A migration learning method and system suitable for optical fiber abnormal disturbance identification

Abstract

This invention discloses a transfer learning method and system for identifying optical fiber anomalies, relating to the field of optical fiber vibration monitoring technology. The method includes acquiring backscattered Rayleigh signals, constructing a multi-channel feature image that integrates time, frequency, and spatial location information, introducing a residual neural network containing residual units and a classification layer, obtaining an attention-enhanced feature map through channel weighting, and identifying the disturbance under test through a classification loss term. This invention improves the accuracy and robustness of optical fiber anomaly identification through time-frequency spatial feature fusion and attention channel weighting, enhancing the model's transfer learning ability and generalization performance across datasets and event categories under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of fiber optic vibration monitoring technology, and in particular to a transfer learning method and system suitable for identifying abnormal disturbances in optical fibers. Background Technology

[0002] Φ-OTDR distributed vibration monitoring, by injecting probe pulses into sensing fibers and acquiring backscattered Rayleigh signals, enables the location and identification of disturbance events over long distances. It has accumulated significant engineering applications and research experience in areas such as perimeter security, oil and gas pipeline and power cable condition monitoring. Regarding disturbance event identification, existing technologies generally extract distinguishable features from the acquired spatiotemporal disturbance data. Early studies often relied on time-domain waveform statistics, frequency-domain energy distribution, or simple time-frequency transformation results to construct features, which were then judged by traditional classifiers. Subsequent research has converted disturbance data into time-frequency maps or two-dimensional feature maps and input them into deep models such as convolutional neural networks and residual networks to reduce manual feature design and improve recognition performance under fixed operating conditions. Some literature has also attempted to introduce transfer learning or domain adaptation concepts to mitigate distribution differences between different data sources and under different scenario conditions.

[0003] However, in engineering sites, the data collected by the Φ-OTDR system is often affected by factors such as the monitoring environment, the composition of disturbance types, and the sample size. The distribution and characteristics of disturbance events vary significantly under different acquisition scenarios. Especially when the number of samples is limited or new disturbance types are introduced, models trained on a single dataset are prone to overfitting or degraded generalization performance. Existing transfer learning and domain adaptation methods are mostly built around general images, and their feature alignment and constraint forms are difficult to directly match the characteristics of time, frequency, and spatial location coupling in Φ-OTDR data. Existing solutions mainly treat the input as a single time-frequency image, lacking explicit expression and utilization of spatial location information, and also lacking fine-grained noise reduction and enhancement mechanisms for channel-level feature differences. As a result, it is difficult for the model to simultaneously achieve recognition robustness and cross-condition transfer performance under small sample conditions and new disturbance event scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a transfer learning method suitable for identifying fiber optic anomalies and disturbances, which solves the problems of low accuracy and poor generalization ability in identifying fiber optic anomalies and disturbances under different datasets, small sample conditions, and scenarios with varying disturbance event categories.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a transfer learning method suitable for identifying optical fiber anomalies, comprising: detecting disturbances in the optical fiber under test, acquiring backscattered Rayleigh signals, and forming two-dimensional disturbance data in temporal and spatial order.

[0007] The time and space two-dimensional perturbation data are digitally demodulated, time-frequency analyzed, and spatial coordinate normalized to construct a multi-channel feature image representing time information, frequency information, and spatial location information.

[0008] The multi-channel feature image corresponding to the fiber disturbance data to be determined is input into the residual neural network model trained with the classification loss function, and the final disturbance determination result is output.

[0009] The residual neural network model includes an input layer, residual units, and a classification layer. An attention-enhanced feature map is introduced into the residual neural network model. The attention-enhanced feature map is used as the update feature map of the first-level residual unit, and the updated feature map is input into the subsequent residual units.

[0010] As a preferred embodiment of the transfer learning method for identifying optical fiber anomalies described in this invention, the optical fiber under test is probed, and the backscattered Rayleigh signal is sampled at a preset sampling period after each probe, and arranged in temporal and spatial order to form two-dimensional perturbation data in time and space.

[0011] As a preferred embodiment of the transfer learning method for identifying optical fiber anomalies described in this invention, the temporal and spatial two-dimensional disturbance data are digitally orthogonally demodulated and phase recovered to obtain phase timing signals at each spatial sampling position along the optical fiber.

[0012] Time-frequency analysis is performed on the phase-series signal to obtain a two-dimensional energy distribution in terms of time and frequency, and the energy values ​​in different frequency ranges are mapped into time-frequency feature maps according to preset rules.

[0013] The coordinates of each spatial sampling position in the fiber along its length are normalized according to the total number of spatial sampling points to obtain normalized spatial coordinates, which are then arranged into a single-channel spatial position matrix according to the order of the spatial sampling positions on the fiber.

[0014] The multi-channel feature image is obtained by analyzing time-frequency features and spatial features through two channels, and then stitching the obtained time-frequency feature map and spatial location matrix together along the channel dimension.

[0015] As a preferred embodiment of the transfer learning method for identifying fiber optic anomalies described in this invention, the residual neural network model includes: receiving multi-channel feature maps at the input layer, sequentially inputting the multi-channel feature maps into each level of residual units, performing convolution operations, nonlinear mapping, and identity mapping superposition on the updated feature maps output by the previous level of residual units, and using the updated feature maps output by the last level of residual units as input to the classification layer, which generates a predicted probability vector for the category of the disturbance event.

[0016] The residual unit includes a convolution operation structure and an activation structure for nonlinear mapping.

[0017] The convolution operation structure includes convolution kernel parameters and convolution bias parameters.

[0018] The classification layer includes a fully connected operation structure; the fully connected operation structure includes fully connected weight parameters and fully connected bias parameters.

[0019] The convolution kernel parameters, convolution bias parameters, fully connected weight parameters, and fully connected bias parameters constitute the internal parameters of the residual neural network model.

[0020] As a preferred embodiment of the transfer learning method for identifying fiber optic anomalies described in this invention, the attention-enhanced feature map includes an updated feature map output by the first-level residual unit as an intermediate feature map, wherein the intermediate feature map is an updated feature map containing the feature information of the multi-channel feature image.

[0021] Global pooling is performed on the intermediate feature map to obtain a comprehensive channel weight. Based on the comprehensive channel weight, the intermediate feature map is weighted channel by channel to obtain an attention-enhanced feature map.

[0022] Global pooling is performed on the temporal features of the intermediate feature map to obtain a temporal direction description vector; global pooling is performed on the frequency direction features of the intermediate feature map to obtain a frequency direction description vector; and the normalized spatial coordinates of each spatial sampling point are processed by one-dimensional fully connected mapping and nonlinear activation to obtain a spatial direction description vector.

[0023] Perform a fully connected operation on the time direction description vector, frequency direction description vector, and spatial direction description vector respectively; and map the time direction description vector, frequency direction description vector, and spatial direction description vector into a time weight vector, a frequency weight vector, and a spatial weight vector according to the preset mapping direction of each channel in three-dimensional space.

[0024] For any channel index Corresponding to part C of the multi-channel feature image, the time weight vector at the channel index Component values ​​at the location Frequency weight vector in channel index Component values ​​at the location and spatial weight vector in channel index Component values ​​at the location ,Will , and The product of the component values ​​is input into the Sigmoid activation function for nonlinear mapping, resulting in the component values ​​of the comprehensive channel weight vector in part C.

[0025] in, This indicates the composite channel weight vector at the channel index. The component at the location; This represents the Sigmoid activation function; Represents the time weight vector, in the channel index The component at the location; Represents the frequency weight vector, in the channel index The component at the location; Represents the spatial weight vector, in the channel index The component at that location.

[0026] Using the comprehensive channel weight vector in the channel index The components at the intermediate feature map in the channel index The attention-enhanced feature map is obtained by weighting the time and frequency feature values ​​at a given location.

[0027] in, Indicates the first The first channel, the first The time point, the first At each frequency point, the feature value of the attention-enhanced feature map; This indicates the composite channel weight vector at the channel index. The component at the location; Indicates the first The first channel, the first The time point, the first The eigenvalues ​​of the intermediate feature map at each frequency point.

[0028] As a preferred embodiment of the transfer learning method for identifying fiber optic anomalies described in this invention, the multi-channel feature image set pre-selected as training samples is defined as the source domain.

[0029] The multi-channel feature images in the source domain are classified according to the preset perturbation event category information. A perturbation event category label is assigned to each multi-channel feature image, and the classification loss term is constructed based on the perturbation event category label.

[0030] in, Represents the classification loss term; Index representing the multi-channel feature image; Indicates the index of disturbance event categories; Indicates the label, when the source domain is... The perturbation event category of the multi-channel feature image is the first When class, When the source domain is The perturbation event category of the multi-channel feature image is not the first. When class, ; Indicates the first The multi-channel feature image belongs to the first... Predicted probability of perturbation event categories.

[0031] The residual neural network model is trained by minimizing the classification loss function. During the training process, the gradient of each internal parameter in the current training iteration is calculated using the chain rule, and the internal parameters are iteratively updated according to the preset learning rate coefficient to obtain the trained residual neural network model.

[0032] As a preferred embodiment of the transfer learning method for identifying optical fiber anomalies described in this invention, the disturbance determination result includes: inputting the multi-channel feature image corresponding to the optical fiber disturbance data to be determined into a residual neural network model trained with classification loss, and outputting the predicted probability vector of each disturbance event category by the classification layer of the model; selecting the disturbance event category with the highest predicted probability from the predicted probability vector, and using the disturbance event category as the disturbance determination result.

[0033] Secondly, the present invention provides a transfer learning system suitable for identifying optical fiber anomalies and disturbances, comprising a disturbance data acquisition module: performing disturbance detection on the optical fiber under test, acquiring backscattered Rayleigh signals, and forming two-dimensional disturbance data in time and space order.

[0034] Multi-channel feature map construction module: performs digital demodulation, time-frequency analysis and spatial coordinate normalization on the time and space two-dimensional perturbation data to construct a multi-channel feature image representing time information, frequency information and spatial location information.

[0035] Judgment Module: Input the multi-channel feature image corresponding to the fiber disturbance data to be judged into the residual neural network model trained by the classification loss function, and output the final disturbance judgment result.

[0036] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the transfer learning method for identifying fiber optic anomalies as described in the first aspect of the present invention.

[0037] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the transfer learning method for identifying fiber optic anomalies as described in the first aspect of the present invention.

[0038] The beneficial effects of this invention are as follows: By acquiring backscattered Rayleigh signals and performing digital demodulation, time-frequency analysis, and spatial coordinate normalization on two-dimensional temporal and spatial disturbance data, a multi-channel feature image is constructed that simultaneously represents temporal, frequency, and spatial information. This expands the feature representation of fiber optic disturbance events from traditional single time-domain or frequency-domain features to an integrated representation of time, frequency, and space, thereby improving the ability of features to characterize complex operating conditions. Based on this, a residual neural network model incorporating residual units and attention channel weighting mechanisms is introduced. By calculating the comprehensive channel weights using time-direction description vectors, frequency-direction description vectors, and spatial-direction description vectors, the intermediate feature map is selectively enhanced. This helps to highlight key features related to the disturbance category and suppress noise and redundant information, improving the accuracy and robustness of fiber optic abnormal disturbance identification. Simultaneously, based on the model parameters trained on the source dataset, transfer learning training is performed on the target dataset, significantly improving the model's transferability and generalization performance under small sample conditions and scenarios with changing disturbance event categories. This reduces the workload of re-collecting and labeling large amounts of training data, facilitating rapid deployment in practical engineering. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a transfer learning method applicable to fiber optic anomaly perturbation identification. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] Reference Figure 1 This is one embodiment of the present invention, which provides a transfer learning method suitable for identifying optical fiber anomalies and disturbances, comprising the following steps: S1: Perturbation detection is performed on the fiber under test, back Rayleigh scattering signals are collected, and two-dimensional perturbation data in time and space are formed in time and space order.

[0045] The fiber under test is repeatedly probed, and the backscattered Rayleigh signal is sampled according to a preset sampling period after each probe. The samples are then arranged in time and space order to form two-dimensional perturbation data in time and space.

[0046] S2: Perform digital demodulation, time-frequency analysis, and spatial coordinate normalization on the two-dimensional time and space disturbance data to construct a multi-channel feature image representing time information, frequency information, and spatial location information.

[0047] The time and space two-dimensional perturbation data are digitally orthogonally demodulated and phase recovered to obtain the phase timing signal at each spatial sampling position along the optical fiber.

[0048] Time-frequency analysis is performed on the phase-series signal to obtain a two-dimensional energy distribution in terms of time and frequency, and the energy values ​​in different frequency ranges are mapped into time-frequency feature maps according to preset rules.

[0049] The coordinates of each spatial sampling position in the fiber along its length are normalized according to the total number of spatial sampling points to obtain normalized spatial coordinates, which are then arranged into a single-channel spatial position matrix according to the order of the spatial sampling positions on the fiber.

[0050] in, Represents normalized spatial coordinates; Indicates the spatial sampling index; This represents the total number of spatial sampling points.

[0051] The multi-channel feature image is obtained by analyzing time-frequency features and spatial features through two channels, and then stitching the obtained time-frequency feature map and spatial location matrix together along the channel dimension.

[0052] S3: Input the multi-channel feature image corresponding to the fiber disturbance data to be determined into the residual neural network model trained by the classification loss function, and output the final disturbance determination result.

[0053] The residual neural network model adopts a structure consisting of an input layer, residual units, and a classification layer connected sequentially. It constructs temporal, frequency, and spatial descriptive vectors on the intermediate feature map output by the first residual unit, adapting to the three-dimensional organization of channels, time, and frequency in multi-channel feature images. While maintaining a lightweight structure, it introduces globally statistical features with physical meaning. The input layer uniformly maps the multi-channel feature image into a three-dimensional feature map arranged in the channel, time, and frequency directions. The residual units stably extract deep features through convolution operations and nonlinear mapping superposition with identity mapping. Based on this, the intermediate features... Global pooling is performed along the frequency and time directions to obtain a time-direction descriptor vector representing the evolution trend of the disturbance over time and a frequency-direction descriptor vector representing the average energy distribution of each frequency band. The normalized spatial coordinates are then mapped through a one-dimensional fully connected layer and nonlinear activation to obtain a spatial direction descriptor vector. This allows the channel weighting process to comprehensively utilize time, frequency, and spatial location information when generating channel weights. While reducing the feature dimension and model parameter size, it focuses on enhancing feature channels related to the disturbance event, suppressing background channels and noise channels, thereby providing a more discriminative feature representation for the classification layer to determine the category of the disturbance event.

[0054] The residual neural network model includes an input layer receiving multi-channel feature maps, which are then sequentially input into each level of residual units. Each level of residual unit performs convolution operations, nonlinear mapping, and identity mapping superposition based on the updated feature maps output by the previous level of residual unit. The updated feature maps output by the last level of residual unit are used as input to a classification layer, which generates a predicted probability vector for the perturbation event category.

[0055] The residual unit includes a convolution operation structure and an activation structure for nonlinear mapping.

[0056] The convolution operation structure includes convolution kernel parameters and convolution bias parameters.

[0057] The classification layer includes a fully connected operation structure; the fully connected operation structure includes fully connected weight parameters and fully connected bias parameters.

[0058] The convolution kernel parameters, convolution bias parameters, fully connected weight parameters, and fully connected bias parameters constitute the internal parameters of the residual neural network model.

[0059] By performing fully connected operations and nonlinear activation on the time, frequency, and spatial description vectors, time channel weights, frequency channel weights, and spatial channel weights are obtained, respectively. The values ​​of the three channel weights are then multiplied on the same feature channel and input into a Sigmoid activation function for nonlinear mapping, resulting in a comprehensive channel weight with a limited value range. This comprehensive channel weight is then used to weight the values ​​of the intermediate feature map at the time and frequency positions in each feature channel, generating an attention-enhanced feature map, which is then input into subsequent residual units. This process ensures that feature channels exhibiting high importance only in the time, frequency, and spatial domains receive larger comprehensive channel weights, while channels that are insignificant in any direction are suppressed. This maintains the time and frequency distribution structure within each channel and improves the residual neural network model's ability to represent fiber optic anomalies and distinguish disturbance event categories.

[0060] The attention-enhanced feature map includes an updated feature map output by the first-level residual unit as an intermediate feature map, wherein the intermediate feature map is an updated feature map containing the feature information of the multi-channel feature image.

[0061] Global pooling is performed on the intermediate feature map to obtain a comprehensive channel weight. Based on the comprehensive channel weight, the intermediate feature map is weighted channel by channel to obtain an attention-enhanced feature map.

[0062] Global pooling is performed on the temporal features of the intermediate feature map to obtain a temporal direction description vector; global pooling is performed on the frequency direction features of the intermediate feature map to obtain a frequency direction description vector; and the normalized spatial coordinates of each spatial sampling point are processed by one-dimensional fully connected mapping and nonlinear activation to obtain a spatial direction description vector.

[0063] The time direction description vector: in, The time direction description vector is located in the channel. Time Index The components of position; The number of discrete frequency points in the frequency direction; Represents the intermediate feature map; Indicates the channel index; Indicates a time index; Indicates frequency index.

[0064] The frequency direction description vector: in, The frequency direction description vector is located in the channel. Frequency Index Components at position; The number of discrete points in the time direction; Indicates the first c The first channel, the first The time point, the first The eigenvalues ​​of the intermediate feature map at each frequency point.

[0065] Perform a fully connected operation on the time direction description vector, frequency direction description vector, and spatial direction description vector respectively; and map the time direction description vector, frequency direction description vector, and spatial direction description vector into a time weight vector, a frequency weight vector, and a spatial weight vector according to the preset mapping direction of each channel in three-dimensional space.

[0066] For any channel index Corresponding to part C of the multi-channel feature image, the time weight vector at the channel index Component values ​​at the location Frequency weight vector in channel index Component values ​​at the location and spatial weight vector in channel index Component values ​​at the location ,Will , and The product of the component values ​​is input into the Sigmoid activation function for nonlinear mapping, resulting in the component values ​​of the comprehensive channel weight vector in part C.

[0067] in, This indicates the composite channel weight vector at the channel index. The component at the location; This represents the Sigmoid activation function; Represents the time weight vector, in the channel index The component at the location; Represents the frequency weight vector, in the channel index The component at the location; Represents the spatial weight vector, in the channel index The component at that location.

[0068] Using the comprehensive channel weight vector in the channel index The components at the intermediate feature map in the channel index The attention-enhanced feature map is obtained by weighting the time and frequency feature values ​​at a given location.

[0069] in, Indicates the first The first channel, the first The time point, the first At each frequency point, the feature value of the attention-enhanced feature map; This indicates the composite channel weight vector at the channel index. The component at the location; Indicates the first The first channel, the first The time point, the first The eigenvalues ​​of the intermediate feature map at each frequency point.

[0070] By introducing a classification loss term, while ensuring the separability of source domain disturbance events, the residual neural network is supervised and constrained using pre-defined disturbance event category labels. This encourages the network to form clear inter-class discrimination boundaries in the feature space, thereby improving its ability to distinguish different disturbance events. The classification loss term guides the network to focus on key features related to disturbance event discrimination by minimizing the difference between the predicted results and the true category labels. This suppresses noise and redundant information while enhancing the model's ability to represent complex disturbance patterns and its recognition stability. By performing gradient updates on the network's internal parameters based on the classification loss function, the generalization ability of the trained residual neural network model can be improved under different datasets, small sample conditions, and scenarios with changing disturbance event categories, while ensuring recognition accuracy. This enhances the reliability of fiber optic anomaly disturbance identification.

[0071] The set of multi-channel feature images pre-selected as training samples is defined as the source domain.

[0072] The multi-channel feature images in the source domain are classified according to the preset perturbation event category information. A perturbation event category label is assigned to each multi-channel feature image, and the classification loss term is constructed based on the perturbation event category label.

[0073] in, Represents the classification loss term; Index representing the multi-channel feature image; Indicates the index of disturbance event categories; Indicates the label, when the source domain is... The perturbation event category of the multi-channel feature image is the first When class, When the source domain is The perturbation event category of the multi-channel feature image is not the first. When class, ; Indicates the first The multi-channel feature image belongs to the first... Predicted probability of perturbation event categories.

[0074] The residual neural network model is trained by minimizing the classification loss term. During the training process, the gradient of each internal parameter in the current training iteration is calculated using the chain rule, and the internal parameters are iteratively updated according to the preset learning rate coefficient to obtain the trained residual neural network model.

[0075] in, Indicates the first The values ​​of the internal parameters after each training iteration update; Indicates the first The values ​​of the internal parameters at the beginning of the next training iteration; The learning rate coefficient represents the step size for updating the control parameters; it is a real number greater than 0 and less than 1. This represents the classification loss term.

[0076] The disturbance determination result includes inputting the multi-channel feature image corresponding to the fiber disturbance data to be determined into a residual neural network model trained with classification loss, and outputting the predicted probability vector of each disturbance event category by the classification layer of the model; selecting the disturbance event category with the highest predicted probability from the predicted probability vector, and taking the disturbance event category as the disturbance determination result.

[0077] This embodiment also provides a transfer learning system suitable for identifying optical fiber anomalies and disturbances, including: a disturbance data acquisition module: to detect disturbances in the optical fiber under test, acquire backscattered Rayleigh signals, and form two-dimensional disturbance data in time and space order.

[0078] Multi-channel feature map construction module: performs digital demodulation, time-frequency analysis and spatial coordinate normalization on the time and space two-dimensional perturbation data to construct a multi-channel feature image representing time information, frequency information and spatial location information.

[0079] Judgment Module: Input the multi-channel feature image corresponding to the fiber disturbance data to be judged into the residual neural network model trained by the classification loss function, and output the final disturbance judgment result.

[0080] This embodiment also provides a computer device applicable to the transfer learning method for identifying fiber optic anomalies, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the transfer learning method for identifying fiber optic anomalies as proposed in the above embodiment.

[0081] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0082] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the transfer learning method for identifying fiber optic anomalies as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0083] In summary, this invention achieves improved transfer learning capability and accuracy under small sample conditions and new disturbance event scenarios by: detecting disturbances in the tested optical fiber and acquiring backscattered Rayleigh signals; constructing a multi-channel feature image that simultaneously represents time, frequency, and spatial location information; combining a residual neural network model containing an input layer, residual units, and a classification layer, and a channel-weighted attention mechanism based on time, frequency, and spatial direction description vectors; further introducing transfer learning training of the residual neural network model using a classification loss function; and using the residual neural network model trained with the classification loss function to classify and determine the disturbance data of the optical fiber to be judged. This effectively improves the accuracy of optical fiber abnormal disturbance identification and the model's generalization ability under complex working conditions.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A transfer learning method suitable for identifying optical fiber anomalies and disturbances, characterized in that: include, The fiber under test is subjected to disturbance detection, backscattered Rayleigh signals are collected, and two-dimensional disturbance data in time and space are formed in time and space order. The time and space two-dimensional perturbation data are digitally demodulated, time-frequency analyzed and spatial coordinate normalized to construct a multi-channel feature image representing time information, frequency information and spatial location information; The multi-channel feature image corresponding to the fiber disturbance data to be determined is input into the residual neural network model trained by the classification loss function, and the final disturbance determination result is output. The residual neural network model includes an input layer, residual units, and a classification layer. An attention-enhanced feature map is introduced into the residual neural network model. The attention-enhanced feature map is used as the update feature map of the first-level residual unit, and the updated feature map is input into the subsequent residual units.

2. The transfer learning method for identifying fiber optic anomalies as described in claim 1, characterized in that: The fiber under test is repeatedly probed, and the backscattered Rayleigh signal is sampled according to a preset sampling period after each probe. The samples are then arranged in time and space order to form two-dimensional perturbation data in time and space.

3. The transfer learning method for identifying fiber optic anomalies as described in claim 2, characterized in that: Digital orthogonal demodulation and phase recovery are performed on the two-dimensional temporal and spatial perturbation data to obtain the phase timing signal at each spatial sampling position along the optical fiber; Time-frequency analysis is performed on the phase timing signal to obtain a two-dimensional energy distribution in time and frequency, and the energy values ​​in different frequency intervals are mapped into time-frequency feature maps according to preset rules; The coordinates of each spatial sampling position in the fiber along the length direction are normalized according to the total number of spatial sampling points to obtain normalized spatial coordinates, and then arranged into a single-channel spatial position matrix according to the arrangement order of the spatial sampling positions on the fiber. The multi-channel feature image is obtained by analyzing time-frequency features and spatial features through two channels, and then stitching the obtained time-frequency feature map and spatial location matrix together along the channel dimension.

4. The transfer learning method for identifying fiber optic anomalies as described in claim 3, characterized in that: The residual neural network model includes an input layer receiving multi-channel feature maps, sequentially inputting the multi-channel feature maps into each level of residual units, each level of residual units performing convolution operations, nonlinear mapping, and identity mapping superposition based on the updated feature maps output by the previous level of residual units, and the updated feature maps output by the last level of residual units serving as input to a classification layer, which generates a predicted probability vector for the perturbation event category. The residual unit includes a convolution operation structure and an activation structure for nonlinear mapping; The convolution operation structure includes convolution kernel parameters and convolution bias parameters; The classification layer includes a fully connected operation structure; the fully connected operation structure includes fully connected weight parameters and fully connected bias parameters; The convolution kernel parameters, convolution bias parameters, fully connected weight parameters, and fully connected bias parameters constitute the internal parameters of the residual neural network model.

5. The transfer learning method for identifying fiber optic anomalies as described in claim 4, characterized in that: The attention-enhanced feature map includes an updated feature map output by the first-level residual unit as an intermediate feature map, wherein the intermediate feature map is an updated feature map containing the feature information of the multi-channel feature image. Global pooling is performed on the intermediate feature map to obtain a comprehensive channel weight, and channel-wise weighting is performed on the intermediate feature map based on the comprehensive channel weight to obtain an attention-enhanced feature map; Global pooling is performed on the temporal features of the intermediate feature map to obtain a temporal direction description vector; global pooling is performed on the frequency direction features of the intermediate feature map to obtain a frequency direction description vector; the normalized spatial coordinates of each spatial sampling point are processed by a one-dimensional fully connected mapping and nonlinear activation to obtain a spatial direction description vector; Perform a fully connected operation on the time direction description vector, frequency direction description vector, and spatial direction description vector respectively; and map the time direction description vector, frequency direction description vector, and spatial direction description vector into a time weight vector, a frequency weight vector, and a spatial weight vector according to the preset mapping direction of each channel in three-dimensional space. For any channel index Corresponding to part C of the multi-channel feature image, the time weight vector at the channel index Component values ​​at the location Frequency weight vector in channel index Component values ​​at the location and spatial weight vector in channel index Component values ​​at the location ,Will , and The product of the component values ​​is input into the Sigmoid activation function for nonlinear mapping to obtain the component values ​​of the comprehensive channel weight vector in part C. in, This indicates the composite channel weight vector at the channel index. The component at the location; This represents the Sigmoid activation function; Represents the time weight vector, in the channel index The component at the location; Represents the frequency weight vector, in the channel index The component at the location; Represents the spatial weight vector, in the channel index The component at the location; Using the comprehensive channel weight vector in the channel index The components at the intermediate feature map in the channel index The attention-enhanced feature map is obtained by weighting the time and frequency feature values ​​at each location; in, Indicates the first The first channel, the first The time point, the first At each frequency point, the feature value of the attention-enhanced feature map; This indicates the composite channel weight vector at the channel index. The component at the location; Indicates the first The first channel, the first The time point, the first The eigenvalues ​​of the intermediate feature map at each frequency point.

6. The transfer learning method for identifying fiber optic anomalies as described in claim 5, characterized in that: The set of multi-channel feature images pre-selected as training samples is defined as the source domain; The multi-channel feature images in the source domain are classified according to the preset perturbation event category information. A perturbation event category label is assigned to each multi-channel feature image, and the classification loss term is constructed based on the perturbation event category label. in, Represents the classification loss term; Index representing the multi-channel feature image; Indicates the index of disturbance event categories; Indicates the label, when the source domain is... The perturbation event category of the multi-channel feature image is the first When class, When the source domain is The perturbation event category of the multi-channel feature image is not the first. When class, ; Indicates the first The multi-channel feature image belongs to the first... Predicted probability of perturbation event categories; The residual neural network model is trained by minimizing the classification loss term. During the training process, the gradient of each internal parameter in the current training iteration is calculated using the chain rule, and the internal parameters are iteratively updated according to the preset learning rate coefficient to obtain the trained residual neural network model.

7. The transfer learning method for identifying fiber optic anomalies as described in claim 6, characterized in that: The disturbance determination result includes inputting the multi-channel feature image corresponding to the fiber disturbance data to be determined into a residual neural network model trained by a classification loss function, and outputting the predicted probability vector of each disturbance event category by the classification layer of the model; selecting the disturbance event category with the highest predicted probability from the predicted probability vector, and taking the disturbance event category as the disturbance determination result.

8. A transfer learning system for identifying fiber optic anomaly disturbances, based on the transfer learning method for identifying fiber optic anomaly disturbances according to any one of claims 1 to 7, characterized in that: Includes a disturbance data acquisition module: to detect disturbances in the fiber under test, acquire backscattered Rayleigh signals, and generate two-dimensional disturbance data in time and space order; Multi-channel feature map construction module: performs digital demodulation, time-frequency analysis and spatial coordinate normalization on the time and space two-dimensional perturbation data to construct a multi-channel feature image representing time information, frequency information and spatial location information; Judgment Module: Input the multi-channel feature image corresponding to the fiber disturbance data to be judged into the residual neural network model trained by the classification loss function, and output the final disturbance judgment result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the transfer learning method for identifying fiber optic abnormal disturbances as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the transfer learning method for identifying fiber optic abnormal disturbances as described in any one of claims 1 to 7.

Images