An operation and maintenance method and system based on causal inference and optimization, and a storage medium

Abstract

The application relates to an operation and maintenance method and system based on causal inference and optimization and a storage medium, the method comprising the following steps: constructing a causal knowledge graph based on local operation and maintenance event information; based on the causal knowledge graph, adopting a privacy-protected federated learning strategy to perform causal knowledge sharing and global model aggregation among multiple service nodes, obtaining a shared model integrating multi-domain causal knowledge; in response to an ability optimization scheme proposed by a local intelligent agent based on the shared model, performing constraint verification on the ability optimization scheme based on a preset safety boundary constraint, and optimizing the local intelligent agent by using the ability optimization scheme that passes the verification; when an operation and maintenance accident occurs, calculating the responsibility weight of each intelligent agent participating in execution on a causal path, so as to complete the responsibility determination of each intelligent agent. The application solves the problems of missing causal inference, knowledge silos, intelligent agent evolution risks and unclear responsibility in data center operation and maintenance, and improves operation and maintenance efficiency and safety.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer operation and maintenance technology, specifically to an operation and maintenance method, system, and storage medium based on causal inference and optimization. Background Technology

[0002] With the scaling up of data centers and the increasing prevalence of multi-cluster and multi-region collaborative operations, artificial intelligence's intelligent agent technology has been widely applied in the field of automated operations and maintenance. However, existing solutions face a series of core challenges.

[0003] The agents in each data center operate independently, and their experience and knowledge in handling faults cannot be effectively shared, forming "knowledge silos." For example, the same or similar faults are repeatedly analyzed and handled in different data centers. Statistics show that the processing time is approximately 60% of the initial processing time, resulting in serious resource waste and inefficiency. Current mainstream root cause analysis systems mostly rely on statistical correlation, failing to establish accurate causal chains, leading to a high false positive rate. The online learning capabilities of agents lack effective safety constraints, posing a risk of major incidents due to their autonomous optimization behavior violating business rules. In multi-agent collaborative scenarios, the lack of a quantitative causal responsibility tracing mechanism makes it difficult to clearly define the specific responsibilities of each agent after an incident, hindering problem improvement. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the existing technologies mentioned above. This invention provides an operation and maintenance method, system, and storage medium based on causal inference and optimization, which can accurately perform causal inference, break down knowledge silos, ensure the evolutionary security of intelligent agents, and achieve clear traceability of responsibilities.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: An operation and maintenance method based on causal inference and optimization includes the following steps: Construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; Based on the causal knowledge graph, a privacy-preserving federated learning strategy is adopted to share causal knowledge and aggregate global models among multiple service nodes, thereby obtaining a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. In response to the capability optimization scheme proposed by the local agent based on the shared model, the capability optimization scheme is constrained and verified based on a preset safety boundary constraint, and the local agent is optimized using the verified capability optimization scheme. When an operational incident occurs, the causal path leading to the incident is extracted based on the causal knowledge graph. The responsibility weight of each intelligent agent participating in the execution on the causal path is calculated by combining the mapping relationship of the execution intelligent agents corresponding to each operational event node on the causal path, so as to complete the determination of responsibility for each intelligent agent.

[0006] The beneficial effects of this invention are as follows: by constructing a knowledge graph with causal strength, the false judgment rate is significantly reduced; by adopting a federated learning architecture, efficient cross-domain knowledge reuse is achieved while ensuring that the original data is not leaked, significantly improving the knowledge reuse rate; by employing a security boundary constraint mechanism to constrain and verify the optimization scheme of the agent, the potential security incident rate caused by agent evolution is significantly reduced. The tracing method based on causal chain quantification improves the accuracy of responsibility identification in multi-agent collaboration, facilitating precise improvements.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, the PETER-CLARK algorithm is used to perform conditional independence tests and direction determination on the operation and maintenance event nodes in the local operation and maintenance event information, and the causal strength is calculated based on Do calculus to obtain the causal knowledge graph that supports intervention reasoning and counterfactual reasoning.

[0009] Furthermore, the formula for calculating the strength of causality is as follows: ; Where c(X→Y) represents the causal strength from event node X to event node Y, and P(Y|do(X=1)) and P(Y|do(X=0)) are the probabilities of event node Y occurring when event node X is forcibly intervened to make its value 1 and 0, respectively.

[0010] Furthermore, the Fisher Z-transform is used to perform a conditional independence test on the operation and maintenance event nodes in the local operation and maintenance event information.

[0011] Furthermore, a privacy-preserving federated learning strategy is adopted to perform causal knowledge sharing and global model aggregation among multiple service nodes, obtaining a shared model that integrates multi-domain causal knowledge, including the following steps: Deploy a federated learning architecture across multiple service nodes; Each service node's local agent trains a local model based on local operation and maintenance data; After the local model is trained, homomorphic encryption algorithm and differential privacy protection method are used to protect the model parameters of the local model for privacy. Based on the deployed federated learning architecture, the privacy-preserving model parameters are aggregated using a secure multi-party computation method to obtain a shared model that integrates multi-domain causal knowledge, and the shared model is published to each service node.

[0012] Furthermore, the homomorphic encryption algorithm employs the Paillier encryption algorithm, with the encryption formula as follows: ; Where Enc(m) is the encrypted ciphertext, m is the plaintext, the encryption public key is (g, n), and r is a random number.

[0013] Furthermore, the differential privacy protection method adds Laplace noise to the encrypted model parameters.

[0014] Furthermore, the local model is trained using the Transformer architecture, whose loss function L includes a causal consistency loss term L_causal, satisfying: ; ; in, ; ; ; For the true value, For input The model parameters are The model's predicted values; Lprediction is the prediction loss, Lprivacy is the privacy loss, Lcausality is the causality loss, and Lregularity is the regularization loss. These are the original model parameters. Here are the privacy-preserving model parameters: clocal(X→Y) and cglobal(X→Y) represent the causal strength from event node X to event node Y in the local and global knowledge graphs, respectively; λ1 is the weight coefficient of the prediction loss, λ2 is the weight coefficient of the causal loss, λ3 is the weight coefficient of the privacy loss, and λ4 is the weight coefficient of the regularization loss; the local knowledge graph is a causal knowledge graph constructed by local service nodes based on local operation and maintenance event information; the global knowledge graph is obtained by sharing causal knowledge among multiple service nodes using a privacy-preserving federated learning strategy.

[0015] Furthermore, during the release of the shared model, a sliding window mechanism is used to monitor operational anomalies in real time, and a rollback operation is triggered when the number of anomalies within the sliding window exceeds a preset number.

[0016] Furthermore, the release method adopts a gray-scale release approach, and the release is carried out in multiple stages according to a preset ratio.

[0017] Furthermore, each service node performs knowledge anonymization processing on the original information before uploading parameters. The anonymization rules include at least removing IP addresses, user accounts, and plaintext data of specific business operations from the operation and maintenance data.

[0018] Furthermore, the security boundary constraints include at least business-dimensional constraints, resource-dimensional constraints, and time-dimensional constraints; wherein, the business-dimensional constraints include at least service availability, maximum response latency, and failure rate threshold; the resource-dimensional constraints include at least CPU utilization, memory utilization, and network bandwidth limit; and the time-dimensional constraints include at least preset optimization operation windows and rules prohibiting operation during peak business periods.

[0019] Furthermore, the local intelligent agent generates audit logs during the capability optimization process, and writes the audit logs to the blockchain for irreversible evidence storage.

[0020] Furthermore, the formula for calculating the responsibility weight of each agent is as follows: ; Where Wj is the responsibility weight of the j-th agent, ci is the causal strength of the i-th edge of the causal path, and pi is the confidence level of the i-th edge. M(vi) is the indicator function, and M(vi) is the operation execution agent corresponding to the operation and maintenance event node vi.

[0021] Furthermore, the method for assigning responsibility to each agent is to determine the responsibility level based on the scope of the responsibility level where the agent's responsibility weight falls.

[0022] To address the aforementioned technical problems, this invention also provides an operation and maintenance system based on causal inference and optimization, the specific technical content of which is as follows: An operation and maintenance system based on causal inference and optimization includes: The causal knowledge graph module is used to construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; The federated knowledge sharing module, connected to the causal knowledge graph module, is used to perform causal knowledge sharing and global model aggregation among multiple service nodes based on the causal knowledge graph and using a privacy-preserving federated learning strategy to obtain a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. The agent safety evolution module is connected to the federated knowledge sharing module and the causal knowledge graph module respectively. It is used to respond to the capability optimization scheme proposed by the local agent based on the sharing model, perform constraint verification on the capability optimization scheme based on the preset safety boundary constraints, and optimize the local agent using the verified capability optimization scheme. The multi-agent responsibility tracing module is connected to the causal knowledge graph module. When an operation and maintenance accident occurs, it extracts the causal path that led to the operation and maintenance accident based on the causal knowledge graph, and calculates the responsibility weight of each agent participating in the execution on the causal path by combining the mapping relationship of the execution agents corresponding to each operation and maintenance event node on the causal path, so as to complete the determination of responsibility for each agent.

[0023] To address the aforementioned technical problems, the present invention also provides a storage medium, the specific technical content of which is as follows: A storage medium storing a computer program or computer instructions, which, when executed by a computer's processor, implement the steps of the aforementioned operation and maintenance method based on causal inference and optimization. Attached Figure Description

[0024] Figure 1 This is a flowchart of an operation and maintenance method based on causal inference and optimization in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the causal knowledge graph in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the construction process of the causal knowledge graph in an embodiment of the present invention. Figure 4 This is a flowchart of the federated learning strategy in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the constraint verification of the capability optimization scheme in an embodiment of the present invention; Figure 6 This is a flowchart illustrating the multi-agent responsibility tracing process in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structure of an operation and maintenance system based on causal inference and optimization in an embodiment of the present invention; Figure 8 This is a schematic diagram of the system architecture of the federated learning architecture in an embodiment of the present invention. Detailed Implementation

[0025] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0026] Example 1 like Figure 1 As shown, this embodiment provides an operation and maintenance method based on causal inference and optimization, including the following steps: S1. Construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; The specific construction method is as follows: The structure of a knowledge causal graph is defined as G=(V,E,C,P); Where V is the set of operation and maintenance event nodes, E is the set of causal relationship edges, C is the set of causal strengths (values ​​from 0 to 1), and P is the set of confidence scores (values ​​from 0 to 1). An improved Peter-Clark algorithm combined with operation and maintenance domain knowledge and a causal discovery algorithm is used to construct a system as follows: Figure 2 The causal knowledge graph shown.

[0027] S100. Data Preprocessing. Local operation and maintenance event information is processed by extracting event data, cleaning the data, and standardizing it.

[0028] S101. Conditional Independence Test. For any two event nodes X and Y, test their independence under a given condition set Z. Calculate the partial correlation coefficient using the Fisher Z-transform, with the formula z = 0.5 × ln((1 + r) / (1-r)), where r is the partial correlation coefficient and z follows a normal distribution. If the p-value is greater than the significance level α, where α = 0.05, then X and Y are considered conditionally independent under the given Z.

[0029] S102. Skeleton Learning. Construct the skeleton of an undirected graph based on the conditional independence test results. If X and Y are not independent under all conditional sets, add an edge (X,Y).

[0030] S103. Direction Determination. Determine the causal direction by combining knowledge from the operations and maintenance domain. Rules include: chronological order (cause must precede effect), business logic (configuration changes may cause failures, but failures may not cause configuration changes), and hierarchical dependency (lower-level failures may cause upper-level failures, but the reverse is not true).

[0031] S104. Causal Strength Calculation. The causal effect strength is calculated based on the Do calculus, with the formula c(X→Y) = P(Y|do(X=1))-P(Y|do(X=0)), where do(X) represents the intervention in X, and P(Y|do(X)) represents the probability of Y occurring after the intervention in X.

[0032] Causal reasoning engine: Supports counterfactual reasoning and intervention reasoning.

[0033] The counterfactual reasoning formula is P(Yx|X=x',Y=y')=ΣzP(Yx|X=x',Y=y',Z=z)×P(Z=z|X=x',Y=y'), where Yx represents the value of Y if X is x, used to answer the question of what would happen if different operations were performed.

[0034] The intervention inference formula is P(Y|do(X=x))=ΣzP(Y|X=x, Z=z)×P(Z=z), which is used to predict the change in system state after performing a certain operation.

[0035] S2. Based on the causal knowledge graph, a privacy-preserving federated learning strategy is adopted to share causal knowledge and aggregate global models among multiple service nodes to obtain a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. The service nodes can be distributed servers such as data centers, computing nodes, and local servers. The federated learning strategy adopts a horizontal federated learning architecture, in which each data center has the same data feature space but a different sample space.

[0036] Adopting such Figure 4 The federated learning strategy shown below has the following specific steps: Federated Learning Architecture: A horizontal federated learning architecture is adopted, in which each data center has the same data feature space but a different sample space.

[0037] Local training process: S200. Data Preparation. Each data center uses local operational data to train local models; the raw data does not leave the domain.

[0038] S201. Model Training. The local model adopts the Transformer architecture, with the input being a sequence of operational events and the output being fault prediction and handling suggestions.

[0039] The loss function L is: ; To ensure consistency between local and global causal knowledge, the loss of causal consistency is addressed. in, Mean squared error is used to predict losses from faults. To mitigate privacy concerns, measure the difference in parameters before and after adding noise; L regularization = λ1=0.5, λ2=0.3, λ3=0.1, λ4=0.1, and λ1, λ2, λ3, and λ4 can all be optimized using the validation set.

[0040] The loss function L ensures the consistency of causal knowledge shared across data centers, thus preventing the propagation of false correlations.

[0041] S202. Parameter Encryption. Homomorphic encryption is used to encrypt model parameters, supporting aggregation operations in ciphertext. The encryption formula is as follows: , where m is the plaintext parameter, g and n are the public key parameters, r is a random number, and the key length is 2048 bits.

[0042] S203. Differential Privacy Protection. Laplace noise is added before the uploaded parameters, with a noise scale of ε / Δf, where ε is the privacy budget and Δf is the sensitivity. The formula is: [Noise formula would be inserted here]. Privacy budget ε=1.0.

[0043] The global aggregation steps are as follows: Parameter collection. The federated server collects encrypted parameters from each node.

[0044] Secure aggregation. Parameter aggregation is performed using secure multi-party computation techniques, ensuring the server cannot access the parameters of individual nodes. The aggregation formula is as follows: Where wi is the model parameter of the i-th node, and ni is the number of samples of the i-th node. Secure multi-party computation supports up to 50 participants.

[0045] Model distribution. The aggregated global model is distributed to each node.

[0046] Local updates. Each node updates its local model using the global model, employing the FedAvg federated averaging algorithm. Privacy protection metrics include a privacy budget of ε=1.0, satisfying ε-differential privacy; a homomorphic encryption key length of 2048 bits; and secure multi-party computation supporting up to 50 participants.

[0047] S3. In response to the capability optimization scheme proposed by the local agent based on the shared model, the capability optimization scheme is constrained and verified based on the preset security boundary constraints, and the local agent is optimized using the verified capability optimization scheme; The evolution of intelligent agent capabilities is divided into three dimensions: knowledge evolution, strategy evolution, and tool evolution, forming a capability evolution or capability optimization framework.

[0048] Safety boundary definition: Safety boundary B = (C, R, T).

[0049] Business constraint set C: C1 Service availability is greater than or equal to 99.9%, C2 Response time is less than or equal to 500ms (P99), C3 Failure rate is less than 0.1% per hour, C4 Data consistency is strong consistency, and C5 Order processing latency is less than or equal to 200ms.

[0050] Resource constraint set R: R1 CPU utilization less than or equal to 80%, R2 memory utilization less than or equal to 85%, R3 network bandwidth less than or equal to 70%, R4 disk I / O less than or equal to 90%, R5 number of connections less than the maximum number of connections multiplied by 80%.

[0051] Time constraint set T: T1 Evolution window 02:00 to 05:00 is the off-peak business period; T2 Change freeze period prohibits evolution during the last 3 days of each month; T3 Business peak period prohibits operation from 09:00 to 11:00 and 14:00 to 16:00; T4 Major event guarantee period prohibits evolution.

[0052] like Figure 5 As shown, the constraint verification steps for the capability optimization scheme are as follows: S300. Pre-evolution assessment. After the agent proposes an evolutionary plan, it is simulated and verified in a sandbox environment. Verification metrics include business impact assessment, resource consumption assessment, and risk assessment. If any metric exceeds the safety boundary, the evolutionary plan is rejected.

[0053] The security intervention mechanism for security boundaries is as follows: Intervention trigger conditions: abnormal business indicators, i.e., availability decreases by more than or equal to 0.1%; abnormal resource indicators, i.e., CPU or memory exceeds the threshold by more than or equal to 10%; abnormal causal chain, i.e., new causal relationship conflicts with historical knowledge by more than or equal to 30%.

[0054] The intervention strategy is divided into three levels: Level 1 intervention is an alarm notification applicable to minor anomalies, Level 2 intervention is a pause in evolution applicable to moderate anomalies, and Level 3 intervention is a rollback to the pre-evolution state applicable to severe anomalies.

[0055] Evolution Audit Log: Records the entire evolution process, including evolution proposals, evaluation results, monitoring data, and intervention records. The logs are stored on the blockchain to ensure they cannot be tampered with.

[0056] The simulation verification strategy is as follows: Evolutionary Proposal: The agent submits a capability evolution plan, including the evolution goals, implementation steps, and expected results.

[0057] Sandbox simulation: Simulate the execution of evolutionary plans in an isolated environment to verify their effectiveness and risks.

[0058] Boundary constraint check: Check whether the evolution scheme meets the three-dimensional security boundaries of business, resources, and time.

[0059] Gray-scale release: Solutions that meet the constraints are first deployed on a small scale and then gradually rolled out in proportions of 5% to 20% to 50% to 100%.

[0060] Monitoring and verification: Continuously monitor key indicators during the gray-scale phase and intervene promptly when anomalies are detected.

[0061] Triggered intervention: When an anomaly is detected, different levels of intervention are performed based on the severity.

[0062] Rollback and recovery: In the event of a severe anomaly, the system will automatically roll back to its pre-evolution state to ensure system stability.

[0063] It should be noted that the evolution in the embodiments of this invention is an optimization within the technical content. Agent evolution or optimization includes, but is not limited to, the following operations: Parameter-driven optimization (PDO) enhances capabilities by modifying the model's internal parameters. This includes supervised fine-tuning (SFT), efficient parameter fine-tuning (such as LoRA), reinforcement learning-based optimization (such as RLHF, DPO, PPO), and hybrid training strategies (SFT and RL iterations). The goal is to make the agent more accurate in decision-making, long-term planning, and tool usage.

[0064] Parameterless optimization: Without changing the model weights, it dynamically improves behavior quality through cue engineering (such as CoT, Reflexion), retrieval augmentation generation (RAG), external memory management, tool learning and automatic generation (such as ToolMaker), multi-agent collaboration / debate (such as CAMEL), and contextual experience reuse.

[0065] Self-evolution mechanism: The intelligent agent constructs a closed loop of perception-decision-execution-feedback, and automatically corrects its trajectory by using feedback from the environment / humans / AI referees (such as Self-Rewarding, STaR), compresses experience into reusable skills, and proactively reconstructs workflows or toolsets to achieve continuous growth across sessions.

[0066] Architecture-level optimization includes workflow topology adjustment (automatic design of nodes and edges), multi-agent role allocation and communication protocol evolution, test-time computation expansion (such as MCTS enhanced inference), and collaborative iteration of the base model and agent system (such as "AI tuning AI").

[0067] S301. Monitoring During Evolution. Key metrics are monitored in real-time during the evolution process, using a sliding window to detect anomalies. The window size is 5 minutes; if the number of anomalies within the window exceeds a threshold of 3, evolution is paused. Monitoring metrics are sampled every 10 seconds.

[0068] S302. Post-evolution verification. After the evolution is complete, perform full regression testing to verify that the pass rate of all historical test cases is greater than or equal to 99%, and the pass rate of newly added test cases is greater than or equal to 95%.

[0069] S4. When an operation and maintenance accident occurs, the causal path leading to the operation and maintenance accident is extracted based on the causal knowledge graph. The responsibility weight of each intelligent agent participating in the execution on the causal path is calculated by combining the mapping relationship of the execution intelligent agents corresponding to each operation and maintenance event node on the causal path, so as to complete the determination of responsibility for each intelligent agent.

[0070] like Figure 6 As shown, the multi-agent responsibility tracing method is as follows when an operational accident occurs: S400. Incident Detection: The monitoring system detects service anomalies or malfunctions.

[0071] S401. Causal Chain Extraction: Extract all causal paths leading to the accident from the causal knowledge graph.

[0072] S402. Agent Mapping: Maps event nodes on a causal path to the agent that performs the operation.

[0073] S403. Responsibility Weight Calculation: Calculate the responsibility contribution of each agent based on causal strength and confidence level.

[0074] S404. Traceability Report: Generates a complete report that includes causal chain visualization, responsibility analysis, and improvement recommendations.

[0075] S405. Query Request: Operations and maintenance personnel can query historical incident tracing information through the interface.

[0076] Compared with the prior art, the beneficial effects of the present invention are reflected in: Root cause analysis has high accuracy: by constructing a knowledge graph with causal strength, the false positive rate is significantly reduced from the current approximately 35% to below 10%.

[0077] High efficiency and security in knowledge sharing: The federated learning architecture enables efficient reuse of cross-domain knowledge while ensuring that the original data is not leaked. The knowledge reuse rate can be increased by more than 80%, and triple privacy protection technology is adopted.

[0078] The risks of agent evolution are controllable: The security boundary mechanism, combined with sandbox pre-launch and canary release, reduces the potential security incident rate caused by agent evolution by more than 90%.

[0079] Clear and transparent definition of responsibility: The traceability method based on causal chain quantification enables the accuracy of responsibility positioning in multi-agent collaboration to reach over 95%, supporting precise improvement. Based on the above technical solution, the present invention can also be improved as follows.

[0080] Example 2 like Figure 7 The present invention provides an operation and maintenance system based on causal inference and optimization, comprising: The causal knowledge graph module is used to construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; The federated knowledge sharing module, connected to the causal knowledge graph module, is used to perform causal knowledge sharing and global model aggregation among multiple service nodes based on the causal knowledge graph and using a privacy-preserving federated learning strategy to obtain a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. The agent safety evolution module is connected to the federated knowledge sharing module and the causal knowledge graph module respectively. It is used to respond to the capability optimization scheme proposed by the local agent based on the sharing model, perform constraint verification on the capability optimization scheme based on the preset safety boundary constraints, and optimize the local agent using the verified capability optimization scheme. The multi-agent responsibility tracing module is connected to the causal knowledge graph module. When an operation and maintenance accident occurs, it extracts the causal path that led to the operation and maintenance accident based on the causal knowledge graph, and calculates the responsibility weight of each agent participating in the execution on the causal path by combining the mapping relationship of the execution agents corresponding to each operation and maintenance event node on the causal path, so as to complete the determination of responsibility for each agent.

[0081] like Figure 8 As shown, a federated learning architecture is adopted. Local agent clusters, causal knowledge graphs, federated learning clients, and security boundary monitoring modules are deployed in each data center at the local layer to ensure localized data processing. At the federated coordination layer, a federated aggregation server, a global causal graph, capability evolution auditing, and a responsibility tracing engine are deployed to coordinate knowledge sharing and agent evolution across multiple data centers. Data flow is as follows: local model parameters are encrypted and uploaded to the federated server; after global model aggregation, data is distributed to each node, ensuring that raw data does not leave the domain and protecting privacy.

[0082] Example 3 Data Center A successfully handled a "database connection pool exhaustion" failure. The local agent transformed its handling experience into causal knowledge and stored it in its causal graph. After anonymization, this knowledge was encoded as incremental model parameters and uploaded to the federated server using homomorphic encryption and differential privacy processing. The federated server aggregated knowledge from multiple centers, updated the global model, and then distributed it. Subsequently, when Data Center B encountered a similar failure, its agent directly applied the validated optimal solution using the federated updated model, reducing the failure handling time from approximately 2 hours in the traditional method to approximately 10 minutes.

[0083] Example 4 The agent proposed an evolutionary solution to improve query performance by "optimizing the database indexing strategy." The solution was first tested in a sandbox environment driven by the capability evolution audit module. The results showed a 28% improvement in query performance but a 15% decrease in write performance. This decrease in write performance triggered the "prohibition of operations during peak business periods" constraint and the potential risk of increased latency. The solution was rejected and required optimization. The agent then adjusted to a "time-sharing indexing strategy" and tested it again. After verifying that it met all security constraints, it was deployed in a phased, gradual, 50%, 50%, and 100% phase through the canary release process of the security boundary monitoring module, ultimately achieving a safe deployment.

[0084] Example 5 A service outage involved three agents: monitoring, diagnosis, and handling. The responsibility tracing module extracted the causal chain "monitoring delay, diagnostic misjudgment, handling error, service interruption" from the causal graph and mapped it to the corresponding agents. Then, based on the causal strength and confidence of each link in the causal chain, the following responsibility weights were calculated: Diagnosis agent W2 = 0.55 (primary responsibility), Monitoring agent W1 = 0.25 (secondary responsibility), and Handling agent W3 = 0.20 (secondary responsibility). Finally, a tracing report was generated, including a visualized causal chain, a responsibility weight table, and targeted improvement suggestions (such as enhancing the causal reasoning capabilities of the Diagnosis agent).

[0085] Example 6 Based on Embodiment 1, this embodiment provides a storage medium that stores a computer program or computer instructions. When the computer program or computer instructions are executed by the processor of a computer, the steps of the above-described operation and maintenance method based on causal inference and optimization are implemented.

[0086] The storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0087] Example 7 Based on Embodiment 1, this embodiment provides a computer, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they implement the steps of the above-described operation and maintenance method based on causal inference and optimization.

[0088] The memory can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or RAM. The memory can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the memory can include both internal storage units and external storage devices of any data processing device. The memory is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the concept and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An operation and maintenance method based on causal inference and optimization, characterized in that, Includes the following steps: Construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; Based on the causal knowledge graph, a privacy-preserving federated learning strategy is adopted to share causal knowledge and aggregate global models among multiple service nodes, thereby obtaining a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. In response to the capability optimization scheme proposed by the local agent based on the shared model, the capability optimization scheme is constrained and verified based on a preset safety boundary constraint, and the local agent is optimized using the verified capability optimization scheme. When an operational incident occurs, the causal path leading to the incident is extracted based on the causal knowledge graph. The responsibility weight of each intelligent agent participating in the execution on the causal path is calculated by combining the mapping relationship of the execution intelligent agents corresponding to each operational event node on the causal path, so as to complete the determination of responsibility for each intelligent agent.

2. The method according to claim 1, characterized in that, The PETER-CLARK algorithm is used to perform conditional independence tests and direction determination on the operation and maintenance event nodes in the local operation and maintenance event information, and the causal strength is calculated based on Do calculus to obtain the causal knowledge graph that supports intervention reasoning and counterfactual reasoning.

3. The method according to claim 2, characterized in that, The formula for calculating the strength of causality is: ； Where c(X→Y) represents the causal strength from event node X to event node Y, and P(Y|do(X=1)) and P(Y|do(X=0)) are the probabilities of event node Y occurring when event node X is forcibly intervened to make its value 1 and 0, respectively.

4. The method according to claim 2, characterized in that, The Fisher Z-transform is used to perform a conditional independence test on the operation and maintenance event nodes in the local operation and maintenance event information.

5. The method according to claim 1, characterized in that, A privacy-preserving federated learning strategy is employed to share causal knowledge and aggregate global models across multiple service nodes, resulting in a shared model that integrates multi-domain causal knowledge. This process includes the following steps: Deploy a federated learning architecture across multiple service nodes; Each service node's local agent trains a local model based on local operation and maintenance data; After the local model is trained, homomorphic encryption algorithm and differential privacy protection method are used to protect the model parameters of the local model for privacy. Based on the deployed federated learning architecture, the privacy-preserving model parameters are aggregated using a secure multi-party computation method to obtain a shared model that integrates multi-domain causal knowledge, and the shared model is published to each service node.

6. The method according to claim 5, characterized in that, The homomorphic encryption algorithm uses the Paillier encryption algorithm, and the encryption formula is: ； Where Enc(m) is the encrypted ciphertext, m is the plaintext, the encryption public key is (g, n), and r is a random number.

7. The method according to claim 5, characterized in that, Differential privacy protection methods involve adding Laplace noise to the encrypted model parameters.

8. The method according to claim 5, characterized in that, The local model is trained using the Transformer architecture, and its loss function L includes a causal consistency loss term L_causal, which satisfies: ; ； in, ; ; ; For the true value, For input The model parameters are The model's predicted values; Lprediction is the prediction loss, Lprivacy is the privacy loss, Lcausality is the causality loss, and Lregularity is the regularization loss. These are the original model parameters. Here are the privacy-preserving model parameters: clocal(X→Y) and cglobal(X→Y) represent the causal strength from event node X to event node Y in the local and global knowledge graphs, respectively; λ1 is the weight coefficient of the prediction loss, λ2 is the weight coefficient of the causal loss, λ3 is the weight coefficient of the privacy loss, and λ4 is the weight coefficient of the regularization loss; the local knowledge graph is a causal knowledge graph constructed by local service nodes based on local operation and maintenance event information; the global knowledge graph is obtained by sharing causal knowledge among multiple service nodes using a privacy-preserving federated learning strategy.

9. The method according to claim 5, characterized in that, During the release of the shared model, a sliding window mechanism is used to monitor operational anomalies in real time. When the number of anomalies within the sliding window exceeds a preset number, a rollback operation is triggered.

10. The method according to claim 9, characterized in that, The release method adopts a gray-scale release approach, and the release is carried out in multiple stages according to a preset ratio.

11. The method according to claim 1, characterized in that, Before uploading parameters, each service node performs knowledge anonymization processing on the original information. The anonymization rules include at least removing IP addresses, user accounts, and specific business plaintext data from the operation and maintenance data.

12. The method according to claim 1, characterized in that, The security boundary constraints include at least business-dimensional constraints, resource-dimensional constraints, and time-dimensional constraints; wherein, the business-dimensional constraints include at least service availability, maximum response latency, and failure rate threshold; the resource-dimensional constraints include at least CPU utilization, memory utilization, and network bandwidth limit; and the time-dimensional constraints include at least preset optimization operation windows and rules prohibiting operation during peak business periods.

13. The method according to claim 1, characterized in that, The process of local intelligent agents optimizing their capabilities generates audit logs, which are then written to the blockchain for irreversible evidence storage.

14. The method according to claim 1, characterized in that, The formula for calculating the responsibility weight of each agent is as follows: ； Where Wj is the responsibility weight of the j-th agent, ci is the causal strength of the i-th edge of the causal path, and pi is the confidence level of the i-th edge. M(vi) is the indicator function, and M(vi) is the operation execution agent corresponding to the operation and maintenance event node vi.

15. The method according to claim 1, characterized in that, The method for assigning responsibility to each agent is to determine the responsibility level by dividing the scope according to the responsibility weight of the agent.

16. A system for implementing the method according to any one of claims 1 to 15, characterized in that, include: The causal knowledge graph module is used to construct a causal knowledge graph carrying structured causal knowledge based on local operation and maintenance event information; The federated knowledge sharing module, connected to the causal knowledge graph module, is used to perform causal knowledge sharing and global model aggregation among multiple service nodes based on the causal knowledge graph and using a privacy-preserving federated learning strategy to obtain a shared model that integrates multi-domain causal knowledge, which is then used by the local agents of each service node to optimize their capabilities. The agent safety evolution module is connected to the federated knowledge sharing module and the causal knowledge graph module respectively. It is used to respond to the capability optimization scheme proposed by the local agent based on the sharing model, perform constraint verification on the capability optimization scheme based on the preset safety boundary constraints, and optimize the local agent using the verified capability optimization scheme. The multi-agent responsibility tracing module is connected to the causal knowledge graph module. When an operation and maintenance accident occurs, it extracts the causal path that led to the operation and maintenance accident based on the causal knowledge graph, and calculates the responsibility weight of each agent participating in the execution on the causal path by combining the mapping relationship of the execution agents corresponding to each operation and maintenance event node on the causal path, so as to complete the determination of responsibility for each agent.

17. A storage medium storing a computer program or computer instructions that, when executed by a processor of a computer, implement the steps of the method as claimed in claims 1 to 15.

Images