A cloud native intelligent operation and maintenance system based on multi-modal log analysis

Abstract

The application discloses a cloud native intelligent operation and maintenance system based on multi-modal log analysis, which utilizes a kernel-level probe to capture system call data of a micro-service instance and generates a multi-dimensional observation vector; a complexity factor is calculated according to connectivity state information of a micro-service orchestration architecture, and a trajectory matrix representing a running state is generated according to modulation of phase space reconstruction parameters; a simplex is constructed using persistent homology theory and a topological invariant sequence is extracted, and an alarm is generated after determining a topological structure mutation; a root cause node is determined by calculating a maximum Lyapunov index, and infrastructure management and control instructions are generated by performing self-healing reverse mapping according to a topological jump dimension to execute. The application realizes unified modeling and time sequence alignment of heterogeneous observation data, avoids distributed clock asynchronous interference, enhances the robustness of operation and maintenance analysis to dynamic changes of the architecture, and completes closed-loop control from topological feature analysis to physical cluster self-healing.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and specifically to a cloud-native intelligent operation and maintenance system based on multimodal log analysis. Background Technology

[0002] With the development of cloud computing technology, cloud-native architecture has become the mainstream model for building and running large-scale software systems. Cloud-native architecture is usually based on the microservice concept, which breaks down complex applications into a large number of independently deployed and mutually calling service entities. In order to ensure the efficient and stable operation of such distributed systems, intelligent operation and maintenance systems have been introduced into cluster infrastructure management. Its core is to realize the perception and analysis of system status by collecting various monitoring data during the operation of microservices.

[0003] In existing industry operation and maintenance practices, cluster monitoring usually relies on multiple independent collection tools or agent components. The operation and maintenance platform collects basic resource usage indicators such as CPU and memory of each microservice instance, text logs during business execution, and cross-service tracing information. After obtaining this information, traditional analysis engines usually store it in their own independent storage space and perform independent trend judgment and alarm triggering for single-type monitoring indicators by setting static thresholds or applying conventional time-series prediction algorithms.

[0004] However, existing intelligent operation and maintenance platforms suffer from technical defects in practical applications, making it difficult to uniformly model multi-source heterogeneous data. The data generated by microservice operation has significant heterogeneity in terms of form and collection frequency. Resource indicators are usually represented as continuous numerical variables with low-frequency sampling, while file write logs and network link tracing spans are discretely triggered transient pulse events. Existing analysis schemes lack a mechanism for the continuous transformation of physical quantities of underlying heterogeneous data and a strict time-series alignment mechanism. They cannot map discrete events and continuous indicators of different dimensions onto a globally unified time-series grid, resulting in the operation and maintenance system being able to perform isolated analysis on fragmented data silos. It is difficult to build a comprehensive observation model that can accurately reflect the overall state of microservices in a highly complex distributed environment. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the technical deficiency of existing cloud-native operation and maintenance platforms, which lack a continuous transformation and strict time-series alignment mechanism for underlying multi-source heterogeneous observation data, resulting in the inability to map discrete events and continuous indicators of different dimensions onto a globally unified time-series grid for comprehensive state modeling. This invention provides a cloud-native intelligent operation and maintenance system based on multimodal log analysis.

[0006] This invention provides a cloud-native intelligent operation and maintenance system based on multimodal log analysis, comprising:

[0007] The multimodal data continuous extraction module is configured to collect system call data of microservice instances in cloud-native clusters, obtain resource indicators, log write rate, log byte length sequence and distributed tracing span duration from the system call data, convert them into one-dimensional continuous time series signals, and aggregate them to generate multi-dimensional observation vectors for the corresponding microservice nodes.

[0008] The orchestration state awareness and parameter modulation module is configured to read the architecture connectivity state information of microservice nodes from the control plane component of the cloud-native cluster, calculate the complexity factor of the microservice nodes based on the architecture connectivity state information, and convert the complexity factor into the embedding dimension and time delay parameters required for nonlinear phase space reconstruction through a preset mapping function.

[0009] The nonlinear phase space reconstruction module is configured to use the embedded dimension and time delay parameters to perform a time delay mapping operation on each single-dimensional time series signal in the multidimensional observation vector to obtain a local state vector, and to spatially stitch together all the local state vectors corresponding to the multidimensional observation vector to generate a trajectory matrix representing the running state of the microservice node.

[0010] The continuous cohomology calculation and anomaly detection module is configured to construct a simple complex based on the set of state vectors inside the trajectory matrix, calculate the continuous cohomology features of the simple complex at a multi-dimensional distance scale to generate a sequence of topological invariants, calculate the temporal gradient jump amplitude of the sequence of topological invariants, and generate a topological structure abrupt change alarm event when the temporal gradient jump amplitude exceeds the preset topological structure abrupt change threshold.

[0011] The root cause localization and reverse control mapping module is configured to receive topology change alarm events, calculate the Lyapunov exponential variability of the trajectory matrix of each microservice node within the alarm time window, identify the microservice node whose variability exceeds the limit and is located upstream of the service call as the root cause node, and generate infrastructure management instructions corresponding to the root cause node based on the topology invariant jump dimension that triggered the alarm event and send them to the service mesh data plane of the cloud-native cluster for execution.

[0012] Specifically, the complexity factor of the aforementioned microservice node is obtained by multiplying the number of instance replicas, the number of in-degree dependencies, and the number of out-degree dependencies by their respective preset weight coefficients and then performing scalar addition.

[0013] The embedding dimension required for the aforementioned nonlinear phase space reconstruction is obtained by multiplying the complexity factor and the dimension expansion coefficient, taking the integer part, and comparing it with the baseline embedding dimension to take the maximum value; the time delay parameter is obtained by calculating the positive exponential growth result of the complexity factor, multiplying it with the baseline time delay, and taking the integer part.

[0014] The aforementioned maximum Lyapunov exponent is specifically obtained by calculating the ratio of the trajectory deviation distance after the initial small perturbation distance has evolved over a preset time step to the initial small perturbation distance, and then taking the natural logarithm of this ratio and dividing it by the time step.

[0015] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0016] 1. This invention captures file write and network transmission call events bound to microservice instances by deploying probes at the kernel layer of the working node operating system, and transforms the extracted discrete resource indicators, logs and link tracing information into one-dimensional continuous time series signals. Then, by combining the global basic time resolution, zero-order hold and temporal convolution smoothing kernel function are used to perform time axis interpolation and discrete sequence grid resampling operations, realizing efficient aggregation of multi-source heterogeneous observation data in cloud-native environment as well as strict temporal alignment and unified state modeling.

[0017] 2. This invention calculates the architecture complexity factor by sensing the number of instance replicas of microservice nodes and the number of in-degree and out-degree dependencies of service calls. Based on this factor, it dynamically calculates the embedding dimension and time delay parameters required for nonlinear phase space reconstruction using mapping rules such as exponential decay and rounding up. This guides the time delay mapping and spatial splicing of local state vectors, enabling adaptive adjustment of phase space reconstruction geometric control variables and accurate capture of microservice dynamic evolution characteristics, thereby improving the robustness of operation and maintenance analysis against dynamic changes in the architecture.

[0018] 3. This invention quantitatively assesses the divergence rate of the state trajectory in the multidimensional phase space by calculating the maximum Lyapunov exponent of the microservice running state trajectory matrix to locate the root cause node that causes a change in topology. It also combines simple complex evolution to extract the topological invariant sequence reflecting spatial connectivity branches and holes to identify the target jump dimension and match the corresponding circuit breaker or rate limiting strategy template. This realizes intelligent operation and maintenance closed-loop control from the discovery of mathematical topological feature anomalies to the determination of the fault source and the issuance of infrastructure self-healing management instructions to the service mesh data plane. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the cloud-native intelligent operation and maintenance system architecture based on multimodal log analysis, as described in this invention.

[0020] Figure 2 This is the main flowchart of the cloud-native intelligent operation and maintenance analysis method based on multimodal log analysis of the present invention;

[0021] Figure 3 This is a flowchart of the simple complex evolution and continuous homology feature extraction of the present invention;

[0022] Figure 4This is the root cause localization and service mesh self-healing reverse mapping control flow graph of the present invention. Detailed Implementation

[0023] 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.

[0024] Example 1

[0025] See attached document Figure 1-4 This embodiment provides a cloud-native intelligent operation and maintenance system based on multimodal log analysis, including:

[0026] The system includes a multimodal data continuous extraction module 100, an orchestration state perception and parameter modulation module 200, a nonlinear phase space reconstruction module 300, a continuous coherence calculation and anomaly detection module 400, and a root cause localization and inverse control mapping module 500.

[0027] The multimodal data continuous extraction module 100 is deployed on each working node of the cloud-native cluster. It obtains system call data of microservice instances in the cloud-native cluster through kernel-level probes, and transforms and aggregates the collected resource indicators, log write rate, log byte length sequence and distributed tracking span duration into one-dimensional continuous time series signals, generates multi-dimensional observation vectors for corresponding microservice nodes, and outputs the multi-dimensional observation vectors to the nonlinear phase space reconstruction module 300.

[0028] The orchestration state awareness and parameter modulation module 200 communicates with the control plane component of the cloud-native cluster. The module 200 obtains the architectural connectivity state information of each microservice node, including the number of instance replicas, the number of in-degree dependencies, and the number of out-degree dependencies. Based on this architectural connectivity state information, the module 200 uses formulas... Compute microservice nodes Complexity factor ,in, The number of instance replicas. For the number of in-degree dependencies, For the number of out-degree dependencies, Using preset weight coefficients, the orchestration state awareness and parameter modulation module 200 assigns complexity factors according to preset mapping rules. The embedded dimension and time delay parameters are converted into the parameters required for phase space reconstruction. The orchestration state perception and parameter modulation module 200 then sends the generated embedded dimension and time delay parameters to the nonlinear phase space reconstruction module 300.

[0029] The nonlinear phase space reconstruction module 300 receives the multidimensional observation vector output by the multimodal data continuous extraction module 100 and the embedding dimension and time delay parameters output by the orchestration state perception and parameter modulation module 200. Based on the embedding dimension and time delay parameters, the nonlinear phase space reconstruction module 300 performs time delay mapping on each single-dimensional signal in the multidimensional observation vector to construct a local state vector. The nonlinear phase space reconstruction module 300 spatially splices the local state vectors of each dimension to generate a trajectory matrix representing the microservice running state. The nonlinear phase space reconstruction module 300 transmits the trajectory matrix to the continuous coherence calculation and anomaly detection module 400. At the same time, the nonlinear phase space reconstruction module 300 synchronously caches the trajectory matrix and transmits it to the root cause localization and reverse control mapping module 500.

[0030] The continuous cohomology calculation and anomaly detection module 400 receives the trajectory matrix output by the nonlinear phase space reconstruction module 300. The continuous cohomology calculation and anomaly detection module 400 transforms the set of state vectors in the trajectory matrix into a simplex and calculates the continuous cohomology features of the simplex at different distance scales to extract a multi-dimensional topological invariant sequence. The continuous cohomology calculation and anomaly detection module 400 monitors the temporal gradient change of the topological invariant sequence. When the temporal gradient jump amplitude of the topological invariant sequence exceeds the preset topological abrupt change threshold, the continuous cohomology calculation and anomaly detection module 400 generates a topological abrupt change alarm event and notifies the root cause localization and reverse control mapping module 500.

[0031] The root cause localization and reverse control mapping module 500 calculates the Lyapunov exponential variability of the trajectory matrix of each microservice node within the alarm time window. Combining the global service dependency directed graph, the root cause localization and reverse control mapping module 500 traces the upstream link of the abnormal propagation and extracts abnormal nodes with excessive variability within the upstream link. The microservice node at the upstream position of the call is identified as the root cause node to avoid chaotic oscillation interference caused by the avalanche effect in downstream nodes. Based on the extracted topological invariant jump dimension, the root cause localization and reverse control mapping module 500 generates infrastructure management instructions for the root cause node. The root cause localization and reverse control mapping module 500 issues the infrastructure management instructions to the service mesh data plane of the cloud-native cluster to execute control operations.

[0032] Example 2

[0033] See attached document Figure 1-4This embodiment provides a cloud-native intelligent operation and maintenance analysis method based on multimodal log analysis, including the following steps:

[0034] The multimodal data continuous extraction module 100 collects system call data of microservice instances on the worker nodes of the cloud-native cluster. The multimodal data continuous extraction module 100 obtains resource indicators, log write rate, log byte length sequence and distributed tracing span duration from the system call data. The multimodal data continuous extraction module 100 converts the resource indicators, log write rate, log byte length sequence and distributed tracing span duration into one-dimensional continuous time series signals. The multimodal data continuous extraction module 100 aggregates the one-dimensional continuous time series signals to generate multi-dimensional observation vectors for the corresponding microservice nodes.

[0035] The orchestration state awareness and parameter modulation module 200 reads the architectural connectivity state information of the microservice nodes from the control plane component of the cloud-native cluster. This architectural connectivity state information includes the number of microservice instance replicas, the number of in-degree dependencies, and the number of out-degree dependencies. Based on this architectural connectivity state information, the orchestration state awareness and parameter modulation module 200 uses formulas... Compute microservice nodes Complexity factor , in the formula, The number of instance replicas. For the number of in-degree dependencies, For the number of out-degree dependencies, Using preset weight coefficients, the orchestration state awareness and parameter modulation module 200 uses a preset mapping function to assign complexity factors. The embedding dimension and time delay parameters required for nonlinear phase space reconstruction are transformed into the parameters required for nonlinear phase space reconstruction.

[0036] The nonlinear phase space reconstruction module 300 receives the multi-dimensional observation vector generated by the multimodal data continuous extraction module 100, and receives the embedding dimension and time delay parameters generated by the orchestration state perception and parameter modulation module 200. The nonlinear phase space reconstruction module 300 uses the embedding dimension and time delay parameters to perform a time delay mapping operation on each single-dimensional time series signal in the multi-dimensional observation vector to obtain a local state vector. The nonlinear phase space reconstruction module 300 spatially splices all the local state vectors corresponding to the multi-dimensional observation vector to generate a trajectory matrix representing the running state of the microservice node. The nonlinear phase space reconstruction module 300 transmits the trajectory matrix to the continuous coherence calculation and anomaly detection module 400 and synchronously stores the trajectory matrix in the cache.

[0037] The continuous coherence calculation and anomaly detection module 400 receives the trajectory matrix output by the nonlinear phase space reconstruction module 300. The continuous coherence calculation and anomaly detection module 400 constructs a simple complex based on the set of state vectors inside the trajectory matrix. The continuous coherence calculation and anomaly detection module 400 calculates the continuous coherence features of the simple complex at a multi-dimensional distance scale and generates a sequence of topological invariants. The continuous coherence calculation and anomaly detection module 400 calculates the temporal gradient jump amplitude of the sequence of topological invariants. When the temporal gradient jump amplitude exceeds the preset topological abrupt change threshold, the continuous coherence calculation and anomaly detection module 400 generates a topological abrupt change alarm event and sends it to the root cause localization and reverse control mapping module 500.

[0038] The root cause localization and reverse control mapping module 500 calculates the Lyapunov exponential variability of the trajectory matrix of each microservice node. The root cause localization and reverse control mapping module 500 performs topology filtering in conjunction with global service dependencies. The root cause localization and reverse control mapping module 500 identifies the microservice node with excessive variability and located at the top upstream as the root cause node. Based on the topology invariant jump dimension that triggers the alarm event, the root cause localization and reverse control mapping module 500 generates infrastructure management instructions corresponding to the root cause node. The root cause localization and reverse control mapping module 500 issues the infrastructure management instructions to the service mesh data plane of the cloud-native cluster for execution.

[0039] Furthermore, the transformation of heterogeneous feature physical quantities based on kernel-level probes includes the following steps:

[0040] The multimodal data continuous extraction module 100 deploys an extended Berkeley packet filter probe at the kernel layer of the working node operating system in the cloud-native cluster. The multimodal data continuous extraction module 100 binds the extended Berkeley packet filter probe to the kernel trace point and performance monitoring counter of the operating system.

[0041] The multimodal data continuous extraction module 100 obtains the process identifier assigned to the microservice instance in the worker node operating system. The multimodal data continuous extraction module 100 uses the extended Berkeley packet filter probe to capture system call events bound to the process identifier. The system call events include file write call events and network transfer call events.

[0042] The multimodal data continuous extraction module 100 reads low-frequency resource indicator data of the microservice instance through a performance monitoring counter. This low-frequency resource indicator data includes CPU utilization and the number of physical pages occupied in memory. The multimodal data continuous extraction module 100 samples the CPU utilization and the number of physical pages occupied in memory according to a preset sampling period, generating continuous low-frequency resource indicator signals, denoted as... ;

[0043] The multimodal data continuous extraction module 100 captures file write call events triggered by microservice instances. It also counts the number of file write call events occurring within a specific time window, and then uses a formula... In the formula for calculating log write rate, For a moment Log write rate, In the time window Total number of file write event calls captured internally. Given the length of the time window, the multimodal data continuous extraction module 100 splices the log write rates at different times along the time axis to generate a log write rate signal.

[0044] When capturing a file write call event, the multimodal data continuous extraction module 100 reads the system call input parameters of the file write call event. The multimodal data continuous extraction module 100 extracts the count parameter value representing the length of the written bytes from the input parameters of the file write system call. The multimodal data continuous extraction module 100 extracts the count parameter values ​​at different times to generate a log byte length sequence signal, denoted as... ;

[0045] The multimodal data continuous extraction module 100 captures network transmission call events of microservice instances. The multimodal data continuous extraction module 100 extracts the first kernel timestamp when the network request is issued and stores the first kernel timestamp and the thread identifier of the currently executing thread as a key-value pair in the kernel mapping table. When a network response arrival event is captured, the multimodal data continuous extraction module 100 retrieves the corresponding first kernel timestamp from the kernel mapping table based on the thread identifier of the currently executing thread.

[0046] And obtain the second kernel timestamp when the network response arrives, the multimodal data continuous extraction module 100 according to the formula In the formula for calculating the duration of distributed tracing span, For a moment Distributed tracing span duration, The first kernel timestamp, As the second kernel timestamp, the multimodal data continuous extraction module 100 maps the distributed tracking span duration into discrete pulse amplitudes, generating a distributed tracking span duration signal. The multimodal data continuous extraction module 100 also converts the low-frequency resource indicator signal... Log write rate signal Log byte length sequence signal and distributed tracking span duration signal Aggregate the data along the time dimension to generate multimodal observation vectors for microservice nodes.

[0047] Furthermore, methods for aligning and constructing multidimensional observation vectors include:

[0048] The multimodal data continuous extraction module 100 acquires low-frequency resource indicator signals, log write rate signals, log byte length sequence signals, and distributed tracking span duration signals. The multimodal data continuous extraction module 100 sets a global basic time resolution based on the period of the highest frequency signal in the multimodal signals. The multimodal data continuous extraction module 100 divides a discrete time series grid containing multiple reference time nodes according to the global basic time resolution.

[0049] The original sampling period of the low-frequency resource index signal is greater than the global basic time resolution. The multimodal data continuous extraction module 100 uses the zero-order hold algorithm to perform time axis interpolation on the low-frequency resource index signal. The multimodal data continuous extraction module 100 maps the value of the low-frequency resource index signal to each reference time node of the discrete time series grid to generate the aligned low-frequency resource index signal.

[0050] The distributed tracking span duration signal is a discrete pulse signal. The multimodal data continuous extraction module 100 uses a smoothing kernel function to perform a time-domain convolution operation with a fixed window size on the distributed tracking span duration signal. The multimodal data continuous extraction module 100 uses the following formula:

[0051]

[0052] Perform continuous waveform reconstruction;

[0053] In the formula, This is the waveform of the distributed tracking span duration after continuous processing. This is the original distributed tracking span duration signal. The Gaussian smoothing kernel function is used. The preset width of the convolution sliding window;

[0054] The multimodal data continuous extraction module 100 performs a resampling operation on the distributed tracking span duration waveform after continuous processing on the discrete time series grid to generate an aligned distributed tracking span duration signal. The multimodal data continuous extraction module 100 uses a linear interpolation algorithm to perform projection mapping to the reference time node on the log write rate signal and the log byte length sequence signal to generate an aligned log write rate signal and an aligned log byte length sequence signal.

[0055] The multimodal data continuous extraction module 100 can extract data at any reference time node in the discrete time series grid. The multimodal data continuous extraction module 100 synchronously reads the instantaneous values ​​of the aligned low-frequency resource indicator signal, the aligned distributed tracking span duration signal, the aligned log write rate signal, and the aligned log byte length sequence signal. It then combines the extracted values ​​of the four independent physical quantities into a column vector form at a single moment, according to the formula... To complete the construction of a multidimensional continuous observation vector, in the formula, For microservice nodes At the reference time node The multidimensional continuous observation vector; , , , These are the aligned low-frequency resource indicator signal, log write rate signal, log byte length sequence signal, and distributed tracing span duration signal at the reference time node. The instantaneous value, The transpose operator is used to output the constructed multidimensional continuous observation vector to the nonlinear phase space reconstruction module 300.

[0056] Furthermore, the architecture connectivity awareness and complexity factor quantification methods include:

[0057] The orchestration state awareness and parameter modulation module 200 establishes a network communication connection with the control plane component of the cloud-native cluster. The control plane component of the cloud-native cluster includes a container orchestration application interface server and a service mesh control plane application interface. The orchestration state awareness and parameter modulation module 200 listens for lifecycle events and routing rule change events of microservice instances by registering callback functions in the control plane component.

[0058] The orchestration state awareness and parameter modulation module 200 obtains microservice nodes from the container orchestration application interface server. The number of container groups in the ready state, and the microservice nodes. The number of container groups in the ready state is recorded as microservice nodes. Number of instance replicas ;

[0059] The orchestration state awareness and parameter modulation module 200 reads the virtual service configuration list and the target rule configuration list from the service mesh control plane application interface. The orchestration state awareness and parameter modulation module 200 extracts the call routing relationship in the virtual service configuration list and the target rule configuration list, constructs the global service dependency directed graph of the cloud-native cluster, and generates the adjacency matrix of the global service dependency directed graph.

[0060] The orchestration state awareness and parameter modulation module 200 is based on the formula Compute microservice nodes The number of in-degree dependencies;

[0061] In the formula, For microservice nodes The number of in-degree dependencies, This represents the total number of microservice nodes within the cloud-native cluster. For microservice nodes in the adjacency matrix Pointing to microservice nodes The connectivity status value, in the presence of microservice nodes Calling microservice nodes When the network link is... The value is assigned to the constant 1 when no network link exists. The value is assigned to the constant 0;

[0062] The orchestration state awareness and parameter modulation module 200 is based on the formula Compute microservice nodes The number of out-degree dependencies, in the formula, For microservice nodes The number of out-degree dependencies, For microservice nodes in the adjacency matrix Pointing to microservice nodes The connectivity state value;

[0063] The orchestration state awareness and parameter modulation module 200 reads the preset instance weight coefficients. In-degree weighting coefficient and out-degree weight coefficient The state perception and parameter modulation module 200 is arranged according to the formula Perform scalar addition to generate microservice nodes. Complexity factor , in the formula, For microservice nodes The complexity factor is determined by the orchestration state awareness and parameter modulation module 200. Stored in memory for subsequent parameter modulation operations.

[0064] Furthermore, adaptive mapping methods for embedding dimension and time delay include:

[0065] The orchestration state awareness and parameter modulation module 200 reads microservice nodes from system memory. Complexity factor The orchestration state awareness and parameter modulation module 200 obtains the preset baseline embedding dimension and baseline time delay, and uses the baseline embedding dimension as the lower limit of the phase space reconstruction dimension when the microservice node is in a single instance state and has no external network dependency state, and uses the baseline time delay as the phase space reconstruction time interval step size when the microservice node is in a single instance state and has no external network dependency state.

[0066] The orchestration state awareness and parameter modulation module 200 obtains the preset dimension expansion coefficients, and then modulates the dimension expansion coefficients with the complexity factor. Multiplication produces a product, and the result is determined according to the formula. Compute microservice nodes The dynamic embedding dimension, in the formula, For microservice nodes The dynamic embedding dimension, As the baseline embedding dimension, For dimensional expansion coefficients, It is a rounding function. This is a function to find the maximum value.

[0067] The orchestration state awareness and parameter modulation module 200 obtains a preset delay growth factor, and then modulates the delay growth factor with a complexity factor. Multiplication yields the exponential term parameters. To ensure accurate indexing in the discrete multidimensional continuous observation vector sequence during phase space reconstruction, the state sensing and parameter modulation module 200 is arranged according to the formula... Compute microservice nodes The dynamic time delay, in the formula, For microservice nodes Dynamic time delay, As a base time delay, This is the delayed growth coefficient. is the base of the natural logarithm. This is the rounding function;

[0068] The orchestration state awareness and parameter modulation module 200 will dynamically embed dimensions. and dynamic time delay Combining into microservice nodes The phase space reconstruction parameter set is sent by the state perception and parameter modulation module 200 to the nonlinear phase space reconstruction module 300, so that the nonlinear phase space reconstruction module 300 can perform time delay mapping operation of multidimensional continuous observation vectors.

[0069] Example 3

[0070] See attached document Figure 1-4This embodiment provides a method for constructing a local time delay vector, including:

[0071] The nonlinear phase space reconstruction module 300 receives the microservice nodes output by the multimodal data continuous extraction module 100. The multidimensional continuous observation vector is obtained, and the microservice node transmitted by the orchestration state awareness and parameter modulation module 200 is acquired. The phase space reconstruction parameter set, which includes the dynamic embedding dimension. and dynamic time delay ;

[0072] The nonlinear phase space reconstruction module 300 decouples the multidimensional continuous observation vector according to the preset feature dimension index, and separates the first phase space from the multidimensional continuous observation vector. One-dimensional continuous time series signal, where, For feature dimension index, and , The total number of feature dimensions of a multidimensional continuous observation vector;

[0073] To ensure that the reconstruction calculation has sufficient historical data support, the nonlinear phase space reconstruction module 300th verification step... The length of a one-dimensional continuous time series signal in the system cache is such that when the historical time span stored in the system cache is greater than or equal to... At that time, the nonlinear phase space reconstruction module 300 uses the current time node... Using the reference starting point, with dynamic time delay The span step size is calculated sequentially in the system cache along the historical time direction. One historical delay node;

[0074] Nonlinear phase space reconstruction module 300 reads the first... A one-dimensional continuous time series signal at the current time node The instantaneous sampled values ​​are obtained, and the first value is extracted simultaneously. The corresponding sampled values ​​of a one-dimensional continuous time series signal at each historical time delay node;

[0075] The nonlinear phase space reconstruction module 300 will reconstruct the current time node. The instantaneous sampled values ​​and the corresponding sampled values ​​at historical delay nodes are concatenated into an array in descending time order, and then processed according to the formula. Constructing the local state vector, in the formula, For microservice nodes At the time point Regarding the first A local state vector constructed from a one-dimensional continuous time series signal; For the first One-dimensional continuous time series signal at time nodes The instantaneous sampled value; This represents a historical sample value shifted backward by one time delay step; For dynamic time delay; For dynamic embedding dimensions; This is the transpose operator;

[0076] Nonlinear phase space reconstruction module 300 incremental feature dimension index The decoupling and extraction operations are performed repeatedly until all dimensions contained in the multidimensional continuous observation vector are traversed, generating a set of local state vectors corresponding to all feature dimensions, and temporarily storing the set of local state vectors in the memory work area.

[0077] Example 4

[0078] See attached document Figure 1-4 This embodiment provides a method for fusing and stitching together the global state manifold of microservices, including:

[0079] The nonlinear phase space reconstruction module 300 reads microservice nodes from the memory workspace. The set of local state vectors is arranged according to a preset feature dimension order, and the local state vectors corresponding to each feature dimension in the set of local state vectors are spatially spliced ​​vertically.

[0080] The nonlinear phase space reconstruction module 300 is based on the formula Constructing time nodes The global phase space state vector, in the formula, For microservice nodes At the time point The global phase space state vector; For the first The transpose matrix of the local state vector corresponding to a one-dimensional continuous time series signal; The total number of feature dimensions of a multidimensional continuous observation vector; This is the transpose operator;

[0081] Global phase space state vectors represent microservice nodes In order to construct a state manifold that represents the continuous operation and evolution of microservice nodes in the instantaneous high-dimensional running state at a single time node, and to meet the real-time online monitoring requirements of cloud-native environment, the nonlinear phase space reconstruction module 300 sets a sliding global observation time window and extracts multiple consecutive reference time nodes within the global observation time window.

[0082] The nonlinear phase space reconstruction module 300 extracts the local state vectors corresponding to each reference time node in ascending order of time, and performs vertical spatial splicing to generate the global phase space state vectors corresponding to each reference time node. The nonlinear phase space reconstruction module 300 arranges all global phase space state vectors within the global observation time window horizontally in the direction of time sequence to complete the construction of the trajectory matrix of the state manifold.

[0083] The nonlinear phase space reconstruction module 300 is based on the formula Generate microservice nodes The trajectory matrix, in the formula, For microservice nodes The trajectory matrix; to These are continuous reference time nodes within the global observation time window; The trajectory matrix represents the microservice nodes in the multidimensional phase space, which is the total number of reference time nodes included in the global observation time window. As the reference time node advances, the nonlinear phase space reconstruction module 300 controls the sliding global observation time window to slide forward and simultaneously updates the trajectory matrix, thus shaping the evolutionary manifold of the operational state. ;

[0084] The nonlinear phase space reconstruction module 300 transmits the generated and continuously updated trajectory matrix to the continuous coherence calculation and anomaly detection module 400, and synchronously stores the trajectory matrix in the system cache so that the root cause localization and reverse control mapping module 500 can read and perform diagnostic calculations when it receives an alarm event.

[0085] Example 5

[0086] See attached document Figure 1-4 This embodiment provides a method for extracting simple complex evolution and persistent homology features, including:

[0087] The continuous coherence computation and anomaly detection module 400 receives the microservice nodes output by the nonlinear phase space reconstruction module 300. trajectory matrix The continuous coherence calculation and anomaly detection module 400 will calculate the trajectory matrix. Each global phase space state vector is mapped to a discrete data point in a high-dimensional phase space, and the discrete data points are aggregated to construct a phase space point cloud set. ;

[0088] The continuous coherence calculation and anomaly detection module 400 obtains the preset maximum distance scale. and distance scale step and at the zero and maximum distance scales Generate discrete scale sequences between them, and set arbitrary distance thresholds within the scale sequences. The continuous coherence computation and anomaly detection module computes 400 phase space point cloud sets. Any two discrete data points and The Euclidean distance between them;

[0089] When phase space point cloud collection middle The pairwise Euclidean distances between all discrete data points are less than or equal to the distance threshold. At that time, the continuous coherence calculation and anomaly detection module 400 will The combination of discrete data points is The 400-dimensional simplex, continuous homology calculation and anomaly detection module aggregates the current distance threshold. Construct the Vietoris-Rips simplex complex from all the simplexes generated below. ;

[0090] The continuous coherence calculation and anomaly detection module 400 increments the distance threshold according to the scale sequence. Furthermore, low-dimensional simplexes are combined and connected to generate high-dimensional simplexes, constructing nested simplex filtering sequences. ;

[0091] In the evolutionary computation of the simple complex filtering sequence, the continuous cohomology computation and anomaly detection module 400 tracks the evolutionary state of topological features and records the topological dimensions. The distance threshold for the initial generation of topological features is used as the birth scale. And record the topology dimensions. The topological features are used as the death metric by the distance threshold at which the high-dimensional simplex forms a boundary and disappears. ;

[0092] The continuous coherence calculation and anomaly detection module 400 calculates the difference between the death scale and the birth scale as the duration of each topological feature.

[0093] The continuous homology calculation and anomaly detection module 400 reads the system's preset noise filtering threshold and deletes topological features with durations less than the noise filtering threshold from the calculation records. The continuous homology calculation and anomaly detection module 400 counts the number of retained topological features and extracts the number of connected components formed by the zero-dimensional simplex to generate a zero-dimensional Betti number. The number of topological holes formed by one-dimensional simplexes is extracted to generate one-dimensional Betti numbers. The number of topological cavities formed by two-dimensional simplexes is extracted to generate two-dimensional Betti numbers. ;

[0094] The continuous coherence calculation and anomaly detection module 400 is based on the formula Construct the latest time node of the current global observation time window The topological invariant sequence, in the formula, For the latest time point topological invariant sequences, , and The latest time nodes are respectively The corresponding zero-dimensional, one-dimensional, and two-dimensional Betti numbers, The transpose operator, continuous cohomology calculation and anomaly detection module 400, will convert the topological invariant sequence... Stored in memory for use in topological mutation detection operations.

[0095] Furthermore, the system call event data recorded by the multimodal data continuous extraction module 100 includes the physical clock offset and network asynchronous transmission latency of the cloud-native cluster nodes. The continuous coherence computation and anomaly detection module 400 defines the baseline running evolution time of the microservice node under ideal conditions as the absolute time base. Due to the influence of physical clock offset and asynchronous network transmission delay, the actual recorded time series nodes are mapped to the observation time. ;

[0096] because and There is an amplitude upper limit in the physical network, and the observation time has time distortion. Distortion trajectory matrix driving reconstruction The internal data points are relative to the baseline trajectory matrix under the absolute time base. In the higher-dimensional manifold of phase space, it constitutes bounded space perturbation;

[0097] The continuous coherence calculation and anomaly detection module has 400 pairs of distorted trajectory matrices. Performing simple complex filtering and evolutionary computation, according to the stability theorem of persistent homology, the bounded perturbation of the phase space point cloud will strictly constrain the bottleneck distance between its persistent homology diagrams. Therefore, the local nonlinear stretching and compression of the data point distance scale will not change the significant hole connectivity state of the overall space.

[0098] This means that the continuous coherence calculation and anomaly detection module 400 obtains information from the distorted trajectory matrix. The extracted Betti number sequences for each topological dimension with durations greater than the noise filtering threshold are compared with the baseline trajectory matrix. The corresponding sequences remain consistent and follow the stability equivalence relation. ;

[0099] The continuous coherence computation and anomaly detection module 400 outputs the Betti number sequence as the benchmark data feature for anomaly determination, directly avoiding the strong alignment dependency on absolute timestamps, and using the stability mechanism of algebraic topological features to achieve computational immunity to distributed environment clock asynchrony and network latency jitter.

[0100] Furthermore, the continuous coherence calculation and anomaly detection module 400 reads the latest time node of the current global observation time window from the system memory. The topological invariant sequence is obtained, and the higher-order Betti numbers are extracted from the topological invariant sequence. The continuous cohomology calculation and anomaly detection module 400 compares the higher-order Betti numbers with the previous reference time nodes. The corresponding Betti number is used for difference calculation according to the formula. Generate topology dimensions Time gradient jump amplitude , in the formula, For topological dimensions At the latest time point The amplitude of the time gradient jump. For the latest time point The number of Betti, Preceding reference time node Betti number,

[0101] This is the absolute value operator;

[0102] The continuous coherence calculation and anomaly detection module 400 will calculate the temporal gradient jump amplitude. The value is compared with the preset topology mutation threshold. When the time gradient jump amplitude is greater than the preset topology mutation threshold, the continuous coherence calculation and anomaly detection module 400 generates a topology mutation alarm event and sends the topology mutation alarm event to the root cause localization and reverse control mapping module 500.

[0103] The root cause localization and reverse control mapping module 500 receives topology change alarm events and reads the trajectory matrix of each microservice node in the cloud-native cluster within the alarm time window from the system cache. For any microservice node... trajectory matrix The root cause localization and reverse control mapping module 500 selects the state vector corresponding to the alarm trigger time in the phase space as the initial state vector.

[0104] And traverse the trajectory matrix The root cause localization and reverse control mapping module 500 selects the neighboring state vector with the closest Euclidean distance to the initial state vector, and calculates the initial small perturbation distance between the initial state vector and the neighboring state vector, denoted as . As the time step progresses, the root cause localization and reverse control mapping module 500 in the trajectory matrix... The system tracks the phase space positions of the initial state vector and its neighboring state vectors after evolution, and calculates the evolution time interval. The deviation distance of the subsequent trajectory is denoted as ;

[0105] Root cause localization and reverse control mapping module 500 according to formula Compute microservice nodes The Lyapunov index is calculated, and the absolute difference of the index before and after the alarm is triggered is further calculated to generate the Lyapunov index variability.

[0106] In the formula, For microservice nodes The Lyapunov index. Let be the time step of the state vector evolution. For the initial small perturbation distance, For the process The distance of the evolved trajectory deviation. This is the 2-norm operator. The operator for natural logarithms;

[0107] Lyapunov exponents are used to quantize trajectory matrices. The divergence of the internal state manifold, while the Lyapunov exponential variability can simultaneously capture the sharp convergence caused by deadlock or the oscillatory divergence caused by anomalous states;

[0108] The root cause localization and reverse control mapping module 500 obtains the Lyapunov variability of all microservice nodes in the cloud-native cluster within the alarm time window, and combines it with the upstream link range of the reverse source propagation of anomalies in the global service dependency directed graph. To eliminate secondary cascading oscillation interference caused by the avalanche effect in downstream nodes, topology sorting and filtering are performed on all nodes within this link range to extract the set of abnormal nodes whose Lyapunov variability exceeds a preset threshold. The microservice node in this set that is at the upstream position of the service call is identified as the root cause node that triggered the topology change alarm event.

[0109] Furthermore, when root cause localization is reverse-mapped to the service mesh self-healing mechanism;

[0110] The root cause localization and reverse control mapping module 500 extracts the target topology dimension that triggers the topology change alarm event. Target topology dimension The root cause localization and inverse control mapping module 500 determines the Betti number, a dimensional parameter that causes temporal gradient jumps out of bounds, as the root cause node and target topology dimension. As an input parameter, the system's pre-set self-healing reverse mapping rule base is queried, based on the formula. Solving for the reverse control action set, in the formula, For reverse control action set, The root cause node with the largest Lyapunov exponent value. For the target topology dimension, It is a self-healing reverse mapping function;

[0111] Self-healing reverse mapping function Discrete mapping rules that include topological distortion patterns and infrastructure control instructions, when the target topology dimension When the value is constant 1, the jump in the one-dimensional Betti number represents the breakage of the one-dimensional cyclic trajectory in the phase space, corresponding to the deadlock state of the synchronous call chain between microservices in the underlying microservice architecture. Its physical mapping principle is as follows: In the deadlock state, the dependency requests between microservices form a closed loop and cannot proceed downwards. The dynamic trajectory of the system's evolution is manifested in the phase space as being trapped in a low-dimensional circular manifold for a long time, causing the original one-dimensional cyclic trajectory (topological hole) to break or abnormally close, thus triggering a drastic jump in the one-dimensional Betti number. The root cause localization and reverse control mapping module 500 matches the preset circuit breaker strategy template through a self-healing reverse mapping function, and sets the reverse control action set... The assignment is a traffic circuit breaking command for the root cause node. The traffic circuit breaking command is configured with error injection parameters, fast failure rules, and outlier eviction configuration.

[0112] When the target topology dimension When the value is constant 2, the jump in the two-dimensional Betti number represents the expansion of the two-dimensional topological cavity in the phase space. This corresponds to the exhaustion and blocking state of the underlying computing resource pool or network connection pool of the microservice node in the underlying microservice architecture. The physical mapping principle is that the exhaustion of the resource pool will block the normal evolution trajectory of the microservice processing requests. The newly arrived request state can only bypass this blocking region and accumulate and evolve, forming an impenetrable "repulsion region" in the high-dimensional manifold of the phase space, namely the two-dimensional cavity, which is manifested as the abnormal expansion of the two-dimensional Betti number. The root cause localization and reverse control mapping module 500 matches the preset rate limiting strategy template through the self-healing reverse mapping function and sets the reverse control action set. The assignment is a resource throttling instruction for the root cause node. The resource throttling instruction sets a maximum number of concurrent connections, a request processing rate threshold, and a request queuing timeout limit.

[0113] The root cause localization and reverse control mapping module 500 will solve for the reverse control action set. It is encapsulated as a service mesh configuration manifest, and a network connection is established through the remote procedure call protocol to send the service mesh configuration manifest to the service mesh control plane component of the cloud-native cluster;

[0114] The service mesh control plane component receives the service mesh configuration list, parses the service mesh configuration list into network data proxy rules, and distributes the network data proxy rules to the network data plane proxy component bound to the root node. The network data plane proxy component directly loads and executes traffic circuit breaking instructions or resource rate limiting instructions on the forwarding path of network data packets, restricting or blocking the business request traffic flowing to the root node, and completing the closed-loop control mapping from topology change alarm to physical cluster self-healing.

Claims

1. A cloud-native intelligent operation and maintenance system based on multimodal log analysis, characterized in that, include: The multimodal data continuous extraction module (100) is configured to collect system call data of microservice instances in the cloud-native cluster, obtain resource indicators, log write rate, log byte length sequence and distributed tracking span duration in the system call data, convert them into one-dimensional continuous time series signals, and aggregate them to generate multi-dimensional observation vectors of the corresponding microservice nodes. The orchestration state awareness and parameter modulation module (200) is configured to read the architecture connectivity state information of the microservice node from the control plane component of the cloud-native cluster, calculate the complexity factor of the microservice node based on the architecture connectivity state information, and convert the complexity factor into the embedding dimension and time delay parameters required for nonlinear phase space reconstruction through a preset mapping function. The nonlinear phase space reconstruction module (300) is configured to use the embedding dimension and the time delay parameter to perform a time delay mapping operation on each single-dimensional time series signal in the multidimensional observation vector to obtain a local state vector, and to spatially stitch together all the local state vectors corresponding to the multidimensional observation vector to generate a trajectory matrix characterizing the running state of the microservice node. The continuous cohomology calculation and anomaly detection module (400) is configured to construct a simple complex based on the set of state vectors inside the trajectory matrix, calculate the continuous cohomology features of the simple complex at a multi-dimensional distance scale to generate a topological invariant sequence, calculate the temporal gradient jump amplitude of the topological invariant sequence, and generate a topological structure abrupt change alarm event when the temporal gradient jump amplitude exceeds a preset topological structure abrupt change threshold. The root cause localization and reverse control mapping module (500) is configured to receive the topology mutation alarm event, calculate the Lyapunov exponential variability of the trajectory matrix of each microservice node within the alarm time window, and combine the upstream link of the reverse source of the abnormal propagation of the global service dependency directed graph of the cloud-native cluster to extract the set of abnormal nodes with excessive variability. The microservice node in the set that is at the upstream position of the service call is determined as the root cause node. Based on the topology invariant jump dimension that triggers the alarm event, the module generates the infrastructure management instruction corresponding to the root cause node and sends it to the service mesh data plane of the cloud-native cluster for execution.

2. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, The multimodal data continuous extraction module (100) is specifically configured to execute the following feature aggregation logic: Deploy an extended Berkeley packet filter probe at the kernel level of the worker node operating system to capture file write call events and network transport call events bound to the process identifier assigned to the microservice instance; The CPU utilization and physical page usage of microservice instances are sampled by reading performance monitoring counters to generate continuous low-frequency resource indicator signals. The log write rate is calculated by counting the number of file write call events within the statistical time window, generating a log write rate signal, and extracting the count parameter value representing the length of written bytes from the system call input parameters of the file write call event to generate a log byte length sequence signal. Extract the first kernel timestamp of the network request and the second kernel timestamp of the corresponding network response, calculate the distributed tracking span duration and map it to discrete pulse amplitude to generate a distributed tracking span duration signal; The low-frequency resource indicator signal, log write rate signal, log byte length sequence signal, and distributed tracking span duration signal are aggregated in the time dimension to generate the multidimensional observation vector.

3. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 2, characterized in that, The multimodal data continuous extraction module (100) is also configured to perform the following alignment operation when generating multidimensional observation vectors: The global base time resolution is set based on the period of the highest frequency signal in the multimodal signal, and a discrete time series grid containing multiple reference time nodes is divided. A zero-order hold algorithm is used to perform time-axis interpolation on the low-frequency resource index signal to generate an aligned low-frequency resource index signal. A continuous waveform reconstruction is performed by performing a time-domain convolution operation with a fixed window size on the distributed tracking span duration signal using a smoothing kernel function, and a resampling operation is performed on the discrete time series grid to generate an aligned distributed tracking span duration signal. A linear interpolation algorithm is used to perform projection mapping to a reference time node on the log write rate signal and the log byte length sequence signal to generate an aligned log write rate signal and an aligned log byte length sequence signal. At any reference time node of the discrete time series grid, the aligned instantaneous values ​​of each signal are synchronously read and combined into a column vector to complete the construction of a multidimensional continuous observation vector.

4. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, The orchestration state awareness and parameter modulation module (200) is specifically configured as follows when calculating the complexity factor of microservice nodes: Get the number of container groups in the ready state of the microservice node and record it as the number of instance replicas; Read the virtual service configuration list and the target rule configuration list, extract the call routing relationship, construct the adjacency matrix of the global service dependency directed graph, and calculate the number of in-degree dependencies and the number of out-degree dependencies; Read the preset instance weight coefficients, in-degree weight coefficients, and out-degree weight coefficients, and then apply the formula... Compute microservice nodes Complexity factor ; In the formula, For the number of instance replicas, For the number of in-degree dependencies, For the number of out-degree dependencies, These are the preset instance weight coefficient, in-degree weight coefficient, and out-degree weight coefficient, respectively.

5. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 4, characterized in that, The orchestration state awareness and parameter modulation module (200) is specifically configured as follows when converting the embedding dimension and time delay parameters required for nonlinear phase space reconstruction: Multiply the preset dimension expansion coefficient by the complexity factor and round up, then take the maximum value after comparing it with the preset baseline embedding dimension as the dynamic embedding dimension. The preset delay growth coefficient is multiplied by the complexity factor to calculate its exponential growth result. The product of the exponential growth result and the preset base time delay is rounded to the nearest integer and compared with the constant 1. The maximum value is taken as the dynamic time delay. The calculated dynamic embedding dimension and dynamic time delay are combined into a phase space reconstruction parameter set and sent to the nonlinear phase space reconstruction module (300).

6. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 4, characterized in that, The nonlinear phase space reconstruction module (300) is specifically configured to execute the following construction logic: According to the preset feature dimension index, the single-dimensional continuous time series signal is separated from the multi-dimensional continuous observation vector; When the historical time span stored in the system cache is greater than or equal to the span calculated by the time delay parameter and the embedding dimension, the corresponding sampled values ​​on the historical time delay nodes are extracted sequentially in the historical time direction, with the current time node as the reference starting point and the time delay parameter as the span step size. Concatenate the instantaneous sampled value at the current time node with the corresponding sampled values ​​at each historical delay node in descending order of time to construct a local state vector. The local state vectors corresponding to each feature dimension are vertically spliced ​​to construct the global phase space state vector. Within the sliding global observation time window, the global phase space state vectors corresponding to multiple consecutive reference time nodes are horizontally arranged according to the time series direction to complete the construction of the trajectory matrix.

7. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, The specific configuration of the continuous cohomology calculation and anomaly detection module (400) when constructing the simplicial complex and generating the topological invariant sequence is as follows: The global phase space state vectors contained in the trajectory matrix are mapped to discrete data points in a high-dimensional phase space, and aggregated to construct a phase space point cloud set; A discrete scale sequence is generated between the zero value and the maximum distance scale. For any distance threshold, when the pairwise Euclidean distances between multiple discrete data points in the phase space point cloud set are all less than or equal to the distance threshold, they are combined into a simplex and aggregated to construct a Vietoris-Rips simplex complex. A nested simple complex filtering sequence is constructed based on an increasing distance threshold of the scale sequence. The birth scale and death scale of the topological feature are recorded and the difference between the two is calculated as the duration. Topological features with a duration less than the noise filtering threshold are deleted. The zero-dimensional Betti number is generated by extracting the number of connected components formed by the zero-dimensional simplex, the one-dimensional Betti number is generated by extracting the number of topological holes formed by the one-dimensional simplex, and the two-dimensional Betti number is generated by extracting the number of topological cavities formed by the two-dimensional simplex, thereby constructing the topological invariant sequence.

8. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, The continuous coherence calculation and anomaly detection module (400) is also configured to perform clock asynchronous disturbance rejection based on continuous coherence stability: For the distorted trajectory matrix generated by observation time-driven reconstruction with physical clock offset and asynchronous network transmission delay of cloud-native cluster nodes, simple complex filtering and evolution calculation are performed. Since the clock offset and network delay constitute bounded perturbation in phase space, the stability theorem of continuous cohomology is used to ensure that the significant topological features extracted from the distorted trajectory matrix (i.e., the Betti number sequence with duration greater than the noise threshold) are consistent with the Betti number sequence of the reference trajectory matrix under the absolute time reference, thus eliminating the interference of nonlinear time distortion on multimodal time series analysis.

9. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, When the root cause localization and reverse control mapping module (500) calculates the maximum Lyapunov exponent, the specific configuration is as follows: In the phase space, the state vector corresponding to the alarm trigger time is selected as the initial state vector, and the trajectory matrix of the microservice nodes is traversed to select the neighboring state vector with the closest Euclidean distance to the initial state vector. Calculate the initial small perturbation distance between the initial state vector and its neighboring state vectors, and trace the evolved phase space position in the trajectory matrix to calculate the evolution time interval. The distance of the trajectory deviation afterward; According to the formula Compute microservice nodes Maximum Lyapunov index ; In the formula, The initial small perturbation distance, For the process The distance of the evolved trajectory deviation. This is the 2-norm operator. This is the operator for the natural logarithm.

10. The cloud-native intelligent operation and maintenance system based on multimodal log analysis according to claim 1, characterized in that, When the root cause localization and reverse control mapping module (500) generates infrastructure management instructions, it is specifically configured to execute the following reverse mapping logic: Extract the target topology dimension that triggers the topology mutation alarm event, and use the root cause node and the target topology dimension as input parameters to query the preset self-healing reverse mapping rule base. When the target topology dimension is a constant 1, the preset circuit breaker strategy template is matched to generate a traffic circuit breaker instruction for the root cause node, so as to limit the business request traffic in the deadlock state of the synchronous call chain between the corresponding microservices. When the target topology dimension is a constant of 2, the preset rate limiting strategy template is matched to generate resource rate limiting instructions for the root cause node, so as to limit the business request traffic in the corresponding microservice node’s underlying computing resource pool or network connection pool that is in a state of exhaustion and blockage. The encapsulated service mesh configuration list is sent to the service mesh control plane component of the cloud-native cluster for execution.

Images