AI-based communication optical cable anomaly prediction and fault positioning method and system

Abstract

The application discloses an AI-based communication optical cable abnormality prediction and fault positioning method and system, and a complete closed loop from data acquisition, abnormality prediction, accurate positioning, root cause analysis to feedback optimization is constructed through full-process closed loop management, the problem of fragmentation of each link in traditional operation and maintenance is solved, through the combination of prediction, positioning, classification and visualization, the change from passive repair to active prevention is realized, the business interruption risk is significantly reduced, the operation and maintenance efficiency is improved, the closed loop feedback mechanism is introduced, so that the system can continuously iterate and optimize the model by using real field data, adapt to network environment changes, and keep long-term high accuracy.

Description

Technical Field

[0001] This invention relates to the field of optical fiber communication cables, and in particular to an AI-based method and system for anomaly prediction and fault location in optical fiber communication cables. Background Technology

[0002] Traditional methods typically only discover faults passively after a complete fiber optic cable break or severe performance degradation has led to service disruptions. Existing OTDR tests are susceptible to noise interference, making it difficult to distinguish between real fault events and test artifacts such as ghosting and noise spikes. This results in inaccurate location or a large number of false alarms, and it is difficult to accurately determine the specific physical location of the fault. Traditional methods often only indicate that the cable is broken or has high loss, but it is difficult to accurately determine the specific fault type, such as micro-bends, loose connectors, or detached pigtails. Furthermore, when faced with complex alarm storms, it is difficult to quickly pinpoint the true root cause from a massive number of related alarms.

[0003] In summary, there is a need for an AI-based method and system for predicting and locating anomalies in optical fiber communication cables to address the shortcomings of existing technologies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based method and system for predicting and locating anomalies in optical fiber communication cables, aiming to solve the aforementioned problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based method for anomaly prediction and fault location in communication optical cables, comprising the following steps:

[0006] Step S1: Multidimensional data acquisition and preprocessing. Real-time acquisition of multi-source heterogeneous data from the communication optical cable link. Cleaning, noise reduction and timestamp alignment of the multi-source heterogeneous data to construct a standardized optical cable status dataset.

[0007] Step S2: Abnormal trend prediction. Input the optical cable status dataset into the pre-trained abnormal prediction model, extract the optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period in the future, and generate an early warning signal when the predicted health index is lower than the dynamic threshold or an abnormal fluctuation pattern is detected.

[0008] Step S3: Accurate fault location and classification. In response to the early warning signal, the OTDR test is automatically triggered to obtain the current backscatter curve. The backscatter curve is image-recognized using a convolutional neural network to distinguish between real fault events and noise artifacts. The physical distance of the fault point is calculated. Combined with the telemetry data of the intelligent optical module and the physical distance of the fault point, the fault type is determined by the fault classification model.

[0009] Step S4: Root cause analysis and visualization mapping. Based on graph neural network, a fault propagation topology is constructed. Cluster analysis is performed on associated alarms to locate the root cause fault. The physical distance of the fault point is mapped to the GIS geographic information system to generate a visualized fault work order containing fault latitude and longitude, fault type and surrounding environment information.

[0010] Step S5: Closed-loop feedback optimization. Receive the actual fault handling results from on-site maintenance personnel, update the optical cable status dataset as labeled data, and perform incremental training and parameter iteration on the anomaly prediction model, fault classification model, and convolutional neural network.

[0011] Optionally, the multi-source heterogeneous data in step S1 includes at least millisecond-level telemetry data from the smart optical module, distributed fiber optic sensing data, optical time domain reflectometer test waveform data, network management system alarm logs, and geographic information system data.

[0012] Optionally, the millisecond-level telemetry data of the intelligent optical module includes: received optical power, transmitted optical power, bias current, module temperature and power supply voltage, and the distributed optical fiber sensing data includes acoustic vibration signals and temperature change signals distributed along the optical cable.

[0013] The timestamp alignment process is as follows: based on the network standard time protocol, multi-source data with different sampling frequencies are uniformly resampled to a preset time granularity.

[0014] Optionally, step S2 is implemented in the following manner:

[0015] Step A1: Multi-source data alignment and sliding window construction. Using linear interpolation or spline interpolation, all data are resampled to a uniform time granularity to form a complete multidimensional time series matrix. The sliding window length and prediction step size are set to cut the continuous time series into overlapping sample pairs.

[0016] Step A2: Automatic extraction and fusion of multidimensional features, using deep neural networks to automatically mine deep patterns in the data;

[0017] Temporal feature extraction: Convolutional neural network is used to scan the input sequence to extract short-term fluctuation patterns of optical power, and multi-head self-attention mechanism is used to capture long-distance dependencies and identify long-term aging trends and seasonal variation patterns of optical power.

[0018] Environmental correlation feature fusion introduces a gating fusion mechanism, using external environmental variables as condition vectors to dynamically adjust the weights of time-series features;

[0019] Step A3: Based on the hybrid architecture, predict the health status by using the trained hybrid model to infer the future state and output specific health indicators;

[0020] The hidden state vector is input into the decoder, and the model autoregressively generates a sequence of optical power predictions for future steps.

[0021] Using Monte Carlo Dropout or quantile regression techniques, the confidence intervals of the predicted values ​​are output, and the health index is calculated.

[0022] Step A4: Automatic adjustment of dynamic thresholds. Based on historical data from the same period and recent trends, statistical methods or extreme value theory are used to calculate the upper and lower bounds of the dynamic normal range at the current moment, and threshold corrections are made based on time, environment, and equipment dimensions.

[0023] Step A5: Abnormal pattern recognition, combining health indicators and dynamic thresholds, performing logical judgments, and generating early warning signals.

[0024] Optionally, step S3 is performed in the following manner:

[0025] Step B1: Receive early warning signals and dynamically adjust the test strategy according to the warning level and link characteristics;

[0026] Step B2: OTDR curve preprocessing and 2D image conversion, converting the one-dimensional electrical signal curve into an image format suitable for convolutional neural network processing, and removing basic noise;

[0027] Step B3: Event detection and artifact filtering based on convolutional neural networks. The trained deep convolutional neural network is used to perform semantic segmentation and object detection on OTDR images to accurately identify real events.

[0028] Step B4: High-precision calculation of the physical distance to the fault point, converting the image pixel coordinates into the actual physical distance, and correcting it by combining the optical cable routing data;

[0029] Step B5: Fault root cause classification by multi-source data fusion, combined with OTDR images and smart optical module data for comprehensive analysis.

[0030] Optionally, in step S3, the fault classification model employs the XGBoost or Random Forest algorithm, and its input feature vector includes:

[0031] The slope of optical power decrease, the change in bias current, the OTDR event loss value, the height of the reflection peak, the geological environment characteristics of the area where the fault point is located, and historical fault records before and after the fault occurred.

[0032] The output results are the root cause categories of the fault, which include at least: complete fiber optic interruption, fiber microbending loss, loose / contaminated connectors, detached pigtails, and damaged optical modules on the equipment side.

[0033] Optionally, step S4 is implemented in the following manner:

[0034] Step C1: Construct a dynamic fault propagation knowledge graph, which integrates static network resource data with dynamic real-time alarm data to build a graph structure that can reflect the fault propagation logic;

[0035] Step C2: Alarm clustering and root cause reasoning based on graph neural networks. Deep learning algorithms are used to pass messages on the graph structure to automatically identify alarm clusters and locate the source.

[0036] Step C3: High-precision spatial mapping of the fault point using GIS, converting the logical root cause node into precise geographic coordinates;

[0037] Step C4: Intelligent work order generation assisted by large models, which uses a large AI language model in the vertical field to transform structured data into natural language processing solutions.

[0038] Optionally, the step S4 of constructing the fault propagation topology based on the graph neural network specifically includes:

[0039] A directed weighted graph is constructed using network elements, ports, optical cable segments, and junction boxes as nodes, and physical connection relationships and signal flow directions as edges.

[0040] The alarm status features of neighboring nodes are aggregated using a graph attention mechanism, and the probability score of each node as the root cause is calculated.

[0041] The physical distance of the node with the highest probability score and its associated fault point is mapped to the GIS map, and information on surrounding manholes, markers, and construction areas is overlaid and displayed.

[0042] Optionally, step S5 is implemented in the following manner:

[0043] Step D1: Standardized collection of multimodal field feedback data. After the fault is repaired, collect the true data of front-line maintenance personnel in a structured manner through mobile APP or handheld terminal.

[0044] Step D2: Data cleaning, labeling, and quality assessment, transforming unstructured field feedback into high-quality training samples that machines can understand, and performing rigorous quality filtering;

[0045] Step D3: Based on the characteristics of different models, adopt differentiated incremental training strategies to update parameters.

[0046] An AI-based communication optical cable anomaly prediction and fault location system, employing the AI-based communication optical cable anomaly prediction and fault location method, includes a multi-dimensional data acquisition and preprocessing module, an anomaly trend prediction module, a fault accurate location and classification module, a root cause analysis and visualization mapping module, and a closed-loop feedback optimization module.

[0047] The multi-dimensional data acquisition and preprocessing module is used to acquire multi-source heterogeneous data from communication optical cable links in real time, clean, denoise and timestamp align the data, and build a standardized optical cable status dataset.

[0048] The abnormal trend prediction module is used to input standardized datasets into the abnormal prediction model, extract optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period, and generate an early warning signal when the predicted health is lower than the dynamic threshold or abnormal fluctuations are detected.

[0049] The fault location and classification module is used to respond to early warning signals, automatically trigger OTDR testing to obtain backscatter curves, use convolutional neural networks to identify curve images, distinguish between real fault events and noise artifacts, calculate the physical distance of the fault point, and combine with telemetry data from the intelligent optical module to determine the specific fault type through a fault classification model.

[0050] The root cause analysis and visualization mapping module is used to construct the fault propagation topology using graph neural networks, perform cluster analysis on associated alarms to locate the root cause fault, map the physical distance of the fault point to the GIS geographic information system, and generate a visualized fault work order containing the fault latitude and longitude, fault type and surrounding environment information.

[0051] The closed-loop feedback optimization module is used to receive the actual fault handling results reported by on-site maintenance personnel as labeled data, update the optical cable status dataset with new data, and perform incremental training and parameter iteration on the anomaly prediction model, fault classification model and convolutional neural network.

[0052] The beneficial effects of this invention are:

[0053] 1. In this invention, the full-process closed-loop management constructs a complete closed loop from data collection, anomaly prediction, precise location, root cause analysis to feedback optimization, which solves the problem of fragmented links in traditional operation and maintenance. Through the combination of prediction, location, classification and visualization, it realizes the transformation from passive emergency repair to proactive prevention, significantly reduces the risk of business interruption, improves operation and maintenance efficiency, and introduces a closed-loop feedback mechanism, enabling the system to continuously iterate and optimize the model using real field data, adapt to changes in the network environment, and maintain long-term high accuracy.

[0054] 2. In this invention, the data dimensions are comprehensive: it integrates data from the electrical layer, physical layer, and management layer, eliminating the blind spots of a single data source. The complementarity of multi-dimensional data (such as combining vibration signals with optical power) can effectively eliminate environmental interference and improve the confidence of fault diagnosis.

[0055] 3. In this invention, information is aggregated on the topology graph using graph neural networks, which can quickly pinpoint the source from massive concurrent alarms, preventing alarm storms from overwhelming the true cause. Abstract fault points are directly mapped to GIS maps and overlaid with the surrounding environment, greatly shortening the on-site addressing time. Natural language solutions are automatically generated using large language models, which lowers the cognitive threshold for operation and maintenance personnel and improves the level of standardized operations. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0057] Figure 2 This is a schematic diagram of step S2 of the present invention.

[0058] Figure 3 This is a schematic diagram of step S3 of the present invention.

[0059] Figure 4 This is a schematic diagram of step S4 of the present invention.

[0060] Figure 5 This is a schematic diagram of step S5 of the present invention.

[0061] Figure 6 This is a schematic diagram of a system structure according to the present invention. Detailed Implementation

[0062] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] like Figures 1 to 5 As shown, an AI-based method for anomaly prediction and fault location in optical fiber communication cables includes the following:

[0064] Step S1: Multidimensional data acquisition and preprocessing. Real-time acquisition of multi-source heterogeneous data from the communication optical cable link. Cleaning, noise reduction and timestamp alignment of the multi-source heterogeneous data to construct a standardized optical cable status dataset.

[0065] Multi-source heterogeneous data includes at least millisecond-level telemetry data from intelligent optical modules, distributed fiber optic sensing data, optical time domain reflectometer test waveform data, network management system alarm logs, and geographic information system data.

[0066] The millisecond-level telemetry data of the intelligent optical module includes: received optical power, transmitted optical power, bias current, module temperature and power supply voltage. The distributed optical fiber sensing data includes acoustic vibration signals and temperature change signals distributed along the optical cable.

[0067] The timestamp alignment process involves resampling multi-source data with different sampling frequencies to a preset time granularity, based on the network standard time protocol.

[0068] Step S2: Abnormal trend prediction. Input the optical cable status dataset into the pre-trained abnormal prediction model, extract the optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period in the future, and generate an early warning signal when the predicted health index is lower than the dynamic threshold or an abnormal fluctuation pattern is detected.

[0069] Before inputting data into the model, the problem of inconsistent sampling frequencies of multi-source data must be solved, and an input structure suitable for time series models must be constructed.

[0070] Time synchronization and resampling: Based on the network standard time, the millisecond-level telemetry data, meteorological data, and construction plan data of the intelligent optical module are uniformly mapped to the same time axis. Using linear interpolation or spline interpolation, all data are resampled to a uniform time granularity (e.g., one data point every 5 minutes) to form a complete multidimensional time series matrix.

[0071] Constructing a sliding window sample: Set the sliding window length L, such as data from the past 7 days, i.e., 2016 time points, and the prediction step size H, such as the next 24 hours. Cut the continuous time series into overlapping sample pairs (Xt−L:t,Yt:t+H), where X is the historical observation sequence and Y is the future optical power sequence to be predicted.

[0072] Data normalization: Z-score normalization is performed on features of different dimensions (such as optical power dBm, temperature ℃, humidity %) to eliminate the influence of dimensions and accelerate model convergence.

[0073] Deep neural networks are used to automatically uncover deep patterns in data, rather than relying on manual rules.

[0074] Temporal feature extraction (local and global):

[0075] The input sequence is scanned using a convolutional neural network (1D-CNN) layer to extract short-term fluctuation patterns of optical power (such as momentary jitter caused by micro-bending and periodic diurnal temperature variation).

[0076] Multi-head self-attention is used to capture long-distance dependencies and identify long-term aging trends in optical power (such as slow attenuation caused by hydrogen loss in optical fibers) as well as seasonal variation patterns.

[0077] Environmental Feature Fusion: A gating fusion mechanism is introduced, using external environmental variables (rainfall, wind speed, and vibration intensity of surrounding construction) as condition vectors to dynamically adjust the weights of time-series features.

[0078] Example logic: When the "heavy rain" feature is detected, the model automatically increases its sensitivity to the "sudden change in light power" feature; when "no construction" and "sunny weather" are detected, the tolerance for short-term fluctuations is reduced, and the focus is on long-term trends.

[0079] Latent state encoding: The extracted temporal features and environmental features are concatenated and mapped into a high-dimensional latent state vector through a fully connected layer. This vector contains a comprehensive mathematical expression of the current optical cable operating state.

[0080] The trained hybrid model is used to predict future states and output specific health indicators.

[0081] Future sequence generation: The hidden state vector is input into the decoder part, and the model autoregressively generates a sequence of optical power prediction values ​​for the next H steps.

[0082] Uncertainty quantification: In addition to outputting a single predicted value, Monte Carlo Dropout or quantile regression techniques are used to output the confidence interval of the predicted value. This can reflect the model's "confidence" in the prediction result. When the data noise is high or the working conditions are complex, the confidence interval will automatically widen.

[0083] Calculation of the Health Index: Definition of Optical Cable Health Index (OHI):

[0084] ,

[0085] In the formula, To actually observe the optical power, To predict optical power, The maximum permissible deviation threshold, This is the deviation weighting coefficient. The trend slope The critical deterioration slope, This is the trend weighting coefficient.

[0086] Abandon fixed alarm thresholds and establish a dynamic threshold system that changes with the environment and equipment status.

[0087] Baseline dynamic modeling: Based on historical data from the same period and recent trends, statistical methods or extreme value theory are used to calculate the upper and lower bounds of the dynamic normal range at the current moment.

[0088] Multi-dimensional threshold correction:

[0089] Time-based adjustments: Thresholds are tightened during off-peak hours at night and appropriately relaxed during peak hours during the day to avoid false alarms.

[0090] Environmental dimension correction: In extreme weather conditions such as typhoons and rainstorms, the allowable threshold for optical power fluctuations is automatically increased because physical disturbances are expected, but the detection of "irreversible degradation" is strengthened.

[0091] Equipment-level correction: For optical modules of different brands and service years, thresholds are set individually based on their historical noise levels.

[0092] The final judgment logic is executed by combining the prediction results with the dynamic threshold.

[0093] Dual trigger mechanism judgment:

[0094] Condition A triggers when the predicted future optical power value is lower than the dynamic lower limit threshold, or the predicted health index (OHI) is lower than the preset warning line, such as 0.8.

[0095] Condition B trigger: Even if the value does not exceed the limit, a specific abnormal fluctuation pattern is detected, such as: a continuous monotonous decrease of N points, a sudden increase in the amplitude of high-frequency oscillations, or a step-like drop.

[0096] Early warning classification and signal generation:

[0097] Level 1 Warning: If condition B is met, or the predicted value is close to the threshold with a margin of <3dB, a "Performance Degradation Alert" will be generated and pushed to the monitoring dashboard. Closer monitoring is recommended.

[0098] Level 2 Early Warning Intervention: If condition A is met and the predicted fault time is within 12-24 hours, a "Maintenance Work Order" is generated, prompting maintenance personnel to prepare fusion splicers or spare fiber cores in advance and arrange inspections.

[0099] Level 3 Emergency Warning: The predicted failure time is within 1-2 hours, or a precipitous drop in performance is detected. This directly triggers the "Emergency Repair Plan," notifying the nearest external team to stand by and automatically initiating the protection switchover process.

[0100] Warning information encapsulation:

[0101] The generated early warning signal includes: predicted failure time, expected failure location, suspected main causes, and suggested remedial measures.

[0102] Step S3: Accurate fault location and classification. In response to the early warning signal, the OTDR test is automatically triggered to obtain the current backscatter curve. The backscatter curve is image-recognized using a convolutional neural network to distinguish between real fault events and noise artifacts. The physical distance of the fault point is calculated. Combined with the telemetry data of the intelligent optical module and the physical distance of the fault point, the fault type is determined by the fault classification model.

[0103] After receiving an early warning signal from the abnormal trend prediction module, the system does not conduct tests blindly, but dynamically adjusts the testing strategy based on the warning level and link characteristics.

[0104] Early warning signal analysis and link locking:

[0105] Analyze the key information in the warning signal: the optical cable route ID involved, the affected optical module port, the estimated direction of degradation, query the resource management system, and locate the specific fiber core sequence and remote equipment information corresponding to the port.

[0106] Adaptive test parameter generation:

[0107] Pulse width selection: The system automatically selects the optimal pulse width based on the estimated fault distance or total fiber optic cable length, using a narrow pulse width for short distances and a wide pulse width for long distances.

[0108] Wavelength selection: Automatically switches to dual-wavelength test mode of 1550nm and 1625nm to distinguish between macrobending loss and breakage.

[0109] Averaging time setting: The number of averages is dynamically adjusted based on the current signal-to-noise ratio. If the initial scan has high noise, the averaging time is automatically increased to improve curve smoothness and ensure AI recognition accuracy.

[0110] Seamless test execution: Through the built-in or external OTDR module, online testing is performed during business idle time slots or using a WDM coupler to obtain the current backscatter curve data.

[0111] The one-dimensional electrical signal curve is converted into an image format suitable for convolutional neural network processing, and the underlying noise is removed.

[0112] Data cleaning and denoising: Wavelet transform or Kalman filter algorithms are applied to filter out high-frequency random noise in the curve, while preserving the edge features of event points, correcting baseline drift, and ensuring that the level of the curve's start and end points is normalized.

[0113] One-dimensional to two-dimensional image mapping: Mapping the cleaned one-dimensional distance-loss sequence to a two-dimensional grayscale image or pseudo-color image.

[0114] Channel enhancement: If dual-wavelength testing is performed, the two curves at 1550nm and 1625nm are used as two channels of the RGB image to construct a multi-channel feature map for CNN to extract wavelength difference features.

[0115] By using a trained deep convolutional neural network to perform semantic segmentation and object detection on OTDR images, real events can be accurately identified.

[0116] Feature extraction and event candidate region generation: Input the two-dimensional map into the backbone network, extract multi-scale feature maps, and use the region proposal network or anchor box mechanism to mark all potential event regions on the curve, including reflection events and non-reflection events.

[0117] True / False Event Classification: Each candidate region is judged using a classification header to distinguish the following categories:

[0118] Real faults: breakpoints, moving connectors, mechanical joints, macro bends.

[0119] Noise artifacts: ghosting, gain phenomena, Rayleigh scattering noise spikes.

[0120] Judgment logic: The features learned by CNN include deep patterns such as "ghosting usually does not have a corresponding secondary reflection" and "gain phenomenon will disappear in reverse testing", thereby automatically eliminating false positives.

[0121] Refined regression of event parameters: Through the regression header, the start and end points of each real event are accurately output.

[0122] Converting image pixel coordinates into actual physical distances and correcting them in conjunction with fiber optic cable routing data is crucial. Sometimes, relying solely on OTDRs is insufficient to distinguish subtle faults, such as slight bending versus dirty connectors. A comprehensive assessment using data from intelligent optical modules is necessary.

[0123] Constructing multidimensional feature vectors:

[0124] OTDR characteristics: event loss value, reflection peak height, event waveform shape factor, and loss difference between two wavelengths. The loss difference between 1550nm and 1625nm is the key to judging macrobending, and the loss is significantly greater at 1625nm.

[0125] Telemetry characteristics of optical modules: slope of Rx / Tx power change before and after the fault, whether the bias current increases abnormally, and sudden increase in bit error rate (BER).

[0126] Environmental characteristics: Are there any construction records or meteorological data at the location of the fault?

[0127] Ensemble classification model inference: The above feature vectors are input into a gradient boosting tree model or a multilayer perceptron. The model outputs a probability distribution of fault types, with the main categories including:

[0128] Fiber breakage: OTDR displays high reflection at the end, and optical power drops sharply to the background noise.

[0129] Macrobending loss: OTDR shows step loss with no obvious reflection, and the loss at 1625nm is much greater than that at 1550nm, resulting in a slow or step-like decrease in optical module power.

[0130] Dirty / loose connectors: OTDR shows high reflection peaks accompanied by large insertion loss, and the optical module power fluctuates greatly.

[0131] Water ingress / aging at junction box: OTDR shows that the loss at the junction box is gradually increasing, with no obvious reflection.

[0132] Equipment-side fault: The OTDR curve is normal, but the optical module bias current is abnormal or there is no light output.

[0133] Confidence assessment: Outputs the confidence score of the classification result. If the confidence score is below the threshold, the system marks it as a suspected complex fault, and manual review is recommended.

[0134] Step S4: Root cause analysis and visualization mapping. Based on graph neural network, a fault propagation topology is constructed. Cluster analysis is performed on associated alarms to locate the root cause fault. The physical distance of the fault point is mapped to the GIS geographic information system to generate a visualized fault work order containing fault latitude and longitude, fault type and surrounding environment information.

[0135] By fusing static network resource data with dynamic real-time alarm data, a graph structure that reflects the fault propagation logic is constructed. Deep learning algorithms are then used to pass messages on the graph structure, automatically identifying alarm clusters and locating their sources.

[0136] Node feature embedding transforms the attributes of each node into a high-dimensional feature vector. For time-series alarms, a time decay factor is introduced to give higher weight to recent alarms.

[0137] Graph attention message passing employs a graph attention network model, which aggregates the feature information of its neighboring nodes for each alarm node in the graph.

[0138] The model automatically learns the importance of different neighbors. For example, when a "fiber optic cable break" alarm is connected to multiple "port LOS" alarms, the model will give the "fiber optic cable break" node a higher attention weight and reduce the weight of derived alarms.

[0139] Root cause scoring and clustering, after passing through multiple GNN convolutions, outputs the probability score of each node as a "root cause". The node with the highest score is the inferred root cause fault point. Based on the structural connectivity and node similarity of the graph, spectral clustering or community detection algorithms are used to divide the originally scattered massive alarms into several independent "fault event clusters". Each cluster corresponds to an independent physical fault source.

[0140] Once the root cause node in a cluster is identified, the system automatically marks all other low-scoring nodes in that cluster as "derived alarms" or "accompanying alarms," ​​and then collapses or mutes them in subsequent displays to avoid interfering with maintenance personnel.

[0141] The logical root cause node is converted into precise geographic coordinates to obtain the fault logic location output by the root cause analysis, such as: 3.45km away from A in the direction from network element A to network element B.

[0142] Call the GIS engine to load the digital routing vector data of the optical cable segment, including the starting point, ending point latitude and longitude, and length of each segment.

[0143] Starting from the coordinates of the initial equipment room, the calculation is performed by accumulating along the vector line of the optical cable route. When the accumulated distance reaches 3.45km, the precise latitude and longitude coordinates of the point on the Earth's surface are calculated using a linear interpolation algorithm.

[0144] If the calculation point falls in a non-road / non-tower area, it will automatically snap to the nearest possible facility point, such as a river-crossing tower or manhole, taking into account the uncertainty range of OTDR measurements.

[0145] Within a radius of 50-100 meters centered on the fault point, query and overlay the following layer information:

[0146] Infrastructure: Manhole number, utility pole number, marker stone location.

[0147] Risk sources: nearby construction sites, high-voltage power lines, and flood-prone areas.

[0148] Transportation network: the nearest accessible roads and parking lots.

[0149] By leveraging AI-powered large language models in vertical fields, we can transform structured data into natural language, ensuring that information reaches the relevant responsible parties accurately and quickly.

[0150] Step S5: Closed-loop feedback optimization. Receive the actual fault handling results from on-site maintenance personnel, update the optical cable status dataset as labeled data, and perform incremental training and parameter iteration on the anomaly prediction model, fault classification model, and convolutional neural network.

[0151] Implemented in the following ways:

[0152] Standardized collection of multimodal field feedback data allows for the structured collection of ground truth data from frontline maintenance personnel via mobile apps or handheld terminals after fault repair is completed.

[0153] Data cleaning, labeling, and quality assessment transform unstructured field feedback into high-quality training samples that machines can understand, and then perform rigorous quality filtering.

[0154] Different incremental training strategies are adopted to update parameters based on the characteristics of different models.

[0155] like Figure 6 As shown, an AI-based communication optical cable anomaly prediction and fault location system employs the AI-based communication optical cable anomaly prediction and fault location method, including a multi-dimensional data acquisition and preprocessing module, an anomaly trend prediction module, a fault accurate location and classification module, a root cause analysis and visualization mapping module, and a closed-loop feedback optimization module.

[0156] The multi-dimensional data acquisition and preprocessing module is used to acquire multi-source heterogeneous data from communication optical cable links in real time, clean, denoise and timestamp align the data, and build a standardized optical cable status dataset.

[0157] The abnormal trend prediction module is used to input standardized datasets into the abnormal prediction model, extract optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period, and generate an early warning signal when the predicted health is lower than the dynamic threshold or abnormal fluctuations are detected.

[0158] The fault location and classification module is used to respond to early warning signals, automatically trigger OTDR testing to obtain backscatter curves, use convolutional neural networks to identify curve images, distinguish between real fault events and noise artifacts, calculate the physical distance of the fault point, and combine with telemetry data from the intelligent optical module to determine the specific fault type through a fault classification model.

[0159] The root cause analysis and visualization mapping module is used to construct the fault propagation topology using graph neural networks, perform cluster analysis on associated alarms to locate the root cause fault, map the physical distance of the fault point to the GIS geographic information system, and generate a visualized fault work order containing the fault latitude and longitude, fault type and surrounding environment information.

[0160] The closed-loop feedback optimization module receives actual fault handling results from on-site maintenance personnel as labeled data, updates the fiber optic cable status dataset with new data, and performs incremental training and parameter iteration on the anomaly prediction model, fault classification model, and convolutional neural network.

[0161] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based method for anomaly prediction and fault location in optical fiber communication cables, characterized in that, Includes the following steps: Step S1: Multidimensional data acquisition and preprocessing. Real-time acquisition of multi-source heterogeneous data from the communication optical cable link. Cleaning, noise reduction and timestamp alignment of the multi-source heterogeneous data to construct a standardized optical cable status dataset. Step S2: Abnormal trend prediction. Input the optical cable status dataset into the pre-trained abnormal prediction model, extract the optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period in the future, and generate an early warning signal when the predicted health index is lower than the dynamic threshold or an abnormal fluctuation pattern is detected. Step S3: Accurate fault location and classification. In response to the early warning signal, the OTDR test is automatically triggered to obtain the current backscatter curve. The backscatter curve is image-recognized using a convolutional neural network to distinguish between real fault events and noise artifacts. The physical distance of the fault point is calculated. Combined with the telemetry data of the intelligent optical module and the physical distance of the fault point, the fault type is determined by the fault classification model. Step S4: Root cause analysis and visualization mapping. Based on graph neural network, a fault propagation topology is constructed. Cluster analysis is performed on associated alarms to locate the root cause fault. The physical distance of the fault point is mapped to the GIS geographic information system to generate a visualized fault work order containing fault latitude and longitude, fault type and surrounding environment information. Step S5: Closed-loop feedback optimization. Receive the actual fault handling results from on-site maintenance personnel, update the optical cable status dataset as labeled data, and perform incremental training and parameter iteration on the anomaly prediction model, fault classification model, and convolutional neural network.

2. The AI-based method for anomaly prediction and fault location of optical fiber communication cables according to claim 1, characterized in that, The multi-source heterogeneous data in step S1 includes at least millisecond-level telemetry data from the intelligent optical module, distributed optical fiber sensing data, optical time domain reflectometer test waveform data, network management system alarm logs, and geographic information system data.

3. The AI-based method for anomaly prediction and fault location of communication optical cables according to claim 2, characterized in that, The millisecond-level telemetry data of the intelligent optical module includes: received optical power, transmitted optical power, bias current, module temperature and power supply voltage; the distributed optical fiber sensing data includes acoustic vibration signals and temperature change signals distributed along the optical cable. The timestamp alignment process is as follows: based on the network standard time protocol, multi-source data with different sampling frequencies are uniformly resampled to a preset time granularity.

4. The AI-based method for anomaly prediction and fault location of optical fiber communication cables according to claim 1, characterized in that, Step S2 is implemented in the following manner: Step A1: Multi-source data alignment and sliding window construction. Using linear interpolation or spline interpolation, all data are resampled to a uniform time granularity to form a complete multidimensional time series matrix. The sliding window length and prediction step size are set to cut the continuous time series into overlapping sample pairs. Step A2: Automatic extraction and fusion of multidimensional features, using deep neural networks to automatically mine deep patterns in the data; Temporal feature extraction: Convolutional neural network is used to scan the input sequence to extract short-term fluctuation patterns of optical power, and multi-head self-attention mechanism is used to capture long-distance dependencies and identify long-term aging trends and seasonal variation patterns of optical power. Environmental correlation feature fusion introduces a gating fusion mechanism, using external environmental variables as condition vectors to dynamically adjust the weights of time-series features; Step A3: Based on the hybrid architecture, predict the health status by using the trained hybrid model to infer the future state and output specific health indicators; The hidden state vector is input into the decoder, and the model autoregressively generates a sequence of optical power predictions for future steps. Using Monte Carlo Dropout or quantile regression techniques, the confidence intervals of the predicted values ​​are output, and the health index is calculated. Step A4: Automatic adjustment of dynamic thresholds. Based on historical data from the same period and recent trends, statistical methods or extreme value theory are used to calculate the upper and lower bounds of the dynamic normal range at the current moment, and threshold corrections are made based on time, environment, and equipment dimensions. Step A5: Abnormal pattern recognition, combining health indicators and dynamic thresholds, performing logical judgments, and generating early warning signals.

5. The AI-based method for anomaly prediction and fault location in communication optical cables according to claim 1, characterized in that, Step S3 is performed in the following manner: Step B1: Receive early warning signals and dynamically adjust the test strategy according to the warning level and link characteristics; Step B2: OTDR curve preprocessing and 2D image conversion, converting the one-dimensional electrical signal curve into an image format suitable for convolutional neural network processing, and removing basic noise; Step B3: Event detection and artifact filtering based on convolutional neural networks. The trained deep convolutional neural network is used to perform semantic segmentation and object detection on OTDR images to accurately identify real events. Step B4: High-precision calculation of the physical distance to the fault point, converting the image pixel coordinates into the actual physical distance, and correcting it by combining the optical cable routing data; Step B5: Fault root cause classification by multi-source data fusion, combined with OTDR images and smart optical module data for comprehensive analysis.

6. The AI-based method for anomaly prediction and fault location in communication optical cables according to claim 1, characterized in that, In step S3, the fault classification model uses either XGBoost or Random Forest algorithm, and its input feature vector includes: The slope of optical power decrease, the change in bias current, the OTDR event loss value, the height of the reflection peak, the geological environment characteristics of the area where the fault point is located, and historical fault records before and after the fault occurred. The output results are the root cause categories of the fault, which include at least: complete fiber optic interruption, fiber microbending loss, loose / contaminated connectors, detached pigtails, and damaged optical modules on the equipment side.

7. The AI-based method for anomaly prediction and fault location of communication optical cables according to claim 1, characterized in that, Step S4 is implemented in the following manner: Step C1: Construct a dynamic fault propagation knowledge graph, which integrates static network resource data with dynamic real-time alarm data to build a graph structure that can reflect the fault propagation logic; Step C2: Alarm clustering and root cause reasoning based on graph neural networks. Deep learning algorithms are used to pass messages on the graph structure to automatically identify alarm clusters and locate the source. Step C3: High-precision spatial mapping of the fault point using GIS, converting the logical root cause node into precise geographic coordinates; Step C4: Intelligent work order generation assisted by large models, which uses a large AI language model in the vertical field to transform structured data into natural language processing solutions.

8. The AI-based method for anomaly prediction and fault location of communication optical cables according to claim 1, characterized in that, The specific steps in step S4, including constructing the fault propagation topology based on the graph neural network, include: A directed weighted graph is constructed using network elements, ports, optical cable segments, and junction boxes as nodes, and physical connection relationships and signal flow directions as edges. The alarm status features of neighboring nodes are aggregated using a graph attention mechanism, and the probability score of each node as the root cause is calculated. The physical distance of the node with the highest probability score and its associated fault point is mapped to the GIS map, and information on surrounding manholes, markers, and construction areas is overlaid and displayed.

9. The AI-based method for anomaly prediction and fault location of communication optical cables according to claim 1, characterized in that, Step S5 is implemented in the following manner: Step D1: Standardized collection of multimodal field feedback data. After the fault is repaired, collect the true data of front-line maintenance personnel in a structured manner through mobile APP or handheld terminal. Step D2: Data cleaning, labeling, and quality assessment, transforming unstructured field feedback into high-quality training samples that machines can understand, and performing rigorous quality filtering; Step D3: Based on the characteristics of different models, adopt differentiated incremental training strategies to update parameters.

10. An AI-based communication optical cable anomaly prediction and fault location system, employing the AI-based communication optical cable anomaly prediction and fault location method as described in any one of claims 1-9, characterized in that, It includes a multi-dimensional data acquisition and preprocessing module, an anomaly trend prediction module, a fault accurate location and classification module, a root cause analysis and visualization mapping module, and a closed-loop feedback optimization module; The multi-dimensional data acquisition and preprocessing module is used to acquire multi-source heterogeneous data from communication optical cable links in real time, clean, denoise and timestamp align the data, and build a standardized optical cable status dataset. The abnormal trend prediction module is used to input standardized datasets into the abnormal prediction model, extract optical power time series features and environmental correlation features, predict the optical cable health index within a preset time period, and generate an early warning signal when the predicted health is lower than the dynamic threshold or abnormal fluctuations are detected. The fault location and classification module is used to respond to early warning signals, automatically trigger OTDR testing to obtain backscatter curves, use convolutional neural networks to identify curve images, distinguish between real fault events and noise artifacts, calculate the physical distance of the fault point, and combine with telemetry data from the intelligent optical module to determine the specific fault type through a fault classification model. The root cause analysis and visualization mapping module is used to construct the fault propagation topology using graph neural networks, perform cluster analysis on associated alarms to locate the root cause fault, map the physical distance of the fault point to the GIS geographic information system, and generate a visualized fault work order containing the fault latitude and longitude, fault type and surrounding environment information. The closed-loop feedback optimization module is used to receive the actual fault handling results reported by on-site maintenance personnel as labeled data, update the optical cable status dataset with new data, and perform incremental training and parameter iteration on the anomaly prediction model, fault classification model and convolutional neural network.

Images