Methods, apparatus, electronic devices and storage media for multi-source crossover anomaly detection and root cause inversion of time series indicators

By constructing a data lineage map and indicator network, combining multiple anomaly detection methods for weighted fusion of anomaly scores, and performing causal inversion along the indicator network, the problem of difficult root cause localization in single-indicator time-series anomaly detection is solved, achieving efficient and automated root cause localization.

CN123116089APending Publication Date: 2026-11-17SHENZHEN CHINESE CABBAGE TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611353432.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing single-indicator time-series anomaly detection technologies cannot effectively locate the root cause, resulting in isolated alarm information and low efficiency due to reliance on manual investigation.

Method used

By constructing a data lineage map and prior features of indicators, a network of related indicators is established. Multiple anomaly detection methods are executed in parallel to perform weighted fusion of anomaly scores and causal inversion along the network of related indicators to achieve automated root cause localization.

Benefits of technology

It has achieved a leap from isolated alarms to automated root cause localization, improving the efficiency and accuracy of anomaly root cause localization and reducing reliance on human experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116089A_ABST
    Figure CN123116089A_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, electronic device, and storage medium for multi-source cross-referenced anomaly detection and root cause inversion of time-series indicators. The method includes: constructing a network of related indicators containing upstream and downstream dependencies for the monitored indicators based on data lineage graphs and prior features of the indicators; performing parallel anomaly detection on the monitored indicators, including at least two types of anomaly detection: statistical anomaly detection, machine learning anomaly detection, and business rule-based anomaly detection, with each type of anomaly detection outputting a corresponding anomaly score; weighted fusion of the anomaly scores to obtain a comprehensive anomaly score; identifying nodes whose comprehensive anomaly scores exceed a preset threshold as anomaly nodes, and performing causal inversion along the related indicator network to obtain the anomaly cause of the anomaly node and the corresponding credibility score. Thus, causal inversion along the related indicator network achieves automated root cause localization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the technical field of data monitoring and operation and maintenance, and more specifically, to a method, apparatus, electronic device and storage medium for multi-source cross-anomaly detection and root cause inversion of time series indicators. Background Technology

[0002] As enterprises deepen their digital transformation, data-driven decision-making has become mainstream. In the daily data monitoring and maintenance process, enterprises typically rely on various monitoring tools to track key business indicators (such as GMV, order volume, active users, etc.) in real time.

[0003] Currently, the anomaly detection technologies commonly used in the industry mainly rely on the time-series anomaly detection of a single indicator. These technologies typically set fixed thresholds or rely on simple statistical rules to determine whether an indicator is abnormal. For example, when the system detects that "today's GMV is down 30% compared to yesterday," it will trigger an alarm to notify relevant personnel.

[0004] However, the aforementioned technologies can only provide isolated alarm information and lack correlation analysis with upstream and downstream related indicators, making it impossible to directly locate the root cause. Furthermore, since the investigation process relies heavily on human experience, it results in low work efficiency and is not reproducible. Summary of the Invention

[0005] In view of the above problems, this application proposes a method, apparatus, electronic device and storage medium for multi-source crossover anomaly detection and root cause inversion of time series indicators, which can solve the above problems.

[0006] This application provides a method, apparatus, electronic device, and storage medium for multi-source cross-analysis anomaly detection and root cause inversion of time-series indicators. The method includes: acquiring a data lineage graph and prior features of the indicators; constructing a related indicator network containing upstream and downstream dependencies for the monitored indicators based on the data lineage graph and prior features of the indicators; the nodes of the related indicator network represent indicators, and the edges represent the correlation between indicators, with each edge associated with correlation strength and time lag; performing anomaly detection on the monitored indicators in parallel, including at least two types of anomaly detection such as statistical anomaly detection, machine learning anomaly detection, and business rule-based anomaly detection, with each type of anomaly detection outputting a corresponding anomaly score; weighting and fusing the anomaly scores to obtain a comprehensive anomaly score; identifying nodes whose comprehensive anomaly scores exceed a preset threshold as anomaly nodes, and performing causal inversion along the related indicator network for the anomaly nodes to obtain the anomaly cause of the anomaly node and the corresponding credibility score of the anomaly cause.

[0007] Therefore, by performing causal inversion on abnormal nodes along the relevant indicator network, a leap from isolated alarms to automated root cause localization is achieved, effectively replacing inefficient and unreproducible manual experience-based investigation. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments and drawings obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0009] Figure 1 The diagram shows a flowchart of a time-series index multi-source crossover anomaly detection and root cause inversion method provided in an embodiment of this application.

[0010] Figure 2 The diagram shows a schematic of the structure of a time-series index multi-source crossover anomaly detection and root cause inversion device provided in an embodiment of this application.

[0011] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown.

[0012] Figure 4 This illustration shows a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0015] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0016] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0017] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0018] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0019] This invention provides a method for multi-source crossover anomaly detection and root cause inversion of time-series indicators. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, this method can be executed by software or hardware installed on a terminal device or a server device. The software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Please see Figure 1 , Figure 1 This illustration shows a flowchart of a time-series index multi-source crossover anomaly detection and root cause inversion method provided in an embodiment of this application. Figure 1 As shown, the method may include steps 110 to 150.

[0021] In step 110, the data lineage map and prior features of the indicators are obtained.

[0022] A data lineage graph, also known as an indicator lineage graph, is a topological structure diagram used to represent the data flow, processing dependencies, and logical relationships among various data indicators within a system. In this data lineage graph, nodes represent specific monitored indicators or data entities, and directed edges represent the direction of data flow or computational dependencies. For example, if indicator A is derived from indicators B and C through specific computational logic, then edges pointing from B to A and from C to A will be generated in the graph.

[0023] When a certain indicator becomes abnormal, the system can trace back along the edges of the data lineage graph to quickly locate the upstream dependent indicators or data processing tasks, thereby accurately converging the global anomaly investigation scope to the relevant local sub-network, which greatly improves the efficiency and accuracy of anomaly root cause localization.

[0024] Prior features of indicators are used to characterize the prior knowledge of the basic background and historical correlation patterns of the monitored indicators. In some implementations, prior features of indicators include at least the indicator metadata and the indicator co-occurrence frequency matrix. The indicator metadata is mainly used to construct the basic node attributes of the related indicator network and to provide a benchmark reference for subsequent business rule-based anomaly detection; the indicator co-occurrence frequency matrix reflects the statistical correlation between indicators in historical operation and is used to assign initial correlation weights to the edges between nodes when constructing the related indicator network, helping to identify potential implicit dependencies.

[0025] The metadata of the metrics can be the meta-information describing the static business attributes and logical definitions of the monitored metrics. For example, the system automatically obtains the metadata of each metric by connecting to the metric management platform. The metadata includes at least the following: the unique identifier of the metric, the metric name, the business definition, the calculation formula (e.g., the calculation logic of derived metrics), the data granularity (e.g., minute-level, hour-level, day-level), and the preset business alarm threshold.

[0026] The co-occurrence frequency matrix of indicators can be a probability matrix that characterizes the dynamic correlation strength between indicators, obtained through statistical analysis based on historical time-series data and historical alarm records. For example, within a historical time window, the system counts the number of times any two indicators simultaneously cause anomalies or trigger alarms, and divides this number by the total number of historical sampling periods to calculate the co-occurrence frequency of these two indicators. The matrix composed of the co-occurrence frequencies of all monitored indicators is the indicator co-occurrence frequency matrix.

[0027] In step 120, based on the data lineage graph and prior features of the indicators, a network of related indicators containing upstream and downstream dependencies is constructed for the monitored indicators; the nodes of the related indicator network represent indicators, the edges represent the correlation between indicators, and each edge is associated with the correlation strength and time lag relationship.

[0028] The construction of the relevant indicator network can integrate information from three sources: First, other indicators output by the same data processing task, that is, related indicators that belong to the same task output as the monitored indicator; Second, key node indicators in the upstream and downstream links, for example, when the monitored indicator is the transaction conversion rate, its upstream arrival volume indicator, downstream order volume indicator, etc. can all be included in the network; Third, indicators that frequently co-occur in historical alarm records, that is, indicator pairs that frequently trigger alarms at the same time within a historical time window.

[0029] In the constructed network of related indicators, nodes are used to represent indicators, and edges are used to represent the correlation between indicators. Each edge is associated with correlation strength and time lag. The correlation strength can be measured by Pearson or Spearman correlation coefficients, while the time lag is used to reflect the time offset characteristics experienced by one indicator after another responds when one indicator becomes abnormal.

[0030] In some implementations, to enable the relevant indicator network to adapt to business changes and system evolution, the time-series indicator multi-source crossover anomaly detection and root cause inversion method may also include an incremental update step for the relevant indicator network: (1) Using the preset update cycle as the trigger condition, retrieve the data processing tasks that have been added or changed within the update cycle from the data scheduling system, and extract the changed parts of their upstream and downstream dependencies. (2) For newly generated alarm logs within the update cycle, the newly added co-occurrence index pairs and their co-occurrence frequencies are statistically analyzed according to the preset time window, and the corresponding items of the index co-occurrence frequency matrix are updated. (3) Perform time decay processing on the original edge weights in the relevant index network. The decay function is exponential decay, so that the contribution of the historical co-occurrence records further away from the current time to the edge weights is smaller, so as to reduce the interference of historical noise. The decayed historical edge weights and the incremental weights calculated based on the new data are merged according to the preset fusion ratio to obtain the updated edge weights. (4) For index pairs whose data lineage has changed, if their task dependency has disappeared, the blood_lineage attribute of the corresponding edge will be set to invalid, and the edge will be deleted from the graph database when the weight of the edge is lower than the preset minimum weight threshold.

[0031] Therefore, by introducing a dynamic update mechanism for edge weights based on time decay, this application enables the relevant indicator network to continuously reflect the current real indicator association structure of the system, avoiding the formation of redundant edges in the graph due to the accumulation of historical dependencies caused by business iteration, thereby improving the accuracy of candidate root cause node selection in the causal inversion process.

[0032] In some implementations, the steps, based on data lineage maps and prior features of indicators, construct a network of related indicators containing upstream and downstream dependencies for the monitored indicators, and may also include the following steps: (1) Based on the data lineage graph, analyze the historical data processing tasks to obtain the upstream and downstream dependencies of the data processing tasks; (2) Obtain the metadata of the monitored indicators from the indicator management platform; (3) From the historical alarm logs, the co-occurrence frequency of co-occurrence index pairs is statistically analyzed according to the preset time window to obtain the index co-occurrence frequency matrix; (4) Based on upstream and downstream dependencies, indicator metadata and indicator co-occurrence frequency matrix, construct relevant indicator network.

[0033] In one specific implementation, the data lineage map is extracted by a lineage collector from task instances of a data scheduling system (e.g., Alibaba Cloud DataWorks, Apache Airflow). Each historical data processing task corresponds to a lineage map. (Task identifier) [] (Upstream Task List) (Output Table) [] (Output Field List) Four fields. By parsing the above fields, the complete upstream and downstream dependencies between data processing tasks can be restored.

[0034] In one specific implementation, the indicator metadata is obtained from the indicator management platform, and each indicator corresponds to... (Indicator label) [] (Input table list) [] (List of input fields) (Business Manager) The five fields (refresh frequency) are used to describe the source of the indicator's processing, the processing scope, and the business entity to which it belongs, among other attribute information.

[0035] In one specific implementation, historical alarm records can be exported from the alarm platform, according to... (Alarm time) ± Within a preset time window t, the index pairs that co-occur with alarms and their frequency of occurrence are counted, thus obtaining the co-occurrence frequency matrix. , and Each represents any two indicators.

[0036] Finally, based on upstream and downstream dependencies, indicator metadata, and indicator co-occurrence frequency matrix, a network of related indicators is constructed. This network comprehensively depicts the possible association paths between the monitored indicators and other indicators by integrating the task-level dependencies, indicator attribute information, and historical co-occurrence patterns reflected by the three types of data sources mentioned above, thus providing a data foundation for subsequent edge weight calculation and graph storage.

[0037] Furthermore, in some implementations, the edge weights between any two indicators in the relevant indicator network are calculated using the following formula: in, As an indicator With indicators The correlation coefficient on historical preset duration data This is a normalized function for co-occurrence frequencies. As a time-delay stability index, , and These are preset weighting coefficients.

[0038] In one specific implementation method It can be represented as That is, the co-occurrence frequency of index i and index j. Divide by the maximum value in the entire co-occurrence frequency matrix Normalization is performed to make the co-occurrence frequencies of different indicator pairs comparable.

[0039] Sij is a time-delay stability index, with a value range of [0,1], used to reflect the index With indicators The consistency of the time offset at the time of the anomaly across multiple historical anomaly events. The larger the value, the more stable the time lag relationship between the two indicators.

[0040] , and These are all preset weighting coefficients, used to adjust the contribution ratios of the correlation coefficient, the normalized co-occurrence frequency, and the time-delay stability index in the overall edge weights, respectively. In a specific implementation, the weight hyperparameters can take example values. =0.5、 =0.3、 =0.2.

[0041] In some implementations, the constructed network of relevant metrics can be stored using a graph database, such as Neo4j 4.x. Nodes in the network represent metrics, and node attributes may include... (Indicator label) (Indicator Name) (Business Manager) (Refresh rate); Edges in the network are used to represent the correlation between indicators, and edge attributes may include weight (edge ​​weight, i.e., the aforementioned ), lag (time delay), corr_type (correlation type).

[0042] Based on this graph database storage structure, the system can support performing K-hop neighbor lookup for any node according to the index ID. In a specific implementation, K can be 2, which is used to limit the range of candidate nodes in the subsequent root cause inversion process and reduce the computational complexity of causal testing.

[0043] In step 130, anomaly detection is performed in parallel on the monitored metrics. The anomaly detection includes at least two of the following: statistical anomaly detection, machine learning anomaly detection, and business rule-based anomaly detection. Each type of anomaly detection outputs a corresponding anomaly score.

[0044] In some implementations, statistical anomaly detection may employ algorithms such as the 3σ criterion, moving average, year-on-year detection, and month-on-month detection.

[0045] In some implementations, machine learning-based anomaly detection can employ the Isolation Forest model and the Prophet time series prediction model.

[0046] In some implementations, business rule-based anomaly detection can be based on calendar features such as weekdays, weekends, holidays, and major promotional days.

[0047] Statistical anomaly detection, machine learning-based anomaly detection, and business rule-based anomaly detection are performed in parallel and independently for the same monitored metric. Each type of anomaly detection outputs a corresponding anomaly score, thereby avoiding the limitations of a single detection method and achieving cross-validation of multi-source signals.

[0048] Specifically, in some implementations, the steps of performing anomaly detection on the monitored metrics in parallel may also include the following steps: (1) Use a preset set of statistical algorithms to detect the indicators to be monitored and determine the output score of each statistical algorithm in the preset set of statistical algorithms; the preset set of statistical algorithms includes at least one of the 3σ criterion algorithm, moving average algorithm, year-on-year detection algorithm and month-on-month detection algorithm; (2) Take the maximum value of the output score as the first abnormal score; (3) Use the preset machine learning model group to detect the indicators to be monitored and determine the output score of each model in the preset machine learning model group; the preset machine learning model group includes the isolated forest model and the prophetic time series prediction model. (4) Perform weighted summation on the output scores of each model to obtain the second anomaly score; (5) Calculate the cosine similarity between the pattern vector and the baseline vector of the indicator to be monitored based on calendar features, and convert the cosine similarity into a third anomaly score using a preset mapping relationship; (6) The steps of weighted fusion of abnormal scores to obtain a comprehensive abnormal score include: weighted fusion of the first abnormal score, the second abnormal score and the third abnormal score to obtain a comprehensive abnormal score.

[0049] For each monitored metric in the relevant metric network, the system calls different types of anomaly detection algorithms in parallel through a multi-source anomaly detector and outputs the corresponding anomaly score respectively.

[0050] For statistical anomaly detection, a pre-set set of statistical algorithms can be used to detect the monitored indicators, and the output score of each statistical algorithm in the pre-set set can be determined. The pre-set set of statistical algorithms may include at least one of the following: 3σ criterion algorithm, moving average algorithm, year-on-year detection algorithm, and month-on-month detection algorithm.

[0051] In one specific implementation, for each time point t, calculate separately , , and The maximum value among the above output scores is taken as the first anomaly score. ,Right now: The first anomaly score can be normalized to the interval [0, 1]. Using the maximum value as the statistical anomaly score ensures that any statistical detection rule that detects a significant deviation will be retained, avoiding the weakening of strong anomaly signals by averaging multiple statistical rules.

[0052] For machine learning-based anomaly detection, a pre-defined set of machine learning models can be used to detect the monitored indicators, and the output score of each model in the pre-defined set of machine learning models can be determined. The pre-defined set of machine learning models includes an isolated forest model and a prophetic time series prediction model. The isolated forest model is used to identify relatively isolated anomalous samples in the multi-dimensional feature space, while the prophetic time series prediction model is used to predict the normal range of the indicators based on historical trends, periodicity, and holidays.

[0053] In one specific implementation, machine learning-based anomaly scoring can be represented as: in, For the isolated forest model at time points The output is the probability or score of anomalies. This represents the anomaly probability or anomaly score output by the Prophet model at time t. This is the second anomaly score. It's understood that the weight of 0.5 used in the machine learning anomaly scoring is an exemplary configuration and can be adjusted based on the model's historical performance in actual implementation.

[0054] For business rule-based anomaly detection, the cosine similarity between the pattern vector and the baseline vector of the monitored metric can be calculated based on calendar features, and the cosine similarity can be converted into a third anomaly score using a preset mapping relationship. Calendar features can include weekdays, weekends, holidays, major promotional days, business activity days, etc.

[0055] In a specific implementation, business rule-based anomaly scoring can be represented as: in, This is the indicator pattern vector at the current time or within the current time window. This is the baseline pattern vector corresponding to the current calendar scene. This represents the cosine similarity between the indicator pattern vector and the baseline pattern vector.

[0056] The less similar the current pattern is to the baseline pattern under holidays or special business scenarios, the more likely it is to be affected. The larger the value, the higher the degree of anomaly in the business rule dimension.

[0057] Furthermore, the first anomaly score, second anomaly score, and third anomaly score are weighted and fused to obtain a comprehensive anomaly score. In one specific implementation, the comprehensive anomaly scorer employs a two-layer weighted fusion method combining static weights and dynamic adjustments. The static weights can be obtained through pre-training using historical AUCs, for example, by setting... =0.4、 =0.4、 =0.2, corresponding to the base weights of statistical anomaly scoring, machine learning anomaly scoring, and business rule-based anomaly scoring, respectively. The comprehensive anomaly score can be expressed as: in, To calculate the overall anomaly score, is the Sigmoid normalization function, used to map the weighted fusion result to the interval [0, 1].

[0058] Therefore, the system can simultaneously utilize three types of signals: statistical deviation, machine learning prediction deviation, and business calendar rule deviation, thereby improving the stability and robustness of anomaly detection. In a specific implementation, when Score(t) > At time t, the indicator is marked as an outlier. To preset the exception trigger threshold, for example, the default value is set to... =0.85. If multiple consecutive time points are marked as outliers, then the sequence of consecutive outliers is merged to form an outlier interval. The abnormal interval will be used as the input for subsequent causal inversion and root cause localization.

[0059] Furthermore, in some implementations, the step of weightedly fusing the first anomaly score, the second anomaly score, and the third anomaly score to obtain a comprehensive anomaly score may also include the following steps: (1) Determine whether the current time period belongs to a preset special time period; (2) If so, obtain the preset weight group corresponding to the target special time period corresponding to the current time period; (3) Based on the first abnormality score, the second abnormality score, the third abnormality score and the preset weighted group, the initial abnormality score is determined; (4) Normalize the initial anomaly score to obtain the comprehensive anomaly score.

[0060] Special periods can include periods that differ significantly from the regular business operation mode, such as major promotional periods, holidays, and the early morning period when data is sparse.

[0061] If the current time period belongs to a preset special time period, then the corresponding preset weight reorganization is selected based on the target special time period corresponding to the current time period. As an optional implementation method, if the current time period is during a major promotional period, such as Double 11, the business rule-based weights can be used because business rules and calendar features have a stronger ability to explain indicator fluctuations. Increased to 0.4, while also adjusting the statistical weight. The value was lowered to 0.2 to reduce the probability of ordinary statistical fluctuations being misjudged as anomalies. If the current time period is during a data-sparse period such as 02:00-06:00 AM, due to the sparse sampling of indicators and large statistical fluctuations, the statistical weights can be reduced. Reduced to 0.2, and machine learning weights The value was increased to 0.5 to enhance the model's ability to utilize historical patterns and trend predictions.

[0062] After determining the corresponding preset weighted reassembly, the initial anomaly score is calculated based on the first anomaly score, the second anomaly score, the third anomaly score, and the preset weighted reassembly. Specifically, the first anomaly score can be... Second abnormal score and the third abnormal score Multiply by the corresponding current time period respectively , and The initial anomaly scores are then summed to obtain the total anomaly score.

[0063] Therefore, by combining static and dynamic weights, this application can adaptively adjust the contribution ratio of various detection algorithms according to different business periods, so that the comprehensive anomaly score can reflect the anomaly detection performance under normal circumstances, and adapt to the index fluctuation patterns under special scenarios such as promotional events, holidays, and sparse data in the early morning.

[0064] In step 140, the anomaly scores are weighted and fused to obtain a comprehensive anomaly score.

[0065] After obtaining various anomaly scores, the anomaly scores are weighted and fused to obtain a comprehensive anomaly score. The weighted fusion process can pre-set weights based on the historical detection results of various algorithms, and dynamically adjust the weights according to the business scenario of the current time period. Finally, the fusion result is mapped to a unified interval through a normalization function to obtain a comprehensive anomaly score, which is used to uniformly measure the degree of anomaly of the monitored indicators at the current time.

[0066] In step 150, nodes whose comprehensive anomaly scores exceed a preset threshold are identified as anomaly nodes, and causal inversion is performed on the anomaly nodes along the relevant index network to obtain the anomaly causes of the anomalies and the corresponding credibility scores.

[0067] Nodes whose comprehensive anomaly scores exceed a preset threshold are identified as anomaly nodes. That is, when the comprehensive anomaly score of a certain indicator exceeds a preset judgment threshold, the network node corresponding to that indicator is marked as an anomaly node. Causal inversion is performed on the anomaly nodes along the relevant indicator network. That is, starting from the anomaly node, the upstream indicators that may have caused the anomaly are backtracked along the edges of the network. The strength of the causal relationship between each candidate indicator and the anomaly node is quantified by combining causal testing methods. Finally, the anomaly cause of the anomaly node and the credibility score corresponding to the anomaly cause are obtained.

[0068] Specifically, in some implementations, the step of performing causal inversion on anomalous nodes along the relevant index network may also include the following steps: (1) Select the nodes in the preset number of hops neighbors of the abnormal node in the relevant indicator network as the initial screening nodes; (2) Calculate the correlation coefficient between the initial screening nodes and the abnormal nodes; (3) The nodes whose correlation coefficients are greater than or equal to the preset test threshold are identified as candidate root cause nodes. (4) Based on the candidate root cause nodes, Granger causality test is performed on the subsequence data within the preset offset time period centered on the preset abnormal interval.

[0069] Since the number of nodes in the relevant indicator network may reach hundreds or even more, if causal testing is performed on all indicator pairs, the computational complexity can reach O(n log n). This approach, however, is insufficient to meet the requirements of real-time root cause analysis. Therefore, in one specific implementation, the system employs a graph-based pruning strategy, performing subsequent analysis only on indicators within the K-hop neighbor range of anomaly nodes, where K can be 2. This graph-based pruning reduces the number of candidate indicator pairs from... This reduces the computational load to the kN level, thus significantly reducing the computational cost.

[0070] After obtaining the initial screening nodes, calculate the correlation coefficient between the initial screening nodes and the outlier nodes. The correlation coefficient can be the Spearman correlation coefficient, and its absolute value can be taken. The correlation between the two is measured near the abnormal range or within the historical window.

[0071] Furthermore, nodes whose correlation coefficients are greater than or equal to a preset test threshold are identified as candidate root cause nodes. In one specific implementation, the preset test threshold can be set to 0.3, meaning only nodes with a correlation coefficient greater than or equal to a preset threshold are considered as candidate root cause nodes. Strongly correlated indicators with a value ≥0.3 are used for subsequent causal tests to filter out weakly correlated or irrelevant indicators, forming a second pre-screening for correlation.

[0072] An abnormal interval is formed during the abnormal triggering phase. After that, the system only extracts subsequence data centered on the abnormal interval and extended forward and backward by a preset time length, for example, extracting [ Data within the time window, and The default setting can be 24 hours. This sliding window pruning method eliminates the need to scan all historical data; instead, it performs causal checks only on the most relevant timeframes before and after the anomaly, improving real-time performance.

[0073] In the above process, the system effectively controls computational complexity by employing three strategies sequentially: graph pruning, correlation pre-screening, and sliding window pruning. Graph pruning limits the range of candidate nodes, correlation pre-screening limits the range of candidate index pairs, and sliding window pruning limits the time series range. After these three pruning steps, Granger causality tests are performed on the subsequence data between candidate root cause nodes and anomalous nodes to determine whether changes in candidate root cause nodes have predictive explanatory power for anomalous nodes over time.

[0074] Furthermore, before performing the Granger causality test, this application also includes identifying the data granularity type of the subsequence data and adaptively setting the time delay parameter according to the data granularity type. In some embodiments, this time-series index multi-source crossover anomaly detection and root cause inversion method may further include the following steps: (1) Identify the data granularity type of the subsequence data; (2) If the data granularity type is hourly, then set the time delay parameter of the Granger causality test to the first preset value; (3) If the data granularity type is daily, then set the time delay parameter of the Granger causality test to the second preset value; (4) If the data granularity type is weekly, then set the time delay parameter of the Granger causality test to the third preset value; (5) Construct a vector autoregression model based on the set time delay parameters, and perform Granger causality test on the subsequence data within the preset offset time period centered on the preset abnormal interval based on the candidate root cause node through the steps of the vector autoregression model.

[0075] The system can be based on the indicators Fields, timestamp intervals, or data sampling periods can be used to identify whether subsequence data belongs to hourly, daily, or weekly data levels.

[0076] If the data granularity is hourly, the time lag parameter for the Granger causality test is set to a first preset value. In one specific implementation, the default value for the lag of hourly data can be set to 24 to cover hourly periodic variations within a day.

[0077] If the data granularity type is daily, the time lag parameter of the Granger causality test is set to the second preset value. In a specific implementation, the default value of lag for daily data can be set to 7 to cover the cyclical changes within a week.

[0078] If the data granularity type is weekly, the time lag parameter of the Granger causality test is set to the third preset value. In a specific implementation, the default value of lag for weekly data can be set to 4 to cover approximately one month of weekly changes.

[0079] For each pair of candidate indices, the system can be configured in lag∈[1, The time delay order is iterated through within the range, and the lag with the smallest AIC value is selected as the time delay order used in the actual test according to the AIC criterion. Then, using the selected lag A vector autoregression model is constructed, and it is determined whether the historical sequence of candidate root cause nodes can significantly improve the predictive ability of the current sequence of abnormal nodes. This adapts to the propagation delay characteristics of different time granularity indicators and avoids the test bias caused by fixed lag.

[0080] In some implementations, the actual time-series propagation characteristics of different candidate index pairs vary significantly. If a fixed maximum time delay upper limit is uniformly applied to all index pairs based on granularity type, this can lead to problems. This could lead to the introduction of unnecessary higher-order parameters for metrics with shorter propagation delays, or the omission of the optimal time delay order for metrics with longer propagation delays due to insufficient search upper bound. To address this, this application further provides an adaptive time delay order selection mechanism driven jointly by mutual information and multi-information criteria, the specific steps of which are as follows: (1) Before performing Granger causality tests, validity checks were set up for standard VAR-Granger and Toda-Yamamoto respectively.

[0081] If it is a standard VAR-Granger path, it is stabilized and the optimal lag order p is determined to construct VAR(p). If it is a Toda-Yamamoto path, the conventional optimal lag order p is determined, VAR(p+dmax) is constructed, and model diagnostics / validity checks suitable for the TY path are performed.

[0082] In one specific implementation, the Granger causality test is used to examine whether historical information of one variable helps predict another variable, and is usually based on a VAR model. When the series is stationary (or stationary after transformation), the standard Granger test can be used, provided that the VAR(p) model is stable and the residuals meet diagnostic requirements such as white noise.

[0083] When a sequence has a unit root (non-stationary) or there may be cointegration between variables, the asymptotic distribution of the standard test is no longer effective. In this case, the Toda-Yamamoto method is recommended. By artificially extending the model to VAR(p + dmax), the excess lag order dmax is used to absorb the influence of the unit root. Thus, it is not necessary to test stationarity and cointegration in advance. A robust causal conclusion can be obtained by directly performing the MWALD test on the first p-order coefficients.

[0084] If all variables are stationary I(0), follow the standard Granger path; if there is an integrated or cointegrated relationship of I(1) or above, directly follow the Toda-Yamamoto path, which can eliminate the pre-test step and make the conclusion more robust.

[0085] Calculate the mutual information value MI(τ) between candidate root cause nodes and anomalous nodes under different time delays τ, where τ ranges from [1, ..., ... ], This represents the system's global maximum search limit. Find the point where MI(τ) first drops to its peak value. The time delay value corresponding to less than 50% ,Will Default corresponding to granularity type Take the smaller value as the actual search upper limit for this candidate index pair. ,Right now: =min( , Therefore, for index pairs with short propagation delays, the search range is narrowed to reduce the risk of overfitting, while for index pairs with long delays, the upper limit is appropriately extended within the system's allowable range to improve the test sensitivity.

[0086] (2) In [1, Within the range of lag orders p, calculate the optimal order given by the Akaike Information Criterion (AIC(p), Bayesian Information Criterion (BIC(p)) and Hannan-Quinn Criterion (HQC(p)) respectively. , , .

[0087] In one specific implementation, if the three criteria yield the same result, that result is directly used as the final time delay order lag. If two of the three criteria are the same, the majority vote rule is used to select the order indicated by the majority criterion; if the three criteria are different, the BIC (Best Income, First Class) criterion is preferred. As the final order, BIC imposes a stronger penalty on model complexity, which helps prevent overfitting when the sample size is limited.

[0088] (3) To evaluate the robustness of the selected lag to the subsequence window division, the system divides the original subsequence window into multiple overlapping sub-windows with a length of 70% of the original window in a sliding manner. Steps (1) and (2) are executed independently for each sub-window, and the optimal order given by each sub-window is recorded.

[0089] If more than a preset proportion (e.g., 70%) of the sub-windows select the same lag, the order selection is considered stable and reliable; otherwise, the lag selection of the candidate indicator pair is marked as "unstable", and a confidence discount factor δ (default δ=0.7) is introduced in the subsequent causal strength calculation. That is, the causal strength score of the final output of the indicator pair is multiplied by δ to reflect the impact of the uncertainty of the time delay order selection on the reliability of the test conclusion.

[0090] Therefore, by independently estimating a reasonable time delay search range for each candidate indicator and using multi-information criteria for joint decision-making and stability verification, this application enables the lag order of the VAR model to fully adapt to the actual time-series propagation characteristics of different indicators. This avoids both excessive parameters and estimation bias caused by too high a fixed order and omission of true causal relationships caused by too low a fixed order, thereby improving the statistical power and reliability of Granger causality tests.

[0091] In some implementations, before performing Granger causality tests on the subsequence data between candidate root cause nodes and anomalous nodes, this time-series index multi-source crossover anomaly detection and root cause inversion method may further include the following steps: (1) Perform Augmented Dickey-Fuller (ADF) unit root test on the subsequence data of candidate root cause nodes and the subsequence data of outlier nodes respectively to determine whether each subsequence is stationary; (2) If the ADF test statistic of a subsequence is greater than the critical value under the preset significance level, the subsequence is determined to be a non-stationary sequence. Then, the subsequence is subjected to first-order differencing until the differencing sequence passes the ADF stationarity test. (3) Use the stabilized subsequence data as input for Granger causality test.

[0092] Therefore, by introducing an adaptive stationarization preprocessing step before Granger causality testing, the spurious regression problem caused by non-stationary time series data violating the presuppositions of Granger testing can be effectively avoided, thus improving the statistical reliability of causality test conclusions. Especially when the monitored indicators exhibit business growth trends or seasonal drift, directly participating in the test with undifferentiated non-stationary series often produces falsely significant causal relationships, leading to misjudgments of root causes. This application, however, eliminates trend components through differencing, ensuring that the judgment of causality is based on genuine short-term fluctuation patterns, thereby improving the accuracy of root cause localization.

[0093] In some implementations, when the subsequence data of both the candidate root cause node and the anomalous node are first-order integrated sequences (i.e., each is I(1), and can only pass the ADF stationarity test after first-order differencing), directly performing first-order differencing on the two sequences and then participating in the Granger test will lose the long-term equilibrium constraint relationship that may exist between the two sequences, resulting in the inability to detect root cause associations driven by long-term cointegration. Therefore, this application further provides a multi-branch stationarization scheme that includes a cointegration relationship processing path for non-stationary sequences: After determining that a subsequence is non-stationary using the ADF test, the minimum number of differencing operations, d, required to pass the stationarity test is recorded. d=0 indicates that the original sequence is stationary (I(0)), d=1 indicates that it is stationary after first differencing (I(1)), and so on. The system records the integration order of the subsequences of the candidate root cause nodes. The integral order of the subsequence of abnormal nodes .

[0094] In determining and Then, proceed according to the following branches: Branch 1: If =0 and =0, meaning both subsequences are stationary, and can directly proceed to the standard Granger causality test procedure without additional processing.

[0095] Branch 2: If =1 and =1, meaning both subsequences are first-order integrated sequences. Perform the Johansen cointegration test, using the trace statistic and the maximum eigenvalue statistic at the 5% significance level to test the number of cointegrating vectors.

[0096] If the Johansen test results show the existence of at least one significant cointegrating vector, then the Toda-Yamamoto causality test method is used: estimate VAR(p+) on the level value sequence (without differencing). )Model, =1, meaning an additional lag of one order is added to the standard lag order p to absorb the effects of non-stationarity. Subsequently, the Modified Wald (MWALD) test is performed on the matrix of coefficients with the first p lags, and the asymptotic chi-square statistic is used to determine whether the Granger non-causal null hypothesis holds. This method avoids the loss of long-term equilibrium information caused by direct differencing when the series have cointegration relationships, and it does not require explicit construction of a vector error correction model (VECM), making it simple to implement and statistically sound.

[0097] If the Johansen test results show that there is no significant cointegration relationship, then perform first-order differencing on each of the two sequences, and substitute the differencing sequences into the standard Granger causality test procedure.

[0098] Branch 3: If (The two subsequences have different integral orders, for example, one is I(0) and the other is I(1)). Perform the necessary difference operation on the sequence with the higher integral order so that the two sequences reach the same integral order. Then, refer to branch one or branch two to perform the subsequent steps.

[0099] Branch 4: If a subsequence fails the ADF test after multiple differencing (d≥2), or if the variance of the differencing sequence tends to zero (over-difference), the sequence is marked as "highly non-stationary or structurally mutated." The VAR-based Granger test process is then abandoned, and a causal measure based on permutation entropy difference or non-parametric transfer entropy is used to evaluate the candidate indicator pairs. The alternative test method used and its applicable limitations are noted in the output results.

[0100] After completing the causality test, the system records the specific test path used (standard Granger test / Toda-Yamamoto test / differenced Granger test / nonparametric substitution method) in the output field. This information will serve as a basis for subsequent review and verification.

[0101] Therefore, by combining the ADF unit root test with the Johansen cointegration test, a multi-branch nonstationary processing mechanism containing cointegration paths is established. This application can retain long-term causal information using the Toda-Yamamoto method when two sequences have a long-term cointegration equilibrium relationship, and eliminate trend interference by using standard deviation processing when there is no cointegration. This comprehensively covers various time series characteristic scenarios such as stationary, nonstationary without cointegration, and nonstationary with cointegration, effectively improving the robustness and accuracy of root cause localization in the context of complex indicator evolution.

[0102] In some implementations, before inputting the stationary subsequence data into the Granger causality test, this time-series index multi-source crossover anomaly detection and root cause inversion method further includes the following validity check steps to ensure the statistical reliability of the Granger test conclusions based on the VAR model: (1) The Granger test is achieved by constructing a VAR model of order p. In the case of bivariate, the number of parameters to be estimated in VAR(p) is 2p+2 (p lag coefficients and one intercept term for each equation).

[0103] Before constructing the VAR model, the system verifies whether the ratio of the effective sample size T to the number of model parameters satisfies the sufficiency condition: T > μ*(2p+2), where μ is the redundancy coefficient, defaulting to 5, to ensure the statistical confidence of the parameter estimates. If the effective sample size does not meet the above condition, a reduction in order is performed: the current time lag order is gradually reduced until the maximum order lag' that satisfies the sufficiency condition is found as the alternative order; if the condition is still not met even when lag'=1, the Granger test for the candidate indicator pair is abandoned, and instead, the absolute value of the Spearman correlation coefficient of the candidate indicator pair near the outlier interval is used. | As an alternative estimate of causal strength, the output is marked "Insufficient sample size, causal strength is a substitute for correlation coefficient".

[0104] (2) After constructing the VAR model with the selected time delay order, the system calculates all the characteristic roots of its characteristic polynomial.

[0105] If the magnitudes of all eigenvalues ​​are strictly less than 1 (i.e., all fall within the unit circle), the VAR model is considered stable, and subsequent tests can continue. If there are eigenvalues ​​with a magnitude greater than or equal to 1, it indicates that the VAR model is unstable or in a unit root boundary state. In this case, a step-back strategy is implemented: gradually reduce the lag order, re-estimate the VAR and test the stability of the eigenvalues ​​after each reduction of 1, until the VAR is stable or the lag order is reduced to 1. If the VAR is still unstable when the lag order is 1, the Granger test for that indicator pair is abandoned, its causal strength score is set to 0, and a warning label "VAR model is unstable, causal test not applicable" is added to the output candidate root cause results, indicating that the correlation of the candidate root cause may come from a non-stationary trend rather than a true causal relationship, and manual review is recommended.

[0106] (3) After the VAR model parameters are estimated, the Portmanteau test (i.e., the Ljung-Box test) is performed on the residual series of each equation. The test lag is H = min(2lag). , 10) Whether there is significant autocorrelation within the order, H is the test order.

[0107] If the p-value of the Ljung-Box statistic for the residuals is less than 0.05 at the 5% significance level, it indicates that the VAR model does not adequately fit the dynamic structure of the data, and systematic patterns still exist in the residuals. In this case, the system attempts to increase the lag order to min(lag*+3, The VAR is re-estimated and the residual autocorrelation test is performed again. If the residual autocorrelation problem is still not eliminated after the order is increased, a "residual autocorrelation warning" mark is added when outputting the causal test results of the candidate index pair, indicating that the current Granger test conclusion has a certain degree of unreliability. The causal strength score of the index pair is multiplied by a penalty factor of 0.8 to reflect the discount of the test reliability.

[0108] Because the subsequence data covers outlier intervals, which may contain extreme deviations, directly using them in VAR parameter estimation would lead to the coefficient matrix being dominated by extreme values, reducing the robustness of causality tests. Therefore, before constructing the VAR: (4) Perform Winsorization on the data points of each index in the subsequence: truncate values ​​below the 1st percentile P1 to P1, and truncate values ​​above the 99th percentile P99 to P99, to preserve the overall dynamic characteristics of the sequence and eliminate the excessive influence of extreme single points on the estimation results. Retain the original subsequence and compare the causal direction / strength before and after processing. The Winsorized sequence serves as a robust estimation sequence. After aggregation, time-lag identifiability is checked, and confidence is judged on the checked relationship to avoid truncation of outliers or aggregation to a coarser granularity that weakens the key information originally used to identify the causal direction.

[0109] Furthermore, if the proportion of abnormal interval data points to the total length of the subsequence exceeds 30%, the system will use the normal data subsequence outside the abnormal interval as a separate training set for the VAR model to estimate the steady-state parameter relationship. The system will use the transition segment data before and after the abnormal interval to evaluate whether the disturbance of the candidate root cause node has a predictable effect on the abnormal node in the Granger sense, thereby reducing the interference of the abnormal interval data itself on the causal direction judgment.

[0110] Therefore, by systematically performing four validity checks before the Granger test—sample adequacy verification, VAR stability test, residual autocorrelation diagnosis, and extreme value robustness preprocessing—this application can effectively avoid false causal conclusions caused by insufficient data, model instability, or residual structure, thereby ensuring the statistical reliability of root cause localization under abnormal data interference conditions in actual business scenarios.

[0111] In some implementations, before performing the aggregation operation at the alignment granularity, the time-series index multi-source crossover anomaly detection and root cause inversion method further includes the following time window integrity judgment step to avoid introducing bias by directly aggregating incomplete data windows: The abnormal range [ , Mapped to the aggregation cycle sequence at the alignment granularity. Calculate the number of aggregation cycles that are fully covered by the anomalous interval. and the number of boundary periods that are only partially covered .

[0112] The value can be 0, 1, or 2, corresponding to the following scenarios: the entire abnormal interval falls within a single aggregation cycle, there is only one partial overlap with the boundary, and there is partial overlap at both the beginning and end, respectively. If... If the value is greater than 0, the corresponding boundary aggregation period will be marked as a "boundary period", indicating that the aggregated value in this period contains data from both abnormal and normal periods, which may dilute the abnormal signal.

[0113] For each aggregation period t of the subsequence, calculate the expected number of data points within that period based on the refreshfrequency field in the indicator metadata. And count the actual number of original data points. It excludes missing values, NaN, and duplicate data points.

[0114] The integrity score for each aggregation cycle is defined as follows: Will Aggregation periods with a value less than 0.7 are marked as "incomplete data periods". Periods with a value <0.3 are marked as "severely missing periods".

[0115] Differential processing strategy for incomplete data: For incomplete periods and boundary periods, the following differential processing is applied according to their severity: When the data in the subsequence is incomplete (including boundary periods), When the proportion of <0.7) is less than 20%, linear interpolation is performed to complete the missing time points within the above period, and confidence weights are assigned to the completed aggregate value. , which serves as the sample weights for subsequent VAR parameter estimation.

[0116] When the proportion of incomplete data periods is between 20% and 50%, interpolation completion is not performed; instead, it is done during VAR model estimation. As a weighting factor for the corresponding data points, the impact of incomplete cycles on parameter estimation is reduced by weighted least squares estimation. When the proportion of incomplete data periods exceeds 50%, or the proportion of abnormal data points within the boundary period is less than 30% (i.e., the abnormal signal is diluted by a large amount of normal data), the aggregation test scheme under the current alignment granularity is abandoned, and the test is regressed to directly extracting the data before and after the abnormal interval from the original subsequence with a finer granularity to participate in the Granger test. If the granularity of the two sequences is still different, the coarser granularity sequence is downsampled and interpolated to align with the finer granularity of the two sequences, and the reason is noted in the output causal test result: "Cross-granularity aggregation failed, and the test has been regressed to fine-granularity local test".

[0117] Aggregate quality information pass-through and confidence discount: This involves passing through the aggregated sequence after integrity checks and processing, along with the data from each period. The vector is passed to the Granger criterion module. After the criterion is completed, the average integrity score of the entire subsequence is calculated. .

[0118] like If the score is below the preset integrity threshold (default is 0.85), the causal strength score of the current candidate indicator is multiplied by the data quality penalty factor. , This is to quantify the impact of missing data on the confidence of the test results; at the same time, a warning label "Insufficient data integrity, causal strength has been discounted" is added to the output results to remind maintenance personnel to pay attention to the data collection quality of the corresponding time period when referring to the candidate root cause.

[0119] Therefore, by introducing anomaly interval boundary alignment checks, cycle-by-cycle integrity score calculations, and differentiated processing strategies before cross-granularity aggregation, this application can effectively ensure the quality and representativeness of the input data used in Granger causality tests in complex scenarios where the data granularity of candidate root cause nodes and anomaly nodes does not match and the anomaly interval spans multiple aggregation cycles or there is data missing, thus preventing misjudgment of causal direction due to incomplete data.

[0120] In some implementations, when the data granularity types of candidate root cause nodes and anomalous nodes are different, the time-series index multi-source cross-linking anomaly detection and root cause inversion method may further include the following steps: (1) Obtain the data granularity type of the candidate root cause node subsequence data and the abnormal node subsequence data respectively; (2) If the data granularity types of the two are different, then a unified alignment granularity shall be determined; the alignment granularity shall be determined according to the following rules: when the granularity of the two is hourly and daily respectively, the daily granularity shall be used as the alignment granularity; when the granularity of the two is hourly and weekly or daily and weekly respectively, the weekly granularity shall be used as the alignment granularity. (3) For fine-grained subsequence data, perform aggregation operations according to the aligned granularity. The aggregation function is adaptively selected according to the indicator type in the indicator metadata: for cumulative indicators (such as order volume, GMV), summation aggregation is used; for mean indicators (such as conversion rate, response time), mean aggregation is used; for status indicators (such as system health), mode aggregation is used. (4) The Granger causality test is performed on the subsequence data after alignment by using the time delay preset value corresponding to the alignment granularity as the upper limit of the time delay parameter of the Granger causality test.

[0121] Therefore, by introducing a cross-granularity data alignment strategy, this application can still effectively perform Granger causality tests in scenarios where the data granularity of candidate root cause nodes and abnormal nodes is inconsistent, avoiding test failures caused by granularity mismatch and expanding the applicability of the root cause inversion method of this application.

[0122] In some implementations, the multi-source crossover anomaly detection and root cause inversion method for time-series indicators may further include the following steps: (1) Obtain the significance probability value and test statistic of Granger causality test output; (2) If the significance probability value is less than the preset significance threshold, the causal strength score is calculated based on the significance probability value and the test statistic; the causal strength score is directly proportional to the test statistic and inversely proportional to the logarithm of the significance probability value; (3) Starting from the abnormal node, traverse the causal edge in reverse for the preset number of hops to construct a candidate propagation chain; (4) Calculate the propagation chain score for each candidate propagation chain; wherein, the propagation chain score is determined based on the product of the causal strength scores of each side in the candidate propagation chain and the length of the candidate propagation chain; (5) Candidate propagation chains whose propagation chain scores are greater than the preset link threshold are taken as root cause propagation paths.

[0123] After performing the Granger causality test, obtain the significance probability value and test statistic of the Granger causality test output. The significance probability value can be expressed as... The test statistic can be expressed as: In one specific implementation, only when A significant temporal causal relationship is considered to exist between corresponding indicator pairs only when the value is less than a preset significance threshold. The preset significance threshold can be set to 0.05, meaning only indicators with a significant temporal causal relationship are considered. Strong causal edges <0.05.

[0124] If the significance probability value is less than a preset significance threshold, a causal strength score is calculated based on the significance probability value and the test statistic. The causal strength score is directly proportional to the test statistic and inversely proportional to the logarithm of the significance probability value. In a specific implementation, the causal strength of index i pointing to index j can be expressed as: in, The smaller the value, the stronger the statistical significance of the causal relationship; The larger the value, the stronger the explanatory power of the candidate root cause node's historical values ​​for the outlier node. Therefore, the above formula can simultaneously combine significance and test strength to obtain a quantifiable causal strength score.

[0125] Furthermore, if the anomalous node is the target indicator, the system backtracks upstream along the preserved strong causal edges to identify possible anomalous propagation paths within a K-hop range. For example, when upstream indicator A first becomes anomalous, followed by indicator B, and ultimately leads to the target indicator C becoming anomalous, the system can form a candidate propagation chain of A→B→C.

[0126] For each candidate propagation chain, its propagation chain score is calculated. In one specific implementation, the propagation chain score can be expressed as: in, This represents the product of the causal strength scores of each causal edge in the candidate propagation chain. This indicates the length of the candidate propagation chain. Dividing by the propagation chain length can penalize excessively long propagation chains, avoiding the introduction of too much uncertainty due to excessively long links.

[0127] Finally, candidate propagation chains with scores greater than a preset link threshold are selected as root cause propagation paths. In one specific implementation, the preset link threshold... It can be set to 0.6, only Only the candidate propagation chains are retained as valid root cause propagation paths.

[0128] Furthermore, this application can also combine a historical case database to recommend and rank root causes. Specifically, in some embodiments, the time-series index multi-source crossover anomaly detection and root cause inversion method may further include the following steps: (1) Execute an anomaly detection algorithm on the detected indicators and select indicators whose values ​​deviate from the preset range as an abnormal indicator set; (2) Construct a first feature vector based on the set of abnormal indicators, and calculate the set similarity between the first feature vector and the second feature vector of each historical case in the preset historical case library; (3) Use a pre-set deep learning encoder to map the time series data corresponding to the abnormal indicator set into a third feature vector, and calculate the time series pattern similarity between the third feature vector and the fourth feature vector of each historical case; (4) The set similarity and temporal pattern similarity are weighted and fused according to the preset weights to obtain the comprehensive similarity; (5) Based on the comprehensive similarity, retrieve a preset number of target historical cases from the preset historical case library, and output the root cause tags associated with the target historical cases as recommended root causes.

[0129] The system first performs an anomaly detection algorithm on the detected indicators and filters out indicators whose values ​​deviate from the preset range as an abnormal indicator set. This abnormal indicator set may include multiple indicators that are abnormal at the same time or in sequence in the relevant indicator network in the same abnormal event, which are used to represent the indicator fingerprint of the current abnormal event.

[0130] A first feature vector is constructed based on a set of anomaly indicators, and the set similarity between the first feature vector and the second feature vectors of each historical case in a preset historical case library is calculated. In one specific implementation, the current set of anomaly indicators can be encoded as a one-hot vector, and the set similarity between the first feature vector and the second feature vectors of each historical case in the historical case library can be calculated. The fields are also encoded into corresponding indicator set vectors, and then the Jaccard similarity between the two is calculated as the indicator fingerprint matching result.

[0131] Furthermore, a pre-set deep learning encoder is used to map the time-series data corresponding to the anomaly indicator set into a third feature vector, and the temporal pattern similarity between the third feature vector and the fourth feature vector of each historical case is calculated. In a specific implementation, the pre-trained deep learning encoder can be a pre-trained ResNet-1D encoder or an LSTM encoder, mapping the temporal pattern corresponding to the current anomaly event into a 128-dimensional embedding vector; each historical case in the historical case library pre-stores... This field is also a 128-dimensional anomaly pattern vector. The system can calculate the 128-dimensional embedding of the current anomaly event and compare it with historical cases. The cosine similarity between them is used as the temporal pattern similarity.

[0132] Subsequently, the set similarity and temporal pattern similarity are weighted and fused according to preset weights to obtain a comprehensive similarity. In a specific implementation, the comprehensive similarity can be expressed as: in, This represents the set similarity between the current set of abnormal indicators and the set of related indicators from historical cases. This represents the cosine similarity between the current abnormal time-series pattern vector and the abnormal pattern vectors of historical cases. Through the above three-stage similarity matching, namely indicator fingerprint matching, time-series pattern matching, and comprehensive similarity fusion, the system can simultaneously use the set of indicators involved in the anomaly and the shape of the anomaly curve to find similar historical cases.

[0133] In one specific implementation, the preset historical case library may include the following fields: Used to represent a unique identifier for alarms; Used to indicate the time when the alarm occurred; Used to represent root cause indicators in historical cases; Used to indicate root cause type labels; Used to represent a set of related indicators; The system uses a 128-dimensional anomaly pattern vector for resolution; resolution represents the handling plan; and outcome represents the handling result, such as a real problem or a false alarm. The system can retrieve a preset number of target historical cases from a pre-defined historical case library based on comprehensive similarity, for example, outputting the Top-K most similar historical cases (K defaults to 5), and outputting the root cause labels associated with the target historical cases as recommended root causes.

[0134] Furthermore, after obtaining candidate propagation chains through causal inversion and similarity results through historical case matching, the system can use a root cause sorter to fuse causal strength, similarity score, and historical prior probability to perform a Top-N ranking of candidate root cause indicators. In a specific implementation, the root cause score of candidate root cause indicator m can be expressed as: in, Let sim(m) represent the propagation chain score of the propagation chain corresponding to the candidate root cause index m, sim(m) represent the comprehensive similarity obtained by matching the candidate root cause index m with historical cases, and prior(m) represent the prior probability obtained based on historical frequency statistics. The system can then... Output the Top-N root cause indicators and their credibility scores from high to low. The default value for N in Top-N is 5.

[0135] In some implementations, after obtaining multiple valid root cause propagation paths, the time-series index multi-source crossover anomaly detection and root cause inversion method may further include the following steps: (1) Extract the starting nodes of all valid root cause propagation paths to form a root cause candidate set; (2) For any two root cause candidate nodes in the root cause candidate set, query their shortest path length in the relevant index network; (3) If the shortest path length between two root cause candidate nodes is greater than the preset isolation hop count threshold, then the current abnormal event is determined to be a multi-root cause concurrent scenario; (4) For multi-root cause concurrent scenarios, each independent root cause candidate node is used as the starting point to independently construct its own corresponding propagation subgraph, and alarm suppression is performed on the downstream abnormal nodes in each propagation subgraph. That is, when outputting the root cause result, only the final abnormal node corresponding to the propagation subgraph is regarded as the affected object of this root cause path, and it is not included in the propagation range of other root cause paths. (5) Output the multiple independent root cause propagation paths and their respective credibility scores.

[0136] Therefore, by actively identifying and decoupling multiple root cause candidate nodes that are isolated from each other in the graph structure, this application can avoid false root cause reports caused by forcibly merging multiple independent fault links into a single root cause in multi-root cause concurrent scenarios, enabling operation and maintenance personnel to know multiple independent fault sources at the same time, and significantly improving the accuracy of root cause localization in complex fault scenarios.

[0137] In some implementations, after obtaining a set of candidate root causes through causal inversion, this application further identifies and processes scenarios where multiple candidate root cause nodes jointly point to the same anomalous node (i.e., multiple causes leading to one effect) to clarify whether each candidate root cause is a truly independent contributing root cause or whether there is a chain-like inclusion relationship, and ultimately determines the attribution result of the anomalous node: For any two candidate root cause nodes in the candidate root cause set and If both have a significant Granger causal relationship with the same anomalous node v (p-value < 0.05), then they constitute a candidate common root cause pair. , (v). Further analysis of common root causes in the middle and Perform a directed Granger causality test between them: if right There is significant Granger causality (i.e., yes In terms of upstream driving factors in time sequence, then ( , ,v) actually represents a → →v follows a chain propagation path, rather than a true multiple causes and one effect; at this point, Remove from the candidate root cause set, retaining only the link root causes. and will It has been reclassified as an intermediate propagation node.

[0138] For common root cause pairs that still exist after checking the chain relationships ( , ,v), Conditional Granger Causality (CGC), also known as Partial Granger Causality Test: In the control... Test under the influence of historical values Does significant Granger causality still exist for v (corresponding to p-value denoted as v)? ); in control Test under the influence of historical values Does significant Granger causality still exist for v (corresponding to p-value denoted as v)? ).

[0139] Based on the above test results, the attribution strategy should be determined according to the following rules: If <0.05 and If the value is less than 0.05, it is determined to be "true multiple causes and one effect". and All of these are independent root causes contributing to the abnormal node v, and the abnormal node v is simultaneously attributed to... and If only If <0.05, then only keep As the root cause, Lacking independent explanatory capacity; if only If <0.05, then only keep For the root cause; if ≥0.05 and If the value is ≥0.05, then neither of them has an independent contribution after controlling the other, and it is judged that they may share an implicit confounding upstream factor. The system will... and The common K-hop upstream nodes are listed as extended candidates for further Granger testing, and are marked with "The root cause relationship is mixed, and it is recommended to check the upstream indicators" when outputting in the current round.

[0140] Based on the partial Granger test results above, the attribution results for each outlier node v are structured and labeled according to the following types: Single Cause Attribution: Only one node in the candidate set passes the partial Granger test, and v is attributed to a single root cause; Multiple nodes pass the partial Granger test, and v is simultaneously attributed to multiple independent contributing root causes, with each root cause's output field containing conditional causality strength. If each candidate root cause node is structurally isolated within the relevant indicator network (shortest path length greater than a preset isolation hop count threshold), then a partial Granger test is unnecessary, and the node is initially identified as a structurally independent concurrent root cause. Further confirmation of its independent concurrent root cause status is based on statistical test results, and confidence levels are added accordingly. This helps avoid misjudgments caused by unmodeled relationships within the network. Set to 1; after the partial Granger test, all candidates lose significance and are marked as mixed candidates to be investigated, triggering upstream expansion candidate exploration.

[0141] In some implementations, for true multi-cause scenarios, Forecast Error Variance Decomposition (FEVD) can be further performed to quantify each independent root cause candidate m. k The proportion of the variance of the prediction error for outlier node v explained The FEVD results are stored in the root cause output field, enabling operations and maintenance personnel to quantify the relative contribution of each root cause in the current abnormal event, providing data basis for prioritizing concurrent fault handling.

[0142] Therefore, by introducing the partial Granger causality test into the multi-root-cause and co-cause scenario, this application can statistically distinguish between three types of situations: true multiple causes and one effect, chain-like pseudo-multiple root causes, and mixed relationships. This avoids misreporting redundant intermediate nodes in the chain as independent root causes, and clarifies the degree of independent contribution of each root cause to the same abnormal node in the multi-attribution scenario, thereby improving the accuracy and interpretability of root cause localization in the multi-root-cause concurrent scenario.

[0143] In some implementations, this application can be executed by a system including a lineage collector, an indicator network builder, a multi-source anomaly detector, a comprehensive anomaly scorer, a causal inversion engine, a historical case matcher, a root cause sorter, and an anomaly classification and handling unit. Specifically, the lineage collector collects data from the scheduling system, the indicator platform, and alarm logs; the indicator network builder constructs a relevant indicator network and stores it in a graph database; the multi-source anomaly detector executes statistical, machine learning, and business rule-based anomaly detection algorithms in parallel; the comprehensive anomaly scorer weights and fuses various anomaly scores and normalizes them to output a comprehensive anomaly score; the causal inversion engine backtracks along the indicator network and performs Granger causality tests; the historical case matcher matches historical alarm cases based on vector similarity; the root cause sorter outputs a Top-N root cause ranking based on causal strength, historical case similarity, and prior probability; and the anomaly classification and handling unit classifies anomalies according to decision rules and triggers differentiated handling processes. The above modules work together to enable the system to form a complete processing chain from data collection, network construction, anomaly detection, causal inversion, historical case matching, root cause sorting to anomaly classification and handling.

[0144] Furthermore, for existing systems where explicit lineage information is unavailable, implicit causal discovery based on correlation can be used as a fallback solution as an alternative. Specifically, implicit causal graphs between indicators can be automatically constructed by dynamically learning Bayesian networks from historical time-series data, and these implicit causal graphs can serve as the basis for related indicator networks or causal inversion networks. Although this approach may be less accurate than implementations based on explicit data lineage maps, it can still support anomaly detection and root cause inversion in scenarios where metadata is lacking, lineage links are incomplete, or historical systems cannot provide task dependencies.

[0145] In some implementations, after determining the root cause propagation path, the time-series index multi-source crossover anomaly detection and root cause inversion method may further include the following steps: (1) Traverse all valid root cause propagation paths, extract all intermediate and terminal nodes except the starting node in each propagation path, and form a propagation coverage node set. (2) For abnormal nodes that belong to the propagation coverage node set, mark their alarm status as "has been traced" and associate them with the corresponding root cause propagation path; (3) For the abnormal nodes that have been traced above, they will no longer be used as new independent alarms to trigger the causal inversion process, so as to avoid repeatedly performing Granger causality tests on multiple affected nodes in the propagation chain of the same abnormal event.

[0146] Therefore, by performing alarm suppression on downstream nodes after causal inversion, this application can significantly reduce the number of duplicate alarms caused by the spread of the same fault event in the relevant indicator network, reduce alarm noise for operation and maintenance personnel, and avoid triggering the computationally complex causal inversion process again for indicators that have been identified as downstream propagation nodes, thereby improving the overall computational efficiency of the system.

[0147] In some implementations, the multi-source crossover anomaly detection and root cause inversion method for time-series indicators may also include an online update step for the historical case database: (1) After the operation and maintenance personnel confirm or correct the root cause localization results output by the system, obtain the root cause confirmation label and handling results fed back by the operation and maintenance personnel. (2) Based on the set of abnormal indicators, time-series pattern vector, root cause confirmation label, handling plan and handling results corresponding to the current abnormal event, construct a new historical case record; (3) Record new historical cases in the preset historical case database; (4) Based on the newly written historical case records, update the historical frequency statistics of relevant root cause candidate indicators, and then update the prior probabilities corresponding to the root cause ranking. (5) In the preset historical case library, if the written processing result is "false alarm", the prior probability associated with the historical case with the comprehensive similarity to the case exceeds the preset similarity threshold is penalized and reduced to reduce the probability of the same candidate root cause being output in subsequent similar abnormal scenarios.

[0148] Therefore, by introducing an online update mechanism for the historical case library based on operation and maintenance feedback, this application can accumulate more and more real alarm cases as the system actually runs, continuously optimize the recall quality of historical case matching and the prior probability estimation of root cause ranking, and actively suppress high-frequency false alarm types through a false alarm feedback penalty mechanism, forming a continuously self-evolving root cause analysis capability.

[0149] To further address the issues of unclear handling priorities caused by the lack of differentiation between anomaly types in existing alarm information, and the mixing of business changes, data bugs, data delays, and seasonal fluctuations in the same alarm, some implementations of this multi-source cross-anomaly detection and root cause inversion method for time-series indicators may also include the following steps: (1) Extract the abnormal features of the cause of the abnormality; (2) Based on the abnormal characteristics, query the abnormal node to which the abnormal type belongs in the preset abnormal type library, and determine the corresponding handling suggestions for the abnormal type.

[0150] Anomaly features can be significant indicators used to characterize the causes of anomalies (i.e., anomalous nodes) in terms of temporal morphology, statistical distribution, and contextual relevance. In some implementations, anomaly features include at least: temporal morphological features, statistical distribution features, and contextual relevance features.

[0151] Among them, temporal morphological features are used to describe the geometry of abnormal waveforms, such as spikes, step changes, trends, seasonality, or stationary noise. Statistical distribution features are used to describe the numerical attributes of abnormal data, such as skewness, kurtosis, coefficient of variation, and duration of the anomaly. Contextual features, based on data lineage maps or indicator metadata, determine whether the anomaly occurred within a business change window, whether there are upstream data delay markers, etc.

[0152] The system extracts temporal morphological features, statistical distribution features, and contextual features from the time-series data of abnormal nodes using a sliding window algorithm or morphological operators, forming a high-dimensional feature vector.

[0153] The preset anomaly type library can be a pre-built knowledge base storing various known fault modes and their corresponding handling strategies. This library contains multiple standard anomaly templates, each associated with a specific anomaly type tag and operational handling suggestions. In some implementations, the preset anomaly type library can be obtained through offline training of historical fault cases, or it can be pre-entered by operations and maintenance experts based on their business experience.

[0154] Anomaly types can be qualitative classifications of the nature of the anomaly's cause. In some implementations, anomaly types include at least data bugs, data delays, business changes, and seasonal / external factors. Specifically, data bugs correspond to characteristics such as a sudden change exceeding 5σ, no associated business change event, and accompanying upstream ETL errors, triggering data rollback. Data delays correspond to characteristics such as delayed data timestamps and scheduling delays, and are handled by marking them as "pending" and waiting for recalculation after the upstream task is completed. Business changes correspond to characteristics such as a gradual trend in indicator patterns and association with business change orders, and are handled by pushing confirmation from the business party within a preset time limit. Seasonal / external factors correspond to characteristics such as strong correlation with the same period in history, and high similarity year-on-year or month-on-month, and are handled by only recording observations, without proactive alerts, and only escalating after the cumulative number exceeds a threshold.

[0155] In some implementations, when the same abnormal node hits multiple features at the same time, the priority is determined according to the order of data bug, data delay, business change, and seasonal / external factors, because the abnormal types ranked first often constitute the root cause of abnormal phenomena ranked later.

[0156] Therefore, this application can transform mixed raw alarms into structured anomaly types and handling suggestions, helping operation and maintenance personnel to quickly identify which system failures require urgent handling, which are negligible seasonal fluctuations, and which are business changes that need to be rolled back, thereby significantly reducing alarm noise and clarifying handling priorities.

[0157] Please see Figure 2 , Figure 2 This illustration shows a schematic diagram of a time series index multi-source crossover anomaly detection and root cause inversion device provided in an embodiment of this application. The time series index multi-source crossover anomaly detection and root cause inversion device 200 includes: an acquisition module 210, a construction module 220, a detection module 230, a fusion module 240, and an inversion module 250. Specifically: The acquisition module 210 is used to acquire data lineage maps and prior features of indicators; Module 220 is used to construct a network of related indicators containing upstream and downstream dependencies for the monitored indicators based on the data lineage graph and prior features of the indicators. The nodes of the related indicator network represent indicators, and the edges represent the correlation between indicators. Each edge is associated with the correlation strength and time lag relationship. The detection module 230 is used to perform anomaly detection on the monitored indicators in parallel. The anomaly detection includes at least two of the following: statistical anomaly detection, machine learning anomaly detection, and business rule anomaly detection. Each type of anomaly detection outputs a corresponding anomaly score. The fusion module 240 is used to perform weighted fusion of anomaly scores to obtain a comprehensive anomaly score. The inversion module 250 is used to identify nodes whose comprehensive anomaly scores exceed a preset threshold as anomaly nodes, and to perform causal inversion on the anomaly nodes along the relevant index network to obtain the anomaly causes of the anomalies and the corresponding credibility scores.

[0158] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0159] In the several embodiments provided in this application, the coupling or direct coupling or communication connection between the modules shown or discussed may be an indirect coupling or communication connection through some interface, device or module, and may be electrical, mechanical or other forms.

[0160] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0161] Please see Figure 3 , Figure 3 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. The electronic device 300 in this application may include one or more of the following components: a processor 310, a memory 320, and one or more application programs. The one or more application programs may be stored in the memory 320 and configured to be executed by one or more processors 310. The one or more programs are configured to execute the time-series index multi-source crossover anomaly detection and root cause inversion method as described in the foregoing method embodiments.

[0162] Processor 310 may include one or more processing cores. Processor 310 connects to various parts within the electronic device 300 using various interfaces and lines, and performs various functions and processes data of the electronic device 300 by running or executing instructions, programs, code sets, or instruction sets stored in memory 320, and by calling data stored in memory 320. Optionally, processor 310 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 310 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 310 and may be implemented separately using a communication chip.

[0163] The memory 320 may include random access memory (RAM) or read-only memory (ROM). The memory 320 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 320 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the various method embodiments described below, etc. The data storage area may also store data created by the electronic device 300 during use.

[0164] Please see Figure 4 , Figure 4 The diagram shows a computer-readable storage medium 400 provided in an embodiment of this application. The computer-readable storage medium 400 stores program code, which can be called by a processor to execute the time-series index multi-source crossover anomaly detection and root cause inversion method described in the above method embodiment.

[0165] The computer-readable storage medium 400 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium 400 includes a non-transitory computer-readable storage medium. The computer-readable storage medium 400 has storage space for program code 410 that performs any of the method steps described above. This program code can be read from or written to one or more computer program devices. The program code 410 may be compressed, for example, in a suitable form.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A time series index multi-source cross-exception detection and root cause inversion method, characterized in that, The method includes: Obtain the data lineage map and prior features of indicators; Based on the data lineage map and the prior features of the indicators, a related indicator network containing upstream and downstream dependencies is constructed for the monitored indicators; the nodes of the related indicator network represent indicators, the edges represent the correlation between indicators, and each edge is associated with the correlation strength and time lag relationship. Anomaly detection is performed in parallel on the monitored metrics. The anomaly detection includes at least two of the following: statistical anomaly detection, machine learning anomaly detection, and business rule-based anomaly detection. Each type of anomaly detection outputs a corresponding anomaly score. The anomaly scores are weighted and fused to obtain a comprehensive anomaly score; The nodes whose comprehensive anomaly scores exceed a preset threshold are identified as anomaly nodes. Causal inversion is then performed on the anomaly nodes along the relevant index network to obtain the anomaly causes of the anomalies and the corresponding credibility scores.

2. The timing index multi-source cross- anomaly detection and root cause inversion method according to claim 1, characterized in that, The method further includes: Extract the abnormal features of the cause of the abnormality; Based on the aforementioned abnormal characteristics, the abnormal node is queried in a preset abnormal type library to determine its corresponding abnormal type and the corresponding handling recommendations.

3. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 1, characterized in that, The prior features of the indicator include at least the indicator metadata and the indicator co-occurrence frequency matrix of the monitored indicator; Based on the data lineage map and the prior features of the indicators, the steps include constructing a network of related indicators containing upstream and downstream dependencies for the monitored indicators, including: Based on the data lineage map, historical data processing tasks are analyzed to obtain the upstream and downstream dependencies of the data processing tasks; Obtain the metadata of the monitored indicators from the indicator management platform; From the historical alarm logs, the co-occurrence frequency of co-occurrence index pairs is statistically analyzed according to a preset time window to obtain the index co-occurrence frequency matrix; Based on the upstream and downstream dependencies, the indicator metadata, and the indicator co-occurrence frequency matrix, the relevant indicator network is constructed.

4. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 3, characterized in that, The edge weights between any two indicators in the relevant indicator network are calculated using the following formula: in, As an indicator With indicators The correlation coefficient on historical preset duration data This is a normalized function for co-occurrence frequencies. As a time-delay stability index, , and These are preset weighting coefficients.

5. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 1, characterized in that, The step involves performing anomaly detection on the monitored metrics in parallel, including: The monitored indicators are detected using a pre-set set of statistical algorithms to determine the output score of each statistical algorithm in the pre-set set of statistical algorithms; the pre-set set of statistical algorithms includes at least one of the following: 3σ criterion algorithm, moving average algorithm, year-on-year detection algorithm, and month-on-month detection algorithm; The maximum value of the output score is taken as the first anomaly score; The monitored indicators are detected using a pre-defined set of machine learning models to determine the output score of each model in the pre-defined set of machine learning models; the pre-defined set of machine learning models includes an isolated forest model and a prophetic time series prediction model. The second anomaly score is obtained by weighted summation of the output scores of each model. Based on calendar features, the cosine similarity between the pattern vector and the baseline vector of the indicator to be monitored is calculated, and the cosine similarity is converted into a third anomaly score using a preset mapping relationship. The step involves weighting and fusing the anomaly scores to obtain a comprehensive anomaly score, including: The comprehensive anomaly score is obtained by weighting and fusing the first anomaly score, the second anomaly score, and the third anomaly score.

6. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 5, characterized in that, The step involves weighted fusion of the first anomaly score, the second anomaly score, and the third anomaly score to obtain the comprehensive anomaly score, including: Determine whether the current time period belongs to a preset special time period; If so, then obtain the preset weight group corresponding to the target special time period corresponding to the current time period; An initial anomaly score is calculated based on the first anomaly score, the second anomaly score, the third anomaly score, and the preset weighted group. The initial anomaly score is normalized to obtain the comprehensive anomaly score.

7. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 1, characterized in that, The step involves performing causal inversion on the abnormal nodes along the relevant index network, including: The abnormal node is selected from the nodes in the relevant indicator network with a preset number of hops as the initial screening nodes; Calculate the correlation coefficient between the initial screening node and the abnormal node; The nodes whose correlation coefficients are greater than or equal to a preset test threshold are identified as candidate root cause nodes. Based on the candidate root cause nodes, Granger causality tests are performed on the subsequence data within a preset offset time period centered on the preset anomaly interval.

8. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 7, characterized in that, The method further includes: Identify the data granularity type of the subsequence data; If the data granularity type is hourly, then the time delay parameter of the Granger causality test is set to the first preset value; If the data granularity type is daily, then the time delay parameter of the Granger causality test is set to the second preset value; If the data granularity type is weekly, then the time delay parameter of the Granger causality test is set to the third preset value; Based on the set time delay parameters, a vector autoregression model is constructed, and the steps of the vector autoregression model are performed to conduct Granger causality tests on the subsequence data within a preset offset time period centered on the preset anomaly interval according to the candidate root cause nodes.

9. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 7, characterized in that, The method further includes: Obtain the significance probability value and test statistic of the Granger causality test output; If the significance probability value is less than a preset significance threshold, a causal strength score is calculated based on the significance probability value and the test statistic; the causal strength score is directly proportional to the test statistic and inversely proportional to the logarithm of the significance probability value. Starting from the abnormal node, traverse the preset number of hops in reverse along the causal edge to construct a candidate propagation chain; Calculate the propagation chain score for each of the candidate propagation chains; wherein the propagation chain score is determined based on the product of the causal strength scores of each edge in the candidate propagation chain and the length of the candidate propagation chain; Candidate propagation chains whose scores are greater than a preset link threshold are selected as root cause propagation paths.

10. The method for multi-source crossover anomaly detection and root cause inversion of time-series indicators according to claim 1, characterized in that, The method further includes: An anomaly detection algorithm is executed on the detected indicators to filter out indicators whose values ​​deviate from the preset range as an abnormal indicator set; A first feature vector is constructed based on the set of abnormal indicators, and the set similarity between the first feature vector and the second feature vectors of each historical case in the preset historical case library is calculated. A preset deep learning encoder is used to map the time series data corresponding to the set of abnormal indicators into a third feature vector, and the time series pattern similarity between the third feature vector and the fourth feature vector of each historical case is calculated. The set similarity and the temporal pattern similarity are weighted and fused according to preset weights to obtain a comprehensive similarity. Based on the comprehensive similarity, a preset number of target historical cases are retrieved from the preset historical case library, and the root cause tags associated with the target historical cases are output as recommended root causes.