Artificial intelligence-based optical communication fault detection method and system

Abstract

The present application relates to the field of optical communication technology, and more particularly to an optical communication fault detection method and system based on artificial intelligence, comprising multi-source data acquisition and management, small sample time series prediction model construction and edge early warning, digital twin driven maintenance strategy simulation and optimization, and decision execution and closed-loop verification; the present application deploys a lightweight time series prediction model based on small sample learning, automatically learns the microscopic degradation mode of device performance from high-frequency time series data such as power and bit error rate, and thus can issue a warning within hours to days before the hardware completely fails; the method realizes a fundamental change from passive alarm triggered by threshold to active warning based on trend prediction, wins valuable time for preventive operations such as replacing components and adjusting configurations, and effectively avoids business interruption caused by sudden failure.

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to an optical communication fault detection method and system based on artificial intelligence. Background Technology

[0002] With the continuous growth in the scale and complexity of optical communication networks, their reliability and operational efficiency have become critical challenges. Traditional network operation and maintenance heavily rely on pre-set static performance threshold alarms and manual judgment, which constitutes the current mainstream maintenance paradigm. Under this paradigm, when indicators such as optical power and bit error rate exceed fixed thresholds, the system triggers alarms, and then maintenance personnel intervene to diagnose and repair them. This method is essentially a passive response mode, and its effectiveness depends on the rationality of the thresholds and the experience level of the engineers, aiming to quickly locate and resolve faults that have occurred or are about to occur.

[0003] However, existing technologies have significant shortcomings. First, alarm mechanisms based on fixed thresholds suffer from severe lag, only issuing alerts when device performance has significantly deteriorated or even failed, failing to provide early warnings of slow performance degradation and missing valuable preventative maintenance window opportunities. Second, when faced with multiple concurrent alerts, existing network management systems lack the ability to globally simulate and quantitatively assess the overall network status and potential cascading effects of faults. Operational decisions are often based on local information and experience, easily leading to resource scheduling conflicts or secondary faults, making it difficult to balance service level agreements with optimized maintenance costs. Finally, the data flow throughout the entire operation and maintenance process is unidirectional and open-loop. The alerting model and decision-making logic cannot obtain feedback from historical execution results and perform self-optimization, resulting in stagnant system intelligence and limited long-term improvement in operational efficiency.

[0004] This invention aims to address the aforementioned problems by proposing an artificial intelligence-based optical communication fault detection method and system. By deploying a small-sample time-series prediction model at the network edge, it achieves early and accurate warnings of the slow degradation of optical device performance. Through the construction of a high-fidelity digital twin in the cloud, it conducts scenario-based simulations and quantitative evaluations of different maintenance strategies under multiple warning scenarios, providing data support for globally optimal decision-making. Finally, through a closed-loop feedback mechanism, it continuously optimizes the warning model and simulation model using actual operation and maintenance results, forming a self-learning and continuously evolving intelligent operation and maintenance system. This elevates optical network operation and maintenance from passive response to a new stage of proactive prediction and global optimization. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes an optical communication fault detection method and system based on artificial intelligence.

[0006] The technical solution of this invention is: an optical communication fault detection method based on artificial intelligence, comprising the following steps:

[0007] S11: Multi-source data acquisition and governance, which collects runtime sequence data, event log data and related data from multiple network nodes in the optical communication network, and performs standardization, alignment and labeling governance on the data to build a health status profile of the device;

[0008] S12: Small sample time series prediction model construction and edge early warning. Based on the treated data, a lightweight time series prediction model is trained and deployed at the network edge to evaluate the health status of single nodes and single optical paths in real time and output predictive early warning information.

[0009] S13: Digital twin-driven maintenance strategy simulation and optimization. In the cloud, based on real network topology, equipment models, real-time operation data and predictive early warning information, a digital twin of the optical communication network is constructed and updated. When multiple early warnings are received, different pre-maintenance strategies are input into the digital twin to conduct sandbox simulation, simulate the reliability indicators and risk changes of the entire network under each strategy, and generate optimized maintenance strategy suggestions.

[0010] S14: Decision execution and closed-loop verification. The system provides optimization and maintenance strategy recommendations to the operation and maintenance system, generates executable early warning work orders and switchover plans; at the same time, the system records the strategy execution results and the subsequent actual network status.

[0011] As a preferred option, the multi-source data acquisition and governance steps specifically include:

[0012] S21: Collect optical signal quality time-series data of each network node at a frequency of not less than minutes, including optical power and bit error rate;

[0013] S22: Collect event log data from each network node, standardize and clean text-descriptive logs, and extract and convert them into structured event tags;

[0014] S23: Collect relevant data related to equipment health, including equipment temperature, power supply voltage, and operating time;

[0015] S24: Time-align and feature-fuse time-series data, structured labels, and associated data from the same time window to form a unified multi-dimensional time-series feature vector that characterizes the health status of the device.

[0016] As a preferred option, the steps for constructing a small-sample time-series prediction model and edge warning specifically include:

[0017] S31: Model building and training: Based on historical multi-source time series data, build and train a lightweight time series prediction model specifically for early identification of equipment performance degradation.

[0018] S32: Real-time edge warning. The trained lightweight time series prediction model is deployed at the network edge to perform online inference on the real-time collected runtime time series data and generate predictive warning information containing failure probability and remaining lifetime.

[0019] As a preferred option, the model building and training steps specifically include:

[0020] S311: Model input feature construction. Time window slicing is performed on the treated multi-source time series data to construct a multi-dimensional feature sequence as model input. The multi-dimensional feature sequence includes: optical power sequence, bit error rate sequence, structured event label sequence extracted from event logs, and associated environmental parameter sequence.

[0021] S312: Model architecture selection: Select a lightweight neural network architecture suitable for small sample time series pattern learning as the core prediction model.

[0022] S313: Model training and optimization, using historical data, taking the feature sequence of the previous time step as input, and using whether a failure will occur in the future as the supervision label to train the model.

[0023] As a preferred option, the simulation and optimization steps for maintenance strategies driven by digital twins specifically include:

[0024] S41: Digital Twin Construction and Synchronization: Based on the real physical information, real-time operation data and predictive early warning information of the optical communication network, a network digital twin is constructed and dynamically updated in the cloud;

[0025] S42: Multi-strategy sandbox simulation. When a concurrent predictive warning is received for multiple entities in the network, different predefined pre-maintenance strategies are loaded into the digital twin and simulation is performed.

[0026] S43: Quantitative evaluation and recommendation generation of strategies. Based on the results of simulation, calculate and compare the network-wide performance indicators and risk costs under each strategy, and output optimization and maintenance strategy recommendations accordingly.

[0027] As a preferred approach, the steps for multi-strategy sandbox simulation include:

[0028] S421: Strategy definition and input. The input includes the objects to be maintained corresponding to the multiple received predictive warnings, the predicted remaining normal working time, and the probability of failure. A set of pre-maintenance strategies to be evaluated is defined. Each strategy contains a sequence of handling instructions for each object to be maintained. The handling instructions include: immediate repair, delayed repair, taking temporary protection measures, and performing protection switching.

[0029] S422: The simulation engine executes, taking the current synchronization state of the digital twin as the initial state. For each pre-maintenance strategy, the simulation engine drives the twin to perform discrete event or time step-by-time simulations according to the strategy instructions and the built-in rule base, simulating the dynamic changes in network status, device health, and service flow over a future simulation period.

[0030] As a preferred approach, the strategy quantitative assessment and recommendation generation steps include:

[0031] S431: Quantitative calculation of indicators. For the results of each strategy, calculate a set of predefined key performance indicators, including the predicted total duration of network-wide service interruption, the predicted number of violations of key service level agreements, the estimated cost of operation and maintenance resources, and the risk index of derivative failures that may be caused by delayed maintenance.

[0032] S432: Strategy comparison and ranking, based on key performance indicators, using multi-objective decision analysis methods to comprehensively evaluate and rank different pre-maintenance strategies;

[0033] S433: Optimization suggestions are generated, and the evaluation results are output, including recommended priority execution strategies, comparative analysis of the advantages and disadvantages of each strategy, and time window suggestions for strategy execution.

[0034] As a preferred option, the decision execution and closed-loop verification steps specifically include:

[0035] S51: Strategy execution and work order generation converts maintenance strategy suggestions derived from digital twin simulation optimization into specific executable instructions, early warning work orders, and switchover plans in the operation and maintenance system;

[0036] S52: Execution tracking and effect verification, tracking and recording the execution process and actual results of instructions, work orders and contingency plans, and continuously collecting data on the actual operating status of relevant network devices after policy execution;

[0037] S53: Data closed-loop feedback and model optimization, which compares and analyzes policy execution records, actual network state changes and prediction information, and uses the analysis results to update and optimize the time series prediction model and digital twin.

[0038] As a preferred option, the strategy execution and work order generation steps specifically include:

[0039] S511: Command conversion and generation, automatically generating a sequence of operation commands that can be executed by network devices based on the recommended scheme suggested by the optimization and maintenance strategy;

[0040] S512: Pre-set and associated contingency plans. When the policy recommendation involves switching optical path protection, a detailed switching contingency plan is automatically generated and associated with the corresponding early warning events, primary and backup optical path information. The contingency plan is pre-set in the network management system and SDN controller for triggering execution.

[0041] S513: Audit information embedding. When generating instructions, work orders, and contingency plans, source information is automatically embedded, including the unique identifier that triggers the warning, the unique identifier of the digital twin simulation simulation on which it is based, and the decision timestamp of the strategy generation.

[0042] An AI-based optical communication fault detection system includes:

[0043] The data acquisition and governance module is deployed in a production environment close to network devices. It is used to collect and standardize runtime sequence data, event logs and related data from multiple network nodes in real time to build device health status characteristics.

[0044] The edge intelligent early warning module is deployed on network nodes or edge computing gateways. It has a built-in lightweight time-series prediction model for real-time analysis of device health status characteristics and outputs predictive early warning information including failure probability and remaining normal working time.

[0045] The digital twin simulation decision-making module, deployed in the cloud, is used to build and synchronize a digital twin based on network topology, real-time data and predictive early warning information. When multiple early warnings are received, it performs sand table simulations and quantitative evaluations for different pre-maintenance strategies to generate recommendations for optimized maintenance strategies.

[0046] The operation and maintenance scheduling and closed-loop management module is used to receive optimization maintenance strategy suggestions and generate executable work orders or instructions; at the same time, it tracks the execution results and collects feedback data to drive the iterative update of the time series prediction model and digital twin, thereby realizing closed-loop optimization of operation and maintenance decisions.

[0047] The beneficial effects of this invention are:

[0048] 1. Compared to existing technologies that mainly rely on fixed performance thresholds for post-event alarms, the disadvantage is that by the time an alarm is triggered, the equipment degradation has often entered a late stage, leaving maintenance personnel with a very limited response window and failing to achieve true preventative maintenance. This invention deploys a lightweight time-series prediction model based on few-sample learning to automatically learn the micro-degradation patterns of equipment performance directly from high-frequency time-series data such as power and bit error rate, thereby issuing early warnings hours to days before complete hardware failure. This method achieves a fundamental shift from threshold-triggered passive alarms to trend-prediction-based proactive early warnings, gaining valuable time for preventative operations such as component replacement and configuration adjustments, and effectively avoiding business interruptions caused by sudden failures.

[0049] 2. Compared to existing technologies that rely on maintenance personnel's personal experience to prioritize responses when facing multiple concurrent alerts, this technology lacks quantitative analysis of the overall network status and potential cascading effects of faults, leading to blind decision-making processes and potential secondary faults or resource scheduling conflicts. This invention constructs a digital twin in the cloud that integrates physical topology, business logic, and rule base, allowing maintenance personnel to conduct "if-then" scenario simulations of different maintenance strategies in a virtual image. The system can simulate and simulate the network-wide business interruption risk, compliance status, and cost changes under various strategies, thereby elevating maintenance decisions from isolated judgments based on local experience to global optimization based on network-wide simulation and data-driven approaches, significantly improving the scientific rigor and reliability of decision-making.

[0050] 3. Compared to existing operation and maintenance systems, whose early warning models and decision-making logic are often static and lack the ability to learn from actual results, the system cannot continuously improve with network changes and operation and maintenance practices. This invention innovatively designs a complete closed-loop feedback link from decision execution to model optimization. The system records the accuracy of each early warning and the actual effect of the maintenance strategy in detail, and uses this feedback data to drive the incremental learning of the time series prediction model and the calibration of the digital twin simulation parameters. This enables the entire system to have the ability to learn and improve itself. The prediction model will become more and more accurate over time, and the simulation will become closer and closer to reality, thus forming a continuously evolving intelligent operation and maintenance system that continuously improves long-term operation and maintenance efficiency.

[0051] 4. Compared to existing technologies that rely primarily on manual document review and checks for verifying complex compliance requirements during optical path switching or maintenance plan development, which is inefficient and prone to oversights in emergency situations leading to compliance risks, this invention internalizes relevant compliance rules, such as data supervision requirements and operator monitoring strategies, into the computable logic of a digital twin rule base. When performing protection switching simulations or setting maintenance windows, the simulation engine automatically compares and verifies each alternative scheme with the built-in rules. This ensures that the final generated switching plan or maintenance strategy, while pursuing optimal performance, naturally meets all preset compliance constraints, freeing manual labor from tedious rule verification and eliminating compliance incidents caused by human negligence. Attached Figure Description

[0052] Figure 1 The diagram shown is a flowchart of the artificial intelligence-based optical communication fault detection method of the present invention.

[0053] Figure 2 The diagram shown is a structural schematic of the artificial intelligence-based optical communication fault detection system of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] Please see Figure 1 The present invention provides an embodiment of an artificial intelligence-based optical communication fault detection method, comprising the following steps:

[0056] S11: Multi-source data acquisition and governance, which collects runtime sequence data, event log data and related data from multiple network nodes in the optical communication network, and performs standardization, alignment and labeling governance on the data to build a health status profile of the device;

[0057] S12: Small sample time series prediction model construction and edge early warning. Based on the treated data, a lightweight time series prediction model is trained and deployed at the network edge to evaluate the health status of single nodes and single optical paths in real time and output predictive early warning information.

[0058] S13: Digital twin-driven maintenance strategy simulation and optimization. In the cloud, based on real network topology, equipment models, real-time operation data and predictive early warning information, a digital twin of the optical communication network is constructed and updated. When multiple early warnings are received, different pre-maintenance strategies are input into the digital twin to conduct sandbox simulation, simulate the reliability indicators and risk changes of the entire network under each strategy, and generate optimized maintenance strategy suggestions.

[0059] S14: Decision execution and closed-loop verification. The system provides optimization and maintenance strategy recommendations to the operation and maintenance system, generates executable early warning work orders and switchover plans; at the same time, the system records the strategy execution results and the subsequent actual network status.

[0060] Specifically, this invention first collects and processes multi-source data to construct a health status profile of the equipment. Then, based on the processed data, it trains a lightweight time-series prediction model and deploys it at the network edge to achieve real-time health assessment and fault early warning for single nodes or optical paths. Next, it constructs and updates a network digital twin in the cloud, performs multi-strategy simulations of multiple early warnings through a sandbox approach, and outputs optimized maintenance strategy suggestions. Finally, it executes the optimization strategies and records the results, forming a closed-loop verification and continuous optimization. This solution enables a paradigm shift from passive alarms to proactive prediction in operations and maintenance, providing an early warning window before hardware failure, and achieving global and forward-looking operational and maintenance decisions through digital twin simulation. This effectively improves network reliability, reduces operational and maintenance costs, and meets compliance requirements.

[0061] As a preferred option, the multi-source data acquisition and governance steps specifically include:

[0062] S21: Collect optical signal quality time-series data of each network node at a frequency of not less than minutes, including optical power and bit error rate;

[0063] S22: Collect event log data from each network node, standardize and clean text-descriptive logs, and extract and convert them into structured event tags;

[0064] S23: Collect relevant data related to equipment health, including equipment temperature, power supply voltage, and operating time;

[0065] S24: Time-align and feature-fuse time-series data, structured labels, and associated data from the same time window to form a unified multi-dimensional time-series feature vector that characterizes the health status of the device.

[0066] Specifically, in the multi-source data acquisition and governance step, this invention acquires optical signal quality time-series data such as optical power and bit error rate at a high frequency of no less than minutes. Combined with structured cleaning of event log data and synchronous acquisition of related data such as equipment temperature and voltage, the multi-source heterogeneous data is time-aligned and feature-fused to form a unified multi-dimensional time-series feature vector representing the equipment's health status. This technical solution solves the problems of scattered data sources and inconsistent formats, constructing a high-quality, high-dimensional equipment health profile. This lays a solid and complete data foundation for subsequent accurate prediction and intelligent analysis, thereby significantly improving the timeliness and accuracy of fault detection.

[0067] As a preferred option, the steps for constructing a small-sample time-series prediction model and edge warning specifically include:

[0068] S31: Model building and training: Based on historical multi-source time series data, build and train a lightweight time series prediction model specifically for early identification of equipment performance degradation.

[0069] S32: Real-time edge warning. The trained lightweight time series prediction model is deployed at the network edge to perform online inference on the real-time collected runtime time series data and generate predictive warning information containing failure probability and remaining lifetime.

[0070] As a preferred option, the model building and training steps specifically include:

[0071] S311: Model input feature construction. Time window slicing is performed on the treated multi-source time series data to construct a multi-dimensional feature sequence as model input. The multi-dimensional feature sequence includes: optical power sequence, bit error rate sequence, structured event label sequence extracted from event logs, and associated environmental parameter sequence.

[0072] S312: Model architecture selection: Select a lightweight neural network architecture suitable for small sample time series pattern learning as the core prediction model.

[0073] S313: Model training and optimization, using historical data, taking the feature sequence of the previous time step as input, and using whether a failure will occur in the future as the supervision label to train the model.

[0074] Among them, the lightweight neural network architecture suitable for learning time-series patterns with few samples is a one-dimensional convolutional neural network or a pruned and compressed long short-term memory network.

[0075] In the model training and optimization steps, a loss function optimized for class imbalance is used during training, and early stopping and mini-batch gradient descent are used to prevent overfitting.

[0076] In the model training and optimization steps, a two-stage strategy is adopted for model training:

[0077] Pre-training phase: The model is pre-trained using large-scale normal operating condition time series data and a small amount of injected synthetic failure mode data, so that it learns basic health state representations and common degradation patterns.

[0078] Fine-tuning phase: Using real, small-scale historical fault sample data and their corresponding preceding time series data, the pre-trained model is fine-tuned to adapt to the actual degradation trajectory of a specific network or device type.

[0079] The real-time edge warning stage includes:

[0080] S321: Model Lightweighting and Encapsulation. The trained model is quantized, pruned, and knowledge distilled, and then encapsulated into an inference engine suitable for edge computing platforms.

[0081] S322: Edge deployment and synchronization. Deploy the encapsulated model inference engine and necessary preprocessing programs to the target network node or the nearest edge gateway, and establish a parameter synchronization mechanism with the cloud model management center.

[0082] S323: Real-time streaming inference. At the edge, the multi-dimensional feature sequences that flow in in real time are preprocessed using a sliding time window and then input into the inference engine to calculate and output prediction results in real time.

[0083] In the real-time streaming inference step, the model output is:

[0084] Classification prediction results: Output the probability value of a specific type of failure occurring within a preset time period in the future;

[0085] Regression prediction results: Based on the current performance degradation trend, extrapolate the remaining normal operating time of the equipment.

[0086] The real-time edge warning stage also includes:

[0087] S33: Early warning decision and reporting. Set dynamic or static early warning thresholds at the edge. When the fault probability value exceeds the threshold or the remaining normal working time is lower than the threshold, automatically generate a structured early warning message. The early warning message should include at least the device identifier, the predicted fault type, the fault probability, the remaining normal working time timestamp, and the key feature values ​​of the current time window, and report it to the upper-level operation and maintenance management system.

[0088] Specifically, this invention first slices the treated multi-source time-series data into time windows, constructing a multi-dimensional feature sequence containing optical power, bit error rate, event labels, and environmental parameters as input. A lightweight architecture, such as a one-dimensional convolutional neural network or a pruned and compressed long short-term memory network, is used as the core model, employing a two-stage training strategy: first, pre-training using normal operating conditions and synthetic fault data, and then fine-tuning using real small-sample fault data. During training, improved loss functions and early stopping methods are used for optimization, effectively overcoming sample imbalance and overfitting problems. In the real-time edge warning stage, after lightweight processing such as quantization, pruning, and knowledge distillation, the model is encapsulated as an inference engine and deployed at the edge. It infers from the real-time data stream through a sliding window, simultaneously outputting future fault probabilities and remaining lifetime estimates. When the predicted value exceeds a preset threshold, the system automatically generates and reports a structured warning message containing key information such as device identifier, prediction type, probability, and remaining time. This technical solution, through small-sample learning and lightweight model design, achieves accurate, real-time early warning of the slow degradation trend of equipment on the resource-constrained edge, transforming the operation and maintenance mode from reactive response to proactive prediction, winning a valuable time window for preventive maintenance, while reducing network transmission dependence and cloud computing load.

[0089] As a preferred option, the simulation and optimization steps for maintenance strategies driven by digital twins specifically include:

[0090] S41: Digital Twin Construction and Synchronization: Based on the real physical information, real-time operation data and predictive early warning information of the optical communication network, a network digital twin is constructed and dynamically updated in the cloud;

[0091] S42: Multi-strategy sandbox simulation. When a concurrent predictive warning is received for multiple entities in the network, different predefined pre-maintenance strategies are loaded into the digital twin and simulation is performed.

[0092] S43: Quantitative evaluation and recommendation generation of strategies. Based on the results of simulation, calculate and compare the network-wide performance indicators and risk costs under each strategy, and output optimization and maintenance strategy recommendations accordingly.

[0093] The steps involved in building and synchronizing a digital twin include:

[0094] S411: Twin Model Construction, creating and maintaining a high-fidelity network digital twin, including:

[0095] The physical layer model is used to map the topology, device attributes, connectivity, and physical state of a real network.

[0096] The business layer model is used to define the business logic, data flow paths, and service level agreement constraints that are carried on top of the physical layer.

[0097] The rule base includes a built-in device fault propagation model, network operation compliance rules, and resource scheduling strategies.

[0098] S412: Data Injection and Synchronization. Real-time operational data from the data acquisition system and predictive early warning information from the edge early warning system are continuously injected into the digital twin to drive the parameter and state updates of the physical layer model and the business layer model, so that they are synchronized with the real network.

[0099] The built-in fault propagation model in the rule base is constructed based on historical fault data, physical device dependencies, and network logical dependencies. It is used to simulate the cascading effects on related devices, paths, and upper-layer services when a node or link fails or degrades in performance.

[0100] As a preferred approach, the steps for multi-strategy sandbox simulation include:

[0101] S421: Strategy definition and input. The input includes the objects to be maintained corresponding to the multiple received predictive warnings, the predicted remaining normal working time, and the probability of failure. A set of pre-maintenance strategies to be evaluated is defined. Each strategy contains a sequence of handling instructions for each object to be maintained. The handling instructions include: immediate repair, delayed repair, taking temporary protection measures, and performing protection switching.

[0102] S422: The simulation engine executes, taking the current synchronization state of the digital twin as the initial state. For each pre-maintenance strategy, the simulation engine drives the twin to perform discrete event or time step-by-time simulations according to the strategy instructions and the built-in rule base, simulating the dynamic changes in network status, device health, and service flow over a future simulation period.

[0103] As a preferred approach, the strategy quantitative assessment and recommendation generation steps include:

[0104] S431: Quantitative calculation of indicators. For the results of each strategy, calculate a set of predefined key performance indicators, including the predicted total duration of network-wide service interruption, the predicted number of violations of key service level agreements, the estimated cost of operation and maintenance resources, and the risk index of derivative failures that may be caused by delayed maintenance.

[0105] S432: Strategy comparison and ranking, based on key performance indicators, using multi-objective decision analysis methods to comprehensively evaluate and rank different pre-maintenance strategies;

[0106] S433: Optimization suggestions are generated, and the evaluation results are output, including recommended priority execution strategies, comparative analysis of the advantages and disadvantages of each strategy, and time window suggestions for strategy execution.

[0107] Specifically, this invention constructs a high-fidelity network digital twin in the cloud, integrating a physical layer model, a business layer model, and a rule base, and continuously injects real-time data and early warning information to maintain virtual-real synchronization. When faced with multiple concurrent early warnings, the system allows the definition of various pre-maintenance strategies containing different handling instructions (such as immediate repair, delayed repair, protection switching, etc.), and dynamically predicts the future network state based on the twin and a built-in fault propagation model. Subsequently, the system quantitatively calculates key indicators under each strategy (such as total service interruption duration, number of compliance violations, operation and maintenance costs, and derivative risks), and uses multi-objective decision analysis for comprehensive evaluation and ranking, ultimately outputting optimization suggestions including priority strategies, advantages and disadvantages comparisons, and execution time windows. This technical solution elevates operation and maintenance decision-making from isolated responses to single early warnings to the level of quantitative evaluation and optimization selection of complex strategies with multiple objectives and constraints in a full-network simulation environment, achieving optimal scheduling of operation and maintenance resources and pre-emptive global risk control, significantly improving network reliability and the scientific nature of operation and maintenance decisions.

[0108] As a preferred option, the decision execution and closed-loop verification steps specifically include:

[0109] S51: Strategy execution and work order generation converts maintenance strategy suggestions derived from digital twin simulation optimization into specific executable instructions, early warning work orders, and switchover plans in the operation and maintenance system;

[0110] S52: Execution tracking and effect verification, tracking and recording the execution process and actual results of instructions, work orders and contingency plans, and continuously collecting data on the actual operating status of relevant network devices after policy execution;

[0111] S53: Data closed-loop feedback and model optimization, which compares and analyzes policy execution records, actual network state changes and prediction information, and uses the analysis results to update and optimize the time series prediction model and digital twin.

[0112] As a preferred option, the strategy execution and work order generation steps specifically include:

[0113] S511: Command conversion and generation, automatically generating a sequence of operation commands that can be executed by network devices based on the recommended scheme suggested by the optimization and maintenance strategy;

[0114] S512: Pre-set and associated contingency plans. When the policy recommendation involves switching optical path protection, a detailed switching contingency plan is automatically generated and associated with the corresponding early warning events, primary and backup optical path information. The contingency plan is pre-set in the network management system and SDN controller for triggering execution.

[0115] S513: Audit information embedding. When generating instructions, work orders, and contingency plans, source information is automatically embedded, including the unique identifier that triggers the warning, the unique identifier of the digital twin simulation simulation on which it is based, and the decision timestamp of the strategy generation.

[0116] The execution tracking and effect verification steps specifically include:

[0117] S521: Execution process record. During the process of work order being dispatched and executed or contingency plan being triggered, the system records key node information, including work order dispatch time, executor, actual start time, situations encountered during execution, and completion time.

[0118] S522: Effect verification data collection. Within a preset time period after the policy is executed, collect actual operating data from relevant network devices, including time-series data of key performance indicators of the devices, event logs, and the continuity status of related services.

[0119] S523: Impact assessment, based on the collected actual operational data, to evaluate the actual effectiveness of this maintenance strategy, to determine whether the predicted failures were effectively avoided, and whether there were any unplanned negative impacts on network operation;

[0120] S524: Early warning accuracy marking. The final result of each predictive early warning is compared with the actual situation, and the accuracy of the early warning is marked, including true positive, false positive, true negative and false negative, and the time difference between the early warning time and the actual time of the fault is recorded.

[0121] The data closed-loop feedback and model optimization steps specifically include:

[0122] S531: Construction of prediction model update dataset. Based on the recorded early warning accuracy label and time difference, construct the model retraining dataset, and align and package the feature sequences used in historical predictions, the predicted probabilities output by the model, and the final true result labels.

[0123] S532: Incremental learning of time series prediction models. When enough feedback data has been accumulated, the retraining dataset is used to incrementally learn and retrain small-sample time series prediction models deployed at the edge or in the cloud to optimize their prediction accuracy and adjust the warning threshold.

[0124] S533: Twin model calibration, based on the differences between the actual scope of fault impact, service interruption duration and the results of digital twin simulation, to calibrate the parameters of the fault propagation model and equipment performance degradation model in the digital twin.

[0125] Specifically, this invention first automatically converts optimization strategies into executable instructions, work orders, or switching plans, embedding complete audit information to ensure traceability. The system then tracks and records the entire execution process and collects actual operational data afterward to evaluate the strategy's effectiveness and assess the accuracy of warnings. Finally, the system utilizes execution feedback and verification results to construct a retraining dataset to drive incremental learning of the time-series prediction model, and calibrates the model parameters of the digital twin based on the differences between simulation and reality. This technical solution achieves an automated closed loop from intelligent decision-making to on-site execution, not only ensuring the auditability and traceability of operation and maintenance operations, but also enabling the prediction model and simulation model to continuously evolve their accuracy through a continuous feedback learning mechanism, thereby improving the overall system's adaptability and long-term operational efficiency.

[0126] like Figure 2 As shown, this embodiment also provides an artificial intelligence-based optical communication fault detection system, including:

[0127] The data acquisition and governance module is deployed in a production environment close to network devices. It is used to collect and standardize runtime sequence data, event logs and related data from multiple network nodes in real time to build device health status characteristics.

[0128] The edge intelligent early warning module is deployed on network nodes or edge computing gateways. It has a built-in lightweight time-series prediction model for real-time analysis of device health status characteristics and outputs predictive early warning information including failure probability and remaining normal working time.

[0129] The digital twin simulation decision-making module, deployed in the cloud, is used to build and synchronize a digital twin based on network topology, real-time data and predictive early warning information. When multiple early warnings are received, it performs sand table simulations and quantitative evaluations for different pre-maintenance strategies to generate recommendations for optimized maintenance strategies.

[0130] The operation and maintenance scheduling and closed-loop management module is used to receive optimization maintenance strategy suggestions and generate executable work orders or instructions; at the same time, it tracks the execution results and collects feedback data to drive the iterative update of the time series prediction model and digital twin, thereby realizing closed-loop optimization of operation and maintenance decisions.

[0131] Example 1: Predictive Maintenance of Core Optical Amplifiers in Metropolitan Area Networks

[0132] 1. Scenario and Data Acquisition: In a metropolitan area backbone network, the erbium-doped fiber amplifier (EDFA, device ID: EDFA-Node7) of a key node experienced slow performance degradation. The system collected its optical power (slowly drifting from -1dBm to -3.5dBm) and bit error rate (increasing from 10^-12 to 10^-9) time-series data once per second. Simultaneously, the event log frequently contained the text "optical power fluctuation alarm," which, after cleaning, was converted into a structured tag "Power_Fluctuation." Correlation data collection showed a slight decrease in the cooling fan speed, causing the internal temperature to slowly rise from 35°C to 42°C. All data was time-aligned on the cloud governance platform and merged into a minute-level multidimensional health feature vector containing optical power, bit error rate, temperature, and power fluctuation tags.

[0133] 2. Small sample margin warning

[0134] A lightweight 1D-CNN early warning model deployed on the edge gateway of the node received the latest temporal feature vector. This model had previously undergone two-stage training: the first stage used one year's worth of data from 100 normal amplifiers across the entire network and 10% of synthetic degradation data for pre-training; the second stage used only 5 historical real-world fault data points for fine-tuning. After real-time analysis, the model output: the probability of EDFA-Node7 experiencing a "gain reduction fault" within the next 48 hours is 87%, with an estimated remaining uptime (RUL) of 60 hours. Since the probability exceeds the preset threshold of 85%, the edge gateway immediately generates an early warning message, including the device ID, predicted fault type, probability value, RUL, and a current feature snapshot, and reports it to the operations and maintenance center.

[0135] 3. Digital twin simulation decision making

[0136] The operations and maintenance center's digital twin platform received the alert. The twin has synchronized the entire network topology, service flow (such as the 10G leased line carrying Financial A to Data Center B), and compliance rules (this leased line must be fully monitorable), including EDFA-Node7. The simulation module was activated, setting two strategies: Strategy 1 (immediate maintenance): simulating the impact of switching services to the backup path within a 4-hour maintenance window; Strategy 2 (delayed maintenance until the weekend): simulating a real equipment failure after 72 hours, triggering the automatic protection switchover process. The simulation results show that Strategy 1 will cause a 4-minute transient interruption on the leased line (permissible under compliance), but it is low-cost and has no derivative risks; Strategy 2, although without planned interruption, may result in 15 minutes of service loss if the prediction is inaccurate, violating the SLA. After comprehensive system evaluation, the following recommendation was generated: adopt Strategy 1, perform preventative replacement during the low-business period from 2-4 AM today, and pre-configure the switchover plan to the optical path PL-Backup-01.

[0137] 4. Decision Implementation and Closed Loop

[0138] The operations and maintenance system adopted the suggestion, automatically generated a preventative maintenance work order, assigned it to engineer Zhang San, and pre-issued the switchover command to the SDN controller. Zhang San performed the replacement at 2:00 AM, and the system recorded the entire operation. Over the following week, the system continuously monitored EDFA-Node7, confirming that its new module was operating normally and the predicted fault was avoided. This warning was marked as a "true positive," and the error between the predicted RUL (60 hours) and the actual maintenance time (20 hours after the warning) was recorded. This data was packaged into a feedback dataset for fine-tuning the 1D-CNN model, making it more sensitive to similar amplifier degradation patterns; simultaneously, the actual result of this fault-free event was also used to calibrate simulation parameters regarding the impact of "amplifier replacement" in the digital twin.

[0139] Example 2: Global Operation and Maintenance Resource Scheduling under High Concurrency Warning

[0140] 1. Scenarios and Data Acquisition

[0141] In another large optical network, the system received two high-level warnings almost simultaneously: 1) Edge warning model A predicted that the optical switch SW-A located in the eastern hub (due to drive motor wear) had a 92% probability of switching failure within the next 24 hours, with a RUL of 30 hours. 2) Edge warning model B predicted that the optical amplifier AMP-B located in the western hub (due to pump laser aging) had an 80% probability of a sudden drop in output power within the next 48 hours, with a RUL of 55 hours. Minute-level optical power, bit error rate, temperature, and event tag data for both devices were synchronized to the cloud in real time.

[0142] 2. Small sample margin warning

[0143] Both alerts were generated independently by lightweight LSTM models deployed at their respective nodes. Both models were trained with a small number of samples and were able to identify the micro-fluctuation pattern of the drive current in SW-A and the slow decline pattern of pump efficiency in AMP-B. Detailed alert information was reported, triggering the "multi-event concurrent processing" process of the digital twin platform.

[0144] 3. Digital twin simulation decision making

[0145] The digital twin loads the current network status, and two early warning devices are highlighted. Maintenance personnel input three pre-maintenance strategies to conduct a simulation exercise.

[0146] Strategy X (Repair SW-A first, then AMP-B): Simulate immediate repair of SW-A (requiring interruption of its ring network), followed by repair of AMP-B 24 hours later.

[0147] Strategy Y (Repair AMP-B first, then SW-A): Simulate immediate repair of AMP-B (affecting a single trunk line), followed by repair of SW-A 36 hours later.

[0148] Strategy Z (Dual-track parallel): Simulates sending out two teams of engineers for maintenance at the same time, but the resource cost doubles.

[0149] The simulation engine, based on a fault propagation model, simulated a week's worth of scenarios: Under Strategy X, SW-A was repaired first, avoiding potential large-scale routing chaos. However, the risk of AMP-B failing before repair caused a short-term interruption of a secondary trunk line. Under Strategy Y, AMP-B was repaired first, stabilizing the trunk line. However, SW-A failed before repair, causing protection switching chaos in its ring network, affecting the services of three major customers and violating the SLA. Strategy Z had the highest cost but the lowest overall risk. Quantitative assessment showed that Strategy X had the shortest "estimated total duration of network-wide service interruption" and the lowest "SLA violation risk index." The system recommends Strategy X and specifically points out that when repairing SW-A, the financial leased lines it is responsible for must be forcibly switched to the backup optical path PL-Audit-05, which meets regulatory audit requirements.

[0150] 4. Decision Implementation and Closed Loop

[0151] The system generated two related work orders: Work Order 1 (Urgent, SW-A) and Work Order 2 (High Priority, AMP-B). Work Order 1 included a forced switchover contingency plan. The operations team executed Strategy X, successfully completing the replacement before the predicted failure of SW-A and repairing AMP-B 24 hours later. Post-event verification showed that the warning for SW-A was a true positive (the actual RUL was 28 hours, close to the predicted 30 hours), while the warning for AMP-B was a false positive (its performance did not plummet after 55 hours). This valuable "false positive" sample, along with the "true positive" sample, was used for incremental learning of the model, particularly helping the model better distinguish between "critical aging" and "stable aging" modes. Simultaneously, the result that "the SW-A failure did not trigger derivative problems" reinforced the model parameters in the twin regarding "good fault isolation of this type of optical switch," making future simulations more accurate. The entire decision-making logic and execution record were stored in the case library as a reference for future handling of similar conflicts between "critical switches" and "trunk amplifiers."

[0152] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An artificial intelligence-based optical communication fault detection method, characterized in that: Includes the following steps: S11: Multi-source data acquisition and governance, which collects runtime sequence data, event log data and related data from multiple network nodes in the optical communication network, and performs standardization, alignment and labeling governance on the data to build a health status profile of the device; S12: Small sample time series prediction model construction and edge early warning. Based on the treated data, a lightweight time series prediction model is trained and deployed at the network edge to evaluate the health status of single nodes and single optical paths in real time and output predictive early warning information. S13: Digital twin-driven maintenance strategy simulation and optimization. In the cloud, based on real network topology, equipment models, real-time operation data and predictive early warning information, a digital twin of the optical communication network is constructed and updated. When multiple early warnings are received, different pre-maintenance strategies are input into the digital twin to conduct sandbox simulation, simulate the reliability indicators and risk changes of the entire network under each strategy, and generate optimized maintenance strategy suggestions. S14: Decision execution and closed-loop verification. The system provides optimization and maintenance strategy recommendations to the operation and maintenance system, generates executable early warning work orders and switchover plans; at the same time, the system records the strategy execution results and the subsequent actual network status.

2. The artificial intelligence-based optical communication fault detection method according to claim 1, characterized in that: The specific steps of multi-source data acquisition and governance include: S21: Collect optical signal quality time-series data of each network node at a frequency of not less than minutes, including optical power and bit error rate; S22: Collect event log data from each network node, standardize and clean text-descriptive logs, and extract and convert them into structured event tags; S23: Collect relevant data related to equipment health, including equipment temperature, power supply voltage, and operating time; S24: Time-align and feature-fuse time-series data, structured labels, and associated data from the same time window to form a unified multi-dimensional time-series feature vector that characterizes the health status of the device.

3. The artificial intelligence-based optical communication fault detection method according to claim 2, characterized in that: The specific steps for building a small-sample time-series prediction model and edge early warning include: S31: Model building and training: Based on historical multi-source time series data, build and train a lightweight time series prediction model specifically for early identification of equipment performance degradation. S32: Real-time edge warning. The trained lightweight time series prediction model is deployed at the network edge to perform online inference on the real-time collected runtime time series data and generate predictive warning information containing failure probability and remaining lifetime.

4. The artificial intelligence-based optical communication fault detection method according to claim 3, characterized in that: The specific steps of model building and training include: S311: Model input feature construction. Time window slicing is performed on the treated multi-source time series data to construct a multi-dimensional feature sequence as model input. The multi-dimensional feature sequence includes: optical power sequence, bit error rate sequence, structured event label sequence extracted from event logs, and associated environmental parameter sequence. S312: Model architecture selection: Select a lightweight neural network architecture suitable for small sample time series pattern learning as the core prediction model. S313: Model training and optimization, using historical data, taking the feature sequence of the previous time step as input, and using whether a failure will occur in the future as the supervision label to train the model.

5. The artificial intelligence-based optical communication fault detection method according to claim 4, characterized in that: The specific steps of digital twin-driven maintenance strategy simulation and optimization include: S41: Digital Twin Construction and Synchronization: Based on the real physical information, real-time operation data and predictive early warning information of the optical communication network, a network digital twin is constructed and dynamically updated in the cloud; S42: Multi-strategy sandbox simulation. When a concurrent predictive warning is received for multiple entities in the network, different predefined pre-maintenance strategies are loaded into the digital twin and simulation is performed. S43: Quantitative evaluation and recommendation generation of strategies. Based on the results of simulation, calculate and compare the network-wide performance indicators and risk costs under each strategy, and output optimization and maintenance strategy recommendations accordingly.

6. The artificial intelligence-based optical communication fault detection method according to claim 5, characterized in that: The steps of a multi-strategy sand table simulation include: S421: Strategy definition and input. The input includes the objects to be maintained corresponding to the multiple received predictive warnings, the predicted remaining normal working time, and the probability of failure. A set of pre-maintenance strategies to be evaluated is defined. Each strategy contains a sequence of handling instructions for each object to be maintained. The handling instructions include: immediate repair, delayed repair, taking temporary protection measures, and performing protection switching. S422: The simulation engine executes, taking the current synchronization state of the digital twin as the initial state. For each pre-maintenance strategy, the simulation engine drives the twin to perform discrete event or time step-by-time simulations according to the strategy instructions and the built-in rule base, simulating the dynamic changes in network status, device health, and service flow over a future simulation period.

7. The artificial intelligence-based optical communication fault detection method according to claim 6, characterized in that: The steps for strategy quantitative assessment and recommendation generation include: S431: Quantitative calculation of indicators. For the results of each strategy, calculate a set of predefined key performance indicators, including the predicted total duration of network-wide service interruption, the predicted number of violations of key service level agreements, the estimated cost of operation and maintenance resources, and the risk index of derivative failures that may be caused by delayed maintenance. S432: Strategy comparison and ranking, based on key performance indicators, using multi-objective decision analysis methods to comprehensively evaluate and rank different pre-maintenance strategies; S433: Optimization suggestions are generated, and the evaluation results are output, including recommended priority execution strategies, comparative analysis of the advantages and disadvantages of each strategy, and time window suggestions for strategy execution.

8. The optical communication fault detection method based on artificial intelligence according to claim 7, characterized in that: The specific steps of decision execution and closed-loop verification include: S51: Strategy execution and work order generation converts maintenance strategy suggestions derived from digital twin simulation optimization into specific executable instructions, early warning work orders, and switchover plans in the operation and maintenance system; S52: Execution tracking and effect verification, tracking and recording the execution process and actual results of instructions, work orders and contingency plans, and continuously collecting data on the actual operating status of relevant network devices after policy execution; S53: Data closed-loop feedback and model optimization, which compares and analyzes policy execution records, actual network state changes and prediction information, and uses the analysis results to update and optimize the time series prediction model and digital twin.

9. The artificial intelligence-based optical communication fault detection method according to claim 8, characterized in that: The specific steps for strategy execution and work order generation include: S511: Command conversion and generation, automatically generating a sequence of operation commands that can be executed by network devices based on the recommended scheme suggested by the optimization and maintenance strategy; S512: Pre-set and associated contingency plans. When the policy recommendation involves switching optical path protection, a detailed switching contingency plan is automatically generated and associated with the corresponding early warning events, primary and backup optical path information. The contingency plan is pre-set in the network management system and SDN controller for triggering execution. S513: Audit information embedding. When generating instructions, work orders, and contingency plans, source information is automatically embedded, including the unique identifier that triggers the warning, the unique identifier of the digital twin simulation simulation on which it is based, and the decision timestamp of the strategy generation.

10. An artificial intelligence-based optical communication fault detection system, used to implement the artificial intelligence-based optical communication fault detection method according to any one of claims 1-9, characterized in that: include: The data acquisition and governance module is deployed in a production environment close to network devices. It is used to collect and standardize runtime sequence data, event logs and related data from multiple network nodes in real time to build device health status characteristics. The edge intelligent early warning module is deployed on network nodes or edge computing gateways. It has a built-in lightweight time-series prediction model for real-time analysis of device health status characteristics and outputs predictive early warning information including failure probability and remaining normal working time. The digital twin simulation decision-making module, deployed in the cloud, is used to build and synchronize a digital twin based on network topology, real-time data and predictive early warning information. When multiple early warnings are received, it performs sand table simulations and quantitative evaluations for different pre-maintenance strategies to generate recommendations for optimized maintenance strategies. The operation and maintenance scheduling and closed-loop management module is used to receive optimization maintenance strategy suggestions and generate executable work orders or instructions; at the same time, it tracks the execution results and collects feedback data to drive the iterative update of the time series prediction model and digital twin, thereby realizing closed-loop optimization of operation and maintenance decisions.

Images