A multi-cloud data anomaly detection method and system based on deep learning

Through the deep learning network, the time series of multi-cloud data is processed and the anomaly score threshold is calculated, which solves the problem of lack of labels and features in multi-cloud data acquisition, and more accurate anomaly detection is achieved, improving the stability and data quality of multi-cloud platforms.

CN114332537BActive Publication Date: 2024-11-26CHANJET INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111654479.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-30
Publication Date
2024-11-26
Estimated Expiration
2041-12-30

AI Technical Summary

Technical Problem

The amount of data collected by multi-cloud data is large and lacks abnormal labels. It is difficult to use existing machine learning algorithms directly, resulting in low abnormal detection efficiency and affecting the stability and robustness of multi-cloud platforms.

Method used

Using a multi-cloud data anomaly detection method based on deep learning, the time series data of the multi-cloud platform is collected, and it is divided into different subsets for hyperparameter selection, training and testing. The reconstruction error and anomaly score threshold are calculated using the deep learning network. Achieve more accurate abnormality detection.

Benefits of technology

It improves the accuracy and efficiency of multi-cloud data anomaly detection, quickly locates data anomalies, and ensures the stability and robustness of multi-cloud platform.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114332537B_ABST
    Figure CN114332537B_ABST
Patent Text Reader

Abstract

The present invention proposes a multi-cloud data anomaly detection method and system based on deep learning. The method includes: inputting the time series of data into a trained deep learning network to obtain a reconstructed sequence of data; applying the time series of data and the reconstructed sequence of data to calculate the reconstruction error, and then applying the reconstruction error to calculate the anomaly score threshold; if the value in the time series of real-time data exceeds the anomaly score threshold, the value is considered to be an anomaly. The solution proposed in the present invention is based on the multi-cloud data anomaly detection implemented by the deep learning network to provide more accurate anomaly alarms, thereby improving the accuracy of multi-cloud data component data, quickly locating data anomalies, and ensuring the stability and robustness of the multi-cloud platform.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of multi-cloud data anomaly detection, and in particular, relates to a multi-cloud data anomaly detection method and system based on deep learning. Background Art

[0002] With the continuous implementation of cloud computing services, hybrid cloud has become the mainstream IT infrastructure architecture for enterprises. There are many hybrid cloud management components. In order to ensure the access efficiency of users, it is necessary to persist the information of multi-cloud components so as to uniformly manage hybrid cloud resources. When the amount of multi-cloud resource data is particularly large, the abnormality of the scheduled synchronization of multi-cloud resource data will inevitably affect the upper-level projects based on multi-cloud management (such as multi-cloud cost management, multi-cloud resource application platform, etc.). Although the alarm method based on custom rules can realize the function of anomaly detection when the multi-cloud resource data is small, when the multi-cloud resources continue to grow, the rule-based alarm method will have the problem of poor maintenance. The multi-cloud data unsupervised anomaly detection based on deep learning is mainly for more accurate anomaly alarm, thereby improving the quality of multi-cloud data, reducing synchronized data anomalies, and ensuring the stability and robustness of the multi-cloud platform.

[0003] There are two pain points in multi-cloud data collection anomaly detection. First, the data collected by multi-cloud is huge and data anomaly labels do not exist, so it can only be manually labeled or modeled by unsupervised learning. Second, the data collected by multi-cloud has characteristics such as time, trend, and cycle, and most machine learning algorithms cannot be used directly.

[0004] Defects of the existing technology

[0005] 1. There is a lot of data collected by multi-cloud and there are no data anomaly labels, so you can only manually label it or choose supervised learning to achieve modeling;

[0006] 2. The data collected by multi-cloud has characteristics such as time, trend and cycle, which most machine learning algorithms cannot use directly. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a technical solution of a multi-cloud data anomaly detection method based on deep learning to solve the above technical problems.

[0008] The first aspect of the present invention discloses a multi-cloud data anomaly detection method based on deep learning; the method comprises:

[0009] Step S1, collecting time series of data from multiple cloud platforms;

[0010] The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing;

[0011] Step S2, inputting the time series of the data of the N2 subset into a deep learning network for training, and outputting a reconstructed sequence;

[0012] Step S3, inputting the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset;

[0013] Step S4, using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset to calculate the reconstruction error of the A1 subset, and then using the reconstruction error of the A1 subset to calculate the abnormal score threshold τ;

[0014] Step S5: If the value in the time series of the real-time data exceeds the abnormal score threshold, the value is considered to be an abnormal value.

[0015] According to the method of the first aspect of the present invention, in step S4, the specific method of calculating the abnormal score threshold τ includes:

[0016] The reconstruction error of the A1 subset is used to calculate the abnormal estimation threshold a(A1) of the A1 subset, and the abnormal estimation threshold is used as the abnormal score threshold τ. The specific formula includes:

[0017] a(·)=(e (i) -μ) T Σ -1 (e (i) -μ)

[0018] in,

[0019] μ and Σ are the parameters of the normal distribution N(μ,Σ);

[0020] e (i) is the reconstruction error of a certain subset;

[0021] e (i) =X (i) -X ’(i) ;

[0022] X (i) A time series of a subset of data;

[0023] X ’(i) is the reconstructed sequence of a subset of data;

[0024] T is transpose.

[0025] According to the method of the first aspect of the present invention, in step S4, the specific method of calculating the abnormal score threshold τ also includes:

[0026] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0027] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0028] The AUC value is used as the anomaly score threshold τ.

[0029] According to the method of the first aspect of the present invention, in step S4, the specific method of calculating the abnormal score threshold τ also includes:

[0030] Calculating an abnormal estimation threshold a(A1) of the A1 subset by applying the reconstruction error of the A1 subset;

[0031] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0032] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0033] The larger value of the anomaly estimation threshold and the AUC value of the A1 subset is taken as the anomaly score threshold τ.

[0034] According to the method of the first aspect of the present invention, in step S2, the input of the deep learning network further includes:

[0035] Statistical indicators of the data in the N2 subset.

[0036] According to the method of the first aspect of the present invention, in step S2, the statistical indicators include:

[0037] The maximum value of the data in the N2 subset, the minimum value of the data in the N2 subset, the mean of the data in the N2 subset, the median of the data in the N2 subset, and the variance of the data in the N2 subset.

[0038] According to the method of the first aspect of the present invention, in step S2, the input of the deep learning network further includes:

[0039] The time series of the data in the N2 subset is in reverse order.

[0040] A second aspect of the present invention discloses a multi-cloud data anomaly detection system based on deep learning; the system comprises:

[0041] A first processing module is configured to collect time series of data from multiple cloud platforms;

[0042] The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing;

[0043] A second processing module is configured to input the time series of the data of the N2 subset into a deep learning network for training, and output a reconstructed sequence;

[0044] A third processing module is configured to input the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset;

[0045] A fourth processing module is configured to calculate a reconstruction error of the A1 subset using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset, and then calculate an abnormal score threshold τ using the reconstruction error of the A1 subset;

[0046] The fifth processing module is configured to consider a value as an abnormal value if the value in the time series of the real-time data exceeds the abnormal score threshold.

[0047] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in any one of the multi-cloud data anomaly detection methods based on deep learning in the first aspect of the present disclosure are implemented.

[0048] A fourth aspect of the present invention discloses a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps in any one of the multi-cloud data anomaly detection methods based on deep learning in the first aspect of the present disclosure are implemented.

[0049] The solution proposed in the present invention provides more accurate anomaly alarms for multi-cloud data anomaly detection based on deep learning networks, thereby improving the accuracy of multi-cloud data component data, quickly locating data anomalies, and ensuring the stability and robustness of the multi-cloud platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 Flow chart of a multi-cloud data anomaly detection method based on deep learning according to an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the encoding and decoding process of the LSTM according to an embodiment of the present invention;

[0053] Figure 3 A structural diagram of a multi-cloud data anomaly detection system based on deep learning according to an embodiment of the present invention;

[0054] Figure 4 The figure is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the 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 making creative work are within the scope of protection of the present invention.

[0056] A first aspect of the present invention discloses a multi-cloud data anomaly detection method based on deep learning. Figure 1 : is a flowchart of a multi-cloud data anomaly detection method based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] Step S1, collecting time series of data from multiple cloud platforms;

[0058] The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing;

[0059] Step S2, inputting the time series of the data of the N2 subset into a deep learning network for training, and outputting a reconstructed sequence;

[0060] In some embodiments, in step S2, the input of the deep learning network further includes:

[0061] Statistical indicators of the data of the N2 subset;

[0062] In some embodiments, in step S2, the statistical indicators include:

[0063] The maximum value of the data in the N2 subset, the minimum value of the data in the N2 subset, the mean of the data in the N2 subset, the median of the data in the N2 subset, and the variance of the data in the N2 subset;

[0064] In some embodiments, count_above_mean and count_below_mean are further selected as features to calculate the number of elements greater than the mean and the number of elements less than the mean, respectively. The method here is to first calculate the mean, then calculate the difference between each time series point and the mean, and then the neural network module calculates the number of differences greater than zero and the number of differences less than zero.

[0065] In some embodiments, in step S2, the input of the deep learning network further includes:

[0066] The time series of the data in the N2 subset is in reverse order;

[0067] In some embodiments, the deep learning network uses LSTM, and the Autoencoder-LSTM consists of two LSTM units, one as an encoder and the other as a decoder; the encoder first encodes the training data set into a code (hidden layer), and then decodes the code through the decoder. During training, the encoder and the decoder are trained at the same time to try to make the decoded information as similar as possible to the information before encoding; the input of the encoder is of dimension<MB,T,D> The dataset is a dataset of , where MB (Batch size) is the number of data windows contained in the mini-batch, T (LSTM unit input window length) is the number of data points in each data window, and D represents the data dimension. In the initialization phase, MB and T are used for parameter learning on the decoder.

[0068] like Figure 2 As shown in the figure, because the architecture contains multiple data windows, in the Encoder stage, the encoded output of the previous data window is used as the input of the next time window; when encoding the last data window, the output can be considered as a hidden layer data structure; a copy of the output data is used as the input of the decoder, and the decoding operation is performed on each previous data window at a time;

[0069] Step S3, inputting the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset;

[0070] Step S4, using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset to calculate the reconstruction error of the A1 subset, and then using the reconstruction error of the A1 subset to calculate the abnormal score threshold τ;

[0071] In some embodiments, in step S4, the specific method of calculating the anomaly score threshold τ includes:

[0072] The reconstruction error of the A1 subset is used to calculate the abnormal estimation threshold a(A1) of the A1 subset, and the abnormal estimation threshold is used as the abnormal score threshold τ. The specific formula includes:

[0073] a(·)=(e (i) -μ) T Σ -1 (e (i) -μ)

[0074] in,

[0075] μ and Σ are the parameters of the normal distribution N(μ,Σ);

[0076] e (i) is the reconstruction error of a certain subset;

[0077] e (i) =X (i) -X ’(i) ;

[0078] X (i) A time series of a subset of data;

[0079] X ’(i) is the reconstructed sequence of a subset of data;

[0080] T is transpose;

[0081] In some embodiments, in step S4, the specific method of calculating the anomaly score threshold τ further includes:

[0082] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0083] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0084] The AUC value is used as anomaly score threshold τ;

[0085] In some embodiments, in step S4, the specific method of calculating the anomaly score threshold τ further includes:

[0086] Calculating an abnormal estimation threshold a(A1) of the A1 subset by applying the reconstruction error of the A1 subset;

[0087] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0088] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0089] The larger value of the abnormal estimation threshold and the AUC value of the A1 subset is taken as the abnormal score threshold τ;

[0090] Step S5: If the value in the time series of the real-time data exceeds the abnormal score threshold, the value is considered to be an abnormal value.

[0091] In some embodiments, the deep learning network is updated online. If the deep learning network is used for streaming data, the accuracy of the existing model for anomaly detection will decrease because the initial training data set is relatively small and simple, and the anomaly detection changes over time. Therefore, the update of the model of the deep learning network is essential.

[0092] After the initial training of the Autoencoder-LSTM model, the preliminary model parameters have been trained and online prediction can be performed. Here, a multi-threaded approach is used to manage the online anomaly detection model.

[0093] In real-time anomaly detection, the child thread collects the data generated in real time, while the main thread works on real-time anomaly detection. For each window in the batch, each data is reconstructed and the anomaly score is calculated by using the formula:

[0094] a(i)=(e (i) -μ) T Σ -1 (e (i) -μ)

[0095] The system maintains two data buffers, one for normal data windows and the other for abnormal data windows. Considering that the window reconstruction error of normal data is small, while the reconstruction error of abnormal data is large / new data features are generated. This reconstruction error level can be measured by a predefined reconstruction error distribution or error distribution.

[0096] In summary, the solution proposed in the present invention can provide more accurate anomaly alarms for multi-cloud data anomaly detection based on deep learning networks, thereby improving the accuracy of multi-cloud data component data, quickly locating data anomalies, and ensuring the stability and robustness of the multi-cloud platform.

[0097] A second aspect of the present invention discloses a multi-cloud data anomaly detection system based on deep learning. Figure 3 is a structural diagram of a multi-cloud data anomaly detection system based on deep learning according to an embodiment of the present invention; Figure 3 As shown, the system 100 includes:

[0098] The first processing module 101 is configured to collect time series of data from multiple cloud platforms;

[0099] The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing;

[0100] The second processing module 102 is configured to input the time series of the data of the N2 subset into a deep learning network for training, and output a reconstructed sequence;

[0101] The third processing module 103 is configured to input the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset;

[0102] The fourth processing module 104 is configured to calculate the reconstruction error of the A1 subset by using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset, and then calculate the abnormal score threshold τ by using the reconstruction error of the A1 subset;

[0103] The fifth processing module 105 is configured to consider a value as an abnormal value if the value in the time series of the real-time data exceeds the abnormal score threshold.

[0104] According to the system of the second aspect of the present invention, the second processing module 102 is specifically configured such that the input of the deep learning network further includes:

[0105] Statistical indicators of the data of the N2 subset;

[0106] The statistical indicators include:

[0107] The maximum value of the data in the N2 subset, the minimum value of the data in the N2 subset, the mean of the data in the N2 subset, the median of the data in the N2 subset, and the variance of the data in the N2 subset;

[0108] Count_above_mean and count_below_mean are also selected as features to calculate the number of elements greater than the mean and the number of elements less than the mean respectively; the method here is to first calculate the mean, then calculate the difference between each time series point and the mean, and then the neural network module calculates the number of differences greater than zero and the number of differences less than zero;

[0109] The input of the deep learning network also includes:

[0110] The time series of the data in the N2 subset is in reverse order;

[0111] The deep learning network uses LSTM. Autoencoder-LSTM consists of two LSTM units, one as an encoder and the other as a decoder. The encoder first encodes the training data set into a code (hidden layer), and then decodes the code through the decoder. During training, the encoder and decoder are trained simultaneously to try to make the decoded information as similar as possible to the information before encoding. The input of the encoder is of dimension<MB,T,D> The dataset is a dataset of , where MB (Batch size) is the number of data windows contained in the mini-batch, T (LSTM unit input window length) is the number of data points in each data window, and D represents the data dimension. In the initialization phase, MB and T are used for parameter learning on the decoder.

[0112] like Figure 2 As shown in the figure, because the architecture contains multiple data windows, in the Encoder stage, the encoded output of the previous data window is used as the input of the next time window; when encoding the last data window, the output can be considered as a hidden layer data structure; the output data is copied as the input of the decoder, and the decoding operation is performed on each previous data window at a time.

[0113] According to the system of the second aspect of the present invention, the fourth processing module 104 is specifically configured as follows: the specific method of calculating the abnormal score threshold τ includes:

[0114] The reconstruction error of the A1 subset is used to calculate the abnormal estimation threshold a(A1) of the A1 subset, and the abnormal estimation threshold is used as the abnormal score threshold τ. The specific formula includes:

[0115] a(·)=(e (i) -μ) T Σ -1 (e (i) -μ)

[0116] in,

[0117] μ and Σ are the parameters of the normal distribution N(μ,Σ);

[0118] e (i) is the reconstruction error of a certain subset;

[0119] e (i) =X (i) -X ’(i) ;

[0120] X (i) A time series of a subset of data;

[0121] X ’(i) is the reconstructed sequence of a subset of data;

[0122] T is transpose;

[0123] The specific method for calculating the abnormal score threshold τ also includes:

[0124] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0125] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0126] The AUC value is used as anomaly score threshold τ;

[0127] The specific method for calculating the abnormal score threshold τ also includes:

[0128] Calculating an abnormal estimation threshold a(A1) of the A1 subset by applying the reconstruction error of the A1 subset;

[0129] Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset;

[0130] The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve;

[0131] The larger value of the anomaly estimation threshold and the AUC value of the A1 subset is taken as the anomaly score threshold τ.

[0132] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of any one of the multi-cloud data anomaly detection methods based on deep learning in the first aspect of the present invention are implemented.

[0133] Figure 4 is a structural diagram of an electronic device according to an embodiment of the present invention, such as Figure 4 As shown, the electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covered on the display screen, or a button, a trackball or a touch pad set on the housing of the electronic device, or an external keyboard, touch pad or mouse, etc.

[0134] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the technical solution of the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0135] A fourth aspect of the present invention discloses a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps in the method for multi-cloud data anomaly detection based on deep learning in any one of the first aspects of the present invention are implemented.

[0136] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several variations and improvements can be made, which all belong to the scope of protection of the present application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A multi-cloud data anomaly detection method based on deep learning, characterized in that: The method comprises: Step S1, collecting time series of data from multiple cloud platforms; The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing; Step S2, inputting the time series of the data of the N2 subset into a deep learning network for training, and outputting a reconstructed sequence; Step S3, inputting the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset; Step S4, using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset to calculate the reconstruction error of the A1 subset, and then using the reconstruction error of the A1 subset to calculate the abnormal score threshold τ; Step S5: If the value in the time series of the real-time data exceeds the abnormal score threshold, the value is considered to be an abnormal value; In step S4, the specific method of calculating the abnormal score threshold τ includes: The reconstruction error of the A1 subset is used to calculate the abnormal estimation threshold a(A1) of the A1 subset, and the abnormal estimation threshold is used as the abnormal score threshold τ. The specific formula includes: a(·)=(e (i) -m) T S -1 (e (i) -m) in, μ and Σ are the parameters of the normal distribution N(μ,Σ); e (i) is the reconstruction error of a certain subset; e (i) =X (i) -X ’(i) ; X (i) is a time series of a subset of data; X ’(i) is the reconstructed sequence of a subset of data; T is transpose; In step S4, the specific method for calculating the abnormal score threshold τ also includes: Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset; The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve; The AUC value is used as anomaly score threshold τ; In step S4, the specific method for calculating the abnormal score threshold τ also includes: Calculating an abnormal estimation threshold a(A1) of the A1 subset by applying the reconstruction error of the A1 subset; Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset; The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve; The larger value of the anomaly estimation threshold and the AUC value of the A1 subset is taken as the anomaly score threshold τ.

2. The multi-cloud data anomaly detection method based on deep learning according to claim 1 is characterized in that: In step S2, the input of the deep learning network also includes: Statistical indicators of the data in the N2 subset.

3. The multi-cloud data anomaly detection method based on deep learning according to claim 2 is characterized in that: In step S2, the statistical indicators include: The maximum value of the data in the N2 subset, the minimum value of the data in the N2 subset, the mean of the data in the N2 subset, the median of the data in the N2 subset, and the variance of the data in the N2 subset.

4. The multi-cloud data anomaly detection method based on deep learning according to claim 1 is characterized in that: In step S2, the input of the deep learning network also includes: The time series of the data in the N2 subset is in reverse order.

5. A multi-cloud data anomaly detection system based on deep learning, characterized in that: The system comprises: A first processing module is configured to collect time series of data from multiple cloud platforms; The time series of normal data is divided into four subsets, of which the N1 subset is used for hyperparameter selection, the N2 subset is used for deep learning network training, the N3 subset is used for learning the constant score threshold, and the N4 subset is used for testing; the time series of abnormal data is divided into two subsets, the A1 subset is used for learning the constant score threshold, and the A2 subset is used for testing; A second processing module is configured to input the time series of the data of the N2 subset into a deep learning network for training, and output a reconstructed sequence; A third processing module is configured to input the time series of the data of the A1 subset into the trained deep learning network to obtain a reconstructed sequence of the data of the A1 subset; A fourth processing module is configured to calculate a reconstruction error of the A1 subset using the time series of the data of the A1 subset and the reconstruction sequence of the data of the A1 subset, and then calculate an abnormal score threshold τ using the reconstruction error of the A1 subset; The calculation of the abnormal score threshold τ specifically includes: The reconstruction error of the A1 subset is used to calculate the abnormal estimation threshold a(A1) of the A1 subset, and the abnormal estimation threshold is used as the abnormal score threshold τ. The specific formula includes: a(·)=(e (i) -m) T S -1 (e (i) -m) in, μ and Σ are the parameters of the normal distribution N(μ,Σ); e (i) is the reconstruction error of a certain subset; e (i) =X (i) -X ’(i) ; X (i) is a time series of a subset of data; X ’(i) is the reconstructed sequence of a subset of data; T is transpose; The calculation of the abnormal score threshold τ specifically includes: Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset; The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve; The AUC value is used as anomaly score threshold τ; The calculation of the abnormal score threshold τ specifically includes: Calculating an abnormal estimation threshold a(A1) of the A1 subset by applying the reconstruction error of the A1 subset; Apply the reconstruction error of the N3 subset to calculate the abnormal estimation threshold a(N3) of the N3 subset; The AUC value is calculated by applying the abnormal estimation threshold a(N3) of the N3 subset, where the AUC value is the area value under the ROC curve; The larger value of the abnormal estimation threshold and the AUC value of the A1 subset is taken as the abnormal score threshold τ; The fifth processing module is configured to consider a value as an abnormal value if the value in the time series of the real-time data exceeds the abnormal score threshold.

6. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in the multi-cloud data anomaly detection method based on deep learning described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps in the multi-cloud data anomaly detection method based on deep learning described in any one of claims 1 to 4 are implemented.