Network fault self-healing and prediction maintenance method based on artificial intelligence

Abstract

The invention relates to the technical field of network fault maintenance, in particular to a network fault self-healing and prediction maintenance method based on artificial intelligence, and the method comprises the steps: S1, constructing a multi-source data real-time collection framework; s2, deploying a lightweight A I model at an edge node; s3, introducing an interpretable AI technology; s4, constructing a causal reasoning module; s5, designing a dynamic self-healing strategy library; s6, establishing an online model learning mechanism; s7, developing a simulation verification environment; and S8, realizing a man-machine cooperative operation and maintenance workflow. According to the scheme, data is subjected to streaming preprocessing and intelligent dimension reduction at the source, the transmission load is greatly reduced, the processing efficiency is improved, real-time, near-real-time and batch processing tasks are further distinguished through the edge side parallel assembly line technology, it is ensured that key indexes are preferentially processed, and compression and acceleration are achieved on the model level through knowledge distillation, quantification and pruning technologies.

Description

Technical Field

[0001] This invention relates to the field of network fault maintenance technology, and in particular to a network fault self-healing and predictive maintenance method based on artificial intelligence. Background Technology

[0002] With the rapid development of information technology, modern communication networks have become a core component of social infrastructure. Their scale and complexity are increasing day by day. Frequent network failures not only affect user experience but may also cause huge economic losses and social problems. Therefore, the automation and intelligence of network operation and maintenance have become an important development direction for the industry.

[0003] In recent years, artificial intelligence technology has been widely applied in the field of network fault management, especially in fault detection, diagnosis, self-healing, and predictive maintenance. Methods based on machine learning and deep learning can automatically learn patterns from massive amounts of network data, achieving rapid anomaly detection and partial self-healing functions, significantly improving the efficiency and reliability of network operation and maintenance. According to the authorization announcement number "CN101945410A", a method for intelligent analysis of mobile packet network faults belongs to the field of mobile communication network maintenance and optimization technology. This method, based on the intelligent analysis of real-time alarm faults, daily network operation status, and weekly network operation status, comprehensively analyzes the causes of network faults, key concerns related to network faults, and solutions to network faults. Furthermore, it achieves proactive, accurate, and comprehensive intelligent analysis of network faults. This invention, through proactive analysis of network operation status at the time of alarm occurrence, at the end of each day, and at the end of each week, realizes a shift from passive to proactive network maintenance and optimization. The introduction of intelligent analysis technology during the analysis process improves the efficiency and accuracy of network maintenance and optimization; it is suitable for scenarios involving intelligent analysis of mobile network faults.

[0004] Currently, the following problems still exist in the application of artificial intelligence:

[0005] 1. Insufficient real-time performance and response speed is a prominent pain point. Although existing AI methods can achieve faster fault detection and response, with the explosive growth of network traffic and the stringent requirements of emerging applications for low latency and high reliability, current systems still cannot meet the millisecond-level response requirements.

[0006] 2. The low interpretability and transparency of artificial intelligence models seriously affect the trust and operational efficiency of operation and maintenance personnel. Most advanced deep learning models are regarded as black boxes, and their decision-making process is difficult to understand and verify. In network fault management scenarios, operation and maintenance administrators not only need to know whether the fault exists, but also need to clarify the root cause of the fault, the basis for self-healing actions, and potential risks. Lack of interpretability may lead to misjudgment, refusal to adopt artificial intelligence suggestions, or delay in processing time.

[0007] 3. Traditional AI methods rely too much on correlation and ignore causation, resulting in inaccurate root cause analysis of faults and limited effectiveness of maintenance strategies. Machine learning models usually learn statistical associations from data, but correlation is not the same as causation. In a network, multiple faults may occur simultaneously or present a chain reaction. Relying solely on correlation analysis may misjudge the root cause, thus leading to ineffective self-healing measures. Summary of the Invention

[0008] The purpose of this invention is to address the aforementioned shortcomings in the existing technology by proposing an artificial intelligence-based method for network fault self-healing and predictive maintenance.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a network fault self-healing and predictive maintenance method based on artificial intelligence, wherein the network fault self-healing and predictive maintenance method based on artificial intelligence includes:

[0010] S1. Construct a multi-source data real-time acquisition framework, integrate network device metrics and log streams, and support millisecond-level data injection and anomaly detection;

[0011] S2. Deploy lightweight AI models on edge nodes to achieve local real-time inference, reduce response latency and improve fault detection speed;

[0012] S3. Introduce explainable AI technology to generate visual reports on fault decision-making basis, enhancing the trust of operations and maintenance personnel in the model results;

[0013] S4. Construct a causal reasoning module, combine domain knowledge graph to distinguish between correlation and causal relationships, and accurately locate the root cause of the fault.

[0014] S5. Design a dynamic self-healing strategy library to automatically match repair actions based on fault type and confidence level, and support manual review and intervention;

[0015] S6. Establish an online model learning mechanism and use feedback data to continuously optimize the accuracy of fault prediction and root cause analysis.

[0016] Online incremental learning model update:

[0017]

[0018] in:

[0019] θ t These are the current model parameters;

[0020] D new This includes new feedback data, such as misclassified-corrected sample pairs and expert-annotated data.

[0021] L is the loss function;

[0022] η is the learning rate;

[0023] S7. Develop a simulation verification environment to conduct sand table simulations and effect evaluations of self-healing actions to avoid chain risks;

[0024] S8 enables human-machine collaborative operation and maintenance workflow, integrates early warning, diagnosis, decision-making and execution closed loop, and improves the overall reliability of the system.

[0025] Preferably, the handling of insufficient real-time performance and response speed in S1 and S2 is as follows:

[0026] A1. At the data level, streaming preprocessing and intelligent dimensionality reduction are introduced. Before the data is sent, data cleaning, format standardization and preliminary filtering are performed in a lightweight agent on network devices or edge nodes. Only valuable data is sent, which greatly reduces the network transmission load. Lightweight principal component analysis or autoencoder models are deployed on edge nodes to perform real-time dimensionality reduction on high-dimensional index data. Several key feature vectors are used to represent most of the information in the original data.

[0027] Real-time data dimensionality reduction based on principal component analysis is specifically as follows:

[0028] Z = X·W

[0029] in:

[0030] X: Original high-dimensional data matrix, n*d, n samples, d features;

[0031] W: Projection matrix obtained from principal component analysis, d*k, where k is the dimension after dimensionality reduction;

[0032] Z: The data matrix after dimensionality reduction, n*k;

[0033] Lightweight scoring for real-time anomaly detection based on S1, S2, and A1:

[0034]

[0035] in:

[0036] x i (t) represents the value of the i-th indicator at time t;

[0037] μ i , σ i These are the mean and standard deviation of the indicator within the sliding window;

[0038] w i These are the feature weights, the principal component weights retained after dimensionality reduction through principal component analysis;

[0039] k is the number of features after dimensionality reduction;

[0040] A2. At the computing level, edge clusters and parallel pipelines are used to form a lightweight Kubernetes or DockerSwarm cluster by combining multiple adjacent edge nodes. Through clustered management and scheduling of computing tasks, load balancing is achieved, serial processes are broken, and multiple parallel pipelines are designed.

[0041] A3. Deploy incremental learning and model slimming techniques at the model level;

[0042] A4. At the response level, predictive early warning and automated closed-loop action are achieved by utilizing time series forecasting models.

[0043] Preferably, the production line in A2 is specifically divided into:

[0044] A2.1 Real-time pipeline, handling the most critical metrics, including latency and packet loss rate, using a simplified model for millisecond-level anomaly detection;

[0045] A2.2 Near real-time pipeline for handling complex log analysis and multi-metric correlation analysis, outputting results in seconds;

[0046] A2.2 Batch processing pipeline performs background tasks such as data compression and model fine-tuning when edge nodes are idle.

[0047] Preferably, the model slimming technology in A3 specifically includes:

[0048] A3.1 Knowledge distillation involves training a large and accurate teacher model in the cloud, and then having it guide the learning of a lightweight student model, allowing the student model to maintain high accuracy while reducing the model size by several times.

[0049] A3.2 Model quantization converts model parameters from 32-bit floating-point numbers to 8-bit integers or even lower precision, which can significantly reduce model size and computational overhead, and improve inference speed.

[0050] A3.3 Model pruning involves removing unimportant weight parameters from the model to generate a sparse model, thereby reducing computational load.

[0051] Preferably, the processing in S3 and S8 for cases with low interpretability and transparency is as follows:

[0052] B1. Construct a dynamic, multimodal, interpretable report generator;

[0053] B2. Establish a collaborative diagnosis and feedback mechanism for people in the loop;

[0054] B3. Develop a transparent simulation and deduction environment based on digital twins.

[0055] Preferably, the multi-interpretation method in B2 and the further processing of increasing the learning sample library are as follows:

[0056] B2.1 Constructing a dynamic adaptive confidence calibration and multi-interpretation method arbitration mechanism;

[0057] B2.2 Develop a closed-loop feedback-driven incremental learning and expert knowledge embedding framework.

[0058] Preferably, the optimization of the digital twin environment and the glass box in B3 specifically includes:

[0059] B3.1 Constructing a dynamic adaptive twin framework and an interactive glass box interpretation layer;

[0060] B3.2 Introducing a multi-agent collaborative game and incremental learning mechanism.

[0061] Preferably, the handling of misjudgment risk in S4, S6, and S7 is as follows:

[0062] C1. Construct a layer for quantifying and assessing the risk of misjudgment;

[0063] C2. Implement a risk-level-based tiered decision-making mechanism, setting three risk level thresholds: high, medium, and low. Tiered decision-making is divided into low risk - high confidence, medium risk - medium confidence, and high risk - low confidence.

[0064] C3. Online risk model learning with integrated feedback data;

[0065] C4. Conduct targeted stress tests for misjudgment scenarios in a simulation environment. Transform the misjudgment cases collected in C3 into executable simulation test cases. In the simulation verification environment, not only are real faults injected, but also interference signals and noise data that may lead to misjudgment are deliberately injected. Observe whether the output confidence of the causal reasoning module is reasonable and whether the decision-making mechanism is correctly triggered when the system faces these traps, and record its performance.

[0066] C5. Establish a full-chain traceability and closed-loop management system for handling misjudgments.

[0067] Preferably, the hierarchical decision in C2 is specifically as follows:

[0068] C2.1 Low risk - high confidence: The conclusion is reliable, allowing direct entry into the self-healing process;

[0069] C2.2, Medium Risk - Medium Confidence: Initiate an enhanced verification process, trigger additional diagnostic probes, cross-validate other monitoring data sources, or initiate a lightweight simulation to obtain more information to improve confidence.

[0070] C2.3 High Risk - Low Confidence: Immediately triggers a manual review alarm, and the system takes the most conservative containment action to prevent the fault from escalating due to erroneous operation.

[0071] Preferably, the manual review and intervention in S5, based on a dynamic self-healing strategy and handling through human-machine collaboration, is specifically as follows:

[0072] S5.1 The system receives fault events from the monitoring platform in real time and automatically identifies the fault type and calculates its confidence level through the built-in root cause analysis engine or AI model.

[0073] S5.2 The dynamic strategy library performs multi-dimensional matching based on the diagnostic results, automatically filters and recommends the optimal repair solution, and the matching logic follows the preset rules C2.1-C2.3;

[0074] S5.3 provides a unified decision-making interface for the operations and maintenance team, which centrally displays the remediation plan to be reviewed, related indicators, scope of impact and prediction results. For complex or high-risk operations, simulation exercises can be conducted in the sandbox environment first to verify effectiveness before going online.

[0075] S5.4 All automated actions are recorded in detail, and the execution status and repair effects are fed back in real time. The results data will be fed back to the strategy knowledge base as the basis for strategy effectiveness evaluation and model optimization, forming a continuous learning loop from practice to knowledge.

[0076] The design scheme proposed in this invention has the following beneficial effects during application:

[0077] 1. This solution significantly reduces transmission load and improves processing efficiency by performing streaming preprocessing and intelligent dimensionality reduction on data at the source. The parallel pipeline technology on the edge further distinguishes between real-time, near real-time and batch processing tasks to ensure that key indicators are prioritized. At the model level, knowledge distillation, quantization and pruning techniques are used to achieve compression and acceleration, maintaining high-precision inference capabilities even in resource-constrained environments. The overall system significantly reduces fault response latency through predictive early warning and automated closed-loop action, effectively improving the continuity and reliability of network services.

[0078] 2. By introducing explainable artificial intelligence technology and a dynamic multimodal report generation mechanism, complex model decisions are transformed into visualized fault diagnosis data, significantly enhancing the trust of maintenance personnel in the automated system. By constructing a human-in-the-loop collaborative feedback framework, the system can integrate expert experience and real-time operational feedback, continuously calibrate confidence and optimize the decision-making process. Combined with a digital twin simulation environment, users can transparently deduce the fault handling process and self-healing effects, achieving system behavior insight from black box to glass box, thereby supporting more reliable human-machine collaborative decision-making. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the network fault self-healing and predictive maintenance process of the present invention;

[0080] Figure 2 This is a schematic diagram of the layered processing of the A2 pipeline architecture of the present invention;

[0081] Figure 3 This is a treatment that results in lower interpretability and transparency of the present invention. Detailed Implementation

[0082] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0083] Example 1

[0084] Reference Figures 1-3 A network fault self-healing and predictive maintenance method based on artificial intelligence, comprising:

[0085] S1. Construct a multi-source data real-time acquisition framework, integrate network device metrics and log streams, and support millisecond-level data injection and anomaly detection;

[0086] S2. Deploy lightweight AI models on edge nodes to achieve local real-time inference, reduce response latency and improve fault detection speed;

[0087] S3. Introduce explainable AI technology to generate visual reports on fault decision-making basis, enhancing the trust of operations and maintenance personnel in the model results;

[0088] S4. Construct a causal reasoning module, combine domain knowledge graph to distinguish between correlation and causal relationships, and accurately locate the root cause of the fault.

[0089] S5. Design a dynamic self-healing strategy library to automatically match repair actions based on fault type and confidence level, and support manual review and intervention;

[0090] S6. Establish an online model learning mechanism and use feedback data to continuously optimize the accuracy of fault prediction and root cause analysis.

[0091] Online incremental learning model update:

[0092]

[0093] in:

[0094] θ t These are the current model parameters;

[0095] D newThis includes new feedback data, such as misclassified-corrected sample pairs and expert-annotated data.

[0096] L is the loss function;

[0097] η is the learning rate.

[0098] S7. Develop a simulation verification environment to conduct sand table simulations and effect evaluations of self-healing actions to avoid chain risks;

[0099] S8 enables human-machine collaborative operation and maintenance workflow, integrates early warning, diagnosis, decision-making and execution closed loop, and improves the overall reliability of the system.

[0100] The specific handling of insufficient real-time performance and response speed in S1 and S2 is as follows:

[0101] A1. At the data level, streaming preprocessing and intelligent dimensionality reduction are introduced. Before the data is sent, data cleaning, format standardization and preliminary filtering are performed in a lightweight agent on network devices or edge nodes. Only valuable data is sent, which greatly reduces the network transmission load. Lightweight principal component analysis or autoencoder models are deployed on edge nodes to perform real-time dimensionality reduction on high-dimensional index data. Several key feature vectors are used to represent most of the information in the original data.

[0102] Real-time data dimensionality reduction based on principal component analysis is specifically as follows:

[0103] Z = X·W

[0104] in:

[0105] X: Original high-dimensional data matrix, n*d, n samples, d features;

[0106] W: Projection matrix obtained from principal component analysis, d*k, where k is the dimension after dimensionality reduction;

[0107] Z: The data matrix after dimensionality reduction, n*k;

[0108] Lightweight scoring for real-time anomaly detection based on S1, S2, and A1:

[0109]

[0110] in:

[0111] x i (t) represents the value of the i-th indicator at time t;

[0112] μ i , σ i These are the mean and standard deviation of the indicator within the sliding window;

[0113] w iThese are the feature weights, the principal component weights retained after dimensionality reduction through principal component analysis;

[0114] k is the number of features after dimensionality reduction.

[0115] A2. At the computing level, edge clusters and parallel pipelines are used to form a lightweight Kubernetes or DockerSwarm cluster by combining multiple adjacent edge nodes. Through clustered management and scheduling of computing tasks, load balancing is achieved, serial processes are broken, and multiple parallel pipelines are designed.

[0116] Based on the performance optimization formula:

[0117] E = N * P * L

[0118] in:

[0119] E: Total system computational probability;

[0120] N: Number of edge nodes;

[0121] P: Number of parallel pipelines;

[0122] L: Load balancing factor.

[0123] A3. Deploy incremental learning and model slimming techniques at the model level. The deployed lightweight AI model can utilize newly incoming local data for online fine-tuning and incremental learning. It applies extreme model slimming techniques, which are divided into knowledge distillation, model quantization, and model pruning.

[0124] A4. At the response level, predictive early warning and automated closed-loop action are achieved. Using time series prediction models, performance trends for a future period are predicted based on historical indicator data. Predictive early warnings are issued in advance when indicators are about to reach the threshold but the fault has not yet occurred. Automated scripts for specific prediction or detection results are preset at edge nodes.

[0125] In A4, for example:

[0126] If it is predicted that a certain port will soon be full of traffic, the traffic scheduling strategy will be automatically triggered to switch some traffic to the backup link.

[0127] Abnormal login behavior detected → Automatically triggers command to temporarily block IP address.

[0128] The pipeline in A2 is specifically divided into:

[0129] A2.1 Real-time pipeline, handling the most critical metrics, including latency and packet loss rate, using a simplified model for millisecond-level anomaly detection;

[0130] A2.2 Near real-time pipeline for handling complex log analysis and multi-metric correlation analysis, outputting results in seconds;

[0131] A2.2 Batch processing pipeline performs background tasks such as data compression and model fine-tuning when edge nodes are idle.

[0132] The model slimming technology in A3 specifically includes:

[0133] A3.1 Knowledge distillation involves training a large and accurate teacher model in the cloud, and then having it guide the learning of a lightweight student model, allowing the student model to maintain high accuracy while reducing the model size by several times.

[0134] A3.2 Model quantization converts model parameters from 32-bit floating-point numbers to 8-bit integers or even lower precision, which can significantly reduce model size and computational overhead, and improve inference speed.

[0135] A3.3 Model pruning involves removing unimportant weight parameters from the model to generate a sparse model, thereby reducing computational load.

[0136] The specific handling of low interpretability and transparency in S3 and S8 is as follows:

[0137] B1. Construct a dynamic multimodal interpretability report generator that no longer relies on a single interpretability method. Instead, it uses attribution analysis, counterfactual interpretation, and key sample comparison in parallel for the same fault prediction result to cross-validate the rationality of the model's decision from different perspectives. The generated visualization report is no longer a static PDF or image, but an interactive dashboard. It uses a large language model to automatically transform complex feature weights and model logic into a concise and easy-to-understand natural language conclusion.

[0138] B2. Establish a collaborative diagnosis and feedback mechanism for human-in-the-loop systems. The system not only outputs prediction results but also confidence scores. When the confidence score is lower than a preset threshold or there is a significant contradiction with the conclusions of multiple interpretation methods, the system automatically changes the warning to a pending review status and actively pushes it to the operation and maintenance expert interface to request manual judgment. A convenient feedback button is provided in the interactive report interface. The feedback from operation and maintenance experts is recorded in real time. The collected feedback data forms a high-quality incremental learning sample library, which is used for model fine-tuning and retraining on a regular or triggered basis.

[0139] B3. Develop a transparent simulation and deduction environment based on digital twins. Based on historical data and real-time data streams, construct a high-fidelity digital twin environment for key system components. When the AI ​​model generates a fault prediction or handling suggestion, automatically inject the same initial conditions into the digital twin environment and run two schemes in parallel. One is to execute according to the decision suggested by the AI, and the other is to execute according to the alternative scheme proposed by the operation and maintenance personnel. This transforms the model from a black box into a glass box.

[0140] The multi-interpretation method in B2 and the further processing of adding a learning sample library are as follows:

[0141] B2.1 Construct a dynamic adaptive confidence calibration and multi-interpretation method arbitration mechanism, introduce a dynamic confidence threshold adjustment mechanism, optimize the threshold setting in real time based on historical feedback data, avoid misjudgment or missed detection caused by fixed threshold, design a multi-interpretation method arbitration module, and integrate locally interpretable and globally interpretable methods.

[0142] B2.2 Develop a closed-loop feedback-driven incremental learning and expert knowledge embedding framework, design a lightweight real-time feedback interface, support operation and maintenance experts to submit feedback in a multimodal way such as drag and drop, annotation, and text annotation, and automatically associate it with the original prediction context. After denoising and enhancement, the feedback data is synchronized to the incremental learning sample library and labeled with expert knowledge tags.

[0143] The digital twin environment and glass box optimizations in B3 are as follows:

[0144] B3.1 Construct a dynamic adaptive twin framework and an interactive glass box interpretation layer. Based on the existing digital twin, introduce dynamic adaptive modeling technology to enable the twin environment to automatically adjust model parameters and boundary conditions according to real-time data streams, ensuring high fidelity and evolution accuracy. Embed the interactive glass box interpretation layer into the AI ​​model. When the AI ​​proposes fault prediction or handling suggestions, the system not only automatically injects initial conditions for parallel deduction, but also visualizes the AI ​​decision-making logic in real time, and allows operation and maintenance personnel to question assumptions, adjust parameters, or inject adversarial scenarios through natural language interaction.

[0145] B3.2 Introducing a multi-agent collaborative game and incremental learning mechanism, a multi-agent system is introduced into the parallel simulation to simulate the collaborative and game process of multiple intelligent agents such as AI, operation and maintenance personnel, system components, and environment. An incremental learning closed loop is constructed to simulate differential data, automatically trigger model fine-tuning, and generate synthetic data to enhance the training set, enabling AI to continuously optimize the decision boundary.

[0146] The handling of misjudgment risk in S4, S6, and S7 is as follows:

[0147] C1. Construct a layer for quantifying the risk of misjudgment and assessing confidence. This requires the causal reasoning model to output not only the root cause node but also a confidence score. Combined with the domain knowledge graph, conduct an impact analysis on each possible misjudgment scenario.

[0148] C2. Implement a risk-level-based tiered decision-making mechanism, setting three risk level thresholds: high, medium, and low. Tiered decision-making is divided into low risk - high confidence, medium risk - medium confidence, and high risk - low confidence.

[0149]

[0150] in:

[0151] The confidence score of the model output is between 0 and 1;

[0152] T h This is a high-risk threshold;

[0153] T l This is the low-risk threshold.

[0154] C3. Online risk model learning with integrated feedback data: systematically collect all diagnostic results that have been overturned after manual review or simulation verification, as well as the finally confirmed correct root causes, to form misjudgment-correction sample pairs, analyze the common characteristics of these misjudgment samples, and regularly use these high-quality negative samples to incrementally train or fine-tune the causal inference model and prediction model.

[0155] C4. Conduct targeted stress tests for misjudgment scenarios in a simulation environment. Transform the misjudgment cases collected in C3 into executable simulation test cases. In the simulation verification environment, not only are real faults injected, but also interference signals and noise data that may lead to misjudgment are deliberately injected. Observe whether the output confidence of the causal reasoning module is reasonable and whether the decision-making mechanism is correctly triggered when the system faces these traps, and record its performance.

[0156] C5. Establish a full-chain traceability and closed-loop management system for handling misjudgments. Record detailed logs of the entire process of each diagnosis, decision, and action, especially key information such as confidence calculation, risk level judgment, and decision rationale. Whenever a misjudgment requiring manual intervention occurs, the system automatically generates a tracking work order. Regularly review all misjudgment work orders, analyze the root causes of misjudgments, and generate optimization tasks to drive the iteration of knowledge graphs, models, and processes.

[0157] The hierarchical decision-making in C2 is specifically as follows:

[0158] C2.1 Low risk - high confidence: The conclusion is reliable, allowing direct entry into the self-healing process;

[0159] C2.2, Medium Risk - Medium Confidence: Initiate an enhanced verification process, trigger additional diagnostic probes, cross-validate other monitoring data sources, or initiate a lightweight simulation to obtain more information to improve confidence.

[0160] C2.3 High Risk - Low Confidence: Immediately triggers a manual review alarm, and the system takes the most conservative containment action to prevent the fault from escalating due to erroneous operation.

[0161] The manual review and intervention in S5, along with the dynamic self-healing strategy and handling based on human-machine collaboration, are detailed below:

[0162] S5.1 The system receives fault events from the monitoring platform in real time and automatically identifies the fault type and calculates its confidence level through the built-in root cause analysis engine or AI model.

[0163] F(x) = (T, C) | x ∈ E

[0164] in:

[0165] F represents the entire root cause analysis function or system;

[0166] x is the input, which is a single fault event from the monitoring platform. x∈E means that the input x belongs to the set of all possible events.

[0167] T is one of the outputs, namely the fault type obtained after system analysis;

[0168] C is the second output, which is the confidence level of the fault type determination result, usually a value between 0 and 1;

[0169] | indicates either given or based on, meaning the output is generated based on the input x;

[0170] S5.2 The dynamic strategy library performs multi-dimensional matching based on the diagnostic results, automatically filters and recommends the optimal repair solution, and the matching logic follows the preset rules C2.1-C2.3;

[0171] S5.3 provides a unified decision-making interface for the operations and maintenance team, which centrally displays the remediation plan to be reviewed, related indicators, scope of impact and prediction results. For complex or high-risk operations, simulation exercises can be conducted in the sandbox environment first to verify effectiveness before going online.

[0172] S5.4 All automated actions are recorded in detail, and the execution status and repair effects are fed back in real time. The results data will be fed back to the strategy knowledge base as the basis for strategy effectiveness evaluation and model optimization, forming a continuous learning loop from practice to knowledge.

[0173] Example 2

[0174] Intelligent handling of network failures on large e-commerce platforms

[0175] Scenario: During the "Double Eleven" shopping festival, the network of a major e-commerce platform's core data center carries millions of transaction requests per second. Any slight network jitter could trigger an avalanche effect, resulting in huge economic losses and damage to brand reputation.

[0176] Application process:

[0177] Real-time detection and localization (S1, S2, S4, A1-A3): Data acquisition (S1, A1): Lightweight agents deployed on switches and routers collect millisecond-level indicators such as port traffic, CPU utilization, and packet loss rate in real time, and perform streaming cleaning and PCA-based dimensionality reduction processing on them, sending only key feature vectors to edge AI nodes.

[0178] Edge Inference (S2, A2): After the "real-time pipeline" in the edge Kubernetes cluster receives data, the lightweight distillation model (A3) immediately calculates the anomaly score Score_anomaly. The system detects that there is a microsecond-level latency spike and slight packet loss in the link between a core switch and the database cluster.

[0179] Causal Localization (S4): The causal reasoning module is activated. Combining the "Network Device Topology Knowledge Graph", the analysis reveals that the problem is not that the switch itself is faulty, but that a downstream server is sending useless broadcast packets due to a software bug, causing port congestion. The system accurately locates the root cause of the fault as this server.

[0180] Decision and self-healing (S5, C2): The system determines that this fault is high-risk (may affect the payment process), but the root cause analysis confidence level is medium. According to the hierarchical decision mechanism (C2.2), the enhanced verification process is triggered.

[0181] The system automatically sends out diagnostic probes to directly detect the network connection status of the target server, confirms the source of abnormal traffic, raises the confidence level to high confidence, and adjusts the risk level to medium risk;

[0182] The dynamic policy library (S5) automatically matches and repairs actions: ① The solution of "isolating the problematic server and migrating its traffic to a healthy node" is verified to be effective and risk-free in the simulation environment (S7); ② The solution is pushed to the operation and maintenance personnel's console for final approval;

[0183] Human-computer collaboration and learning (S3, S8, S6, S5.3-S5.4):

[0184] The operations and maintenance personnel received a notification that they could view the explainable AI report (S3): it highlighted the time sequence diagram, cause-effect path diagram and simulation results of abnormal traffic, which enhanced the trust level;

[0185] When the user clicks "Approve Execution," the system automatically calls the API to remove the server from the load balancer pool and triggers a restart and repair script.

[0186] After successful execution, network metrics returned to normal. The entire process, from discovery to resolution, took only seconds, avoiding a major failure. The processed sample data was fed back to the online learning mechanism (S6) for continuous model optimization.

[0187] Example 3

[0188] Predictive Maintenance via IoT in Smart Manufacturing Plants

[0189] Scenario: A highly automated smart factory relies on thousands of industrial IoT devices to work collaboratively via wireless networks. A network outage will directly cause the production line to stop.

[0190] Application process: Predictive early warning (S1, S2, A4): The model deployed on the edge gateway of the factory continuously analyzes time series data such as signal strength, AP access point connection success rate, and retransmission rate of the AGV navigation system (S1);

[0191] Using the time series prediction model (A4), it was found that the connection success rate of a certain AP in the southwest area of ​​Plant 3 showed a continuous and slow downward trend, predicting that it would fall below the reliable working threshold after 24 hours. The system issued a predictive warning instead of a post-event alarm.

[0192] Root cause analysis and strategy matching (S4, S5): The causal reasoning module (S4) analyzes the data by combining the "plant map and equipment deployment knowledge graph". It eliminates the possibility of AP hardware failure and infers that the root cause may be signal obstruction caused by the newly added large metal rack or interference caused by the addition of equipment in the same frequency band.

[0193] The dynamic policy library (S5) matched multiple suggestions: ① It suggested that the operation and maintenance personnel go to the site to check if there are any new obstacles; ② It provided automated scripts for adjusting the AP antenna angle or channel parameters; ③ It recommended the optimal long-term solution of "adding a Mesh node";

[0194] Simulation and Human-Machine Collaboration (S7, S8, B3): The maintenance personnel chose to perform the low-risk operation of "adjusting the channel" first, but in order to avoid the operation from affecting other areas, he first conducted a simulation in the digital twin simulation environment (S7, B3).

[0195] The digital twin system simulated the wireless radio frequency environment of the entire factory area. After running the script to adjust the channel, it visually showed that the signal coverage in the area was improved without causing interference to the surrounding area.

[0196] Execution and Learning (S6, S5.4): After personnel confirmation, the system automatically executes the script, the AP switches to a better channel, and subsequent monitoring confirms that the connection success rate trend has stopped declining and started to rise.

[0197] The closed-loop data of the entire "prediction-diagnosis-simulation-execution-verification" process is recorded (S5.4) and used to update the incremental learning model (S6), making the system's prediction and root cause analysis of similar problems more accurate in the future.

[0198] In practice

[0199] The principle of multi-source data real-time acquisition and edge intelligent processing: This method constructs a high-throughput, low-latency data acquisition framework, integrates network device indicators and log streams, and achieves millisecond-level data injection and anomaly detection. At the data level, it adopts streaming preprocessing and intelligent dimensionality reduction technology, deploys lightweight agents on network devices or edge nodes to perform data cleaning, format standardization, and preliminary filtering, and transmits only high-value data, significantly reducing network load. Principal component analysis or autoencoder is used to perform real-time dimensionality reduction of high-dimensional indicators, extract key feature vectors, and retain the core information of the original data. At the computing level, multiple edge nodes are combined into a lightweight cluster through edge cluster and parallel pipeline technology to achieve load balancing and task scheduling. Three types of pipelines are designed: real-time, near real-time, and batch processing, to handle millisecond-level anomaly judgment, second-level multi-indicator correlation analysis, and background model optimization tasks, respectively, thereby comprehensively improving the system response speed and resource utilization efficiency.

[0200] Explainable AI and Collaborative Decision Enhancement Principles: The system introduces explainable AI technology to generate visualized reports on fault decision-making basis. Through a dynamic multimodal report generator, it integrates feature contribution, decision path, and semantic interpretation to enhance the trust of maintenance personnel in the model results. At the same time, it establishes a human-in-the-loop collaborative diagnosis and feedback mechanism, and constructs a dynamic adaptive confidence calibration and multi-interpretation method arbitration mechanism to ensure the reliability of the output conclusions. By developing a transparent simulation and deduction environment based on digital twins, it constructs a dynamic adaptive twin framework and an interactive glass box interpretation layer to support maintenance personnel in intuitively understanding the system's reasoning process and conducting interactive verification. Furthermore, it introduces a multi-agent collaborative game and incremental learning mechanism, enabling the system to continuously absorb expert feedback, optimize decision logic, and form a highly transparent human-machine collaborative maintenance workflow.

[0201] Causal Reasoning and Risk Layering Control Principles: The system constructs a causal reasoning module, combines domain knowledge graphs to distinguish between correlation and causal relationships, accurately locates the root cause of the fault, and avoids misjudgment. Through misjudgment risk quantification and confidence assessment layers, uncertainty is measured for each diagnostic result, and a risk-level-based layered decision-making mechanism is implemented. Based on high, medium, and low risk thresholds, automated repair, enhanced verification, or manual review processes are triggered respectively: low-risk, high-confidence results directly execute self-healing actions; medium-risk, medium-confidence results initiate additional diagnostic probes or cross-validation; high-risk, low-confidence results immediately trigger alarms and take conservative containment measures. An online risk model learning mechanism integrating feedback data is used to continuously optimize judgment accuracy using misjudgment-correction samples, and targeted stress tests are conducted in a simulation environment. Interference data is injected to verify the robustness of the system. Finally, a full-link traceability and closed-loop management mechanism is established to ensure the reliability and safety of fault handling.

[0202] Dynamic self-healing and continuous optimization closed-loop principle: The system is designed with a dynamic self-healing strategy library, which automatically matches repair actions based on fault type and confidence level, and supports manual review and intervention. During the operation and maintenance process, the system receives fault events in real time, identifies fault types and calculates confidence levels through the root cause analysis engine, and recommends the optimal repair solution after multi-dimensional matching by the dynamic strategy library. The repair solution, impact scope and prediction results are displayed to the operation and maintenance team through a unified decision interface. It supports simulation of complex or high-risk operations in a sandbox environment. All actions are recorded and the effects are fed back in real time. The result data flows back to the strategy knowledge base, forming a continuous learning closed loop from practice to knowledge. At the same time, the system establishes an online incremental learning mechanism, using feedback data such as misjudgment-correction samples and expert-annotated data to continuously improve model parameters through loss function optimization and learning rate adjustment, thereby improving the accuracy of fault prediction and root cause analysis, and realizing the continuous enhancement of system self-evolution and operation and maintenance capabilities.

[0203] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A network fault self-healing and predictive maintenance method based on artificial intelligence, characterized in that: The AI-based network fault self-healing and predictive maintenance method includes: S1. Construct a multi-source data real-time acquisition framework and integrate network device metrics and log streams; S2. Deploy lightweight AI models on edge nodes to achieve localized real-time inference; S3. Introduce explainable AI technology to generate visual reports that provide a basis for fault decision-making; S4. Construct a causal reasoning module, combine domain knowledge graph to distinguish between correlation and causal relationships, and accurately locate the root cause of the fault. S5. Design a dynamic self-healing strategy library to automatically match repair actions based on fault type and confidence level; S6. Establish an online model learning mechanism and use feedback data to continuously optimize the accuracy of fault prediction and root cause analysis. S7. Develop a simulation verification environment to conduct sand table simulation and effect evaluation of self-healing actions; S8 enables human-machine collaborative operation and maintenance workflow, integrating early warning, diagnosis, decision-making and execution closed loop.

2. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 1, characterized in that: The methods for addressing insufficient real-time performance and response speed in S1 and S2 include: A1. At the data level, streaming preprocessing and intelligent dimensionality reduction are introduced. Before the data is sent, data cleaning, format standardization and preliminary filtering are performed in a lightweight agent on network devices or edge nodes. Lightweight principal component analysis or autoencoder models are deployed on edge nodes to perform real-time dimensionality reduction on high-dimensional index data. Real-time data dimensionality reduction based on principal component analysis is specifically as follows: Z = X·W in: X: Original high-dimensional data matrix, n*d, n samples, d features; W: Projection matrix obtained from principal component analysis, d*k, where k is the dimension after dimensionality reduction; Z: The data matrix after dimensionality reduction, n*k; Lightweight scoring for real-time anomaly detection based on S1, S2, and A1: in: x i (t) represents the value of the i-th indicator at time t; μ i , σ i These are the mean and standard deviation of the indicator within the sliding window; w i These are the feature weights, the principal component weights retained after dimensionality reduction through principal component analysis; k is the number of features after dimensionality reduction; A2. At the computing level, edge clustering and parallel pipeline technology are adopted to form a lightweight Kubernetes or DockerSwarm cluster by combining multiple adjacent edge nodes. Through clustered management and scheduling of computing tasks, multiple parallel pipelines are designed. A3. Deploy incremental learning and model slimming at the model level; A4. At the response level, predictive early warning and automated closed-loop action are achieved by utilizing time series forecasting models.

3. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 2, characterized in that: The production line in A2 is specifically divided into: A2.1 Real-time pipeline, handling the most critical metrics, including latency and packet loss rate, using a simplified model for millisecond-level anomaly detection; A2.2 Near real-time pipeline for handling complex log analysis and multi-metric correlation analysis, outputting results in seconds; A2.2 Batch processing pipeline: Process background tasks when edge nodes are idle.

4. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 3, characterized in that: The methods for slimming down the model in A3 include: A3.1 Knowledge distillation to reduce model size; A3.2 Model quantization converts model parameters from 32-bit floating-point numbers to 8-bit integers to reduce model size and computational overhead; A3.3 Model pruning: Remove unimportant weight parameters from the model to generate a sparse model.

5. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 4, characterized in that: In S3 and S8, the methods for handling low interpretability and transparency include: B1. Construct a dynamic, multimodal, interpretable report generator; B2. Establish a collaborative diagnosis and feedback mechanism for people in the loop; B3. Develop a transparent simulation and deduction environment based on digital twins.

6. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 5, characterized in that: The multiple interpretation method in B2 and the further processing methods for increasing the learning sample library include: B2.1 Constructing a dynamic adaptive confidence calibration and multi-interpretation method arbitration mechanism; B2.2 Develop a closed-loop feedback-driven incremental learning and expert knowledge embedding framework.

7. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 6, characterized in that: In B3, the optimization methods for the digital twin environment and the glass box include: B3.1 Constructing a dynamic adaptive twin framework and an interactive glass box interpretation layer; B3.2 Introducing a multi-agent collaborative game and incremental learning mechanism.

8. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 7, characterized in that: In step S6, the online incremental learning model is updated: in: θ t These are the current model parameters; D new Added feedback data; L is the loss function; η is the learning rate.

9. The network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 8, characterized in that: The methods for handling the risk of misjudgment in S4, S6, and S7 include: C1. Construct a layer for quantifying and assessing the risk of misjudgment; C2. Implement a risk-level-based tiered decision-making mechanism, setting three risk level thresholds: high, medium, and low. Tiered decision-making is divided into low risk - high confidence, medium risk - medium confidence, and high risk - low confidence. C3. Online risk model learning with integrated feedback data; C4. Conduct targeted stress tests for misjudgment scenarios in a simulation environment; C5. Establish a full-chain traceability and closed-loop management system for handling misjudgments; The hierarchical decision-making method in C2 includes: C2.1 Low risk - high confidence: The conclusion is reliable, allowing direct entry into the self-healing process; C2.2, Medium Risk - Medium Confidence: Initiate an enhanced verification process, trigger additional diagnostic probes, cross-validate other monitoring data sources, or initiate a lightweight simulation to obtain more information to improve confidence. C2.3 High Risk - Low Confidence: Immediately triggers a manual review alarm, and the system takes the most conservative containment action to prevent the fault from escalating due to erroneous operation.

10. A network fault self-healing and predictive maintenance method based on artificial intelligence according to claim 9, characterized in that: The manual review and intervention in S5, based on a dynamic self-healing strategy and handling through human-machine collaboration, includes the following steps: S5.1 The system receives fault events from the monitoring platform in real time and automatically identifies the fault type and calculates its confidence level through the built-in root cause analysis engine or AI model. S5.2 The dynamic strategy library performs multi-dimensional matching based on the diagnostic results, automatically filters and recommends the optimal repair solution, and the matching logic follows the preset rules C2.1-C2.3; S5.3 Provides a unified decision-making interface, which centrally displays the remediation plan to be reviewed, related indicators, scope of impact and prediction results; S5.4 Record all automated actions and provide real-time feedback on execution status and repair effects. Feed the results back to the strategy knowledge base as a basis for strategy effectiveness evaluation and model optimization, forming a closed loop.

Images