482, if Then continue; otherwise, return to step S481;

[0048] S483, S iRepresented as a set {a1,a2,...,a Y}; then calculate the vector pS i , denoted as [p1,p2,...,p Y ], represents the set S i The corresponding discrete distribution vector, i.e., p Y =ProCompute(S i ,a Y ), where y∈[1,Y];

[0049] S484, when and Then end and return 0; otherwise, set variable count = count + 1;

[0050] S485, when count == k, go to S410; otherwise return to step 481;

[0051] S49, if it has not returned yet, return to step S47;

[0052] S410, end and output result 1.

[0053] In the above-mentioned method for intelligent recommendation of anomaly detection algorithms supporting online rapid evaluation, in S45, the method for constructing data value range buckets is:

[0054] make

[0055] For any j∈[1,k-1], let

[0056]

[0057] In the above-mentioned method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation, in step 5, the anomaly detection model recommendation method The establishment method is:

[0058] S51. Obtaining preprocessed data and

[0059] S52, will As step three Model input and output The corresponding discrete distribution p dec , which is the output p of step three;

[0060] S53, initialize score = 0, A = empty;

[0061] S54, traverse i∈[1,h], use the method in step 4, for each data set Determine whether it is filtered;

[0062] S55. Return model A as a result.

[0063] In the above-mentioned method for intelligent recommendation of anomaly detection algorithms supporting online rapid evaluation, in S54, the method of judgment and processing is:

[0064] S541, when When , go to S542; otherwise, repeat this step to process the next round;

[0065] S542, Calculation The corresponding discrete distribution function p i ;

[0066] S543, calculate score i =1 / KL(p i ||p dec ), where KL(·||·) is the standard KL divergence;

[0067] S544, when score i > score, let score = score i , let A=A j Make

[0068] S545. When i==h, continue; otherwise, return to S541.

[0069] The beneficial effects of the present invention compared with the prior art are:

[0070] (1) The present invention proposes an intelligent recommendation method for anomaly detection that supports online rapid evaluation. This method addresses the deficiency of unsupervised anomaly detection algorithms that lack unified evaluation indicators, and considers the current situation that data labels are difficult to obtain in scenarios of autonomous health management and intelligent and precise operation and maintenance of spacecraft. Based on meta-learning and collaborative filtering methods, the intrinsic relationship between data similarity and algorithm matching is established based on detectability theory, and a solution combining offline learning with online recommendation is established, which supports intelligent recommendation of multiple existing spacecraft anomaly detection algorithms.

[0071] (2) Aiming at the intelligent recommendation requirements of anomaly detection algorithms, the present invention takes into account the actual situation of limited spacecraft computing resources, proposes a data element feature description method that combines statistical characteristics with time series characteristics, further designs a sublinear algorithm, establishes an online fast extraction method for data element features, supports online fast evaluation of data characteristics, and further supports the construction of efficient intelligent recommendation algorithms;

[0072] (3) The present invention first abstracts and describes the problem of intelligent recommendation of unsupervised operation and maintenance algorithms; then, by using the detection capability (detectability) of the anomaly detection algorithm on the data, the matching between the algorithm and the data and the similarity measurement index between the data are established, and an intelligent recommendation scheme based on meta-learning is designed; secondly, a data meta-feature extraction method supporting online evaluation is designed; thirdly, based on the idea of ​​matrix decomposition, the similarity meta-features of the data are transformed into fused meta-features supporting intelligent selection of the algorithm, and the data features of the collaborative filtering method are fused; finally, a data statistical characteristic maintenance algorithm based on an incremental algorithm is designed to support rapid online evaluation of data statistical characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of intelligent recommendation of the anomaly detection algorithm of the present invention;

[0074] Figure 2 This is a diagram of the intelligent recommendation results of the anomaly detection algorithm of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below in conjunction with the embodiments.

[0076] The present invention provides an intelligent recommendation method for anomaly detection that supports online rapid evaluation, which can realize intelligent recommendation of algorithms under the condition of limited computing resources. According to the detection capability (detectability) of the anomaly detection algorithm, intelligent recommendation based on meta-learning is realized, the data features of the collaborative filtering method are integrated, the performance of the anomaly detection algorithm is quantitatively ranked, and a sublinear calculation method considering the characteristics of time series data is designed to support rapid online evaluation of the performance of anomaly detection algorithms based on time series data.

[0077] Anomaly detection intelligent recommendation methods that support online rapid evaluation, such as Figure 1 As shown, the anomaly detection recommendation method specifically includes the following steps:

[0078] Step 1: Determine the required representation symbols, including the data set used to build the anomaly detection intelligent model Represents a collection of anomaly detection models Represents the data set D to be tested in the online detection process dec , represents the anomaly detection model and data matching information set

[0079] The method to determine the representation symbol is:

[0080] make represents the data set used to build an anomaly detection intelligent model; where h represents the number of data; D i is the i-th data. i Including training and testing data.

[0081] make represents the set of anomaly detection models; k represents the number of anomaly detection models; A j is the jth anomaly detection model. Each anomaly detection model A j They are all built using machine learning algorithms on different data.

[0082] Through offline training and testing, each anomaly detection model A is obtained j exist Performance of all datasets in .

[0083] Let the anomaly detection model and data matching information set be gather Each of the j ,D i >The matching pairs of each item are the anomaly detection model A j In the dataset D i The performance is better than other models.

[0084] Let parameter S represent the discretization parameter required for the preprocessing process; S is a positive integer; set the accuracy requirement of the model filtering algorithm to ε; ε∈(0,1); set the filtering threshold of discrete distribution data items to γ, γ∈(0,1); set the parameter for controlling the granularity of discrete distribution description to α, α>0; set the parameter used to control the sampling probability of the training data set to β, β∈[0,1]; set the parameter used to control the size of the sampling block of the data to be detected to η, The parameter used to control the upper limit of uneven bucketing is set to κ, κ>0.

[0085] Step 2: The data set used to build the anomaly detection intelligent model in step 1 And the dataset D to be tested dec Preprocess the detection data and output the preprocessed data and

[0086] The data set used to build the anomaly detection intelligent model in step 1 And the dataset D to be tested dec The method for preprocessing the detection data is:

[0087] Establishing a detection data preprocessing model

[0088] S11. Settings The input of the model is a data set And the dataset D to be tested dec , the output is the preprocessed data and ​

[0089] S12. For each data Do the following processing: Data D i The time series data obtained by the anomaly detection process is represented by {x1,x2,...,x n}; where n represents the detection time length; the detection data corresponding to time t∈[1,n] is an m-dimensional real vector; let x t,q Represents x t The data of the q∈[1,m]th dimension of ; the data of each dimension are real numbers between 0 and 1.

[0090] S13. Data D i Discretize, calculate and output data The data corresponding to each moment t and dimension q in is

[0091] S14. Data Perform windowed transformation representation, that is, combine the sliding window data of length l; for time t∈[1,n], let the corresponding data for The superscript T indicates the transpose operation; is a vector of length; when t<l, the insufficient data is filled with the constant 0.

[0092] S15. Obtaining preprocessed data

[0093] S16. Data to be detected D dec Do the same processing as steps S12 to S15 to obtain the preprocessed data

[0094] S17. Output preprocessed data and

[0095] Step 3: Calculate the preprocessed data in step 2 similarity score; where a is a preset value; the output is p.

[0096] Calculate the preprocessed data in step 2 The method for the similarity score is:

[0097] Establishing a model for calculating distributed parameters

[0098] Let n D For the dataset Scale;

[0099] a is the preset value; traverse Calculate the number of data that is exactly the same as a, denoted as c a Output

[0100] Step 4: Preprocess the data in step 2 and Perform filtering and judgment processing; wherein, It is the data set to be tested that has been preprocessed in the online detection process; the preprocessed data is judged Whether to be filtered.

[0101] After preprocessing in step 2, and The method for filtering and judging is:

[0102] Building a Filter Model

[0103] S41. Order

[0104] S42, from The data set D1 is constructed by deleting the data that falls in the set Z, and let Z′ be the set consisting of the remaining data values.

[0105] S43, Order Where n = |Z′| is the size of the effective value range, and α is a parameter used to adjust the granularity of data distribution description.

[0106] S44, from Evenly select samples, and remove the data that fall in set Z to obtain data D2.

[0107] S45. Complete the bucketing construction of the data value range.

[0108] In S45, the method for constructing the data value range bucket is:

[0109] make

[0110] For any j∈[1,k-1], let

[0111]

[0112] S46, Order

[0113] S47, obtain n data samples from D1, denoted as D S .

[0114] S48, to D SEach data sample d in is processed as follows:

[0115] S481, from Find the bucket B containing d in i , add d to the set S i .

[0116] S482, if Continue; otherwise, return to step S481.

[0117] S483, S i Represented as a set {a1,a2,...,a Y}; then calculate the vector pSi, denoted as [p1,p2,...,p Y ], represents the set S i The corresponding discrete distribution vector, i.e., p Y =ProCompute(S i ,a Y ), where y∈[1,Y].

[0118] S484, when and Then end and return 0; otherwise, set variable count = count + 1.

[0119] S485. When count == k, go to S410; otherwise return to step 481.

[0120] S49. If not returned yet, return to step S47.

[0121] S410, end and output result 1.

[0122] Step 5: According to the output p in step 3 and the preprocessed data in step 4 Whether it is filtered or not, establish an anomaly detection recommendation model Method to get final recommendation.

[0123] Anomaly Detection Model Recommendation Method The establishment method is:

[0124] S51. Obtaining preprocessed data and

[0125] S52, will As step three Model input and output The corresponding discrete distribution p dec , which is the output p of step three.

[0126] S53. Initialize score=0, A=empty.

[0127] S54, traverse i∈[1,h], use the method in step 4, for each data set Perform a determination process to determine whether or not the content has been filtered.

[0128] The judgment and processing methods are:

[0129] S541, when When , go to S542; otherwise, repeat this step to process the next round;

[0130] S542, Calculation The corresponding discrete distribution function p i ;

[0131] S543, calculate score i =1 / KL(p i ||p dec ), where KL(·||·) is the standard KL divergence;

[0132] S544, when score i > score, let score = score i , let A=A j Make

[0133] S545. When i==h, continue; otherwise, return to S541.

[0134] S55. Return model A as a result.

[0135] Example

[0136] Five data sets are used: DMDS, MSL, SMAP, SMD and WADI. The six traditional intelligent anomaly detection models considered in the experiment are: UAE, LSTM-AE, TCN-AE, LSTM-VAE, MSCRED and NASALSTM. The recommendation results of the patent of this invention (abbreviated as RecAD) are tested. The comparison standard used in the embodiment is the F1 score. A higher F1 score indicates a higher detection accuracy and a lower false alarm rate, that is, the recommended result should be the anomaly detection algorithm with the highest F1 score.

[0137] The F1 score is calculated as follows:

[0138]

[0139] in: TP represents the number of abnormal points identified as abnormal, FP represents the number of normal points identified as abnormal, and FN represents the number of abnormal points identified as normal. These three numbers can be obtained by comparing the normal / abnormal labels of the data output by the algorithm with the true labels.

[0140] From the experimental results, it can be found that the intelligent recommendation method (RecAD) proposed in this patent can obtain the best anomaly detection algorithm, which is completely consistent with the actual situation. Figure 2 shown.

[0141] The present invention proposes an intelligent recommendation method for anomaly detection that supports online rapid evaluation. This method targets the current situation that unsupervised anomaly detection algorithms lack unified evaluation indicators, and takes into account the current situation that data labels are difficult to obtain in scenarios of autonomous health management and intelligent and precise operation and maintenance of spacecraft. Based on meta-learning, the collaborative filtering method is used, and the intrinsic relationship between data similarity and algorithm matching is established based on detectability theory. A solution combining offline learning with online recommendation is established, and the intelligent recommendation of various existing spacecraft anomaly detection algorithms is supported. Aiming at the intelligent recommendation needs of anomaly detection algorithms, the present invention takes into account the actual situation that spacecraft computing resources are limited, proposes a data element feature description method that combines statistical characteristics with time series characteristics, and further designs a sublinear algorithm, establishes an online rapid extraction method for data element features, supports online rapid evaluation of data characteristics, and further supports the construction of efficient intelligent recommendation algorithms.

[0142] The present invention firstly abstracts and describes the problem of intelligent recommendation of unsupervised operation and maintenance algorithms; then, by using the detection capability (detectability) of anomaly detection algorithms on data, the matching between algorithms and data and the similarity measurement indicators between data are established, and an intelligent recommendation scheme based on meta-learning is designed; secondly, a data meta-feature extraction method supporting online evaluation is designed; thirdly, based on the idea of ​​matrix decomposition, the similarity meta-features of data are transformed into fused meta-features supporting intelligent selection of algorithms, and the data features of collaborative filtering methods are fused; finally, a data statistical characteristic maintenance algorithm based on an incremental algorithm is designed to support rapid online evaluation of data statistical characteristics.

[0143] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. An intelligent recommendation method for anomaly detection supporting online rapid evaluation, characterized in that: include: Step 1: Determine the required representation symbols, including the data set used to build the anomaly detection intelligent model Represents a collection of anomaly detection models Represents the data set D to be tested in the online detection process dec , represents the anomaly detection model and data matching information set Step 2: The data set used to build the anomaly detection intelligent model in step 1 And the dataset D to be tested dec Preprocess the detection data and output the preprocessed data and Step 3: Calculate the preprocessed data in step 2 Similarity score; where a is the preset value; the output is p; Step 4: Preprocess the data in step 2 and Perform filtering and judgment processing; wherein, It is the data set to be tested that has been preprocessed in the online detection process; the preprocessed data is judged Whether it is filtered; Step 5: According to the output p in step 3 and the preprocessed data in step 4 Whether it is filtered or not, establish an anomaly detection recommendation model Method to get final recommendation.

2. According to claim 1, the method for intelligent recommendation of anomaly detection supporting online rapid evaluation is characterized in that: In the step 1, the method for determining the representation symbol is: make represents the data set used to build an anomaly detection intelligent model; where h represents the number of data; D i is the i-th data; make represents the set of anomaly detection models; k represents the number of anomaly detection models; A j is the jth anomaly detection model; Through offline training and testing, each anomaly detection model A is obtained j exist Performance on all data sets in ; Let the anomaly detection model and data matching information set be gather Each of the j ,D i The matching pairs of each item are the anomaly detection model A j In the dataset D i The performance is better than other models.​ Let parameter S represent the discretization parameter required for the preprocessing process; S is a positive integer; set the accuracy requirement of the model filtering algorithm to ε; ε∈(0,1); set the filtering threshold of discrete distribution data items to γ, γ∈(0,1); set the parameter that controls the granularity of discrete distribution description to α, α>0; set the parameter used to control the sampling probability of the training data set to β, β∈[0,1]; set the parameter used to control the size of the sampling block of the data to be detected to The parameter used to control the upper limit of uneven bucketing is set to κ, κ>

0.

3. The method for intelligent recommendation of anomaly detection supporting online rapid evaluation according to claim 2, characterized in that: The D i Including training and testing data.

4. The method for intelligent recommendation of anomaly detection supporting online rapid evaluation according to claim 2, characterized in that: Each anomaly detection model A j They are all built using machine learning algorithms on different data.

5. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 2, characterized in that: In step 2, the data set used to construct the anomaly detection intelligent model in step 1 is And the dataset D to be tested dec The method for preprocessing the detection data is: Establishing a detection data preprocessing model S11. Settings The input of the model is a data set And the dataset D to be tested dec , the output is the preprocessed data and S12. For each data Do the following processing: Data D i The time series data obtained by the anomaly detection process is represented by {x1,x2,...,x n }; where n represents the detection time length; the detection data corresponding to time t∈[1,n] is an m-dimensional real vector; let x t,q Represents x t The q∈[1,m]th dimension data; each dimension data is a real number between 0 and 1; S13. Data D i Discretize, calculate and output data The data corresponding to each moment t and dimension q in is S14. Data Perform windowed transformation representation, that is, combine the sliding window data of length l; for time t∈[1,n], let the corresponding data for The superscript T indicates the transpose operation; is a vector of length; when t<l, the insufficient data is filled with a constant 0; S15. Obtaining preprocessed data S16. Data to be detected D dec Do the same processing as steps S12 to S15 to obtain the preprocessed data S17. Output preprocessed data and 6. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 5, characterized in that: In step 3, the data preprocessed in step 2 is calculated. The method for the similarity score is: Establishing a model for calculating distributed parameters Let n D For the dataset Scale; a is the preset value; traverse Calculate the number of data that is exactly the same as a, denoted as c a ; Output 7. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 6, characterized in that: In the step 4, the data preprocessed in step 2 is and The method for filtering and judging is: Building a Filter Model S41. Order S42, from Delete the data that falls in the set Z to construct the data set D1, and let Z′ be the set of the remaining data values; S43, Order Where n = |Z′| is the size of the effective value range, and α is a parameter used to adjust the granularity of data distribution description; S44, from Evenly select samples, and remove the data that falls in set Z to obtain data D2; S45, completing the bucket construction of the data value range; S46, Order S47, obtain n data samples from D1, denoted as D S ; S48, to D S Each data sample d in is processed as follows: S481, from Find the bucket B containing d in i , add d to the set S i ; S482, if Then continue; otherwise, return to step S481; S483, S i Represented as a set {a1,a2,...,a Y }; then calculate the vector pS i , denoted as [p1,p2,...,p Y ], represents the set S i The corresponding discrete distribution vector, i.e., p Y =ProCompute(S i ,a Y ), where y∈[1,Y]; S484, when and Then end and return 0; Otherwise, let variable count = count + 1; S485, when count == k, enter S410; Otherwise, return to step 481; S49, if it has not returned yet, return to step S47; S410, end and output result 1.

8. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 7, characterized in that: In S45, the method for constructing the data value range bucket is: make For any j∈[1,k-1], let 9. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 8, characterized in that: In step 5, the anomaly detection model recommendation method The establishment method is: S51. Obtaining preprocessed data and S52, will As step three Model input and output The corresponding discrete distribution p dec , which is the output p of step three; S53, initialize score = 0, A = empty; S54, traverse i∈[1,h], use the method in step 4, for each data set Determine whether it is filtered; S55. Return model A as a result.

10. The method for intelligently recommending anomaly detection algorithms supporting online rapid evaluation according to claim 9, characterized in that: In S54, the method of determining and processing is: S541, when When , go to S542; otherwise, repeat this step to process the next round; S542, Calculation The corresponding discrete distribution function p i ; S543, calculate score i =1 / KL(p i ||p dec ), where KL(·||·) is the standard KL divergence; S544, when score i > score, let score = score i , let A=A j Make S545. When i==h, continue; otherwise, return to S541.