A cloud service fault diagnosis method and device based on granger causality analysis and a medium

Abstract

This invention discloses a cloud service fault diagnosis method, apparatus, and medium based on Granger causality analysis. The method includes: acquiring time-series observation data of all fault variables of the cloud service system within a fault-related time period; using the current observation data of any one of the fault variables as the output variable of a feedforward-gated causal network model, and using the time-series observation data of the remaining fault variables as the input variables of the feedforward-gated causal network model to obtain the weight values ​​of each neuron in the feedforward-gated causal network model; inputting the weight values ​​into an improved nonlinear Granger causal calculation model to output an estimated value of the impact of the output variable on each of the input variables; and based on the estimated impact value, determining the root cause variable of the fault among the fault variables. This invention improves the interpretability and accuracy of cloud service fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of cloud service fault diagnosis technology, and in particular to a cloud service fault diagnosis method, apparatus and medium based on Granger causality analysis. Background Technology

[0002] With the rapid development of information technology, cloud services have become a crucial infrastructure for driving enterprise digital transformation. As an internet-based computing model, cloud services dynamically provide users with scalable virtualized resources and services via the network. However, cloud service systems are characterized by high distribution, strong coupling, and multi-tenant sharing, resulting in complex system structures, numerous variables, and a distinctly dynamic evolutionary nature in their operation. Once a system anomaly or failure occurs, it often creates a cascading propagation effect across multiple service components, leading to data loss, service interruptions, or performance degradation, thereby severely impacting the operational stability and user experience of enterprises. Therefore, accurately identifying fault propagation paths and locating the root cause of faults in a complex environment with intertwined variables is a key issue in ensuring the continuous and stable operation of cloud service systems.

[0003] Existing machine learning-based cloud service fault root cause diagnosis technologies construct fault prediction and diagnosis models by training on historical fault data. For example, attention mechanisms or deep neural networks are introduced to model system operating indicators to predict and identify fault occurrences. While this method has certain advantages in prediction accuracy, its model structure is usually quite complex, and the internal decision-making process lacks clear causal explanation capabilities. This makes it difficult to clearly characterize the causal relationships between variables and the sequence of fault evolution, thus blurring the determination of the root cause of the fault and the order of its causal relationships, making it difficult to accurately identify the root cause fault. Summary of the Invention

[0004] The purpose of this invention is to propose a cloud service fault diagnosis method, device, and medium based on Granger causality analysis. It models fault variables by constructing a feedforward-gated causal network model and calculates the first-order Jacobian matrix of the output variable relative to the input variable based on an improved nonlinear Granger causal calculation model to quantify the causal relationship between fault variables. Furthermore, it performs significance analysis on the causal relationship and constructs a causal relationship matrix between variables for causal analysis, thereby achieving accurate identification of the root cause variables of the fault and improving the interpretability and accuracy of cloud service fault diagnosis.

[0005] To achieve the above objectives, a first aspect of the present invention provides a cloud service fault diagnosis method based on Granger causality analysis, the method comprising: S1: Obtain time-series observation data of all fault variables of the cloud service system during the fault-related time period; S2: Using the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and using the time-series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, the weight values ​​of each neuron in the feedforward gated causal network model are obtained. S3: Input the weight values ​​into the improved nonlinear Granger causal calculation model, and output the estimated value of the impact of the output variable on each of the input variables, wherein the estimated value of the impact is used to characterize the degree of contribution of the input variable to the fault evolution of the output variable; S4: For each of the fault variables, its observed value at the current time is used as the output variable and input into the feedforward gated causal network model. Steps S2 and S3 are executed respectively to obtain the estimated value of the influence between the fault variables. S5: Based on the estimated impact value, determine the root cause variable of the fault among the several fault variables.

[0006] This embodiment uses a feedforward-gated causal network model and an improved nonlinear Granger causal calculation model to estimate the influence between various fault variables, so as to accurately screen root cause propagation faults.

[0007] Furthermore, the acquisition of time-series observation data of all fault variables of the cloud service system within the fault-related time period includes: The observation data of any one of the fault variables at the current time is determined as the output variable of the feedforward gated causal network model, wherein the output variable serves as a supervision signal to constrain the predicted value output by the output layer. The time-series observation data of the remaining fault variables are used as the input variables of the feedforward gated causal network model and input into the input layer. The input variables are subjected to feature encoding processing, and an enhanced input vector is output. The enhanced input vector is input into the hidden layer, and feature transformation is performed on the enhanced input vector to output a hidden feature vector. The hidden feature vector is input into the gating layer, and the hidden feature vector is weighted by the gating unit to output the gating enhanced feature vector; The gated enhanced feature vector is input into the fusion layer, and a nonlinear mapping is performed on the gated enhanced feature vector to output the fused feature vector; The fused feature vector is input into the output layer to obtain the predicted value of the output variable; Based on the error between the predicted value of the output variable and the supervision signal, the weight parameters of each neuron in the feedforward gated causal network model are iteratively updated through a parameter update mechanism. When the weight parameters meet the preset convergence conditions, the final value of the weight parameters of the fusion layer is the weight value.

[0008] In this embodiment, the hidden features are weighted by a gating layer, which enables the network to suppress input features that are irrelevant or weakly correlated with the current output variable during training, thereby reducing the interference of redundant variables on weight learning. The fusion layer is composed of a fully connected network and serves as a nonlinear causal modeling layer. It receives the gated enhanced feature vector output by the gating layer and obtains the causal relationship between variables through nonlinear mapping. The weight parameters that the network model finally converges to provide a reliable data foundation for the subsequent calculation of influence estimates.

[0009] Further, the estimated impact of the output variable on the input variable is output through the following steps: The weight values ​​are input into the improved nonlinear Granger causal computation model, and the first-order Jacobian matrix of the output variable relative to the input variable is calculated based on the improved nonlinear Granger causal computation model. Based on the values ​​of each column element of the first-order Jacobian matrix, the estimated influence value of the input variable associated with the current column is calculated respectively. The improved nonlinear Granger causal calculation model introduces a time lag term into the input variables.

[0010] This embodiment calculates the first-order Jacobian matrix based on an improved nonlinear Granger causal computation model, and uses each column of the Jacobian matrix as an estimate of the influence of the corresponding input variable on the output variable, thereby converting the weight parameters obtained from training the gated causal network model into quantifiable inter-variable influence indicators; at the same time, a time lag term is introduced into the input variable to improve the modeling ability of the time delay effect in the fault propagation link and improve the accuracy of the influence estimate.

[0011] Further, the step of determining the root cause variable of the fault among the plurality of fault variables based on the impact estimate includes: Obtain a set of estimated impact values ​​for each of the fault variables within several fault-related time periods; Based on each set of estimated impact values, determine whether the fault variable is a significant causal variable. If so, the fault variable is marked as a first preset value; otherwise, the fault variable is marked as a second preset value. Based on the first preset value and the second preset value, a causal analysis is performed on all the fault variables to determine the root cause variables of the fault.

[0012] This embodiment performs statistical calculations on the estimated impact values ​​of each fault variable formed in multiple fault-related time periods, and determines whether the fault variable is a significant causal variable, thereby reducing the randomness brought about by a single estimation judgment. Furthermore, based on the determination of significant causal impact, the causal relationship between fault-related variables is obtained, and the root cause fault is accurately screened.

[0013] Furthermore, obtaining the set of impact estimates for each of the fault variables within several fault-related time periods includes: Calculate the estimated impact values ​​between the fault variables within several fault-related time periods; Based on all the estimated impact values ​​of each of the fault variables, a set of estimated impact values ​​for each fault variable is obtained.

[0014] This embodiment calculates the estimated impact values ​​between variables and aggregates them into a set of impact estimates by variable. This provides a reusable and comparable basic data structure for subsequent significance judgment and root cause screening, avoiding the omission of root causes caused by relying solely on local judgments of single variables.

[0015] Further, the step of determining whether the fault variable is a significant causal variable based on each of the respective sets of impact estimates includes: Based on each set of impact estimates, the empirical null distribution of the fault variable is calculated accordingly. If the empirical null distribution of the fault variable does not follow a Gaussian distribution, then the fault variable is determined to be a significant causal variable.

[0016] This embodiment uses empirical zero distribution to statistically discriminate the impact estimate, removes spurious fault variables caused by noise, thereby reducing erroneous variables introduced by spurious causal relationships and improving the accuracy of subsequent root cause fault location.

[0017] Further, the step of performing causal analysis on all the fault variables based on the first preset value and the second preset value to determine the root cause variable of the fault includes: A causal relationship matrix is ​​constructed based on each of the aforementioned fault variables, and the matrix elements of the causal relationship matrix are used to characterize the causal relationship between each of the aforementioned fault variables; If the fault variable is marked as the first preset value, then the corresponding matrix element in the causal relationship matrix is ​​assigned the first preset value; otherwise, the matrix element is assigned the second preset value. Based on the causal relationship matrix, the root cause variables of the fault are selected.

[0018] This embodiment constructs a causal relationship matrix between variables based on a first preset value and a second value, and expresses the variables with significant causal influence in the matrix in a unified symbolic way, thereby improving the effectiveness of screening root cause faults.

[0019] To achieve the above objectives, a second aspect of the present invention further provides a cloud service fault diagnosis device based on Granger causality analysis, used to implement the cloud service fault diagnosis method based on Granger causality analysis as described in any of the first aspects above, the device comprising: The data acquisition module is used to acquire time-series observation data of all fault variables of the cloud service system during the fault-related time period; The model training module is used to take the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and take the time series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, and input them into the feedforward gated causal network model to obtain the weight values ​​of each neuron in the feedforward gated causal network model. The impact estimation module is used to input the weight values ​​into an improved nonlinear Granger causal calculation model and output an estimated value of the impact of the output variable on each of the input variables, wherein the estimated value of the impact is used to characterize the degree of contribution of the input variable to the failure evolution of the output variable; The root cause determination module is used to determine the root cause variable among the several fault variables based on the estimated impact value.

[0020] A third aspect of the present invention also provides a computer-readable storage medium comprising a stored computer program; wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a cloud service fault diagnosis method based on Granger causality analysis as described in any of the first aspects. Attached Figure Description

[0021] Figure 1 This is a flowchart of a preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of the present invention; Figure 2 This is a schematic diagram of a feedforward gated causal network structure, representing another preferred embodiment of a cloud service fault diagnosis method based on Granger causal analysis provided in the first aspect of the present invention. Figure 3 This is a schematic diagram of a significantly unrelated distribution in another preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of the present invention; Figure 4This is a schematic diagram of the significant correlation distribution of another preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of the present invention; Figure 5 This is a schematic diagram of the causal relationship matrix of another preferred embodiment of a cloud service fault diagnosis method based on Granger causal analysis provided in the first aspect of the present invention; Figure 6 This is a structural block diagram of a preferred embodiment of a cloud service fault diagnosis device based on Granger causality analysis provided in the second aspect of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0023] It should be noted that the data involved in this invention (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0024] In this embodiment of the invention, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplarily" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0025] In this invention description, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In this invention description, unless otherwise stated, "a plurality of" means two or more. In this invention description, the term "comprising" and its variations are open-ended, meaning "including but not limited to." The term "based on" means "at least partially based on." The term "according to" means "at least partially according to." The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments."

[0026] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0027] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] The Granger causality analysis method and its basic principles involved in the embodiments of the present invention are explained below: Granger causality analysis is a causal inference method based on time series analysis. Originally proposed by Clive Granger in 1969 as a statistical method in economics, it is fundamentally based on Wiener's definition of causality: only the past can cause the present and future; future events do not cause past or present events, and the future does not cause the past or present. Extending this to time series, Granger causality describes predictability. Given two time series, X and Y, if the combined history of X and Y is more predictive of the future value of Y than the history of Y alone, then a Granger causal relationship exists between variables X and Y. Variable X is called the Granger cause of Y, and variable Y is the Granger effect of X. Granger causality analysis determines Granger causal relationships between variables by establishing an autoregressive model. The classic Granger causality analysis model is as follows: (1) (2) Among them, model (1) is the Granger baseline model, and model (2) is the Granger extended model. Represents zero-mean noise and Representing variables respectively and The value at time t, and This represents the historical data of the corresponding variable from time 1 to time t-1. This represents the fitted function. If... Under the same circumstances, The addition of The predictions are helpful; the prediction accuracy of the extended model is significantly higher than that of the baseline model. Therefore, it is believed that... and There exists a Granger causal relationship between them, which is the definition of Granger causality.

[0029] Classical Granger causality analysis models can only effectively identify Granger causal relationships between linear data. However, in cloud service scenarios, server monitoring data (such as CPU, memory, network traffic, etc.) generally exhibit complex nonlinear relationships, and their potential fitting functions... Accurate modeling of Granger causality using simple nonlinear functions is difficult. Therefore, improvements to the Granger causality analysis model are needed. For nonlinear autoregressive models, variables are allowed to evolve according to nonlinear dynamics. Therefore, the Granger causality problem can be explained from a functional perspective, allowing the target variable to... Evolutionary modeling is used for all other variables in the system. nonlinear functions Its formula is:

[0030] in, Indicates zero-mean noise. For variables from Time until a certain maximum lag Historical data. Based on variables. The evolutionary formula, Granger noncausality, can be defined as follows: hypothetical variables... for and ( Indicating a hypothesis Another unequal value) satisfies the following formula:

[0031] Right now For any value The values ​​are all unaffected, so for The constants of the evolution formula, Not the target variable Granger causality variables. Based on this, considering the impact of time lag between actual cloud service data, variables can be defined with time lag. Under Granger causality, the formula changes to:

[0032] If the formula applies to all And any If all are true, then it indicates that the lagged terms are true. For target variable The evolution of [variable name] has no substantial impact and is considered a redundant variable, i.e., it exists in time delay. Below, variables and There is no Granger causality. Therefore, determining which variables are important... The core issue of how evolution truly impacts the world can be equivalently transformed into understanding the set of all lagged terms. In the middle, find one that can fully model The minimum possible subset of variables is defined as follows: each variable in the subset is necessary for modeling and there is no redundancy. Variables in the minimum possible subset are considered Granger causal variables, while those not in the subset are considered irrelevant variables. Based on this, non-Granger causality between variables is defined, and this formula is subsequently used to determine whether a Granger causal relationship exists between variables.

[0033] The first aspect of this invention provides a cloud service fault diagnosis method based on Granger causality analysis, see [link to relevant documentation]. Figure 1 The diagram shown is a flowchart of a preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided by the first aspect of the present invention. The method includes steps S1 to S5, as follows: S1: Obtain time-series observation data of all fault variables of the cloud service system during the fault-related time period; In one example, the cloud service monitoring and acquisition system continuously collects the operational metrics of the cloud service system, and after detecting an abnormal or fault event, sends the fault variable observation data within the fault-related time period to the diagnostic execution unit on the fault diagnosis side.

[0034] It should be noted that the fault variables can be indicators that characterize the operating status of cloud services, such as CPU utilization, memory usage, disk I / O, network throughput, latency, and packet loss rate. The time-series observation data can be equally spaced sampled data, non-equal spaced sampled data, or data after alignment processing.

[0035] This embodiment provides a data foundation for subsequent causal modeling by uniformly aggregating time-series observation data of multidimensional fault variables within the fault-related time period, thereby reducing the probability of misjudgment caused by missing dependent variables or inconsistent sampling.

[0036] S2: Using the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and using the time-series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, the weight values ​​of each neuron in the feedforward gated causal network model are obtained. In one example, the diagnostic execution unit selects the observation value of any fault variable at the current moment from the time-series observation data obtained in step S1 as the output variable, and uses the time-series observation data of the remaining fault variables as the input variables to initiate a training request to the feedforward gated causal network model in order to obtain the weight parameters after the model converges.

[0037] It should be noted that the current time is the end time within the fault-related time period; the training can be a single training session or multiple rounds of iterative training; the training data can be multi-tenant data or single-tenant data.

[0038] This embodiment uses a feedforward gated causal network to nonlinearly fit the causal relationships of multiple fault variables, enabling the learning of weight parameters for variable associations in complex coupled scenarios, thus providing a parameter basis for subsequent causal impact estimation.

[0039] S3: Input the weight values ​​into the improved nonlinear Granger causal calculation model, and output the estimated value of the impact of the output variable on each of the input variables, wherein the estimated value of the impact is used to characterize the degree of contribution of the input variable to the fault evolution of the output variable; In one example, the diagnostic execution unit sends the weight values ​​obtained in step S2 to the improved nonlinear Granger causal computation model, which outputs an estimate of the impact of the output variable on each input variable.

[0040] It should be noted that the estimated impact value can be understood as the column vector corresponding to the first-order Jacobian matrix of the output variable relative to the input variable or its statistics, or as any quantitative index that can characterize the contribution of the input to the evolution of the output; the improved nonlinear Granger causal calculation model can introduce a time lag term for each input variable to characterize the influence relationship between past inputs and current outputs, but does not limit the specific lag order or specific construction method.

[0041] This embodiment maps network weight parameters to influence estimates, realizing the transformation from black-box fitting to interpretable causal contribution, which facilitates direct comparison of the influence of different input variables on output variables.

[0042] S4: For each of the fault variables, its observed value at the current time is used as the output variable and input into the feedforward gated causal network model. Steps S2 and S3 are executed respectively to obtain the estimated value of the influence between the fault variables. In one example, the diagnostic execution unit sequentially selects each fault variable as the current output variable and repeats steps S2 and S3, summarizing the impact estimates obtained each time to form the impact estimate results between the fault variables.

[0043] It should be noted that this traversal process can be executed sequentially according to the order of variables, or it can be executed in parallel; the influence estimation results can be organized into a matrix, list, graph structure, or any data structure that facilitates subsequent analysis.

[0044] S5: Based on the estimated impact value, determine the root cause variable of the fault among the several fault variables.

[0045] In one example, the diagnostic execution unit performs causal analysis based on the impact estimation results obtained in step S4, and outputs the root cause variables of the fault as diagnostic conclusions, which are then sent to the alarm handling system, the maintenance personnel terminal, or the automatic repair module.

[0046] It should be noted that the determination of root cause variables can be based on significance testing, threshold screening, matrix structure analysis, propagation direction analysis, or any combination thereof; there is no specific limitation on the specific rules for determining variables that have a significant impact on other variables but are not significantly affected by other variables as root cause variables.

[0047] This embodiment uses the estimated value of the influence between variables to determine the root cause, which can accurately locate the root cause of the fault that triggers the propagation of the fault.

[0048] In another preferred embodiment, obtaining time-series observation data of all fault variables of the cloud service system within the fault-related time period includes: The fault records within the relevant time period are obtained from the cloud service monitoring and collection system. The fault observation records include at least a time identifier, a fault variable identifier, a fault observation value, and the tenant identifier information. Based on the tenant identification information, the fault observation records are classified to generate a set of fault observation records divided by tenant; In the fault observation record set corresponding to each tenant, the fault observation values ​​are classified according to the fault variable identifier, and the classified fault observation values ​​are sorted in chronological order according to the time identifier to obtain the time series observation data of each fault variable.

[0049] In a specific example, after the monitoring and acquisition system detects an alarm event or a fault window trigger condition, it sends fault observation records for the relevant time period to the diagnostic server. These fault observation records carry at least a time stamp, a fault variable identifier, a fault observation value, and tenant identifier information. Upon receiving these fault observation records, the diagnostic server first classifies them based on the tenant identifier information, thereby generating a set of fault observation records divided by tenant.

[0050] Understandably, this step ensures that subsequent modeling data maintains consistent semantics across the tenant dimension, preventing semantic noise caused by identical metrics from different tenants being mixed into the same training sample.

[0051] It should be noted that the tenant identification information can be either a tenant ID, a business instance identifier, an account identifier, or other information that can distinguish the tenant boundary. This embodiment does not limit this.

[0052] Subsequently, in the fault observation record set corresponding to each tenant, the diagnostic server classifies the fault observations according to the fault variable identifier, and further sorts the classified fault observations in chronological order according to the time identifier, so as to construct the time series observation data of each fault variable.

[0053] It is understood that the time-series observation data can be organized as a vector sequence, a table structure sequence, or a matrix structure.

[0054] Furthermore, the diagnostic server can align multiple fault variables under the same tenant on the same time axis to form a multivariate time series observation matrix, which can then be used as input for the subsequent feedforward gated causal network model.

[0055] It should be noted that the alignment method can be interpolation, resampling, window aggregation, or maintaining the original form of non-equally spaced sequences. This embodiment does not limit the specific implementation.

[0056] Furthermore, in an optional implementation, to better reflect tenant differences during model training, such as differences between tenant SLA semantics and resource usage habits, the diagnostic server can construct a tenant-aware mapping space at the input layer. Specifically, discrete tenant identification information is mapped into a dense vector in a continuous space through an embedding layer, and this dense vector is concatenated with the original indicator matrix as an additional feature to form an enhanced input vector, as shown in the following equation:

[0057] in, This can be understood as the original indicator observation vector or observation matrix under the same tenant. This indicates a splicing operation. This represents an embedding mapping. It should be noted that the dimension of the embedding vector can be 64, 128, or other dimensions; this embodiment does not limit the specific expression form.

[0058] It should be noted that this tenant perception mapping is an optional enhancement method and does not affect the basic process of obtaining time-series observation data by tenant classification and sorting as defined in weight 2.

[0059] After completing the above data preparation, the diagnostic server can feedforward a causal network model with time-series observation data constructed per tenant (and optional tenant-aware augmentation inputs).

[0060] This embodiment introduces tenant identification information and classifies data by tenant during the data acquisition stage, aggregates data by variable within a tenant, and sorts data by time. This ensures that the data entering subsequent causal modeling maintains consistency in tenant semantics and temporal order, reducing the impact estimation bias caused by cross-tenant indicator aliasing and temporal misalignment, thereby improving the stability of impact estimation results and the accuracy of root cause fault location. Furthermore, in the optional implementation, a tenant-aware mapping space is introduced, which can further enhance the model's ability to express the differences in the operating modes of different tenants, thereby reducing the interference of semantic ambiguity of tenant indicators on causal analysis.

[0061] In another preferred embodiment, the feedforward gated causal network model includes an input layer, a hidden layer, a gated layer, a fusion layer, and an output layer. The weight values ​​of each neuron in the feedforward gated causal network model are obtained through the following steps: The observation data of any one of the fault variables at the current time is determined as the output variable of the feedforward gated causal network model, wherein the output variable serves as a supervision signal to constrain the predicted value output by the output layer. The time-series observation data of the remaining fault variables are used as the input variables of the feedforward gated causal network model and input into the input layer. The input variables are subjected to feature encoding processing, and an enhanced input vector is output. The enhanced input vector is input into the hidden layer, and feature transformation is performed on the enhanced input vector to output a hidden feature vector. The hidden feature vector is input into the gating layer, and the hidden feature vector is weighted by the gating unit to output the gating enhanced feature vector; The gated enhanced feature vector is input into the fusion layer, and a nonlinear mapping is performed on the gated enhanced feature vector to output the fused feature vector; The fused feature vector is input into the output layer to obtain the predicted value of the output variable; Based on the error between the predicted value of the output variable and the supervision signal, the weight parameters of each neuron in the feedforward gated causal network model are iteratively updated through a parameter update mechanism. When the weight parameters meet the preset convergence conditions, the final value of the weight parameters of the fusion layer is the weight value.

[0062] In a specific example, see Figure 2The diagram shown is a schematic diagram of a feedforward-gated causal network structure, representing another preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of this invention. The network parameters after convergence of the feedforward-gated causal network model are used for interpretable nonlinear modeling of the target fault variable. It can be understood that the top of the diagram... Indicates the first variable selected as the output variable. The fault variables at the current moment The observed values; the bottom , … This represents the time-series observation data of the remaining fault variables within a historical window prior to the current moment; the tenant ID Embedding in the lower right corner represents the dense vector obtained by embedding the tenant identification information associated with this sample.

[0063] When constructing training samples, the diagnostic side first determines that the output variable is the target fault variable. The current observation value is used as a monitoring signal. Subsequently, historical observation data of the remaining fault variables are used. Construct the input data for the input layer.

[0064] In one specific embodiment, the discrete tenant ID is transformed into a high-dimensional dense vector in a continuous space through an embedding layer. The enhanced input vector is obtained by tensor concatenation of the original index matrix with the control variable. :

[0065] in, (This is the original time series matrix of indicators, such as monitoring metrics like memory usage and CPU usage). This represents a vector concatenation operation. It enables the model to automatically decouple the semantics of metrics across different tenants through end-to-end training, achieving global model sharing and local feature adaptation while maintaining privacy.

[0066] Augmented input vectors After input to the hidden layer, the hidden layer performs feature transformation on it and outputs a hidden feature vector. It should be noted that the number of hidden layers, the number of neurons per layer, and the type of activation function can all be configured according to engineering requirements; this embodiment does not limit the specific structure.

[0067] Hidden Feature Vector After being input into the gating layer, the gating units in the gating layer weight the hidden features and output a gating enhanced feature vector. It should be noted that the calculation method for the gating coefficients is not limited; they can be obtained using sigmoid / softmax / normalization functions, or selective enhancement and suppression can be achieved through thresholding / sparsening. Through this gating mechanism, input variables with weak correlation to the output variable are suppressed, while key input variables are enhanced, thereby improving the accuracy of subsequent causal effect estimation.

[0068] Gated augmented feature vectors After input to the fusion layer, the fusion layer performs nonlinear mapping and fusion across variables, outputting a fused feature vector. Understandably, the fusion layer, composed of a fully connected network, serves as a nonlinear causal modeling layer. It receives weighted features from the gating layer and captures complex causal dependencies between variables through nonlinear transformations, providing a high-order causal representation for the final prediction.

[0069] Fusion feature vectors After the input and output layers, the output layer provides the predicted values ​​of the output variables. Subsequently, based on the predicted values With monitoring signals The errors between the layers, such as squared error, absolute error, or other loss forms, are iteratively updated through a parameter update mechanism to the weight parameters of each layer of the network until the preset convergence condition is met.

[0070] It should be noted that the parameter update mechanism can be backpropagation combined with gradient descent, or optimization strategies such as Adam / RMSProp; this embodiment is not limited to any one of these. The convergence condition can be any one or a combination of the following: the change in the loss function is less than a threshold, the maximum number of iterations is reached, or the validation set index is stable. When the preset convergence condition is met, the final value of the weight parameters of the fusion layer can be output as the weight value for subsequent Jacobian matrix calculation.

[0071] This embodiment introduces a gating mechanism and tenant embedding to enhance feature representation, thereby enabling effective screening of key variables and nonlinear causal modeling, improving the accuracy and stability of causal impact estimation, and providing a reliable parameter basis for fault root cause localization.

[0072] In yet another preferred embodiment, the estimated impact of the output variable on the input variable is output through the following steps: The weight values ​​are input into the improved nonlinear Granger causal computation model, and the first-order Jacobian matrix of the output variable relative to the input variable is calculated based on the improved nonlinear Granger causal computation model. Based on the values ​​of each column element of the first-order Jacobian matrix, the estimated influence value of the input variable associated with the current column is calculated respectively. The improved nonlinear Granger causal calculation model introduces a time lag term into the input variables.

[0073] In a specific example, after the diagnostic side completes the training of the feedforward gated causal network model, the network's weight parameters have converged, thus forming a nonlinear function model that characterizes the evolutionary relationship between the historical states of the input variables and the current state of the output variables. It should be noted that this nonlinear function model is not limited to a specific network implementation; it can be implemented using a feedforward neural network or other models capable of representing nonlinear mapping relationships. This embodiment does not impose any limitations.

[0074] Furthermore, for any fault variable selected as an output variable... and any input variable When a time lag term is introduced for each input variable in the improved nonlinear Granger causality model, the input variables are... The form of participation in modeling, in which This indicates the preset lag order or lag time index. In other words, the improved nonlinear Granger causality model does not only utilize the current value of the input variables, but also treats the historical lag values ​​of each input variable prior to the current time as independent modeling variables.

[0075] Based on this, a certain variable Given time lag Below The contribution of evolution can be written as a first-order Jacobian matrix in terms of the input:

[0076] in, This represents a trained neural network. For judgment and Does a causal relationship exist among the variables? Due to the approximate nature of neural networks, in practical calculations, a trained neural network can be used to calculate the Jacobian matrix estimate of each input variable relative to the output. This estimate is an approximation of the true partial derivative.

[0077] Based on the conclusions derived from the Granger causality method—determining Granger causality between variables can be equivalent to determining causality from the set of all lagged terms— In the middle, find one that can fully model The smallest subset of variables—each variable in the subset is necessary for modeling, eliminating redundant variables. Therefore, when When =0, then for To be honest As redundant variables, this precisely satisfies the definition of noncausality in the Granger causal analysis model.

[0078] This embodiment calculates the Jacobian matrix on the converged model weight parameters and assigns each column of the Jacobian matrix to an estimated value of the input variable's influence, making the causal influence between variables quantifiable as a comparable contribution indicator. At the same time, it explicitly introduces a time lag term for each input variable, accurately modeling the time delay effect commonly seen in fault propagation, thereby improving the accuracy of fault causal link identification.

[0079] In yet another preferred embodiment, determining the root cause variable of the fault among the plurality of fault variables based on the impact estimate includes: Obtain a set of impact estimates for each of the fault variables within several fault-related time periods; Based on each set of estimated impact values, determine whether the fault variable is a significant causal variable. If so, the fault variable is marked as a first preset value; otherwise, the fault variable is marked as a second preset value. Based on the first preset value and the second preset value, a causal analysis is performed on all the fault variables to determine the root cause variables of the fault.

[0080] In a specific example, the root cause fault diagnosis unit aggregates the estimated impact values ​​of each fault variable over several fault-related time periods to obtain a set of impact estimates for that fault variable. It can be understood that this set of impact estimates reflects the statistical distribution characteristics of the fault variable's contribution to other fault variables in different fault time periods.

[0081] Furthermore, the diagnostic side determines whether the corresponding fault variable is a significant causal variable based on each set of impact estimates. It should be noted that the determination of significance can be based on threshold determination, statistical test determination, ranking screening determination, or a combination thereof. For example, it can be based on the mean, variance, quantiles, extreme values, or significance test statistics of the set of impact estimates; it can also be based on the stable occurrence frequency under multiple rounds of training or multiple windows. This embodiment does not limit the specific determination method.

[0082] If a fault variable is determined to be a variable with significant causal influence, then the fault variable is marked as the first preset value; otherwise, it is marked as the second preset value. It is understood that the first preset value / second preset value can be implemented using binary labeling (such as 1 / 0), or it can be implemented using different encoding methods, as long as it can distinguish between the two categories of significant causal influence and non-significant causal influence.

[0083] After the labeling is completed, the diagnostic side performs causal analysis on all fault variables based on the first and second preset values ​​to determine the root cause variables of the fault.

[0084] It is understood that this causal analysis is not limited to a single implementation method. In one implementation method, a causal relationship matrix or set of causal relationships between variables can be constructed based on the labeling results, and the driving relationship between variables can be determined by combining the directional information of the influence estimates. In another implementation method, the role of each variable in the causal relationship can be further analyzed, for example, identifying variables that have a significant impact on other variables but are not significantly affected by other variables as root cause candidates. When the number of candidates is greater than 1, indicators such as influence strength, stability, or frequency of occurrence can be combined for screening to determine the final root cause variable of the failure. It should be noted that the above determination can be achieved by any indicator or combination thereof, such as the number of incoming relationships, the number of outgoing relationships, the intensity of incoming influence, and the intensity of outgoing influence. This embodiment does not limit this.

[0085] This embodiment combines multi-time period convergence, significance determination, and causal analysis to establish the identification of root cause variables based on statistical stability and overall causal structure analysis, thereby improving the accuracy and reliability of root cause localization.

[0086] In yet another preferred embodiment, obtaining the set of impact estimates for each of the fault variables over several fault-related time periods includes: Calculate the estimated impact values ​​between the fault variables within several fault-related time periods; Based on all the estimated impact values ​​of each of the fault variables, a set of estimated impact values ​​for each fault variable is obtained.

[0087] In a specific example, the diagnostic side selects multiple fault-related time periods for repeated modeling for the same fault event or the same type of fault phenomenon. These fault-related time periods can be multiple fault windows on different dates, or multiple sub-windows divided within the same fault window; this embodiment does not impose any limitations. For each fault-related time period, the diagnostic side can execute a training model once based on the observed fault variable data within that time period and calculate the influence estimate, thereby obtaining the influence estimate between each fault variable within that time period.

[0088] Furthermore, the diagnostic side aggregates the impact estimates obtained from multiple fault-related time periods according to the fault variable. For example, for a certain fault variable... In other words, it can be used as a driving variable for other fault variables (such as...) The impact estimates, along with the corresponding impact estimates obtained by repeated calculations at different time periods, are collected and compiled into a unified system. The set of impact estimates; similarly, they can be formed separately. , The set of estimated impact values ​​for each fault variable.

[0089] This embodiment reduces the impact of single training fluctuations, sampling noise, or occasional interference on the judgment results by repeatedly calculating the estimated values ​​of the influence between variables in multiple fault-related time periods and aggregating them by variable. This provides a reliable data foundation for subsequent determination of significant causal influence variables and screening of root cause variables of faults.

[0090] In yet another preferred embodiment, the step of determining whether the fault variable is a significant causal variable based on each of the respective sets of impact estimates includes: Based on each set of impact estimates, the empirical null distribution of the fault variable is calculated accordingly. If the empirical null distribution of the fault variable does not follow a Gaussian distribution, then the fault variable is determined to be a significant causal variable.

[0091] In a specific example, the diagnostic side targets a pair of variables, such as input variables. With output variables 100 were accumulated through multiple training sessions. right The Jacobian matrix estimates were then obtained. Subsequently, the diagnostic side statistically modeled these 100 estimates to obtain an empirical null distribution, which was presented as a histogram. See [link / reference]. Figure 3 The diagram shown is a schematic representation of a significantly uncorrelated distribution, representing another preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of this invention. The horizontal axis represents the range of estimated values, and the vertical axis represents the frequency density of estimated values ​​falling within the corresponding range. Furthermore, a Gaussian curve (as shown by the dashed line in the diagram) can be fitted to the histogram to assess the degree of fit between the empirical null distribution and the Gaussian distribution.

[0092] As can be understood from Figure 3, the estimated value set presents an approximately unimodal and symmetrical distribution, and the fitted Gaussian curve and histogram are in good agreement. This indicates that the estimated impact of this variable on the system is more likely to be mainly composed of noise fluctuations, and its empirical null distribution is closer to the Gaussian distribution. In this case, it can be determined that the causal impact of this variable is not significant.

[0093] In another case, see Figure 4 The figure shows a schematic diagram of the significant correlation of another preferred embodiment of the cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of the present invention. The estimated value set in the figure shows obvious skewness, heavy tails or multimodal characteristics, and the Gaussian fitting curve deviates significantly from the histogram, indicating that the variable repeatedly produces a stable non-zero contribution component in multiple fault time periods, and its empirical null distribution does not follow a Gaussian distribution. Therefore, the variable can be determined to be a significant causal variable.

[0094] It should be noted that the determination of non-Gaussian distribution can be achieved through goodness-of-fit test, quantile difference judgment or other equivalent methods. This embodiment does not limit the specific test method.

[0095] It should be further noted that the training data corresponding to the set of impact estimates is usually taken from observation data at the time of the fault or within the fault-related time period; the length of the training window can be flexibly set according to the scenario, for example, a model can be trained from one day's data or one hour's data. In some simple fault scenarios with short fault propagation links and clear variable coupling relationships, the set of impact estimates trained based on a short time window (such as one hour) can also form a sufficiently stable empirical null distribution, thereby supporting the determination of significant causal influence variables and further locating the root cause of the fault; while in more complex fault scenarios, the reliability of the empirical null distribution estimation can be improved by increasing the number of fault-related time periods or extending the training window length.

[0096] This embodiment constructs an empirical zero distribution on the set of impact estimates and performs Gaussian consistency discrimination, which can distinguish between pseudo-fault variables caused by random noise and real fault variables that appear stably across time periods. This reduces the probability of misjudgment caused by pseudo-causal relationships and provides more reliable data for subsequent root cause variable screening.

[0097] In yet another preferred embodiment, the step of performing causal analysis on all the fault variables based on the first preset value and the second preset value to determine the root cause variable of the fault includes: A causal relationship matrix is ​​constructed based on each of the aforementioned fault variables, and the matrix elements of the causal relationship matrix are used to characterize the causal relationship between each of the aforementioned fault variables; If the fault variable is marked as the first preset value, then the corresponding matrix element in the causal relationship matrix is ​​assigned the first preset value; otherwise, the matrix element is assigned the second preset value. Based on the causal relationship matrix, the root cause variables of the fault are selected.

[0098] In a specific example, it can be understood that after the diagnostic side completes the determination of the significant causal influencing variables and marks the first and / or second preset values, it further expresses the influence relationship between the various fault variables in a structured matrix form, thereby facilitating the overall analysis of the fault propagation direction.

[0099] See Figure 5The diagram shown is a causal relationship matrix diagram of another preferred embodiment of a cloud service fault diagnosis method based on Granger causality analysis provided in the first aspect of the present invention. It consists of three variables: the horizontal axis represents the driving variable (Cause / Driver), and the vertical axis represents the result variable (Effect / Result). Each cell in the matrix corresponds to the determination result of whether the driving variable has a causal influence on the result variable. When the cell takes a first preset value (e.g., 1), it can be understood that the corresponding driving variable has a significant causal influence on the result variable, i.e., the driving variable is a Granger cause of the result variable; when the cell takes a second preset value (e.g., 0), it can be understood that the corresponding driving variable does not have a significant causal influence on the result variable.

[0100] In the row corresponding to the result variable X1, only the cell corresponding to the driving variable X2 has the first preset value, while the cells corresponding to the driving variables X1 and X3 have the second preset value, indicating that X2 affects X1. In the row corresponding to the result variable X2, the cells corresponding to the driving variables X1 and X3 both have the first preset value, indicating that X1 and X3 affect X2. In the row corresponding to the result variable X3, the cell corresponding to the driving variable X1 has the first preset value, indicating that X1 affects X3. Therefore, it can be seen that variable X1 has a causal influence on both X2 and X3 (i.e., the column corresponding to X1 takes the first preset value in multiple result variable rows), and X1 itself is not affected by other variables (i.e., all cells in the row corresponding to X1, except for the self-item, have the second preset value). Therefore, X1, which has a causal influence on other variables but is not causally affected by other variables, can be identified as the root cause variable of the system's fault.

[0101] It should be noted that the first preset value / second preset value can be implemented using 1 / 0, True / False or other equivalent coding methods; the rules for screening root cause variables are not limited to strict conditions that are completely unaffected by other variables. In some scenarios, a small number of weak influence relationships may be allowed, and a comprehensive judgment can be made by combining the influence intensity, frequency of occurrence or stability indicators. This embodiment does not impose any limitations.

[0102] This embodiment presents a structured representation of significant causal relationships between variables in the form of a causal relationship matrix, which facilitates a holistic analysis of the fault propagation path. Furthermore, it rapidly identifies root cause fault variables based on the matrix, thereby improving the intuitiveness and interpretability of fault root cause localization.

[0103] A second aspect of this invention provides a cloud service fault diagnosis device based on Granger causality analysis, used to implement the cloud service fault diagnosis method based on Granger causality analysis described in any of the first aspects above. See also... Figure 6 The diagram shown is a structural block diagram of a preferred embodiment of a cloud service fault diagnosis device based on Granger causality analysis provided in the second aspect of the present invention. The device includes: The data acquisition module 11 is used to acquire time-series observation data of all fault variables of the cloud service system during the fault-related time period; The model training module 12 is used to take the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and take the time series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, and input them into the feedforward gated causal network model to obtain the weight values ​​of each neuron in the feedforward gated causal network model. The impact estimation calculation module 13 is used to input the weight values ​​into the improved nonlinear Granger causal calculation model and output the impact estimate of the output variable on each of the input variables, wherein the impact estimate is used to characterize the degree of contribution of the input variable to the failure evolution of the output variable; The root cause determination module 14 is used to determine the root cause variable of the fault among the plurality of fault variables based on the estimated impact value.

[0104] Preferably, the data acquisition module 11 acquires time-series observation data of all fault variables of the cloud service system within the fault-related time period through the following steps: The fault records within the relevant time period are obtained from the cloud service monitoring and collection system. The fault observation records include at least a time identifier, a fault variable identifier, a fault observation value, and the tenant identifier information. Based on the tenant identification information, the fault observation records are classified to generate a set of fault observation records divided by tenant; In the fault observation record set corresponding to each tenant, the fault observation values ​​are classified according to the fault variable identifier, and the classified fault observation values ​​are sorted in chronological order according to the time identifier to obtain the time series observation data of each fault variable.

[0105] Preferably, the feedforward gated causal network model includes an input layer, a hidden layer, a gated layer, a fusion layer, and an output layer. The model training module 12 obtains the weight values ​​of each neuron in the feedforward gated causal network model through the following steps: The observation data of any one of the fault variables at the current time is determined as the output variable of the feedforward gated causal network model, wherein the output variable serves as a supervision signal to constrain the predicted value output by the output layer. The time-series observation data of the remaining fault variables are used as the input variables of the feedforward gated causal network model and input into the input layer. The input variables are subjected to feature encoding processing, and an enhanced input vector is output. The enhanced input vector is input into the hidden layer, and feature transformation is performed on the enhanced input vector to output a hidden feature vector. The hidden feature vector is input into the gating layer, and the hidden feature vector is weighted by the gating unit to output the gating enhanced feature vector; The gated enhanced feature vector is input into the fusion layer, and a nonlinear mapping is performed on the gated enhanced feature vector to output the fused feature vector; The fused feature vector is input into the output layer to obtain the predicted value of the output variable; Based on the error between the predicted value of the output variable and the supervision signal, the weight parameters of each neuron in the feedforward gated causal network model are iteratively updated through a parameter update mechanism. When the weight parameters meet the preset convergence conditions, the final value of the weight parameters of the fusion layer is the weight value.

[0106] Preferably, the impact estimation calculation module 13 outputs the estimated impact of the output variable on the input variable through the following steps: The weight values ​​are input into the improved nonlinear Granger causal computation model, and the first-order Jacobian matrix of the output variable relative to the input variable is calculated based on the improved nonlinear Granger causal computation model. Based on the values ​​of each column element of the first-order Jacobian matrix, the estimated influence value of the input variable associated with the current column is calculated respectively. The improved nonlinear Granger causal calculation model introduces a time lag term into the input variables.

[0107] Preferably, the root cause determination module 14 specifically includes: The first root cause determination unit is used to obtain a set of impact estimates for each of the fault variables within several fault-related time periods; The second root cause determination unit is used to determine whether the fault variable is a significant causal influence variable based on each of the aforementioned sets of influence estimates. The third root cause determination unit is used to mark the fault variable as a first preset value if the fault is true, and otherwise mark the fault variable as a second preset value. The fourth root cause determination unit is used to perform causal analysis on all the fault variables based on the first preset value and the second preset value, and determine the fault root cause variables therefrom.

[0108] Preferably, the first root cause determination unit obtains a set of impact estimates for each fault variable within several fault-related time periods through the following steps: Calculate the estimated impact values ​​between the fault variables within several fault-related time periods; Based on all the estimated impact values ​​of each of the fault variables, a set of estimated impact values ​​for each fault variable is obtained.

[0109] Preferably, the second root cause determination unit determines whether the fault variable is a significant causal variable based on each of the estimated impact values ​​through the following steps: Based on each set of impact estimates, the empirical null distribution of the fault variable is calculated accordingly. If the empirical null distribution of the fault variable does not follow a Gaussian distribution, then the fault variable is determined to be a significant causal variable.

[0110] Preferably, the fourth root cause determination unit performs causal analysis on all the fault variables based on the first preset value and the second preset value through the following steps to determine the fault root cause variable: A causal relationship matrix is ​​constructed based on each of the aforementioned fault variables, and the matrix elements of the causal relationship matrix are used to characterize the causal relationship between each of the aforementioned fault variables; If the fault variable is marked as the first preset value, then the corresponding matrix element in the causal relationship matrix is ​​assigned the first preset value; otherwise, the matrix element is assigned the second preset value. Based on the causal relationship matrix, the root cause variables of the fault are selected.

[0111] It should be noted that the cloud service fault diagnosis device based on Granger causality analysis provided in the second aspect embodiment of the present invention can realize all the processes of the cloud service fault diagnosis method based on Granger causality analysis described in the first aspect above. The functions and technical effects of each module and unit in the device are the same as those of the cloud service fault diagnosis method based on Granger causality analysis described in the first aspect above, and will not be repeated here.

[0112] A third aspect of the present invention also provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program; wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute a cloud service fault diagnosis method based on Granger causality analysis as described in any of the first aspects above.

[0113] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary hardware platforms, and of course, it can also be implemented entirely by hardware. Based on this understanding, all or part of the technical solution of the present invention that contributes to the background technology can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM (Read-Only Memory) / RAM (Random Access Memory), magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0114] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cloud service fault diagnosis method based on Granger causality analysis, characterized in that, include: S1: Obtain time-series observation data of all fault variables of the cloud service system during the fault-related time period; S2: Using the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and using the time-series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, the weight values ​​of each neuron in the feedforward gated causal network model are obtained. S3: Input the weight values ​​into the improved nonlinear Granger causal calculation model, and output the estimated value of the impact of the output variable on each of the input variables, wherein the estimated value of the impact is used to characterize the degree of contribution of the input variable to the fault evolution of the output variable; S4: For each of the fault variables, its observed value at the current time is used as the output variable and input into the feedforward gated causal network model. Steps S2 and S3 are executed respectively to obtain the estimated value of the influence between the fault variables. S5: Based on the estimated impact value, determine the root cause variable of the fault among the several fault variables.

2. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 1, characterized in that, Obtain time-series observation data of all fault variables of the cloud service system within the fault-related time period, including: The fault records within the relevant time period are obtained from the cloud service monitoring and collection system. The fault observation records include at least a time identifier, a fault variable identifier, a fault observation value, and the tenant identifier information. Based on the tenant identification information, the fault observation records are classified to generate a set of fault observation records divided by tenant; In the fault observation record set corresponding to each tenant, the fault observation values ​​are classified according to the fault variable identifier, and the classified fault observation values ​​are sorted in chronological order according to the time identifier to obtain the time series observation data of each fault variable.

3. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 1, characterized in that, The feedforward-gated causal network model includes an input layer, hidden layers, gating layers, fusion layers, and an output layer. The weight values ​​of each neuron in the feedforward-gated causal network model are obtained through the following steps: The observation data of any one of the fault variables at the current time is determined as the output variable of the feedforward gated causal network model, wherein the output variable serves as a supervision signal to constrain the predicted value output by the output layer. The time-series observation data of the remaining fault variables are used as the input variables of the feedforward gated causal network model and input into the input layer. The input variables are subjected to feature encoding processing, and an enhanced input vector is output. The enhanced input vector is input into the hidden layer, and feature transformation is performed on the enhanced input vector to output a hidden feature vector. The hidden feature vector is input into the gating layer, and the hidden feature vector is weighted by the gating unit to output the gating enhanced feature vector; The gated enhanced feature vector is input into the fusion layer, and a nonlinear mapping is performed on the gated enhanced feature vector to output the fused feature vector; The fused feature vector is input into the output layer to obtain the predicted value of the output variable; Based on the error between the predicted value of the output variable and the supervision signal, the weight parameters of each neuron in the feedforward gated causal network model are iteratively updated through a parameter update mechanism. When the weight parameters meet the preset convergence conditions, the final value of the weight parameters of the fusion layer is the weight value.

4. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 1, characterized in that, The following steps are used to output an estimated value of the impact of the output variable on the input variable: The weight values ​​are input into the improved nonlinear Granger causal computation model, and the first-order Jacobian matrix of the output variable relative to the input variable is calculated based on the improved nonlinear Granger causal computation model. Based on the values ​​of each column element of the first-order Jacobian matrix, the estimated influence value of the input variable associated with the current column is calculated respectively. The improved nonlinear Granger causal calculation model introduces a time lag term into the input variables.

5. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 1, characterized in that, The step of determining the root cause variable of the fault among the plurality of fault variables based on the impact estimate includes: Obtain a set of impact estimates for each of the fault variables within several fault-related time periods; Based on each set of estimated impact values, determine whether the fault variable is a significant causal variable. If so, the fault variable is marked as a first preset value; otherwise, the fault variable is marked as a second preset value. Based on the first preset value and the second preset value, a causal analysis is performed on all the fault variables to determine the root cause variables of the fault.

6. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 5, characterized in that, The process of obtaining a set of impact estimates for each of the fault variables over several fault-related time periods includes: Calculate the estimated impact values ​​between the fault variables within several fault-related time periods; Based on all the estimated impact values ​​of each of the fault variables, a set of estimated impact values ​​for each fault variable is obtained.

7. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 5, characterized in that, The step of determining whether a fault variable is a significant causal variable based on each of the aforementioned sets of impact estimates includes: Based on each set of impact estimates, the empirical null distribution of the fault variable is calculated accordingly. If the empirical null distribution of the fault variable does not follow a Gaussian distribution, then the fault variable is determined to be a significant causal variable.

8. The cloud service fault diagnosis method based on Granger causality analysis as described in claim 5, characterized in that, The step of performing causal analysis on all the fault variables based on the first preset value and the second preset value to determine the root cause variables of the fault includes: A causal relationship matrix is ​​constructed based on each of the aforementioned fault variables, and the matrix elements of the causal relationship matrix are used to characterize the causal relationship between each of the aforementioned fault variables; If the fault variable is marked as the first preset value, then the corresponding matrix element in the causal relationship matrix is ​​assigned the first preset value; otherwise, the matrix element is assigned the second preset value. Based on the causal relationship matrix, the root cause variables of the fault are selected.

9. A cloud service fault diagnosis device based on Granger causality analysis, used to implement a cloud service fault diagnosis method based on Granger causality analysis as described in any one of claims 1 to 8, the device comprising: The data acquisition module is used to acquire time-series observation data of all fault variables of the cloud service system during the fault-related time period; The model training module is used to take the current observation data of any one of the fault variables as the output variable of the feedforward gated causal network model, and take the time series observation data of the remaining fault variables as the input variables of the feedforward gated causal network model, and input them into the feedforward gated causal network model to obtain the weight values ​​of each neuron in the feedforward gated causal network model. The impact estimation module is used to input the weight values ​​into an improved nonlinear Granger causal calculation model and output an estimated value of the impact of the output variable on each of the input variables, wherein the estimated value of the impact is used to characterize the degree of contribution of the input variable to the failure evolution of the output variable; The root cause determination module is used to determine the root cause variable among the several fault variables based on the estimated impact value.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program; wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform a cloud service fault diagnosis method based on Granger causality analysis as described in any one of claims 1 to 8.

Images