A method and device for locating a cable break based on deep learning and knowledge fusion

Abstract

The present application relates to the field of optical fiber communication network testing, and particularly relates to a cable breakpoint positioning method and device based on deep learning and knowledge fusion. Feature extraction is performed on the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the cable to be positioned, the feature vector and the measurement parameters are input into a preset positioning model to obtain a breakpoint correction distance interval and a corresponding breakpoint ranging confidence; interval intersection calculation is performed on the breakpoint correction distance interval and the cumulative distance range of each cable section to determine candidate cable sections; in each point position corresponding to each candidate cable section, a point position in which the cumulative distance falls within the breakpoint correction distance interval is taken as a candidate point position; and the point position confidence of each candidate point position is calculated based on the breakpoint correction distance interval, the cumulative distance of each candidate point position and the breakpoint ranging confidence. Thus, the influence of system error and parameter setting deviation on the point finding result is reduced, and the on-site repeated trial and error and excavation rework are reduced.

Description

Technical Field

[0001] This invention relates to the field of optical fiber communication network testing, specifically to a method and apparatus for locating optical cable breakpoints based on deep learning and knowledge fusion. Background Technology

[0002] As a crucial transmission medium in fiber optic communication networks, optical cables may experience breakage or severe loss events during long-term operation due to external forces, construction damage, poor splicing quality, bending and compression, and environmental aging, leading to communication interruptions or performance degradation. To quickly locate faults, the operation and maintenance field commonly uses optical time-domain reflectometry (OTDR) for testing. This involves acquiring backscattered and reflected signals along the fiber optic link to form a measurement curve, and then estimating the fault distance based on the event characteristics within the curve. Building upon this, some operation and maintenance systems have begun to correlate test results with optical cable resource data, using information on locations such as equipment rooms, optical distribution boxes, junction boxes, manholes, and pole lines, as well as the connection relationships of optical cable segments, to assist in determining the fault range. Simultaneously, with the accumulation of data and improved computing power, curve feature extraction and event localization methods based on machine learning or deep learning have gradually emerged to improve the automation level of curve analysis and the efficiency of fault location.

[0003] The existing technology still has the following shortcomings:

[0004] Distance measurement results are easily affected by factors such as refractive index settings, pulse width, range, sampling interval, and equipment system errors, leading to distance axis deviation and consequently, large breakpoint location errors, resulting in high costs for on-site point finding and excavation rework. Some solutions only output a single breakpoint distance estimate, lacking uncertainty or confidence level expressions, and are not sufficiently integrated with the optical cable resource topology, making it difficult to stably map the distance results into executable candidate optical cable segments and candidate points for ranking and decision-making.

[0005] Therefore, a method for locating optical cable breaks that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and device for locating optical cable breaks based on deep learning and knowledge fusion, which solves the problems in the prior art.

[0007] This invention is achieved through the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a method for locating optical cable breaks based on deep learning and knowledge fusion, including:

[0009] Feature vectors are obtained by extracting features from the OTDR curves and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned.

[0010] The feature vector and the measurement parameters are input into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label.

[0011] The intersection calculation is performed between the breakpoint correction distance range and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point.

[0012] In each point corresponding to each candidate optical cable segment, the points whose cumulative distance falls within the breakpoint correction distance interval are taken as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point.

[0013] The point location confidence of each candidate point is calculated based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence. The candidate point list is then output in order of point location confidence.

[0014] Preferably, the step of extracting features from the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned, to obtain a feature vector, includes:

[0015] Obtain the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned;

[0016] Convert the time axis of the OTDR curve to the distance axis based on the refractive index in the measurement parameters;

[0017] The distance axis is resampled according to a uniform distance sampling interval to obtain a preprocessed OTDR curve;

[0018] Based on the preprocessed OTDR curve, feature vectors are extracted through a pre-trained feature extraction network, which is obtained by training a deep neural network through self-supervised contrastive learning.

[0019] Preferably, the feature extraction network is trained in the following manner:

[0020] Construct positive sample pairs based on historical preprocessed OTDR curves of the same fiber optic link under different measurement parameters;

[0021] Negative sample pairs are constructed based on the historical preprocessed OTDR curves of different fiber optic links;

[0022] The deep neural network is trained using the positive and negative sample pairs and a contrastive learning loss to obtain a feature extraction network that is robust to changes in measurement parameters.

[0023] Preferably, the localization model is trained in the following manner:

[0024] The feature vector extracted from the historical preprocessed OTDR curve and the corresponding measurement parameters are input into the three-layer fully connected structure of the localization network to be trained. The network outputs the cumulative distance prediction value of the breakpoint, the attenuation coefficient prediction value, and the reflection intensity prediction value. The attenuation coefficient prediction value is used to describe the attenuation coefficient of back Rayleigh scattering with distance, and the reflection intensity prediction value is used to describe the amplitude intensity of the breakpoint reflection event term.

[0025] The measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity are input into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve.

[0026] The physical consistency loss is calculated based on the difference between the predicted OTDR curve and the historical preprocessed OTDR curve.

[0027] The location loss is calculated based on the difference between the predicted cumulative distance of the breakpoint and the cumulative distance label of the breakpoint obtained from the repair record;

[0028] The positioning network parameters are updated based on the joint loss of the physical consistency loss and the positioning loss to obtain the positioning model.

[0029] Preferably, the step of inputting the measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve includes:

[0030] A scattering baseline term is constructed based on the back Rayleigh scattering and attenuation model and the predicted attenuation coefficient.

[0031] Based on the predicted cumulative distance of the breakpoint and the predicted reflection intensity, a reflection event item is constructed at the corresponding distance position;

[0032] The reflection event term is convolved with the instrument impulse response function determined by the pulse width in the measurement parameters to obtain the convolution result;

[0033] The convolution result is added to the scattering baseline term to obtain the predicted OTDR curve.

[0034] Preferably, the step of inputting the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, starting from the OTDR test end, includes:

[0035] The feature vector and the measurement parameters are input into a preset positioning model, and the cumulative distance prediction value of the breakpoint is output.

[0036] Based on the predicted cumulative distance of the breakpoint and its prediction uncertainty, a distance interval centered on the predicted cumulative distance of the output breakpoint and extended to both sides by a certain multiple of uncertainty is determined as the breakpoint correction distance interval.

[0037] The breakpoint ranging confidence level corresponding to the breakpoint correction distance interval is calculated based on the prediction uncertainty.

[0038] Preferably, before performing the interval intersection calculation between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end, the method further includes:

[0039] Based on the routing data of the optical cable to be located, obtain all point data and optical cable segment data on the optical cable route. The point data includes point identifier, point type and location information. The optical cable segment data includes optical cable segment identifier, optical cable segment length and the identifiers of the two points connected to the optical cable segment.

[0040] Based on the location data, corresponding location entities are created in the knowledge graph, and each location entity is associated with its location information and location type attribute.

[0041] Based on the optical cable segment data, a corresponding optical cable segment entity is created in the knowledge graph, and each optical cable segment entity is associated with its optical cable segment length attribute.

[0042] Based on the location identifiers connected to the optical cable segment, a connection relationship between the optical cable segment entity and the location entities at both ends is established in the knowledge graph;

[0043] Based on the sequential relationship of the optical cable routes to be located, a sequential relationship is established for adjacent optical cable segment entities in the knowledge graph, resulting in a constructed optical cable resource knowledge graph.

[0044] Preferably, the step of calculating the intersection of the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment includes:

[0045] Based on the routing order relationship in the optical cable resource knowledge graph, determine the optical cable segment sequence starting from the OTDR test end;

[0046] The cumulative distance range of each optical cable segment relative to the OTDR test end is obtained by sequentially accumulating the lengths of each optical cable segment.

[0047] The distance interval for breakpoint correction is intersected with the cumulative distance range of each optical cable segment. Optical cable segments whose interval intersection results are not empty are identified as candidate optical cable segments.

[0048] Preferably, the step of calculating the point location confidence of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence includes:

[0049] For each candidate point, the cumulative distance deviation is calculated based on the deviation between the cumulative distance of the candidate point and the center of the breakpoint correction distance interval.

[0050] Based on the cumulative distance deviation and the breakpoint ranging confidence level, the point location confidence level of the candidate point is calculated using an exponential decay function.

[0051] Secondly, embodiments of the present invention provide a fiber optic cable breakpoint location device based on deep learning and knowledge fusion, comprising:

[0052] The feature extraction module is used to extract features from the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned, and obtain feature vectors.

[0053] The positioning module is used to input the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label.

[0054] The interval matching module is used to perform interval intersection calculations between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point.

[0055] The candidate point module is used to select points whose cumulative distance falls within the breakpoint correction distance range for each point corresponding to each candidate optical cable segment as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point.

[0056] The point location reliability module is used to calculate the point location reliability of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence, and outputs a list of candidate points sorted by point location reliability.

[0057] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0058] This invention addresses the issue of large breakpoint location errors caused by ranging deviations in optical time-domain reflectometers (OTDRs) by combining self-supervised deep learning, physically constrained deep networks, and knowledge fusion. It employs a closed-loop improvement mechanism: First, OTDR curves under different test parameters are uniformly preprocessed. Then, positive sample pairs are constructed using curves with different measurement parameters on the same link for comparative learning and self-supervised training, making the feature representation more robust to changes in refractive index, pulse width, range, and sampling interval. Subsequently, a cumulative breakpoint distance label is formed by combining repair records. During training, a joint loss is introduced, combining the consistency constraint between the predicted and measured curves with the breakpoint distance location constraint. This ensures that the model output not only provides the cumulative breakpoint distance but also the breakpoint correction distance interval and ranging confidence level, thereby reducing the impact of system errors and parameter setting deviations on the point-finding results and minimizing repeated on-site trials and rework.

[0059] This invention proposes targeted improvements: First, a differentiable OTDR forward computation layer is set in the positioning network. A scattering baseline term is constructed based on the backward Rayleigh scattering and attenuation model. A reflection event term is introduced at the breakpoint and convolved with the instrument impulse response function to generate a predicted OTDR curve. This makes the positioning results subject to physical mechanism constraints and can further output distance axis correction parameters to compensate for OTDR ranging system errors. Second, a two-level mapping structure is adopted to first map the "distance interval" to candidate optical cable segments and then to candidate points. The point position confidence is calculated by combining the ranging confidence and the cumulative distance deviation of the point, and the results are sorted and output. This transforms the model results from abstract distance values ​​into executable segment-level and point-level positioning schemes, which is more conducive to improving emergency repair efficiency and positioning reliability. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0061] Figure 1 A flowchart illustrating the optical cable breakpoint localization method based on deep learning and knowledge fusion provided by this invention;

[0062] Figure 2 A schematic diagram of the optical cable breakage location device based on deep learning and knowledge fusion provided by the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0065] It should be noted that all actions involving the acquisition of signals, information, or data in this invention are carried out in compliance with the relevant data protection laws and regulations of the locality and with authorization from the owner of the relevant device.

[0066] Example 1

[0067] Please see Figure 1 This invention provides a method for locating optical cable breakpoints based on deep learning and knowledge fusion, comprising:

[0068] S1. Extract features from the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned, and obtain the feature vector;

[0069] Specifically, the OTDR curve is a waveform diagram showing the change in backscattered and reflected optical power as a function of distance after an optical pulse is injected into an optical fiber by an optical time domain reflectometer. The horizontal axis of the curve represents distance, and the vertical axis represents relative optical power (expressed in decibels).

[0070] Measurement parameters are a set of configuration parameters used by an OTDR device when performing measurements. These parameters include at least refractive index, pulse width, range, and sampling interval. Together, these parameters determine the distance conversion, resolution, range, and data point density of the OTDR curve.

[0071] Feature extraction refers to the process of processing and transforming the original OTDR curve and measurement parameters to extract a high-level abstract representation that can characterize the curve shape and is suitable for subsequent model calculations.

[0072] This step involves first converting the time axis of the OTDR curve into a distance axis based on the refractive index in the measurement parameters, and then standardizing and resampling the distance axis according to the sampling interval to eliminate distance scale differences caused by different devices or settings. Subsequently, the curve is denoised and its amplitude normalized to suppress noise interference and unify the signal amplitude reference. Finally, the standardized curve is processed by a deep neural network trained to capture key patterns related to breakpoint events in the curve and output a low-dimensional feature vector. This process transforms the raw, high-dimensional, noisy measurement data into a stable, compact feature representation, laying a reliable data foundation for subsequent model inference and reducing the direct impact of original measurement errors and parameter differences on the positioning process.

[0073] In some embodiments, S1, the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned are subjected to feature extraction to obtain a feature vector, including:

[0074] S11. Obtain the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned;

[0075] Specifically, acquisition refers to the process of receiving or reading data from an optical time domain reflectometer or storage system.

[0076] The OTDR test end is the starting point for this optical time domain reflectance measurement, and it is usually located at one end of the optical cable link, such as a specific connection port in the equipment room or optical junction box.

[0077] This step marks the beginning of the localization method and involves data acquisition. The system first needs to simultaneously acquire two core data points from the optical time-domain reflectometer (OTDR) performing the measurement, or from a stored test file. The first is the OTDR curve, which is the raw signal waveform data obtained by the instrument. The second is the specific settings parameters used by the OTDR during this measurement; these parameters correspond one-to-one with the curve. Ensuring that both are obtained simultaneously and correctly correlated is fundamental to all subsequent processing.

[0078] S12. Convert the time axis of the OTDR curve to the distance axis based on the refractive index in the measurement parameters;

[0079] Specifically, the time axis is the original horizontal axis of the OTDR curve, representing the time elapsed after the light pulse is emitted. The distance axis is the horizontal axis obtained by converting the time axis according to the speed of light propagation in the optical fiber, representing the physical distance traveled along the optical fiber.

[0080] The core of this step is the physical transformation of the coordinate system. Since the optical time-domain reflectometer directly measures the return time of the optical signal, while the actual requirement is the spatial location of the fault point, a conversion is necessary. In practice, based on the refractive index value from the simultaneously acquired measurement parameters and the known speed of light in a vacuum, the physical principle that the speed of light in an optical fiber is inversely proportional to the refractive index is used to calculate the time value corresponding to each sampling point on the curve, converting it into the one-way fiber length represented by that sampling point relative to the test end. After this conversion, the horizontal axis of the curve changes from time units to distance units, allowing the curve shape to directly reflect the loss and reflection at various points along the fiber.

[0081] S13. Resample the distance axis according to a uniform distance sampling interval to obtain a preprocessed OTDR curve;

[0082] Specifically, the uniform distance sampling interval is a pre-set fixed value that defines the distance difference represented by adjacent sampling points on the processed OTDR curve.

[0083] Resampling involves calculating the corresponding signal amplitude at new, equally spaced distance coordinate points using an interpolation algorithm, thereby generating a new curve.

[0084] The purpose of this step is to standardize the data specifications. Different OTDR devices or measurement settings may generate raw data with varying distance intervals, which can complicate subsequent processing by deep neural networks based on a fixed architecture. Therefore, the system predefines a standard distance sampling interval. Then, for the curve already converted to the distance axis, a series of new, equidistant distance coordinate points are generated within this standard interval, ranging from zero to the measurement range. For each new coordinate point, the signal amplitude at that new location is estimated using linear interpolation and other calculation methods, based on the amplitudes of adjacent data points on the original curve. Ultimately, all the new coordinate points and their estimated amplitudes constitute a preprocessed OTDR curve with a uniform sampling interval and fixed length. This operation eliminates scale differences in the input data and ensures the stability of the model processing.

[0085] For example, the original discrete amplitude sequence of the OTDR curve to be located is obtained from an optical time domain reflectometer (OTDR), and measurement parameters corresponding one-to-one with the OTDR curve to be located are simultaneously obtained. The measurement parameters are determined by the refractive index. Pulse width Range With sampling interval Composition, in which the refractive index Used for converting from the time axis to the distance axis, pulse width Used to determine the time scale and range of noise reduction and smoothing. Used to define the effective distance range, sampling interval Used to determine the time interval between adjacent sampling points;

[0086] Subsequently, denoising processing is performed on the OTDR curve to be located to obtain a denoised OTDR curve. The denoising processing is implemented using a one-dimensional moving average filter, and the filter window length is set to the pulse width. Divide by sampling interval The obtained sample number is rounded up and then doubled and added to construct an odd number of sample points of not less than 3. Mirror padding is used at the sequence boundaries to ensure that the filter output length is consistent with the input length and to avoid abrupt changes in boundary amplitude.

[0087] Then, amplitude normalization is performed based on the amplitude range of the denoised OTDR curve to obtain a normalized OTDR curve, specifically within the range. Within the corresponding effective distance range, the minimum and maximum amplitudes of the denoised OTDR curves are statistically analyzed, and the amplitude of each sampling point is linearly mapped to the minimum and maximum values. This range eliminates amplitude scale differences caused by different test echo intensities and gain settings;

[0088] Then based on the refractive index Convert the time axis of the normalized OTDR curve to a distance axis, specifically for the th... Construct the distance coordinates of each sampling point. and in accordance with:

[0089] ;

[0090] in Indicates the first The distance of each sampling point relative to the OTDR test end. Represents the speed of light in a vacuum and takes the value of , This indicates the sampling point number, which is counted starting from 0. Indicates the sampling interval. The value represents the refractive index, and the coefficient 2 in the formula indicates that the OTDR measurement is converted to the distance corresponding to the round-trip propagation time.

[0091] Finally, based on the sampling interval The distance axis is resampled to achieve uniform distance axis processing. The system pre-fixes a uniform distance sampling interval. And based on this, a range from 0 to... A uniform distance grid is used to normalize the OTDR curve in its original distance coordinates. Piecewise linear interpolation is used to map the data to the uniform distance grid and truncate the data beyond the range. The part that yields the preprocessed OTDR curve with consistent distance axis sampling and amplitude scale is obtained.

[0092] S14. Based on the preprocessed OTDR curve, feature vectors are extracted through a pre-trained feature extraction network, which is obtained by training a deep neural network through self-supervised contrastive learning.

[0093] Specifically, pre-training refers to the state of a model that has been trained using other relevant data before performing the current localization task.

[0094] Self-supervised contrastive learning is a machine learning training method that does not require manual labeling. It allows the model to learn the intrinsic feature representation of the data by constructing similarity and dissimilarity relationships between samples.

[0095] This step is responsible for extracting high-level abstract features from the standardized curve. A pre-trained feature extraction network, typically composed of convolutional layers, is deployed in the system. The pre-processed OTDR curve is input into this network, which automatically analyzes details such as the curve's shape, event pattern, and attenuation trend through its multi-layered nonlinear transformations, encoding this information into a low-dimensional, dense numerical vector—the feature vector. This feature extraction network is not randomly initialized for the current task; instead, it is pre-trained using a large number of unlabeled historical OTDR curves through self-supervised contrastive learning. In this training, the network is taught to classify curves obtained from the same fiber under different measurement parameters as similar, while classifying curves from different fibers as dissimilar. This forces the network to learn to ignore the interference of measurement parameter differences and focus on extracting the inherent, stable features of the fiber link itself. Therefore, the feature vector output from this step is more robust to changes in measurement conditions and can more fundamentally represent the current state of the measured fiber.

[0096] In some implementations, the feature extraction network is trained in the following manner:

[0097] Construct positive sample pairs based on historical preprocessed OTDR curves of the same fiber optic link under different measurement parameters;

[0098] Negative sample pairs are constructed based on the historical preprocessed OTDR curves of different fiber optic links;

[0099] The deep neural network is trained using the positive and negative sample pairs and a contrastive learning loss to obtain a feature extraction network that is robust to changes in measurement parameters.

[0100] Specifically, historical preprocessed OTDR curves refer to the set of curve data measured by an optical time domain reflectometer and preprocessed according to the methods described in S11 to S13 during the testing, maintenance, or fault repair of the optical fiber network over a period of time.

[0101] A positive sample pair is a pair consisting of two different OTDR curve data. The core criterion for its determination is that the two curves come from the same physical fiber optic link, but correspond to two or more independent tests using different measurement parameters (such as different refractive indices and different pulse widths).

[0102] Negative sample pairs are also pairs composed of two different OTDR curve data, and the core criterion for their determination is that these two curves come from two different physical fiber optic links.

[0103] The training process aims to teach the feature extraction network to focus on the inherent properties of the optical fiber itself, rather than representational differences introduced by variable measurement settings. In implementation, training sample pairs are first selected and constructed from historical datasets. For positive sample pairs, the system searches for multiple test records for the same optical fiber, ensuring that the measurement parameters corresponding to these records differ—for example, one test uses a 1550nm wavelength, another uses a 1310nm wavelength, or different pulse widths—and then uses these paired curves as positive samples. For negative sample pairs, the system randomly selects two test curves belonging to different optical fiber links for pairing. During training iterations, batches of data containing multiple such positive and negative sample pairs are input into the deep neural network to be trained. The network encodes each curve and outputs the corresponding feature vector. The objective function for training is a contrastive learning loss, which is designed to calculate the similarity between the feature vectors of the sample pairs. Its optimization direction is to reduce the distance between the two feature vectors in the representation space of positive sample pairs, while increasing the distance between the two feature vectors in negative sample pairs. By training with a large number of such sample pairs, the network gradually adjusts its internal parameters, ultimately developing a feature extraction capability: it can map various curve shapes of the same optical fiber under different test conditions to very close positions in the feature space; while mapping curves of different optical fibers to positions that are far apart. Therefore, the trained network is robust to changes in measurement parameters, and its extracted feature vectors better reflect the essential identity and state of the optical fiber link, rather than the appearance of a specific test. This provides a stable and discriminative input for subsequent localization models.

[0104] For example, the set of historical OTDR curves corresponding to the optical cable to be located and the measurement parameters of each historical OTDR curve are obtained from the OTDR historical test library, and a one-to-one correspondence is established. The historical OTDR curve is denoted as Record its measurement parameters as ,in Indicates refractive index, Indicates pulse width. Indicates the range, This indicates the sampling interval, and also assigns a fiber optic link identifier to each historical OTDR curve. To indicate the specific fiber optic link corresponding to the curve;

[0105] For each historical OTDR curve Perform preprocessing, including one-dimensional moving average filtering for noise reduction, and in-range... Amplitude normalization is performed within the corresponding effective range, based on refractive index. Convert the time axis to a distance axis and resample to a uniform distance sampling interval. This yields historical preprocessed OTDR curves of consistent length. And represent it as an amplitude sequence arranged in ascending order of distance;

[0106] Subsequently, training sample pairs were constructed based on the historical preprocessed OTDR curves, and positive and negative sample pairs were distinguished. Positive sample pairs were defined as those originating from the same fiber optic link. And measurement parameters An ordered pair consisting of two not identical historical preprocessing OTDR curves The measurement parameters are not completely the same. If at least one of the values ​​is different, a negative sample pair is defined as one that comes from a different fiber optic link. An ordered pair consisting of two historical preprocessed OTDR curves And in each training iteration, construct a structure containing Batch data of anchor samples, where each anchor sample Corresponding to a unique positive sample Furthermore, all other samples within the batch are considered negative samples of the anchor sample;

[0107] Next, the training samples are input to the feature extraction network of the deep neural network and subjected to a two-branch encoding with shared weights. The feature extraction network is denoted as... And its network parameters are The network structure consists of a sequentially connected one-dimensional convolutional layer, three sets of residual blocks, and a global average pooling layer. The one-dimensional convolutional layer has a kernel length of 7 and 64 output channels. The three sets of residual blocks have the following output channel numbers: Each residual block contains two residual units, and the global average pooling layer outputs a 256-dimensional feature vector, which is then connected to a projection head at the output of the feature extraction network. And its network parameters are The projection head consists of two fully connected layers with dimensions mapped as follows: and perform on the projection output. Normalization yields the feature vectors used for contrastive learning. ;

[0108] Then, the contrastive learning loss is calculated based on the feature vectors of positive and negative sample pairs and used to update the network parameters. The contrastive learning loss adopts the temperature-scaled InfoNCE loss and is calculated as follows:

[0109] ;

[0110] in Indicates the contrast learning loss. Indicates the number of anchor samples within a batch and takes the value of , Indicates the anchor sample index. Indicates by Obtained and passed Normalized anchor sample feature vector, Represents anchor sample Corresponding positive samples Obtained through the same network and through Normalized positive sample feature vectors Represents anchor sample The set of indexes of all samples except the positive samples in the same batch of data. Indicates that the index is The negative sample feature vector, Represents the temperature coefficient and takes the value of Represents cosine similarity and is calculated as the inner product of normalized vectors;

[0111] Using the Adam optimizer and Perform joint updates, with the learning rate set to a value. The weight decay value is The training epochs are set to 200 epochs, and the feature extraction network is retained after training. And discard the projection head Thus, a self-supervised feature extraction network was obtained.

[0112] S2. Input the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label.

[0113] Specifically, the localization model is a trained machine learning model whose core function is to predict the possible distance range of the fiber optic break point and the reliability of the prediction based on the input OTDR curve features and measurement parameters.

[0114] The breakpoint correction distance interval is a numerical range in units of distance. It represents the range of physical locations where the breakpoint predicted by the model is most likely to exist. This interval is determined by the predicted center distance value and its uncertainty.

[0115] The confidence level of a breakpoint distance prediction is a value or percentage between 0 and 1, used to quantify the reliability of the predicted distance range after the breakpoint correction. A higher confidence level usually means that the model is more confident in the distance prediction.

[0116] This step involves inputting the feature vector generated in the previous step, along with the corresponding measurement parameters, into a pre-trained localization model. Internally, the model, through learned mapping relationships, first estimates a predicted value for the breakpoint distance and its corresponding prediction uncertainty. Subsequently, based on this predicted value and uncertainty, the model calculates a distance interval as the breakpoint correction distance interval according to predetermined rules, and simultaneously outputs a confidence score reflecting the quality of this prediction. This step transforms abstract curve features into distance estimates with clear physical meaning and reliability indicators, enabling subsequent localization processes to no longer rely on a single distance point, but rather to make more robust decisions based on a reasonable distance range and its confidence level.

[0117] In some implementations, the localization model is trained in the following manner:

[0118] The feature vector extracted from the historical preprocessed OTDR curve and the corresponding measurement parameters are input into the three-layer fully connected structure of the localization network to be trained. The network outputs the cumulative distance prediction value of the breakpoint, the attenuation coefficient prediction value, and the reflection intensity prediction value. The attenuation coefficient prediction value is used to describe the attenuation coefficient of back Rayleigh scattering with distance, and the reflection intensity prediction value is used to describe the amplitude intensity of the breakpoint reflection event term.

[0119] The measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity are input into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve.

[0120] The physical consistency loss is calculated based on the difference between the predicted OTDR curve and the historical preprocessed OTDR curve.

[0121] The location loss is calculated based on the difference between the predicted cumulative distance of the breakpoint and the cumulative distance label of the breakpoint obtained from the repair record;

[0122] The positioning network parameters are updated based on the joint loss of the physical consistency loss and the positioning loss to obtain the positioning model.

[0123] Specifically, a three-layer fully connected structure is a deep neural network component with three successively connected linear transformation layers and nonlinear activation functions, used to achieve complex mappings from input features to multidimensional predicted values.

[0124] Differentiable forward computation layer is a computational module embedded in a neural network and built based on the physical principle of light temporal reflection. Its internal operations all support gradient calculation, so that physical laws can participate in network training in the form of loss functions.

[0125] The goal of this training process is to construct a localization model that can accurately predict breakpoint distances while adhering to physical laws. During the training phase, each historical preprocessed OTDR curve is first converted into a feature vector through a feature extraction network with fixed parameters. Subsequently, this feature vector is concatenated with its corresponding measurement parameters and input into the three-layer fully connected structure to be trained in the localization network. This three-layer fully connected structure undergoes a nonlinear transformation, outputting three scalar prediction values ​​with clear physical meanings. Specifically, the cumulative breakpoint distance prediction directly corresponds to the fault location estimate; the attenuation coefficient prediction describes the exponential decay rate of the backscattered Rayleigh signal in the optical fiber, determined by material properties, with increasing transmission distance; and the reflection intensity prediction quantifies the signal amplitude of the Fresnel reflection event caused by the abrupt change in refractive index at the breakpoint.

[0126] Next, the differentiable forward computation layer in the localization network is invoked. This layer receives the three predicted values ​​and the original measurement parameters as input. Internally, it first simulates and generates a smooth baseline curve containing only Rayleigh scattering attenuation based on the attenuation coefficient prediction and measurement parameters. Simultaneously, based on the cumulative distance prediction and reflection intensity prediction, simulated reflection events are superimposed at the corresponding distance positions on the baseline curve. To more realistically simulate instrument effects, this reflection event is convolved with the instrument response function determined by the pulse width in the measurement parameters to reflect the pulse broadening phenomenon in actual OTDRs. Finally, the convolved reflection event is superimposed onto the baseline curve to generate a complete predicted OTDR curve.

[0127] To ensure the training process aligns with physical reality, the model calculates two losses. The physical consistency loss is calculated by comparing the overall shape difference between the generated predicted OTDR curve and the historical preprocessed OTDR curves (i.e., the actual measurement curves) input to the network; for example, by calculating the mean absolute error of the amplitude at each sampling point. The localization loss is calculated by comparing the difference between the network's output cumulative distance prediction for breakpoints and the actual cumulative distance labels for breakpoints parsed from historical repair work orders. These two loss terms are weighted and summed according to preset weights to form the joint loss. Through backpropagation, the gradient generated by the joint loss updates all trainable parameters in the localization network, including the weights and biases of the three fully connected layers. Through iterative training with a large number of samples, the network parameters are gradually optimized, enabling the model to ultimately output predictions that are numerically close to the actual labels and can reasonably interpret the measurement curves through the physical model, thus obtaining the trained localization model.

[0128] In some embodiments, the step of inputting the measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve includes:

[0129] A scattering baseline term is constructed based on the back Rayleigh scattering and attenuation model and the predicted attenuation coefficient.

[0130] Based on the predicted cumulative distance of the breakpoint and the predicted reflection intensity, a reflection event item is constructed at the corresponding distance position;

[0131] The reflection event term is convolved with the instrument impulse response function determined by the pulse width in the measurement parameters to obtain the convolution result;

[0132] The convolution result is added to the scattering baseline term to obtain the predicted OTDR curve.

[0133] Specifically, the backscattering and attenuation model is a mathematical model that describes the exponential attenuation of the intensity of scattered light signal as it propagates in an optical fiber due to the microscopic inhomogeneities of the material.

[0134] The scattering baseline term refers to the smooth and monotonically decaying curve calculated under this model, representing the shape that the OTDR curve should present under ideal conditions without discrete reflection events (such as breakpoints or connectors).

[0135] A reflection event is a signal pulse with a specific amplitude that occurs at a specific distance location, used to simulate Fresnel reflection caused by abrupt changes in the refractive index of an optical fiber (such as a break or connection end face).

[0136] The instrument pulse response function is a function that describes the time / space broadening characteristics of the light pulse emitted by the OTDR device. It can usually be approximated as a Gaussian distribution, and its width is determined by the pulse width in the measurement parameters.

[0137] This step, within the internal computation of the differentiable forward computation layer, first constructs the scattering baseline term. This process is based on the backscattering Rayleigh scattering and attenuation model, inputting the predicted attenuation coefficient, which determines the rate attenuation of the scattered light signal with distance. The calculation proceeds along a uniform distance axis starting from the test end. For each distance sampling point, its baseline amplitude is calculated using the model formula (which can be the initial amplitude multiplied by a negative exponential function with the attenuation coefficient and distance as exponents), thus forming a curve that continuously attenuates from the starting point. Next, the reflection event term is constructed. The system determines a specific position coordinate on the distance axis based on the cumulative distance prediction values ​​at breakpoints. At this position, an initial pulse amplitude is set based on the predicted reflection intensity value. At this point, the reflection event is represented as an idealized, extremely narrow pulse on the distance axis. Then, to simulate the event broadening effect caused by the finite pulse width in a real OTDR device, the reflection event term needs to be convolved. The instrument pulse response function is obtained by converting the pulse width in the measurement parameters; this function defines the broadening shape of the pulse in space. The constructed ideal reflection event term is discretically convolved with the impulse response function. This convolution smooths and broadens the originally sharp ideal reflection pulse, making its shape closer to the smoother reflection peaks observed in actual OTDR curves. Finally, the convolved, more realistic reflection event signal is added point-by-point to the previously calculated scattering baseline term. The result is a composite curve that includes both the inherent Rayleigh scattering attenuation background of the fiber and the characteristic peaks generated by the predicted intensity reflection event at the predicted location. This composite curve is the final generated predicted OTDR curve. The entire process uses differentiable operators, ensuring gradient backtracking along the entire computational path from input physical parameters to the output predicted curve, thus allowing the physical consistency loss to effectively guide the optimization of the localization network parameters.

[0138] For example, obtain the OTDR curve for each historical preprocessing line. The corresponding repair records are then used to establish a data index relationship aligned with the fiber optic link identifier, where Indicates the first The historical preprocessed OTDR curves and their distance axes have been sampled at a uniform distance interval. Resampling yields a discrete amplitude sequence. This indicates the sample sequence number; the repair record contains the location identifier of the breakpoint. The optical cable segment marker where the break point is located ,in The identifier of the point entity where the breakpoint occurred. This indicates the physical identifier of the optical cable segment carrying the breakpoint and is used to eliminate ambiguity when the point is reused or multiple segments are converged;

[0139] Then, based on the given optical cable routing sequence and optical cable segment length in the optical cable resource data, the point sequence and optical cable segment sequence starting from the OTDR test endpoint are determined, and the breakpoint markers are located on this sequence. Its associated optical cable segment identifier The location index is obtained by sequentially summing the lengths of all optical cable segments from the starting point to that location to obtain the historical breakpoint cumulative distance label. ,in This represents the cumulative distance from the OTDR test terminal along the route to the breakpoint, in meters.

[0140] Subsequently, a self-supervised feature extraction network was used. The output end connects to the positioning network. ,in This indicates a self-supervised feature extraction network with network parameters of... Location network The network parameters are The input to the localization network consists of historical preprocessed OTDR curves. and its measurement parameters Constituting and

[0141] Location network right Output 256-dimensional feature vector and measurement parameters (in Indicates refractive index, Indicates pulse width. Indicates the range, The normalized values ​​(representing the sampling interval) are concatenated and then input into a three-layer fully connected structure. Accumulated distance prediction value from output breakpoints Attenuation coefficient prediction value Compared with the predicted value of reflection intensity ,in This represents the cumulative distance of the breakpoints output by the localization network. This represents the attenuation coefficient used to describe the decrease in backscattering Rayleigh scattering with distance, and its unit is 1. It represents the magnitude intensity of the breakpoint reflection event term and has the same dimensions as the normalized OTDR magnitude;

[0142] Setting up a differentiable OTDR forward computation layer in the positioning network And it is used as the end physical generation module of the positioning network, in which by With measurement parameters Input and generate predicted OTDR curves on a uniform distance axis Specifically, the length is determined by the range. With uniform distance sampling interval Construct a scattering baseline term formed by backscattering and attenuation on a jointly determined discrete range grid and at the range position. Construct a reflection event item, and associate the reflection event item with the pulse width. The one-dimensional Gaussian impulse response kernel corresponding to the converted spatial impulse width is discretized and convolved to characterize the instrument's impulse broadening effect. The convolution result is then added to the scattering baseline term to obtain the predicted OTDR curve. The length of the Gaussian impulse response kernel is taken as the spatial impulse width divided by . The number of sample points obtained by rounding up is set to one-sixth of the length, and the kernel coefficients are normalized to a sum of 1 to ensure the conservation of reflected energy. Meanwhile, the differentiable OTDR forward computation layer is implemented using a one-dimensional convolution operator and maintains the... End-to-end differentiability;

[0143] Then based on and Point-by-point difference calculation of physical consistency loss The mean absolute error is used as a measure of difference to constrain the consistency of the shape between the predicted curve and the measured curve, while based on... Cumulative distance from historical breakpoints (label) Difference calculation of positioning loss Huber loss was used with a threshold of 50m to maintain training stability in the presence of a small amount of label noise;

[0144] Finally, the combined losses For positioning network parameters Perform backpropagation to update and obtain the localization model, where Indicates joint loss, This represents the weight of the physical consistency loss and takes the value of . Indicates the loss of physical consistency. This represents the positioning loss weights and their values ​​are... This represents the localization loss, and the training uses the Adam optimizer with a learning rate of [value]. The batch size is set to 128, the number of training epochs is set to 120, and the self-supervised feature extraction network is fixed during training. parameters Without updating, the positioning network The mapping relationship between the cumulative distance of breakpoints and physical parameters is learned in a stable representation space.

[0145] The forward computation layer of a differentiable OTDR is denoted as And deployed in the positioning network described in step S3 for generating predicted OTDR curves. To participate in the calculation of physical consistency loss;

[0146] The input to the differentiable OTDR forward computation layer includes the breakpoint cumulative distance output by the positioning network. attenuation coefficient Reflection intensity and measurement parameters ,in This represents the cumulative distance from the breakpoint starting from the OTDR test terminal, with units of 1. This represents the attenuation coefficient of backscattered Rayleigh scattering with distance, and its unit is 1. This represents the amplitude intensity of the breakpoint reflection event term and has the same dimensions as the amplitude of the preprocessed OTDR curve. Indicates refractive index, Indicates pulse width and the unit is Indicates the measurement range and the unit is meters (m). Indicates the sampling interval and the unit is seconds (s);

[0147] The differentiable OTDR forward computation layer operates at a uniform sampling interval. Constructing Discrete Distance Axes and satisfy ,in Indicates the first The distance from each sampling point to the OTDR test terminal. This represents the distance from the sampling point index, counted starting from 0, and the total number of sampling points. Take as To ensure coverage range ;

[0148] The differentiable OTDR forward computation layer constructs a scattering baseline term based on a backward Rayleigh scattering and attenuation model and a reflection event term based on the cumulative distance of the breakpoint, wherein the initial amplitude coefficient of the scattering baseline term is denoted as... Furthermore, based on the measured historical preprocessed OTDR curves Before the starting end The amplitude values ​​at each sampling point are taken as an arithmetic mean to ensure that the amplitude scale of the predicted curve is consistent with... Consistent, the reflection event term is implemented using differentiable range interpolation pulses and at distance Nearby sampling points are assigned to adjacent sampling points according to a triangular kernel, thus maintaining alignment on the discrete grid. Differentiability;

[0149] The differentiable OTDR forward computation layer is also based on pulse width. Determined instrument impulse response function Convolution is performed on the reflection event term, where the impulse response function... Take a discrete Gaussian kernel and use the spatial standard deviation Determine the kernel width and Represents the speed of light in vacuum, and is the half-width index of the discrete Gaussian kernel. Take as To cover The main energy range is normalized with the kernel coefficients summed to 1 to maintain the energy conservation of the reflection event;

[0150] Under the above definition, the differentiable OTDR forward computation layer for the first... The amplitude of the predicted OTDR curve at a distance from the sampling point Generate using the following formula:

[0151] ;

[0152] in Indicates the predicted OTDR curve At the distance from the sampling point The amplitude at that point, This represents the initial amplitude coefficient of the scattering baseline term. This represents an exponential function with the natural constant as its base. Indicates the attenuation coefficient. Indicates the first The distance to the sampling point is indicated by the symbol 2, which represents the two-way attenuation introduced by the OTDR backscattering propagation. This represents the discrete summation operation. Represents the kernel index. This represents the half-width index of the convolution kernel. Indicates the intensity of reflection. The operator that takes the larger value is used to achieve non-negative truncation of the triangular kernel. This represents the absolute value operation. Indicates a uniform distance sampling interval. This indicates the distance sampling point position aligned with the convolution index. Indicates the cumulative distance from the breakpoint. Indicates the impulse response function at index Discrete kernel coefficients at the specified location and satisfying ;

[0153] The differentiable OTDR forward computation layer is implemented by using a one-dimensional convolution operator for the summation term and a tensor absolute value and truncation operator for the triangular kernel term, thereby ensuring... and Gradients are backpropagable during end-to-end training and during generation. Later according to Amplitude clipping is performed to match the amplitude normalization result in step S1.

[0154] In this specific embodiment, the positioning network accumulates distance at the output breakpoint. attenuation coefficient With reflection intensity Based on this, the distance axis correction parameters are further output, and the differentiable OTDR forward calculation layer corrects the distance axis determined by the refractive index based on the distance axis correction parameters to compensate for the OTDR ranging system error;

[0155] Among them, the cumulative distance of the breakpoint This represents the cumulative distance from the breakpoint starting from the OTDR test end, in meters, and the attenuation coefficient. This represents the two-way equivalent attenuation coefficient of backscattering Rayleigh scattering, with units of 1. Reflection intensity The amplitude of the discontinuity reflection event term is represented and has the same dimensions as the preprocessed OTDR curve amplitude. The distance axis correction parameter is determined by the relative scale correction amount. With distance offset correction Both are constituted and generated by the linear output nodes of the positioning network, where It represents the correction amount for the proportional error of the distance axis and is a dimensionless quantity. It represents the correction amount for the fixed offset error of the distance axis, and the unit is meters (m).

[0156] The output layer dimension of the positioning network is fixed at 5 and corresponds sequentially to... ,in and Each consists of two raw outputs and The constraint mapping is obtained to ensure controllable correction amplitude and avoid training divergence. The constraint mapping is implemented using the hyperbolic tangent function and... Limited to Within range and will Limited to Within this range, common ranging error scales caused by refractive index setting deviations, sampling clock errors, and equipment system delays are covered;

[0157] The differentiable OTDR forward computation layer generates the predicted OTDR curve. First, follow step S1 based on the measurement parameters Obtain the uncorrected distance axis ,in Indicates a set of measurement parameters. Indicates refractive index, Indicates pulse width. Indicates the range, Indicates the sampling interval. Indicates the first The uncorrected distance from each sampling point to the OTDR test terminal, in units of Indicates the distance from the sampling point number and is counted starting from 0;

[0158] Then, the uncorrected distance axis is corrected based on the distance axis correction parameters to obtain the corrected distance axis. And calculate using the following formula:

[0159] ;

[0160] in Indicates the first The calibration distance from each sampling point to the OTDR test terminal, in units of... This represents the relative scale correction amount. Indicates uncorrected distance. Indicates the distance offset correction amount;

[0161] The differentiable OTDR forward computation layer is based on the corrected distance axis. Construct a scattering baseline term and accumulate distance at breakpoints Construct a reflection event term, where the distance independent variable of the scattering baseline term is used. To ensure that the attenuation pattern changes consistently with the correction distance, the position matching of the reflection event term is adopted. To align breakpoint events on the corrected distance axis, while the reflection event items remain aligned with the pulse width. Discrete convolution is performed on the determined instrument impulse response function to generate the predicted OTDR curve. And compared with the measured historical preprocessed OTDR curve Calculate the physical consistency loss to drive and During end-to-end training, it automatically converges to a value that can compensate for the ranging system error.

[0162] In some implementations, S2 involves inputting the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval starting from the OTDR test end and the corresponding breakpoint ranging confidence level, including:

[0163] S21. Input the feature vector and the measurement parameters into a preset positioning model, and output the predicted value of the cumulative distance of the breakpoint;

[0164] Specifically, the localization model has been pre-trained, internally encoding a nonlinear mapping from OTDR curve features and measurement parameters to the breakpoint location. In implementation, the feature vector extracted in step S14 and the original measurement parameters obtained in step S11 are provided as a complete set of inputs to the loaded localization model. The model first processes these inputs through its network forward propagation process. Components such as the fully connected structure within the model perform nonlinear transformations and synthesis of the input information, ultimately generating a specific numerical value at its output node—the breakpoint cumulative distance prediction. This value represents the model's estimate of the straight fiber length from the OTDR test end to the fiber breakpoint based on the current input information. This step completes the conversion from raw signal features to a single distance estimate, providing a foundation for subsequently constructing a more comprehensive distance description.

[0165] S22. Based on the predicted cumulative distance of the breakpoint and its prediction uncertainty, determine the breakpoint correction distance interval as the distance interval centered on the predicted cumulative distance of the output breakpoint and extended to both sides by a certain multiple of uncertainty.

[0166] Specifically, prediction uncertainty is an auxiliary scalar output by the positioning model along with the distance prediction value. It quantifies the model's confidence level or estimate of the possible error range of its own distance prediction result. A certain multiple is a pre-defined, positive-zero constant coefficient used to extend the uncertainty value to a symmetrical distance range.

[0167] This step aims to transform a single distance point estimate into a more robust and practical distance interval. The localization model outputs a distance prediction value along with a corresponding prediction uncertainty value. This uncertainty value is typically generated by another output node of the model and processed to ensure its positive value. It reflects the model's self-assessment of the reliability of the prediction result when processing the current specific input data. After receiving these two values, the system calculates based on a predefined expansion coefficient. Specifically, the prediction uncertainty is multiplied by this coefficient to obtain the expansion amount. Then, using the cumulative distance prediction value at the breakpoint as the center point of the interval, the expansion amount is subtracted and added in the directions of distance decrease and distance increase, respectively, to calculate the lower and upper limits of the distance interval, i.e., the breakpoint-corrected distance interval, which expresses the range of physical distances where the breakpoint is more likely to exist, considering the model's own estimation errors. This method transforms the localization result from a point that may be biased due to various errors into an interval that encompasses uncertainty, improving the fault tolerance and practicality of the result in real-world applications.

[0168] S23. Calculate the breakpoint ranging confidence level corresponding to the breakpoint correction distance interval based on the prediction uncertainty.

[0169] Specifically, there exists a pre-defined, monotonic mathematical mapping function that uses prediction uncertainty as the primary input variable. This function is designed to output a higher confidence value when the prediction uncertainty is small and a lower confidence value when the prediction uncertainty is large. A common implementation is to use an exponential decay function, where the prediction uncertainty is the negative exponent. The system substitutes the prediction uncertainty value into this pre-defined function for calculation, and the resulting output value is the breakpoint distance confidence score. This confidence score is between 0 and 1 or expressed as a percentage, intuitively indicating the overall reliability of the model's prediction and the resulting correction distance interval. Through this step, the positioning model not only provides distance information but also provides a self-evaluation of the quality of this information, allowing subsequent decisions to comprehensively consider the predicted location and the reliability of the prediction, such as prioritizing positioning results with high confidence.

[0170] For example, the preprocessed OTDR curve is denoted as The measurement parameters corresponding one-to-one with the preprocessed OTDR curve are recorded as follows: ,in Indicates sampling interval at a uniform distance Discrete OTDR amplitude sequence, Indicates refractive index, Indicates pulse width. Indicates the range, Indicates the sampling interval;

[0171] Then and A common input localization model, wherein the localization model is composed of a self-supervised feature extraction network. Location network The localization model consists of a differentiable OTDR forward computation layer, which performs forward computation during inference using the same distance axis processing method as during training. The output of the localization model is the cumulative distance prediction value of the breakpoint starting from the OTDR test end. and the prediction uncertainty of the cumulative distance prediction value of the breakpoint. ,in This represents the cumulative distance from the breakpoint, in meters. The uncertainty is expressed in the form of the standard deviation of the cumulative distance from the breakpoint, with the unit being meters. The distance uncertainty output node of the positioning network is obtained by activating it via softplus to ensure... And avoid negative uncertainty;

[0172] Then based on and Determine the breakpoint correction distance interval and generate the breakpoint ranging confidence score. The breakpoint correction distance interval is denoted as... The confidence level of the breakpoint ranging is denoted as And calculate using the following formula:

[0173] ;

[0174] in Indicates the breakpoint correction distance range. This represents the cumulative distance prediction value of the breakpoints output by the localization model. This represents the prediction uncertainty output by the positioning model. This represents the interval expansion coefficient and has a value of 2. Indicates the range and is related to the measurement parameters The ranges are consistent. This represents the confidence level of the breakpoint ranging and its value range is [value range missing]. , This represents the confidence scaling factor, with a value of 200m. This represents an exponential function with the natural constant as its base. and These represent operators that take the larger and smaller values, respectively. and Cut the breakpoint correction distance interval to Within the range to ensure that the interval is consistent with the effective ranging range of the OTDR;

[0175] Correct the breakpoint distance range The endpoints of the interval are sampled at a uniform distance. Perform alignment and output the aligned breakpoint correction distance interval and the breakpoint ranging confidence level corresponding to that interval. .

[0176] S3. Perform interval intersection calculation between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point.

[0177] Specifically, the cumulative distance range is a distance interval calculated for each optical cable segment in the optical cable route. Its lower limit is the cumulative distance from the starting point of the optical cable segment to the OTDR test end, and its upper limit is the cumulative distance from the ending point to the OTDR test end.

[0178] Intersection calculation is a set operation used to determine whether two numerical intervals overlap. In this context, it determines whether the possible distance interval of the breakpoint intersects with the actual distance interval occupied by a certain optical cable segment.

[0179] Candidate optical cable segments refer to optical cable segments whose cumulative distance range intersects with the breakpoint correction distance range, meaning that the breakpoint is considered to be likely located within these optical cable segments.

[0180] The implementation of this step relies on pre-constructed optical cable resource topology data. First, based on the known optical cable routing structure, starting from the OTDR test end, the length of each optical cable segment is sequentially accumulated along the fiber transmission path to calculate the cumulative distance between the start and end points of each segment, forming the cumulative distance range for each segment. Then, the breakpoint correction distance interval output from step S2 is compared sequentially with the cumulative distance range of each segment. The overlap between the lower and upper limits of the interval is determined. Optical cable segments that overlap with the breakpoint correction distance interval are filtered out and marked as candidate segments. This step completes the mapping from the abstract distance values ​​predicted by the model to specific physical optical cable segments, narrowing the fault search scope from the entire link to one or more specific optical cable segments, providing structured constraints for subsequent precise location.

[0181] In some embodiments, before calculating the intersection of the breakpoint correction distance interval with the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end, the method further includes:

[0182] Based on the routing data of the optical cable to be located, obtain all point data and optical cable segment data on the optical cable route. The point data includes point identifier, point type and location information. The optical cable segment data includes optical cable segment identifier, optical cable segment length and the identifiers of the two points connected to the optical cable segment.

[0183] Based on the location data, corresponding location entities are created in the knowledge graph, and each location entity is associated with its location information and location type attribute.

[0184] Based on the optical cable segment data, a corresponding optical cable segment entity is created in the knowledge graph, and each optical cable segment entity is associated with its optical cable segment length attribute.

[0185] Based on the location identifiers connected to the optical cable segment, a connection relationship between the optical cable segment entity and the location entities at both ends is established in the knowledge graph;

[0186] Based on the sequential relationship of the optical cable routes to be located, a sequential relationship is established for adjacent optical cable segment entities in the knowledge graph, resulting in a constructed optical cable resource knowledge graph.

[0187] Specifically, the point data and cable segment data on the route of the optical cable to be located are extracted from the optical cable resource management system according to the optical cable identifier to be located, and a structured resource dataset is formed. The point data includes the point identifier corresponding to the OTDR test end, the equipment room point identifier, the optical distribution box point identifier, the junction box point identifier, the pipe manhole point identifier, the pole line point identifier, and the location information corresponding to each point identifier. The location information is expressed in the WGS-84 coordinate system latitude and longitude and consists of longitude and latitude. The optical cable segment data includes the optical cable segment identifier, the optical cable segment connection relationship between adjacent points, the optical cable segment length, and the route number used to characterize the sequential relationship of the optical cable route to be located. The unit of the optical cable segment length is meters and is based on the completed length field of the resource system. The route number is a positive integer that increases in the transmission direction starting from the OTDR test end.

[0188] Then, consistency verification and cleaning are performed on the structured resource dataset. Specifically, the uniqueness of each point identifier is verified, the uniqueness of each optical cable segment identifier is verified, the connection relationship of each optical cable segment can be found in the point data with corresponding two endpoint identifiers, the length of the optical cable segment is verified to be a value greater than 0, and the longitude and latitude of the location information are verified to be valid values ​​and not empty. For data that fails the verification, the resource management system is checked back according to the point identifier or optical cable segment identifier to fill in the fields until the verification passes.

[0189] Next, an attribute graph model is used to construct a knowledge graph of optical cable resources and represent it as follows: ,in This represents a knowledge graph of optical fiber cable resources. This represents a set of entities, consisting of both point entities and fiber optic cable segment entities. It represents a set of relationships, consisting of the connection relationships between point entities and optical cable segment entities, as well as the sequential relationships between adjacent optical cable segments. The entity type mapping is used to map each entity to either a point entity type or a fiber optic segment entity type. Attribute mapping is used to bind entities and relationships to their attribute key-value pairs;

[0190] Based on the location identifier Each point entity is created and its attributes are written. The point entity attributes include point identifier, point type and location information. The point type is one of six categories: "OTDR test terminal, equipment room, optical distribution box, junction box, pipe manhole, pole line" and is determined by the category field in the point data. The location information consists of longitude and latitude and is used as a searchable attribute of the point entity.

[0191] Based on the optical cable segment identification Each optical cable segment entity is created and its attributes are written. The optical cable segment entity attributes include the optical cable segment identifier and the optical cable segment length.

[0192] Based on the aforementioned optical cable segment connection relationship For each optical cable segment entity, two directed connections are established with the entities at both ends, and the relationship type is uniformly recorded as "connection". Each "connection" relationship carries at least the endpoint role attribute to identify the endpoint role of the point in the optical cable segment, thereby ensuring that the entry point and exit point can be uniquely determined according to the routing direction.

[0193] Then, based on the sequential relationship of the optical cable routes to be located, In the diagram, a directed sequential relationship is established between adjacent optical cable segments and the relationship type is uniformly recorded as "sequence". Each "sequence" relationship carries a route sequence number attribute and satisfies the requirement of forming a single chain structure by increasing the route sequence number from the OTDR test end. At the same time, the point entity corresponding to the OTDR test end is marked as the starting entity in the diagram so that subsequent route traversal and cumulative distance calculation can be performed from the starting entity along the "connection" and "sequence" relationships.

[0194] The The data is persistently stored in a graph database, and unique indexes are created for point identifiers and optical cable segment identifiers, and sorting indexes are created for route sequence numbers. This results in an optical cable resource knowledge graph that can be used for segment-level matching and point-level landing point calculation in subsequent steps S6 and S7.

[0195] In some embodiments, S3 involves performing an interval intersection calculation between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine candidate optical cable segments, including:

[0196] S31. Determine the optical cable segment sequence starting from the OTDR test end based on the routing order relationship in the optical cable resource knowledge graph;

[0197] Specifically, the routing order relationship is a directed connection relationship stored in the optical cable resource knowledge graph, which clarifies the order in which optical cable segments are connected on the physical path.

[0198] This step extracts ordered structural information from the resource topology data. The system queries the constructed optical cable resource knowledge graph, starting from a specific point entity representing the OTDR test end, and traverses along the edges defined in the graph that indicate the route direction. During the traversal, each optical cable segment entity is visited and recorded sequentially, forming an ordered list of optical cable segments, i.e., an optical cable segment sequence, according to their order of appearance in the actual physical route. This sequence completely describes the physical path taken by the optical signal originating from the test end.

[0199] S32. Accumulate the lengths of each optical cable segment sequentially to obtain the cumulative distance range of each optical cable segment relative to the OTDR test end;

[0200] Specifically, the cumulative distance range is a numerical interval calculated for each optical cable segment in the sequence. Its lower limit is equal to the sum of the lengths of all preceding optical cable segments traversed from the OTDR test end to the starting point of the optical cable segment, and its upper limit is equal to the lower limit plus the length of the current optical cable segment.

[0201] This step assigns a precise distance metric to the physical routes. The system processes the fiber optic segment sequence obtained in S31 sequentially. Starting from distance 0 (the test end location), the length of the first fiber optic segment in the sequence is read. The lower limit of the cumulative distance range for this segment is 0, and the upper limit is its length value. Next, for the second fiber optic segment in the sequence, the lower limit of its cumulative distance range is equal to the upper limit of the first fiber optic segment (i.e., the distance to the endpoint of the previous segment), and the upper limit is equal to the lower limit plus the length of the second fiber optic segment itself. This process continues iteratively, calculating the precise distances between the start and end points of each fiber optic segment relative to the test end, thus obtaining the cumulative distance range for each segment. This process transforms the abstract topological connectivity into a set of intervals with numerical distance information.

[0202] S33. Perform interval intersection judgment between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment, and determine the optical cable segments whose interval intersection result is not empty as candidate optical cable segments.

[0203] Specifically, interval intersection judgment is a logical operation used to determine whether two numerical intervals have a common part. An interval intersection result that is not empty means that the two intervals at least partially overlap, that is, the lower limit of one interval is less than or equal to the upper limit of the other interval, while its upper limit is greater than or equal to the lower limit of the other interval.

[0204] This step matches the distance information predicted by the model with the physical network structure. The system sequentially compares the breakpoint correction distance interval obtained in step S22 with the cumulative distance range of each optical cable segment calculated in step S32. During the comparison, it checks whether the lower limit of the breakpoint correction distance interval is less than or equal to the upper limit of the cumulative distance range of the optical cable segment, and whether the upper limit of the breakpoint correction distance interval is greater than or equal to the lower limit of the cumulative distance range of the optical cable segment. If both conditions are met, the two intervals are considered to intersect, meaning that the possible location of the breakpoint predicted by the model falls within the physical length range of the optical cable segment. All optical cable segments that pass this judgment are selected to form a candidate optical cable segment set. This step realizes the mapping from the distance domain to the physical network domain, locking the possible fault range into one or several specific optical cable segments.

[0205] For example, from the optical cable resource knowledge graph The system reads the corresponding point entity of the OTDR test terminal that has been marked as the starting entity, and performs a directed traversal of the optical cable segment entities along the "sequence" relationship to obtain the optical cable segment sequence that is consistent with the optical cable route to be located. and the point sequence that corresponds to the endpoint of the optical cable segment sequence. ,in Indicates the first in the route Optical cable segment physical and This indicates the optical cable segment number, starting from 1. and Represents the physical entity of the optical cable segment Point entities connected at both ends and For OTDR test endpoint bit entity, Indicates the total number of optical cable segments;

[0206] Then from each optical cable segment entity Read the fiber optic cable segment length from the attributes and record it as ,in Represents the physical entity of the optical cable segment The length of the optical cable segment is given in meters, and the cumulative distance range of each optical cable segment relative to the OTDR test end is calculated according to the routing order. ,in Represents the physical entity of the optical cable segment The cumulative distance from the starting point relative to the OTDR test terminal, in meters. Represents the physical entity of the optical cable segment The cumulative distance at the endpoint relative to the OTDR test end, in meters, and expressed as... As the starting condition and in sequence let and This creates a continuous, non-overlapping cumulative distance range across the optical cable segment sequence;

[0207] Next, read the breakpoint correction distance range. ,in This indicates the breakpoint correction distance range, with the unit being meters. This represents the cumulative distance from the left endpoint of the breakpoint correction distance interval. This represents the cumulative distance to the right endpoint of the breakpoint correction distance interval, and... and Sampling interval at a uniform distance Endpoint alignment is performed to ensure consistency with the distance benchmark used for calculating the cumulative distance range;

[0208] For each optical cable segment entity Performing interval intersection judgment to determine candidate optical cable segments will satisfy... Optical cable segment entity Candidate optical cable segments whose interval intersection results are not empty are identified, among which and Representing the physical entities of optical cable segments The cumulative distance range between the left and right endpoints, and These represent the left and right endpoints of the breakpoint correction distance interval, respectively, with symbols... This represents the logical AND operation and is used to satisfy two inequalities simultaneously, thus being equivalent to an interval... and interval There is an intersection;

[0209] Write the optical cable segment identifiers of all optical cable segment entities that are identified as candidate optical cable segments into the candidate optical cable segment identifier set and output the candidate optical cable segment identifier set.

[0210] S4. In each point corresponding to each candidate optical cable segment, the points whose cumulative distance falls within the breakpoint correction distance interval are taken as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point.

[0211] Specifically, candidate locations refer to those physical locations on the candidate optical cable segment whose cumulative distance value falls within the breakpoint correction distance range.

[0212] This step is carried out based on the established set of candidate fiber optic cable segments. For each candidate segment, all points located on that segment can be identified based on the connection information at both ends. The system needs to calculate the cumulative distance of each of these points relative to the OTDR test end, which is equal to the sum of the lengths of all fiber optic cable segments traversed from the test end along the route to that point. Subsequently, the cumulative distance value of each point is compared with the upper and lower bounds of the breakpoint correction distance interval output in step S2. Only points whose cumulative distance values ​​are greater than or equal to the lower bound of the interval and less than or equal to the upper bound are selected as candidate points. This step further focuses the search from possible fiber optic cable segments to more specific physical location nodes, clarifying the positioning target to a limited number of installation or splicing locations on the route with practical engineering significance, such as junction boxes or manholes.

[0213] S5. Calculate the point location confidence of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point and the breakpoint ranging confidence, and output a list of candidate points sorted by point location confidence.

[0214] Specifically, the point location confidence score is a score calculated for each candidate point to comprehensively reflect the probability that the point is the true breakpoint location. Its calculation usually combines the model's ranging confidence score and the proximity of the point to the model's predicted center.

[0215] This step calculates a quantified confidence score for each candidate point selected in step S4. The calculation considers two main factors: first, the breakpoint distance measurement confidence score output by the model in step S2, which represents the reliability of the overall distance prediction; and second, the deviation between the cumulative distance of the candidate point and the center point of the breakpoint correction distance interval. A smaller deviation generally means that the point matches the model's best prediction better. These two factors are combined using a predetermined mathematical formula (e.g., multiplying the distance measurement confidence score by a decay function based on distance deviation) to calculate a final point location confidence score for each candidate point. Finally, the system sorts all candidate points according to their point location confidence scores from highest to lowest and outputs the sorted list. This step, by prioritizing candidate points, provides clear action suggestions for on-site repair personnel, placing the most likely fault points first, which helps improve the efficiency of fault diagnosis and reduce the time spent on blind excavation and testing.

[0216] In some implementations, S5, calculating the point location confidence of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence, includes:

[0217] S51. For each candidate point, calculate the cumulative distance deviation based on the deviation between the cumulative distance of the candidate point and the center of the breakpoint correction distance interval;

[0218] Specifically, the cumulative distance deviation is a non-negative scalar value that represents the absolute value of the distance difference between the cumulative distance of the candidate point itself and the geometric center of the breakpoint correction distance interval.

[0219] This step aims to quantify the degree of fit between each candidate point and the model's best predicted location. First, the center point of the breakpoint correction distance interval is calculated; its value is equal to the arithmetic mean of the lower and upper limits of the interval. This center point represents the ideal distance location where the model-predicted breakpoint is most likely to exist. Then, for each candidate point, its cumulative distance value, determined in step S4, is read. The absolute value of the difference between this cumulative distance value and the interval center point value is calculated. The absolute value obtained is the cumulative distance deviation of the candidate point. The smaller this deviation value, the closer the point is to the model's center prediction in the distance dimension; from a distance perspective, its probability of being a true breakpoint is relatively higher.

[0220] S52. Based on the cumulative distance deviation and the breakpoint ranging confidence, calculate the point location confidence of the candidate point using an exponential decay function.

[0221] Specifically, the exponential decay function is an exponential function with the natural constant e as its base, and its form is exp(-x / s), where x is the input variable and s is the scale parameter. The function value decreases monotonically as x increases.

[0222] The point location reliability is a final score that combines the overall predictive reliability of the model with the accuracy of the point location, and is used to prioritize candidate points.

[0223] This step is responsible for synthesizing the final confidence score. It has two inputs: the cumulative distance deviation calculated in step S51 and the breakpoint ranging confidence score obtained in step S23. The calculation follows a predetermined formula, which uses the breakpoint ranging confidence score as the base factor and multiplies it by a decay factor calculated from the cumulative distance deviation using an exponential decay function. Specifically, the cumulative distance deviation value is divided by a preset scaling factor, which is then used as the negative exponent input to the exponential decay function. This function works as follows: when the distance deviation is 0, the decay factor is 1, without weakening the base confidence score; as the distance deviation increases, the decay factor gradually decreases from 1 to 0, discounting the base confidence score. Finally, the breakpoint ranging confidence score is multiplied by this decay factor, and the result is the point location confidence score for that candidate point. The physical meaning of this calculation method is that the final confidence score is determined by the overall reliability of the model's prediction (ranging confidence score) and the proximity of the point to the prediction center (distance deviation). The more reliable the model as a whole, and the closer the points are to the prediction center, the higher their final point location reliability score. This score provides a comparable ranking criterion for all candidate points.

[0224] For example, the candidate optical cable segment identifier set is read and denoted as ,in A set representing candidate optical cable segment identifiers;

[0225] Then in the optical fiber resource knowledge graph For each candidate optical cable segment identifier, query the point entities that have a "connection" relationship with it and extract the point identifiers. Form the union of the point identifiers corresponding to the candidate optical cable segment and remove duplicates to obtain the point identifier set. ,in This represents the set of point identifiers that participate in the point-level landing point calculation;

[0226] Next, according to the order of the optical cable routes to be located... For each point marker in the calculation, its cumulative distance relative to the OTDR test terminal is calculated and recorded as . ,in Indicates the location identifier and Indicates location identifier The cumulative distance, measured in meters, is relative to the OTDR test endpoint. The calculation method for the cumulative distance is based on the knowledge graph. Starting from the OTDR test endpoint entity, a unique sequence of points is determined along a "sequence" relationship. The cumulative distance for each point is obtained by sequentially adding the corresponding optical cable segment lengths between adjacent points in the sequence. Sampling interval at a uniform distance Perform distance alignment to ensure consistency with the distance reference in the breakpoint correction distance range;

[0227] Read breakpoint correction distance range confidence level of breakpoint ranging ,in This indicates the breakpoint correction distance range, with the unit being meters. This indicates the cumulative distance from the left endpoint of the breakpoint correction interval. This indicates the cumulative distance from the right endpoint of the breakpoint correction distance interval. This represents the confidence level of the breakpoint ranging and its value range is [value range missing]. And label each point. The execution interval includes the judgment if and only if satisfy Point marking Identify it as a candidate point identifier and add it to the candidate point identifier set. ,in Represents the set of candidate point identifiers;

[0228] Then, for each candidate point, an identifier is created. Calculate the cumulative distance deviation and generate point location confidence. The cumulative distance deviation is used to determine the interval center by the interval boundary of the breakpoint correction distance interval and measures the degree of deviation of the cumulative distance of the point from the interval center. The point location confidence is combined with the breakpoint distance measurement confidence and the cumulative distance deviation to form a sortable score and is calculated by the following formula:

[0229] ;

[0230] in This indicates the cumulative distance from the center of the breakpoint correction interval. Indicates candidate point identifier The cumulative distance deviation, in meters. This represents the absolute value operation. Indicates candidate point identifier The point location reliability and the range of values ​​are This represents an exponential function with the natural constant as its base. This represents the deviation scaling factor, with a value of 100m, used to control the rate at which the cumulative distance deviation decays the point location confidence.

[0231] Reliability by point location From largest to smallest The candidate point identifiers are sorted, and when the point location confidence is the same, the cumulative distance deviation is used as the criterion. Sort the candidate points from smallest to largest to ensure stable output. Output the sorted candidate point identifiers and their corresponding point location confidence scores.

[0232] Example 2

[0233] Please see Figure 2 This invention provides a fiber optic cable breakpoint location device based on deep learning and knowledge fusion, comprising:

[0234] The feature extraction module is used to extract features from the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned, and obtain feature vectors.

[0235] The positioning module is used to input the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label.

[0236] The interval matching module is used to perform interval intersection calculations between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point.

[0237] The candidate point module is used to select points whose cumulative distance falls within the breakpoint correction distance range for each point corresponding to each candidate optical cable segment as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point.

[0238] The point location reliability module is used to calculate the point location reliability of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence, and outputs a list of candidate points sorted by point location reliability.

[0239] It should be noted that each module and unit in the optical cable breakage location device based on deep learning and knowledge fusion in this embodiment corresponds one-to-one with each step in the optical cable breakage location method based on deep learning and knowledge fusion in the aforementioned embodiment. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned optical cable breakage location method based on deep learning and knowledge fusion, and will not be repeated here.

[0240] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0241] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for locating optical cable breaks based on deep learning and knowledge fusion, characterized in that, include: Feature vectors are obtained by extracting features from the OTDR curves and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned. The feature vector and the measurement parameters are input into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label. The intersection calculation is performed between the breakpoint correction distance range and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point. In each point corresponding to each candidate optical cable segment, the points whose cumulative distance falls within the breakpoint correction distance interval are taken as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point. The point location confidence of each candidate point is calculated based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence. The candidate point list is then output in order of point location confidence.

2. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 1, characterized in that, The OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned are subjected to feature extraction to obtain a feature vector, including: Obtain the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned; Convert the time axis of the OTDR curve to the distance axis based on the refractive index in the measurement parameters; The distance axis is resampled according to a uniform distance sampling interval to obtain a preprocessed OTDR curve; Based on the preprocessed OTDR curve, feature vectors are extracted through a pre-trained feature extraction network, which is obtained by training a deep neural network through self-supervised contrastive learning.

3. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 2, characterized in that, The feature extraction network was trained in the following manner: Construct positive sample pairs based on historical preprocessed OTDR curves of the same fiber optic link under different measurement parameters; Negative sample pairs are constructed based on the historical preprocessed OTDR curves of different fiber optic links; The deep neural network is trained using the positive and negative sample pairs and a contrastive learning loss to obtain a feature extraction network that is robust to changes in measurement parameters.

4. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 1, characterized in that, The localization model is trained in the following way: The feature vector extracted from the historical preprocessed OTDR curve and the corresponding measurement parameters are input into the three-layer fully connected structure of the localization network to be trained. The network outputs the cumulative distance prediction value of the breakpoint, the attenuation coefficient prediction value, and the reflection intensity prediction value. The attenuation coefficient prediction value is used to describe the attenuation coefficient of back Rayleigh scattering with distance, and the reflection intensity prediction value is used to describe the amplitude intensity of the breakpoint reflection event. The measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity are input into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve. The physical consistency loss is calculated based on the difference between the predicted OTDR curve and the historical preprocessed OTDR curve. The location loss is calculated based on the difference between the predicted cumulative distance of the breakpoint and the cumulative distance label of the breakpoint obtained from the repair record; The positioning network parameters are updated based on the joint loss of the physical consistency loss and the positioning loss to obtain the positioning model.

5. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 4, characterized in that, The step of inputting the measurement parameters, the predicted cumulative distance of the breakpoint, the predicted attenuation coefficient, and the predicted reflection intensity into the differentiable forward computation layer of the positioning network to generate a predicted OTDR curve includes: A scattering baseline term is constructed based on the back Rayleigh scattering and attenuation model and the predicted attenuation coefficient. Based on the predicted cumulative distance of the breakpoint and the predicted reflection intensity, a reflection event item is constructed at the corresponding distance position; The reflection event term is convolved with the instrument impulse response function determined by the pulse width in the measurement parameters to obtain the convolution result; The convolution result is added to the scattering baseline term to obtain the predicted OTDR curve.

6. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 1, characterized in that, The step of inputting the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, starting from the OTDR test end, includes: The feature vector and the measurement parameters are input into a preset positioning model, and the cumulative distance prediction value of the breakpoint is output. Based on the predicted cumulative distance of the breakpoint and its prediction uncertainty, a distance interval centered on the predicted cumulative distance of the output breakpoint and extended to both sides by a certain multiple of uncertainty is determined as the breakpoint correction distance interval. The breakpoint ranging confidence level corresponding to the breakpoint correction distance interval is calculated based on the prediction uncertainty.

7. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 1, characterized in that, Before performing the interval intersection calculation between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end, the method further includes: Based on the routing data of the optical cable to be located, obtain all point data and optical cable segment data on the optical cable route. The point data includes point identifier, point type and location information. The optical cable segment data includes optical cable segment identifier, optical cable segment length and the identifiers of the two points connected to the optical cable segment. Based on the location data, corresponding location entities are created in the knowledge graph, and each location entity is associated with its location information and location type attribute. Based on the optical cable segment data, a corresponding optical cable segment entity is created in the knowledge graph, and each optical cable segment entity is associated with its optical cable segment length attribute. Based on the location identifiers connected to the optical cable segment, a connection relationship between the optical cable segment entity and the location entities at both ends is established in the knowledge graph; Based on the sequential relationship of the optical cable routes to be located, a sequential relationship is established for adjacent optical cable segment entities in the knowledge graph, resulting in a constructed optical cable resource knowledge graph.

8. A method for locating optical cable breaks based on deep learning and knowledge fusion according to claim 1 or 7, characterized in that, The step of calculating the intersection of the breakpoint correction distance interval with the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment includes: Based on the routing order relationship in the optical cable resource knowledge graph, determine the optical cable segment sequence starting from the OTDR test end; The cumulative distance range of each optical cable segment relative to the OTDR test end is obtained by sequentially accumulating the lengths of each optical cable segment. The distance interval for breakpoint correction is intersected with the cumulative distance range of each optical cable segment. Optical cable segments whose interval intersection results are not empty are identified as candidate optical cable segments.

9. The optical cable breakpoint localization method based on deep learning and knowledge fusion according to claim 1, characterized in that, The calculation of the point location confidence of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence level includes: For each candidate point, the cumulative distance deviation is calculated based on the deviation between the cumulative distance of the candidate point and the center of the breakpoint correction distance interval. Based on the cumulative distance deviation and the breakpoint ranging confidence level, the point location confidence level of the candidate point is calculated using an exponential decay function.

10. A fiber optic cable break point location device based on deep learning and knowledge fusion, characterized in that, include: The feature extraction module is used to extract features from the OTDR curve and measurement parameters measured by the OTDR at the OTDR test end of the optical cable to be positioned, and obtain feature vectors. The positioning module is used to input the feature vector and the measurement parameters into a preset positioning model to obtain the breakpoint correction distance interval and the corresponding breakpoint ranging confidence level, with the OTDR test end as the starting point. The positioning model is obtained by supervised training with the feature vector and measurement parameters corresponding to the historical OTDR curve as input and the cumulative breakpoint distance obtained from the repair record as the label. The interval matching module is used to perform interval intersection calculations between the breakpoint correction distance interval and the cumulative distance range of each optical cable segment in the optical cable to be located relative to the OTDR test end to determine the candidate optical cable segment. The start and end points of the optical cable segment are defined by the point, and adjacent optical cable segments are connected through a shared point. The candidate point module is used to select points whose cumulative distance falls within the breakpoint correction distance range for each point corresponding to each candidate optical cable segment as candidate points. The cumulative distance refers to the total physical distance accumulated from the OTDR test end along the optical cable route to the point. The point location reliability module is used to calculate the point location reliability of each candidate point based on the breakpoint correction distance interval, the cumulative distance of each candidate point, and the breakpoint ranging confidence, and outputs a list of candidate points sorted by point location reliability.

Images