A Time-Series-Based Multi-Dimensional Index Anomaly Detection Method and Device

By improving the STMV vector and time series decomposition methods, and combining PCA summarization and VAE models, the accuracy problem of anomaly detection in microservice systems under resource-sufficient environments in existing technologies is solved, achieving higher detection accuracy and lower false alarm rate.

CN117201353BActive Publication Date: 2026-11-17INSTITUTE OF INFORMATION ENGINEERING CHINESE ACADEMY OF SCIENCES
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311179785.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-13
Publication Date
2026-11-17
Estimated Expiration
2043-09-13

AI Technical Summary

Technical Problem

Existing technologies, even in resource-rich environments, struggle to effectively combine monitoring metrics with time series for anomaly detection, leading to decreased accuracy. This is particularly true in microservice systems, where time series changes and resource utilization variations are not adequately considered, resulting in high false positive and false negative rates.

Method used

In a containerized microservice environment, an improved STMV vector is constructed, and unsupervised learning is performed by combining PCA summarization and VAE model to decompose time series into trend, random components and seasonal series for multi-dimensional anomaly detection. Prometheus is used to monitor disk I/O and database metrics, and a time series anomaly detection module is used for secondary detection.

Benefits of technology

It improves the accuracy of anomaly detection when physical resources are sufficient, reduces false alarm and false negative rates, and enables more accurate judgment of anomalies, especially when there are changes in time series and resource utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117201353B_ABST
    Figure CN117201353B_ABST
Patent Text Reader

Abstract

This invention discloses a time-series-based multi-dimensional anomaly detection method and apparatus. The method includes: acquiring all nodes of a target microservice backup, and constructing a time series for each node and a PCA digest for each microservice call chain by monitoring the physical resource data of these nodes; forming an improved STMV based on the PCA digest, and classifying the improved STMV to obtain a first anomaly detection result for each microservice call chain; extracting the corresponding time series according to the anomaly detection accuracy requirements; decomposing the corresponding time series into trend series, random component series, and seasonal series, and performing statistical tests on outliers to generate an outlier set; and detecting the abnormal microservice call chains in the first anomaly detection result based on the outlier set to obtain the final detection result for the microservice call chain. This invention can detect environmental anomalies other than response time and reduces false positives or false negatives.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of time series data mining and machine learning, and in particular to a method and apparatus for detecting anomalies in multi-dimensional indicators based on time series. Background Technology

[0002] With the rapid growth of Internet of Things (IoT) networks and many other connected data sources, time series data has seen tremendous growth. It is much larger in volume than traditional time series data and often has long seasonal periods. One of the fundamental problems in managing and utilizing this time series data is decomposing seasonality and trend. A good seasonality and trend decomposition can reveal fundamental insights into the time series and play a role in further analysis [Hochenbaum J, Vallis OS, Kejariwal A. Automatic Anomaly Detection in the Cloud ViaStatistical Learning:, 10.48550 / arXiv.1704.07706 [P]. 2017.]. For example, in anomaly detection, a local anomaly might be a spike during a period of inactivity. Without seasonality and trend decomposition, it would be missed because its value would still be much lower than the baseline anomaly value during busy periods. Furthermore, different types of anomalies correspond to different patterns in different components after decomposition.

[0003] Therefore, anomaly detection in time series data monitoring [Zeyan Li, Junjie Chen, Rui Jiao, Nengwen Zhao, Zhijun Wang, Shuwei Zhang, Yanjun Wu, Long Jiang, Leiqin Yan, Zikai Wang, Zhekang Chen, Wenchi Zhang, Xiaohui Nie, Kaixin Sui, Dan Pei: Practical Root Cause Localization for Microservice Systems via Trace Analysis. IWQoS2021:1-10] faces many technical challenges. For example, the periods are very complex, possibly aperiodic, single-periodic, or multi-periodic. Even within a single period, there are many variations in the magnitude of period offsets. The trend of these periods also includes gradual and abrupt changes, and significant noise is common in real-world data. Furthermore, the operations and maintenance platform collects predefined metrics and compares historical data with the latest collected data. If there is a significant deviation from historical data, anomalies may occur. However, if physical resources are sufficient, anomalies cannot be detected solely through these predefined metrics. Furthermore, this kind of intelligent operation and maintenance itself has high requirements for real-time processing, requiring the algorithm to have low computational complexity and be able to provide timely service responses. Also, since there is no baseline for the indicators, reducing the false positive rate of anomaly detection is also a challenge.

[0004] Currently, log-based online anomaly detection typically relies on unsupervised machine learning algorithms [Nandi A, Mandal A, Atreja S, et al. Anomaly Detection Using Program Control Flow Graph Mining From Execution Logs[C] / / the 22nd ACM SIGKDD International Conference. ACM, 2016.]. These algorithms process logs generated by application services to detect anomalies within those services. Specifically, unsupervised learning algorithms are used to learn a baseline model of the log behavior of application services during fault-free operation. This baseline model is then used to detect online whether newly recorded events differ from the baseline model, indicating that the service has experienced anomalies. However, this method has significant limitations due to the large size of the log system, the high resource consumption required, the lack of high accuracy, and the inability to guarantee the usability of historical logs as a baseline.

[0005] Anomaly detection based on traces encompasses unsupervised methods [Liu P, Xu H, Ouyang Q, et al. Unsupervised Detection of Microservice Trace Anomalies through Service-Level Deep Bayesian Networks[C] / / 2020 IEEE 31st International Symposium on Software Reliability Engineering (ISSRE). IEEE, 2020.], supervised machine learning [Nedelkoski S, Cardoso J, Kao O. Anomaly Detection and Classification using Distributed Tracing and Deep Learning[C] / / 2019 19th IEEE / ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID). ACM, 2019.], and trace data comparison [Wang T, Zhang W, Xu J, et al. Workflow-Aware Automatic Fault Diagnosis for Microservice-Based Applications With Statistics[J]. IEEE Transactions on Network and ServiceManagement, 2020, PP(99):1-1.]. In all these cases, the common assumption is that the trace data generated by the application during the training run is consistent with the trace data generated during runtime, thus providing a baseline for comparing newly generated trace data. This is almost identical to the assumption of log-based anomaly detection techniques discussed earlier, and therefore brings the same limitation: if the application's running conditions differ from those during the training run, and timing issues are not considered, the accuracy of anomaly detection may decrease.

[0006] The most common technique for anomaly detection based on monitoring information is to use unsupervised [Gulenko A, Schmidt F, Acker A, et al. Detecting Anomalous Behavior of Black-Box Services Modeled with Distance-Based Online Clustering[C] / / 2018:912-915.] or supervised [Du Q, Xie T, He Y. Anomaly Detection and Diagnosis for Container-Based Microservices with Performance Monitoring[C] / / International Conference on Algorithms and Architectures for Parallel Processing. Springer, Cham, 2018.] machine learning algorithms to process the monitored KPIs, train a baseline model, and detect whether any application service has experienced anomalies by comparing newly monitored KPIs with this baseline. Such methods can identify functional and performance anomalies at the application and service levels, but only if the application operates under the same conditions during training and runtime. When application runtime conditions change, such as higher load due to peak end-user requests or the deployment of new services to replace outdated ones, the accuracy of the anomaly detection may decrease.

[0007] In summary, there is currently no algorithm that combines monitoring metrics with time series data in a resource-rich environment. Summary of the Invention

[0008] To address the technical problems existing in the prior art, this invention provides a time-series-based multi-dimensional anomaly detection method and apparatus. In a containerized microservice environment, where microservices employ redundancy backups and load balancing strategies, this method incorporates non-parametric estimations of indicators such as disk I / O and database read / write operations into the STMV (Service Call Vector) model for retraining. Simultaneously, a time-series decomposition method is used to perform secondary detection on the anomaly traces obtained from the model, yielding the final anomaly trace (microservice call chain). This invention can reconstruct the STMV model in existing research when physical resources are sufficient, and performs multi-dimensional anomaly detection based on time series, optimizing previous methods and improving accuracy.

[0009] The technical solution of the present invention includes:

[0010] A time-series-based multi-dimensional indicator anomaly detection method includes:

[0011] Obtain all nodes of the target microservice backup, and construct the time series corresponding to each node and the PCA summary corresponding to each microservice call chain by monitoring the physical resource data of these nodes;

[0012] After forming an improved STMV based on the PCA digest, classification is performed based on the improved STMV to obtain the first anomaly detection result for each microservice call chain.

[0013] Extract the corresponding time series based on the anomaly detection accuracy requirements;

[0014] After decomposing the corresponding time series into trend series, random component series and seasonal series, statistical tests on outliers are performed to generate an outlier set;

[0015] Based on the set of outliers, the abnormal microservice call chains in the first anomaly detection result are detected to obtain the final detection result of the microservice call chains.

[0016] Furthermore, the physical resource data includes: CPU utilization, memory utilization, and hard disk utilization.

[0017] Furthermore, the classification based on the improved STMV to obtain the first anomaly detection result for each microservice call chain includes:

[0018] The VAE model is trained using unsupervised learning to obtain the trained VAE model.

[0019] The improved STMV is classified based on the trained VAE model to obtain the first anomaly detection result for each microservice call chain.

[0020] Furthermore, the step of extracting the corresponding time series according to the anomaly detection accuracy requirement includes:

[0021] If the anomaly detection accuracy requirement is greater than a certain accuracy threshold, extract the time series corresponding to all nodes.

[0022] If the anomaly detection accuracy requirement is not greater than the accuracy threshold, extract the time series corresponding to the nodes covered by the normal microservice call chain in the first anomaly detection result.

[0023] Further, the step of decomposing the corresponding time series into a trend series, a random component series, and a seasonal series includes:

[0024] Periodic extraction is performed on the corresponding time series to obtain all period lengths contained in the corresponding time series;

[0025] Periodic decomposition is performed on each period length to obtain trend sequence, random component sequence and seasonal sequence.

[0026] Further, periodicity extraction is performed on the corresponding time series to obtain all period lengths contained in the corresponding time series, including:

[0027] The corresponding time series is preprocessed, and the preprocessing process includes: filling missing values ​​in the corresponding time series using linear interpolation, and extracting and removing trend terms in the time series using HP filtering;

[0028] For the preprocessed time series, the periodicity at different levels is separated using the maximum repeating discrete wavelet transform;

[0029] Based on the decomposed periodic components and the wavelet variance corresponding to each periodic component, all periodic components are ranked by strength.

[0030] Based on the order of all periodic components from strongest to weakest, output periodicity at different levels;

[0031] For each level of periodicity, after generating candidate period lengths for each level based on the Fisher test, the candidate period lengths are verified and optimized using the ACF to include all period lengths contained in the periodicity of that level.

[0032] By combining all the periodic lengths contained in the periodicity of all levels, we can obtain all the period lengths contained in the corresponding time series.

[0033] Furthermore, the process of decomposing each period length to obtain the trend sequence, random component sequence, and seasonal sequence includes:

[0034] Based on the minimum absolute deviation loss with sparse regularization, the trend sequence contained in this period length is obtained;

[0035] Combining the aforementioned trend sequence, a nonlocal seasonality filter is applied to extract the seasonal sequence;

[0036] The non-trend and non-seasonal portions of the cycle length are set as random component sequences.

[0037] Furthermore, the statistical testing of outliers to generate an outlier set includes:

[0038] For the trend sequence, calculate the mean and standard deviation of the trend sequence, compare each data point in the trend sequence with the mean of the trend sequence to obtain a first difference, and determine whether the data point in the trend sequence is an outlier by comparing the first difference with the standard deviation of the trend sequence.

[0039] For a random component sequence, the random component sequence is used as a residual sequence. The mean and standard deviation of the residual sequence are calculated. Each data point in the random component sequence is compared with the mean to obtain a second difference. By comparing the second difference with the standard deviation of the residual sequence, it is determined whether the data point in the random component sequence is an outlier.

[0040] For a seasonal series, the data points in the seasonal series are placed in a bin, and the upper and lower quartiles and median are calculated based on the data points in the bin. Then, a box plot is drawn, and the interquartile range of the box plot is used to determine whether the data points in the seasonal series are outliers.

[0041] A time-series-based multi-dimensional indicator anomaly detection device includes:

[0042] The data acquisition module is used to obtain all nodes of the target microservice backup, and construct the time series corresponding to each node and the PCA summary corresponding to each microservice call chain by monitoring the physical resource data of these nodes.

[0043] The vector construction module is used to form an improved STMV based on the PCA digest, and then classify it based on the improved STMV to obtain the first anomaly detection result for each microservice call chain.

[0044] The time-series anomaly detection module is used to extract the corresponding time series according to the anomaly detection accuracy requirements; after decomposing the corresponding time series into trend series, random component series and seasonal series, perform statistical tests on outliers to generate an outlier set; and detect the abnormal microservice call chain in the first anomaly detection result based on the outlier set to obtain the final detection result of the microservice call chain.

[0045] A computer device, characterized in that the computer device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the time-series-based multi-dimensional indicator anomaly detection method described above.

[0046] Compared with existing technical solutions, the present invention has at least the following advantages:

[0047] 1. This invention can detect environmental anomalies other than response time when physical resources are sufficient.

[0048] 2. This invention takes into account timing, which can more accurately determine whether an abnormal false alarm or missed alarm is caused by a certain promotional event. Attached Figure Description

[0049] Figure 1 Memory changes with sufficient physical resources.

[0050] Figure 2 A graph showing the changes in disk I / O when physical resources are sufficient.

[0051] Figure 3 The time series of increments per second over the past 2 minutes.

[0052] Figure 4 CPU variation diagram considering timing.

[0053] Figure 5 Timing detection flowchart.

[0054] Figure 6 Time series construction.

[0055] Figure 7 Periodic decomposition diagram. Detailed Implementation

[0056] This invention combines temporal anomaly detection with system machine-level metrics. Based on the anomaly detection method described in [Shi Yuan. Research and Implementation of Root Cause Algorithm for Microservices [D]. Beijing: University of Chinese Academy of Sciences, 2023], it expands the dimensions of machine metrics and incorporates temporal anomaly detection for anomaly detection in microservice systems. The detection process mainly consists of three modules: data acquisition and vector construction, and temporal anomaly detection.

[0057] 1. Data Acquisition and Vector Construction

[0058] 1) Data Acquisition

[0059] Previous studies only collected and calculated metrics for physical machine CPU and memory. In a microservice multi-backup scenario, if one service fails, Kubernetes will quickly activate other backup microservices based on the calculated metrics. In this case, the response time of the microservices alone cannot detect the anomaly. Machine metrics can be used to determine if an anomaly has occurred; for example, if the physical machine's CPU and memory fluctuate abnormally or experience sustained high load after switching microservices, an anomaly can be identified. However, this approach does not consider that when physical resources are sufficient, changes in CPU and memory are minimal and cannot detect machine anomalies. Figure 1The graph showing memory usage with sufficient physical resources simulates a scenario where memory load is excessively high (450GB RAM, 50GB used). Under these conditions, normal memory usage is around 10GB, so 50GB represents a persistent anomaly that goes undetected. The same applies to the CPU. Therefore, this invention proposes incorporating the rate of change in disk I / O. Figure 2 The graph shows the changes in disk I / O under conditions of sufficient physical resources. The rate of change in disk I / O at this point clearly indicates an anomaly. Furthermore, most multi-service platforms use MongoDB as their database, but existing research has not considered the possibility that database anomalies could lead to eventual microservice anomalies. Therefore, this invention will utilize Prometheus to monitor MongoDB data and collect the rate(mongodb_mongod_wiredtiger_transactions_checkpoint_milliseconds_total[5m]) metric, representing the time series of increments per second over the past two minutes. This metric is used to monitor and analyze the changing trends of time series data. By calculating the increments per second, the average rate of change of this metric within a given time range can be understood, helping to assess system performance, resource utilization, or other metrics of interest. Figure 3 As shown. Current research also does not consider the impact of temporal changes on anomaly detection, such as... Figure 4 The CPU load variation is shown here. This situation can be simulated as the CPU load variation when an e-commerce platform has a major promotion or when train tickets are purchased during holidays. When the CPU load is consistently too high, this situation should be normal. In the past, this would have been identified as an anomaly. If this data is used as the training set in the dataset and then applied to the VAE model, the accuracy is not high enough, and the false positive rate and false negative rate are too high.

[0060] 2) Vector Construction

[0061] Existing STMV studies only consider response time and CPU and memory usage, neglecting metrics such as disk I / O and database per-second increments. Therefore, this invention re-binds these metrics into a new STMV forming vector. Furthermore, considering the current advanced internet era where each microservice in a microservice system has multiple backups distributed across different nodes, this invention expands the dimensions of STMV processing methods and constructs a PCA summary for each microservice, assuming each microservice has four backups, which may be distributed across the same node or different nodes. The PCA sequence is a sequence summary of CPU, Memory, Disk, and MongoDB usage on each backup node, as shown in Table 1 (Microservice PCA Summary).

[0062] Table 1 Summary of Microservice PCA

[0063]

[0064] Add the above PCA summary to STMV to form the improved STMV. There are two microservice backups in this table, as shown in Table 2. This table has covered various scenarios in case of anomalies.

[0065] Table 2: Real-world STMV Examples

[0066]

[0067]

[0068] 2. Timing Anomaly Detection

[0069] This module employs temporal anomaly detection, using the improved STMV vectors of this invention and unsupervised learning via a VAE model to obtain the first round of anomaly detection data. The need for secondary detection of all normal and anomalous data in the first round can be determined based on requirements. If higher accuracy and recall are required, secondary temporal detection should be performed on both correct and anomalous data; if only higher recall is needed, secondary temporal detection should be performed on all correct data. Figure 5 As shown.

[0070] 1) Data collection

[0071] This invention monitors the physical resource data of each node using Prometheus. This physical resource data includes CPU utilization, memory utilization, and disk utilization. Data is collected every minute and stored in a MySQL database. The system records the values ​​of each metric for the current time series. After obtaining the start time of the trace, it queries which different nodes the microservice is backed up on. It extracts the time from the current time to the previous 120 minutes as a time series and performs time-series testing on each of these different time points. If the last time point is an abnormal time point, then the microservice is considered abnormal. The query order is as follows: Figure 6 As shown.

[0072] 2) Periodic extraction

[0073] The first stage is time series preprocessing. Linear interpolation is used to impute missing values ​​to address large data gaps. HP filtering is then used to extract and remove trend components from the time series. Next, multi-period decomposition is performed. First, the Maximum Overlap Discrete Wavelet Transform (MODWT) is used to separate periodicity at different levels. Based on the decomposed periodic components and their corresponding wavelet variances, these components are ranked by strength, prioritizing the output of more significant periods. Finally, single-period detection is performed. After multi-period decomposition, Fisher's Test based on Huber-Periodogram is used to generate candidate period lengths for each level. These candidates are then validated and optimized using ACF (Active Flow Factor), ultimately yielding all period lengths for the sequence used in period decomposition.

[0074] 3) Periodic decomposition

[0075] This section decomposes the complex time series into trend, seasonality, and residual components, following the extracted periods described above. The trend is the long-term, persistent change pattern in the time series. By decomposing the time series into its trend component, long-term trend information can be extracted. Time series anomalies often manifest as sudden changes or unusual fluctuations that are clearly inconsistent with the trend. Seasonality is the periodic recurring pattern in the time series, such as annual, monthly, or weekly seasonal fluctuations. By decomposing the time series into its seasonal component, seasonal information of periodic changes can be extracted. If observations in the time series show anomalies or deviate from seasonal patterns in a particular season, this may be an indication of anomaly. The residual component refers to the part of the time series that cannot be explained by the trend and seasonality, i.e., the portion that cannot be explained by the trend and seasonality. This part typically includes random fluctuations, noise, or non-periodic changes that cannot be captured by the model.

[0076] By using a minimum absolute bias loss with sparse regularization to solve the regression problem, the trend component is robustly extracted. Based on the extracted trend, this invention applies a nonlocal seasonality filter to extract the seasonal component. This process is repeated until an accurate decomposition is obtained, which can then be used for further analysis. Figure 7 This is an example obtained after periodic decomposition.

[0077] 4) Statistical tests

[0078] This module performs statistical tests on the trend series, random component series, and seasonal series derived from periodic decomposition. For the trend series, the mean and standard deviation are used to detect outliers. The mean and standard deviation are calculated, and each data point is compared to the mean. If the absolute value is greater than three times the standard deviation, the data point is considered an outlier. For the random component series, the mean and standard deviation of the residuals are used to detect outliers. The decomposed random component series is used as the residual series, and its mean and standard deviation are calculated. Each data point is compared to the mean, and if the absolute value is greater than three times the standard deviation, the data point is considered an outlier. For the seasonal series, box plots are used to detect outliers. Data points from each season are placed in a box, and the upper and lower quartiles and median are calculated based on the data points in the box, and then a box plot is drawn. The IQR (interquartile range) is used to determine whether the data points in the box are outliers. If the value of a data point is less than Q1 - 1.5IQR or greater than Q3 + 1.5IQR, the data point is considered an outlier.

[0079] 5) Anomaly detection

[0080] All outliers from the above statistical tests are added to the outlier set, representing all outliers within the first 120 minutes of this Trace. If the last time point is an outlier, then this Trace is considered to have a time-series anomaly.

[0081] To demonstrate the feasibility of the method of this invention, the idea proposed in this invention was compared with previous methods. An open-source testing platform, TrainTicket, was built. This platform contains forty-one microservices. In the test of this invention, each microservice has three backups. For data collection, the open-source tracing system Jaeger was used to extract each trace information. The data was collected through ElasticSearch, and Fluentbit was used as a buffer mechanism to send the data to Kibana for display. At the same time, the Prometheus monitoring system was used to collect the indicator information of each machine, and the ChaosMesh chaos testing platform was used to simulate real failure scenarios for testing.

[0082] In the experiments of this invention, historical and real-time data were collected to test its effectiveness in two aspects. The first experiment was to detect anomalies under conditions of sufficient physical resources, and the second experiment was to detect anomalies considering temporal changes. The results of comparing multiple models are shown in Table 3.

[0083] Table 3 Comparison results between this method and other methods

[0084]

[0085]

[0086] As can be seen, there is a significant improvement in detecting load anomalies and timing anomalies in microservice systems with sufficient physical resources.

[0087] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. A time-series-based multi-dimensional indicator anomaly detection method, characterized in that, The method includes: Obtain all nodes of the target microservice backup, and construct the time series corresponding to each node and the PCA summary corresponding to each microservice call chain by monitoring the physical resource data of these nodes; After forming an improved STMV based on the PCA digest, classification is performed based on the improved STMV to obtain the first anomaly detection result for each microservice call chain. Extract the corresponding time series based on the anomaly detection accuracy requirements; After decomposing the corresponding time series into trend series, random component series and seasonal series, statistical tests on outliers are performed to generate an outlier set; Based on the set of outliers, the abnormal microservice call chains in the first anomaly detection result are detected to obtain the final detection result of the microservice call chains. The step of extracting the corresponding time series according to the anomaly detection accuracy requirement includes: If the anomaly detection accuracy requirement is greater than a certain accuracy threshold, extract the time series corresponding to all nodes. If the anomaly detection accuracy requirement is not greater than the accuracy threshold, extract the time series corresponding to the nodes covered by the normal microservice call chain in the first anomaly detection result; The step of performing statistical tests on outliers to generate an outlier set includes: For the trend sequence, calculate the mean and standard deviation of the trend sequence, compare each data point in the trend sequence with the mean of the trend sequence to obtain a first difference, and determine whether the data point in the trend sequence is an outlier by comparing the first difference with the standard deviation of the trend sequence. For a random component sequence, the random component sequence is used as a residual sequence. The mean and standard deviation of the residual sequence are calculated. Each data point in the random component sequence is compared with the mean to obtain a second difference. By comparing the second difference with the standard deviation of the residual sequence, it is determined whether the data point in the random component sequence is an outlier. For a seasonal series, the data points in the seasonal series are placed in a bin, and the upper and lower quartiles and median are calculated based on the data points in the bin. Then, a box plot is drawn, and the interquartile range of the box plot is used to determine whether the data points in the seasonal series are outliers.

2. The method as described in claim 1, characterized in that, The physical resource data includes: CPU utilization, memory utilization, and hard disk utilization.

3. The method as described in claim 1, characterized in that, The classification based on the improved STMV to obtain the first anomaly detection result for each microservice call chain includes: The VAE model is trained using unsupervised learning to obtain the trained VAE model. The improved STMV is classified based on the trained VAE model to obtain the first anomaly detection result for each microservice call chain.

4. The method as described in claim 1, characterized in that, The step of decomposing the corresponding time series into trend series, random component series, and seasonal series includes: Periodic extraction is performed on the corresponding time series to obtain all period lengths contained in the corresponding time series; Periodic decomposition is performed on each period length to obtain trend sequence, random component sequence and seasonal sequence.

5. The method as described in claim 4, characterized in that, Periodic extraction is performed on the corresponding time series to obtain all period lengths contained in the corresponding time series, including: The corresponding time series is preprocessed, and the preprocessing process includes: filling missing values ​​in the corresponding time series using linear interpolation, and extracting and removing trend terms in the time series using HP filtering; For the preprocessed time series, the periodicity at different levels is separated using the maximum repeating discrete wavelet transform; Based on the decomposed periodic components and the wavelet variance corresponding to each periodic component, all periodic components are ranked by strength. Based on the order of all periodic components from strongest to weakest, output periodicity at different levels; For each level of periodicity, after generating candidate period lengths for each level based on the Fisher test, the candidate period lengths are verified and optimized using the ACF to include all period lengths contained in the periodicity of that level. By combining all the periodic lengths contained in the periodicity of all levels, we can obtain all the period lengths contained in the corresponding time series.

6. The method as described in claim 4, characterized in that, The process of decomposing each period length to obtain the trend sequence, random component sequence, and seasonal sequence includes: Based on the minimum absolute deviation loss with sparse regularization, the trend sequence contained in this period length is obtained; Combining the aforementioned trend sequence, a nonlocal seasonality filter is applied to extract the seasonal sequence; The non-trend and non-seasonal portions of the cycle length are set as random component sequences.

7. A time-series-based multi-dimensional indicator anomaly detection device, characterized in that, The device includes: The data acquisition module is used to obtain all nodes of the target microservice backup, and construct the time series corresponding to each node and the PCA summary corresponding to each microservice call chain by monitoring the physical resource data of these nodes. The vector construction module is used to form an improved STMV based on the PCA digest, and then classify it based on the improved STMV to obtain the first anomaly detection result for each microservice call chain. The time-series anomaly detection module is used to extract the corresponding time series according to the anomaly detection accuracy requirements; after decomposing the corresponding time series into trend series, random component series and seasonal series, perform statistical tests on outliers to generate an outlier set; and detect the abnormal microservice call chain in the first anomaly detection result based on the outlier set to obtain the final detection result of the microservice call chain. The step of extracting the corresponding time series according to the anomaly detection accuracy requirement includes: If the anomaly detection accuracy requirement is greater than a certain accuracy threshold, extract the time series corresponding to all nodes. If the anomaly detection accuracy requirement is not greater than the accuracy threshold, extract the time series corresponding to the nodes covered by the normal microservice call chain in the first anomaly detection result; The step of performing statistical tests on outliers to generate an outlier set includes: For the trend sequence, calculate the mean and standard deviation of the trend sequence, compare each data point in the trend sequence with the mean of the trend sequence to obtain a first difference, and determine whether the data point in the trend sequence is an outlier by comparing the first difference with the standard deviation of the trend sequence. For a random component sequence, the random component sequence is used as a residual sequence. The mean and standard deviation of the residual sequence are calculated. Each data point in the random component sequence is compared with the mean to obtain a second difference. By comparing the second difference with the standard deviation of the residual sequence, it is determined whether the data point in the random component sequence is an outlier. For a seasonal series, the data points in the seasonal series are placed in a bin, and the upper and lower quartiles and median are calculated based on the data points in the bin. Then, a box plot is drawn, and the interquartile range of the box plot is used to determine whether the data points in the seasonal series are outliers.

8. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the time-series-based multi-dimensional indicator anomaly detection method as described in any one of claims 1-6.