A Micro-Disturbance Identification Method for Optical Cables Based on Physical Simulation and Self-Supervised Temporal Decoupling

Abstract

This invention discloses a method for identifying micro-disturbances in optical cables based on physical simulation and self-supervised temporal decoupling, relating to the field of optical cable identification technology. The method includes: constructing a physical digital twin simulator to generate a high-fidelity training set; constructing a deep learning model using a lightweight temporal decoupling network; training the model using a phased training strategy, sequentially performing self-supervised noise distribution pre-training, simulation-supervised training, and spectral domain physical consistency fine-tuning; inputting real-time acquired DAS temporal data into the trained model, outputting the optical cable identification ID and physical location; the lightweight temporal decoupling network includes a physically guided preprocessing module, a lightweight U-Net separation module, a sparse gating module, and an intelligent parallel decoding module. This invention, through simulation-driven data generation and a phased training strategy, addresses the shortcomings of existing technologies in reducing data costs, improving detection capabilities in low SNR environments, and achieving multi-source blind source separation.

Description

Technical Field

[0001] This invention relates to the field of optical cable identification technology, specifically to an optical cable micro-perturbation identification method based on physical simulation and self-supervised temporal decoupling. Background Technology

[0002] Distributed Acoustic Sensing (DAS) technology, as a novel fiber optic sensing technology, utilizes the Rayleigh scattering effect in optical fibers to achieve continuous distributed monitoring of vibration and acoustic signals along the entire optical cable. Due to its advantages such as resistance to electromagnetic interference, long-distance monitoring, and passive intrinsic safety, it has been widely applied in fields such as power cable monitoring and traffic flow monitoring. However, in practical engineering applications, especially for monitoring micro-disturbances in underground utility tunnels, cable trenches, and complex direct-buried environments, existing DAS signal processing and recognition technologies still face many challenges, mainly in the following aspects:

[0003] 1. Heavy Reliance on Large-Scale Labeled Data and the "Cold Start" Problem: Current intelligent recognition solutions are mainly based on deep learning methods, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). These supervised learning models typically require massive amounts of on-site measured data with clearly labeled categories for training to achieve the desired recognition accuracy. In real-world power or fiber optic cable maintenance scenarios, collecting a large amount of representative event data (such as manual knocking, damage, or various disturbance conditions) is not only costly but also impractical. Furthermore, the manual labeling process for on-site data is time-consuming and labor-intensive, requiring expert knowledge to ensure labeling quality. These factors lead to a "cold start" dilemma when deploying new systems or migrating models to new scenarios: a lack of sufficient labeled samples to complete model training or rapid fine-tuning, thus preventing the achievement of usable recognition performance in a short period.

[0004] 2. Challenges in Feature Extraction of Weak Signals under Extremely Low Signal-to-Noise Ratios: In engineering-grade "non-destructive testing," active disturbance signals used for optical cable identification or disturbance marking typically need to maintain extremely low amplitudes to avoid physical damage or impact on communication services. However, in complex laying environments such as underground utility tunnels, there are often strong background interferences such as vehicle vibration, power frequency electromagnetic interference, and ventilation equipment noise, causing effective disturbance signals to be submerged by noise, resulting in extremely low signal-to-noise ratios (SNR). SNR is usually defined as the logarithmic expression of the ratio of useful signal power to noise power. When the SNR is below a certain threshold, traditional detection methods based on energy thresholds (such as constant false alarm rate (CFAR)) or feature extraction methods based on simple time-frequency analysis often struggle to distinguish weak disturbances from background noise, leading to high false alarm and missed detection rates. CFAR methods rely on estimating the statistical characteristics of background noise to set the detection threshold, but under conditions of non-stationary noise, multi-source interference, and significant multipath effects, background statistics are difficult to estimate stably, resulting in decreased detection performance.

[0005] 3. Insufficient Blind Source Separation Capability for Multi-Source Aliasing Signals: In multi-cable co-channel or array-based laying scenarios, the single-channel time-series signal acquired by the DAS system is often a linear superposition of vibration signals from multiple optical cables and environmental noise. However, most existing technologies assume that only one event occurs at a single moment, lacking an effective blind source separation mechanism. When multiple optical cables are disturbed simultaneously, or when concurrent events occur at different locations on the same optical cable, severe aliasing occurs in the spatiotemporal domain. Existing algorithms struggle to effectively decouple the mixed signals into independent event sources, thus failing to achieve independent identification and precise location of each optical cable.

[0006] Based on this, we now provide a method for identifying optical cable micro-disturbances based on physical simulation and self-supervised temporal decoupling, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0007] The purpose of this invention is to provide a method for identifying micro-disturbances in optical cables based on physical simulation and self-supervised temporal decoupling, so as to solve the problems of the existing technical solutions in the background art in terms of data acquisition cost, robustness in low signal-to-noise ratio environments, and multi-target dealiasing capability.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling includes the following steps:

[0010] Step S1: Construct a physical digital twin simulator based on physical channel response and real noise, and generate a high-fidelity training set containing multipath effects, dispersion and complex noise;

[0011] Step S2: Construct a deep learning model. The deep learning model uses a physically guided lightweight temporal decoupling network to extract and separate multi-path micro-perturbation features from single-channel aliased temporal signals.

[0012] Step S3: Train the deep learning model using a phased training strategy, sequentially performing self-supervised noise distribution pre-training, simulation-supervised training, and spectral domain physical consistency fine-tuning.

[0013] Step S4: Input the real-time collected DAS time-series data into the trained deep learning model, and output the optical cable identity ID and physical location.

[0014] The lightweight temporal decoupling network consists of four cascaded modules, including a physical guidance preprocessing module, a lightweight U-Net separation module, a sparse gating module, and an intelligent parallel decoding module.

[0015] Step S1 specifically includes:

[0016] Step S11: Based on the preset event types in the event type library, use a linear frequency modulation signal as the basic encoding to generate an ideal excitation signal;

[0017] Step S12: For different laying states of optical cables in underground utility tunnels, construct a random channel impulse response filter that conforms to Rayleigh fading distribution, and perform one-dimensional convolution operation between the ideal excitation signal and the channel impulse response to obtain a signal containing multipath reverberation characteristics.

[0018] Step S13: Based on the optical fiber transmission loss characteristics, construct an exponential attenuation function, multiply the convolved signal containing multipath reverberation characteristics with the attenuation factor, and simulate signal attenuation over long distance transmission.

[0019] Step S14: Randomly extract environmental noise segments from the real DAS background noise library, calculate the noise scaling factor according to the preset target signal-to-noise ratio, and superimpose the scaled noise onto the signal.

[0020] Step S15: Output two-dimensional spatiotemporal matrix data containing physical propagation characteristics and real environmental noise, and use it as a high-fidelity training set and validation set for the deep learning model.

[0021] The expression for the two-dimensional spatiotemporal matrix data in step S15 is:

[0022] ;

[0023] in, It is a two-dimensional spatiotemporal matrix data. In terms of time dimension, For spatial dimensions, For ideal incentive signals, For convolution operation nodes, For channel impulse response, The fiber attenuation coefficient, This is the noise scaling factor. This is a segment of environmental noise.

[0024] The expression for the ideal excitation signal is:

[0025] ;

[0026] in, The signal amplitude, The starting frequency, For the termination frequency, For signal duration, This is a window function used to smooth signal edges;

[0027] The expression for the channel impulse response is:

[0028] ;

[0029] in, For the number of multipaths, , , The first The attenuation coefficient, phase shift, and time delay of each path. The imaginary unit;

[0030] The expression for the noise scaling factor is:

[0031] ;

[0032] in, For signal power, This represents the original noise power. The preset target signal-to-noise ratio.

[0033] Furthermore, the physical guidance preprocessing module, serving as the input to the lightweight temporal decoupling network, is used to receive data with a dimension of [missing information]. The original DAS spatiotemporal matrix, where, For batch size, This represents the number of spatial sampling points. To determine the number of time sampling points, the physical guidance preprocessing module employs an embedded learnable time-frequency transform convolutional layer. Through end-to-end backpropagation, it optimizes the convolutional kernel parameters, mapping the original one-dimensional time-domain vibration signal into a high-dimensional time-frequency feature tensor containing rich physical information, thereby capturing the transient frequency texture of the micro-perturbation signal.

[0034] Furthermore, the lightweight U-Net separation module is used to perform blind source separation tasks and adopts an encoder-decoder U-shaped architecture. The convolution operation adopts depthwise separable convolution. The shallow high-resolution features of the encoder are directly passed to the decoder through skip connections to preserve the fine structural information of the signal in the spatiotemporal domain and avoid information loss during downsampling. The lightweight U-Net separation module finally outputs several independent feature channels, and the number of feature channels is the preset maximum number of concurrent sources.

[0035] Furthermore, the sparse gating module is located between the lightweight U-Net separation module and the intelligent parallel decoding module, and introduces a learnable soft thresholding function to perform sparsification processing on the feature channels.

[0036] Furthermore, the intelligent parallel decoding module includes several parallel decoding units, the number of which is consistent with the number of feature channels. Each decoding unit has a multi-head self-attention sub-layer for capturing the long-term dependencies and global context information of the signal. The end of each decoding unit includes two output branches, namely a classification head and a regression head. The classification head outputs the confidence probability of the corresponding channel signal belonging to various events, and the regression head outputs the normalized physical location coordinates of the corresponding channel signal on the optical fiber.

[0037] Furthermore, the self-supervised noise distribution pre-training in step S3 specifically includes:

[0038] Using unlabeled real background noise data or single-source time-series signals processed by random masking as input, a masked signal reconstruction task is performed to initialize the model's noise representation and basic denoising capability.

[0039] Let the input timing signal be Masking operation yields ,in, For the mask, The timing signal after masking;

[0040] Deep learning models predict the masked portion from the masked temporal signal. And minimize the reconstruction error by reconstructing the loss: ,in, To reconstruct the loss, For mathematical expectation operators, It is the square of the L2 norm.

[0041] Furthermore, in step S3, the simulation-supervised training enables the deep learning model to master decoupling target features from aliased signals and mapping abstract features to physical coordinates. The simulation-supervised training specifically includes:

[0042] Fully supervised training is performed using the simulation data generated in step S1. The training objectives include two tasks: identity classification and location regression after blind source separation. The cross-entropy loss function is used for the identity classification task, and the Smooth L1 function is used for the location regression task. The expressions are as follows:

[0043] ;

[0044] ;

[0045] in, For classifying losses, Represents the cross-entropy loss function. To categorize identities using real-world tags, For identity classification prediction results, For position loss, To revert to the true labels for location, The result is the location regression prediction.

[0046] The joint optimization objective is expressed as: ;

[0047] in, To jointly optimize objectives, is the weighting coefficient used to balance the classification and localization losses.

[0048] Furthermore, the spectral domain physical consistency fine-tuning operation in step S3 specifically includes:

[0049] To alleviate the distribution difference between simulation data and measured data, a two-dimensional Fourier transform is applied to the spatiotemporal matrix output by the lightweight U-Net separation module to obtain a frequency-wavenumber domain representation. Based on the physical relationship of sound wave propagation, a physically allowed region is constructed in the f–k domain.

[0050] Under the ideal dispersionless approximation, sound waves satisfy a linear dispersion relation, and the corresponding theoretical wavenumber is expressed as: A physical mask is defined based on the linear dispersion relation of sound waves, and is expressed as:

[0051] ;

[0052] in, For the allowed spectral width, For physical mask, For frequency, The central theoretical wavenumber is derived based on the linear dispersion relation of sound waves, i.e. , For the spatial frequency variable wavenumber, For the speed of transmission;

[0053] The spectral domain physical consistency loss is defined as the weighted sum of the out-of-mask energies, expressed as:

[0054] ;

[0055] in, Let be the spectral domain physical consistency loss function. The frequency-wavenumber domain representation after two-dimensional Fourier transform and satisfying , For two-dimensional Fourier transform operators, It is a spacetime matrix.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] 1. This invention constructs a physical digital twin simulator to generate high-fidelity training samples containing multipath effects and complex background interference. With the help of this simulation-driven data augmentation and learning strategy, deep learning model pre-training and fine-tuning can be completed with only limited on-site labeled data or no large-scale labeled data, thereby reducing the need for on-site experiments and manual labeling, and reducing the investment of manpower and material resources.

[0058] 2. In terms of signal preprocessing and modeling, this invention introduces a physical guidance preprocessing module and a sparse gating module. The physical guidance preprocessing module extracts time-frequency features based on learnable short-time Fourier transform, and the sparse gating module uses a soft threshold function to suppress irrelevant noise channels, thereby avoiding noise interference with the network attention mechanism and enhancing the response capability to weak perturbation features.

[0059] 3. This invention proposes a lightweight temporal decoupling network to achieve blind source separation of single-channel mixed temporal signals. It can decouple multiple optical cable vibration signals that are linearly or nearly linearly superimposed into several independent feature streams, and support independent identification and location estimation of multiple disturbance sources within the same time window, thereby improving the concurrent monitoring capability and event parsing accuracy of the DAS system in complex laying scenarios.

[0060] 4. This invention introduces a spectral domain physical consistency fine-tuning operation into the training objective, applying physical constraints such as propagation speed and dispersion characteristics to the model output in the frequency-wavenumber domain, prompting the network to conform to the basic laws of sound wave propagation while fitting the data, which helps to narrow the gap between simulation data and measured data. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the steps of the optical cable micro-disturbance identification method of the present invention.

[0062] Figure 2 This is a schematic diagram illustrating the simulation data generation principle of the physical digital twin simulator of the present invention.

[0063] Figure 3 This is a schematic diagram of the lightweight temporal decoupling network of the present invention.

[0064] Figure 4 This is a schematic diagram of the module architecture of the lightweight temporal decoupling network of the present invention.

[0065] Figure 5 This is a schematic diagram of the phased training strategy of the present invention.

[0066] Figure 6 This is a schematic diagram illustrating the mechanism of the spectral domain physical consistency loss function of the present invention.

[0067] Figure 7 This is a schematic diagram of multi-source signal dealiasing according to the present invention.

[0068] Figure 8 This is a comparison chart showing the application effects of the model of the present invention.

[0069] Figure label annotations: Physical guidance preprocessing module 10, lightweight U-Net separation module 20, sparse gating module 30, intelligent parallel decoding module 40. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0071] Example 1

[0072] In this embodiment, as Figure 1 As shown, this invention proposes a method for identifying micro-disturbances in optical cables based on physical simulation and self-supervised temporal decoupling, specifically including the following steps:

[0073] Step S1: Construct a physical digital twin simulator based on physical channel response and real noise, and generate a high-fidelity training set containing multipath effects, dispersion and complex noise;

[0074] Step S2: Construct a deep learning model. The deep learning model uses a physically guided lightweight temporal decoupling network to extract and separate multi-path micro-perturbation features from single-channel aliased temporal signals.

[0075] Step S3: Train the deep learning model using a phased training strategy, sequentially performing self-supervised noise distribution pre-training, simulation-supervised training, and spectral domain physical consistency fine-tuning. A progressive training scheme is adopted to improve the model's noise suppression capability, decoupling capability, and physical consistency.

[0076] Step S4: Input the real-time collected DAS time-series data into the trained deep learning model, and output the optical cable identification ID and physical location. Deploy the trained deep learning model on the monitoring terminal, collect the DAS time-series data of the optical cable in real time, and input the data into the model after slicing. The model directly outputs the identification ID and precise physical location of each optical cable at the current moment from end to end, realizing intelligent identification of multi-source micro-disturbance events.

[0077] In this embodiment, the present invention is based on generating high-fidelity training data using physical digital twins, and combines a training process of self-supervised noise distribution pre-training, simulation-supervised training and spectral domain physical consistency fine-tuning. It also uses a physically guided lightweight temporal decoupling network to achieve blind source separation, identity recognition and location regression of single-channel aliased signals.

[0078] Example 2

[0079] In this embodiment, as Figure 2 As shown, to address the difficulty in obtaining real optical cable micro-disturbance samples, this invention constructs a physical digital twin engine, i.e., a physical digital twin simulator, which includes signal excitation, channel response, physical attenuation, and environmental noise superposition. The data generation process of this physical digital twin simulator strictly follows the physical propagation law of sound waves in optical fiber media. The specific steps are as follows:

[0080] Step S11, corresponding Figure 2 The left-side signal input is based on the preset event types (such as manual excavation, mechanical vibration, and knocking) in the event type library, which is stored in the DAS system. In this embodiment, a linear frequency modulation (LFM / Chirp) signal is used as the basic encoding to generate an ideal excitation signal. Alternatively, an exponentially decaying pulse can be used as the basic encoding.

[0081] Step S12, corresponding Figure 2 Convolution operation nodes in For different laying states of optical cables in underground utility tunnels (such as suspended, wall-mounted, or directly buried), a random channel impulse response (CIR) filter conforming to Rayleigh fading distribution is constructed. The CIR filter consists of a direct main path and multiple reflection multipaths with random delays and attenuations. The ideal excitation signal is convolved with the channel impulse response (CIR) in one dimension to obtain a signal containing multipath reverberation characteristics. This signal is used to simulate the real physical distortions that occur when sound waves propagate in complex media. For convolution operation nodes;

[0082] Step S13, corresponding Figure 2 Multiplication node in Based on the optical fiber transmission loss characteristics of light or sound waves, a spatial dimension is applied to the convolutional signal. Energy decay on the surface, constructing an exponential decay function. , The exponential attenuation term for fiber optic transmission is used to convolve the signal, which contains multipath reverberation characteristics. Multiply by the attenuation factor to simulate signal attenuation over long distances;

[0083] Step S14, corresponding Figure 2 The addition operation node in To eliminate the domain gap between simulation data and the real environment, this embodiment establishes a real DAS background noise library that includes various operating conditions such as rain, passing vehicles, and silence. The real DAS background noise library is constructed in the following way: In actual monitoring scenarios without target disturbances, DAS background signals are collected under multiple time periods and operating conditions (such as silence, wind and rain, traffic, etc.). After preprocessing such as detrending and downsampling, the signals are sliced ​​into several noise segments of fixed length. This library can be expanded and updated according to specific application scenarios, or it can be supplemented by publicly available distributed fiber optic sensor noise datasets. A segment of environmental noise with the same length as the signal is randomly extracted from the real DAS background noise library. The noise scaling factor is calculated according to the preset target signal-to-noise ratio. The scaled noise is superimposed on the signal to generate extreme operating condition data with a signal-to-noise ratio as low as -15dB.

[0084] Step S15, corresponding Figure 2 The right-hand output, after the above processing, produces a two-dimensional spatiotemporal matrix data containing physical propagation features and real environmental noise. This data is used as a high-fidelity training and validation set for the deep learning model, supporting subsequent self-supervised and supervised training stages.

[0085] The expression for two-dimensional spatiotemporal matrix data is:

[0086] ;

[0087] in, It is a two-dimensional spatiotemporal matrix data. In terms of time dimension, For spatial dimensions, For ideal incentive signals, For convolution operation nodes, For channel impulse response, The fiber attenuation coefficient, This is the noise scaling factor. This is an environmental noise segment collected by an actual DAS system;

[0088] The expression for the ideal excitation signal is:

[0089] ;

[0090] in, The signal amplitude, The starting frequency, For the termination frequency, For signal duration, For window functions used to smooth signal edges (such as the Hann window);

[0091] ;

[0092] in, For the number of multipaths, , , The first The attenuation coefficient, phase shift, and time delay of each path. As an imaginary unit, it is used in fields such as communications and electronic engineering to avoid confusion with the symbol for electric current. Representing the imaginary unit, functionally, this formula, using Euler's formula, transforms the first... Phase shift of multipath It is incorporated into the complex expression of the channel response, thus simultaneously describing the amplitude attenuation and phase shift characteristics of the signal;

[0093] The expression for the noise scaling factor is:

[0094] ;

[0095] in, For signal power, This represents the original noise power. The preset target signal-to-noise ratio.

[0096] In summary, to alleviate the heavy reliance on large-scale labeled data and the "cold start" problem, this invention uses physics-driven digital twin to generate high-fidelity synthetic simulation data with multipath effects and background noise characteristics. It also combines self-supervised learning with a fine-tuning strategy of a small amount of on-site annotation, thereby reducing the reliance on large-scale manually labeled data, lowering data acquisition and annotation costs, and shortening the time from deployment to availability of the DAS system.

[0097] Example 3

[0098] In this embodiment, considering the challenges of existing technologies in extracting weak signal features under extremely low signal-to-noise ratios, this invention improves the separability of weak signals in the feature space through physically guided time-frequency preprocessing, sparse gating, and spectral domain physical consistency constraints. This enhances the detection capability of useful disturbances and reduces false alarms and missed alarms under low SNR conditions. Figure 3 and Figure 4 As shown, the present invention constructs a lightweight temporal decoupling network called Phys-DecoupleNet. The lightweight temporal decoupling network consists of four cascaded modules, including a physical guidance preprocessing module 10, a lightweight U-Net separation module 20, a sparse gating module 30, and an intelligent parallel decoding module 40.

[0099] The physical guidance preprocessing module 10 serves as the input to the lightweight temporal decoupling network, used to receive data with dimensions of... The original DAS spatiotemporal matrix, where, For batch size, This represents the number of spatial sampling points. To improve the number of time sampling points, the physical guidance preprocessing module 10 abandons the traditional fixed STFT structure and adopts an embedded Learnable Time-Frequency Conv Layer. Through end-to-end backpropagation, the convolution kernel parameters are optimized to map the original one-dimensional time-domain vibration signal into a high-dimensional time-frequency feature tensor containing rich physical information, thereby capturing the transient frequency texture of the micro-perturbation signal.

[0100] The lightweight U-Net separation module 20 is used to perform blind source separation tasks and adopts an encoder-decoder U-shaped architecture. The convolutional operation uses depthwise separable convolution, which significantly reduces the number of parameters while maintaining feature extraction capabilities. Shallow high-resolution features from the encoder are directly passed to the decoder through skip connections to preserve the fine structural information of the signal in the spatiotemporal domain and avoid information loss during downsampling. The lightweight U-Net separation module 20 ultimately outputs several independent feature channels. The number of feature channels is the preset maximum number of concurrent sources. This number can be preset according to the actual application scenario and fixed during training. This parameter is determined before network training and remains unchanged during model deployment. It can also be adjusted by reconfiguring the network structure according to actual monitoring needs. An example is shown below. This decouples the aliased mixed features into independent single-source feature streams. This is the preset maximum number of concurrent sources;

[0101] The sparse gating module 30 is located between the lightweight U-Net separation module 20 and the intelligent parallel decoding module 40. It introduces a learnable soft thresholding function to sparsify the feature channels. Its mathematical processing logic is as follows:

[0102] ;

[0103] in, As input features, The dynamic threshold parameter is automatically learned by the lightweight temporal decoupling network. This process is used to automatically detect the energy intensity of each channel, set the feature value of the channel below the threshold to zero, effectively optimize the feature flow, prevent noise signal from propagating backward and interfering with the attention mechanism, and significantly improve the robustness under low signal-to-noise ratio.

[0104] The intelligent parallel decoding module 40 contains several parallel decoding units, the number of which is the same as the number of feature channels. The decoding units are shown as stacked rectangles, each independently processing several gated feature streams. Each decoding unit has a multi-head self-attention sub-layer, as shown by the triangular arrow symbol in the figure, which is used to capture the long-term dependencies of the signal and global context information. The end of each decoding unit includes two output branches: a classification head and a regression head. The classification head outputs the confidence probability of the corresponding channel signal belonging to various events (such as mining, vehicles, noise, etc.), and the regression head outputs the normalized physical location coordinates of the corresponding channel signal on the optical fiber, i.e., floating-point numbers between 0 and 1.

[0105] In this embodiment, to address the spatiotemporal signal aliasing problem caused by multi-cable co-testing or concurrent events, the present invention constructs a lightweight timing decoupling network and a matching intelligent decoding structure to achieve blind source separation and feature decoupling of multiple disturbance sources in a single-channel mixed signal. It can independently identify multiple overlapping optical cable micro-disturbance events and achieve high-precision physical location positioning.

[0106] Example 4

[0107] In this embodiment, wherein, as Figure 5 As shown, the present invention also constructs a phased training strategy, which consists of self-supervised noise distribution pre-training, simulation-supervised training, and spectral domain physical consistency fine-tuning operation;

[0108] Phase 1, Self-Supervised Noise Distribution Pre-training (Masked Signal Reconstruction), specifically includes:

[0109] Using unlabeled real background noise data or single-source time-series signals processed by random masking as input, a masked signal reconstruction task is performed to initialize the model's noise representation and basic denoising capability.

[0110] Let the input timing signal be Masking operation yields ,in, For the mask, The timing signal after masking;

[0111] Deep learning models predict the masked portion from the masked temporal signal. And minimize the reconstruction error by reconstructing the loss: ,in, To reconstruct the loss, For mathematical expectation operators, The square of the L2 norm;

[0112] Phase 2, Multi-task Supervised Learning, specifically includes:

[0113] Fully supervised training is performed using the simulation data generated in step S1. The training objectives include two tasks: identity classification and location regression after blind source separation. The cross-entropy loss function is used for the identity classification task, and the Smooth L1 function is used for the location regression task. The expressions are as follows:

[0114] ;

[0115] ;

[0116] in, For classifying losses, Represents the cross-entropy loss function. To categorize identities using real-world tags, For identity classification prediction results, For position loss, To revert to the true labels for location, The result is the location regression prediction.

[0117] The joint optimization objective is expressed as: ;

[0118] in, To jointly optimize objectives, These are weighting coefficients used to balance classification and localization losses;

[0119] This stage enables deep learning models to decouple target features from aliased signals and map abstract features to physical coordinates;

[0120] Phase 3, Spectral Physics Fine-tuning, specifically includes:

[0121] To mitigate the distribution discrepancy between simulation and measured data, the spatiotemporal matrix output by the lightweight U-Net separation module 20 is... Applying a two-dimensional Fourier transform yields a frequency-wavenumber domain representation. Based on the physical relationship of sound wave propagation, a physically allowed region is constructed in the f–k domain;

[0122] Under the ideal dispersionless approximation, sound waves satisfy a linear dispersion relation, and the corresponding theoretical wavenumber is expressed as: A physical mask is defined based on the linear dispersion relation of sound waves, and is expressed as:

[0123] ;

[0124] in, For the allowed spectral width, For physical mask, For frequency, The central theoretical wavenumber is derived based on the linear dispersion relation of sound waves, i.e. , For the spatial frequency variable wavenumber, For the speed of transmission;

[0125] The spectral domain physical consistency loss is defined as the weighted sum of the out-of-mask energies, expressed as:

[0126] ;

[0127] in, Let be the spectral domain physical consistency loss function. The frequency-wavenumber domain representation after two-dimensional Fourier transform (2D-FFT) and satisfying , For two-dimensional Fourier transform operators, It is a spacetime matrix;

[0128] The total loss during the final staged training can be expressed as:

[0129] ;

[0130] in, For the total loss function, These are weighting coefficients used to control the strength of physical consistency constraints, forcing the output of deep learning models to strictly follow the physical laws of the wave equation;

[0131] In this embodiment, to reduce the “simulation-reality” difference between training based on simulation data and measured data from actual engineering scenarios, the present invention introduces a spectral domain physical consistency loss function during the training process. This guides the deep learning model to follow the basic physical laws of sound wave propagation and fiber optic sensing, thereby alleviating the generalization performance degradation problem to a certain extent and improving the model’s adaptability and robustness under different application scenarios and environmental conditions.

[0132] In summary, in this embodiment, the present invention is based on generating high-fidelity training data using physical digital twins, combined with a training process of self-supervised noise distribution pre-training, simulation-supervised training and spectral domain physical consistency fine-tuning, and uses a physically guided lightweight temporal decoupling network to achieve blind source separation, identity recognition and location regression of single-channel aliased signals.

[0133] Specifically, to verify the effectiveness of the aforementioned staged training strategy and lightweight temporal decoupling network architecture, this invention analyzes and verifies the performance of the deep learning model from three dimensions: interpretability of the physical mechanism, signal decoupling performance in multi-source aliasing scenarios, and robustness under extremely low signal-to-noise ratio conditions, as shown below:

[0134] Figure 6 This is a schematic diagram illustrating the mechanism of the spectral domain physical consistency loss function. Figure 6 Figure a shows the spatiotemporal waveform matrix output by the model. The tilted stripe texture represents the propagation of the vibration signal along the optical fiber, and its tilt slope is related to the propagation speed of the sound wave in the medium. There is a correspondence; after performing a two-dimensional Fourier transform on this spatiotemporal matrix, we can obtain the following: Figure 6 Figure b shows the frequency-wavenumber domain energy distribution.

[0135] According to the wave equation theory, vibrational energy that satisfies the laws of physical propagation should be mainly concentrated in a specific sector-shaped region centered at the origin in the fk domain. This sector-shaped region corresponds to the speed of sound in the medium. Based on the reasonable range of values, this invention defines the region as the Physically Allowed Zone (Valid Zone) and the region outside it as the Penalty Zone (Penalty), such as... Figure 6 As shown in Figure b, the light gray noise scattered in the background represents artifacts or noise components that do not conform to the aforementioned physical propagation laws. Based on this, this invention constructs a spectral domain physical consistency loss function, which weights and penalizes energy components falling into the penalty region. By minimizing the energy proportion in this region, the model output is constrained to satisfy the given physical propagation characteristics in the frequency domain. This constraint mechanism guides the model to suppress non-physical artifacts during training, making the final output spatiotemporal waveform closer to the physical propagation characteristics. Figure 6 The ideal physical form shown in Figure a improves the physical consistency and interpretability of the model's prediction results;

[0136] For scenarios involving multiple cables being tested together or multiple points of concurrent disturbance on the same optical cable, Figure 7 The model's performance in decoupling multi-source aliased signals is demonstrated. Figure 7 Figure a shows the original aliasing spatiotemporal matrix as input. Figure 7 Figure b shows the spatiotemporal matrix of the separated signal source A. Figure 7 Figure c shows the spatiotemporal matrix of the separated signal source B, from which we can deduce:

[0137] exist Figure 7 In Figure a, two vibration events with opposite directions of motion and different waveform characteristics exist simultaneously within the same spatiotemporal window. They exhibit a distinct "X"-shaped interferometric texture on the spatiotemporal image, superimposed with strong environmental random noise. In the overlapping region, the signal characteristics undergo severe nonlinear aliasing. Traditional time-frequency filtering-based processing methods struggle to effectively distinguish the various signal sources. After processing by the deep learning model of this invention, Figure 7 The rightward propagating signal source A, as extracted in Figure b, corresponds to... Figure 7 The right-leaning diagonal texture in Figure a corresponds to the backpropagation waveform and most of the background noise, where the backpropagation waveform and most of the background noise are effectively suppressed. Figure 7 The signal source B, which is extracted and propagates to the left, is basically unaffected by the signal source A in Figure c.

[0138] according to Figure 7 Figures b and c show that the lightweight U-Net separation module 20, combined with a sparse gating mechanism, can effectively decouple complex aliased signals with nonlinear superposition, realize the separation and independent characterization of multi-target vibration events, and thus improve the accuracy of identification and positioning under multi-source co-measurement conditions.

[0139] To evaluate the model's adaptability to extremely harsh noise environments, Figure 8 The image shows a verification diagram of signal recognition performance under extremely low signal-to-noise ratio (SNR) conditions (-10dB). Figure 8 Figure a shows the original DAS spatiotemporal heatmap of the model input. Figure 8 Figure b shows the single-channel waveform output by the model. Figure 8 Figure c shows the frequency domain analysis of the input signal;

[0140] from Figure 8 As can be seen in Figure a, the effective vibration signal features are extremely weak in the strong background noise, and the overall image shows a dense speckle texture, making it difficult for the human eye to identify the specific vibration mode.

[0141] Figure 8Figure c shows the corresponding frequency domain analysis results. It can be seen that the background noise has obvious pink noise characteristics, with a high proportion of low-frequency energy, which can easily mask the low-amplitude target vibration signal. Under such conditions, traditional methods often have difficulty extracting effective features stably. However, after introducing the physical guidance preprocessing module 10 and the sparse gating mechanism, the deep learning model in this invention can still extract the target response from the complex noise background under the above conditions.

[0142] Figure 8 Figure b shows the time-domain response of the single-channel waveform output by the model at the target location. Although it still contains some residual noise and nonlinear distortion, its overall envelope and main vibration modes are relatively complete and can be reliably used by subsequent detection and recognition algorithms. The experimental results show that under extremely low signal-to-noise ratio conditions, the model of this invention can still maintain high detection sensitivity and localization robustness.

[0143] In summary, this invention, through simulation-driven data generation, phased training strategies, physics-guided time-frequency processing, temporal decoupling, and fine-tuning of spectral domain physical consistency, demonstrates significant advantages over existing technologies in reducing data costs, improving detection capabilities in low SNR environments, achieving multi-source blind source separation, and enhancing model generalization and engineering deployability, and has promising application prospects.

[0144] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying micro-disturbances in optical cables based on physical simulation and self-supervised temporal decoupling, characterized in that, Specifically, the following steps are included: Step S1: Construct a physical digital twin simulator based on physical channel response and real noise, and generate a high-fidelity training set containing multipath effects, dispersion and complex noise; Step S2: Construct a deep learning model. The deep learning model uses a physically guided lightweight temporal decoupling network to extract and separate multi-path micro-perturbation features from single-channel aliased temporal signals. Step S3: Train the deep learning model using a phased training strategy, sequentially performing self-supervised noise distribution pre-training, simulation-supervised training, and spectral domain physical consistency fine-tuning. Step S4: Input the real-time collected DAS time-series data into the trained deep learning model, and output the optical cable identity ID and physical location. The lightweight temporal decoupling network consists of four cascaded modules, including a physical guidance preprocessing module, a lightweight U-Net separation module, a sparse gating module, and an intelligent parallel decoding module. Step S1 specifically includes: Step S11: Based on the preset event types in the event type library, use a linear frequency modulation signal as the basic encoding to generate an ideal excitation signal; Step S12: For different laying states of optical cables in underground utility tunnels, construct a random channel impulse response filter that conforms to Rayleigh fading distribution, and perform one-dimensional convolution operation between the ideal excitation signal and the channel impulse response to obtain a signal containing multipath reverberation characteristics. Step S13: Based on the optical fiber transmission loss characteristics, construct an exponential attenuation function, multiply the convolved signal containing multipath reverberation characteristics with the attenuation factor, and simulate signal attenuation over long distance transmission. Step S14: Randomly extract environmental noise segments from the real DAS background noise library, calculate the noise scaling factor according to the preset target signal-to-noise ratio, and superimpose the scaled noise onto the signal. Step S15: Output two-dimensional spatiotemporal matrix data containing physical propagation characteristics and real environmental noise, and use it as a high-fidelity training set and validation set for the deep learning model. The expression for the two-dimensional spatiotemporal matrix data in step S15 is: ； in, It is a two-dimensional spatiotemporal matrix data. In terms of time dimension, For spatial dimensions, For ideal incentive signals, For convolution operation nodes, For channel impulse response, The fiber attenuation coefficient, This is the noise scaling factor. This is a segment of environmental noise. The expression for the ideal excitation signal is: ； in, The signal amplitude, The starting frequency, For the termination frequency, For signal duration, This is a window function used to smooth signal edges; The expression for the channel impulse response is: ； in, For the number of multipaths, , , The first The attenuation coefficient, phase shift, and time delay of each path. The imaginary unit; The expression for the noise scaling factor is: ； in, For signal power, This represents the original noise power. The preset target signal-to-noise ratio.

2. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, The physical guidance preprocessing module serves as the input to the lightweight temporal decoupling network, and is used to receive data with a dimension of [missing information]. The original DAS spatiotemporal matrix, where, For batch size, This represents the number of spatial sampling points. To determine the number of time sampling points, the physical guidance preprocessing module employs an embedded learnable time-frequency transform convolutional layer. Through end-to-end backpropagation, it optimizes the convolutional kernel parameters, mapping the original one-dimensional time-domain vibration signal into a high-dimensional time-frequency feature tensor containing rich physical information, thereby capturing the transient frequency texture of the micro-perturbation signal.

3. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, The lightweight U-Net separation module is used to perform blind source separation tasks and adopts an encoder-decoder U-shaped architecture. The convolution operation uses depthwise separable convolution. The shallow high-resolution features of the encoder are directly passed to the decoder through skip connections to preserve the fine structural information of the signal in the spatiotemporal domain and avoid information loss during downsampling. The lightweight U-Net separation module finally outputs several independent feature channels, and the number of feature channels is the preset maximum number of concurrent sources.

4. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, The sparse gating module is located between the lightweight U-Net separation module and the intelligent parallel decoding module, and introduces a learnable soft threshold function to perform sparsification processing on the feature channels.

5. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 3, characterized in that, The intelligent parallel decoding module contains several parallel decoding units, the number of which is the same as the number of feature channels. Each decoding unit has a multi-head self-attention sub-layer to capture the long-term dependencies and global context information of the signal. The end of each decoding unit includes two output branches: a classification head and a regression head. The classification head outputs the confidence probability of the corresponding channel signal belonging to various events, and the regression head outputs the normalized physical location coordinates of the corresponding channel signal on the optical fiber.

6. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, The self-supervised noise distribution pre-training in step S3 specifically includes: Using unlabeled real background noise data or single-source time-series signals processed by random masking as input, a masked signal reconstruction task is performed to initialize the model's noise representation and basic denoising capability. Let the input timing signal be Masking operation yields ,in, For the mask, The timing signal after masking; Deep learning models predict the masked portion from the masked temporal signal. And minimize the reconstruction error by reconstructing the loss: ,in, To reconstruct the loss, For mathematical expectation operators, It is the square of the L2 norm.

7. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, In step S3, the simulation-supervised training enables the deep learning model to decouple target features from aliased signals and map abstract features to physical coordinates. The simulation-supervised training specifically includes: Fully supervised training is performed using the simulation data generated in step S1. The training objectives include two tasks: identity classification and location regression after blind source separation. The cross-entropy loss function is used for the identity classification task, and the Smooth L1 function is used for the location regression task. The expressions are as follows: ； ； in, For classifying losses, Represents the cross-entropy loss function. To categorize identities using real-world tags, For identity classification prediction results, For position loss, To restore the true labels to the location, This is the result of location regression prediction; The joint optimization objective is expressed as: ; in, To jointly optimize objectives, is the weighting coefficient used to balance the classification and localization losses.

8. The optical cable micro-disturbance identification method based on physical simulation and self-supervised temporal decoupling according to claim 1, characterized in that, The spectral domain physical consistency fine-tuning operation in step S3 specifically includes: To alleviate the distribution difference between simulation data and measured data, a two-dimensional Fourier transform is applied to the spatiotemporal matrix output by the lightweight U-Net separation module to obtain a frequency-wavenumber domain representation. Based on the physical relationship of sound wave propagation, a physically allowed region is constructed in the f–k domain. Under the ideal dispersionless approximation, sound waves satisfy a linear dispersion relation, and the corresponding theoretical wavenumber is expressed as: A physical mask is defined based on the linear dispersion relation of sound waves, and is expressed as: ； in, For the allowed spectral width, For physical mask, For frequency, The central theoretical wavenumber is derived based on the linear dispersion relation of sound waves, i.e. , For the spatial frequency variable wavenumber, For the speed of transmission; The spectral domain physical consistency loss is defined as the weighted sum of the out-of-mask energies, expressed as: ； in, Let be the spectral domain physical consistency loss function. The frequency-wavenumber domain representation after two-dimensional Fourier transform and satisfying , For two-dimensional Fourier transform operators, It is a spacetime matrix.

Images