A distribution communication network fault location method and system based on deep learning

Abstract

The present invention discloses a method and system for locating faults in a power distribution communication network based on deep learning. The method includes fitting acquired time series data of optical fiber links to obtain an actual power attenuation curve, calculating the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve to obtain standardized difference data; inputting the standardized difference data into a nonlinear attenuation model to obtain nonlinear change characteristic data; inputting the distribution network topology characteristic data and the nonlinear change characteristic data into a fault identification model to obtain a failure probability value for each optical fiber link; comparing the failure probability value with a preset failure probability threshold to obtain a potential fault link; performing feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data; processing the attenuation mutation node data and determining fault point data based on the obtained stress value data. The method of the present invention achieves rapid location of loss faults in optical fiber networks.

Description

Technical Field

[0001] The present invention relates to the technical field of fault analysis, and in particular to a method and system for locating faults in a power distribution communication network based on deep learning. Background Art

[0002] As a key infrastructure for smart grids, the power distribution communication network undertakes core tasks such as data transmission, equipment control, and status monitoring in the power system. Fiber optic links, due to their high bandwidth and interference resistance, have become the primary transmission medium for these networks. However, fiber optic links are susceptible to external forces, temperature fluctuations, and connector aging, leading to optical signal power degradation and even interruption, seriously threatening the safe and stable operation of the distribution network.

[0003] Existing fiber optic fault location methods rely on hardware devices such as optical time-domain reflectometers (OTDRs) to detect fault points by analyzing the attenuation curve of backscattered light signals. This method requires a single detection cycle of tens of seconds to several minutes. However, the speed at which faults evolve in distribution and communication networks far exceeds the frequency of data updates, making it difficult to locate faults in a timely manner. Summary of the Invention

[0004] The present invention provides a distribution communication network fault location method and system based on deep learning to solve the technical problem of how to improve the existing distribution communication network fault location method and achieve the effect of reducing the distribution communication network fault processing time.

[0005] To solve the above technical problems, an embodiment of the present invention provides a method for locating faults in a power distribution communication network based on deep learning, comprising:

[0006] Obtaining optical power data of optical fiber links in a target distribution network, calculating optical power differences between adjacent sampling points in the optical power data, and generating time series data;

[0007] Performing fitting processing on the time series data to obtain an actual power attenuation curve, and calculating the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve to obtain standardized difference data;

[0008] Inputting the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear variation characteristic data reflecting the attenuation of optical signal power in the optical fiber link;

[0009] Extracting features of the physical topology of the target distribution network to obtain distribution network topology feature data, inputting the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model to obtain a failure probability value of each optical fiber link;

[0010] Comparing the fault probability value with a preset fault probability threshold to obtain a potential fault link, performing feature extraction and mutation node identification on the potential fault link in sequence to obtain attenuated mutation node data;

[0011] The attenuation mutation node data is processed according to a stress inversion method, and the fault point data in the target distribution network is determined based on the obtained stress value data.

[0012] As one preferred solution, the step of obtaining optical power data of an optical fiber link in a target distribution network, calculating optical power differences between adjacent sampling points in the optical power data, and generating time series data includes:

[0013] Collecting raw optical power data of optical fiber links in the target distribution network in real time and preprocessing the raw optical power data;

[0014] Time-sorting the pre-processed raw optical power data to obtain an optical power sequence data set;

[0015] Calculating the optical power difference between adjacent sampling points in the optical power sequence data set to obtain difference data; wherein if the optical power difference exceeds a preset optical power fluctuation threshold, it is marked as an abnormal data point;

[0016] Non-abnormal data points in the difference data are screened to obtain time series data.

[0017] As one preferred solution, before fitting the time series data, the method further includes:

[0018] Performing linear regression processing on the time series data to obtain optical power trend data, and calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data;

[0019] If the mean square error exceeds a preset error threshold, fitting processing is performed on the time series data.

[0020] As one preferred solution, the calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data includes:

[0021] generating stress distribution data according to the obtained stress distribution parameters of the optical fiber link, and performing frequency domain feature extraction on the stress distribution data using Fourier transform to obtain the theoretical stress characteristic data;

[0022] Performing linear regression processing on the time series data using the least squares method to obtain an optical power trend function, and obtaining a trend function value at each sampling time point based on the optical power trend function to form the optical power trend data;

[0023] The theoretical stress characteristic data and the optical power trend data are time-aligned, and the mean square error of the time-aligned theoretical stress characteristic data and the optical power trend data is calculated.

[0024] As one preferred solution, the fitting process is performed on the time series data to obtain an actual power attenuation curve, and the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve is calculated to obtain standardized difference data, including:

[0025] The time series data is subjected to curve fitting based on the cubic spline interpolation method to obtain the actual power attenuation curve;

[0026] Constructing a theoretical power attenuation curve based on the physical parameters of the optical fiber link;

[0027] Calculating the absolute difference of each sampling point on the time axis of the actual power attenuation curve and the theoretical power attenuation curve to obtain an absolute difference sequence reflecting the degree of attenuation abnormality;

[0028] The absolute difference sequence is normalized to obtain the standardized difference data.

[0029] As one preferred solution, extracting features of the physical topology of the target distribution network to obtain distribution network topology feature data includes:

[0030] Based on the obtained physical connection relationship of the optical fiber link, constructing a node adjacency matrix and a link attribute table;

[0031] Summing each row of the node adjacency matrix to obtain a node degree vector reflecting the importance of the node;

[0032] Calculating link weight data according to the link attribute table, and constructing a link weight matrix based on the link weight data;

[0033] The node degree vector and the link weight matrix are concatenated to obtain the distribution network topology characteristic data.

[0034] As one preferred solution, the comparison of the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially performing feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data include:

[0035] Comparing the failure probability value with a preset failure probability threshold, screening out links whose failure probability values are greater than or equal to the preset failure probability threshold, and generating a list of potential failure links;

[0036] Acquire optical power sequence data of links corresponding to the faulty link list, and convert the optical power sequence data into a two-dimensional matrix;

[0037] The two-dimensional matrix is spliced with the standardized difference data and the nonlinear change feature data to obtain a fused feature tensor; and the fused feature tensor is input into a pre-built deep residual network for feature extraction to obtain attenuation mutation feature data;

[0038] Cluster analysis is performed on the attenuation mutation characteristic data to obtain the attenuation mutation node data where the attenuation trend has changed.

[0039] As one preferred solution, processing the attenuation mutation node data according to the stress inversion method and determining the fault point data in the target distribution network based on the obtained stress value data includes:

[0040] Constructing a stress attenuation relationship model based on the acquired stress-sensitive parameters, material stress threshold, and physical length of the optical fiber link;

[0041] Extracting the power attenuation and the physical length of the link corresponding to each attenuation mutation node data, inputting the power attenuation and the physical length into the stress attenuation relationship model to obtain a stress value data set;

[0042] Comparing each stress value data in the stress value data set with the material stress threshold, and screening the comparison results to obtain a list of high stress mutation nodes;

[0043] Fault point data in the target distribution network is determined based on the high stress mutation node list.

[0044] As one preferred solution, determining the fault point data in the target distribution network based on the high stress mutation node list includes:

[0045] Construct a candidate fault point set based on each node and its adjacent nodes in the high stress mutation node list;

[0046] Obtaining coordinate data of each node in the candidate fault point set, and calculating the relative distance between each high stress mutation node and its adjacent nodes based on the coordinate data;

[0047] A multi-condition screening is performed based on the acquired stress value of each high stress mutation node and the relative distance to the fault point, and the fault point data is obtained based on the result of the multi-condition screening.

[0048] Another embodiment of the present invention provides a power distribution communication network fault location system based on deep learning, including:

[0049] An acquisition module is used to obtain optical power data of optical fiber links in the target distribution network, calculate the optical power difference between adjacent sampling points in the optical power data, and generate time series data;

[0050] a calculation module, configured to perform fitting processing on the time series data to obtain an actual power attenuation curve, and calculate an absolute difference between the actual power attenuation curve and a theoretical power attenuation curve to obtain standardized difference data;

[0051] a first extraction module, configured to input the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear variation characteristic data reflecting the attenuation of optical signal power in the optical fiber link;

[0052] a second extraction module, configured to extract features of the physical topology of the target distribution network to obtain distribution network topology feature data, input the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model to obtain a failure probability value of each optical fiber link;

[0053] an identification module, configured to compare the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially perform feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data;

[0054] A positioning module is used to process the attenuation mutation node data according to a stress inversion method, and determine the fault point data in the target distribution network based on the obtained stress value data.

[0055] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0056] 1) This invention uses a deep learning model to accurately capture nonlinear fault characteristics and integrate multi-dimensional data, breaking through the traditional method's reliance on linear attenuation characteristics and significantly improving fault identification accuracy in complex scenarios. Specifically, this invention uses a nonlinear attenuation model to deeply process standardized difference data, effectively extracting the nonlinear variation characteristics of optical signal power attenuation. Combined with the distribution network topology feature data, this invention constructs a fault identification model that can accurately identify progressive faults such as microbends and joint aging that are easily missed by traditional methods, increasing the early fault detection rate by more than 40%. At the same time, it reduces the misjudgment rate in multi-branch networks through topological structure analysis.

[0057] 2) This invention establishes a closed-loop intelligent positioning process encompassing data acquisition, model inference, and stress inversion, enabling automated processing from fault probability assessment to precise positioning, significantly improving fault location efficiency and engineering practicality. Specifically, this invention uses stress inversion to convert attenuation mutation node data into specific stress values. This, combined with material stress thresholds, accurately locates the fault point, avoiding the time-consuming segment-by-segment detection inherent in traditional OTDRs. This reduces positioning time by over 60% in complex topologies such as ring networks, and can adapt to changes in network topology in real time, providing direct guidance for the rapid repair of distribution and communication networks, significantly reducing operational costs and troubleshooting time. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 1 is a flow chart of a method for locating a fault in a power distribution communication network based on deep learning in one embodiment of the present invention;

[0059] Figure 2 is a schematic diagram of a power distribution communication network fault location system based on deep learning in one embodiment of the present invention;

[0060] Reference numerals:

[0061] Among them, 11, acquisition module; 12, calculation module; 13, first extraction module; 14, second extraction module; 15, recognition module; 16, positioning module. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0063] In the description of the present invention, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0064] In the description of the present invention, it should be noted that, unless otherwise expressly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the two components. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are for illustrative purposes only, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0065] In describing the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0066] An embodiment of the present invention provides a method for locating faults in a power distribution communication network based on deep learning. Figure 1 , Figure 1 The figure shows a flow chart of a method for locating a fault in a power distribution communication network based on deep learning in one embodiment of the present invention, which includes steps S1 to S6:

[0067] S1: Obtain optical power data of optical fiber links in the target distribution network, calculate the optical power differences between adjacent sampling points in the optical power data, and generate time series data;

[0068] Preferably, in one embodiment of the present invention, obtaining optical power data of an optical fiber link in a target distribution network, calculating optical power differences between adjacent sampling points in the optical power data, and generating time series data includes:

[0069] Collecting raw optical power data of optical fiber links in the target distribution network in real time and preprocessing the raw optical power data;

[0070] Time-sorting the pre-processed raw optical power data to obtain an optical power sequence data set;

[0071] Calculating the optical power difference between adjacent sampling points in the optical power sequence data set to obtain difference data; wherein if the optical power difference exceeds a preset optical power fluctuation threshold, it is marked as an abnormal data point;

[0072] Non-abnormal data points in the difference data are screened to obtain time series data.

[0073] Optical power data refers to the energy intensity monitoring data of optical signals in optical fiber links, typically measured in decibel milliwatts (dBm). It reflects signal transmission quality and is a core indicator for determining the attenuation status of optical fibers. Time series data refers to a one-dimensional data sequence arranged in chronological order. In this embodiment, it is an ordered set of optical power differences between adjacent sampling points, used to characterize the dynamic changes in optical power over time. Preprocessing involves filtering the raw optical power data for noise and removing outliers to ensure data reliability. Common methods include filtering algorithms (such as median filtering and SG filtering) and outlier detection. The preset optical power fluctuation threshold is a critical value set based on the power fluctuation range during normal operation of the optical fiber link. It is used to identify abnormal data points that significantly deviate from the normal state (such as sudden power changes caused by sudden interference or equipment failure).

[0074] It's important to note that raw optical power data can be affected by environmental noise (such as electromagnetic interference and temperature fluctuations) and equipment errors. Direct use can lead to misjudgments in subsequent model analysis, requiring preprocessing to remove invalid data. Furthermore, the time series of optical power differences can reflect the dynamics of signal attenuation and is fundamental for identifying nonlinear attenuation (such as sudden changes and periodic fluctuations). However, anomalous data points can interfere with feature extraction, requiring threshold screening to ensure sequence validity.

[0075] Specifically, in this embodiment, optical power data of the optical fiber link is collected in real time by using devices such as an optical time domain reflectometer (OTDR) and an ultra-weak fiber grating array (uwFBG), and the sampling timestamp (with microsecond accuracy) and spatial position information are recorded synchronously.

[0076] Median filtering (3×3 window) is used to smooth the raw data, suppressing high-frequency noise and improving the signal-to-noise ratio to over 20dB. The mean and standard deviation of the data are calculated based on the 3σ principle, and points with absolute values exceeding ±3σ of the mean are marked as abnormal. This is combined with a dual screening method using physical thresholds (such as sudden changes in optical power exceeding ±0.5dB) to ensure an anomalous data rejection rate of over 99%.

[0077] The preprocessed data is sorted in ascending order by sampling timestamp to generate an optical power sequence data set. The power values at adjacent sampling points are subtracted to obtain differential data reflecting instantaneous power changes. A preset fluctuation threshold (which can be dynamically adjusted based on historical data) is set to remove anomalous data points whose differences exceed the threshold, retaining continuous and stable non-anomalous data points, ultimately generating time series data for subsequent analysis.

[0078] S2: performing fitting processing on the time series data to obtain an actual power attenuation curve, and calculating the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve to obtain standardized difference data;

[0079] Preferably, in one embodiment of the present invention, before performing fitting processing on the time series data, the method further includes:

[0080] Performing linear regression processing on the time series data to obtain optical power trend data, and calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data;

[0081] If the mean square error exceeds a preset error threshold, fitting processing is performed on the time series data.

[0082] Preferably, in one embodiment of the present invention, the calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data includes:

[0083] generating stress distribution data according to the obtained stress distribution parameters of the optical fiber link, and performing frequency domain feature extraction on the stress distribution data using Fourier transform to obtain the theoretical stress characteristic data;

[0084] Performing linear regression processing on the time series data using the least squares method to obtain an optical power trend function, and obtaining a trend function value at each sampling time point based on the optical power trend function to form the optical power trend data;

[0085] The theoretical stress characteristic data and the optical power trend data are time-aligned, and the mean square error of the time-aligned theoretical stress characteristic data and the optical power trend data is calculated.

[0086] It should be noted that optical power variations in fiber links can be influenced by a variety of factors, including regular trends caused by normal stress (such as slow changes in ambient temperature) and abnormal, sudden changes caused by faults. Linear regression is used to first extract the overall optical power trend (such as long-term attenuation or slight fluctuations). This is then compared with the expected trend under theoretical stress (theoretical stress characteristic data) to determine whether the trend in the current data conforms to normal physical laws. If the deviation (mean square error) between the two exceeds a preset threshold, it indicates that the actual trend may contain abnormal variations due to non-stress factors (such as poor contact and fiber microbends). Further fitting processing is required to explore more complex nonlinear features to prevent normal stress fluctuations from interfering with fault detection.

[0087] Specifically, in this embodiment, stress distribution data is generated based on the laying environment parameters of the optical fiber link (such as the preset temperature variation range, the allowable tension value, etc.), and the frequency characteristics of these stress data are then analyzed by Fourier transform (such as whether there are regular fluctuations caused by periodic temperature changes) to obtain the frequency domain characteristics that the optical power change should theoretically have under the influence of stress (i.e., the theoretical stress characteristic data).

[0088] The time series data is processed using the least squares method to find a straight line or low-order curve (trend function) that best fits the data. The output value of this function (trend function value) is the optical power trend data, which represents the long-term change trend of the data after removing short-term fluctuations.

[0089] The theoretical stress characteristic data and optical power trend data are aligned at each time point (time alignment). The difference between the two at each time point is calculated, and the squares of all differences are averaged (mean square error). If this error exceeds a preset threshold, it indicates that the actual trend deviates significantly from the normal trend under the influence of theoretical stress. More complex fitting processing (such as polynomial fitting or nonlinear fitting) is required on the time series data to capture possible fault-related abnormal characteristics.

[0090] Preferably, in one embodiment of the present invention, the fitting process is performed on the time series data to obtain an actual power attenuation curve, and the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve is calculated to obtain standardized difference data, including:

[0091] The time series data is subjected to curve fitting based on the cubic spline interpolation method to obtain the actual power attenuation curve;

[0092] Constructing a theoretical power attenuation curve based on the physical parameters of the optical fiber link;

[0093] Calculating the absolute difference of each sampling point on the time axis of the actual power attenuation curve and the theoretical power attenuation curve to obtain an absolute difference sequence reflecting the degree of attenuation abnormality;

[0094] The absolute difference sequence is normalized to obtain the standardized difference data.

[0095] Cubic spline interpolation is a mathematical method used to construct a smooth curve based on known discrete data points (here, the optical power differences in time series data). Cubic spline interpolation uses a cubic polynomial to fit between every two adjacent data points, ensuring that the curve has continuous first- and second-order derivatives throughout the entire interval. The resulting curve is both smooth and closely approximates the original data.

[0096] The actual power attenuation curve is obtained by curve fitting time series data, reflecting the actual attenuation of optical power in a fiber link over time. The theoretical power attenuation curve is constructed based on the physical parameters of the fiber link (such as fiber material, length, and degree of curvature), according to physical principles and theoretical models. It represents the expected attenuation of optical power over time under ideal, fault-free conditions.

[0097] It should be noted that in actual fiber optic link operation, optical power attenuation is affected by a variety of factors, such as ambient temperature, external compression, and loose connectors, resulting in differences between actual attenuation and theoretical attenuation. These differences can be identified by fitting the actual power attenuation curve and comparing it with the theoretical power attenuation curve.

[0098] While absolute difference series can reflect the degree of attenuation anomalies, their numerical ranges can vary across different fiber links or time periods, making them difficult to analyze and compare directly. Normalized difference data, obtained through normalization, can measure attenuation anomalies on a unified scale, providing a more reliable basis for determining the presence and severity of fiber link faults.

[0099] Specifically, in this embodiment, each data point in the time series data is treated as a discrete point on the two-dimensional plane of time-optical power difference. Then, using cubic spline interpolation, a cubic polynomial is determined between every two adjacent data points. The coefficients of this cubic polynomial are calculated by satisfying certain conditions (such as the continuity of the first and second derivatives of the curve at the junction of adjacent intervals). Finally, the cubic polynomials for all adjacent intervals are combined to form a smooth curve, namely the actual power attenuation curve. The actual power attenuation curve can effectively fit the time series data and demonstrate the actual trend of optical power attenuation over time.

[0100] Collect the physical parameters of the optical fiber link, including fiber type (single-mode fiber, multimode fiber), length, loss coefficient (related to the fiber material), bend radius, and so on. Based on the physical principles and theoretical models of optical fiber transmission (e.g., fiber attenuation is proportional to length), establish a functional relationship between time and optical power attenuation. Using this functional relationship, calculate the theoretical optical power attenuation values at different time points. These theoretical attenuation values are then connected to form a theoretical power attenuation curve. The theoretical power attenuation curve represents the optical power attenuation over time under ideal conditions with no external interference or faults.

[0101] Based on the time axis, the optical power attenuation values corresponding to each sampling point in the actual power attenuation curve are subtracted from the theoretical power attenuation curve. The absolute value of the subtraction result is taken to form an absolute difference sequence. Each value in the absolute difference sequence reflects the difference between the actual attenuation and the theoretical attenuation at that sampling point. The larger the difference, the more significant the attenuation anomaly at that point.

[0102] Normalize each value in the difference sequence. Combine all normalized values into standardized difference data. This data more intuitively reflects the degree of attenuation anomaly at each sampling point, facilitating subsequent analysis and processing.

[0103] S3: inputting the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear change characteristic data reflecting the attenuation of optical signal power in the optical fiber link;

[0104] Standardized difference data is the optical power attenuation anomaly data that has been preprocessed and normalized to eliminate the impact of dimension and link differences. The value range is [0, 1] or the interval with a mean of 0 and a standard deviation of 1. It is the core input data that reflects the degree to which actual attenuation deviates from the theoretical model.

[0105] The nonlinear attenuation model is a neural network model based on deep learning. It can capture the nonlinear and non-stationary change characteristics of optical power attenuation (such as mutations, exponential decay, and periodic fluctuations). Different from the traditional linear model (such as the least squares method) that fits linear trends, it is suitable for feature extraction in complex fault scenarios.

[0106] The purpose of this step is to automatically extract the fault features hidden in the standardized difference data, avoid the one-sidedness of manually designed features, and convert the abstract numerical sequence into interpretable physical features, providing fault characterization information for subsequent topological structure analysis to improve positioning accuracy.

[0107] Specifically, the standardized difference data is converted into a format suitable for model input; then a nonlinear attenuation model is constructed. Taking the CNN-LSTM combination model as an example, the CNN layer is first used to extract local features, and then the LSTM layer is used to capture long-term dependent features. Finally, the fully connected layer is used to map the features into a nonlinear change feature vector containing information such as mutation position probability, attenuation mode label, and feature confidence; during model training, historical fault data sets are used and fault simulation data is artificially injected to enhance generalization ability, and a multi-task loss function combining cross entropy loss and mean square error is used for optimization; during inference, real-time standardized difference data is input into the trained model, and nonlinear change feature data such as "a certain type of mutation is detected at a certain position with a certain degree of confidence" is output.

[0108] S4: extracting features of the physical topology of the target distribution network to obtain distribution network topology feature data, inputting the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model to obtain a fault probability value of each optical fiber link;

[0109] Preferably, in one embodiment of the present invention, extracting features of the physical topology of the target distribution network to obtain distribution network topology feature data includes:

[0110] Based on the obtained physical connection relationship of the optical fiber link, constructing a node adjacency matrix and a link attribute table;

[0111] Summing each row of the node adjacency matrix to obtain a node degree vector reflecting the importance of the node;

[0112] Calculating link weight data according to the link attribute table, and constructing a link weight matrix based on the link weight data;

[0113] The node degree vector and the link weight matrix are concatenated to obtain the distribution network topology characteristic data.

[0114] The physical topology of the target distribution network refers to the actual connection method and layout between various electrical devices (such as transformers, switches, lines, etc.) in the target distribution network, which describes the path and structure of electric energy transmission in the network.

[0115] The node adjacency matrix is a two-dimensional matrix that represents the connectivity between nodes in a distribution network. The rows and columns of the matrix correspond to nodes in the distribution network. If there is a direct connection between two nodes, the corresponding position in the matrix is 1; otherwise, it is 0. The link attribute table records the relevant attribute information of each link (e.g., transmission line) in the distribution network, such as link length, capacity, and resistance.

[0116] The node degree vector is calculated by summing each row of the node adjacency matrix. Each element in the vector represents the number of connections between the corresponding node and other nodes, reflecting the node's importance in the distribution network. The greater the number of connections, the more important the node. Link weight data is calculated based on the information in the link attribute table and is used to measure the importance or transmission capacity of each link in the distribution network. The link weight matrix is constructed from the link weight data and further describes the relative importance of links in the distribution network.

[0117] It's important to note that relying solely on the nonlinear variation characteristics of optical signal power attenuation to determine whether a fiber link is faulty is insufficient. The physical topology of the distribution network significantly influences the occurrence and propagation of faults. Nodes and links at different locations have varying importance within the network, and the impact of a fault on the entire distribution network varies accordingly. By extracting distribution network topology characteristics and combining them with nonlinear variation characteristics to feed into a fault identification model, we can comprehensively consider both topology and optical signal attenuation to more accurately determine the failure probability of each fiber link, improving the accuracy and reliability of fault identification.

[0118] Specifically, in this embodiment, first, the physical connection relationship of the optical fiber link in the target distribution network is obtained, which can be obtained by consulting the design drawings, equipment records and other information of the distribution network. Then, a node adjacency matrix is constructed based on these connection relationships. Assuming there are n nodes, the node adjacency matrix is an n×n matrix. If there is a direct connection between node i and node j, the element value of the i-th row and j-th column and the j-th row and i-th column in the matrix is 1; if there is no connection, it is 0. At the same time, the relevant attribute information of the optical fiber link, such as the length, transmission capacity, resistance, etc. of the link, is sorted out, and this information is recorded in the link attribute table.

[0119] Perform a sum operation on each row of the node adjacency matrix. For example, for row i, add all elements in that row. The resulting sum is the degree of node i. Arrange the degrees of all nodes in order to form a vector, the node degree vector. This vector intuitively reflects the importance of each node in the distribution network. A larger degree indicates that the node is connected to more nodes and is more critical in the network.

[0120] Based on the information in the link attribute table, an appropriate method is used to calculate the weight of each link. For example, the weight can be calculated based on factors such as link length and transmission capacity. Links with shorter lengths and greater transmission capacity are likely to have higher weights. The calculated link weights are constructed into a matrix based on the connection relationships between nodes, namely the link weight matrix. The element in the i-th row and j-th column of the matrix represents the weight of the link between nodes i and j.

[0121] Concatenate the node degree vector and the link weight matrix. You can add the node degree vector as a new column to the link weight matrix, or use other appropriate concatenation methods. The resulting concatenated data set is the distribution network topology feature data, which integrates information such as the importance of nodes and the relative importance of links.

[0122] S5: Comparing the fault probability value with a preset fault probability threshold to obtain a potential fault link, performing feature extraction and mutation node identification on the potential fault link in sequence to obtain attenuated mutation node data;

[0123] Preferably, in one embodiment of the present invention, the comparing the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially performing feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data includes:

[0124] Comparing the failure probability value with a preset failure probability threshold, screening out links whose failure probability values are greater than or equal to the preset failure probability threshold, and generating a list of potential failure links;

[0125] Acquire optical power sequence data of links corresponding to the faulty link list, and convert the optical power sequence data into a two-dimensional matrix;

[0126] The two-dimensional matrix is spliced with the standardized difference data and the nonlinear change feature data to obtain a fused feature tensor; and the fused feature tensor is input into a pre-built deep residual network for feature extraction to obtain attenuation mutation feature data;

[0127] Cluster analysis is performed on the attenuation mutation characteristic data to obtain the attenuation mutation node data where the attenuation trend has changed.

[0128] The failure probability value is a numerical value (between 0 and 1) calculated by the fault identification model to indicate the likelihood of failure for each fiber link. A higher value indicates a greater likelihood of failure. The preset failure probability threshold is a manually set critical value (e.g., 0.7) to determine whether a link is likely to fail. This value is used to screen high-risk links and avoid indiscriminate analysis of all links.

[0129] Potentially faulty links are fiber links whose failure probability exceeds a preset threshold and are the focus of subsequent investigation. The Deep Residual Network (ResNet) is a deep learning model that uses residual connections to address the vanishing gradient problem in deep networks. It excels at extracting multi-level abstract features from data (such as short-term mutations and long-term trends). Attenuation mutation node data includes information such as node number, degree of attenuation anomaly, and spatial location, and is used to identify specific nodes in the fiber link experiencing significant attenuation changes.

[0130] It's important to note that this step quickly filters out low-probability links through threshold comparison, concentrating computing resources on potentially faulty links with a high probability of failure. This avoids the "needle-in-a-haystack" approach of troubleshooting and improves efficiency. It also narrows the scope of the fault from "potential links" to "specific nodes," providing direct clues for subsequent precise fault location, forming a progressive troubleshooting logic of "probability screening → feature location → node identification."

[0131] Specifically, in this embodiment, the failure probability values for each link output by the fault identification model (e.g., [0.65, 0.82, 0.51, ...]) are compared with a preset threshold (e.g., 0.7). Links with probabilities ≥ the threshold (e.g., 0.82 ≥ 0.7 for the second link) are retained, generating a list of potential faulty links (e.g., ["Link_03", "Link_17"]). This list includes basic information such as the link number, region, and historical fault records, allowing for rapid identification of the link being investigated. For example, prioritizing "Link_03," which has a history of frequent faults.

[0132] The optical power data (including timestamps and power values) for each potential faulty link is arranged along the "time-power" dimension to form a two-dimensional matrix (rows = number of sampling points, columns = 1). For example, 1000 sampling points form a 1000×1 matrix, visually displaying the power variation trend over time. The two-dimensional matrix is then concatenated with the standardized difference data (reflecting the degree of anomaly) and nonlinear variation features (such as mutation location probability and failure mode labels) along the channel dimension to form a fused feature tensor (e.g., 1000×3 dimensions), enabling the model to simultaneously analyze the original signal, the degree of anomaly, and the failure mode.

[0133] The deep residual network of the present invention adopts 34-layer ResNet, which contains multiple residual blocks. Each residual block extracts multi-layer features through "convolution-batch normalization-ReLU-residual connection":

[0134] Shallow layer: captures short-term power fluctuations (such as sudden changes within 1 second);

[0135] Deep layer: learn long-term decay trends (such as decay acceleration that lasts for 1 hour).

[0136] Input and output: Input the fused feature tensor and output the attenuation mutation feature vector (including information such as "mutation occurrence time point" and "abnormal energy ratio"). For example, it can be identified that "the attenuation mutation energy ratio reaches 85% at the 500th time point (corresponding to 1500 meters of the link)".

[0137] For the feature vectors output by the residual network, cosine similarity is used to calculate the similarity between nodes, forming a similarity matrix (the closer the value is to 1, the more similar the node decay trends are). This similarity matrix is converted into a graph structure, and the main eigenvectors are extracted through eigendecomposition. K-means clustering is then used to classify the nodes into categories such as "normal decay," "gradual change," and "mutation." Nodes in the "mutation" category are selected to generate decay mutation node data (for example, containing the node ID "Node_24," the mutation occurrence distance "1520 meters," and the abnormality level "high").

[0138] S6: Processing the attenuation mutation node data according to a stress inversion method, and determining the fault point data in the target distribution network based on the obtained stress value data.

[0139] Preferably, in one embodiment of the present invention, processing the attenuation mutation node data according to the stress inversion method and determining the fault point data in the target distribution network based on the obtained stress value data includes:

[0140] Constructing a stress attenuation relationship model based on the acquired stress-sensitive parameters, material stress threshold, and physical length of the optical fiber link;

[0141] Extracting the power attenuation and the physical length of the link corresponding to each attenuation mutation node data, inputting the power attenuation and the physical length into the stress attenuation relationship model to obtain a stress value data set;

[0142] Comparing each stress value data in the stress value data set with the material stress threshold, and screening the comparison results to obtain a list of high stress mutation nodes;

[0143] Fault point data in the target distribution network is determined based on the high stress mutation node list.

[0144] Preferably, in one embodiment of the present invention, determining the fault point data in the target distribution network based on the high stress mutation node list includes:

[0145] Construct a candidate fault point set based on each node and its adjacent nodes in the high stress mutation node list;

[0146] Obtaining coordinate data of each node in the candidate fault point set, and calculating the relative distance between each high stress mutation node and its adjacent nodes based on the coordinate data;

[0147] A multi-condition screening is performed based on the acquired stress value of each high stress mutation node and the relative distance to the fault point, and the fault point data is obtained based on the result of the multi-condition screening.

[0148] The stress inversion method uses information such as the power attenuation of the optical signal in the optical fiber link to reversely deduce the stress on the optical fiber. Because there is a certain relationship between the optical signal attenuation and the stress on the optical fiber, this relationship can be used to infer the stress.

[0149] The stress sensitivity parameter reflects the optical fiber's sensitivity to applied stress. Optical fibers of different materials and structures have different stress sensitivity parameters, and they serve as a crucial basis for constructing stress attenuation relationship models. The material stress threshold is the maximum stress the optical fiber can withstand. When the stress on the fiber exceeds this threshold, the fiber may experience failures such as breakage or bending. The stress attenuation relationship model describes the mathematical relationship between optical fiber stress and optical signal power attenuation. This model allows the corresponding stress value to be calculated based on the power attenuation.

[0150] Power attenuation, the reduction in optical signal power during transmission in an optical fiber link, is a key input for calculating stress. The high stress mutation node list consists of nodes whose stress exceeds the material stress threshold. These nodes are key targets for potential faults. The relative distance to the fault point is the distance between a high stress mutation node and its adjacent nodes, which is used to further determine the location of the fault.

[0151] It's important to note that the data obtained in the previous steps regarding nodes with attenuation mutations only indicates possible abnormal nodes and cannot precisely determine the fault location. However, the stress experienced by the optical fiber is closely related to the occurrence of faults. Using stress inversion methods, power attenuation information can be converted into stress information, allowing for more accurate fault location. By comparing stress values with material stress thresholds, nodes with high stress mutations can be identified, narrowing the scope of troubleshooting and focusing on the nodes most likely to have failed. Multi-criteria screening, combined with node coordinate data and relative distance to the fault point, can further improve the accuracy of fault location.

[0152] Specifically, in this embodiment, information such as the stress-sensitive parameters of the optical fiber link, the material stress threshold, and the physical length of the link is first obtained. This information can be obtained by consulting the optical fiber product manual, conducting experimental measurements, or other methods. A stress-attenuation relationship model is then constructed based on this information. This model can be based on physical principles, for example, by deriving the functional relationship between stress and power attenuation based on the mechanical properties of the optical fiber material and the principles of optical transmission. A common example is a linear or nonlinear functional model that describes the corresponding relationship between stress value and power attenuation.

[0153] The power attenuation corresponding to each node is extracted from the attenuation mutation node data, and the physical length of the link to which the node belongs is also obtained. The extracted power attenuation and physical length are substituted into the stress attenuation relationship model, and the model calculations are used to obtain the stress value corresponding to each node. The stress values of all nodes are collected to form a stress value dataset.

[0154] Compare each stress value in the stress data set to the material stress threshold. If a node's stress value exceeds the material stress threshold, it indicates excessive stress at that node and a possible fault. Filter out all nodes with stress values exceeding the material stress threshold and compile their relevant information into a list, known as the High Stress Sudden Change Node List. The nodes in this list are the ones that require special attention.

[0155] A set of candidate fault points is constructed based on each node in the high stress mutation node list and its adjacent nodes. Adjacent nodes can be determined based on the fiber link topology and are typically nodes directly connected to the high stress mutation node. This further narrows the scope of troubleshooting to these candidate nodes.

[0156] Obtain the coordinate data for each node in the candidate fault point set. This coordinate data can be obtained through a geographic information system (GIS) or other means. Then, based on this coordinate data, calculate the relative distance between each high stress mutation node and its adjacent nodes. This distance information can help further determine the possible location of the fault point.

[0157] Multi-condition screening is performed based on the stress values and relative distances to the fault point for each node with a high stress mutation. For example, nodes with high stress values and close proximity to adjacent nodes can be prioritized as the likely fault point. This multi-condition screening approach ultimately identifies the fault point data, including its specific location, associated stress values, and other information, providing an accurate basis for subsequent fault repair.

[0158] Through the above steps, the stress inversion method combined with multi-condition screening can accurately determine the fault point data in the target distribution network from the attenuation mutation node data, thereby improving the efficiency of fault detection and repair.

[0159] An embodiment of the present invention provides a distribution communication network fault location system based on deep learning. For details, see Figure 2 , Figure 2 The figure shows a flow chart of a power distribution communication network fault location system based on deep learning in one embodiment of the present invention, which includes:

[0160] An acquisition module 11 is configured to acquire optical power data of an optical fiber link in a target distribution network, calculate optical power differences between adjacent sampling points in the optical power data, and generate time series data;

[0161] a calculation module 12, configured to perform fitting processing on the time series data to obtain an actual power attenuation curve, and calculate an absolute difference between the actual power attenuation curve and a theoretical power attenuation curve to obtain standardized difference data;

[0162] A first extraction module 13 is configured to input the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear variation characteristic data reflecting the attenuation of optical signal power in the optical fiber link;

[0163] A second extraction module 14 is configured to extract features of the physical topology of the target distribution network to obtain distribution network topology feature data, input the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model, and obtain a failure probability value of each optical fiber link;

[0164] An identification module 15 is configured to compare the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially perform feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data;

[0165] The positioning module 16 is configured to process the attenuation mutation node data according to a stress inversion method, and determine the fault point data in the target distribution network based on the obtained stress value data.

[0166] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0167] 1) This invention uses a deep learning model to accurately capture nonlinear fault characteristics and integrate multi-dimensional data, breaking through the traditional method's reliance on linear attenuation characteristics and significantly improving fault identification accuracy in complex scenarios. Specifically, this invention uses a nonlinear attenuation model to deeply process standardized difference data, effectively extracting the nonlinear variation characteristics of optical signal power attenuation. Combined with the distribution network topology feature data, this invention constructs a fault identification model that can accurately identify progressive faults such as microbends and joint aging that are easily missed by traditional methods, increasing the early fault detection rate by more than 40%. At the same time, it reduces the misjudgment rate in multi-branch networks through topological structure analysis.

[0168] 2) This invention establishes a closed-loop intelligent positioning process encompassing data acquisition, model inference, and stress inversion, enabling automated processing from fault probability assessment to precise positioning, significantly improving fault location efficiency and engineering practicality. Specifically, this invention uses stress inversion to convert attenuation mutation node data into specific stress values. This, combined with material stress thresholds, accurately locates the fault point, avoiding the time-consuming segment-by-segment detection inherent in traditional OTDRs. This reduces positioning time by over 60% in complex topologies such as ring networks, and can adapt to changes in network topology in real time, providing direct guidance for the rapid repair of distribution and communication networks, significantly reducing operational costs and troubleshooting time.

[0169] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for locating faults in a power distribution communication network based on deep learning, characterized in that: include: Obtaining optical power data of optical fiber links in a target distribution network, calculating optical power differences between adjacent sampling points in the optical power data, and generating time series data; Performing fitting processing on the time series data to obtain an actual power attenuation curve, and calculating the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve to obtain standardized difference data; Inputting the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear variation characteristic data reflecting the attenuation of optical signal power in the optical fiber link; Extracting features of the physical topology of the target distribution network to obtain distribution network topology feature data, inputting the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model to obtain a failure probability value of each optical fiber link; Comparing the fault probability value with a preset fault probability threshold to obtain a potential fault link, performing feature extraction and mutation node identification on the potential fault link in sequence to obtain attenuated mutation node data; Processing the attenuation mutation node data according to a stress inversion method, and determining the fault point data in the target distribution network based on the obtained stress value data; The processing of the attenuation mutation node data according to the stress inversion method and determining the fault point data in the target distribution network based on the obtained stress value data includes: Constructing a stress attenuation relationship model based on the acquired stress-sensitive parameters, material stress threshold, and physical length of the optical fiber link; Extracting the power attenuation and the physical length of the link corresponding to each attenuation mutation node data, inputting the power attenuation and the physical length into the stress attenuation relationship model to obtain a stress value data set; Comparing each stress value data in the stress value data set with the material stress threshold, and screening the comparison results to obtain a list of high stress mutation nodes; Determining fault point data in the target distribution network based on the high stress mutation node list includes: Construct a candidate fault point set based on each node and its adjacent nodes in the high stress mutation node list; Obtaining coordinate data of each node in the candidate fault point set, and calculating the relative distance between each high stress mutation node and its adjacent nodes based on the coordinate data; A multi-condition screening is performed based on the acquired stress value of each high stress mutation node and the relative distance to the fault point, and the fault point data is obtained based on the result of the multi-condition screening.

2. The method for locating faults in a power distribution communication network based on deep learning according to claim 1, wherein: The step of obtaining optical power data of an optical fiber link in a target distribution network, calculating optical power differences between adjacent sampling points in the optical power data, and generating time series data includes: Collecting raw optical power data of optical fiber links in the target distribution network in real time and preprocessing the raw optical power data; Time-sorting the pre-processed raw optical power data to obtain an optical power sequence data set; Calculating the optical power difference between adjacent sampling points in the optical power sequence data set to obtain difference data; wherein if the optical power difference exceeds a preset optical power fluctuation threshold, it is marked as an abnormal data point; Non-abnormal data points in the difference data are screened to obtain time series data.

3. The method for locating faults in a power distribution communication network based on deep learning according to claim 1, wherein: Before fitting the time series data, the method further includes: Performing linear regression processing on the time series data to obtain optical power trend data, and calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data; If the mean square error exceeds a preset error threshold, fitting processing is performed on the time series data.

4. The method for locating faults in a power distribution communication network based on deep learning according to claim 3, wherein: The calculating the mean square error between the theoretical stress characteristic data of the optical fiber link and the optical power trend data includes: generating stress distribution data according to the obtained stress distribution parameters of the optical fiber link, and performing frequency domain feature extraction on the stress distribution data using Fourier transform to obtain the theoretical stress characteristic data; Performing linear regression processing on the time series data using the least squares method to obtain an optical power trend function, and obtaining a trend function value at each sampling time point based on the optical power trend function to form the optical power trend data; The theoretical stress characteristic data and the optical power trend data are time-aligned, and the mean square error of the time-aligned theoretical stress characteristic data and the optical power trend data is calculated.

5. The method for locating faults in a power distribution communication network based on deep learning according to claim 1, wherein: The fitting process is performed on the time series data to obtain an actual power attenuation curve, and the absolute difference between the actual power attenuation curve and the theoretical power attenuation curve is calculated to obtain standardized difference data, including: The time series data is subjected to curve fitting based on the cubic spline interpolation method to obtain the actual power attenuation curve; Constructing a theoretical power attenuation curve based on the physical parameters of the optical fiber link; Calculating the absolute difference of each sampling point on the time axis of the actual power attenuation curve and the theoretical power attenuation curve to obtain an absolute difference sequence reflecting the degree of attenuation abnormality; The absolute difference sequence is normalized to obtain the standardized difference data.

6. The method for locating faults in a power distribution communication network based on deep learning according to claim 1, wherein: The extracting features of the physical topology structure of the target distribution network to obtain distribution network topology feature data includes: Based on the obtained physical connection relationship of the optical fiber link, constructing a node adjacency matrix and a link attribute table; Summing each row of the node adjacency matrix to obtain a node degree vector reflecting the importance of the node; Calculating link weight data according to the link attribute table, and constructing a link weight matrix based on the link weight data; The node degree vector and the link weight matrix are concatenated to obtain the distribution network topology characteristic data.

7. The method for locating faults in a power distribution communication network based on deep learning according to claim 1, wherein: The comparing the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially performing feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data, including: Comparing the failure probability value with a preset failure probability threshold, screening out links whose failure probability values are greater than or equal to the preset failure probability threshold, and generating a list of potential failure links; Acquire optical power sequence data of links corresponding to the faulty link list, and convert the optical power sequence data into a two-dimensional matrix; The two-dimensional matrix is spliced with the standardized difference data and the nonlinear change feature data to obtain a fused feature tensor; and the fused feature tensor is input into a pre-built deep residual network for feature extraction to obtain attenuation mutation feature data; Cluster analysis is performed on the attenuation mutation characteristic data to obtain the attenuation mutation node data where the attenuation trend has changed.

8. A distribution communication network fault location system based on deep learning, characterized in that: The method for locating a fault in a power distribution communication network based on deep learning according to claim 1 is applied, comprising: An acquisition module is used to obtain optical power data of optical fiber links in the target distribution network, calculate the optical power difference between adjacent sampling points in the optical power data, and generate time series data; a calculation module, configured to perform fitting processing on the time series data to obtain an actual power attenuation curve, and calculate an absolute difference between the actual power attenuation curve and a theoretical power attenuation curve to obtain standardized difference data; a first extraction module, configured to input the standardized difference data into a pre-built nonlinear attenuation model to obtain nonlinear variation characteristic data reflecting the attenuation of optical signal power in the optical fiber link; a second extraction module, configured to extract features of the physical topology of the target distribution network to obtain distribution network topology feature data, input the distribution network topology feature data and the nonlinear change feature data into a pre-built fault identification model to obtain a failure probability value of each optical fiber link; an identification module, configured to compare the fault probability value with a preset fault probability threshold to obtain a potential fault link, and sequentially perform feature extraction and mutation node identification on the potential fault link to obtain attenuation mutation node data; A positioning module is used to process the attenuation mutation node data according to a stress inversion method, and determine the fault point data in the target distribution network based on the obtained stress value data.

Images