Root cause analysis method and device based on chaos experiment and electronic equipment

Abstract

The invention discloses a root cause analysis method and device based on a chaos experiment and electronic equipment. Relates to the chaos engineering field. The method comprises the following steps: obtaining experiment data generated in a chaos experiment process, and obtaining multi-source time sequence data; constructing a causal graph based on the multi-source time sequence data; performing causal effect quantification on variables in the causal graph by adopting a double-machine learning model to obtain a causal effect evaluation result, the causal effect evaluation result comprising causal effect estimation values of multiple groups of causal variables, each group of causal variables comprising a result variable and a processing variable causing the generation of the result variable, the causal effect estimation value of each group of causal variables is used for representing the influence degree of different processing variables on the result variable. The technical problem that in the prior art, in a chaos experiment scene, a dynamic risk analysis system based on a graph neural network performs causal reasoning, and the accuracy of a root cause analysis result is low is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chaos engineering, and in particular, to a root cause analysis method and device based on chaos experiments and electronic equipment. BACKGROUND

[0002] In the chaos experiment scenario, root cause analysis (RCA) is a systematic problem solving method, aiming to find out the root cause of the problem, so as to eliminate the problem and prevent the problem from happening again. With the development of information technology, root cause analysis technology has been gradually applied in the IT field, and has shown great potential in system stability optimization.

[0003] In the related art, the technology frontier focuses on the fusion application of deep learning and knowledge graph, and proposes a dynamic risk analysis system based on graph neural network, which realizes multi-hop reasoning by constructing a domain knowledge graph. The technical solution includes the following steps:

[0004] Step 1) Use dynamic graph embedding technology to process time series data; Step 2) Introduce attention mechanism to improve key path identification accuracy; Step 3) Combine reinforcement learning to generate optimization strategy.

[0005] However, there are the following defects: 1) Confusion between causal reasoning and correlation analysis, mistaking accompanying phenomena as root causes; 2) Knowledge graph construction relies on manual annotation, with high domain adaptation cost; 3) The execution of optimization suggestions based on root cause analysis results is poor, and the actual adoption rate is low.

[0006] In view of the above problems, no effective solution has been proposed so far. SUMMARY

[0007] The embodiments of the present application provide a root cause analysis method and device based on chaos experiments and electronic equipment, to at least solve the technical problem of low accuracy of root cause analysis results in the related art in the chaos experiment scenario, based on the dynamic risk analysis system of the graph neural network.

[0008] According to an aspect of an embodiment of the present application, there is provided a root cause analysis method based on chaotic experiments, comprising: obtaining experimental data generated in a chaotic experiment, to obtain multi-source time series data, wherein the chaotic experiment is used to verify the stability of a target system in a manner of simulating and manufacturing faults and instability factors in the target system; constructing a causal graph based on the multi-source time series data, wherein the causal graph includes nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables include variables representing data features of the multi-source time series data; and quantifying causal effects of variables in the causal graph using a double-machine learning model to obtain a causal effect evaluation result, wherein the causal effect evaluation result includes causal effect estimation values of a plurality of groups of causal variables, each group of causal variables includes a result variable and a processing variable causing the result variable, and the causal effect estimation value of each group of causal variables represents the influence degree of different processing variables on the result variable.

[0009] Further, the causal effect evaluation result further includes a confidence interval of each group of causal variables, the confidence interval of each group of causal variables is used to represent the uncertainty degree of the causal effect estimation value of the group of causal variables, and after the double-machine learning model is used to quantify the causal effects of the variables in the causal graph to obtain the causal effect evaluation result, the method further includes: obtaining optimization suggestions for the target system to obtain a plurality of optimization suggestions, wherein each optimization suggestion is associated with at least one group of causal variables; determining a weight of each optimization suggestion based on the causal effect estimation value and the confidence interval of at least one group of causal variables associated with the optimization suggestion; and determining a target optimization strategy based on the weight of each optimization suggestion, wherein the target optimization strategy is used to optimize the target system.

[0010] Further, constructing the causal graph based on the multi-source time series data comprises: pre-processing the multi-source time series data to obtain processed multi-source time series data, wherein the pre-processing includes at least one of the following: data alignment, noise removal, missing value processing, and data standardization; extracting features from the processed multi-source time series data to obtain the plurality of variables; identifying a dependency relationship between the plurality of variables to obtain a dependency relationship set, and constructing the causal graph based on the dependency relationship set and the plurality of variables.

[0011] Further, based on the weight of each optimization suggestion, a target optimization strategy is determined, including: based on the weight of each optimization suggestion, screening a plurality of optimization suggestions to obtain screened optimization suggestions; based on the screened optimization suggestions and a multi-objective combination optimization algorithm, determining an initial optimization strategy, and verifying a causal relationship in the initial optimization strategy to obtain a first verification result, wherein the first verification result is used to indicate whether the initial optimization strategy conforms to the causal logic of the target system, and the target of the multi-objective combination optimization algorithm includes at least one of the following: causal effect maximization, implementation cost minimization; based on the first verification result, the target optimization strategy is determined.

[0012] Further, based on the set of dependency relationships and the plurality of variables, the causal graph is constructed, including: based on the set of dependency relationships and the plurality of variables, an initial causal skeleton is generated using a causal inference strategy, wherein the causal inference strategy includes: conditional independence test, Granger causality verification and target rule, the target rule is used to direct the potential edge representing the causal relationship, and the initial causal skeleton includes: nodes representing the plurality of variables and edges representing the causal relationship to be verified; verifying the causal relationship in the initial causal skeleton to obtain a second verification result; based on the second verification result and the initial causal skeleton, the causal graph is obtained.

[0013] Further, the dual machine learning model includes: a result variable model for predicting the result variable, and a processing variable model for predicting the processing variable corresponding to the result variable. The dual machine learning model is used to quantify the causal effect of the variables in the causal graph to obtain a causal effect evaluation result, including: inputting the variables in the causal graph into the dual machine learning model, predicting the predicted value of the result variable through the result variable model, and predicting the predicted value of the processing variable corresponding to the result variable through the processing variable model, to obtain a plurality of groups of predicted values of the causal variables; based on the predicted value of each group of causal variables, a linear regression model is established; based on the linear regression model, the causal effect evaluation result is obtained.

[0014] Further, after establishing the linear regression model based on the predicted value of each group of causal variables, it further includes: determining a plurality of individuals of the target system, wherein a plurality of individuals are used to constitute the target system, and each individual is associated with at least one group of causal variables; using a causal forest algorithm to evaluate the treatment effect of each individual, wherein the treatment effect of the individual is used to represent the difference between the first type of treatment of accepting the target processing variable and the second type of treatment of not accepting the target processing variable of the individual, and the target processing variable is the processing variable associated with the individual.

[0015] According to another aspect of the embodiments of the present application, there is also provided a root cause analysis device based on chaotic experiments, comprising: a first obtaining unit configured to obtain experimental data generated in a chaotic experiment, to obtain multi-source time series data, wherein the chaotic experiment is configured to verify stability of a target system by simulating and manufacturing factors of failure and instability in the target system; a constructing unit configured to construct a causal graph based on the multi-source time series data, wherein the causal graph comprises nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables comprise variables representing data features of the multi-source time series data; and a processing unit configured to quantize causal effects of variables in the causal graph by using a double-machine learning model, to obtain a causal effect evaluation result, wherein the causal effect evaluation result comprises causal effect estimation values of a plurality of groups of causal variables, each group of the causal variables comprises a result variable and a processing variable causing the result variable, and the causal effect estimation value of each group of the causal variables represents an influence degree of different processing variables on the result variable.

[0016] Further, the causal effect evaluation result further comprises a confidence interval of each group of the causal variables, the confidence interval of each group of the causal variables represents an uncertainty degree of the causal effect estimation value of the group of the causal variables, and the root cause analysis device based on chaotic experiments comprises: a second obtaining unit configured to obtain suggestions for optimizing the target system after the processing unit quantizes the causal effects of the variables in the causal graph by using the double-machine learning model to obtain the causal effect evaluation result, to obtain a plurality of optimization suggestions, wherein each optimization suggestion is associated with at least one group of the causal variables; a first determining unit configured to determine a weight of each optimization suggestion based on the causal effect estimation value and the confidence interval of at least one group of the causal variables associated with each optimization suggestion; and a second determining unit configured to determine a target optimization strategy based on the weight of each optimization suggestion, wherein the target optimization strategy is configured to optimize the target system.

[0017] Further, the constructing unit comprises: a first processing subunit configured to pre-process the multi-source time series data to obtain processed multi-source time series data, wherein the pre-processing comprises at least one of the following: data alignment, noise removal, missing value processing, and data standardization; an extracting subunit configured to extract features from the processed multi-source time series data to obtain the plurality of variables; and a constructing subunit configured to identify a dependency relationship between the plurality of variables to obtain a dependency relationship set, and construct the causal graph based on the dependency relationship set and the plurality of variables.

[0018] Further, the second determining unit comprises: a screening subunit, configured to screen the multiple optimization suggestions based on the weight of each optimization suggestion to obtain screened optimization suggestions; a verification subunit, configured to determine an initial optimization strategy based on the screened optimization suggestions and a multi-objective combined optimization algorithm, and verify the cause-effect relationship in the initial optimization strategy to obtain a first verification result, wherein the first verification result is used to indicate whether the initial optimization strategy conforms to the cause-effect logic of the target system, and the target of the multi-objective combined optimization algorithm includes at least one of the following: cause-effect maximization and implementation cost minimization; and a first determining subunit, configured to determine the target optimization strategy based on the first verification result.

[0019] Further, the constructing subunit comprises: a generating module, configured to generate an initial cause-effect skeleton based on the set of dependency relationships and the multiple variables by using a cause-effect inference strategy, wherein the cause-effect inference strategy includes conditional independence test, Granger cause-effect relationship verification and target rule, the target rule is used to direct potential edges representing cause-effect relationships, and the initial cause-effect skeleton includes nodes representing the multiple variables and edges representing cause-effect relationships to be verified; a verification module, configured to verify the cause-effect relationships in the initial cause-effect skeleton to obtain a second verification result; and a processing module, configured to obtain the cause-effect graph based on the second verification result and the initial cause-effect skeleton.

[0020] Further, the double-machine learning model comprises a result variable model and a processing variable model, the result variable model is used to predict the result variable, and the processing variable model is used to predict the processing variable corresponding to the result variable, the processing unit comprises: a second processing subunit, configured to input the variables in the cause-effect graph into the double-machine learning model, predict the predicted value of the result variable by using the result variable model, and predict the predicted value of the processing variable corresponding to the result variable by using the processing variable model to obtain multiple groups of predicted values of the cause-effect variables; a establishing subunit, configured to establish a linear regression model based on the predicted value of each group of cause-effect variables; and a third processing subunit, configured to obtain the cause-effect effect evaluation result based on the linear regression model.

[0021] Further, the processing unit further comprises: a second determining sub-unit, configured to determine a plurality of individuals of the target system after establishing the linear regression model based on the predicted values of each group of the causal variables, wherein the plurality of individuals are used to constitute the target system, and each of the individuals is associated with at least one group of the causal variables; and an evaluation sub-unit, configured to evaluate the processing effect of each of the individuals by using the causal forest algorithm, wherein the processing effect of the individual is used to represent the difference between a first type of processing in which the individual receives a target processing variable and a second type of processing in which the individual does not receive the target processing variable, and the target processing variable is a processing variable associated with the individual.

[0022] According to another aspect of the embodiments of the present application, an electronic device is also provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the chaos experiment-based root cause analysis method of any one of the above by executing the executable instructions.

[0023] According to another aspect of the embodiments of the present application, a computer readable storage medium is also provided, which stores a computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the chaos experiment-based root cause analysis method of any one of the above when the computer program is running.

[0024] In the present application, by the following steps: obtaining experimental data generated in a chaos experiment process to obtain multi-source time series data, wherein the chaos experiment is used to verify the stability of a target system in a manner of simulating and manufacturing faults and instability factors in the target system; constructing a causal graph based on the multi-source time series data, wherein the causal graph includes nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables include variables representing data features of the multi-source time series data; and quantifying causal effects of the variables in the causal graph by using a dual-machine learning model to obtain a causal effect evaluation result, wherein the causal effect evaluation result includes causal effect estimation values of a plurality of groups of causal variables, each group of causal variables includes a result variable and a processing variable causing the result variable, and the causal effect estimation value of each group of causal variables is used to represent the degree of influence of different processing variables on the result variable. Thus, the technical problem of low accuracy of root cause analysis results in the related art in which a dynamic risk analysis system based on a graph neural network performs causal reasoning in a chaos experiment scenario is solved.

[0025] In the present application, by constructing a causal graph and performing causal effect evaluation by using a dual-machine learning algorithm to obtain a causal effect evaluation result, the situation of low accuracy of root cause analysis results in the related art in which a dynamic risk analysis system based on a graph neural network performs causal reasoning is avoided, the purpose of accurately distinguishing direct causes from root causes in a chaos experiment is achieved, and the technical effect of improving the accuracy of causal analysis results is achieved. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0027] Figure 1 This is a flowchart of an optional root cause analysis method based on chaotic experiments according to an embodiment of the present invention;

[0028] Figure 2 This is a flowchart of an optional root cause analysis based on a chaotic experiment according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of an optional root cause analysis device based on a chaotic experiment according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] For ease of description, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0034] Chaos engineering is a method that applies chaos theory to the field of engineering, aiming to test and improve the reliability and robustness of systems by simulating and introducing faults and instabilities into the system. It does this by intentionally causing failures and anomalies in the production environment to verify the system's response and recovery capabilities to these failures, in order to identify and address weaknesses in the system, thereby reducing the risk and cost of the system. The basic idea is to verify the performance of the system under fault conditions by intentionally injecting faults. These faults can be hardware faults, service faults, software faults, etc. According to the chaos engineering theory, a chaos engineering platform suitable for civil aviation business characteristics can be developed according to the characteristics of the software and the needs of the customers. This platform can realize the docking of three existing tool platforms of domain monitoring, operation and maintenance, and testing, meet the development and use needs of customers, and can be actively controlled.

[0035] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.), the collected information and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards in relevant regions, necessary security measures are taken, do not violate public order and good customs, and provide corresponding operation portal for user to choose authorization or refusal.

[0036] Embodiment one

[0037] According to the embodiment of the present application, an optional root cause analysis method based on chaos experiment is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from here.

[0038] Figure 1 is a flowchart of an optional root cause analysis method based on chaos experiment according to an embodiment of the present application, as shown in Figure 1 The method comprises the following steps:

[0039] Step S101, obtaining experimental data generated in the chaos experiment process to obtain multi-source time series data, wherein the chaos experiment is used to simulate and introduce faults and instabilities in the target system to verify the stability of the target system.

[0040] Data collection is the first step of the chaos experiment process. In this stage, experimental data can be obtained according to the research object and experimental data through experimental equipment. Common experimental equipment can include sensors, data recorders, etc.

[0041] The experimental data described above can be multi-source time series data, chaos experiment observation indicators, system configuration parameters, etc. Chaos experiment observation data can include but is not limited to: CPU (Central Processing Unit), memory, IO (Input / Output), request delay, error rate, etc. System configuration parameters can include but are not limited to: thread pool size, connection pool upper limit, K8s resource quota, etc. In order to ensure the integrity and accuracy of the data, the method of multiple repeated experiments can be used to reduce experimental errors.

[0042] Step S102, based on the multi-source time series data, a causal graph is constructed, wherein the causal graph includes nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables include variables representing data features of the multi-source time series data.

[0043] In this embodiment, the multi-source time series data can be preprocessed and feature extracted to obtain variables reflecting the data features of the multi-source time series data. The extracted variables can be used as nodes, and the causal relationships defined based on domain knowledge or automatic discovery (such as PC algorithm) can be used as edges. An improved FCI (Fast Causal Inference) algorithm is used to construct the causal graph.

[0044] Step S103, using a double machine learning model to quantify the causal effect of the variables in the causal graph, obtaining a causal effect evaluation result, wherein the causal effect evaluation result includes a plurality of groups of causal variable causal effect estimation values, each group of causal variables includes a result variable and a processing variable that causes the result variable, and the causal effect estimation value of each group of causal variables is used to represent the influence degree of different processing variables on the result variable.

[0045] For example, in this embodiment, double machine learning can be applied to quantify the causal effect to obtain a causal effect reasoning result (i.e., a causal effect evaluation result). Double machine learning can eliminate confounding variable bias through phased modeling, achieve causal effect estimation in high-dimensional scenarios, effectively improve estimation accuracy, and meet the system's fine-grained attribution needs.

[0046] Through the above steps, in this embodiment, the causal effect evaluation result is obtained by constructing a causal graph and using a double-machine learning algorithm to evaluate the causal effect, which avoids the situation that the root cause analysis result accuracy is low in the related art dynamic risk analysis system based on a graph neural network for causal reasoning, and achieves the purpose of accurately distinguishing the direct cause and the root cause in the chaotic experiment, thereby realizing the technical effect of improving the accuracy of the causal analysis result. Further, the technical problem of low root cause analysis result accuracy in the related art dynamic risk analysis system based on a graph neural network for causal reasoning in a chaotic experiment scenario is solved.

[0047] Optionally, the causal effect evaluation result further includes a confidence interval of each group of causal variables, and the confidence interval of each group of causal variables is used to represent the uncertainty degree of the causal effect estimation value of the group of causal variables. After the double-machine learning model is used to quantify the causal effect of the variables in the causal graph, the causal effect evaluation result is obtained, and the following steps are further included: obtaining a suggestion for optimizing the target system to obtain a plurality of optimization suggestions, wherein each optimization suggestion is associated with at least one group of causal variables; determining the weight of each optimization suggestion based on the causal effect estimation value and the confidence interval of at least one group of causal variables associated with each optimization suggestion; and determining a target optimization strategy based on the weight of each optimization suggestion, wherein the target optimization strategy is used to optimize the target system.

[0048] The causal effect evaluation result can further include a causal effect estimation value (e.g., thread pool + 1→ CPU utilization increases by 2.3%) and its statistical properties, such as a confidence interval (reflecting the uncertainty of the estimation), a p-value (indicating whether the causal effect is statistically significant), etc. Different effect quantities represent different causal influences of a treatment variable T (e.g., a controllable system parameter) on a result variable Y (e.g., a performance indicator), based on which direct causes and root causes can be located.

[0049] In this embodiment, the weight of each optimization suggestion (referred to as a suggestion) in the knowledge base can be calculated according to the causal effect estimation value and its statistical properties (such as a confidence interval, a p-value, etc.), and based on the weight of each suggestion, a target optimization strategy can be determined, as follows:

[0050] The weight of each suggestion in the knowledge base is calculated (using a knowledge entry matching degree), a candidate set (e.g., an initial optimization strategy) is selected according to the calculation result, and based on the causal effect estimation value and the confidence interval [θ low ,θ high], the matching degree of each suggestion is calculated as the weight Wi of each suggestion, where each suggestion can be associated with at least one set of causal variables, for example, the treatment variable can be the "atomic operation unit" of the suggestion, such as suggestion = treatment variable + adjustment direction / target value + action range, for example, increasing the thread pool of the order service from 50 to 80, one suggestion can be associated with multiple treatment variables T, and the weight calculation formula of each suggestion is:

[0051]

[0052] wherein, represents the causal effect estimate value of the jth treatment variable on the result variable, which can be calculated by a causal inference method (such as DoWhy); e ij is the effect enhancement factor of the jth treatment variable related to the ith suggestion, which can reflect the effect intensity of the treatment variable under the specific suggestion; represents the standard deviation of the causal effect estimate value , indicating the uncertainty of the estimate value, wherein the smaller the standard deviation, the more reliable the estimate value; CI j represents the confidence level of the causal effect estimate value . The CI j closer to 1, the higher the accuracy of the estimate value; γ represents an adjustment parameter (default 0.1) that controls the degree of influence of the confidence level on the weight; N represents the number of treatment variables.

[0053] In this embodiment, a multi-objective combinatorial optimization algorithm can also be defined (for example, the objectives include maximizing the overall causal effect, minimizing the implementation cost, reducing the risk, etc.), and a branch and bound method is used to solve the algorithm to output a preliminary optimization scheme (i.e., an initial optimization strategy). Afterwards, the preliminary optimization scheme can be subjected to causal path verification to check whether the scheme conforms to the causal logic of the system, and according to the conforming scheme, a final executable scheme list (i.e., a target optimization strategy) is output, including the implementation sequence, the expected effect, and the key considerations. The implementation sequence can be dynamically adjusted according to the causal dependency relationship and the priority.

[0054] Optionally, based on the multi-source time series data, a causal graph is constructed, including: preprocessing the multi-source time series data to obtain processed multi-source time series data, wherein the preprocessing includes at least one of the following: data alignment, noise removal, missing value processing, and data standardization; extracting features from the processed multi-source time series data to obtain a plurality of variables; identifying dependency relationships between the plurality of variables to obtain a dependency relationship set, and constructing a causal graph based on the dependency relationship set and the plurality of variables.

[0055] In the embodiment, data preprocessing is an important link of data analysis, including data alignment, noise removal, missing value processing, data standardization and other operations. Data alignment can use dynamic time warping (DTW) algorithm to align multi-source time series data, realize data alignment. Noise removal can use filtering, smoothing and other techniques to improve data quality. Processing missing values can delete records containing missing values, fill missing values using multiple imputation (MICE) algorithm, etc. Data standardization can eliminate the differences between different dimensions. Through data preprocessing, high-quality data foundation can be provided.

[0056] In the embodiment, feature extraction can be performed on the preprocessed data to convert the multi-source time series data into a more easily analyzed form, the purpose being to extract key variables that can reflect the characteristics of the data and obtain the plurality of variables described above.

[0057] In the embodiment, the causal dependency relationship between variables can be preliminarily identified based on domain knowledge or automatic discovery (such as PC algorithm), and a candidate edge set (corresponding to a dependency relationship set) is generated. The domain knowledge includes an extensible constraint rule library, including two types of prior knowledge: hard constraint: irreversible causal relationship determined according to system architecture (such as "thread pool size → CPU utilization"); soft constraint: trusted causal relationship based on historical operation and maintenance experience (such as "database connection number → request delay" confidence 0.8). In an optional example, an automatic discovery algorithm can be used in combination with domain knowledge for iterative processing. In the algorithm iteration process, the confidence score of the candidate causal edge is corrected through Bayesian posterior probability. By optimizing the algorithm and integrating the domain knowledge, it is beneficial to distinguish between direct causes and root causes, effectively reduces the misjudgment rate, and meets the system risk analysis requirements.

[0058] After obtaining the generated candidate edge set, domain knowledge constraints can be injected to handle potential hidden variables (for example, indirect causal variables) and feedback loops to obtain verified causal relationships (edges). Finally, the nodes representing the variables and the verified causal relationships (edges) are integrated to form a directed acyclic graph (DAG) or an extended causal graph containing hidden variables (corresponding to the causal graph described above).

[0059] Optionally, based on the weight of each optimization suggestion, a target optimization strategy is determined, including: based on the weight of each optimization suggestion, screening the plurality of optimization suggestions to obtain screened optimization suggestions; based on the screened optimization suggestions and a multi-objective combination optimization algorithm, determining an initial optimization strategy, and verifying the causal relationship in the initial optimization strategy to obtain a first verification result, wherein the first verification result is used to indicate whether the initial optimization strategy conforms to the causal logic of the target system, and the target of the multi-objective combination optimization algorithm includes at least one of the following: causal effect maximization, implementation cost minimization; based on the first verification result, determining the target optimization strategy.

[0060] Based on the weight of each optimization suggestion, a set of optimization suggestions that are most likely to bring about the desired outcome, e.g., suggestions with higher weights, are filtered from the suggestion set. The filtered optimization suggestions can then be combined with a multi-objective combinatorial optimization algorithm to consider multiple optimization objectives simultaneously, e.g., both to reduce the frequency of CPU overload as much as possible (maximizing causal effect) and to ensure the lowest implementation cost (minimizing implementation cost). The preliminary optimization strategy output by the algorithm can also be verified for causal relationships, i.e., checking whether each suggestion in the strategy is consistent with the actual causal logic of the system. For example, if the strategy suggests increasing the thread pool size to address CPU overload, the verification process will check whether this suggestion actually addresses the root cause of CPU overload, rather than merely alleviating a surface symptom. The verification process can produce a first verification result reflecting the consistency of the preliminary optimization strategy with the intrinsic causal relationships of the system. If the causal relationships in the strategy are verified to be correct, the first verification result will indicate that the strategy is consistent with the logic of the system. Based on the first verification result, the preliminary optimization strategy is adjusted. If some suggestions are found to have incorrect causal relationships or to be inconsistent with the target system logic, these suggestions are excluded, and ultimately, the target optimization strategy is formed from the verified optimization suggestions, which effectively address the problems in the system while taking into account the implementation cost and expected effect.

[0061] For example, a multi-objective combinatorial optimization algorithm can be defined (e.g., objectives include maximizing overall causal effect, minimizing implementation cost, reducing risk, etc.), and a branch-and-bound method can be used to solve the algorithm to output a preliminary optimization solution (for the filtered optimization suggestions). Specifically, after the multi-objective combinatorial optimization algorithm is defined, a Pareto front optimization problem can be constructed based on the multi-objective combinatorial optimization algorithm, and then a branch-and-bound method can be used to solve the Pareto front optimization problem to output a preliminary optimization solution. Thereafter, the preliminary optimization solution can be verified for causal path, checking whether the solution is consistent with the causal logic of the system, and based on the consistent solution, a final executable solution list is output, including the implementation order, expected effect, and key considerations. The implementation order can be dynamically adjusted based on causal dependencies and priorities.

[0062] In an alternative example, the branch and bound method can be an algorithm widely applied to solve combinatorial optimization problems, especially when the problem (e.g., recommendation) scale is large and the search space is extensive. It works by creating a search tree, breaking down the problem into subproblems, and gradually narrowing down the search range until the optimal solution that meets all objectives is found. The main advantage of the branch and bound method is that it can avoid enumerating all possible options, reducing the computation time and resource consumption. Specifically, it includes: (1) problem decomposition: the original problem is decomposed into several subproblems, each corresponding to a branch of the decision tree. For example, for the selection of optimization recommendations, all possible combinations can be considered as different branches. (2) boundary setting: set the upper and lower bounds for each branch (subproblem), i.e., the best and worst objective values that the branch can achieve. This is estimated in advance through local optimization or heuristic methods. (3) branch expansion: expand those branches that look the most promising. This means selecting those subproblems that are likely to bring the best solution for further exploration according to the boundary estimates. (4) bounding pruning: once the solution (or boundary estimate) of a branch cannot surpass the currently known optimal solution, it is pruned. This avoids wasting computational resources on subproblems that are unlikely to find better solutions. (5) iterative solution: the above (1) to (4) are iterated until all branches of the search tree are evaluated or pruned. In this way, the algorithm can eventually find one or more sets of optimization recommendations that take into account all objective functions and achieve the best balance. Through the branch and bound method, one or more preliminary optimization schemes can be output, each based on the comprehensive evaluation of multiple objective functions (such as overall causal effect, implementation cost, risk). These schemes can include combinations of multiple optimization recommendations, aiming to improve system performance, cost-effectiveness, and risk management.

[0063] Incorporating causal effect estimates and resource constraints into the same optimization framework can break through the limitations of local optimization caused by traditional rules and innovatively integrate causal effect confidence intervals into weight calculation, automatically recommending conservative solutions in high-uncertainty scenarios. In addition, by associating the optimization database, executable optimization recommendations can be obtained, improving the actual adoption rate.

[0064] Optionally, based on the set of dependency relationships and the plurality of variables, a causal graph is constructed, including: based on the set of dependency relationships and the plurality of variables, an initial causal skeleton is generated using a causal inference strategy, wherein the causal inference strategy includes: conditional independence test, Granger causality check, and target rule, the target rule is used to direct potential edges representing causal relationships, and the initial causal skeleton includes: nodes representing the plurality of variables and edges representing causal relationships to be verified; verifying the causal relationships in the initial causal skeleton to obtain a second verification result; and based on the second verification result and the initial causal skeleton, a causal graph is obtained.

[0065] For example, first, an initial causal skeleton can be generated using a modified FCI algorithm (corresponding to a causal inference strategy) with conditional independence tests based on the node and candidate edge set (corresponding to the set of dependency relationships), a Granger causality test is used to identify the temporal causal relationship of the nodes in the initial causal skeleton, and a V-structure rule (corresponding to a target rule describing the conditional dependency relationship between variables) is applied to orient the potential causal edges;

[0066] Then, domain knowledge constraints can be injected to handle potential hidden variables and feedback loops, and verified causal relationships (edges) are obtained.

[0067] Finally, the nodes and verified causal relationships (edges) can be integrated to form a directed acyclic graph (DAG) or an extended causal graph containing hidden variables (i.e., a causal graph).

[0068] It should be noted that the confidence and main contributing factors of each edge in the causal relationship graph (i.e., the causal graph) can be annotated to improve interpretability.

[0069] In this embodiment, the improved FCI algorithm is implemented based on the improved FCI algorithm, which introduces three-stage optimization based on the traditional PC algorithm:

[0070] Stage 1: Based on conditional independence tests (G 2 tests) to generate an initial causal skeleton, set the significance threshold α = 0.01 to reduce false edges.

[0071] Stage 2: Identify the temporal causal relationship by Granger causality test, set the maximum lag step L = 5 (determined by spectral analysis).

[0072] Stage 3: Apply V-structure rule to orient potential causal edges and solve Markov equivalence ambiguity problem.

[0073] By using the improved FCI algorithm, the final causal graph is determined, which achieves the technical effect of improving the accuracy of causal relationships in the causal graph.

[0074] Optionally, the double machine learning model includes a result variable model and a processing variable model, the result variable model is used to predict the result variable, and the processing variable model is used to predict the processing variable corresponding to the result variable. The double machine learning model is used to quantify the causal effect of the variables in the causal graph to obtain a causal effect evaluation result, including: inputting the variables in the causal graph into the double machine learning model, predicting the predicted value of the result variable through the result variable model, and predicting the predicted value of the processing variable corresponding to the result variable through the processing variable model, to obtain multiple sets of predicted values of causal variables; based on the predicted value of each set of causal variables, a linear regression model is established; based on the linear regression model, the causal effect evaluation result is obtained.

[0075] The double machine learning is applied to quantify the causal effect to obtain a causal effect reasoning result (i.e., a causal effect evaluation result). The double machine learning eliminates the bias of confounding variables through phased modeling to realize the estimation of the causal effect in a high-dimensional scene, effectively improve the estimation accuracy, and meet the fine attribution demand of the system.

[0076] Specifically, first, variable division can be performed: a processing variable T (such as a thread pool size), a result variable Y (such as CPU utilization), and a confounding variable X (such as request QPS and memory occupancy) are obtained; then, through causal graph reasoning, the influence path of the confounding variable on T and Y is identified and verified, wherein the processing variable T is a controllable system parameter; the result variable Y is a performance indicator; the confounding variable X is a common factor affecting T and Y, that is, T←X→Y (for example, request QPS has an impact on the thread pool size and also has an impact on the CPU utilization); and the auxiliary variable W is a covariate (non-confounding variable) that may only affect Y or T (such as environmental temperature and network delay), and the introduction of W can improve the model accuracy of the first stage.

[0077] The specific implementation process of the double machine learning framework is as follows:

[0078] (1) Data division and model training:

[0079] Sample division: K-fold cross-validation (K=5) is adopted to divide the data into a training set (80%) and an auxiliary set (20%) to avoid overfitting.

[0080] First-stage modeling:

[0081] Result variable model: a gradient boosting tree (such as XGBoost) is used to predict the result variable Y (such as CPU utilization), and the input is the confounding variable X (request quantity, memory occupancy, etc.) and the auxiliary variable W (environmental temperature, network state, etc.).

[0082] Processing variable model: a logistic regression or a random forest is used to predict the processing variable T (such as thread pool configuration adjustment), and the input is the same group of variables X and W.

[0083] Residual calculation: the model prediction value is subtracted from the original data to obtain the residual after bias elimination and Y and T are original data, Y pred and T pred are prediction values.

[0084] The result variable model is used to obtain the predicted value of Y, and the processing model is used to obtain the predicted value of T. The residual error is calculated by the two predicted values, and two residual errors are obtained to control the pure fluctuation after the influence of X and W, that is, to filter the influence of X and W, and the remaining is the change brought by T and Y itself. The significance of this step is to more accurately estimate the true causal effect of the processing variable T on the result variable Y, mainly serving the second stage.

[0085] In combination with the second stage, the following is illustrated by way of example:

[0086] Suppose that the data given to the training model in the first stage is: the result variable Y is CPU utilization, the actual observed original value Y of the result variable model is 70%, and the predicted value T pred is 60%. Then the residual error Since the input value takes into account X and W, the additional 10% is considered to be caused by T or other unobserved factors. Taking the processing variable T as the thread pool size, the actual observed original value T of the processing variable model is 80, and the predicted value T pred is 50. Then the residual error As above, the additional 30 is considered to be part of the artificial adjustment or abnormal configuration in the thread pool independent of X and W.

[0087] (2) Causal effect estimation:

[0088] Second stage regression: establish a linear regression model Where θ is the regression coefficient, that is, the estimator of the average treatment effect (ATE), and ε is a random variable. (The regression model here assumes that it reflects the correlation between T and Y, not X and W);

[0089] Heterogeneity analysis: estimate the individual treatment effect (ITE) through the causal forest algorithm (Causal Forest) to identify the effect difference under different individuals (for example, virtual machine service instances in the target system can all be individuals, and all individuals form the total population (i.e., the target system)).

[0090] (3) Analyze the preliminary causal effect estimation result (ATE estimator and ITE estimator) using the causal inference library (causal inference tool) to obtain the causal effect reasoning result (corresponding to the causal effect evaluation result) and distinguish the direct cause (such as CPU overload) from the root cause (such as unreasonable thread pool configuration) in the chaotic experiment.

[0091] Specifically, the causal inference library and other causal inference tools are used for analysis, and the output causal effect reasoning result can include the causal effect estimation value (e.g. thread pool + 1 CPU utilization rises 2.3%) and its statistical properties, such as confidence interval (reflecting the uncertainty of the estimate), p-value (indicating whether the causal effect is statistically significant), etc. Different effect sizes represent different causal impacts of the treatment variable T on the outcome variable Y, based on which direct and root causes can be located.

[0092] In an optional example, an optimization knowledge base can also be associated to give optimization suggestions (corresponding to target optimization strategies, e.g. "thread pool queue backlog -> suggest dynamically adjusting core thread count") in combination with the causal effect inference results.

[0093] Optionally, after establishing the linear regression model based on the predicted values of each group of causal variables, the method further comprises: determining a plurality of individuals of the target system, wherein the plurality of individuals are used to constitute the target system, and each individual is associated with at least one group of causal variables; and using a causal forest algorithm to evaluate a treatment effect of each individual, wherein the treatment effect of the individual represents a difference between a first type of treatment in which the individual receives a target treatment variable and a second type of treatment in which the individual does not receive the target treatment variable, and the target treatment variable is a treatment variable associated with the individual.

[0094] In this embodiment, a plurality of individuals in the target system can also be obtained, wherein each virtual machine or service instance in the target system can be an individual, and all individuals constitute a population (i.e. the target system). Then, a heterogeneity analysis can be performed on each individual. Specifically, an individual treatment effect (ITE) can be estimated by a causal forest algorithm to identify the effect difference under different individuals.

[0095] It should be noted that both ATE and ITE describe the impact of the treatment variable on the outcome variable, but focus on different levels:

[0096] The individual treatment effect ITE refers to the difference between the potential outcomes of Y when an individual i receives treatment T = 1 (corresponding to the first type of treatment) and does not receive treatment T = 0 (corresponding to the second type of treatment), which represents the true causal impact of T on individual i. However, the individual level cannot be directly observed, so it can only be estimated by a model.

[0097] The average treatment effect ATE refers to the mathematical expectation (or average value) of ITE on the entire population. It represents the causal impact of treatment T on the overall average. ATE is at the population level.

[0098] For example, suppose we evaluate the impact of "increasing the thread pool size from 50 to 100" on "CPU utilization". The ATE estimate value θ = +5% indicates that, on average, the expansion will increase CPU by 5%. Based on this, a heterogeneity analysis using a causal forest can be performed to obtain the following conclusions:

[0099] On servers with QPS < 500, ITE ≈ +1% (very little impact).

[0100] On servers with 500 ≤ QPS < 1000, ITE ≈ +4%.

[0101] On servers with QPS ≥ 1000, ITE ≈ +12% (huge impact).

[0102] It can also be known that the effect is greater on machines with sufficient memory and smaller on old models.

[0103] If only ATE is considered, it may be concluded that expansion increases CPU pressure. However, in combination with ITE, the following strategy can be developed: boldly expand high-QPS core services (the benefits far outweigh CPU costs), and do not need to adjust low-QPS edge services. This is much more efficient than the "one-size-fits-all" decision based on average effects.

[0104] In the embodiment, causal reasoning is used to distinguish between direct causes (such as CPU overload) and root causes (such as unreasonable thread pool configuration) in chaotic experiments. This can reduce false positives and does not require manual annotation, which can improve adaptability. The association optimization knowledge base (such as "thread pool queue backlog → suggest dynamically adjusting the number of core threads") can obtain executable optimization suggestions (corresponding to target optimization strategies) to improve the actual adoption rate and improve the optimization effect of the system.

[0105] Embodiment Two

[0106] The embodiment two of the present application provides a kind of optional root cause analysis method based on chaotic experiment, Figure 2 It is according to the flow chart of optional root cause analysis based on chaotic experiment of the embodiment of the present application, as Figure 2 Shown, comprising:

[0107] Step 1: obtain multi-source time series data in chaotic experiment process, and pre-process multi-source time series data.

[0108] Data acquisition is the first step of chaotic experiment process, in this stage, can be according to research object and experimental data, and through experimental equipment obtains experimental data. Common experimental equipment can include sensor, data logger and the like.

[0109] The experimental data can be multi-source time series data, chaotic experimental observation indicators, system configuration parameters, etc. Chaotic experimental observation data includes but is not limited to: CPU / memory / IO, request delay, error rate, etc. System configuration parameters include but are not limited to: thread pool size, connection pool upper limit, K8s resource quota, etc. In order to ensure the integrity and accuracy of the data, the experimental error can be reduced by repeating the experiment multiple times.

[0110] Data preprocessing is an important part of data analysis, which can include but is not limited to: data alignment, noise removal, missing value processing, data standardization, etc. Data alignment can use dynamic time warping (DTW) algorithm to align multi-source time series data and achieve data alignment. Noise removal can use filtering, smoothing, etc. to improve data quality. Missing value processing can delete records containing missing values, use multiple imputation (MICE) algorithm to fill missing values, etc. Data standardization can eliminate differences between different dimensions. Through data preprocessing, a high-quality data foundation can be provided.

[0111] Optionally, feature extraction can also be performed on the preprocessed data to convert multi-source time series data into a more easily analyzed form, with the goal of extracting key variables that reflect the characteristics of the data (corresponding to the multiple variables in Example One).

[0112] Step 2: The variables in the multi-source time series data after preprocessing can also be used as nodes, and the causal relationships defined based on domain knowledge or automatic discovery (such as PC algorithm) can be used as edges. An improved FCI algorithm (corresponding to the causal inference strategy in Example One) can be used to construct a causal graph.

[0113] Step 2.1: First, based on domain knowledge or automatic discovery (such as PC algorithm), the causal dependence relationship between variables is preliminarily identified, and a candidate edge set is generated (corresponding to the dependence relationship set in Example One);

[0114] Domain knowledge can include an extensible constraint rule library, including two types of prior knowledge: hard constraints: irreversible causal relationships determined by system architecture (such as "thread pool size → CPU utilization"); soft constraints: trusted causal relationships based on historical operation and maintenance experience (such as "database connection number → request delay" with a confidence of 0.8).

[0115] In Step 2.1, an automatic discovery algorithm can be used in combination with domain knowledge for iterative processing. During the algorithm iteration process, the confidence score of the candidate causal edge is corrected through Bayesian posterior probability. By optimizing the algorithm and integrating domain knowledge, it is beneficial to distinguish between direct causes and root causes, effectively reducing the misjudgment rate, and meeting the system risk analysis requirements.

[0116] Step 2.2: Based on the nodes and candidate edge set, use the improved FCI algorithm to generate an initial causal skeleton using conditional independence tests, identify the temporal causal relationship of nodes in the initial causal skeleton through Granger causality test, and then apply the V-structure rule (describing the conditional dependence relationship between variables) to orient the potential causal edges;

[0117] Step 2.3: Then inject domain knowledge constraints to handle potential hidden variables and feedback loops to obtain verified causal relationships (edges);

[0118] Step 2.4: Finally, integrate the nodes and verified causal relationships (edges) to form a directed acyclic graph (DAG) or an extended causal graph containing hidden variables.

[0119] The confidence and main contributing factors of each edge in the causal relationship graph can be annotated to improve interpretability.

[0120] This step 2 is based on the improved FCI algorithm, which can introduce three-stage optimization on the basis of the traditional PC algorithm:

[0121] Stage 1: Based on conditional independence tests (G 2 tests) to generate an initial causal skeleton, set the significance threshold α = 0.01 to reduce false connections.

[0122] Stage 2: Identify temporal causal relationships through Granger causality tests, set the maximum lag length L = 5 (determined by spectral analysis).

[0123] Stage 3: Apply the V-structure rule to orient potential causal edges and solve the ambiguity problem of Markov equivalence classes.

[0124] Step 3: Apply dual machine learning to quantify causal effects and obtain causal effect reasoning results. Dual machine learning eliminates confounding variable bias through phased modeling, enabling causal effect estimation in high-dimensional scenarios, effectively improving estimation accuracy and meeting the system's need for refined attribution.

[0125] Specifically, in this step 3, first perform variable division: process variables T (such as thread pool size), result variables Y (such as CPU utilization), and confounding variables X (such as request QPS, memory occupancy); then through causal graph reasoning, identify and verify the influence path of confounding variables on T and Y.

[0126] Where, T is the control variable, Y is the performance indicator, X is the confounding variable, W is the covariate.

[0127] The implementation process of the dual machine learning framework is as follows:

[0128] 3-1. Data division and model training:

[0129] Sample division: K-fold cross-validation (K=5) is used to divide the data into training set (80%) and auxiliary set (20%) to avoid overfitting.

[0130] First stage modeling:

[0131] Result variable model: Gradient boosting tree (such as XGBoost) is used to predict the result variable Y (such as CPU utilization), with confounding variables X (request volume, memory usage, etc.) and auxiliary variables W (environmental temperature, network status, etc.) as input.

[0132] Process variable model: Logistic regression or random forest is used to predict the process variable T (such as thread pool configuration adjustment), with the same group of variables X and W as input.

[0133] Residual calculation: Subtract the model prediction value from the original data to get the de-biased residual and Y, T are original data, Y pred and T pred are predicted values.

[0134] The result variable model can be used to obtain the predicted value of Y, and the process model can be used to obtain the predicted value of T. Through the residual calculation of the two predicted values, two residuals are obtained to control the pure fluctuation after the influence of X and W, which is equivalent to filtering the influence of X and W. The remaining is the change brought by T and Y itself. The significance of this step is to more accurately estimate the true causal effect of the process variable T on the result variable Y, mainly to serve the second stage.

[0135] Combined with the second stage, the following example is given:

[0136] Assume that the data given to the training model in the first stage is:

[0137] Taking the result variable Y as the CPU utilization, the original value Y observed by the result variable model is 70%, and the predicted value T pred is 60%, then the residual Since the input value takes into account X and W, the extra 10% is considered to be caused by T or other unobserved factors.

[0138] Taking the processing variable T as the thread pool size, the actual observed value T of the processing variable model is 80, and the predicted value T pred is 50, so the residual As above, the extra 30 is considered to be part of the artificial adjustment or abnormal configuration in the thread pool independent of X and W.

[0139] 3-2. Causal Effect Estimation:

[0140] Second-stage regression: Establish a linear regression model where θ is the regression coefficient, i.e. the estimator of the average treatment effect (ATE), and ε is a random variable. (The regression model here can reflect the correlation between T and Y, rather than X and W)

[0141] Heterogeneity analysis: Estimate the individual treatment effect (ITE) through the causal forest algorithm (Causal Forest) to identify the effect differences under different individuals.

[0142] Both ATE and ITE describe the impact of the treatment variable on the outcome variable, but they focus on different levels:

[0143] The individual treatment effect ITE refers to the difference between the potential outcomes of Y when an individual i receives treatment T = 1 and does not receive treatment T = 0, representing the true causal effect of T on individual i. However, individual level cannot be directly observed, so it can only be estimated through the model.

[0144] The average treatment effect ATE refers to the mathematical expectation (or average value) of ITE in the entire population. It represents the causal impact of treatment T on the overall average. ATE is at the population level.

[0145] For example, suppose we evaluate the impact of "increasing the thread pool size from 50 to 100" on "CPU utilization". Through the ATE estimate θ = +5%, it means that on average, the expansion will increase the CPU by 5%. On this basis, the heterogeneity analysis using the causal forest can draw the following conclusions:

[0146] On servers with QPS < 500, ITE ≈ +1% (very small impact).

[0147] On servers with 500 ≤ QPS < 1000, ITE ≈ +4%.

[0148] On servers with QPS ≥ 1000, ITE ≈ +12% (huge impact).

[0149] It can also be found that the effect is greater on machines with sufficient memory and smaller on old models.

[0150] If only looking at ATE, it may be concluded that expansion increases CPU pressure. But combined with ITE, the following strategy can be developed: boldly expand high-QPS core services (the benefits far outweigh the CPU cost), and there is no need to adjust low-QPS edge services. This is much more efficient than the "one-size-fits-all" decision based on average effects.

[0151] 3-3. Analyze the preliminary causal effect estimates (including the ATE and ITE estimates in 3-2) using causal inference tools to obtain causal effect reasoning results, and distinguish between direct causes (such as CPU overload) and root causes (such as unreasonable thread pool configuration) in the chaos experiment.

[0152] Using causal inference tools for analysis, the causal effect reasoning results output can include causal effect estimates such as thread pool + 1 → CPU utilization up 2.3% and their statistical properties, such as confidence intervals (reflecting the uncertainty of the estimate), p-values (indicating whether the causal effect is statistically significant), etc. Different effect measures represent different causal effects of processing variables T on outcome variables Y, based on which direct and root causes can be located.

[0153] Step 4: Correlate the optimization knowledge base and combine the causal effect reasoning results to give optimization suggestions (e.g., "thread pool queue backlog → suggest dynamically adjusting core thread count").

[0154] Specifically, in this step 4, the following steps are included:

[0155] 4-1. Receive causal effect estimates and their statistical properties (such as confidence intervals, p-values, etc.) from the causal inference tool output. These effect measures represent different causal effects of processing variables T on outcome variables Y.

[0156] 4-2. Dynamic weight allocation model, calculate the weights of each suggestion in the knowledge base (use knowledge entry matching degree calculation), filter the candidate set according to the calculation results, for example, the processing variable is the "atomic operation unit" of the suggestion, such as suggestion = processing variable + adjustment direction / target value + scope of action, for example: increase the thread pool of the order service from 50 to 80, one suggestion can be related to multiple processing variables T;

[0157] Based on the causal effect estimates and confidence intervals [θ low ,θhigh ], the matching degree of each proposed knowledge item is calculated as the weight Wi of each proposal:

[0158]

[0159] represents the causal effect estimate of the jth treatment variable on the outcome variable. It is calculated by causal inference methods (such as DoWhy);

[0160] e ij : is the effect enhancement factor of the jth treatment variable related to the ith proposal. It reflects the strength of the effect of the treatment variable under the specific proposal.

[0161] causal effect estimate standard deviation, indicating the uncertainty of the estimate. The smaller the standard deviation, the more reliable the estimate.

[0162] CI j : confidence level of the causal effect estimate . The closer CI j to 1, the higher the accuracy of the estimate.

[0163] γ: adjustment parameter (default 0.1), controlling the degree of influence of confidence level on weight. n: number of treatment variables.

[0164] 4-3. Define a multi-objective combinatorial optimization algorithm (e.g. objectives include maximizing overall causal effect, minimizing implementation cost, reducing risk, etc.), solve the algorithm using branch and bound method, and output the preliminary optimization scheme.

[0165] In this step 4-3, after defining the multi-objective combinatorial optimization algorithm, a Pareto frontier optimization problem can be constructed based on the multi-objective combinatorial optimization algorithm, and then the Pareto frontier optimization problem is solved using the branch and bound method, and the preliminary optimization scheme is output.

[0166] 4-4. Verify the causal path of the optimized scheme, check whether the scheme meets the causal logic of the system, and output the final executable scheme list according to the scheme, including implementation sequence, expected effect and key points. The implementation sequence can be dynamically adjusted according to the causal dependence relationship and priority.

[0167] In this step, the causal effect estimate and resource constraints are included in the same optimization framework, which can break through the local optimal limit caused by traditional rules, and innovatively integrate the causal effect confidence interval into the weight calculation, which can automatically recommend conservative schemes in high uncertainty scenarios. In addition, by associating the optimization database, we can get executable optimization recommendations and improve the actual adoption rate.

[0168] In the embodiment, the direct cause (such as CPU overload) and the root cause (such as unreasonable thread pool configuration) in the chaotic experiment are distinguished by the causal reasoning. Misjudgment can be reduced, manual annotation is not required, adaptability can be improved. The correlation optimization knowledge base (such as 'thread pool queue backlog -> suggest dynamic adjustment of core thread number') is obtained, and the executable optimization suggestion (corresponding to the target optimization strategy) is obtained, the actual adoption rate is improved, and the optimization effect of the system is improved.

[0169] Embodiment three

[0170] The embodiment three of the application provides a kind of optional root cause analysis device based on chaotic experiment, each implementation unit in the root cause analysis device corresponds to each implementation step in embodiment one.

[0171] Figure 3 It is according to the schematic diagram of an optional root cause analysis device based on chaotic experiment of the embodiment of the application, as shown in Figure 3 First acquisition unit 31, construction unit 32 and processing unit 33 are included.

[0172] Wherein, the first acquisition unit 31 is used to obtain the experimental data generated in the chaotic experiment process, to obtain multi-source time series data, wherein the chaotic experiment is used to adopt the way of simulating and manufacturing fault and instability factor in target system, to verify the stability of target system;

[0173] The construction unit 32 is used to construct causal diagram based on multi-source time series data, wherein the causal diagram includes: node representing a plurality of variables and edge representing causal relationship between a plurality of variables, a plurality of variables include: variable representing data characteristics of multi-source time series data;

[0174] The processing unit 33 is used to quantize the variables in the causal diagram by double machine learning model, to obtain causal effect evaluation result, wherein the causal effect evaluation result includes: causal effect estimate value of a plurality of groups of causal variables, each group of causal variables includes: result variable and processing variable leading to result variable, and causal effect estimate value of each group of causal variables is used to represent the influence degree of different processing variables on result variable.

[0175] In the root cause analysis device based on the chaos experiment provided in the embodiment, the experimental data generated in the chaos experiment process can be obtained by the first acquisition unit 31, and multi-source time series data is obtained, wherein the chaos experiment is used to verify the stability of the target system by adopting a way of simulating and manufacturing faults and unstable factors in the target system. The causal graph is constructed based on the multi-source time series data by the construction unit 32, wherein the causal graph includes nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables include variables representing data characteristics of the multi-source time series data. The causal effect quantification of the variables in the causal graph is performed by the processing unit 33 using a double-machine learning model to obtain a causal effect evaluation result, wherein the causal effect evaluation result includes causal effect estimation values of a plurality of causal variables, each group of causal variables includes a result variable and a processing variable causing the result variable, and the causal effect estimation value of each group of causal variables is used to represent the influence degree of different processing variables on the result variable. Furthermore, the technical problem of low accuracy of root cause analysis result in the related art in the chaos experiment scene based on the dynamic risk analysis system of the graph neural network for causal reasoning is solved. In the embodiment, the causal graph is constructed, the causal effect evaluation is performed using a double-machine learning algorithm, and the causal effect evaluation result is obtained, which avoids the situation of low accuracy of root cause analysis result in the related art in the dynamic risk analysis system based on the graph neural network for causal reasoning. The purpose of accurately distinguishing the direct cause and the root cause in the chaos experiment is achieved, thereby realizing the technical effect of improving the accuracy of the causal analysis result.

[0176] Optionally, in the root cause analysis device based on the chaos experiment provided in the embodiment, the causal effect evaluation result further includes a confidence interval of each group of causal variables, the confidence interval of each group of causal variables is used to represent the uncertainty degree of the causal effect estimation value of the group of causal variables, and the root cause analysis device based on the chaos experiment includes: a second acquisition unit, configured to acquire suggestions for optimizing the target system after the causal effect quantification of the variables in the causal graph is performed using the double-machine learning model to obtain the causal effect evaluation result, and obtain a plurality of optimization suggestions, wherein each optimization suggestion is associated with at least one group of causal variables; a first determination unit, configured to determine the weight of each optimization suggestion based on the causal effect estimation value and the confidence interval of the at least one group of causal variables associated with each optimization suggestion; and a second determination unit, configured to determine a target optimization strategy based on the weight of each optimization suggestion, wherein the target optimization strategy is used to optimize the target system.

[0177] Optionally, in the root cause analysis device based on chaos experiment provided in the embodiment, the constructing unit comprises: a first processing subunit, configured to preprocess the multi-source time series data to obtain processed multi-source time series data, wherein the processing manner of preprocessing comprises at least one of the following: data alignment, noise removal, missing value processing, and data standardization; an extraction subunit, configured to extract features from the processed multi-source time series data to obtain a plurality of variables; and a constructing subunit, configured to identify a dependency relationship between the plurality of variables to obtain a dependency relationship set, and construct a causal diagram based on the dependency relationship set and the plurality of variables.

[0178] Optionally, in the root cause analysis device based on chaos experiment provided in the embodiment, the second determining unit comprises: a screening subunit, configured to screen the plurality of optimization suggestions based on the weight of each optimization suggestion to obtain screened optimization suggestions; a verification subunit, configured to determine an initial optimization strategy based on the screened optimization suggestions and the multi-objective combined optimization algorithm, and verify the causal relationship in the initial optimization strategy to obtain a first verification result, wherein the first verification result is used to indicate whether the initial optimization strategy conforms to the causal logic of the target system, and the target of the multi-objective combined optimization algorithm comprises at least one of the following: maximizing causal effect, and minimizing implementation cost; and a first determining subunit, configured to determine a target optimization strategy based on the first verification result.

[0179] Optionally, in the root cause analysis device based on chaos experiment provided in the embodiment, the constructing subunit comprises: a generation module, configured to generate an initial causal skeleton based on the dependency relationship set and the plurality of variables by using a causal inference strategy, wherein the causal inference strategy comprises: conditional independence test, Granger causality verification, and target rule, the target rule is used to direct potential edges representing causal relationships, and the initial causal skeleton comprises: nodes representing the plurality of variables and edges representing causal relationships to be verified; a verification module, configured to verify the causal relationships in the initial causal skeleton to obtain a second verification result; and a processing module, configured to obtain the causal diagram based on the second verification result and the initial causal skeleton.

[0180] Optionally, in the root cause analysis device based on chaos experiment provided in the embodiment, the dual-machine learning model comprises: a result variable model and a processing variable model, the result variable model is used to predict a result variable, and the processing variable model is used to predict a processing variable corresponding to the result variable, the processing unit comprises: a second processing subunit, configured to input variables in the causal diagram into the dual-machine learning model, predict a predicted value of the result variable by using the result variable model, and predict a predicted value of the processing variable corresponding to the result variable by using the processing variable model to obtain a plurality of groups of predicted values of causal variables; a establishing subunit, configured to establish a linear regression model based on the predicted value of each group of causal variables; and a third processing subunit, configured to obtain a causal effect evaluation result based on the linear regression model.

[0181] Optionally, in the chaos experiment-based root cause analysis apparatus provided in the embodiment, the processing unit further includes: a second determination subunit, configured to determine a plurality of individuals of the target system after establishing the linear regression model based on the predicted values of each group of causal variables, wherein the plurality of individuals are used to constitute the target system, and each individual is associated with at least one group of causal variables; and an evaluation subunit, configured to evaluate the processing effect of each individual by using the causal forest algorithm, wherein the processing effect of the individual is used to represent the difference between a first type of processing in which the individual receives the target processing variable and a second type of processing in which the individual does not receive the target processing variable, and the target processing variable is the processing variable associated with the individual.

[0182] The chaos experiment-based root cause analysis apparatus described above can further include a processor and a memory, and the first acquisition unit 31, the construction unit 32, the processing unit 33 and the like described above are stored in the memory as program units, and the corresponding functions are realized by the processor executing the program units stored in the memory.

[0183] The processor described above includes a core, and the core calls the corresponding program units in the memory. The core can be one or more, and the causal graph is constructed by adjusting the core parameters, the causal effect evaluation is performed by using the double machine learning algorithm, the causal effect evaluation result is obtained, the situation that the dynamic risk analysis system based on the graph neural network in the related art performs causal reasoning and the accuracy of the root cause analysis result is low is avoided, the purpose of accurately distinguishing the direct cause and the root cause in the chaos experiment is achieved, and the technical effect of improving the accuracy of the causal analysis result is realized.

[0184] The memory described above can include a non-permanent memory in a computer readable medium, a random access memory (RAM) and / or a non-volatile memory such as a read-only memory (ROM) or a flash memory (flash RAM), and the memory includes at least one memory chip.

[0185] According to another aspect of the embodiment of the present application, an electronic device is further provided, including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the chaos experiment-based root cause analysis method of any one of the above by executing the executable instructions.

[0186] According to another aspect of the embodiment of the present application, a computer readable storage medium is further provided, and the computer readable storage medium stores a computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the chaos experiment-based root cause analysis method of any one of the above when the computer program runs.

[0187] Figure 4 is a schematic diagram of an electronic device according to the embodiment of the present application, as Figure 4As shown, the embodiment of the present application provides an electronic device 40, which comprises a processor, a memory, and a program stored in the memory and executable on the processor, and the processor implements the chaos experiment based root cause analysis method of any one of the above when executing the program.

[0188] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0189] In the above-mentioned embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0190] In several embodiments provided in the present application, it should be understood that the disclosed technical contents can be implemented by other ways. Among them, the above-mentioned device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division way, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or modules shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0191] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed to multiple units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0192] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be realized in the form of hardware or in the form of software functional unit.

[0193] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0194] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for root cause analysis based on chaotic experiments, characterized in that, The method comprises: obtaining experimental data generated in a chaotic experiment, to obtain multi-source time series data, wherein the chaotic experiment is used to verify the stability of a target system by simulating and manufacturing faults and unstable factors in the target system; based on the multi-source time series data, a causal diagram is constructed, wherein the causal diagram includes nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables include variables representing data features of the multi-source time series data; using a double machine learning model to quantify the causal effect of the variables in the causal diagram, to obtain a causal effect evaluation result, wherein the causal effect evaluation result includes a plurality of sets of causal variable causal effect estimation values, each set of causal variables includes a result variable and a processing variable that causes the result variable, and each set of causal variable causal effect estimation values represents the influence degree of different processing variables on the result variable.

2. The root cause analysis method of claim 1, wherein, The causal effect evaluation result also includes a confidence interval for each set of causal variables, which represents the uncertainty of the causal effect estimation value of the set of causal variables. After using a double machine learning model to quantify the causal effect of the variables in the causal diagram to obtain a causal effect evaluation result, it also includes: obtaining suggestions for optimizing the target system to obtain a plurality of optimization suggestions, wherein each optimization suggestion is associated with at least one set of causal variables; determining the weight of each optimization suggestion based on the causal effect estimation value and the confidence interval of at least one set of causal variables associated with each optimization suggestion; determining a target optimization strategy based on the weight of each optimization suggestion, wherein the target optimization strategy is used to optimize the target system.

3. The root cause analysis method of claim 1, wherein, Based on the multi-source time series data, a causal diagram is constructed, comprising: preprocessing the multi-source time series data to obtain processed multi-source time series data, wherein the preprocessing method includes at least one of the following: data alignment, noise removal, missing value processing, and data standardization; extracting features from the processed multi-source time series data to obtain the plurality of variables; identifying the dependency relationship between the plurality of variables to obtain a dependency relationship set, and constructing the causal diagram based on the dependency relationship set and the plurality of variables.

4. The root cause analysis method of claim 2, wherein, Based on the weight of each optimization suggestion, the target optimization strategy is determined, comprising: based on the weight of each optimization suggestion, screening a plurality of optimization suggestions to obtain screened optimization suggestions; based on the screened optimization suggestions and a multi-objective combinatorial optimization algorithm, an initial optimization strategy is determined, and the causal relationship in the initial optimization strategy is verified to obtain a first verification result, wherein the first verification result indicates whether the initial optimization strategy conforms to the causal logic of the target system, and the target of the multi-objective combinatorial optimization algorithm includes at least one of the following: causal effect maximization, implementation cost minimization; based on the first verification result, the target optimization strategy is determined.

5. The root cause analysis method of claim 3, wherein, Based on the dependency relationship set and the plurality of variables, the causal diagram is constructed, comprising: Based on the dependency set and the plurality of variables, an initial causal skeleton is generated by employing a causal inference strategy, wherein the causal inference strategy comprises conditional independence testing, Granger causality checking, and a target rule for orienting potential edges representing causal relationships, and the initial causal skeleton comprises nodes representing the plurality of variables and edges representing causal relationships to be verified; The causal relationships in the initial causal skeleton are verified to obtain a second verification result; Based on the second verification result and the initial causal skeleton, the causal graph is obtained.

6. The root cause analysis method of claim 1, wherein, The double-machine learning model comprises a result variable model for predicting the result variable and a processing variable model for predicting the processing variable corresponding to the result variable. The double-machine learning model is employed to quantify the causal effect of the variables in the causal graph to obtain a causal effect evaluation result, comprising: The variables in the causal graph are input into the double-machine learning model, the predicted value of the result variable is predicted by the result variable model, and the predicted value of the processing variable corresponding to the result variable is predicted by the processing variable model to obtain a plurality of groups of predicted values of the causal variables; Based on the predicted values of each group of causal variables, a linear regression model is established; Based on the linear regression model, the causal effect evaluation result is obtained.

7. The root cause analysis method of claim 6, wherein, After establishing the linear regression model based on the predicted values of each group of causal variables, further comprising: A plurality of individuals of the target system are determined, wherein the plurality of individuals are used to constitute the target system, and each individual is associated with at least one group of causal variables; A causal forest algorithm is employed to evaluate the processing effect of each individual, wherein the processing effect of the individual represents the difference between the first type of processing in which the individual receives a target processing variable and the second type of processing in which the individual does not receive the target processing variable, and the target processing variable is the processing variable associated with the individual.

8. A root cause analysis apparatus based on chaotic experiments, characterized by, Comprising: A first acquisition unit is configured to acquire experimental data generated in a chaotic experiment to obtain multi-source time series data, wherein the chaotic experiment is configured to simulate and manufacture faults and unstable factors in a target system to verify the stability of the target system; A construction unit is configured to construct a causal graph based on the multi-source time series data, wherein the causal graph comprises nodes representing a plurality of variables and edges representing causal relationships between the plurality of variables, and the plurality of variables comprise variables representing data features of the multi-source time series data; A processing unit is configured to employ a double-machine learning model to quantify the causal effect of the variables in the causal graph to obtain a causal effect evaluation result, wherein the causal effect evaluation result comprises a plurality of groups of causal effect estimation values of causal variables, each group of causal variables comprises a result variable and a processing variable leading to the result variable, and the causal effect estimation value of each group of causal variables represents the influence degree of different processing variables on the result variable.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein the computer program controls a device where the computer readable storage medium is located to perform the root cause analysis method based on chaos experiment in any one of claims 1 to 7 when the computer program is running.

10. An electronic device, comprising: The device comprises one or more processors and a memory for storing one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the root cause analysis method based on chaos experiment in any one of claims 1 to 7.

Images