A data analysis-based submarine optical cable monitoring and early warning method and system

Abstract

The present application belongs to the field of monitoring and early warning of submarine optical cable, and particularly relates to a method and system for monitoring and early warning of submarine optical cable based on data analysis. The method comprises: obtaining a state feature vector by using historical signals of each sensing unit and monitoring signals of adjacent sensing units, initializing each state feature vector as an atomic cluster; calculating a fusion distance between any two atomic clusters, and determining a merging priority of the two atomic clusters based on the fusion distance and a topological proximity factor, repeating the merging of the two atomic clusters with the highest priority, and terminating the merging when a set condition is reached; after termination of the merging, if both the static abnormal score of any new cluster formed after termination of the merging and the dynamic curvature of the motion trajectory of the cluster centroid exceed their respective thresholds, the new cluster is determined as an abnormal event and an early warning is generated. The present application can more accurately and reliably early warn physical disturbances caused by various external factors that may lead to interruption of submarine optical cable.

Description

Technical Field

[0001] This invention belongs to the field of monitoring and early warning of submarine optical cables, and particularly relates to a method and system for monitoring and early warning of submarine optical cables based on data analysis. Background Technology

[0002] Submarine optical cables, as a core infrastructure for international communications, carry the majority of transoceanic data transmission tasks globally, and their operational stability directly determines the connectivity of international communication networks. However, submarine optical cables operate in complex marine environments characterized by high pressure, high salinity, and darkness, facing multiple external threats: First, mechanical damage threats, such as trawls during fishing operations or impacts from anchoring equipment, can directly cause damage to the cable's outer sheath and fiber breakage. Second, geological threats, such as subsurface displacement caused by submarine earthquakes and turbidity current erosion, can strain the cable through geological stress, leading to increased fiber loss. Third, biological interference threats, such as the attachment of marine shellfish and algae, can alter the acoustic properties of the cable surface, interfering with monitoring signals, and long-term attachment may corrode the cable's outer sheath. Fourth, environmental fluctuation threats, such as sudden changes in seawater temperature and ocean current impacts, can cause minute changes in the cable's physical properties, creating persistent signal noise. If these threats are not detected in time, they may lead to cable outages, causing large-scale communication paralysis and resulting in significant economic losses and social impact.

[0003] To address the aforementioned risks, the industry widely employs distributed fiber optic sensing technology for submarine optical cable monitoring. The core advantage of this technology lies in its ability to acquire long-distance, high-spatiotemporal resolution vibration or acoustic signals along the cable without requiring additional sensor deployments, providing high-dimensional raw data support for cable condition monitoring. However, in practical applications, existing monitoring technologies also face the following shortcomings:

[0004] The signals acquired by distributed fiber optic sensing technology are multi-dimensional spatiotemporal data, with a massive amount of data and containing a large amount of environmental noise. Traditional data processing methods struggle to effectively filter this noise, causing useful anomalous signals to be overwhelmed and making it impossible to accurately extract fault-related features. Furthermore, after acquiring signals using distributed fiber optic sensing technology, traditional clustering algorithms such as K-Means or DBSCAN are typically used for cluster analysis to automatically separate anomalous event signals from normal background noise. However, these algorithms lack sufficient differentiation for different types of events. Moreover, in terms of similarity measurement and clustering logic, traditional clustering algorithms often use simple metrics such as Euclidean distance and fail to consider the continuity of the sensing units along the fiber optic cable. This may lead to the erroneous aggregation of spatially discontinuous but occasionally similar noise points, making it difficult to distinguish between instantaneous, harmless disturbances and ongoing, real threats, thus affecting the reliability of early warning systems. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a data analysis-based method and system for monitoring and early warning of submarine optical cables, in order to solve the technical problem that the existing technology is insufficient in dealing with the complexity and dynamism of submarine optical cable monitoring, resulting in low reliability of early warning.

[0006] To address the aforementioned problems, the present invention provides a technical solution for a data analysis-based method for monitoring and early warning of submarine optical cables:

[0007] A data analysis-based method for monitoring and early warning of submarine optical cables includes the following steps:

[0008] Acquire the time-series monitoring signal of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, extract the historical signal of the sensing unit and the synchronous monitoring signal of the geographically adjacent sensing units from the time-series monitoring signal, and fuse them to construct a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information; initialize each state feature vector as an atom cluster;

[0009] For any two atomic clusters, the weights are determined based on the differences in the dispersion of their intra-cluster data in the temporal and spatial dimensions, and the weighted fusion distance is calculated by combining the point set distribution divergence and sequence morphological similarity. The merging priority of the two atomic clusters is determined based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing unit. The two atomic clusters with the highest merging priority among all cluster pairs are merged into a new cluster.

[0010] When the merging priority of all cluster pairs is lower than the termination threshold or the global partitioning quality index tends to stabilize, the merging is terminated. For any new cluster formed after the termination of merging, the following judgment is performed: the reconstruction error of the new cluster based on the local generation model is calculated and used as the static anomaly score; the motion trajectory of the centroid of the new cluster in the feature space is tracked at continuous monitoring time and its curvature is calculated; when the static anomaly score and the curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

[0011] Furthermore, the process of constructing the state feature vector includes:

[0012] Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients.

[0013] The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols.

[0014] Computation space second difference value ,in, , , These are the average signal values ​​of the current sensing unit and the two sensing units geographically adjacent to the current sensing unit;

[0015] The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

[0016] Furthermore, the calculation process for the fusion distance between any two atomic clusters includes:

[0017] The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ;

[0018] The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ;

[0019] Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ;

[0020] The formula for calculating the fusion distance D is: .

[0021] Furthermore, the determination of the merging priority of the two atomic clusters based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing units includes:

[0022] Calculate the minimum interval distance d between the sets of geographic location numbers of the sensing units contained in these two atomic clusters;

[0023] According to the formula Calculate the topological proximity factor T;

[0024] If the fusion distance D between two atomic clusters is 0, then these two atomic clusters are merged preferentially; if the fusion distance D between the two atomic clusters is not 0, then according to the formula... Obtain the merge priority P.

[0025] Furthermore, the method for determining whether the global partitioning quality index tends to stabilize is as follows:

[0026] After each merge, calculate the average silhouette coefficient value of all current clusters;

[0027] If, after three consecutive merging operations, the growth rate of the average contour coefficient is lower than the preset growth rate threshold, then the global segmentation quality index is determined to be stable.

[0028] Furthermore, the calculation steps for the static anomaly score include:

[0029] For any new cluster formed after the termination of merging, train an independent local generative model based on principal component analysis;

[0030] Project all high-dimensional state feature vectors within the new cluster onto a subspace composed of principal components whose cumulative contribution rate in the local generative model reaches a preset contribution rate threshold, and then reconstruct them back to the original dimensions.

[0031] Calculate the average Euclidean distance between the original vector and the reconstructed vector, and use the average Euclidean distance as the static anomaly score of the new cluster.

[0032] Furthermore, the step of tracking the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring times and calculating its curvature includes:

[0033] Record the centroid position vectors of any new cluster at three consecutive monitoring times as follows: , , ;

[0034] Using the coordinates of three high-dimensional points, the trajectory is calculated. Curvature of time The calculation formula is:

[0035]

[0036] in, Represents the dot product of vectors. Denotes the Euclidean norm of a vector; if , and If at least one of them is 0, then the curvature is directly set to 0. Set to 0.

[0037] The technical solution of the submarine optical cable monitoring and early warning system based on data analysis provided by this invention is as follows:

[0038] A data analysis-based submarine optical cable monitoring and early warning system includes the following modules:

[0039] An initialization module is used to acquire the time-series monitoring signals of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, the historical signal of the sensing unit and the synchronous monitoring signal of the sensing unit with its geographically adjacent sensing units are extracted from the aforementioned time-series monitoring signals, and a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information is constructed by fusing them; each state feature vector is initialized as an atom cluster.

[0040] The merging module is used to determine the weights of any two atomic clusters based on the differences in the dispersion of their intra-cluster data in the temporal and spatial dimensions, and to calculate the weighted fusion distance by combining the point set distribution divergence and sequence morphological similarity. Based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing unit, the merging priority of the two atomic clusters is determined. The two atomic clusters with the highest merging priority among all cluster pairs are merged into a new cluster.

[0041] The early warning module terminates merging when the merging priority of all cluster pairs is lower than the termination threshold or the global partitioning quality index tends to stabilize. For any new cluster formed after the termination of merging, the following judgment is performed: the reconstruction error of the new cluster based on the local generation model is calculated and used as the static anomaly score; the motion trajectory of the centroid of the new cluster in the feature space is tracked at continuous monitoring time and its curvature is calculated; when the static anomaly score and the curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

[0042] Furthermore, the process of constructing the state feature vector includes:

[0043] Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients.

[0044] The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols.

[0045] Computation space second difference value ,in, , , These are the average signal values ​​of the current sensing unit and the two sensing units geographically adjacent to the current sensing unit;

[0046] The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

[0047] Furthermore, the calculation process for the fusion distance between any two atomic clusters includes:

[0048] The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ;

[0049] The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ;

[0050] Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ;

[0051] The formula for calculating the fusion distance D is: .

[0052] Furthermore, the determination of the merging priority of the two atomic clusters based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing units includes:

[0053] Calculate the minimum interval distance d between the sets of geographic location numbers of the sensing units contained in these two atomic clusters;

[0054] According to the formula Calculate the topological proximity factor T;

[0055] If the fusion distance D between two atomic clusters is 0, then these two atomic clusters are merged preferentially; if the fusion distance D between the two atomic clusters is not 0, then according to the formula... Obtain the merge priority P.

[0056] Furthermore, the method for determining whether the global partitioning quality index tends to stabilize is as follows:

[0057] After each merge, calculate the average silhouette coefficient value of all current clusters;

[0058] If, after three consecutive merging operations, the growth rate of the average contour coefficient is lower than the preset growth rate threshold, then the global segmentation quality index is determined to be stable.

[0059] Furthermore, the calculation steps for the static anomaly score include:

[0060] For any new cluster formed after the termination of merging, train an independent local generative model based on principal component analysis;

[0061] Project all high-dimensional state feature vectors within the new cluster onto a subspace composed of principal components whose cumulative contribution rate in the local generative model reaches a preset contribution rate threshold, and then reconstruct them back to the original dimensions.

[0062] Calculate the average Euclidean distance between the original vector and the reconstructed vector, and use the average Euclidean distance as the static anomaly score of the new cluster.

[0063] Furthermore, the step of tracking the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring times and calculating its curvature includes:

[0064] Record the centroid position vectors of any new cluster at three consecutive monitoring times as follows: , , ;

[0065] Using the coordinates of three high-dimensional points, the trajectory is calculated. Curvature of time The calculation formula is:

[0066]

[0067] in, Represents the dot product of vectors. Denotes the Euclidean norm of a vector; if , and If at least one of them is 0, then the curvature is directly set to 0. Set to 0.

[0068] The beneficial effects of this invention are:

[0069] This invention calculates a fusion distance that combines the dispersion of point set distribution and the morphological similarity of sequences during the clustering process. This allows for a more accurate measurement of the similarity between data clusters. Furthermore, it introduces a topological proximity factor to calculate merging priority, prioritizing the merging of geographically proximate regions with similar signal characteristics. This overcomes the limitations of traditional metrics such as Euclidean distance, preventing the erroneous aggregation of physically distant but coincidentally similar noise points, thus ensuring the accuracy and reliability of the clustering results. Moreover, by combining a dual discrimination mechanism of static reconstruction error and cluster centroid trajectory curvature, this invention can detect potential anomalies at both the state and trend levels. This helps distinguish between continuously developing anomalies and transient, harmless disturbances, thereby reducing the false alarm rate and improving the reliability of early warnings. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of a data analysis-based submarine optical cable monitoring and early warning method according to the present invention. Detailed Implementation

[0071] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0072] A specific embodiment of the submarine optical cable monitoring and early warning method based on data analysis of the present invention:

[0073] like Figure 1 As shown, a data analysis-based method for monitoring and early warning of submarine optical cables includes the following steps:

[0074] S1, acquire the time-series monitoring signal of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, extract the historical signal of the sensing unit and the synchronous monitoring signal of the sensing unit with its geographically adjacent sensing units from the above time-series monitoring signal, and fuse them to construct a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information; initialize each state feature vector as an atom cluster.

[0075] Specifically, in this step, a distributed acoustic sensing demodulation device deployed at the shore station is used to emit coherent laser pulses into the submarine optical cable and receive Rayleigh backscattered light signals caused by external vibrations in the optical fiber. By demodulating the phase information of the Rayleigh backscattered light signals, the strain or vibration waveform of each spatial sampling point (i.e., each sensing unit) along the submarine optical cable is reconstructed over time, forming a data matrix with time as the vertical axis, cable distance as the horizontal axis, and signal amplitude as the value. For the signal of any sensing unit i at the current time t, a time series segment of length L consisting of its past L sampling points is extracted. M principal basis vectors are pre-extracted from historical normal operation data using principal component analysis. The current time series segment is projected onto these M basis vectors to obtain M time-domain projection coefficients. A symbolic aggregation approximation method is used to reduce the dimensionality of the time series segment and convert it into a symbolic sequence composed of letters. The current time t is then obtained, along with the data of sensing unit i and its geographically adjacent sensing units. and The signal amplitude is used to calculate the spatial second-order difference value. The M projection coefficients, the numerical representation of the symbol sequence, and the spatial second-order difference value are concatenated into a vector, which is the current state feature vector of the sensing unit.

[0076] In the initial stage of the clustering process, assuming there are N sensing units in the current monitoring area, N state feature vectors will be generated. Each independent state feature vector is treated as an independent atom cluster. Before the iterative merging begins, N atom clusters are initialized, and each atom cluster contains only one data point.

[0077] In a preferred embodiment, the process of constructing the state feature vector in step S1 includes:

[0078] Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients.

[0079] The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols.

[0080] Computation space second difference value ,in, , , These are the average signal values ​​of the current sensing unit and the two sensing units geographically adjacent to the current sensing unit;

[0081] The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

[0082] For example, assuming a sensing unit has 8 monitoring channels, collecting data once per second for 12 consecutive seconds, a 12x8 signal matrix is ​​obtained. Principal component analysis (PCA) compresses this 96-dimensional data space into the three most important dimensions. These three principal components contain over 90% of the original data variance, resulting in 3-dimensional vectors such as 2.5, -1.1, and 0.4, which serve as time-domain projection coefficients. The signal values ​​from the 8 monitoring channels per second over the 12 seconds are averaged to obtain a time series of length 12. To simplify the analysis, this is first shortened to 6 data points through segmented aggregation averaging. Based on the quantiles of a Gaussian distribution derived from extensive historical data, the data range is divided into six intervals: extremely low, relatively low, normal, relatively high, extremely high, and abnormally high. These 6 data points are mapped to corresponding symbol sequences, such as A, B, C, B, D, C, and then converted into integer codes, such as 0, 1, 2, 1, 3, 2. For spatial information, if the signal average of the current sensing unit i is 15.2, then the previous sensing unit... The average value is 15.0, the next sensing unit The mean is 15.5, so the second-order difference of the space is calculated to be 0.1 according to the formula above. Concatenating these three parts, we get a high-dimensional vector such as 2.5, -1.1, 0.4, 0, 1, 2, 1, 3, 2, 0.1.

[0083] S2. For any two atomic clusters, determine the weights based on the differences in the dispersion of their intra-cluster data in the temporal and spatial dimensions, and calculate the weighted fusion distance by combining the point set distribution divergence and sequence morphological similarity. Determine the merging priority of the two atomic clusters based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing unit. Merge the two atomic clusters with the highest merging priority among all cluster pairs into a new cluster.

[0084] In a preferred embodiment, the calculation process for the fusion distance between any two atomic clusters includes:

[0085] The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ;

[0086] The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ;

[0087] Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ;

[0088] The formula for calculating the fusion distance D is: .

[0089] For example, let's denote two atomic clusters as cluster A and cluster B. When determining whether to merge clusters A and B, we first assess the overall distribution difference between them in the feature space. Assuming cluster A has 20 data points and cluster B has 30 data points, the point set distribution divergence calculated using the maximum mean difference is 0.75, indicating a moderate degree of difference in the overall distribution of the state feature vectors of the two clusters. By calculating and averaging the dynamic time-normalized distances between all cross-cluster paired sequences, we obtain a sequence morphological similarity of 4.2. Assuming the variance of the temporally correlated feature (i.e., the time-domain projection coefficient) is 1.5, and the variance of the spatially correlated feature (i.e., the spatial second-order difference value) is 0.5, according to the above formula, the weighting coefficient α is 0.75. Therefore, a greater weight should be given to the point set distribution divergence when calculating the fusion distance, resulting in a final fusion distance D of 1.6125.

[0090] In a preferred embodiment, determining the merging priority of the two atom clusters based on fusion distance and a topological proximity factor representing the geographical continuity of sensing units includes:

[0091] Calculate the minimum interval distance d between the sets of geographic location numbers of the sensing units contained in these two atomic clusters;

[0092] According to the formula Calculate the topological proximity factor T;

[0093] If the fusion distance D between two atomic clusters is 0, then these two atomic clusters are merged preferentially; if the fusion distance D between the two atomic clusters is not 0, then according to the formula... Obtain the merge priority P.

[0094] For example, cluster X contains sensor units numbered 3, 4, and 5, and cluster Y contains sensor units numbered 7 and 8. The minimum distance d between them is 2. Substituting this into the above formula, the topological proximity factor T is approximately 0.018, very close to 0, indicating that the two clusters are geographically separated. Assuming the calculated merging distance D between clusters X and Y is 0.5, their merging priority score P is 0.036. Meanwhile, another pair of clusters M and N are evaluated, where cluster M contains units numbered 10 and 11, and cluster N contains unit numbered 12. Their minimum distance d is 1, and the calculated topological proximity factor T is approximately 0.368. Even if the merging distance D between clusters M and N is slightly larger, say 0.8, their merging priority score P is 0.46. 0.46 is much larger than 0.036, therefore, merging clusters M and N will be preferred.

[0095] S3, when the merging priority of all cluster pairs is lower than the termination threshold or the global partitioning quality index tends to stabilize, the merging is terminated; for any new cluster formed after the termination of merging, the following judgment is performed: calculate the reconstruction error of the new cluster based on the local generation model and use it as the static anomaly score; track the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring time and calculate its curvature; when the static anomaly score and the curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

[0096] In this step, before the start of each iteration, the mean and standard deviation of the merge priorities of all cluster pairs are calculated. The termination threshold is preferably the mean plus twice the standard deviation. If the highest merge priority calculated in this round is still lower than the termination threshold, it means there are no mergeable objects, and the merging process terminates.

[0097] In a preferred embodiment, the method for determining whether the global partitioning quality index tends to stabilize is as follows:

[0098] After each merge, calculate the average silhouette coefficient value of all current clusters;

[0099] If, after three consecutive merging operations, the growth rate of the average contour coefficient is lower than the preset growth rate threshold, then the global segmentation quality index is determined to be stable.

[0100] For example, in the initial stage of hierarchical clustering, each merge typically improves the rationality of the cluster structure, specifically reflected in a significant increase in the average silhouette coefficient value. For instance, with a preset growth rate threshold of 1%, after the 10th merge, the average silhouette coefficient is 0.650. After the 11th merge, the average silhouette coefficient increases to 0.680, a growth rate of 4.6%, which is higher than the termination threshold of 1%, indicating that this merge is effective. As the clustering process progresses, suppose that after the 50th merge, the average silhouette coefficient is 0.812. After the 51st merge, this value becomes 0.815, a growth rate of 0.37%, which is lower than the termination threshold of 1%, and this is recorded as the first low growth. After the 52nd merge, the average silhouette coefficient becomes 0.817, with a growth rate of approximately 0.25%, also lower than the termination threshold of 1%. After the 53rd merge, the average silhouette coefficient becomes 0.818, with a growth rate of approximately 0.12%, which is the third consecutive time it is below the termination threshold of 1%. At this point, the clustering quality has reached a plateau, and the improvement brought by subsequent merging is almost negligible. Therefore, the merging process is terminated, and the cluster structure formed after the 50th merging is taken as the final partitioning result.

[0101] In a preferred embodiment, the calculation steps of the static anomaly score include:

[0102] For any new cluster formed after the termination of merging, train an independent local generative model based on principal component analysis;

[0103] Project all high-dimensional state feature vectors within the new cluster onto a subspace composed of principal components whose cumulative contribution rate in the local generative model reaches a preset contribution rate threshold, and then reconstruct them back to the original dimensions.

[0104] Calculate the average Euclidean distance between the original vector and the reconstructed vector, and use the average Euclidean distance as the static anomaly score of the new cluster.

[0105] In a preferred embodiment, tracking the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring times and calculating its curvature includes:

[0106] Record the centroid position vector of any new cluster at three consecutive monitoring times. , , ;

[0107] Using the coordinates of three high-dimensional points, the trajectory is calculated. Curvature of time The calculation formula is:

[0108]

[0109] in, Represents the dot product of vectors. Denotes the Euclidean norm of a vector; if , and If at least one of them is 0, then the curvature is directly set to 0. Set to 0.

[0110] Calculation in The average value of the state feature vectors of all members of the stable cluster identified at each time step yields the centroid position vector of the cluster in the feature space. In the next monitoring cycle The member sensing units of this cluster will generate new state feature vectors, and their average values ​​will be calculated again to obtain the new centroid. Similarly, in Always get the center of mass Under normal operating conditions, the state of a group of devices changes slowly and smoothly, and its center of mass trajectory may approximate a straight line. For example, arrive The displacement vector is v1. arrive Let the displacement vector be v2. If the directions of v1 and v2 are very close, their dot product will be large, causing the numerator of the curvature formula to approach 0, resulting in a calculated curvature. The value will be very small. If a sudden event, such as a collective equipment failure, causes the cluster's state to... arrive Things are constantly changing, so the center of mass... It will deviate from the original direction of motion, making the direction of displacement vector v2 significantly different from v1, which will lead to curvature. An increase in the value of indicates an abnormal event that causes a significant change in the system's behavior pattern.

[0111] This invention addresses the issue of information loss and incomplete characterization of submarine optical cables caused by single feature extraction in traditional techniques. It utilizes multi-dimensional spatiotemporal data acquired through distributed optical fiber sensing technology to construct state feature vectors. This is achieved by fusing temporal projection coefficients reflecting historical fluctuation patterns, sequence symbolization preserving dynamic morphological characteristics, and spatial second-order difference information capturing the correlation characteristics along the cable route. This provides comprehensive data support for subsequent analysis. Furthermore, in the clustering stage, this invention overcomes the limitations of traditional algorithms that rely on single similarity metrics and ignore geographical continuity. It accurately measures inter-cluster differences by using a weighted fusion distance of point set distribution divergence and sequence morphological similarity. Combined with a topological proximity factor, it prioritizes merging geographically contiguous and feature-similar clusters. A dual-condition control mechanism for cluster termination prevents the erroneous aggregation of spatially irrelevant noise points, thus improving clustering accuracy. Finally, in the anomaly determination stage, this invention trains an independent local generation model for each new cluster after termination and merging, calculates a static anomaly score through reconstruction error, and adapts to the differences in the basic state of different optical cable segments. On the other hand, it tracks the motion trajectory of the cluster centroid in the feature space, uses curvature to reflect dynamic trend changes, and then performs dual threshold collaborative determination through static anomaly score and trajectory curvature to accurately distinguish between instantaneous harmless disturbances and continuous real threats, significantly reducing the false alarm and missed alarm rates. Ultimately, it achieves accurate identification and timely early warning of abnormal events in submarine optical cables, effectively ensuring the operational safety of core international communication infrastructure.

[0112] A specific embodiment of the submarine optical cable monitoring and early warning system based on data analysis according to the present invention:

[0113] A data analysis-based submarine optical cable monitoring and early warning system includes the following modules:

[0114] An initialization module is used to acquire the time-series monitoring signals of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, the historical signal of the sensing unit and the synchronous monitoring signal of the sensing unit with its geographically adjacent sensing units are extracted from the aforementioned time-series monitoring signals, and a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information is constructed by fusing them; each state feature vector is initialized as an atom cluster.

[0115] The merging module is used to determine the weights of any two atomic clusters based on the differences in the dispersion of their intra-cluster data in the temporal and spatial dimensions, and to calculate the weighted fusion distance by combining the point set distribution divergence and sequence morphological similarity. Based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing unit, the merging priority of the two atomic clusters is determined. The two atomic clusters with the highest merging priority among all cluster pairs are merged into a new cluster.

[0116] The early warning module terminates merging when the merging priority of all cluster pairs is lower than the termination threshold or the global partitioning quality index tends to stabilize. For any new cluster formed after the termination of merging, the following judgment is performed: the reconstruction error of the new cluster based on the local generation model is calculated and used as the static anomaly score; the motion trajectory of the centroid of the new cluster in the feature space is tracked at continuous monitoring time and its curvature is calculated; when the static anomaly score and the curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

[0117] In a preferred embodiment, the process of constructing the state feature vector includes:

[0118] Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients.

[0119] The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols.

[0120] Computation space second difference value ,in, , , These are the signal averages of the current sensing unit and the two adjacent sensing units in terms of geographical location.

[0121] The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

[0122] In a preferred embodiment, the calculation process for the fusion distance between any two atomic clusters includes:

[0123] The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ;

[0124] The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ;

[0125] Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ;

[0126] The formula for calculating the fusion distance D is: .

[0127] In a preferred embodiment, determining the merging priority of the two atom clusters based on fusion distance and a topological proximity factor representing the geographical continuity of sensing units includes:

[0128] Calculate the minimum interval distance d between the sets of geographic location numbers of the sensing units contained in these two atomic clusters;

[0129] According to the formula Calculate the topological proximity factor T;

[0130] If the fusion distance D between two atomic clusters is 0, then these two atomic clusters are merged preferentially; if the fusion distance D between the two atomic clusters is not 0, then according to the formula... Obtain the merge priority P.

[0131] In a preferred embodiment, the method for determining whether the global partitioning quality index tends to stabilize is as follows:

[0132] After each merge, calculate the average silhouette coefficient value of all current clusters;

[0133] If, after three consecutive merging operations, the growth rate of the average contour coefficient is lower than the preset growth rate threshold, then the global segmentation quality index is determined to be stable.

[0134] In a preferred embodiment, the calculation steps of the static anomaly score include:

[0135] For any new cluster formed after the termination of merging, train an independent local generative model based on principal component analysis;

[0136] Project all high-dimensional state feature vectors within the new cluster onto a subspace composed of principal components whose cumulative contribution rate in the local generative model reaches a preset contribution rate threshold, and then reconstruct them back to the original dimensions.

[0137] Calculate the average Euclidean distance between the original vector and the reconstructed vector, and use the average Euclidean distance as the static anomaly score of the new cluster.

[0138] In a preferred embodiment, tracking the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring times and calculating its curvature includes:

[0139] Record the centroid position vector of any new cluster at three consecutive monitoring times. , , ;

[0140] Using the coordinates of three high-dimensional points, the trajectory is calculated. Curvature of time The calculation formula is:

[0141]

[0142] in, Represents the dot product of vectors. Denotes the Euclidean norm of a vector; if , and If at least one of them is 0, then the curvature is directly set to 0. Set to 0.

[0143] In addition, a data analysis-based submarine optical cable monitoring and early warning system also includes other components well known to those skilled in the art, such as a processor, a memory, a communication bus, and a communication interface. The memory stores a computer program, and the processor executes the computer program through the aforementioned initialization module, merging module, and early warning module to realize the data analysis-based submarine optical cable monitoring and early warning method of the present invention.

[0144] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A data analysis-based method for monitoring and early warning of submarine optical cables, characterized in that, Includes the following steps: Acquire the time-series monitoring signal of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, extract the historical signal of the sensing unit and the synchronous monitoring signal of the sensing unit with its geographically adjacent sensing units from the time-series monitoring signal, and fuse them to construct a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information; initialize each state feature vector as an atom cluster; For any two atomic clusters, weights are determined based on the differences in the temporal and spatial dispersion of their intra-cluster data, and a weighted fusion distance is calculated by combining the point set distribution divergence and sequence morphological similarity. The merging priority of these two atomic clusters is determined based on the fusion distance and a topological proximity factor representing the geographical continuity of sensing units, including: Calculate the minimum interval distance d between the sets of geographic location numbers of the sensing units contained in these two atomic clusters; According to the formula Calculate the topological proximity factor T; If the fusion distance D between two atomic clusters is 0, then these two atomic clusters are merged preferentially; if the fusion distance D between the two atomic clusters is not 0, then according to the formula... Obtain the merge priority P; Merge the two atomic clusters with the highest merging priority in all cluster pairs into a new cluster; Merging is terminated when the merging priority of all cluster pairs is below the termination threshold or the global partitioning quality index tends to stabilize. For any new cluster formed after terminating the merging, the following judgment is performed: the reconstruction error of the new cluster based on the local generation model is calculated and used as a static anomaly score, including: For any new cluster formed after the termination of merging, train an independent local generative model based on principal component analysis; Project all high-dimensional state feature vectors within the new cluster onto a subspace composed of principal components whose cumulative contribution rate in the local generative model reaches a preset contribution rate threshold, and then reconstruct them back to the original dimensions. Calculate the average Euclidean distance between the original vector and the reconstructed vector, and use the average Euclidean distance as the static outlier score of the new cluster; The motion trajectory of the centroid of the new cluster in the feature space is tracked during continuous monitoring, and its curvature is calculated. When the static anomaly score and curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

2. The method for monitoring and early warning of submarine optical cables based on data analysis according to claim 1, characterized in that, The process of constructing the state feature vector includes: Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients. The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols. Computation space second difference value ,in, , , These are the average signal values ​​of the current sensing unit and the two sensing units geographically adjacent to the current sensing unit; The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

3. The method for monitoring and early warning of submarine optical cables based on data analysis according to claim 1, characterized in that, The calculation process for the fusion distance between any two atomic clusters includes: The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ; The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ; Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ; The formula for calculating the fusion distance D is: .

4. The method for monitoring and early warning of submarine optical cables based on data analysis according to claim 1, characterized in that, The method for determining whether the global partitioning quality index tends to stabilize is as follows: After each merge, calculate the average silhouette coefficient value of all current clusters; If, after three consecutive merging operations, the growth rate of the average contour coefficient is lower than the preset growth rate threshold, then the global segmentation quality index is determined to be stable.

5. The method for monitoring and early warning of submarine optical cables based on data analysis according to claim 4, characterized in that, The process of tracking the motion trajectory of the centroid of the new cluster in the feature space at continuous monitoring times and calculating its curvature includes: Record the centroid position vectors of any new cluster at three consecutive monitoring times as follows: , , ; Using the coordinates of three high-dimensional points, the trajectory is calculated. Curvature of time The calculation formula is: in, Represents the dot product of vectors. Denotes the Euclidean norm of a vector; if , and If at least one of them is 0, then the curvature is directly set to 0. Set to 0.

6. A system for implementing the data analysis-based submarine optical cable monitoring and early warning method according to any one of claims 1 to 5, characterized in that, Includes the following modules: An initialization module is used to acquire the time-series monitoring signal of each sensing unit distributed along the coastal bottom optical cable; for each sensing unit, the historical signal of the sensing unit and the synchronous monitoring signal of the geographically adjacent sensing units are extracted from the time-series monitoring signal, and a state feature vector containing time-domain projection coefficients, sequence symbolic representation and spatial second-order difference information is constructed; each state feature vector is initialized as an atom cluster. The merging module is used to determine the weights of any two atomic clusters based on the differences in the dispersion of their intra-cluster data in the temporal and spatial dimensions, and to calculate the weighted fusion distance by combining the point set distribution divergence and sequence morphological similarity. Based on the fusion distance and the topological proximity factor representing the geographical continuity of the sensing unit, the merging priority of the two atomic clusters is determined. The two atomic clusters with the highest merging priority among all cluster pairs are merged into a new cluster. The early warning module terminates merging when the merging priority of all cluster pairs is lower than the termination threshold or the global partitioning quality index tends to stabilize. For any new cluster formed after the termination of merging, the following judgment is performed: the reconstruction error of the new cluster based on the local generation model is calculated and used as the static anomaly score; the motion trajectory of the centroid of the new cluster in the feature space is tracked at continuous monitoring time and its curvature is calculated; when the static anomaly score and the curvature exceed the preset first threshold and second threshold respectively, the new cluster is determined to be an abnormal event and an early warning is generated.

7. The system for implementing a data analysis-based method for monitoring and early warning of submarine optical cables according to claim 6, characterized in that, The process of constructing the state feature vector includes: Principal component analysis is used to project the multi-channel signal of each sensing unit over 12 consecutive time points onto the first 3 principal components to obtain 3D time-domain projection coefficients. The time series composed of the average signal values ​​of the sensing unit in each channel is reduced in dimension by using the segmented aggregation averaging method. Using the five preset Gaussian distribution quantiles as breakpoints, the dimension-reduced time series is symbolized into a symbol sequence composed of six different symbols. Computation space second difference value ,in, , , These are the signal averages of the current sensing unit and the two adjacent sensing units in terms of geographical location. The state feature vector is obtained by concatenating the 3D time-domain projection coefficients, the integer encoding of the symbol sequence, and the second-order spatial difference value.

8. The system for implementing a data analysis-based method for monitoring and early warning of submarine optical cables according to claim 6, characterized in that, The calculation process for the fusion distance between any two atomic clusters includes: The maximum mean difference between all high-dimensional state eigenvector point sets in any two atomic clusters is calculated using the radial basis kernel function to obtain the point set distribution divergence. ; The average dynamic time-warped distance of the sequence symbolic representations of the feature vectors in these two atomic clusters is calculated to obtain the sequence morphological similarity. ; Calculate the variance of the data for the two atomic clusters in the time dimension. variance in spatial dimensional features And according to the formula Calculate the weighting coefficients ,like and If all are 0, then it is Assign a default value; where, ; The formula for calculating the fusion distance D is: .

Images